Applying domain knowledge to clinical predictive models

Abstract

Clinical predictive models are useful in predicting a patient's risk of developing adverse outcomes and in guiding patient therapy. In this thesis, we explored two different ways to apply domain knowledge to improve clinical predictive models. We first applied knowledge about the heart to engineer better frequency-domain features from electrocardiograms (ECG). The standard frequency domain (in Hz) quantifies events that repeat with respect to time. However, this may be misleading because patients have different heart rates. We hypothesized that quantifying frequency with respective to heartbeats may adjust for these heart rate differences. We applied this beat-frequency to improve two existing ECG predictive models, one based on ECG morphology, and the other based on instantaneous heart rate. We then used machine learning to find predictive frequency bands. When evaluated on thousands of patients after an acute coronary syndrome, our method significantly improved prediction performance (e.g., area under curve, AUC, from 0.70 to 0.75). In addition, the same bands were found to be predictive in different patients for beat-frequency, but not for the standard frequency domain. Next, we developed a method to transfer knowledge from published biomedical articles to improve predictive models when training data are scarce. We used this knowledge to estimate the relevance of features to a given outcome, and used these estimates to improve feature selection. We applied our method to predict the onset of several cardiovascular diseases, using training data that contained only 50 adverse outcomes. Relative to a standard approach (which does not transfer knowledge), our method significantly improved the AUC from 0.66 to 0.70. In addition, our method selected 60% fewer features, improving interpretability of the model by experts, which is a key requirement for models to see real-world use.