Implementation and prospective performance evaluation of an intraoperative duration prediction model using high...

Area of Science:

• Anesthesiology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Accurate real-time prediction of intraoperative duration is crucial for enhancing perioperative outcomes.
• A data pipeline was developed to extract real-time data from anesthesia records.
• A predictive machine learning (ML) algorithm was implemented and deployed.

Purpose of the Study:

• To implement and evaluate a real-time ML algorithm for predicting intraoperative surgery duration.
• To assess the accuracy and stability of the ML model's predictions over time.

Main Methods:

• Clinical variables were extracted from electronic health records via a third-party platform.
• A previously developed ML model was utilized, trained on 3 months of data.
• Model performance was evaluated over 10 months using the continuous ranked probability score (CRPS).

Main Results:

• The ML model made over 6 million predictions across 62,000 procedures.
• The ML model achieved a mean CRPS of 27.19 min, significantly outperforming the scheduled duration (51.66 min).
• Linear regression confirmed no performance drift over the 10-month testing period.

Conclusions:

• A real-time ML algorithm for predicting surgery duration was successfully implemented and deployed.
• Prospective evaluation confirmed that the model's predictive performance remained stable over a 10-month period.