Using a Multi-Task Recurrent Neural Network With Attention Mechanisms to Predict Hospital Mortality of Patients.

Area of Science:

• * Computational biology and health informatics.
• * Application of artificial intelligence in clinical prediction.

Background:

• * Accurate estimation of hospital mortality is crucial for clinical decision-making and resource management.
• * Existing methods like Simplified Acute Physiology Score (SAPS-II) have limitations in prediction accuracy.

Purpose of the Study:

• * To develop and evaluate a multi-task recurrent neural network with attention mechanisms for predicting patient hospital mortality.
• * To utilize physiological time series reconstruction as an auxiliary task to enhance mortality prediction.

Main Methods:

• * Implementation of a multi-task learning framework incorporating recurrent neural networks and attention mechanisms.
• * Training and validation on the MIMIC-III electronic health record database.
• * Utilizing 15 physiological measurements from the initial 24 hours of critical care.

Main Results:

• * The proposed multi-task learning model demonstrated superior sensitivity (0.503 ± 0.020) compared to SAPS-II (0.365 ± 0.021) for predicting hospital mortality within 24 hours.
• * The model's performance is comparable or superior to SAPS-II with daily updates and a minimum 6-hour observation period.

Conclusions:

• * The developed multi-task learning model offers a promising advancement in predicting hospital mortality.
• * The model's enhanced accuracy can aid clinicians in making informed decisions and optimizing resource allocation.
• * Future work could incorporate 'need for intervention' as an additional predictive task.