A reinforcement learning-guided interpretable method for postoperative sepsis prediction with Hilbert-Schmidt...

Area of Science:

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

Background:

• Sepsis poses a significant threat to patient outcomes post-surgery, necessitating accurate and efficient risk stratification.
• Current sepsis prediction models often lack transparency and rely on computationally intensive data streams.
• Electronic Health Records (EHR) offer rich data but present challenges in processing high-dimensional, heterogeneous information.

Purpose of the Study:

• To develop a reinforcement learning-guided framework for interpretable feature engineering in postoperative sepsis prediction.
• To enhance the accuracy and computational efficiency of sepsis risk stratification using perioperative data.
• To provide clinically interpretable insights into sepsis prediction models.

Main Methods:

• An Actor-Critic reinforcement learning approach was employed, treating feature selection as an action.
• A self-attention-based classifier was utilized for downstream prediction.
• An auxiliary baseline model incorporating temporal convolutional networks (TCN) and Hilbert-Schmidt Independence Criterion (HSIC) regularization was introduced for benchmarking.

Main Results:

• The proposed framework achieved performance comparable to or exceeding existing machine learning baselines on a real-world surgical cohort.
• The model demonstrated superior efficiency by selecting a reduced set of input features.
• Perfect scores (1.00) for F1-score, Sensitivity, and Specificity were attained on the experimental dataset.

Conclusions:

• The developed method accurately predicts postoperative sepsis with enhanced interpretability.
• The framework offers a novel and efficient approach to postoperative sepsis prediction using perioperative EHR data.
• Instance-level explanations enhance the clinical utility and trustworthiness of the sepsis prediction model.