Development and validation of machine learning models to improve prediction of surgical duration for cataract surgeries

Abstract

Introduction: Operating theatres significantly contribute to hospital expenditures. Traditional surgery scheduling, which is often based on manual methods or historical averages, lacks precision. This inefficiency impacts cost, timely care delivery and patient experience, particularly in high-volume surgeries. To address this, we aimed to create data-driven predictive models for cataract surgery durations, a largely unexplored area in local literature.

Methods: We utilised supervised machine learning models, including linear regression, random forest and extreme gradient boosting (XGBoost). Performance metrics were accuracy (within ± 10 min of actual surgery time) and mean squared error (MSE). The dataset, post-outlier removal, comprised 4242 cataract cases (80% training sets and 20% test sets).

Results: Compared to the embedded scheduling algorithm in EPIC (an electronic health records system) and surgeons' predictions, machine learning models displayed superior performance, achieving 40% and 20% greater accuracy compared to EPIC and surgeons, respectively, with a significantly lower MSE. Machine learning models' error margin primarily ranged from 0 to 5 min. Notably, underestimation beyond the ± 10-min threshold occurred in about 9% of cases.

Conclusion: In this study, machine learning models have been shown to be more effective in predicting cataract surgery durations compared to current methods, offering practical benefits for optimising operating theatre management. The use of machine learning significantly improves the accuracy of surgery duration estimates.

Keywords: Cataract surgery; healthcare optimisation; mathematical modelling; operating theatre scheduling; predictive analytics.