MATCHING HIDDEN NON-MARKOVIAN MODELS : DIAGNOSING ILLNESSES BASED ON RECORDED SYMPTOMS

Discrete stochastic models (DSM) can be used to accurately describe many natural and technical processes. The simulation algorithms usually require the system parts of interest to be completely observable in order to analyze the model. Hidden non-Markovian models (HnMM) have been applied successfully to the analysis of partially observable systems. They can determine the unobserved most likely system behavior that caused an observed output. The analysis can be done by the state space-based Proxel algorithm, which on-the-fly generates the reachable model state space at discrete points in time. In the current paper, we compute the unconditional probability of a given model having produced a given output. This can be used to find the most likely one of different possible system configurations to produce the given output. In our application we want to find the illness that most likely caused the recorded symptoms of a patient. Experiments are performed to determine the accuracy and limitations of the applicability of the approach. This paper increases the application area of HnMM analysis twofold. We can now perform model matching tasks for HnMM, and we have tested an application example from medical diagnosis.