Phenotyping with Prior Knowledge using Patient Similarity
Proceedings of the 5th Machine Learning for Healthcare Conference, PMLR 126:331-351, 2020.
Prior medical knowledge, like the relationships between diseases or treatments and their corresponding risk factors are widely available in electronic health records (EHR), can be generated by domain experts, and extracted from knowledge graphs. Although informative for predictive modeling tasks, most of the patient-specific knowledge in EHR are not utilized because of practical constraints on data availability or cost of acquiring the data to make inferences. We present a method to learn from prior knowledge using a mixture of experts model where gating probabilities are tuned by an adjacency matrix created using side information available during training, like comorbidities, interventions, outcomes, vital signs and laboratory measurements. The adjacency matrix of a nearest neighbor graph is used to discover subgroups of intensive care unit (ICU) patients. Experts are shown to specialize based on how patients are grouped in the adjacency matrix on two real-world decision support tasks: predicting hemodynamic interventions and stratifying patients at risk for developing a sustained period of hypoxemia. The proposed prior knowledge-guided learning (PKL) model discovers clinically meaningful cohorts in patients with respiratory compromise that match well known sub-phenotypes described in the literature.