Transparent Early ICU Mortality Prediction with Clinical Transformer and Per-Case Modality Attribution

TL;DR

This work introduces a transparent, lightweight multimodal ensemble for early ICU mortality prediction using the first 48 hours of physiological time-series and unstructured clinical notes. A Bidirectional LSTM handles vitals while a finetuned ClinicalModernBERT processes notes; their probability outputs are combined by a linear logistic regression meta-learner on standardized logits, yielding a calibrated risk score. The architecture provides multilevel interpretability, including per-case modality attribution and within-branch Integrated Gradients explanations, and demonstrates robustness to missing modalities via calibrated fallbacks. On MIMIC-III, the ensemble achieves AUPRC 0.565 and AUROC 0.891, outperforming single modalities, with well-calibrated predictions and actionable explanations suitable for clinical governance and deployment considerations.

Abstract

Early identification of intensive care patients at risk of in-hospital mortality enables timely intervention and efficient resource allocation. Despite high predictive performance, existing machine learning approaches lack transparency and robustness, limiting clinical adoption. We present a lightweight, transparent multimodal ensemble that fuses physiological time-series measurements with unstructured clinical notes from the first 48 hours of an ICU stay. A logistic regression model combines predictions from two modality-specific models: a bidirectional LSTM for vitals and a finetuned ClinicalModernBERT transformer for notes. This traceable architecture allows for multilevel interpretability: feature attributions within each modality and direct per-case modality attributions quantifying how vitals and notes influence each decision. On the MIMIC-III benchmark, our late-fusion ensemble improves discrimination over the best single model (AUPRC 0.565 vs. 0.526; AUROC 0.891 vs. 0.876) while maintaining well-calibrated predictions. The system remains robust through a calibrated fallback when a modality is missing. These results demonstrate competitive performance with reliable, auditable risk estimates and transparent, predictable operation, which together are crucial for clinical use.

Figures (9)

• Figure 1: Transparent multimodal ensemble for early ICU mortality prediction. Two specialist models, a bidirectional LSTM for time-series vitals (TS), and a finetuned ClinicalModernBERT transformer for clinical notes (CN), produce probability outputs that are converted to standardized logits. A logistic regression model (meta-learner) combines these logits to generate a calibrated in-hospital mortality (IHM) risk score. Orange blocks indicate modular, independently trainable components. The system provides multilevel explainable AI (XAI): Integrated Gradients attributions identify influential vitals variables/time-steps and note tokens within each specialist, while per-case modality shares quantify how vitals and notes influence a decision. When a modality is unavailable, the system falls back to the calibrated probability, ensuring graceful degradation.
• Figure 2: Precision–Recall Curve for all specialist classifiers.
• Figure 3: AUPRC discrimination (mean [95% CI]) comparison of six meta-learning algorithms for the LSTM + ft_ClinicalModernBERT pairing.
• Figure 4: AUPRC discrimination (mean [95% CI]) for all specialist models and the selected ensemble on the test set. Models emb_ModernBERT, emb_ClinicalModernBERT, and ft_ClinicalModernBERT denoted as emb_MBERT, emb_CMBERT, and ft_CMBERT respectively.
• Figure 5: Reliability diagrams before and after isotonic regression for (a) the LSTM vitals specialist, (b) the ft_ClinicalModernBERT notes specialist, and (c) the logistic stacker ensemble. Each panel shows pre-calibration (raw probabilities) and post-calibration (isotonic regression) curves; points closer to the diagonal indicate better calibration.