Generalizing intensive care AI across time scales in resource-limited settings

Generalization remains a major barrier to the clinical deployment of artificial intelligence in critical care. Models developed using high-frequency physiological monitoring often experience performance degradation when applied across hospitals, patient populations, and monitoring environments that differ from those used during training. Although variability in temporal resolution is widespread in critical care, its impact on representation learning and model transferability remains poorly understood. Here, we introduce the resolution-transfer task for physiological time series and present PhysioStack, a self-supervised physiological representation learning framework designed to evaluate whether representations learned from high-frequency monitoring data remain robust across clinically relevant monitoring frequencies without retraining. We introduce SafeICU, a longitudinal pediatric intensive care dataset spanning ten years of routinely collected clinical and physiological data from a tertiary-care hospital in India, including high-resolution vital signs recorded at 15-second intervals. Transformer-based masked language models were trained on 144,271 patient-hours of high-resolution physiological signals from 984 pediatric ICU stays to learn representations of heart rate, respiratory rate, oxygen saturation, and arterial blood pressure. Representations learned at high temporal resolution demonstrated strong transferability across lower monitoring frequencies, consistently outperforming models trained directly on temporally aggregated data. Importantly, these representations generalized across patient populations, maintaining performance when evaluated on independent adult intensive care cohorts derived from the MIMIC-III and eICU databases without retraining. In a downstream early shock prediction task, PhysioStack achieved AUROC values of 0.82-0.86 and AUPRC values of 0.91-0.94 across 5-60-minute monitoring frequencies without retraining. Representations trained on pediatric data generalized to independent adult ICU cohorts, achieving AUROC 0.91-0.95 and AUPRC 0.94-0.97 without cohort-specific fine-tuning. These findings demonstrate that PhysioStack learns transferable physiological representations that generalize across temporal resolutions and patient cohorts, supporting robust AI for critical care. To support further research, we publicly release the SafeICU database, comprising longitudinal vital signs, laboratory measurements, treatment records, microbiology, and admission and discharge, together with pretrained physiological representation models.