Identification of predictive subphenotypes for clinical outcomes using real world data and machine learning.

Predicting treatment response is an important problem in real-world applications, where the heterogeneity of the treatment response remains a significant challenge in practice. Unsupervised machine learning methods have been proposed to address this challenge by clustering patients with similar electronic health record (EHR) data. However, they cannot guarantee coherent outcomes within the groups. Here, we propose Graph-Encoded Mixture Survival (GEMS) as a general machine learning framework to identify distinct predictive subphenotypes that guarantee coherent survival and baseline characteristics within each subphenotype. We apply our method to a real-world dataset of advanced non-small cell lung cancer (aNSCLC) patients receiving first-line immune checkpoint inhibitor (ICI) therapy to predict overall survival (OS). Our method outperforms baseline methods for predicting OS and identifies three reproducible subphenotypes associated with distinct baseline clinical characteristics and OS. Our results demonstrate that our method can provide insights in the heterogeneity of treatment response and potentially influence treatment selection.