Deep Learning Identifies Novel Subphenotypes of Multiple Organ Dysfunction Syndrome in Critical Care

Background : Multiple organ dysfunction syndrome (MODS) is associated with high mortality in critically ill patients, but traditional severity scores do not capture dynamic patterns of organ failure over time. Objectives : To identify novel MODS subphenotypes using unsupervised deep learning on high-resolution time-series data. Methods : Retrospective observational cohort study using the Medical Infor- mation Mart for Intensive Care (MIMIC)-IV version 3.1 database. We included 30,950 adult ICU stays with SOFA score ≥2 involving at least two organ systems and ICU length of stay > 48 hours. Hourly data from seven organ systems were extracted for the first 72 hours after ICU admission. A long short-term memory variational autoencoder was trained to generate latent representations, followed by UMAP visualization and K-means clustering. Results : Unsupervised clustering of VAE-derived latent representations iden- tified three distinct subphenotypes comprising 9,355 (30.2%), 9,590 (31.0%), and 12,005 (38.8%) ICU stays. Hospital mortality rates were 22.7%, 21.4%, and 11.2%, respectively (log-rank p < 0.001). In multivariable Cox regression adjusted for age, sex, baseline SOFA score, Charlson Comorbidity Index, APS III score, and surgical admission status, the two higher-mortality subphenotypes remained independently associated with increased risk compared with the ref- erence subphenotype (adjusted hazard ratios 1.60 [95% CI 1.49–1.71] and 1.38 [95% CI 1.29–1.48], respectively; both p < 0.001). The model showed good discrimination (C-index 0.725). Conclusions : Application of a variational autoencoder to high-resolution ICU time-series data identifies three clinically meaningful MODS subphenotypes with significantly different mortality risk, providing a foundation for precision critical care interventions.