Анестезия и клинические исследования
Открытый доступ

Абстрактный

Improving and Interpreting Surgical Case Duration Prediction with Machine Learning Methodology

Jesyin Lai, Jhao-Yu Huang, Shu-Cheng Liu, Der-Yang Cho, Jiaxin Yu

Objective: Hospitals encounter challenges in performing efficient scheduling and good resource management to ensure advanced healthcare quality is provided to patients. Operating room (OR) scheduling is important as it affects workflow efficiency, critical care and OR optimization. Automatic scheduling and accurate surgical case duration prediction have critical roles in improving OR utilization. To estimate surgical case duration, most hospitals rely on historic averages obtained from the electronic medical record (EMR) scheduling systems. However, this produces low accuracy leading to negative impacts, e.g. rescheduling and cancellation.

Methods: A large date set, which covered various details on patients, surgeries, specialties and surgical teams, was obtained. Surgical cases within 60-600 min from 14 specialties were selected for predictive model development. These data included over 500 different procedure types. All models were evaluated with R-square (R2), mean absolute error (MAE), percentage overage (actual duration > prediction), underage (actual duration < prediction) and within. Subsequently, all selected cases were separated into cases with 1 procedure or ≥ 2 procedures and retrained with the best model.

Results: The extreme gradient boosting (XGB) model was superior, achieving a higher R2, lower MAE and higher percentage within on a time-wise testing set (not in the original data). The errors (actual - predictions) could be reduced using model retrained on cases with ≥ 2 procedures (XGB2). Interpretation of XGB predictions with Shapley additive explanations showed that procedure type, anesthesia type, and procedure no. were the top 3 most important features. Specific and higher interactions between anesthesia type, procedure no. and specialty were also identified in a subset of complicated cases.

Conclusions: The XGB and XGB2 models outperformed other models in predicting surgical case durations. They are deployed as a stand-alone machine intelligence server connected by the EMR system for scheduling. This will eventually lead to reduce medical and financial burden for healthcare management.