Interpretable laboratory-trajectory phenotypes for risk stratification of in-hospital critical events in cancer inpatients

Hospitalized cancer patients accumulate dozens of routine laboratory results within the first week of admission, but these are typically used only to flag individual abnormal values, and interpretability is a prerequisite for the bedside adoption of any clinical decision-support tool. Existing laboratory-based scores use single timepoints and lack event-specificity, whereas machine-learning models are difficult to interpret at the bedside. An interpretable, event-specific approach to risk stratification from routine laboratory data is therefore needed. We retrospectively assembled 3,083 cancer inpatients (October 2021–July 2025) across seven cancer types. Twenty-six laboratory indicators within 7 days of admission were summarized into five trajectory features each. Phenotypes were derived using principal component analysis and K-means clustering in the training cohort and evaluated in random hold-out, temporal, and MIMIC-IV cohorts. Candidate K = 2–6 solutions were assessed using internal validity and bootstrap stability criteria. Performance was benchmarked against five clinical scores, four outcome-specific three-variable expert-informed logistic models, and a 130-feature CatBoost model. Associations were estimated using Firth logistic regression, and reporting followed TRIPOD+AI. Four reproducible phenotypes were identified. C2 and C4 were associated with obstructive jaundice and GI bleeding, respectively, across random and temporal cohort definitions. Validation AUCs for Phenotype, Expert Logistic, and CatBoost were 0.929, 0.970, and 0.936 for obstructive jaundice and 0.694, 0.766, and 0.812 for GI bleeding. For GI bleeding, CatBoost had higher AUC than Phenotype (q = 0.002) and Expert Logistic (q = 0.007). Four outcome-agnostic laboratory-trajectory phenotypes provided transparent risk stratification, with the strongest evidence for obstructive jaundice and GI bleeding. Predictive performance varied by outcome. Prospective multicenter evaluation is required before clinical implementation. Not applicable.