A decision support system for medical diagnosis using a large probabilistic network (NNN.5322)

Computer-based diagnostic decision support systems will play an increasingly important role in health care. They may improve the quality of the diagnostic process in accuracy and efficiency, while costs and burden of patients may be reduced. In addition, they can play an invaluable role in medical education. Potential users include general internists, super specialists, residents in internal medicine, and medical students.
The modern view is that decision support systems should be based on a probabilistic model, e.g. a belief network. This approach has the advantage that it can deal with uncertainty in a consistent and mathematically correct way. A drawback is that complex probabilistic models are intractable for exact computation. This has obstructed the development of a useful system for internal medicine, which aim to cover a broad medical domain at a detailed level. Several studies have shown that existing systems for internal medicine fail in part due to lack in the required level of detail at which the domain is modelled.
Advances in approximation techniques, in particular using variational methods, have opened new possibilities for computation in a large class of complex belief networks which are intractable for exact methods. These methods are developed by our project team as well as by other groups. The question is whether these methods are sufficiently powerful to make a useful decision support system for internal medicine feasible. This is to a large extend a user defined (medical) question, since

1. comparison with exact inference is not possible due to the complexity of the network and
2. errors in the approximation will be judged as acceptable not just on their numerical values but more importantly on their medical implications.

1. We aim to build a detailed probabilistic model for a prototype decision support system in haematology and endocrinology. The model will be based on domain knowledge and fine-tuned by retraining using historical as well as prospective patient data.
2. We aim to further improve the approximation schemes for probabilistic models to make the decision support system as accurate as possible in a feasible amount of computing time.
3. We will evaluate the decision support system in cooperation with users at the Utrecht University Medical Centre (Division of Internal Medicine and Dermatology (DIGD), departments of Internal Medicine and Endocrinology).
4. Software development for the support of the other 3 parts.

Er zijn nog geen resultaten bekend.

Six companies and three hospitals are involved in this project.