TGRA-P: Task-driven model predicts 90-day mortality from ICU clinical notes on mechanical ventilation

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Natural Language Processing

Background:

• Limited ICU ventilators during COVID-19 highlight the need for efficient patient management.
• Existing clinical models struggle with data flexibility and normalization.
• Advancements in AI offer potential for analyzing unstructured clinical text.

Purpose of the Study:

• To develop an innovative, task-driven predictive model for early mortality risk assessment in ICU patients.
• To introduce the Task-driven Gated Recurrent Attention Pool (TGRA-P) model for mechanical ventilation patients.
• To assist clinicians in diagnosis and decision-making through accurate risk prediction.

Main Methods:

• Developed a Task-driven Gated Recurrent Attention Pool (TGRA-P) model.
• Proposed a Task-Specific Embedding Module for efficient, static embedding fine-tuning.
• Introduced Gated Recurrent Attention Unit (GRA) for enhanced text sequence dependency.
• Utilized Residual Max Pool (RMP) to incorporate all word-level features for prediction.

Main Results:

• TGRA-P achieved an AUROC of 0.8245±0.0096 and AUPRC of 0.7532±0.0115 for 90-day mortality prediction.
• The model demonstrated superior accuracy (0.7422±0.0028) and F1-score (0.6612±0.0059) compared to previous studies.
• Statistical validation confirmed the model's superiority over baseline models using Cohen's d effect sizes.

Conclusions:

• The TGRA-P model achieves state-of-the-art performance in predicting early mortality risk.
• The model offers a balance of performance and efficiency, reducing prediction costs.
• TGRA-P enhances clinical decision-making by integrating challenging textual patient data.