GPT-4o and the quest for machine learning interpretability in ICU risk of death prediction.

Abstract

Background: Clinical utilization of machine learning is hampered by the lack of interpretability inherent in most non-linear black box modeling approaches, reducing trust among clinicians and regulators. Advanced large language models offer a potential framework for integrating medical knowledge into these models, potentially enhancing their interpretability.

Methods: A hybrid mechanistic/data-driven modeling framework is presented for developing an ICU risk of death prediction model for mechanically ventilated patients. In the mechanistic modeling part, GPT-4o is used to generate detailed medical feature descriptions, which are then aggregated into a comprehensive corpus and processed with TF-I DF vectorization. Fuzzy C-means clustering is subsequently applied to these vectorized features to identify significant mortality cause-specific feature clusters, and a physician reviewed the resulting clusters to validate their relevance to actionable insights for clinical decision support. In the data-driven part, the identified clusters inform the creation of XGBoost-based weak classifiers, whose outcomes are combined into a single XGBoost-based strong classifier through a hierarchically structured feed-forward network. This process results in a novel GPT hybrid model for ICU risk of death prediction.

Results: This study enrolled 16,018 mechanically ventilated ICU patients, divided into derivation (12,758) and validation (3,260) cohorts, to develop and evaluate a GPT hybrid model for predicting in-ICU death. Leveraging GPT-4o, we implemented an automated process for clustering mortality cause-specific features, resulting in six feature clusters: Liver Failure, Infection, Renal Failure, Hypoxia, Cardiac Failure, and Mechanical Ventilation. This approach significantly improved upon previous manual methods, automating the reconstruction of structured hybrid models. While the GPT hybrid model showed similar predictive accuracy to a Global XGBoost model, it demonstrated superior interpretability and clinical relevance by incorporating a wider array of features and providing a hierarchical structure of feature importance aligned with medical knowledge.

Conclusion: We introduce a novel approach to predicting in-ICU risk of death for mechanically ventilated patients using a GPT hybrid model. Our methodology demonstrates the potential of integrating large language models with traditional machine learning techniques to create interpretable and clinically relevant predictive models.