Monthly Archives: November 2013

Examples of embedded systems

Examples of embedded systems

PC Engines’ ALIX.1C Mini-ITX embedded board with an x86 AMD Geode LX 800 together with Compact Flash, miniPCI and PCI slots, 22-pin IDE interface, audio, USB and 256MB RAM

Examples of embedded systems

An embedded RouterBoard 112 with U. FL-RSMA pigtail and R52 miniPCI Wi-Fi card widely used by wireless Internet service providers (WISPs) in the Czech Republic.

Embedded systems span all aspects of modern life and there are many examples of their use.

Telecommunications systems employ numerous embedded systems from telephone switches for the network to mobile phones at the end-user. Computer networking uses dedicated routers and network bridges to route data.

Consumer electronics include personal digital assistants (PDAs), mp3 players, mobile phones, videogame consoles, digital cameras, DVD players, GPS receivers, and printers. Many household appliances, such as microwave ovens, washing machines and dishwashers, are including embedded systems to provide flexibility, efficiency and features. Advanced HVAC systems use networked thermostats to more accurately and efficiently control temperature that can change by time of day and season. Home automation uses wired – and wireless-networking that can be used to control lights, climate, security, audio/visual, surveillance, etc., all of which use embedded devices for sensing and controlling.

Transportation systems from flight to automobiles increasingly use embedded systems. New airplanes contain advanced avionics such as inertial guidance systems and GPS receivers that also have considerable safety requirements. Various electric motors — brushless DC motors, induction motors and DC motors — are using electric/electronic motor controllers. Automobiles, electric vehicles, and hybrid vehicles are increasingly using embedded systems to maximize efficiency and reduce pollution. Other automotive safety systems such as anti-lock braking system (ABS), Electronic Stability Control (ESC/ESP), traction control (TCS) and automatic four-wheel drive.

Medical equipment is continuing to advance with more embedded systems for vital signs monitoring, electronic stethoscopes for amplifying sounds, and various medical imaging (PET, SPECT, CT, MRI) for non-invasive internal inspections.

In addition to commonly described embedded systems based on small computers, a new class of miniature wireless devices called motes are quickly gaining popularity as the field of wireless sensor networking rises. Wireless sensor networking, WSN, makes use of miniaturization made possible by advanced IC design to couple full wireless subsystems to sophisticated sensor, enabling people and companies to measure a myriad of things in the physical world and act on this information through IT monitoring and control systems. These motes are completely self contained, and will typically run off a battery source for many years before the batteries need to be changed or charged.

Knowledge-Based Systems for Intelligent Patient Monitoring and Management in Critical Care Environments


186.1

186.2

Benoit M. Dawant

Vanderbilt University

186.3

Patrick R. Norris

Vanderbilt University 186.4

Introduction

Introduction

Intelligent Patient Monitoring and Management Signal Acquisition and Low-Level Processing • Data Validation • Signal-to-Symbol Transformation and Giving Advice

Moving Toward Computer Architectures for Real-Time Intelligent Monitoring Discussion and Conclusions

Patient monitoring and management in critical care environments such as intensive care units (ICUs) and operating rooms (ORs) involves estimating the status of the patient, reacting to events that may be life-threatening, and taking actions to bring the patient to a desired state. This complex process includes the interaction of physicians and nurses with diverse data (ranging from clinical observations to labora­tory results to online data) provided by bedside medical equipment. New monitoring devices provide health care professionals with unsurpassed amounts of information to support decision making. Ironi­cally, rather than helping these professionals, the amount of information generated and the way the data is presented may overload their cognitive skills and lead to erroneous conclusions and inadequate actions. New solutions are needed to manage and process the continuous flow of information and provide efficient and reliable decision support tools.

Patient monitoring can be conceptually organized in four layers [Coiera, 1993]: (1) the signal level, which acquires and performs low-level processing of raw data; (2) the validation level, which removes data artifacts; (3) the signal-to-symbol transform level, which maps detected features to symbols such as normal, low, or high; and (4) the inference level, which relies on a computer representation of medical knowledge to derive possible diagnoses, explanations of events, predictions about future physiologic states, or to control actions. In addition to these four layers, medical decision support systems need data interfaces to other clinical information systems as well as carefully designed user interfaces to facilitate rapid and accurate situation assessment by care providers.

A number of systems have been developed to address problems faced by clinicians in critical care environments. These range from low-level signal analysis applications for detecting specific features in monitored signals to complete architectures for signal acquisition, processing, interpretation, and decision support. As always, specificity and generality are conflicting requirements. Systems developed for specific applications are usually successful in their limited domain of expertise. However, the lessons and problem­solving strategies learned in one domain are often difficult to generalize. Conversely, generic architectures aim at providing support for modeling and developing a wide range of applications. They also strive for flexibility, modularity, and ease of expansion. This generality is often at the expense of expertise and performance in specific domains [Uckun et al., 1993]. This chapter reviews some of the problems and proposed solutions associated with the four layers of intelligent monitoring systems.

Intelligent Patient Monitoring and Management

Signal Acquisition and Low-Level Processing

Typical signals found in critical care environments include vital signs (ECG, EEG, arterial pressure, intracranial pressure, etc.) and information provided by therapeutic devices, infusion pumps, and drain­age devices. Modern monitoring devices are also capable of providing derived and computed information in addition to raw data. Heavily instrumented patients frequently have up to 20 medical devices moni­toring them, producing up to 100 pieces of clinically relevant information. These instruments are often stand-alone, and interconnection requires developing dedicated software in-house, usually with substan­tial effort. To address this problem, the IEEE standards committee, IEEE 1073, has been working since 1984 on a standard for medical device data communication in critical care environments called the Medical Information Bus (MIB). Although much of the standard has been adopted by ANSI, and despite long-term efforts by individual institutions [Gardner et al., 1992], a standard adhered to by major medical device manufacturers is still lacking. Today, this absence of interconnectivity remains a major obstacle to the development and implementation of intelligent monitoring systems.

Current bedside monitors typically provide instantaneous values for the monitored variables. To complement this information, numerous algorithms have been proposed to detect features in the signals. In particular, detection of significant trends has received much attention [Avent and Charlton, 1990, Haimowitz and Kohane, 1996]. Methods based on median filters [Makivirta et al., 1991] and fuzzy logic [Sittig et al., 1992; Steinman and Adlassnig, 1994] have also been proposed. More recently, techniques such as sub-band adaptive filtering, chaos analysis [Karrakchou et al. 1996], and wavelet transform [Unser and Aldroubi, 1996] have been investigated. Multivariate and data fusion methods have been used to reveal interactions between signals and applied to problems such as ventricular rhythm tracking [Thoraval and Carrault, 1997]. Traditional waveform analysis also continues to be applied to signals such as intracranial pressure [Czosnyka et al., 1996] and arterial blood pressure [Karamanoglu, 1997]. Interpret­ing the output of these algorithms and assessing their impact on patient care is an ongoing task.

Data Validation

The risk of noise contamination, inadequate wiring, or instrument failure is significant, especially as the amount of sampled data increases. Unreliable information can drastically reduce the practitioner’s ability to rapidly assess and act on monitor data. In addition, false alarms due to erroneous or incomplete data reduce clinicians’ confidence in the instrumentation resulting in alarms being disabled or ignored. Tsien and Fackler [1997] report that after 298 monitored hours in a pediatric ICU, 86% of the 2,942 alarms were found to be false-positive alarms while an additional 6% were deemed irrelevant true alarms.

Methods relying on data redundancy and correlation (thus assuming some level of interconnectivity between bedside instruments) as well as contextual information to eliminate false alarms, validate data, and diagnose malfunctions in the monitoring equipment have been proposed. Rule-based systems [Jiang, 1991; van der Aa, 1990] and neural networks [Orr, 1991] have been developed for detecting faults and malfunctions in the breathing circuit during anesthesia. Other intelligent alarming systems focusing on data validation have been implemented in the operating room, and include rule-based systems for respiratory-circulatory management [Garfinkel et al., 1988] and for patient state-dependent data collec­tion and processing [Arcay-Varela and Hernandez-Sande, 1993]. Rule-based systems have been combined with fuzzy logic [Schecke et al., 1992] and neural networks for false alarm reduction [Navabi et al., 1991]. Laursen [1994] investigated causal probabilistic networks to reduce false alarms in the ICU. In addition to systems specifically designed for data validation and false-alarm rejection, diagnostic and advice-giving systems often include a data-validation step in their overall strategy.

Signal-to-Symbol Transformation and Giving Advice

The next layer in information processing involves transforming numerical values into symbolic infor­mation and using these symbols for diagnosis and giving advice. A representative sample of these systems is discussed below. For a more complete compilation the reader is referred to Uckun [1994]. Although a strict taxonomy of diagnostic and advice-giving systems is difficult to create, systems with so-called shallow and deep knowledge will be distinguished. Shallow knowledge expresses empirically based, experiential knowledge [Mora et al., 1993] generally called heuristics. Reasoning in these systems is mostly associative, and knowledge is usually represented in terms of an IF (antecedent) THEN (action) form. Shallow knowledge systems exhibit a high degree of competence in their limited domain of application, but performance usually degrades quickly near the edge of the domain of expertise. At the other end of the spectrum, deep knowledge refers to knowledge about the structure, behavior, or function of a system. Model-based reasoning involves using this knowledge to diagnose, predict, or explain the behavior of the system over time.

Shallow Knowledge Systems

Typically, these systems perform a specific task and have similar architectures: (1) a module responsible for transforming quantitative data into qualitative or symbolic information, (2) a module capable of translating symbolic information into patient state, and (3) a module designed to provide therapeutic advice based on the estimated patient state.

Ventilator management has received much attention in intelligent monitoring research, following the pioneering efforts of Fagan et al. [1984] on the VM project. VM was developed in the late 1970s as an extension of the rule based MYCIN formalism to interpret online quantitative data in the ICU setting and to provide assistance for postoperative patients undergoing mechanical ventilation. VM addressed the importance of interpreting data differently depending on patient status.

COMPAS [Sittig et al., 1989] was designed to assist in management of patients with adult respiratory distress syndrome (ARDS). COMPAS is a rule-based system incorporating a blackboard architecture with three main types of knowledge sources (KSs): (1) the data-processing KSs compute derived information from the primary data; (2) the data-classification KSs transform numerical information into symbolic information; and (3) the protocol-instruction KSs suggest therapeutic actions. These knowledge sources roughly correspond to the upper three layers of intelligent monitoring systems (see above). The raw data level consists of information provided by clinical or blood gas laboratories, x-ray departments, and respiratory therapists. Although COMPAS is no longer used the knowledge it contains has served as a basis for a system used clinically [Henderson et al., 1991].

The AIRS system provides three levels of decision support to match three phases of ventilation [Summers et al., 1993]: (1) the startup phase matches the patient state to one of 14 predetermined states and suggests initial ventilator settings; (2) the maintaining phase provides context-dependent values for set point and trend alarms; and (3) the weaning phase contains knowledge represented as PROLOG premise-actions rules to provide advice in the weaning process.

Vie-Vent [Miksch et al., 1993] is an advice-giving system for the monitoring and management of mechanically ventilated neonates. It consists of three modules: (1) the data-acquisition and validation module performs simple range checking and measurement correlation; (2) the data-abstraction module translates numerical values into normal, low, and high ranges based on predetermined rules; and (3) the rule-based therapy-planning module proposes ventilator settings based on the qualitative values for pH, pCO2, and PO2 provided by the data abstraction module. These rules are adapted to reflect the site of measurement and the mode of ventilation. Vie-Vent is designed to operate in real-time and it includes a data validation component using time-point, time-interval, and trend-based methods [Horn et al., 1997].

PATRICIA [Moret-Bonillo et al., 1993] is a system designed to assist in monitoring and managing mechanically ventilated patients in ICUs. It consists of two main components: the deterministic module, which transforms incoming information into symbolic ranges, and the heuristic module, which interprets these symbolic ranges using rules to infer patient status and prescribe therapeutic advice. It uses so-called natural contexts, which include demographic information, patient history, and diagnosis information, to provide patient-dependent ranges for the symbolic variables. Based on these symbolic values, the rule – based system performs data validation both in the deterministic and the heuristic modules. It is also capable of detecting “prealarm” situations by predicting a temporal evolution of monitored parameters. PATRICIA is currently working off-line and underwent an extensive retrospective validation [Moret – Bonillo et al., 1997]

Other systems developed for ventilator management include ESTER [Hernandez-Sande et al., 1989], KUSIVAR (developed for the management of patients with ARDS) [Rudowski et al., 1989], and WEAN – PRO (designed for assisting physicians in weaning postoperative cardiovascular patients) [Tong, 1991]. Expert-system methodology also has been applied to the monitoring and management of patients in the cardiac care unit (CCU) [Mora et al., 1993, Sukuvaara et al., 1993b].

As opposed to the systems discussed before, GANESH [Dojat et al., 1992] has been designed for use as a closed-loop controller. Its domain of application is the weaning of patients under pressure-supported ventilation (PSV). Based on the respiratory rate, tidal volume, and end-tidal partial pressure of carbon dioxide, a rule-based controller determines the patient status and action to be taken and then acts on the ventilator. Preliminary clinical evaluations of the system have shown its ability to maintain patients in the region of comfort and to progressively reduce the pressure of weanable patients. Recently, the theory of Event Calculus has been proposed to capture and use temporal knowledge for this application [Chittaro and Dojat, 1997].

Blom [1991] also proposes an expert-system approach for closed-loop control of arterial pressure during surgery. In this approach, heuristic rules supervise a PID controller and adjust its parameters. The result is a robust controller capable of coping with artifacts in the signal, low signal-to-noise ratio, and patient sensitivity to nitroprusside, a drug used to achieve controlled hypotension during surgery. Huang and Rozear [1998] also describe a fuzzy logic-based system to control administration of multiple IV drugs for hemodynamic regulation. It consists of: (1) a fuzzy decision analysis module to assign the patient to one of six possible states using information derived from monitor data; (2) a hemodynamic management module to determine the dosage changes of each drug based on the current state; and (3) a therapeutic assessment module for scheduling drug administration based on drug properties such as pharmacological delay.

Deep Knowledge Systems

Despite its advantage over associative reasoning for explanation, robustness, and prediction, model-based reasoning has been used in few medical systems. Uckun [1992] suggests this may be attributed to the relatively shallow level of knowledge regarding most disease processes and the complexity and inherent variability of human anatomy, physiology, and pathophysiology. These difficulties still challenge our ability to develop accurate quantitative, qualitative, or hybrid models. Despite the difficulties involved, several model-based reasoning systems have been proposed for patient monitoring.

Rutledge and colleagues’ VentPlan [1993] recommends ventilator settings for ICU patients. It has three main components: a belief network, a mathematical modeling module, and a plan evaluator. The belief network computes probability distributions for the physiologic model parameters from quantitative and qualitative information. The model predicts the distribution of future values for the dependent variables based on this data and current ventilator settings, and can simulate the evolution of important variables under various ventilator settings to evaluate proposed therapy plans. The plan evaluator predicts paCO2 and O2 delivery to rank therapy plans using a multiattribute scoring system and to propose ventilator settings.

The patient model component of SIMON [Uckun et al., 1993; Dawant et al., 1993] was built on an extension of the qualitative process theory (QPT) [Forbus, 1984] to diagnose neonates suffering from respiratory distress syndrome (RDS), to predict the patient’s evolution, and to dynamically adjust alarm parameters based on patient state. As opposed to QPT, this approach does not exhaustively predict all distinct states the system may reach, since this is of little use when external variables (i. e., ventilator settings) are not constant. It determines patient status based on the temporal evolution of clinical events, and can modify the model based on discrepancies between observed and predicted values. SIMON underwent a limited retrospective evaluation using neonatal ICU data.

KARDIO [Lavrac et al., 1985] and CARDIOLAB [Siregar et al., 1993] are examples of models developed for model-based ECG simulation, prediction, and interpretation. KARDIO generates qualitative descrip­tions of the ECG signals corresponding to various arrhythmia combinations, based on the relation between the ECG and 30 elementary arrhythmias. Observed ECG can then be compared with the generated descriptions, and the associated arrhythmias can be retrieved from the database. Without a model permitting its automatic generation, the creation of an exhaustive database associating arrhythmias with ECG description would be a daunting, if not impossible, task.

Moving Toward Computer Architectures for Real-Time Intelligent Monitoring

Previously highlighted is the fact that patient monitoring is a complex, multifaceted task, and that much research has been done to facilitate the monitoring tasks of clinicians. But despite such efforts, few of the systems described earlier are in clinical use. None of them have been fielded or tested in other institutions. This may be due to the specificity of these systems and to the difficulty of modifying their structure for slightly different applications. Rather than focusing on specific tasks, other researchers are attempting to develop generic architectures that can support the development and implementation of a wider range of applications. These architectures are designed to support diverse reasoning schemes, guarantee real-time response, and permit the integration of the entire spectrum of tasks required for intelligent patient monitoring. This section describes a few such systems.

GUARDIAN [Hayes-Roth et al., 1992] is a blackboard system. Based on a general architecture for intelligent agents, it has been applied in respiratory and cardiovascular monitoring domains. GUARDIAN emerges from a long history of research on reasoning and problem solving, and much emphasis is put on the generic problem of coordinating and selecting one of possible perceptual, cognitive, or action operations under real-time constraints. It integrates heuristic, structural, or functional knowledge of (1) common respiratory problems, (2) respiratory, circulatory, metabolic, and mechanical ventilator systems, and (3) generic flow, diffusion, and metabolic systems. A system capable of instantiating treat­ment protocols based on (1) current patient context, (2) executing plans and close-loop control actions,

Monitoring the execution of these plans and actions, and (4) modifying plan execution as necessitated by patient response has also been designed to work with GUARDIAN.

The Intelligent Cardiovascular Monitor (ICM) [Factor et al., 199l] is an application of the multitrellis software architecture, which supports modeling of real-time tasks in terms of an acyclic hierarchical network of decision processes. In ICM, the lowest level corresponds to raw data, and processes at this level implement signal processing algorithms to extract features from the signal. Information is abstracted along the hierarchy in a way reminiscent of information flow in a blackboard architecture, until reaching the highest level corresponding to physiologic processes or events. As opposed to GUARDIAN, the process trellis architecture assumes sufficient resources to execute the entire program and meet real-time con­straints. This architecture has also been used for the development of DYNASCENE [Cohn et al., 1990], which models hemodynamic abnormalities in terms of “scenes” and temporal relationships between scenes. Each scene corresponds to one physiologic process (e. g., increased pericardial pressure) and is related to modules for computing specific information from the incoming data. DYNASCENE and ICM do not work in conjunction with patient models and hence are not capable of predicting patient behavior.

SIMON [Dawant et al., 1993] is a framework designed to support the development of real-time intelligent patient monitoring systems. It is organized around three main modules: the patient model, the data-acquisition and abstraction module, and a flexible user interface. The patient model is respon­sible for (1) estimating the physiologic state of the patient, (2) predicting the temporal evolution of the monitored variables, and (3) defining a dynamic monitoring context. The context includes normal and abnormal ranges for monitored variables, expected temporal evolutions of these variables, significant values and patterns to be detected in the incoming signals (events), and display configuration information (i. e., critical information to be displayed). The data abstraction module adjusts alarm limits and thresh­olds based on the current context and implements strategies to detect complex events. These may involve multiple variables and temporal relationships between events detected on each variable. SIMON has been applied off-line to the problem of monitoring neonates in the intensive care unit.

Following the recommendations of the INFORM project [Hunter et al., 1991], Sukuvaara et al. [1994] proposed an object-oriented framework to implement systems based on the blackboard approach to problem solving. Data on the blackboard are separated into four main categories: (1) source data (raw data acquired from monitors, hospital systems, or users), (2) preprocessed variables, (3) physiologically interpreted variables, and (4) patient status. Preprocessing knowledge sources (KSs) span the first two levels and extract features from the incoming data. Transformation KSs span the second and third levels in the blackboard hierarchy, and map values of preprocessed variables onto (low, high) intervals. Patient status assessment KSs get input from physiologically interpreted data and determine the patient state. The system also assumes sufficient resources to execute all its knowledge sources and meet real-time constraints.

186.4 Discussion and Conclusions

Intelligent patient monitoring and management are complex tasks involving all aspects of information processing. They require context-dependent acquisition, processing, analysis, and interpretation of large amounts of possibly noisy and incomplete data. These systems also need to interface to hospital-wide patient data management systems (PDMS) to gather data from the laboratories, the pharmacy, or radiology. Finally, these systems should include efficient mechanisms to present relevant and timely information to clinical users.

The systems described here illustrate that it is not easy to choose between specificity and generality. Systems that undergo clinical evaluation are typically developed for a very narrow, specific application which lends itself to rule-based approaches. However, these systems are often of limited clinical utility due to their narrow scope. Research aimed at developing deeper models to infer a wider range of behaviors, on the other hand, has failed to produce overwhelming results. Models for predicting patient behavior or response to therapeutic actions are clearly desirable, but currently are only available in very limited or constrained situations. Original modeling methods and approaches have been proposed, but require assumptions that either drastically reduce the scope of the model or make it unrealistic for clinical use.

Complete architectures aim at identifying the essence of patient monitoring tasks and at providing support for the development of a class of systems. These systems should permit experimentation with a number of problem-solving strategies, the integration of diverse reasoning mechanisms, and should ultimately span the entire spectrum of tasks required for efficient and reliable patient monitoring. However, to move from the research laboratory to the clinical arena, these systems will have to be fitted with rich and flexible modeling tools capable of overcoming the complexity of generic architectures. In addition, they will have to interface with a variety of devices and systems, efficiently manage vast amounts of diverse data, and successfully interact with clinical users.

To address the issue of information presentation, web-based technology is being integrated with clinical information systems, and prototypes have been developed to provide integrated clinical user interfaces to ICU patient data [Norris 1997, Nenov 1995] at or close to real time, via a web browser.

However, the fact remains that despite years of research and development and despite the potential of techniques being developed, intelligent monitoring systems have failed to make a significant impact on the quality of patient care. Saranummi [1997] contends that the main reason for this phenomenon is a lack of interaction between clinicians, industry, and academic researchers. Clinicians and academic researchers must interact closely to define the problem, develop solutions, and evaluate these solutions. When this is done, industry needs to be involved to assess the commercial viability of the project. To foster interaction Saranummi advocates the use of new technical platforms that support “plug-in” mod­ules to existing monitoring and PDMS systems. These modules extend the functionality of existing monitoring systems without interfering with routine operation, permitting the in situ evaluation of new technical solutions. Efforts such as the IMPROVE project sponsored by the European Union, aimed at gathering large annotated data sets of ICU patients, could also facilitate the evaluation of novel infor­mation processing methods in close to real world situations [Nieminen et al., 1997].

Coiera [1994] also suggests that current systems are not adapted to their task. He notes that a major effort has been put into developing systems to assist the physician in making diagnoses and taking therapeutic actions. He suggests, however, that the main problem facing health care professionals is not choosing a diagnosis or therapeutic action, but rather establishing a clear picture of the world leading to these choices. Following this observation, Coiera advocates the creation of a “task layer” designed to provide a good match between displayed information and the cognitive processes of physicians and nurses. More research on the way clinicians interact with intelligent monitoring devices is necessary, though, to support the design of these interfaces.

The design of intelligent monitoring systems is thus an ongoing task. Context-based acquisition, processing, analysis, and display of the information are an essential concept in these systems. Elements defining this context are, however, complex and diverse. On the one hand, patient status and history define an interpretation and monitoring context that can be used to determine normal and abnormal limit values, trends, normal relations between observed variables, and significant events to be detected in the signals. Models could automatically establish these contexts, which, in turn, could be used to configure the monitoring system. On the other hand, cognitive tasks faced by the clinician, such as organ system function, treatment course, or resource utilization define the type and form of information presented. This information-processing context may be more difficult to automate and will require development of better interaction mechanisms between the system and its users.

References

Arcay-Varela B, Hernandez-Sande C. 1993. Adaptive monitoring in the ICU: A dynamic and contextual study. J. Clin. Eng. 18:67.

Avent RK and Charlton JD. 1990. A critical review of trend-detection methodologies for biomedical monitoring systems. Crit. Rev. Biomed. Eng. 17:621.

Blom JA. 1991. Expert control of the arterial blood pressure during surgery. Int. J. Clin. Monit. Comput. 8:25.

Chittaro L and Dojat M. 1997. Using a general theory of time and change in patient monitoring: experiment and evaluation. Compu. Biol. Med. 27:435 Cohn AI, Rosenbaum S, Factor M, and Miller PL. 1990. DYNASCENE: An approach to computer-based intelligent cardiovascular monitoring using sequential clinical “scenes”. Meth. Inform. Med. 29:122. Coiera E. 1993. Intelligent monitoring and control of dynamic physiological systems. Artif. Intell. Med. 5:1. Coiera E. 1994. Design for decision support in a clinical monitoring environment. In: Proceedings of the International Conference on Medical Physics and Biomedical Engineering, Nicosa, Cyprus 130-142. Czosnyka M, Guazzo E, Whitehouse M, et al. 1996. Significance of intracranial pressure waveform analysis after head injury. Acta Neurochirurgica 138:531.

Dawant BM, Uckun S, Manders EJ, and Lindstrom DP. 1993. The SIMON project: Model-based signal acquisition, analysis, and interpretation in intelligent patient monitoring. IEEE Eng. Med. Biol. Mag. 12:92.

Dojat M, Brochard L, Lemaire F, and Harf A. 1992. A knowledge-based system for assisted ventilation of patients in intensive care units. Int. J. Clin. Monit. Comput. 9:239.

Dojat M, Pachet F, Guessoum Z, et al. 1997. NeoGanesh: a working system for the automated control of assisted ventilation in ICUs. Artif. Intell. Med. 11:97.

Factor M, Gelertner DH, Kolb CE, et al. 1991. Real-time data fusion in the intensive care unit. IEEE Computer 24:45.

Fagan LM, Shortliffe EH, Buchanan BG. 1984. Computer-based medical decision making: From MYCIN to VM. In: Readings in Medical Artificial Intelligence: The First Decade, WJ Clancey and EH Shortliffe, (Eds.), pp. 241-255. Reading, MA, Addison-Wesley.

Forbus KD. 1984. Qualitative Process Theory. Ph. D. Dissertation, Massachusetts Institute of Technology, Cambridge, MA.

Gardner RM, Hawley WL, East TD, et al. 1991. Real time data acquisition: Recommendations for the Medical Information Bus (MIB). Int. J. Clin. Monit. Comput. 8:251.

Garfinkel D, Matsiras P, Lecky JH, et al. 1988. PONI: An intelligent alarm system for respiratory and circulation management in the operating rooms. In: Proceedings of the 12th Symposium on Computer Applications in Medical Care, pp. 13-17.

Haimowitz IJ and Kohane IS. 1996. Managing temporal worlds for medical trend diagnosis. Artif. Intell. Med. 8:229.

Hayes-Roth B, Washington R, Ash D, et al. 1992. Guardian: A prototype intelligent agent for intensive – care monitoring. Artif. Intell. Med. 4:165.

Henderson S, Crapo RO, Wallace CJ, et al. 1991. Performance of computerized protocols for the man­agement of arterial oxygenation in an intensive care unit. Int. J. Clin. Monit. Comput. 8:271.

Hernandez-Sande C, Moret-Bonillo V, Alonso-Betanzos A. 1989. ESTER: An expert system for manage­ment of respiratory weaning therapy. IEEE Trans. Biomed. Eng. 36:559.

Horn W, Miksch S, Egghart G, et al. 1997. Effective data validation of high-frequency data: Time-point-, time-interval-, and trend-based methods. Comput. Biol. Med. 27:389.

Huang JW and Rozear MP. 1998. Multiple drug hemodynamic control using fuzzy decision theory. IEEE Trans. Biomed. Eng. 45:213.

Hunter J, Chambrin MC, Collinson P, et al. 1991. INFORM: Integrated support for decisions and activities in intensive care. Int. J. Clin. Monit. Comput. 8:189.

Jiang A. 1991. The design and development of a knowledge-based ventilatory and respiratory monitoring system. Ph. D. Dissertation, Vanderbilt University, Nashville, TN.

Karamanoglu M. 1997. A system for analysis of arterial blood pressure waveforms in humans. Comput. Biomed. Res. 30:244.

Karrakchou M, Vibe-Rheymer K, Vesin J-M, et al. 1996. Improving cardiovascular monitoring through modern techniques. IEEE Eng. Med. Biol. Mag. 1996:68.

Laursen P. 1994. Event detection on patient monitoring data using causal probabilistic networks. Meth. Inform. Med. 33:111.

Lavrac N, Bratko I, Mozetic I, et al. 1985. KARDIO-E: An expert system for electrocardiographic diagnosis of cardiac arrhythmias. Expert Syst. 2:46.

Makivirta A, Koski E, Kari A, and Sukuvaara T. 1991. The median filter as a preprocessor for a patient monitor limit alarm system in intensive care. Comput. Methods. Progr. Biomed. 34:139.

Miksch S, Horn W, Popow C, Paky F. 1993. VIE-VENT: Knowledge-based monitoring and therapy planning of the artificial ventilator of newborn infants. In: Artificial Intelligence in Medicine: Proceedings of the European Conference. (AIME), S Andreassen, et al. (Eds.) pp. 218-229. Amster­dam, IOS Press.

Mora A, Passariello G, Carrault G, Le Pichon J-P. 1993. Intelligent patient monitoring and management systems: A review. IEEE Eng. Med. Biol. Mag. 12(4):23.

Moret-Bonillo V, Alonso-Betanzos A, Martin EG, et al. 1993. The PATRICIA project: A semantic-based methodology for intelligent monitoring in the ICU. IEEE Eng. Med. Biol. Mag. 12:59.

Moret-Bonillo V, Mosqueira-Rey E, Alonso-Betanzos A. 1997. Information analysis and validation of intelligent monitoring systems in intensive care units. IEEE Trans. Inf. Tech. Biomed. 1:89.

Navabi MJ, Watt RC, Hameroff SR, and Mylrea KC. 1991. Integrated monitoring can detect critical events and improve alarm accuracy. J. Clin. Eng. 16:295.

Nenov VI, Klopp J. 1995. Remote analysis of physiologic data from neurosurgical ICU patients. J. Amer. Med. Informatics. Assoc. 2:273.

Nieminen K, Langford RM, Morgan CJ, et al. 1997. A clinical description of the IMPROVE data library. IEEE Eng. Med. Biol. Mag. 16(6):25.

Norris PR, Dawant BM, Geissbuhler A. 1997. Web-Based Data Integration and Annotation in the Intensive Care Unit. In: Proceedings of the American Medical Informatics Association Annual Fall Symposium, pp. 794-798.

Orr JA. 1991. An anesthesia alarm system based on neural networks. Ph. D. dissertation, University of Utah, Salt Lake City, UT.

Rudowski R, Frostell C, and Gill H. 1989. A knowledge-based support system for mechanical ventilation of the lungs: The KUSIVAR concept and prototype. Comput. Methods. Progr. Biomed. 30:59.

Rutledge GW, Thomsen GE, Farr BR, et al. 1993. The design and implementation of a ventilator­management advisor. Artif. Intell. Med. 5:67.

Saranummi N, Korhonen I, van Gils M, and Kari A. 1997. Framework for biosignal interpretation in intensive care and anesthesia. Meth. Inform. Med. 36:340.

Schecke T, Langen M, Popp HJ, et al. 1992. Knowledge-based decision support for patient monitoring in cardioanesthesia. Int. J. Clin. Monit. Comput. 9:1.

Siregar P, Coatrieux JL, and Mabo P. 1993. How can deep knowledge can be used in CCU monitoring. IEEE Eng. Med. Biol. Mag. 12(4):92.

Sittig DF, Pace NL, Gardner RM, et al. 1989. Implementation of a computerized patient advice system using the HELP clinical information system. Comput. Biomed. Res. 22:474.

Sittig DF, Cheung KH, Berman L. 1992. Fuzzy classification of hemodynamic trends and artifacts: Experiments with the heart rate. Int. J. Clin. Monit. Comput. 9:251.

Steinman F and Adlassnig K-P. 1994. Clinical monitoring with fuzzy automata. Fuzzy Sets Syst. 61:37.

Sukuvaara T, Sydanma M, Nieminen H, et al. 1993. Object-oriented implementation of an architecture for patient monitoring. IEEE Eng. Med. Biol. Mag. 12:69.

Sukuvaara T, Koski E, and Mдkivirta A. 1993. A knowledge-based alarm system for monitoring cardiac operated patients: Technical construction and evaluation. Int. J. Clin. Monit. Comput. 10:117.

Summers R, Carson ER, and Cramp DG. 1993. Ventilator management: The role of knowledge-based technology. IEEE Eng. Med. Biol. Mag. 12(4):50.

Thoraval L, Carrault G, Schleich, JM, et al. 1997. Data fusion of electrophysiological and hemodynamic signals for ventricular rhythm tracking. IEEE Eng. Med. Biol. Mag. 16(6):48.

Tong DA. 1991. Weaning patients from mechanical ventilation: A knowledge-based system approach. Comput. Meth. Progr. Biomed. 35:267.

Tsien, CL and Fackler, JC. 1997. Poor prognosis for existing monitors in the intensive care unit. Crit. Care. Med. 25:614.

Uckun S. 1992. Model-based reasoning in biomedicine. Crit. Rev. Biomed. Eng. 19(4):261.

Uckun S, Dawant BM, Lindstrom DP. 1993. Model-based diagnosis in intensive care monitoring: The YAQ approach. Artif. Intell. Med. 5:31.

Uckun S. 1994. Intelligent systems in patient monitoring and therapy management: A survey of research projects. Int. J. Clin. Monit. Comput. 11:241.

Uckun S. 1996. Instantiating and monitoring skeletal treatment plans. Meth. Inform. Med. 35:324.

Unser M and Aldroubi A. 1996. A review of wavelets in biomedical applications. Proc. IEEE 84(4):626.

Van der Aa JJ. 1990. Intelligent alarms in anesthesia: A real time expert system application. Ph. D. dissertation, Eindhoven University of Technology.

Kerkhof, P. L. M. “Medical Terminology and Diagnosis Using Knowledge Bases.” The Biomedical Engineering Handbook: Second Edition.

Ed. Joseph D. Bronzino

Boca Raton: CRC Press LLC, 2000

Embedded system

Embedded system

Picture of the internals of a Netgear ADSL modem/router. A modern example of an embedded system. Labelled parts include a microprocessor (4), RAM (6), and flash memory (7).

An embedded system is a computer system designed to perform one or a few dedicated functions, often with real-time computing constraints. It is usually embedded as part of a complete device including hardware and mechanical parts. In contrast, a general-purpose computer, such as a personal computer, can do many different tasks depending on programming. Embedded systems control many of the common devices in use today.

Since the embedded system is dedicated to specific tasks, design engineers can optimize it, reducing the size and cost of the product, or increasing the reliability and performance. Some embedded systems are mass-produced, benefiting from economies of scale.

Physically, embedded systems range from portable devices such as digital watches and MP3 players, to large stationary installations like traffic lights, factory controllers, or the systems controlling nuclear power plants. Complexity varies from low, with a single microcontroller chip, to very high with multiple units, peripherals and networks mounted inside a large chassis or enclosure.

In general, "embedded system" is not an exactly defined term, as many systems have some element of programmability. For example, Handheld computers share some elements with embedded systems — such as the operating systems and microprocessors which power them — but are not truly embedded systems, because they allow different applications to be loaded and peripherals to be connected.

Non-electronic logic

Unconventional computing

It is possible to construct non-electronic digital mechanisms. In principle, any technology capable of representing discrete states and representing logic operations could be used to build mechanical logic. MIT students Erlyne Gee, Edward Hardebeck, Danny Hillis (co­author of The Connection Machine), Margaret Minsky and brothers Barry and Brian Silverman, built two working computers from Tinker toys, string, a brick, and a sharpened pencil. The Tinkertoy computer is supposed to be in the Houston Museum of Natural Science.

Hydraulic, pneumatic and mechanical versions of logic gates exist and are used in situations where electricity cannot be used. The first two types are considered under the heading of fluidics. One application of fluidic logic is in military hardware that is likely to be exposed to a nuclear electromagnetic pulse (nuclear EMP, or NEMP) that would destroy electrical circuits.

Mechanical logic is frequently used in inexpensive controllers, such as those in washing machines. Famously, the first computer design, by Charles Babbage, was designed to use mechanical logic. Mechanical logic might also be used in very small computers that could be built by nanotechnology.

Another example is that if two particular enzymes are required to prevent the construction of a particular protein, this is the equivalent of a biological "NAND" gate.

Knowledge Acquisition and Representation

Introduction

Medical Expertise: Domain and Control

Knowledge

Domain Knowledge • Control Knowledge

Knowledge Acquisition

Human-Centered Approaches • Data-driven Approaches

Knowledge Representation

Basic Representation Schemes • Advanced Representation Catherine Garbay Schemes • Second Generation Systems

Laboratoire TIMC/IMAG 185.5 Conclusion

Introduction

The scope of designing knowledge-based systems in medicine has evolved during these last few years, from pure decision making automata to open environments able to acquire, formalize, validate, as well as handle and disseminate elements of the human expertise. The application fields have expanded in a correlated manner, from the sole field of symbolic processing to data analysis, information management, simulation, and planning.

Therefore, the scope of knowledge acquisition and representation has been shifted from the mere transcription of expert knowledge to the modeling of pathophysiological phenomena of increasing complexity [Stefanelli, 1993]. To reflect such evolution, the notion of models is made central in this chapter and the nature and specificity of medical expertise is addressed in the first section. Knowledge acquisition is considered afterwards, by successively considering human-centered and data-driven approaches to knowledge extraction. Knowledge representation issues are finally discussed in the frame­work of their evolution from basic representation schemes to second generation systems.

Medical Expertise: Domain and Control Knowledge

Expert knowledge is usually described as involving domain knowledge and control knowledge [Schreiber, 1993]. Domain knowledge is meant to represent the various elements, e. g., facts, features, relations, and concepts, that are used to represent a case, while control knowledge is meant to represent the expert know-how, that is, the way to reason about a case.

Domain Knowledge

Domain knowledge may be seen as a kind of formal knowledge tying names, properties and relations to the whole empirium of data, facts, observations, hypotheses or results that are extracted from a case.

Knowledge Acquisition and Representation

FIGURE 185.1 Vertigo taxonomy (partial view). (Source. Schmid [1987].)

Knowledge Acquisition and Representation

FIGURE 185.2 A “part of” hierarchy example (partial view). (Source: Garbay and Pesty [1988].)

Domain ontologies are used to structure and organize domain knowledge according to well-identified knowledge types or relations.

Taxonomic ontologies have been widely used in medicine because of their capacity to classify and organize diseases into hierarchical structures. As may be seen iN Fig. 185.1, Paroxismal vertigo is classified as a particular kind of vertigo, and may be further characterized as benign paroxismal vertigo of child­hood, benign paroxismal positional vertigo, or paroxismal vertigo due to vascular disorder [Schmid, 1987].

“Part of” hierarchies make it possible to describe the compositional characteristics of concepts, and are particularly useful for describing anatomical properties. As may be seen in Fig. 185.2, A mammary gland lobule may be described as composed of stroma cells and acini, which in turn are composed of a basement membrane, epithelial and myoepithelial cells, and a lumen [Garbay and Pesty, 1988].

Knowledge may also be organized according to its role in the reasoning process. According to Patel et al. [1989], this process is decomposed into several levels, from the empirium (the whole universe of medical problem solving data) to the global complex level (complete understanding of the patient case). Several intermediate levels are distinguished in terms of observations, findings, facets (complex gathering of findings understood as the consequence of a pathophysiologic process) and diagnostic, as summarized in Fig. 185.3.

FIGURE 185.3 Organizing knowledge according to its role in the reasoning process. (Source. Patel et al. [1989].)

подпись: 
figure 185.3 organizing knowledge according to its role in the reasoning process. (source. patel et al. [1989].)
Current research emphasizes the notion of causal ontologies (see Fig. 185.11), where a deep functional understanding of the process leading to a disease is sought [Jensen et al., 1987; Ste – fanelli, 1993].

Control Knowledge

Reasoning may be seen as the process of handling facts, observations and hypotheses, to reach a final understanding of a situation. Var­ious mental abstraction levels are traversed in this process, which
deduction ^

Abstraction Problem features

подпись: abstraction problem features – Data

Induction

T

Hypotheses

Abduction

FIGURE 185.4 An epistemological model of diagnostic reasoning. (Source. Lanzola and Stefanelli [1992].)

May involve different inference schemes, depending on the type of data to be processed (e. g., uncertain, geometric or temporal), and on the type of inferencing to be applied (e. g., deductive, inductive, or non­monotonic).

An epistemological model of diagnostic reasoning is proposed by Lanzola and Stefanelli [1992] in terms of abstraction, abduction, deduction, and eliminative induction (Fig. 185.4). The abstraction corresponds to identification of the clinical features that diagnosis should be able to explain. The abduc­tion corresponds to formulation of a set of hypotheses from which expected manifestations are deduced. Eliminative induction is finally used to reduce uncertainty of inferred hypotheses by matching their consequences against observed data.

The global approach to a case nowadays is often formalized in terms of tasks [Bylander and Chan – drasekaran, 1987; Steels, 1990; Schreiber et al., 1993], which makes it possible to structure the overall reasoning process in terms of goals and sub-goals, and thus to organize knowledge and know-how according to a problem-solving rationale [Newell, 1982].

Knowledge Acquisition

Knowledge acquisition has been traditionally considered as the process of extracting and transcribing the expert knowledge into a computer-usable form. This complex process has been shown to critically depend on the knowledge engineering tool at hand, since most representation formalisms convey restrictive and rigid semantics, which do not refer to a clear ontology [Schreiber, 1993]. Current trends are to consider knowledge acquisition as an integral part of the designing process, centered on the notion of a model.

Knowledge discovery, data mining, and machine learning techniques have recently attracted consid­erable attention, due to the growing amount of data available, and to the growing necessity to base the reasoning on evidence taken from physical measurements. Data-driven approaches to knowledge extrac­tion have been developed as a consequence, complementing the more traditional human-centered approaches, by enabling systems to create new knowledge, update existing knowledge, and improve their performance without intervention and reprogramming.

Human-Centered Approaches

Interviewing Techniques

Knowledge is usually acquired through direct interaction with a human expert, but may also be obtained from many other sources, such as textbooks, reports, databases, or case studies [Buchanan, 1983]. A systematic and structured approach is necessary to ensure the completion and consistency of the inter­viewing process, which is usually driven by a knowledge engineer. Other techniques are often used to complement the mere interviewing approach, such as observing the expert in daily working situations, writing scenarios and reports, or filling out questionnaires.

Knowledge Acquisition Tools

Several computerized tools have been designed in the past to assist the knowledge acquisition/transcrip­tion process. The reference work done in TEIRESIAS [Davis and Lenat, 1982] and ETS [Boose, 1985]
appears to bring complementary solutions to this difficult problem. TEIRESIAS is rather meant to support the testing and refinement of a “MYCIN like” system. It proceeds by searching for missing or inconsistent information and may suggest the generation of new inference rules. ETS, on the contrary, is meant to support direct elicitation of expert knowledge. It is based on bipolar constructs structuring the perception of the world (e. g., the grade of a tumor allows the organization of diseases according to their prognostic significance). The acquired knowledge is then represented by means of so-called repertory grids tying concepts and their properties. Distances between concepts may then be computed, as well as the dis­criminatory power of a given characteristic. Production rules are then generated and submitted to the expert to be modified or tested.

Knowledge Acquisition Environments

Knowledge acquisition environments have been designed to integrate several acquisition techniques. The expert is usually assisted by a cognitive engineer to use the software environment. AQUINAS [Boose and Bradshaw, 1987] and MORE [Kahn et al., 1985] are particularly representative of this approach. AQUINAS is based on the cooperation between several interviewing methods and has been designed as an extension of ETS. MORE may interview the expert according to eight different strategies and integrate several facilities such as a rule generator, an inference engine to test the rules, and online assistance to the user.

Knowledge Acquisition Methodologies

The term “knowledge acquisition methodologies” refers to approaches aiming at a close integration between knowledge acquisition and system design, in which knowledge acquisition is considered as a modeling process rather than as an extraction/transcription problem. The modeling approach permits the investigator to work at the “right” abstraction level, often called the knowledge level [Newell, 1982], and to refer a level at which knowledge may be expressed without referring to any implementation constraint [Schreiber et al., 1993]. It is formulated in opposition to symbol level approaches, which support implementation level modeling.

KADS (KBS Analysis and Design Structured Methodology) is among the best known methodologies [Breuker and Wielinga, 1987], and has been conceived to guide the whole process of expert system design. It permits the progressive evolution from real world to implementation by the building of four successive models that organize the domain ontology into four knowledge levels: domain level (description of concepts and the relationship between them), inference level (reasoning primitives), task level (reasoning steps, goals and ways to reach them, structured task) and strategy level (control and goal scheduling).

Data-Driven Approaches

Knowledge Discovery and Data Mining

Knowledge discovery may be defined as “the non-trivial extraction of implicit, previously unknown, and potentially useful information from given data” [W. J. Frawley et al., 1991]. Such a process makes use of various techniques, involving segmentation, filtering, statistical as well as model-based analysis, or super­vised induction.

It is most often considered as a pre-processing step in the design of decision-support systems, by providing computerized support to the identification of salient attributes, to the computation of repre­sentative values for these attributes, or to the structurization of the database into semantically meaningful clusters, thus improving performance. Factor analysis may, for example, be used in this respect. Such an approach has been used in a system devoted to the development of in vitro fertilization treatment plans, where a case-based reasoning module has been extended by a knowledge mining component [Juristica et al., 1998].

Learning

Three main approaches to machine learning are usually distinguished, i. e., inductive learning, analogical reasoning and explanation-based learning. Inductive learning techniques have been widely used in medicine, in the framework of decision-trees or version space/conceptual clustering methods, and more recently by the introduction of the evolutionary and connectionist paradigms.

A typical example of inductive learning is ID3, a decision-tree learning technique developed by Quinlan [1986]. ID3 constructs a classification tree in a recursive, top-down, divide and conquer fashion, using the concept of entropy as a quantitative measure of the “information” conveyed by each attribute. Conversely to ID3 which proceeds by division, the approach of version space/conceptual clustering proceeds in a bottom-up way, by grouping sets of domain concepts to form more general ones.

Genetic-based learning has been conceived as an extension of the evolutionary programming para­digm: it has been designed as a generate-and-test process that maintains a pool of competing problem­solving procedures (most often decision rules), tests them to a limited extent, creates new ones from those that performed well in the past, and prunes poor ones from the pool. Connectionist approaches are revealed to be particularly useful in domains where it is impossible to structure and represent the domain knowledge in logic form.

Machine learning is now introduced in an increasing number of systems and application domains in medicine, especially in domains where the knowledge is difficult to acquire by hand. Current approaches make use of real data and domain models to drive the knowledge acquisition process.

Hau and Coiera [1997] for example, describe a learning system that takes real-time patient data obtained during cardiac bypass surgery to create models of normal and abnormal cardiac physiology. Conversely, qualitative models of heart disease were used early in Kardio to learn diagnostic rules [Bratko et al., 1989]. Examples of “faulty” and “non-faulty” behaviors are generated by the model and transmitted to a learning program, which can then learn diagnostic rules for each particular fault.

Machine learning may also be used to refine existing knowledge bases in cases where there is a

Disagreement between the expert-provided domain knowledge and actual cases. An application to acute

Abdominal pain has been experimented with for effective performance improvement [Dzeroski et al., 1997], using the theory revision system NEITHER. As a consequence of this process, rule R1, for example, has been specialized into rules R6 and R7:

Rule R1: OPERATE ^ (RIGIDITY YES)

Rules R6: OPERATE ^ (RIGIDITY YES) (TYPE STEADY) (SITE-RUQ NO) rules R7: OPERATE ^ (RIGIDITY YES) (TENDER-RLQ SURFACE)

Knowledge Representation

The representation of medical knowledge is a very active research field, characterized by a wide range of tools, models, and languages, which, together with the presence of increasing computer capabilities, allows one to specify and emulate systems of a growing complexity. Knowledge representation schemes indeed have known an important evolution, from basic schemes supporting a rather heuristic approach, to advanced schemes involving a deeper consideration of the various dependencies between knowledge elements. Second genera­tion systems have finally been introduced to cope with complex integration and modeling constraints.

Basic Representation Schemes

Basic schemes for the representation of domain and control knowledge are considered in this section. The representation of concepts, their structure and properties is considered first. How to model the expert way of reasoning is examined afterwards. It is examined successively as the ability to apply well-experi­mented inference schemes (usually represented by rules) and as the capacity to compare and retrieve learned situations (so-called case-based reasoning). One main feature of human reasoning, e. g., its capacity to handle uncertain and imprecise information, is finally considered.

Frame Representations and Object-Oriented Languages

Frame-based representations have been introduced as a way to organize prototypical knowledge about the world, and are currently among the most widely used tools for the construction of AI systems. A

Frame Cytomegalovirus

Sort-of

Epidemiological data

Other kind of pneumonia

CD4

Clinical picture

Symptoms

Signs

Laboratory tests

P02

LDH

Intestitial pneumonia yes

Less than 200/mmc

Fever, non-productive cough, dyspnea cyanosis tachypnea

Lowering

Growing

FIGURE 185.5 Using frames to represent diagnostic knowledge. (Source: Fiore et al. [1993]).

Frame defines the prototypical description of concepts sharing similar properties and behavior. It is defined as a set of slots describing the concept attributes and their values. A variety of other slots may be introduced, including procedures, default and current values, or type/value constraints. The frames are organized in hierarchies, tying classes and instances. The basic inference mechanism is instantiation, in which the attribute values of the new instance are obtained either by inheritance, by computation (using the procedural attachments), or by default.

In a knowledge-based system to diagnose HIV-pneumonias [Fiore et al., 1993], frames are used to describe the physiopathological states of the disease for each particular pneumonia, in terms of several slots describing the epidemiological data, clinical picture, laboratory tests, and diagnosis (Fig. 185.5).

The object-oriented paradigm is nowadays increasingly considered as a way to cope simultaneously with knowledge modeling and software engineering issues, as in HELIOS [Coignard et al., 1992]. HELIOS has been conceived as a software engineering environment to facilitate the development of medical applications. It provides a core set of medical classes, from which to build a new application. Several services are also provided, which include the acquisition and management of medical information, the processing of natural language queries, as well as data driven decision support.

Semantic Networks and Conceptual Graphs

Semantic Networks [Quillian, 1968] have been proposed as a flexible formalism to represent the semantics of concepts (the nodes in the network) and their relations (the arcs) in the framework of a graph-like structure. The reasoning proceeds by unification between an unknown fact and a known concept or subgraph in the network.

The approach has rapidly been extended to represent causal relationhips as well as hierarchical tax­onomies and to include frames. It has been applied successfully to numerous applications in medicine, as in CASNET [Weiss et al., 1978], a consultation system for glaucoma, INTERNIST [Pople, 1977], a diagnostic consultant in internal medicine, and PIP [Pauker et al., 1976] a system devoted to renal disease. The reasoning strategy in these systems is of the event-driven type: initial data triggers a number of hypotheses, which are then to be confirmed [Kulikowski, 1984]. The subgraph of confirmed or undeter­mined hypotheses constitute a patient-specific model. Various weights and scores are usually introduced to render the reasoning strategy more flexible.

Conceptual graphs [Sowa, 1984] have been designed as an extension of the previous formalism in which an explicit representation of the links between concepts is sought. A conceptual graph is a directed graph comprising two kinds of nodes: concept nodes and conceptual relationship nodes. The represen­tation is grounded on first-order logic and is currently being applied in a number of projects for the modeling and representation of medical terminology, as in GALEN [Alpay et al., 1993]. One major goal of GALEN (Generalized Architecture for Languages Encyclopedias and Nomenclatures in Medicine) is to form a general Terminology Server able to conciliate a number of classifications and nomenclatures. It is based on a Semantical Encyclopedia of Terminology connected to a Multilingual Information Module.

Knowledge Acquisition and Representation

| simple

FIGURE 185.6 Conceptual graph representing the concept “simple oblique diaphyseal fracture of the right femur.” (Source: Bernauer and Goldberg [1993]).

Conceptual graphs are used in this project to generate natural language expressions from information

Coded according to the GALEN Master Notation.

Conceptual graphs may also be used as a formalism that supports both expressiveness and classification purposes, i. e., as a knowledge representation formalism able to integrate both terminological and domain knowledge, as shown in Fig. 185.6 [BErnauer and Goldberg, 1993].

In the example in Fig. 185.6, tHe composite concept “simple oblique diaphyseal fracture of the right femur” is described as a nested set of 2-tuples each tying a relation (:loc, for example) and a composite

Concept (diaphyis, for example). Taxonomic as well as compositional relations are used to support

Classification. The equivalent nested notation for the graph in Fig. 185.6 Is the following:

(fracture

(:loc (diaphysis (:part (femur (:side (right))))))

(:morph (oblique))

(:compl (simple)))

Rule-Based Reasoning

The use of if-then rules (also called production rules or condition-action pairs) is a straightforward and popular way to represent the expert know-how. A rule example is given in Fig. 185.7, iN which a deduction is made about the presence of bacteroids based on a composite observation.

This type of knowledge is often said to be shallow or heuristic, since it is a largely empirical and nonformalized representation mechanism. Given a set of facts, the reasoning process is then modeled as the successive activation of rules, which in turn produce new facts to be considered, until no applicable rule can be found.

The rules are handled by means of an inference engine, which may work under two inferencing modes, the data driven or goal driven mode. In the data-driven mode, the reasoning is modeled as a deductive process proceeding from a given premise to some conclusion. In the goal-driven mode, on the contrary,

And the suspected portal of entry of the organism is the gastro-intestinal tract

THEN there is a suggestive evidence (0.7) that the identity of the organism is

The site of the culture is one of the nonsterile sites

IF

And

THEN

подпись: if
and
then
There are rules which mention in their premise a previous organism

Which may be the same as the current organism

It is definite (1.0) that each of them is not going to be useful

FIGURE 185.8 A meta-rule example from MYCIN. (Source. Shortlife [1976].)

The reasoning is modeled as an inductive process proceeding backwards from some hypothesized con –

Elusion to the conditions to be verified. Both inference schemes are usually combined to implement

Complex hypotheses and test strategies. In addition, metarules are often used to structure the reasoning process by constraining the application of the rules. A metarule example is given in Fig. 185.8, Which permits one to predict that some rules will be poorly informative, given the non-sterility of the culture site.

MYCIN, a system for diagnosing and treating infectious blood diseases, is one of the earliest and most widely known expert system in medicine [Shortliffe, 1976].

Uncertain and Imprecise Reasoning

Knowledge Acquisition and RepresentationHow to model the uncertainty and imprecision that is central to the human reasoning style has raised a number of research activities during the last decades [van Ginneken and Smeulders, 1988]. The Bayesian model has been for a long time the primary numerical approach to uncertain reasoning. It is based on assigning a probability distribution to each of the variables representing the problem at hand. These probabilities express the uncertainty and likelihood of occurrence of symptoms as well as diseases. Given P(Di), the a priori probability of disease Di, given P(SIDi), the probability that symptom S may occur in the context of disease Di, Bayes’ rule (1) allows one to compute P(DiIS), the probability that disease Di occurs when S is observed, according to the following formula:

(185.1)

The most important assumption underlying the use of Bayes’ rule is that all disease hypotheses must be mutually exclusive and exhaustive. It means that only one disease is assumed to be present. If this condition is met, Bayes has a completely predictable performance and guarantees a conclusion with minimum overall error [van Ginneken and Smeulders, 1988].

Some intuitive methods for the combination and propagation of these numbers have been suggested and used, as in MYCIN (see the rules in Figs. 185.7 And 185.8, Where coefficients are used to weight their conclusions). The empirical character of the approach has often been criticized. However, it is consistent with the heuristic style that is inherent to rule-based modeling.

The theory of possibility [Zadeh, 1978] has been used in medicine to represent the vagueness of clinical predicates. These vague predicates are represented by means of a fuzzy set and a possibility distribution (or membership function). Fuzzy reasoning may then be implemented by means of rules handling fuzzy facts.

Examples of membership function modeling criteria for the classification of rheumatoid arthritis are given iN Fig. 185.9 [Leitich et al., 1996]. Seven criteria are defined in total, from which at least four must be present in conjunction to establish a diagnosis.

Case-Based Reasoning

CBR (Case-based reasoning) emerged in recent years as a powerful problem-solving technique applicable to a wide range of tasks in Artificial Intelligence. It is based on the hypothesis that new situations are analyzed by reference to past experience, i. e., by considering their similarity to well-experimented situ­ations, a notion that is made central in this kind of reasoning [Campbell and Wolstencroft, 1990]. New problem-solution pairs may then be learned as a consequence: this inherent combination of problem

Criterion Sj

Membership function |isi(x)

0

If t <15

S: morning stiffness

(t-15)/45

If 15 < t < 60

Lasting at least one hour

Hsi« =

1

If t >60

V

If t is unknown

Where t is the duration of morning

Stiffness in minutes

0

If n < 1

S2- arthritis of 3

0.5

If n = 2

Or more joint areas

1

If n > 3

V

If n is unknown

Where n is the number of involved joint areas

FIGURE 185.9 Membership functions examples in the field of rheumatoid arthritis. (Source. [Leitich et al., 1996].)

Knowledge Acquisition and Representation

Revise

FIGURE 185.10 The Case-based reasoning cycle, a general view. (Source. [Aamodt and Plaza, 1994].)

Solving with learning through problem solving experience gives a particular strength to CBR over most other methods [Aamodt and Plaza, 1994].

Case-based reasoning finally appears as a cycle comprising four stages, as illustrated in Fig. 185.10: retrieve the most similar case, reuse the retrieved information to solve the problem by analogical rea­soning, revise the proposed solution, and finally retain the part of the experience likely to be useful for future problem solving.

The advantage of the approach is two-fold. first of all, it is based on the description of experiences, rather than on the modeling of generic or abstract knowledge, second, such a system may evolve in an incremental way, as cases grow [Goos and Schewe, 1993]. However, the choice of the similarity measure turns out to be critical with respect to the reasoning accuracy. The structurization of the case base also has to be carefully considered in order to avoid combinatorial search [Gierl et al., 1993].

Current trends [Reateguiet al., 1997] are to combine CBR with complementary forms of reasoning such as rule-based, model-based, or neural networks (NN) to solve diagnosis problems. The NN here are used during the consultation process to make hypotheses of possible diagnosis solutions and to guide the search of similar cases, restricting the research to cases with similar interpretations (e. g., diagnoses).

Advanced Representation Schemes

Recent approaches to knowledge representation consider as central the capacity to better situate the reasoning with respect to contextual, causal, and temporal modeling of knowledge elements. A number of knowledge domains are concerned, involving therapy and monitoring, pathophysiology, or planning.

Context-Based Reasoning

It is now well recognized that advanced medical decision systems should situate their reasoning with respect to the context in which a problem is considered, including the patient clinical context, but also the chronology of events together with their causal relationships.

In connection with the development of the knowledge-based system TheMPO (Therapy Management in Pediatric Oncology), which supports therapy and monitoring in pediatric oncology, a graph-grammar approach has been used to design and implement a graph-oriented patient model which allows the representation of non-trivial (causal, temporal etc.) clinical contexts [Mьller et al., 1996]. A graphical interface has been designed to facilitate the specification and retrieval of contexts by the physician.

Causal Reasoning

The last decade has known the development of solid foundations for diagnostic systems employing logic, probability theory, and set theory, thus providing rich and robust modeling tools to the developer, and allowing him to cope with the increasing necessity to base the diagnosis on models of the disease process and models of the structure and function of the human body [Lucas, 1997].

There are at least two alternative ways to represent causal knowledge [Ramoni et al., 1992]. The first method is based on a network representation, the second implies the representation of both the structure and behavior of a physiological system.

Network representations, called CPN (causal probabilistic netwoks) or BBN (Bayesian belief networks), are in the form of a direct acyclic graph in which nodes represent stochastic variables and arcs represent conditional dependencies among the variables. These representations make explicit the dependency and independency assumptions among variables. They are based on sound semantics and easily extend to compact representations called influence diagrams [Ramoni et al., 1995].

MUNIN [Jensen et al., 1987], an expert system in electromyography, is representative of this approach. The domain knowledge is embedded in a causal probabilistic network and is further divided into three levels, representing diseases, pathophysiological features, and findings. These levels are linked by causal relations: diseases cause certain affections in muscles. These affections, in turn, cause expectations for certain findings. A partial view of the causal network is given in Fig. 185.11. Diseases are grouped into a single node (on the left), and characterized by their grades. The disease node is connected to several pathophysiologic nodes describing the pathophysiologic status of a muscle in terms of discrete properties (the muscle structure is rather normal, there is no myotonia, and no muscle loss). Some expectations regarding, for example, the muscle force and atrophy or the presence of spontaneous myotonic distrophy are finally obtained. Probabilities are used to characterize each state in the network (the horizontal bars in the figure) and propagated through the causal links.

Despite its effective representation power, the approach still bears strong limitations, due to the large amount of information required: current propagation algorithms require that all the conditional prob­abilities defining a conditional dependency be known, together with all the a priori probabilities attached to the root variables. MUNIN, for example, consists of about 1100 discrete random variables linked together, thus rendering its managing hardly tractable by current computer technology. Several methods have been investigated by the author of MUNIN to reduce the memory space and calculation time [Suojanen et al., 1997]. A methodology for automatically inducing Bayesian network has been introduced by Larranaga et al. [1997], based on genetic algorithms, and applied to the prediction of survival in malignant skin melanoma.

An alternative way is to represent both the structure and behavior of a physiological system [Ironi et al., 1993]. The system structure is simply given in terms of variables and relations. The system behavior is then modeled by a set of mathematical equations tying related variables, which represent potential
Perturbations of the system state. Compartmental system theory is often used in this respect because it provides a robust modeling framework based on differential calculus. It implies, however, that a precise quantitative modeling of the pathophysiologic phenomena is possible. Qualitative models, on the con­trary, allow one to cope with the fuzziness and incompleteness of pathophysiological knowledge. Among these, QSIM is the most widely applied formalism in medicine [Stefanelli, 1993]. This formalism provides a descriptive language to represent the structure of a physiologic system and a simulation algorithm to infer its qualitative behavior. The language consists in qualitative constraints that abstract the relationships in a differential equation. This approach has been applied to a variety of fields in medicine [Ironi et al., 1990]

It should be noted that representing pathophysiologic knowledge requires a special emphasis on the notion of time [Stefanelli, 1993], a problem early addressed by Fagan [1979]. In addition, specific knowledge acquisition issues are raised that have been only recently addressed [Ironi et al., 1993].

Temporal Knowledge

Time constitutes an integral and important aspect of medical concepts, and its explicit modeling is increasingly considered as central to the design of advanced medical systems. However, such modeling still remains challenging, due to the necessity to consider compound objects (e. g., disorders, treatments, or patient states) exhibiting different temporal existences and complex interactions, through mechanisms that are not completely understood. The definition of adapted temporal ontologies conveying a clear semantic is currently an important subject of debate among the community [Keravnou, 1995; Shahar and Musen, 1998].

The ability to reason about time is known to depend on the ability to provide short, informative, and context-sensitive summaries of time-oriented clinical data, in the form of temporal abstractions able to summarize clinical features over a period of time (Fig. 185.12). Such an approach has been considered in RESUME [Shahar and Musen, 1998], a system for forming high-level concepts from raw time-oriented clinical data which has been applied in three domains: monitoring of children’s growth, care of diabetes patients, and protocol-based care.

In RESUME, a temporal abstraction is characterized as a quadruplet: {(parameter, value, context), interval}, meaning that a given parameter has exhibited a certain value in a certain context for a certain

Knowledge Acquisition and Representation

FIGURE 185.12 A temporal-abstraction task example. an evolution of the platelet and granulocyte counts is observed for a patient suffering from chronic graft-versus-host disease (CGVHD). The observation starts with a bone marrow transplantation (BMT), and is performed in a certain open context interval (shaded arrow). The solid bars represent abstracted intervals figuring the evolution of the myelotoxicity grade M(n). (Source. Shahar and Musen, 1998].)

Knowledge Acquisition and Representation

FIGURE 185.13 The temporal abstraction task. Grey arrow: decomposed into, thin arrow: solved by, bold arrow: used-by (Source. Shahar and Musen, 1998].)

Period of time. The temporal abstraction task is modeled as a complex task decomposed into specific subtasks and problem-solving mechanisms, themselves depending on various knowledge types (Fig. 185.13). Such decomposition defines a general problem solving method (as defined in the section titled “Task Oriented Modeling”) for interpreting data in time-oriented domains.

Current efforts concern the integration of temporal and causal modeling, thus allowing the modeling of multiple interacting mechanisms operating over a variety of time periods. Such models have been applied to the domain of heart disease [Long, 1996] and to the interpretation of time-series of blood glucose measurements coming from home monitoring [Riva and Bellazzi, 1996]. However, severe prob­lems come from the computational burden induced by the learning and the managing of these models.

Protocols & Guidelines

Various attempts have been made to design knowledge representation systems which are capable of capturing and handling clinical procedures, from the early work on ONCOCYN [Tu et al., 1989]. The

Knowledge Acquisition and Representation

FIGURE 185.14 Model of a prescription activity in primary care (partial view). (Source: Gordon et al. [1993].)

Objectives are to guide the user in the accurate and correct application of protocols but also to disseminate shared consensus guidelines among the community.

A generic modeling of clinical protocol has been developed in DILEMMA [Gordon et al., 1993], a system delivering guideline-based decision support for shared care in oncology, cardiology, and primary care. A protocol is a formal description of a way to achieve an objective or goal, given a class of problems. It is used as a template from which to derive the specific activity to be performed for a given clinical context. Figure 185.14 depicts a data model of a prescription activity, in which precise information is given about the patient, the agent responsible for the drug prescription and administration, together with administration guidelines, and a description of potential side effects and expected consequences.

The development of standard languages for guideline modeling is witnessing a growing interest, from mere procedural approaches like the Arden syntax [Hripsack et al., 1994] to high-level representation languages based on task ontologies like in the Proforma approach [Fox et al., 1997].

Current developments concern the integration of time-dependent patient information like in EON (Musen et al., 1995), as well as the development of web-based environments, the aim being to produce clinical guidelines which can be widely shared between humans from different institutions, thus dealing with patient and organization preferences [Quaglini et al., 1997]. GLIF (GuideLine Interchange Format), for example, has been introduced to allow the sharing of representations among different sites [Patel et al., 1997].

Second Generation Systems

Second generation systems organize their knowledge into modules and multilevel structures to solve problems of increasing complexity. The objectives are (i) to better structure the knowledge, facilitate its management and allow its reuse, and (ii) to allow the design of hybrid systems combining heterogeneous knowledge and reasoning schemes.

Task-Oriented Modeling

Task-oriented modeling has been proposed to model general problem solving methods, based on two classes of reusable components [Bylander and Chandrasekaran, 1987]: (i) domain-independent problem­solving methods, e. g., standard methods to perform prototypical tasks (see Fig. 185.13) And

Domain ontologies, e. g., description of the main concepts and their relations (see Fig. 185.15). These generic components are then instantiated to cope with the specificity of given expertise domains. This notion may be used as a model to drive knowledge acquisition and system design, since knowledge and

Knowledge Acquisition and Representation

FIGURE 185.15 A domain ontology example, from FreeCall, a system for emergency-call-handling support. (Source. Post et al. [1996].) Note that the knowledge is organized according to the classification proposed by Patel et al. [1989] (see Fig. 185.3).

Know-how specific to a task are specified at the right abstraction level [Steels, 1990]. Several knowledge acquisition tools specific to different task types may indeed be used to guide the expert through a structured elicitation protocol, for that specific problem type. Second generation expert systems also allow the reuse of knowledge, since some knowledge is generic and can be applied across several domains.

Hybrid Systems

There is a growing need for the integration of multimodal knowledge representation and reasoning schemes. Such integration is needed for at least two complementary reasons, i. e., the need to analyze data from different modalities and abstraction levels and the need to dynamically adapt the knowledge representation formalism and inference scheme to the situation at hand and to the current problem solving state.

The first viewpoint is adopted in AMNESIA [O’Hare et al., 1992], a system devoted to the diagnosis of memory related disorder illnesses. In such systems, the knowledge and reasoning skills are distributed among several specialists that cooperate towards the solving of the problem, each specialist being devoted to handling a precise knowledge domain or task.

The second viewpoint has been emphasized in NEOANEMIA, a system able to recognize disorders causing anemia [Ramoni et al., 1992]. An abductive inference scheme is first of all applied, to generate an initial set of hypotheses. A deductive inference scheme is then applied to compute the expected manifestations, which is based on a separate and explicit representation of taxonomic and causal ontology.

Normal

Demyelination

Conduction

Demyelination

Without cond. block

Axonal loss

Structural

Attribute

Functional

Attribute

2

Abstraction/refinement

Anatomical

Concept

подпись: anatomical
concept

Nerve

Fibre

подпись: nerve
fibre

Normal I

Demyelination

Conduction velocity

Demyelination

Amplitude

Without cond. block

Axonal loss

Nerve

Bundle

Abstraction/refinement

FIGURE 185.16 A hybrid knowledge representation scheme in the field of EMG diagnosis. (Source: Cruz et al. [1997].)

Inductive reasoning is finally used to prune the set of hypotheses, by matching their expected manifes­tations against the available data.

Structural, functional, and causal knowledge have been combined at various abstraction levels in a system dedicated to EMG diagnosis [Cruz et al., 1997]. Various anatomical structures like the nervous fiber and the nervous bundle are described in terms of their state and function, as illustrated in Fig. 185.16. Abstraction and refinement operations are associated with these attributes and defined in a probabilistic way to cope with the non-deterministic character of this knowledge.

Changes caused by normal or abnormal functioning of structures, as well as those caused by external agents, are modeled by means of processes expressed at the appropriate level of abstraction. A local neuropathy for example is a lesion that causes deep pathologic processes of local demyelination and axon loss, as modeled below:

Process: local neuropathy

Causes: neuropathy effects: segmentkl of fiber bundleFB

(pointi of nerveN, severity) normal = N

Constraints: transversal dispersion demyelination = D

Longitudinal dispersion demyelination without cond. block = DB

Axonal loss = A

To characterize a neuropathy depends on the proportion of damaged fibers, on one hand, and on the extension of the damage along the nerve fibers, on the other hand. Transversal and longitudinal dispersion functions are moreover used to compute the lesion severity. The diagnosis finally proceeds through successive hypothesis refinement steps, progressively exploiting the various knowledge elements at hand.

185.5 Conclusion

The acquisition and representation of knowledge has been presented as an actively evolving research field characterized by modeling and software engineering issues of increasing complexity. Representing domain ontologies has been shown to raise a variety of issues, such as using conceptual graphs to represent the semantics of medical terms, using frames to describe domain taxonomies or modeling pathophysiological processes through causal networks. Modeling the expert control knowledge has been considered in terms of reasoning as well as planning abilities, which in turn leads to specific modeling issues.

In addition, the scope of designing knowledge-based systems in medicine currently evolves from the mere diagnostic task to the broader issue of patient management, which implies a better integration in existing hospital information systems.

New challenges are now to be faced, due to the rapid development of the internet communication facilities, which increase the possibility of communication and cooperation among health care profes­sionals. It is clear, however, that the lack of a shared vocabulary and a shared understanding of medical terminology is currently impairing such dissemination [Lenat, 1995]. A stable ontological foundation is needed to allow the sharing of knowledge between healthcare professionals [Rossi Mori et al., 1997].

Defining Terms

Control knowledge: The know-how by which to handle domain knowledge and to reason about a case.

Domain knowledge: Formal knowledge about the application domain theory as well as actual knowl­

Edge about the data, facts, observations, hypotheses, or results that are attached to a given case. Generic modeling, Generic task: A way to describe a concept, or a task, that is domain dependent in

The sense that it is related to a medical domain, but that is generic in the sense that it may apply to several application areas in medicine.

Heuristic knowledge: Knowledge that reflects an empirical and non-formalized way to reason about

A situation or solve a problem.

Knowledge level: A notion introduced by Newell, to express the rationale behind the exploitation of

Knowledge, in terms that are free from any implementation constraint.

Model, Modelling approach: A formal representation of the concepts, relationships, and reasoning

Schemes that constitue the domain and control knowledge of the application domain at hand; a designing approach that is driven by a formal representation of the domain and control knowledge, in which a close integration of the knowledge acquisition process is performed. This approach allows one to work at a level that is free from any implementation constraints, the knowledge level. Ontology (domain, taxonomic, causal): The domain ontology characterizes the structure of the

Domain knowledge, in terms of concepts and relationships. Taxonomic and causal ontologies are two major methods for organizing domain knowledge; taxonomic ontologies emphasize the hier­archical relations between concepts, while causal ontologies emphasize their deep causal relations. Shallow (vs deep) knowledge: Shallow knowledge is often used in opposition to deep knowledge, a

Knowledge based on a formal theory of the domain of interest (knowledge of a pathophysiological process for example). The term shallow refers to shortened reasoning schemes, that are often gained by experience.

References

Aamodt A, Plaza E. 1994. Case-based reasoning: foundational issues, methodological variation, and system approaches, AI Commun, 7(1): 39-59.

Alpay L, Baud R, Rassinoux AM, et al. 1993. Interfacing Conceptual Graphs and the Galen Master Notation for Medical Knowledge Representation and Modelling. In: S Andreassen et al., (Eds.), Artificial Intelligence in Medicine, pp. 337-347, Amsterdam, IOS Press.

Bernauer J, Goldberg H. 1993. Compositional Clasification based on Conceptual Graphs. In: Artificial Intelligence in Medicine, S. Andreassen, R. Engelbrecht and J. Wyatt, (Eds.) pp. 348-359. Amsterdam, IOS Press.

Boose JH, Bradshaw JM. 1987. Expertise Transfer and Complex Problems: Using AQUINAS as a Knowl­edge Acquisition Workbench for Knowledge-Based System, Int. J. Man-Mach. Stud., 26:3-28. Boose JH. 1985. A Knowledge Acquisition Program for Expert Systems Based on Personal Construct Psychology. Int. J. Man-Mach. Stud., 23: 495-525.

Bratko I, Mozetic I, Lavrac N. 1989. KARDIO: A Study in Deep and Qualitative Knowledge for Expert Systems. Cambridge, MA: MIT Press.

Breuker J, Wielinga B. 1987. Use of Models in Interpreting Verbal Data. In: Knowledge Elicitation for Expert Systems: a Practical Handbook. A. Kidd. (Ed.) pp. 17-44. Plenum Press.

Buchanan BG, Barstow D, Bechtel et al. 1983. Construction of an Expert System. In: Building Expert Systems. F. Hayes-Roth, D. A. Waterman and D. B. Lenat. (Eds.) pp. 127-167. Reading, MA, Addison – Wesley Publishing Company, Inc.

Bylander T, Chandrasekaran B. 1987. Generic Tasks for Knowledge-Based Reasoning: the “Right” Level of Abstraction for Knowledge Acquisition. Int. J. Man-Mach. Stud., 26:231-243.

Campbell JA, Wolstencroft J. 1990. Structure and Significance of Analogical Reasoning. Art. Intell. Med.. 2:103-118.

Coignard J, Jean F, Degoulet P, et al. 1992. HELIOS: Object-Oriented Software Engineering in Medicine, In: Health Technology and Informatics. vol. 2. pp. 141-149. Amsterdam, IOS Press.

Cruz J, Barahona P. 1997. A causal functional model applied to EMG diagnosis. In: AIME 97: Proc. Euro. Conf. Art. Intell. Med., Lecture Notes in Artificial Intelligence. E Keravnou et al. (Eds.), pp. 249-260. Berlin, Springer-Verlag.

Davis R, Lenat D. 1982. Knowledge-Based System in Artificial Intelligence. New York, MacGraw-Hill.

Dzeroski S, Potamias G, Moustakis V, et al. 1997. Automated revision of expert rules for treating acute abdominal pain in children. In: E Keravnou et al. (Eds.), AIME 97: Proc. Euro. Conf. Art. Intell. Art. Intell. Med., Lecture Notes in Artificial Intelligence. pp. 98-109. Berlin, Springer-Verlag.

Fagan LM. 1979. Representation of Dynamic Clinical Knowledge: Measurement Interpretation in the Intensive Care Unit. In: Proc. of the 6th IJCAI. pp. 260-262. Stanford, CA, Stanford University, Department of Computer Science.

Fiore M, Sicurello F, Vigano M et al. 1993. A Knowledge-Based System to Classify and Diagnose HIV – Pneumonias. In: Artificial Intelligence in Medicine, S. Andreassen, R. Engelbrecht and J. Wyatt, (Eds.), pp. 185-190. Amsterdam, IOS Press, Inc.

Fox J, Johns N, Rahmanzadeh A. 1997. Protocols for medical procedures and therapies: a provisional description of the Proforma language and tools, In: AIME 97: Proc. Euro. Conf. Art. Intell. Med., Lecture Notes in Artificial Intelligence. E Keravnou et al., (Eds.) pp. 21-38. Berlin, Springer-Verlag.

Frawley W, Piatesky-Shapiro G, Matheus C. 1991. Knowledge discovery in databases: an overview, In: Knowledge Discovery in Databases, G Piatessky-Shapiro, W Frawley (Eds.), pp. 1-27, MIT Press.

Garbay C, Pesty S. 1988. Expert Systems for Biomedical Image Interpretation. In: Artificial Intelligence and Cognitive Sciences, J. Demongeot, T. Hervй, V. Rialle, and C. Roche (Eds.) pp. 323-345. Manchester University Press.

Gierl L, Schmidt R, Pollwein B. 1993. ICONS: Cognitive Basic Functions in a Case-Based Consultation System for Health Care. In: Artificial Intelligence in Medicine, S. Andreassen, R. Engelbrecht and J. Wyatt, (Eds.) pp. 230-236. Amsterdam, IOS Press, Inc.

Goos K, Schewe S. 1993. Case Based Reasoning in Clinical Evaluation. In: Artificial Intelligence in Medicine,

S. Andreassen, R. Engelbrecht and J. Wyatt (Eds.), pp. 445-448. Amsterdam, IOS Press.

Gordon C, Herbert SI, Jackson-Smale A et al. 1993. Care Protocols and Health Informatics. In: Art. Intell. Med., S. Andreassen, R. Engelbrecht and J. Wyatt (Eds.), pp. 289-309. Amsterdam, IOS Press, Inc.

Hau DT, Coiera EW. 1997. Learning Qualitative Models of Dynamic Systems, Mach. Learn., 26:177-212.

Hripsack G, Ludeman P, Pryor TA et al. 1994. Rationale for the Arden syntax, Comp. Biomed. Res., 27:291-324.

Ironi L, Cattaneo A, Stefanelli M. 1993. A Tool for Pathophysiological Knowledge Acquisition. In: Artificial Intelligence in Medicine, S. Andreassen, R. Engelbrecht and J. Wyatt (Eds.), pp. 13-31. Amsterdam, IOS Press, Inc.

Ironi L, Stefanelli M, Lanzola G. 1990. Qualitative Models in Medical Diagnosis. Art. Intell. Med. 2: 85-101.

Jensen FV, Andersen SK, Kjaerulff U et al. 1987. MUNIN – On the Case for Probabilities in Medical Expert Systems – a Practical Excercise. In: AIME 87: Proc. Euro. Conf. Art. Intell. Med., Lecture Notes in Medical Informatics. J. Fox, M. Fieschi and R. Engelbrecht (Eds.). pp. 149-160. Berlin Heidelberg, G. Springer-Verlag.

Juristica I, Mylopoulos J, Glasgow J et al. 1998. Case-based reasoning in IVF: prediction and knowledge mining, Art. Intell, 12: 1-24.

Kahn G, Nowlan S, McDermott J. 1985. MORE: an Intelligent Knowledge Acquisition Tool Proc. AAAI ’85. pp. 581-584.

Keravnou ET. 1995. Modelling medical concepts as time-objects. In: AIME 95: Proc. Euro. Conf. Art. Intell.

Med. P. Barahona et al. (Eds.) pp. 67-78. Berlin, Springer-Verlag.

Kulikowski CA. 1984. Artificial Intelligence Methods and Systems for Medical Consultation. In: Readings in Medical Artificial Intelligence. W. J. Clancey and E. H. Shortliffe (Eds.). p.72-97. Reading, MA, Addison-Wesley Publishing Company.

Lanzola G, Stefanelli M. 1992. A specialized Framework for Medical Diagnosis Knowledge-Based Systems.

Comput. Biomed. Res. 25:351-365.

Larranaga P, Sierra B, Gallego MJ, et al. 1997. Learning Bayesian networks by genetic algorithm: a case study in the prediction of survival in malignant skin melanoma. In: AIME 97: Proc. Euro. Conf. Art. Intell. Med. E Keravnou et al. (Eds.) pp. 261-272. Berlin, Springer-Verlag.

Leitich H, Adlassning K-P, Kolarz G. 1996. Development and evaluation of fuzzy criteria for the diagnosis of rheumatoid arthritis, Meth. Inform. Med., 35: 334-342.

Lenat DB. 1995. Steps to sharing knowledge, In: Towards Very Large Knowledge Bases. NJI Mars (Ed.) pp. 3­

IOS Press.

Long W. 1996. Temporal reasoning for diagnosis in a causal probabilistic knowledge base, Art. Intell. Med. 8: 193-215.

Lucas PJF. 1997. Model-based diagnosis in medicine, Art. Intell. Med., 10:201-208.

Mc Graw KL, Harbison-Briggs K. 1989. Knowledge Acquisition: Principles and Guidelines. New Jersey, Prentice-Hall.

Mьller R, Thews O, Rohrbach C, et al. 1996. A Graph-Grammar Approach to Represent Causal, Temporal and Other Contexts in an Oncological Patient Record. Meth. Inform. Med., vol. 35 (2): 127-141. Musen MA, Tu SW, Das AK et al. 1995. A component-based architecture for automation of protocol – directed therapy. In: AIME 95: Proc. Euro. Conf. Art. Intell. Med. P. Barahona et al. (Eds.) pp. 3-13. Berlin, Springer-Verlag.

Newell A. 1982. The knowledge level. Art. Intell., 18:87-127.

O’Hare et al. 1992. AMNESIA—Implementing a Distributed KBS using RAPIDO. In: Proc. 12th. Ann. Inter. Conf. of the British Computer Society Specialist Group on ES. Bramer and Milne (Eds.). pp. 1-29, Cambridge University Press.

Patel VL, Allen VG, Arocha JF et al. 1997. Representing Clinical Guidelines in GLIF: Individual and Collaborative Expertise. SMI Report Number: SMI-97-0694.

Patel VL, Evans DA, Kaufman, D. R. 1989. A Cognitive Framework for Doctor-Patient Interaction. In: Cognitive Science in Medicine. D. A. Evans and V. L. Patel (Eds.). pp. 257-312. MIT Press.

Pauker SG, Gorry GA, Kassirer JP et al. 1976. Towards the Simulation of Clinical Cognition – Taking a Present Illness by Computer. Am. J. Med. 60:981-996.

Pople H. 1977. The Formation of Composite Hypothesis in Diagnostic Problem Solving: an Exercise in Synthetic Reasoning. In: Proc. of the 5th IJCAI. pp. 1030-1037. Pittsburgh, PA, Carnegie-Mellon University, Department of Computer Science.

Post WM, Koster W, Sramek M et al. 1996. FreeCall, a system for emergency-call-handling support, Meth.

Inform. Med., 35(3):242-255.

Quaglini S, Saracco R, Stefanelli M et al. 1997. Supporting tools for guideline development and dissem­ination. In: AIME 97: Proc. Euro. Conf. Art. Intell. Med. E Keravnou et al. (Eds.) pp. 39-50. Berlin, Springer-Verlag.

Quillian MR. 1968. Semantic Memory. In: Semantic Information Processing. M. Minsky (Ed.). p.216-270, Cambridge, MA, MIT Press.

Quinlan JR. 1986. Induction of Decision Trees, Mach. Learn. 1 (1):81-106.

Ramoni M, Riva A, Stefanelli M, et al. 1995. Medical decision making using ignorant influence diagrams. In: AIME 95: Proc. Euro. Conf. Art. Intell. Med. P Barahona et al. (Eds.) pp. 139-150. Berlin, Springer-Verlag.

Ramoni M, Stefanelli M, Barosi G et al. 1992. An Epistemological Framework for Medical Knowledge – based Systems. IEEE Trans. on Sys, Man and Cyb. 22:1361-1375.

Reategui EB, Campbell JA, Leao BF. 1997. Combining a neural network with case-based reasoning in a diagnostic system, Art. Intell. Med. 9:5-27.

Riva A, Bellazzi R. 1996. Learning temporal probabilistic causal models form longitudinal data, Art. Intell. Med. 8: 217-234.

Rossi Mori A, Gangemi A, Steve G, et al. 1997. An ontological analysis of surgical deeds. In: AIME 97: Proc. Euro. Conf. Art. Intell. Med. E Keravnou et al. (Eds.) pp. 361-372. Berlin, Springer-Verlag. Schmid R. 1987. An Expert System for the Classification of Dizziness and Vertigo. In: AIME 87: Proc. Euro. Conf. Art. Intell. Med., Lecture Notes in Medical Informatics. J. Fox, M. Fieschi and R. Engel­brecht (Eds.). pp. 45-53. Berlin, Heidelberg, Germany, Springer-Verlag.

Schreiber ATh, van Heijst G, Lanzola G et al. 1993. Knowledge Organisation in Medical KBS Construction. In: Artificial Intelligence Medicine, S. Andreassen, R. Engelbrecht and J. Wyatt (Eds.), pp. 394-405. Amsterdam, IOS Press, Inc.

Shahar Y, Musen MA. 1996. Knowledge-based temporal abstraction in clinical domains, Art. Intell. 8: 267-298.

Shortliffe EH. 1976. Computer-Based Medical Consultations: MYCIN. New York, Elsevier.

Sowa JF. 1984. Conceptual Structures: Information Processing in Mind and Machine. Reading, MA, Addison- Wesley.

Steels L. 1990. Components of Expertise. AI Magazine. Summer 1990.

Stefanelli M. 1993. European Research Efforts in Medical Knowledge-Based Systems. Art. Intell. Med. 5:107-124.

Suojanen M, Olesen KG, Andreassen S. 1997. A method for diagnosing in large medical expert systems based on causal probabilistic networks. In: AIME 97: Proc. Euro. Conf. Art. Intell. Med. E Keravnou et al. (Eds.) pp. 285-295. Berlin, Springer-Verlag.

Tu SW, Kahn MG, Musen MA, et al. 1989. Episodic skeletal plan-refinement on temporal data, Commu­nication of the ACM, 32:1439-1455. van Ginneken AM, Smeulders AWM. 1988. An Analysis of Five Strategies for Reasoning in Uncertainties and their Suitability in Pathology. In: Pattern Recognition and Artificial Intelligence. E. S. Gelsema and L. N. Kanal LN (Eds.). pp. 367-379. North Holland, Elsevier Science Publisher B. V.

Weiss SM, Kulikowski C, Safir, A. 1978. Glaucoma Consultation by Computer. Comp. Biol. Med. 8(1):25. Wyatt J. 1997. Computer-assisted decision support. In: Internet, Telematics and Health, M Sosa-Iudicissa et al. (Eds.) pp. 229-237. IOS Press.

Further Information

Artificial Intelligence in Medicine is a bimonthly journal published by Elsevier Science Publishers B. V. This international journal publishes articles concerning the theory and practice of medical artificial intelligence.

Studies in Health Technology and Informatics is a series published by IOS Press, in which two volumes are of particular interest to the reader interested in the new trends in health telematics:

Volume 16, edited by C. Gordon and J. P. Christensen, entitled Health Telematics for Clinical Guidelines and Protocols (1995);

Volume 36, edited by M. Sosa-Iudicissa et al., entitled Internet, Telematics and Health (1997).

Lecture Notes in Artificial Intelligence is a sub-series of Lecture Notes in Computer Science published by Springer-Verlag, in which the proceedings of the 5th and the 6th Conference on Artificial Intelligence in Medicine Europe, edited by P. Barahona et al. in 1995, and E. Keravnou et al. in 1997, gives an excellent outlook of the most recent advances of the field throughout Europe.

Dawant, B. M., Norris, P. R. “Knowledge-Based Systems for Intelligent Patient Monitoring and Management in Critical Care Environments"

The Biomedical Engineering Handbook: Second Edition.

Ed. Joseph D. Bronzino

Boca Raton: CRC Press LLC, 2000