Session MC.1

Finding Cardiac Disease Correlations through Time-Varying Disease Modeling for Decision Support

T Syeda-Mahmood*, S Adelman, A Hobbs, J Terdiman, M London,
A Amir, H Greenspan, D Gruhl

IBM Almaden Research Center
San Jose, CA , USA

An important aspect of diagnostic decision support is finding meaningful associations between diseases, also called co-morbidities. Research studies on co-morbidities have been primarily focused on generic disease pairs, such as heart disease with diabetes within population of patients diagnosed with those diseases. With electronic medical records now housing patient information in large hospital networks, it is possible to do large scale disease-association studies through automated data mining methods such as association mining. Such mining methods, however, count primarily the frequency of co-occurrences of diseases which can sometimes lead to incorrect causative inferences. A more refined model would involves looking at the evolution of the diseases over time per patient to infer functional and causal dependencies.
In this paper, we present a method to discover meaningful co-morbidity associations between cardiac diseases using time-varying data modeling. Specifically, we represent cardiac diseases as a two-dimensional function of disease per patient over time. A similarity metric is defined per pair of diseases as a functional correlation of the corresponding time series normalized over the patient population. The correlation takes into account both the presence and lack of overlap between the disease time series as well as the time shift needed for causative dependencies. The similarity metric is used to produce a ranked list of correlations per disease.
The disease correlation algorithm has been applied to a collection of over 5000 patients whose longitudinal records were available for a period of 15 years. A set of 123 ICD9 codes relating to cardiovascular diseases were used as queries to a search engine to produce a ranked list of correlation for each disease. The predictions of correlations made by the algorithm matched those known in literature 85% of the time (e.g. mitral stenosis correlated with hypertension and atrial fibrillation).In addition, the algorithm produced plausible correlations, such as bundle branch blocks with ventricular septal defect, based on causal dependencies detected. Results of comparison to straightforward database association algorithms also revealed a 30% improvement in performance of the time-correlation-based algorithms.

(Abstract Control Number: 239)