An interpretable deep learning framework for predictive modeling of postoperative infections in ICU patients

Area of Science:

• Medical Informatics
• Computational Biology
• Clinical Medicine

Background:

• Postoperative infections are a significant concern for intensive care unit (ICU) patients undergoing surgery.
• Predicting infection risk and identifying contributing factors is vital for patient management but complicated by multifactorial influences.
• Existing machine learning models often lack interpretability, limiting clinical application.

Purpose of the Study:

• To develop an interpretable deep neural network (DNN) model for predicting postoperative infection risk in ICU patients.
• To identify key risk factors contributing to postoperative infections using a permutation feature importance test (PermFIT).
• To enhance predictive accuracy by utilizing only the most significant identified features.

Main Methods:

• Utilized electronic health records (EHRs) from the Medical Information Mart for Intensive Care (MIMIC-III) database.
• Developed an interpretable DNN model integrated with PermFIT for rigorous feature importance evaluation.
• Applied statistical inference to assess the impact of each feature on infection risk.

Main Results:

• The DNN-PermFIT framework successfully identified significant factors associated with postoperative infections.
• The model achieved highly accurate predictions for postoperative infection risk.
• Using only identified important features further improved the DNN's predictive performance.

Conclusions:

• The proposed interpretable DNN framework offers clinical utility for managing postoperative care in ICU surgical patients.
• Accurate risk prediction and identification of key factors can lead to improved patient health outcomes.
• This approach addresses the interpretability challenge in machine learning for clinical risk prediction.