Development and prospective validation of postoperative pain prediction from preoperative EHR data using...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Pain Management

Background:

• Accurate prediction of postoperative pain is crucial for effective perioperative pain management.
• Current prediction methods are often subjective, labor-intensive, or lack sufficient accuracy.
• A need exists for automated, reliable tools to identify patients at high risk for acute postoperative pain.

Purpose of the Study:

• To develop and validate a machine learning model (POPS) for predicting postoperative pain.
• To utilize routinely collected electronic health record (EHR) data for automated pain prediction.
• To assess the performance of POPS against clinician predictions in a prospective cohort.

Main Methods:

• Developed a machine learning method (POPS) using a multicenter dataset of 234,274 adult non-cardiac surgical patients.
• Trained POPS on pre-operative EHR data to predict maximum pain scores for the day of surgery and four subsequent days.
• Validated POPS prospectively, comparing its predictions to actual patient pain scores and clinician assessments.

Main Results:

• POPS achieved state-of-the-art performance in predicting maximum postoperative pain scores (0-10 NRS).
• The model outperformed clinician predictions across all postoperative days in prospective validation.
• While calibration was slightly degraded, POPS demonstrated interpretability, highlighting key comorbidities influencing pain.

Conclusions:

• POPS offers a fully automated, pre-operative solution for predicting postoperative pain using readily available EHR data.
• This machine learning approach enhances pain management strategies by identifying at-risk patients.
• POPS provides valuable, interpretable insights into factors contributing to acute postoperative pain.