Using machine learning techniques to develop forecasting algorithms for postoperative complications: protocol for a...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Outcomes Research

Background:

• Postoperative mortality and morbidity are significant public health issues in the USA.
• Current prediction models for adverse surgical outcomes lack real-time individual patient forecasting capabilities.
• There is a need for advanced methods to predict patient trajectories during the perioperative period.

Purpose of the Study:

• To develop and validate machine learning algorithms for predicting adverse perioperative outcomes.
• To leverage perioperative time-series data for real-time risk assessment.
• To enhance the accuracy of individual patient outcome prediction following surgery.

Main Methods:

• Utilized a 4-year dataset of adult surgical patients from a tertiary care hospital.
• Extracted patient history, lab values, intraoperative vital signs, and medications from electronic medical records.
• Employed density-based logistic regression with Nadaraya-Watson kernel density estimation, feature extraction, shapelet methods, and convolutional neural networks for algorithm construction.

Main Results:

• Algorithms were constructed to forecast in-hospital mortality, postoperative acute kidney injury, and postoperative respiratory failure.
• Time-series variables were analyzed using advanced feature extraction and machine learning techniques.
• Algorithm performance was validated using precision and recall metrics.

Conclusions:

• Developed machine learning algorithms show promise for real-time prediction of adverse perioperative outcomes.
• Accurate dynamic risk profiling can potentially enable clinicians to tailor care plans and improve patient trajectories.
• Further validation through a randomized controlled trial is planned to test the clinical utility of these forecasting algorithms.