A machine learning pipeline trained on electronic health records from 163 patients at Hospital Universitario Parc Tauli in Sabadell, Spain, identified individuals with advanced chronic conditions (MACA) with a sensitivity of 0.91 and AUC-ROC of 0.90 using a bagging ensemble classifier. From 80 candidate variables spanning clinical, functional, and healthcare utilization domains, feature selection narrowed the model to ten key predictors — chief among them absolute dependency, advanced frailty, functional decline, and hospitalization frequency. Cross-validation yielded mean AUC values exceeding 0.95, suggesting internal stability.
This is a preprint posted on medRxiv and has not yet been peer-reviewed; the results should be interpreted cautiously until independent validation occurs. The most significant limitation is the small cohort: 80 MACA patients versus 83 controls is insufficient to establish generalizability across diverse health systems, ethnicities, or care settings. The model is also internally validated only — external validation against separate hospital populations is essential before clinical deployment. That said, the finding is contextually meaningful. Current MACA identification typically relies on retrospective clinical judgment, creating dangerous lag time before palliative or proactive care escalates. Frailty-predictive algorithms are a growing subspecialty in geriatric AI, and a high-sensitivity tool could reduce that lag meaningfully. The ten-variable economy of the model is also practically attractive for resource-limited settings. Overall, this is a promising but preliminary signal — incremental rather than paradigm-shifting — that warrants larger prospective trials.
