Using hidden Markov models to characterize disease trajectories

Carsten Peterson, Mattias Ohlsson

Research output: Chapter in Book/Report/Conference proceedingBook chapterResearchpeer-review


A novel approach is developed for predicting body trajectories for cancer progression, where conditional probabilities of clinical data are modeled using Hidden Markov Model techniques. Basically, each potential body site is encoded by an N-letter code, and a disease trajectory is described in terms of a string of letters. Patient data base records are then represented by such strings with different lengths, start points and end points. The approach is explored using pathology data for non-Hodgkin lymphoma augmented with an artificial data base generated according to observed distributions in the clinical data. For the Hidden Markov Models a Bayesian approach is taken using the Hybrid Monte Carlo method, producing an ensemble of models rather than a single one. Using a test set consisting of both real and random trajectories, we estimate the performance of our Hidden Markov Model models and also extract most probable profiles. Given the limited data set size the results are very encouraging.
Original languageEnglish
Title of host publicationProceedings of the Neural Networks and Expert Systems in Medicine and Healthcare Conference, 324-326 (2001), eds. G.M. Papadourakis
EditorsG. M. Papadourakis
Number of pages3
Publication statusPublished - 2001

Subject classification (UKÄ)

  • Probability Theory and Statistics
  • Other Medical Engineering


Dive into the research topics of 'Using hidden Markov models to characterize disease trajectories'. Together they form a unique fingerprint.

Cite this