Random forest model predicts oxygenation impairment during one-lung ventilation in thoracic surgery

An interpretable machine learning approach has shown strong performance in identifying which thoracic surgery patients are likely to experience intraoperative oxygenation impairment during one-lung ventilation, according to a retrospective cohort study published in Scientific Reports. The condition, in which blood oxygen levels fall while a single lung is deliberately collapsed to give surgeons access to the chest, remains one of the more difficult challenges in thoracic anesthesia. Traditional statistical models often struggle to capture the complex, non-linear relationships between patient characteristics and ventilation settings that determine whether oxygenation holds up during the procedure.

The research team analyzed clinical data from 519 patients undergoing thoracic surgery and evaluated five machine learning algorithms for their ability to predict oxygenation impairment during one-lung ventilation. Model performance was measured using the area under the receiver operating characteristic curve. The random forest model achieved the highest score, at 0.924, clearly outperforming logistic regression, which reached 0.790. That gap suggests the non-linear patterns captured by tree-based methods carry meaningful predictive information that conventional linear models miss in this setting.

To make the model’s internal logic transparent, the researchers applied SHAP analysis, a technique that quantifies how each feature contributes to individual predictions. This identified positive end-expiratory pressure, preoperative arterial oxygen partial pressure, and body mass index as the primary predictors of intraoperative oxygenation impairment. The analysis also revealed potential U-shaped protective associations for positive end-expiratory pressure in the range of 4 to 6 centimeters of water and for tidal volume, alongside synergistic interactions between baseline pulmonary conditions and mechanical ventilation settings. In other words, the apparent benefit of these ventilation parameters was not uniform across their ranges, and the effect of any given setting appeared to depend on the patient’s underlying lung status.

The authors are careful to frame these findings as exploratory rather than definitive. SHAP analysis describes statistical associations and feature interactions within the data without implying biological causality, and the retrospective design means the model’s discriminative performance has not yet been tested prospectively. Even so, the study indicates that interpretable machine learning can generate hypothesis-generating insights into complex non-linear thresholds and interactions that may serve as preliminary reference information for future prospective studies on individualized anesthetic management. If such work confirms the identified thresholds, the approach could eventually help anesthesia teams anticipate which patients are at risk of oxygenation impairment during one-lung ventilation and adjust ventilatory strategies accordingly, rather than reacting to desaturation after it begins.

Subject of Research: Interpretable machine learning for predicting intraoperative oxygenation impairment during one-lung ventilation in thoracic surgery

Article Title: Interpretable machine learning for predicting intraoperative oxygenation impairment during one-lung ventilation in thoracic surgery: a retrospective cohort study

Article References: Deng, Y., Chen, J., Liu, W., & Zhu, F. (2026). Interpretable machine learning for predicting intraoperative oxygenation impairment during one-lung ventilation in thoracic surgery: a retrospective cohort study. Scientific Reports. https://doi.org/10.1038/s41598-026-74032-2

Image Credits: AI Generated

DOI: 10.1038/s41598-026-74032-2

Keywords: machine learning, one-lung ventilation, thoracic surgery, anesthesia, random forest, SHAP, oxygenation impairment, retrospective cohort study, Interpretable, machine, learning, predicting