Incorporating Patient Similarity and Clinical Temporality in Disease Prognostic Modeling

Health recommender systems (HRSs) enhance prognostication by leveraging clinical information. Existing HRSs often fail to capture the intrinsic correlations between patient phenotypes with similar clinical profiles, necessitating approaches that incorporate patient similarity into prognostic modeling. This work explores such correlations within three classes of biomedical information: diagnosis, procedure, and medication, and the clinical temporality during patient visits to improve diagnostic accuracy. Our approaches include both static and dynamic scenarios. In the static scenario, we propose SIM-PR, which integrates patient similarity and PageRank centrality on a personalized patient graph, and can operate with or without sequential hospital visit information. In the dynamic setting, we develop temporal prediction models based on a multilayer perceptron and a long short-term memory network to learn evolving diagnostic patterns from longitudinal visit histories. Standard supervised machine learning, including logistic regression, random forest, and support vector machines, is employed as comparative baselines. Experiments on the MIMIC-III database demonstrate that the proposed static and temporal models effectively reduce false positives and negatives and achieve superior predictive accuracy compared to existing HRS approaches.