Early Prediction of Sepsis Onset Using Neural Architecture Search Based on Genetic Algorithms.

Area of Science:

• * Computational biology and medical informatics.
• * Development of advanced machine learning models for healthcare.

Background:

• * Sepsis presents a critical threat to patient survival, necessitating effective early detection strategies.
• * Current prediction methods often lack the required speed and accuracy for timely intervention.

Purpose of the Study:

• * To develop and evaluate a computationally efficient neural architecture search (NAS) model for predicting sepsis onset.
• * To assess the model's performance across various prediction timeframes using real-world clinical data.

Main Methods:

• * Implementation of a novel NAS model integrated with a genetic algorithm (GA) for optimized neural network architecture discovery.
• * Weight sharing across internal network connections and genotypes to reduce computational search cost.
• * Validation using the Medical Information Mart for Intensive Care III (MIMIC-III) time-series dataset.

Main Results:

• * Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.94 for 3-hour sepsis onset prediction.
• * Demonstrated superior performance compared to SOFA, qSOFA, and SAPS II scoring systems (0.31-0.26 higher AUROC).
• * Outperformed a standard long short-term memory (LSTM) model by 12% in AUROC and showed robustness to input data variations.

Conclusions:

• * The proposed NAS-GA model offers a highly effective and computationally efficient approach for early sepsis prediction.
• * This method significantly enhances predictive accuracy over conventional clinical scoring systems and baseline deep learning models.
• * The model's robustness and performance make it a promising tool for improving sepsis patient outcomes.