Deep Patient Representation of Clinical Notes via Multi-Task Learning for Mortality Prediction

Area of Science:

• * Artificial Intelligence in Medicine
• * Clinical Informatics
• * Machine Learning for Healthcare

Background:

• * Accurate patient mortality prediction is crucial for clinical decision-making and resource allocation.
• * Traditional single-task learning (STL) models may not fully leverage the rich information within electronic health records.
• * Developing generalizable patient representations is key to improving predictive model performance.

Purpose of the Study:

• * To propose and evaluate a deep learning-based multi-task learning (MTL) architecture for patient mortality prediction using clinical notes.
• * To demonstrate the benefits of MTL in improving performance on specific classification tasks compared to STL.
• * To explore the generalizability of MTL-derived patient representations for various clinical prediction tasks.

Main Methods:

• * Development of a multi-level Convolutional Neural Network (CNN) integrated with a multi-task learning loss component.
• * Training and evaluation of the MTL model across multiple related clinical prediction tasks, including in-hospital, 30-day, and 1-year mortality.
• * Comparison of MTL performance against single-task learning (STL) classifiers.

Main Results:

• * The proposed MTL architecture consistently outperformed STL classifiers across evaluated tasks.
• * Incorporating related mortality prediction tasks led to small but significant performance gains on individual tasks.
• * The MTL model generated high-quality patient representations beneficial for simpler downstream models.

Conclusions:

• * Multi-task learning (MTL) offers an efficient and generalizable approach for clinical prediction tasks using deep learning.
• * MTL effectively leverages information across related tasks to enhance patient mortality prediction accuracy.
• * The developed MTL framework provides robust patient representations applicable to diverse clinical outcomes.