naviDCN: Navigator-Guided Multi-Modal Deep Clustering for Sepsis Phenotyping in Early ICU Admission.

Sepsis is a life-threatening and heterogeneous disease characterized by dysregulated host responses to infection. Although recent studies applied unsupervised algorithms to uncover sepsis phenotypes, the clustering process lacks the ability to incorporate clinical knowledge, potentially resulting in phenotypes with limited interpretability. We propose a novel clustering framework, naviDCN, which integrates a navigator component to align clusters with clinical significance. The naviDCN architecture comprises multi-modal encoders, a deep clustering network (DCN) with reconstruction tasks, and a navigator module. We first encode electronic health records into representative embeddings by introducing an attention mechanism on multi-modal information. The framework then iteratively optimizes network weights of reconstruction and navigator modules, and updates cluster centroids of the clustering module. The navigator incorporates clinical knowledge into the embedding through backpropagation, thereby guiding clustering toward clinically meaningful outcomes. We discover four sepsis phenotypes with unique clinical characteristics, SOFA trajectories, and mortality patterns. Notably, while both α and δ phenotypes show severe conditions in the early stage, naviDCN effectively differentiates between patients likely to show clinical improvement (α) and those at risk of deterioration (δ).Furthermore, the navigator effectively enhances phenotype interpretability without compromising objective clustering performance. This study offers insights into understanding the heterogeneity of sepsis phenotypes.Clinical Relevance-This study integrates a navigator module into the clustering framework to identify phenotypes with distinct short-term organ dysfunction trajectories and long-term survival status, thus improving the interpretability of sepsis phenotypes. Unraveling the latent information in demographics, laboratory test results, and vital signs with deep learning, our framework identifies characteristic phenotypes and opens new avenues for exploring targeted treatment for sepsis patients across phenotypes.