Can machine learning predict oxygenation trouble during one-lung ventilation?
A retrospective Chinese cohort reported AUC 0.924 for a random forest, compared with 0.790 for logistic regression, but it remains an internally tested hypothesis rather than a bedside tool.
The short version
- What happened
- The retrospective study analysed 519 thoracic-surgery patients and compared five algorithms for predicting oxygenation impairment during one-lung ventilation.
- Who it affects
- Patients may eventually benefit from earlier warning during thoracic surgery, but an unvalidated alert could prompt unnecessary changes.
- What remains unproven
- Retrospective single-centre cohort
Keep in mind
Retrospective single-centre cohort
No independent external validation or intervention trial
ROC-AUC does not establish calibration
SHAP importance is associative
Source analysis · 1 linked source. Primary analysed source date: 25 October 2026.
Read the full report · 2 min The evidence, methods, implications and limitations
The model found non-linear patterns in 519 operations
Researchers at the Affiliated Hospital of Panzhihua University analysed records from 519 people undergoing thoracic surgery with one-lung ventilation. They compared five algorithms using ROC-AUC. Random forest performed best at 0.924, while conventional logistic regression reached 0.790.
SHAP explanations ranked positive end-expiratory pressure, preoperative arterial oxygen partial pressure and body-mass index as leading predictors. The model suggested U-shaped patterns around positive end-expiratory pressure of 4–6 centimetres of water and tidal volume, plus interactions between lung conditions and ventilation settings.[1]
Prediction is not a prescription for changing ventilation
An accurate warning could focus preparation during a vulnerable part of surgery. But SHAP describes how a model used correlated variables; it cannot show that changing one setting will prevent impairment. Some inputs may also reflect care decisions rather than information available at the intended warning time.
The evidence is an internal retrospective comparison from one hospital, without an external cohort, prospective alert test or patient-outcome evidence. Confidence requires pre-specified time-zero features, external testing across centres and devices, calibration reporting and a controlled impact study.[1]
The distinction that matters
What the evidence supports
The paper moves from recorded data to prediction and explanation; it does not test a treatment rule.
Records
Retrospective data from 519 operations
Single-centre cohort
Prediction
Compare five algorithms
Random-forest AUC 0.924
Action
Change ventilation or care
Not tested prospectively