Diagnosis-based mortality prediction for intensive care unit patients via transfer learning

2025-12-06

AI Summary

To address poor generalizability of mortality prediction models in ICU settings caused by diagnostic heterogeneity, this paper proposes a diagnosis-aware transfer learning framework. First, patients are grouped by primary diagnosis; then, a shared-specific feature modeling mechanism jointly learns both population-level patterns and diagnosis-specific risk signals. The framework is implemented using generalized linear models (GLMs) and XGBoost, with decision thresholds optimized via the Youden index. Empirical evaluation on the eICU database demonstrates statistically significant improvements over baseline models—including diagnosis-specific models trained only on subset data and the APACHE IVa score—across AUC, Brier score, and calibration metrics, particularly enhancing reliability for rare diagnoses. The core contribution lies in embedding clinical diagnostic priors into the transfer learning architecture, thereby synergistically improving both diagnosis-specific interpretability and cross-diagnostic generalizability.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

In the intensive care unit, the underlying causes of critical illness vary substantially across diagnoses, yet prediction models accounting for diagnostic heterogeneity have not been systematically studied. To address the gap, we evaluate transfer learning approaches for diagnosis-specific mortality prediction and apply both GLM- and XGBoost-based models to the eICU Collaborative Research Database. Our results demonstrate that transfer learning consistently outperforms models trained only on diagnosis-specific data and those using a well-known ICU severity-of-illness score, i.e., APACHE IVa, alone, while also achieving better calibration than models trained on the pooled data. Our findings also suggest that the Youden cutoff is a more appropriate decision threshold than the conventional 0.5 for binary outcomes, and that transfer learning maintains consistently high predictive performance across various cutoff criteria.

Problem

• Develops diagnosis-specific mortality prediction models for ICU patients
• Evaluates transfer learning to address diagnostic heterogeneity in critical illness
• Compares model performance against traditional severity scores and pooled data

Innovation

• Transfer learning for diagnosis-specific mortality prediction
• GLM and XGBoost models applied to eICU database
• Youden cutoff improves decision threshold over conventional 0.5