Probabilistic forecasting of surgical case duration using machine learning: model development and validation.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Operations Management

Background:

• Accurate surgical case duration estimation is crucial for efficient operating room utilization and cost-effectiveness.
• Traditional methods often lack the precision needed for complex scheduling and operational decisions.

Purpose of the Study:

• To develop and evaluate a novel machine learning (ML) approach for predicting the continuous probability distribution of surgical case durations.
• To compare the performance of the ML model against existing statistical and tree-based methods.

Main Methods:

• A mixture density network (MDN), a type of ML model, was trained using structured (e.g., ASA status, inpatient status, age) and unstructured (e.g., procedure name, diagnosis) features from 53,783 surgical cases.
• The MDN's predictive accuracy was assessed using continuous ranked probability score (CRPS) and pinball loss (PL), and compared to tree-based and Bayesian methods.
• Feature importance was analyzed to identify key predictors for surgical case duration.

Main Results:

• The MDN achieved the best performance with a CRPS of 18.1 minutes, outperforming tree-based methods (19.5-22.1 minutes) and a Bayesian method (21.2 minutes).
• The MDN demonstrated superior accuracy across all quantiles for both common and rare surgical procedures.
• Scheduled surgery duration and procedure name were identified as the most influential features for the MDN's predictions.

Conclusions:

• Machine learning, particularly the MDN approach utilizing natural language processing of surgical descriptors, can effectively predict the probability distribution of surgical case durations.
• This advanced predictive capability offers significant potential for optimizing surgical schedule design and informing day-of-surgery operational choices.