Esin Yiğit, Prediction of Surgical Durations Using Machine Learning Methods

M.S. Candidate: Esin Yiğit
Program: Bioinformatics
Date: 13.01.2026 / 10:00
Place: A-212

Abstract: Accurate prediction of surgical case durations is essential for effective operating room (OR) scheduling and hospital resource management. However, many hospitals still rely on manually entered surgery times, which contain errors and cannot be a part of proper OR scheduling. This thesis proposes a machine learning (ML) framework that uses Radio Frequency Identification (RFID) derived operational data to estimate surgical durations more accurately. The dataset consists of more than thirty thousand surgeries collected through passive RFID tags attached to patients’ wrists, providing automated timestamps which have no manual effect, alongside metadata such as surgeon, anaesthesiologist, patient demographics and surgery information. Data preprocessing included outlier removal, numerical encoding of categorical features, frequency-based encodings and creation of a historical mean duration feature. Four predictive models, Linear Regression, Decision Tree (CART), Random Forest, and XGBoost, were trained and evaluated using MAE, MSE, and RMSE metrics. Linear Regression has the most limited suitability among the models (MAE ~65 minutes), while non-linear models captured the variability of surgical workflows better. CART reduced MAE to ~52 minutes while Random Forest reduced it to ~46 minutes, and XGBoost achieved the best performance with a MAE of ~45 minutes, which was still too high for short surgeries. Finally, stratified modeling is applied and results were significantly accurate such as MAE of ~11 minutes for short surgeries and an overall MAE of ~30 minutes. The results indicate that ML models which are trained on RFID-based operational data outperform traditional estimation methods and can provide more reliable OR scheduling. This study highlights the value of automated time capture systems and demonstrates how combining RFID data with ensemble methods can improve prediction accuracy and reduce inefficiencies in perioperative management.