DICE: Deep Significance Clustering for Outcome-Driven Stratification

We present deep significance clustering (DICE), a framework for jointly performing representation learning and clustering for "outcome-driven" stratification. Motivated by practical needs in medicine to risk-stratify patients into subgroups, DICE brings self-supervision to unsupervised tasks to generate cluster membership that may be used to categorize unseen patients by risk levels. DICE is driven by a combined objective function and constraint which require a statistically significant association between the outcome and cluster membership of learned representations. DICE also performs a neural architecture search to optimize cluster membership and hyper-parameters for model likelihood and classification accuracy. The performance of DICE was evaluated using two datasets with different outcome ratios extracted from real-world electronic health records of patients who were treated for coronavirus disease 2019 and heart failure. Outcomes are defined as in-hospital mortality (15.9%) and discharge home (36.8%), respectively. Results show that DICE has superior performance as measured by the difference in outcome distribution across clusters, Silhouette score, Calinski-Harabasz index, and Davies-Bouldin index for clustering, and Area under the ROC Curve for outcome classification compared to baseline approaches.