NeoTriFuse: Reliability-Aware Multimodal Fusion under Missingness Heterogeneity for Neonatal Mortality Risk Prediction
NeoTriFuse introduces a reliability-aware multimodal fusion framework that treats missingness as an explicit reliability signal rather than a preprocessing problem, dynamically modulating modality contributions during fusion The architecture combines static perinatal variables, local-global temporal encoders, and patient-level statistical summaries through reliability-guided gating mechanisms Joint optimization of mortality prediction and an auxiliary length-of-stay objective improves overall pr
Analysis
TL;DR
- NeoTriFuse introduces a reliability-aware multimodal fusion framework that treats missingness as an explicit reliability signal rather than a preprocessing problem, dynamically modulating modality contributions during fusion
- The architecture combines static perinatal variables, local-global temporal encoders, and patient-level statistical summaries through reliability-guided gating mechanisms
- Joint optimization of mortality prediction and an auxiliary length-of-stay objective improves overall predictive performance
- NeoTriFuse achieves an F1 score of 0.6736 ± 0.0216 and an AUROC of 0.9454 ± 0.0056 on neonatal mortality risk prediction
- Ablation studies confirm the local-global temporal architecture and patient-level summary branch as the most impactful components, with reliability-aware gating providing additional gains under heterogeneous observation completeness
Why It Matters
This work addresses a critical gap in clinical AI: handling missing data not as a nuisance to impute away, but as an informative signal about data quality and observation reliability. For AI practitioners working in healthcare, the reliability-aware fusion paradigm offers a practical blueprint for building robust multimodal systems in real-world clinical settings where data completeness varies significantly across patients and time.
Technical Details
- Reliability-Aware Gating Mechanism: Missingness is modeled as an explicit reliability signal that dynamically modulates how much each modality contributes during fusion, rather than relying on traditional imputation or deletion strategies
- Three-Branch Architecture: Integrates (1) static perinatal variables, (2) local-global temporal encoders capturing multi-scale temporal dynamics in bedside monitoring data, and (3) patient-level statistical summaries
- Multi-Task Learning: Jointly optimizes the primary mortality prediction task with an auxiliary length-of-stay prediction objective to improve representation learning
- Performance Metrics: F1 score of 0.6736 ± 0.0216 and AUROC of 0.9454 ± 0.0056, with sensitivity analyses confirming stable performance across nearby hyperparameter settings
- Ablation Findings: The local-global temporal architecture and patient-level summary branch contribute most substantially to predictive performance, while reliability-aware gating provides additional improvements specifically on threshold-dependent metrics under heterogeneous observation completeness
Industry Insight
- The reliability-aware fusion paradigm could generalize beyond neonatal care to other clinical domains with heterogeneous missingness patterns, such as ICU monitoring, chronic disease management, and longitudinal electronic health record analysis
- Multi-task learning with auxiliary clinical objectives (e.g., length-of-stay prediction) offers a practical regularization strategy that may improve generalization in data-scarce or imbalanced medical prediction tasks
- The approach challenges the dominant imputation-first mindset in clinical AI, suggesting that missingness patterns themselves carry predictive value and should be preserved and modeled explicitly rather than erased
Disclaimer: The above content is generated by AI and is for reference only.