PhD Thesis information

Knowledge Management and Decision Support in Diabetes Care through Multi Modal Reasoning
Extended Abstract


In order to overcome the limitations shown by classical Expert Systems, today the trend in decision support systems definition in the medical domain is the one of integrating the existing Hospital Information System (HIS) with different methodologies and technologies, able to co-operate in a transparent way with respect to the user, and to promote information exchange among the human agents involved in the patients management process. This goal is achieved by coupling the decision support functionality with the Knowledge Management (KM) task, being KM a discipline for promoting knowledge growth, communication and preservation within an organization. Instead of being stand-alone solutions to afford particular problems, decision support systems can today be considered as building blocks of the HIS within a larger KM perspective.

In medical applications, two knowledge types are generally available: explicit knowledge, corresponding to the already well established domain knowledge, and implicit knowledge, consisting of individual expertise, organizational practices, and past cases. Both implicit and explicit knowledge need to be managed and maintained, in order to support the needs and the complexity of medical decision making. To take advantage from the exploitation of both implicit and explicit knowledge, the recent advances in Information Technology have led to the design of KM and decision support systems able to integrate a set of information sources, relying on different knowledge representation formalisms. Various reasoning methodologies can then be applied, defining a Multi Modal Reasoning (MMR) strategy meant to overcome the single approaches limitations, and to better cope with the target domain problems. Particular attention has received the combination of Rule Based Reasoning (RBR) and Case Based Reasoning (CBR), being rules the most successful knowledge representation formalism for intelligent systems, and being CBR well suited for integration with formalisms grounded on general, declarative knowledge. In particular, CBR is a reasoning paradigm able to exploit the information embedded into already solved instances of problems, called cases. The common basis of most of the MMR approaches described in the literature is that CBR and RBR are used in a quite exclusive way. On the contrary, only a very tight integration, taking place within the general problem solving cycle, can effectively overcome the intrinsic weaknesses that the two reasoning paradigms may show. Realizing a tight integration between RBR and CBR has been our aim while defining and implementing the MMR system described in this PhD thesis, meant to perform KM in the context of type 1 diabetic patients management, and to provide decision support to physicians when dealing with therapy revision.

Diabetes Mellitus is a major chronic disease in the industrialized countries. In particular, type 1 diabetic patients patients need insulin injections to regulate blood glucose metabolism, in order to prevent acute episodes, such as ketoacidosis and coma, as well as later life invalidating complications. Intensive Insulin Therapy, consisting in 3 to 4 injections every day, or in the use of sub-cutaneous insulin pumps, may reduce the outlined risks, but, on the other hand, increases costs and the risk of severe hypoglycemias. Optimizing insulin administration and therapy implementation is therefore a complex task: patients are required to collect their monitoring data (e.g. Blood Glucose Level measurements and insulin doses) several times a day. Moreover, every 2-4 months they undergo control visits, during which physicians are expected to asses their metabolic condition, and, if needed, to adjust the current insulin therapy. In order to revise the therapeutic protocol, physicians may rely both on structured knowledge (e.g. on insulin pharmacodynamics), and on previous experience (e.g. the observation that a certain protocol has been applied on that patient or to patients with similar characteristics in the past, with a particular outcome). Integration among RBR and CBR seems therefore a natural solution when building a decision support system to be applied in this context.

From an implementation perspective, the reasoning process carried out by our system consists in a series of tasks, whose execution is scheduled by the RBR system. First, the possible patient's metabolic alterations are identified from the monitoring data; then, a set of reliable suggestions to cope with them are proposed, and finally the most suitable suggestions are applied to the current insulin therapy. RBR is meant to be general enough to be safely applicable to any patient category. Therefore, it typically proposes quantitatively small changes to the insulin therapy presently adopted by the patient.

On the other hand, rule parameters can be tuned on the basis of past experience and of specific features, extracted from Case Based retrieval, thus tailoring the therapeutic advice to the single patient's needs. In particular, the case library has been structured resorting to a taxonomy of mutually exclusive prototypical classes, that express typical problems that may occur to type 1 diabetic patients in the age of infancy and puberty. Therefore, before retrieving past cases similar to the current one, the system classifies the current case as belonging to one of such classes. Classification implements context detection: the patient is categorized as being experiencing a particular clinical course condition or associated disease, and her/his data can be better interpreted, in the light of the specific situation she/he is currently living. Moreover, classification is crucial in making efficient the retrieval by restricting search only to relevant parts of the whole case library. Retrieval can in fact be limited to the most probable class, or to a subset of all the classes in the taxonomy. Classification and retrieval results are then exploited to specialize the behavior of the metabolic alteration identification task and of the therapy adjustment task respectively.

The clinical correctness of the suggestions proposed by the MMR system has been tested through a formal evaluation procedure. In particular, two diabetologists were asked to perform a fully-crossed, blind review of the therapies proposed by RBR, by MMR, and by two colleagues, to 30 real patient cases. 98/118 therapies were judged as acceptable for the physicians, and 101/120 for the systems. The MMR and the RBR tools hence proved to be able to work at the physician level; on the other hand, no benefit could be found by exploiting MMR in comparison to RBR. The reason for such an unsatisfactory result was due to a too small case library, not sufficiently representative of some patients categories (presence of "competence gaps"). To cope with the problem of misleading retrieval information, we plan to implement a control strategy, that will enable the exploitation of retrieval results only if a sufficiently large number of similar cases has been retrieved, and if they are similar enough to the input case to justify their use. In this way, MMR will support decision making at various levels of complexity, as the case library grows: first, when the stored information is poor, RBR will come out to be the most competent methodology, and it will be applied without relying on CBR results. As far as new information is stored in the case library, retrieval results will become more reliable, and will be exploited by the MMR methodology.

Moreover, by simply introducing the system into clinical practice, it will be possible to enrich the case library content with further information, since new patients' cases will automatically be stored during every day activity, without requiring an additional workload from physicians. In this way, the MMR therapeutic suggestions reliability will be enhanced; our application is in fact a (anytime) continuously learning system, whose competence is increased as new data are collected.

The system was developed in the context of the EU-funded project T-IDDM (HC-1047), and will be exploited as a KM and decision support functionality within EU-funded project M2DM (IST-1999-10315).

Back