Abstract / Summary
Objective: Sepsis and heart failure produce overlapping early physiological signs in intensive care but call for opposite treatments, yet existing models predict one condition at a time. We built one model for both. Methods: DualStream-MTCA combines two bidirectional long short-term memory encoders, for 48 hours of hourly vital signs and 7 days of daily laboratory results, with multi-head cross-attention, gradient-boosted-tree leaf embeddings and two task heads trained under uncertainty weighting. A refined configuration adds per-window aggregates, static patient features, self-supervised encoder pretraining and a five-seed ensemble with Monte Carlo dropout. Development used 53,229 adult stays from MIMIC-IV version 3.1; the frozen models were applied to 102,695 stays from the eICU Collaborative Research Database. Results: At a prediction time 48 hours after admission, the refined ensemble reached an area under the receiver operating characteristic curve of 0.867 for sepsis and 0.899 for heart failure internally, against 0.832 and 0.857 for the baseline model. Isotonic regression brought expected calibration error below 0.015. For positives whose onset followed the prediction time by 4 to 12 hours, the area dropped to 0.66 to 0.70. Externally, the refined ensemble scored 0.750 for sepsis and 0.786 for heart failure, against 0.711 and 0.821 for the baseline model. Conclusion: A single multi-task model matched its single-task variants on both conditions, but discrimination fell under strictly anticipatory prediction and across sites, where static features carried little usable signal. Significance: To our knowledge, this is the first model built specifically to predict both conditions jointly, with calibration, decision-curve and external evaluation.