Self-Supervised Early Warning of Critical Cardiorespiratory Events in Neonatal Vital-Sign Telemetry: A Pilot Feasibility Study

Continuous bedside monitoring generates dense vital-sign telemetry for hospitalised neonates, yet most of this signal is consumed by threshold alarms that fire only once an event is already underway. We investigate whether short-horizon prediction of critical cardiorespiratory events is feasible from routinely collected pulse-oximetry and temperature telemetry, using self-supervised labels derived from the signal itself rather than scarce clinical outcome labels. Using continuous recordings from three neonates (838,613 raw samples; 1,113 cumulative monitoring hours), we frame the task as predicting whether a desaturation event (SpO₂ < 85%) will occur within a 60-second horizon, given a 30-second window of preceding telemetry. We evaluate a gradient-boosting baseline on engineered window features and a gated recurrent unit (GRU) on raw sequences, under strict leave-one-patient-out validation. The GRU achieves an area under the receiver operating characteristic curve (AUROC) of 0.70–0.96 across held-out patients, outperforming the gradient-boosting baseline (0.69–0.91). A leakage-safe lead-time analysis shows that flagged desaturations are warned a median of 80 seconds in advance, while also exposing a gap between ranking performance and event-level sensitivity that ROC alone hides. We position this as a proof-of-concept: with three patients the results establish feasibility and surface methodological constraints, but not population-level performance. Code and a reproducible training notebook are released.