Health State Modeling and Prediction based on Hidden Markov Models

The collective health clinic data of people in a society is surmised to have a variety of characteristic health states and certain dynamics governing health state changes over time. Given such a collection of samples, we propose a way of estimating a set of health states in dynamic context using the tool of hidden Markov model (HMM). We also present a method of predicting the future health states based on the Markov dynamics or a set of state duration statistics derived from the model parameters and data clusters. In the proposed method we design a number of HMMs, each for a set of sequences with a particular disease history of interest. They are used to predict the future health states on the basis of corresponding diseases or health problems. Health state prediction as a service can be presented with a set of potential future states and over a number of years into the future. Experimental results have shown a baseline performance of 48\% for single year prediction with a model of sixteen states with single hypothesis for the apoplexy history data set. The performance quickly increases to 70\% with two hypotheses. When applied to multiple year prediction, the accuracy decreases by about 12 percentage points or less over the course of five years.