We introduce the TRACER model, a Transformer equipped with Contrastive Event Representation, designed to forecast timelines leading to rare hospitalization events from low‑resolution, irregularly sampled telemonitoring data. Each biomarker receives a time‑aware embedding, contrastive pre‑training enhances anomaly detection via representation learning, and independent binary classifiers perform event detection.
The study leveraged biomarker sequences recorded remotely from 276 heart‑failure patients. Sequences were segmented into overlapping windows according to temporal rules and labeled based on whether a heart‑failure‑related hospitalization occurred at the window’s trailing edge. On a highly imbalanced real‑world dataset, TRACER correctly predicted 66.7% of the hospitalization timelines with an overestimation rate of 7.9%. Reformulating TRACER’s training as an event‑detection task, rather than direct forecasting, markedly improved performance and made better use of the scarce hospitalization events.
Compared with baseline models, TRACER achieved superior detection of worsening signs, indicating promise for generating alerts and guiding counteractive treatment in real‑world telemonitoring scenarios. Review