Prediction of peak oxygen uptake using interpretable machine learning on routinely available preoperative assessments
Se-Hun Kim1, Sa-Eun Park1, Cho Hui Hong2
1Department of Industrial Engineering, Pusan National University, Busan, Republic of Korea.
Background:
Cardiopulmonary exercise testing (CPET) is the reference standard for preoperative functional assessment before lung resection, but its use is limited by resource and practical constraints. This study developed and internally evaluated an interpretable machine learning model to estimate preoperative peak oxygen uptake (VO2peak) from routine clinical assessments, quantify the contribution of individual predictors to model estimation, and evaluate classification performance at the clinically relevant threshold of 20 mL·kg-1·min-1.
Methods:
This single-centre retrospective study included 320 consecutive patients in South Korea who underwent preoperative treadmill CPET between April 2018 and March 2024. Thirty-three routinely available predictors-including demographic characteristics, anthropometric measures, pulmonary function results, and bioimpedance-derived indices-were used to train predictive models. Model development employed 10 repeats of 5-fold nested cross-validation. The best-performing model was interpreted using Shapley additive explanations. Potential prioritisation performance was assessed by classifying patients with VO2peak below the prespecified threshold.
Results:
Random forest showed the best performance, with a root mean square error of 3.750±0.731, a mean absolute error of 2.901±0.599, and a coefficient of determination of 0.323±0.153, indicating moderate explanatory performance for VO2peak estimation. The most influential predictors were forced expiratory volume in 1 second, forced vital capacity, age, and bioimpedance-derived phase angle. At the threshold of 20 mL·kg-1·min-1, the model achieved an area under the receiver operating characteristic curve of 0.842±0.072.
Conclusions:
This predictive model showed moderate performance for estimating VO2peak, identified clinically relevant contributors, and may support prioritisation for CPET and consideration of prehabilitation assessment when resources are limited.