NeoTriFuse Predicts Neonatal Mortality with Reliability-Aware Multimodal Fusion

Predicting neonatal mortality risk from bedside monitoring data is a complex challenge in healthcare. This difficulty stems from several factors, including extreme class imbalance, the presence of diverse clinical risk factors, multi-scale temporal dynamics in the data, and crucially, significant amounts of missing data, which are common in real-world clinical settings.To address these issues, researchers propose NeoTriFuse, a novel reliability-aware multimodal fusion framework. Unlike conventional approaches that treat missingness merely as a preprocessing step, NeoTriFuse explicitly models missingness as a signal of data reliability. This signal then dynamically adjusts the contribution of different data modalities during the fusion process.The framework integrates static perinatal variables, local-global temporal encoders, and patient-level statistical summaries through these reliability-guided gating mechanisms. It jointly optimizes for mortality prediction and an auxiliary length-of-stay objective. NeoTriFuse demonstrated competitive performance, with strong F1 and AUROC scores. Ablation studies confirmed that the temporal architecture and patient-level summaries were key contributors, with reliability-aware gating further enhancing performance, especially for metrics sensitive to observation completeness.