Development and multicenter validation of machine learning models for predicting postoperative pulmonary...

Area of Science:

• Neurosurgery
• Pulmonary Medicine
• Medical Informatics

Background:

• Postoperative pulmonary complications (PPCs) are significant adverse events following neurosurgery.
• Predicting and preventing PPCs is crucial for improving patient outcomes.
• Current prediction methods may lack accuracy and generalizability.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting PPCs in neurosurgical patients.
• To identify key risk factors associated with PPCs.
• To compare the performance of ML models against existing clinical scores.

Main Methods:

• Retrospective analysis of anesthesia records from multiple hospitals (2018-2023).
• Development of ML models using six algorithms (e.g., DNN, LR) with 35 or 11 selected features.
• Validation using temporal and external datasets, comparing against ARISCAT and LAS VEGAS scores.

Main Results:

• PPCs occurred in approximately 9-10% of patients across datasets.
• Developed ML models demonstrated good discrimination (AUC ~0.84), with DNN and LR performing best.
• An 11-feature logistic regression nomogram outperformed ARISCAT and LAS VEGAS scores (AUC 0.824 vs. 0.672/0.663).

Conclusions:

• Machine learning models, particularly DNN and the LR-based nomogram, show strong predictive performance for PPCs in neurosurgery.
• These validated models can serve as valuable clinical decision support tools.
• Identification of independent risk factors aids in targeted prevention strategies and improved surgical outcomes.