ORAHS2025
Abstract Submission

116. Evaluation of the performance of machine learning models for the prediction of the operating room occupation time of non-elective surgeries

Invited abstract in session TA-1: Surgery scheduling 1, stream Sessions.

Tuesday, 9:00-10:30
Room: NTNU, Realfagbygget R5

Authors (first author is the speaker)

1. Anem Dupré
CentraleSupelec
2. Oualid Jouini
Laboratoire Genie Industriel, Ecole Centrale Paris
3. Guillaume Lamé
Laboratoire de Génie Industriel, CentraleSupélec
4. Thomas Botrel
CHU Pitié-Salpêtrière, APHP

Abstract

The optimal planning of non-elective surgeries is made difficult by a lack of visibility on patients’ future resource needs. Our study aims at improving the pre-operative visibility of schedulers through the prediction of the operating room occupation time (OT) and post-operative length of stay. We report initial results on OT prediction.

Four machine learning models (MLMs) were selected based on their performance on similar problems: Ridge Regression (RR), Random Forest (RF), XGBoost (XGB), and Multilayer Perceptron (MLP). Using nested cross-validation and a 20% validation set, the models were fitted on the data of patients that underwent non-elective surgery in a large French public teaching hospital between 2015 and 2018.

We included OT for 3,053 patients. On the validation set, the algorithms predicted durations within 20% of actual times (within 20 min for OTs below 100 min) for 58.6% of cases (95% CI [58.1, 590]) for RR, 59.2% [58.5, 59.9] for RF, 59.3% [58.8, 59.8] for XGB and 56.7% [55.8, 57.8] for MLP, respectively. In comparison, using the average of the previous 5 similar surgeries yielded a 48.6% [48.0, 49.4] performance. Surgeons’ predictions before surgery obtained 45.5% [44.3, 46.5] performance. Mean Absolute Percentage Error followed similar trends. Performance varied between surgical specialties.

MLMs improve predictions compared to surgeon’s estimations by around 10%. Further work is needed to better practical usefulness.

Keywords

Status: accepted


Back to the list of papers