Can machine learning predict intractable hypoxemia before repeat lung surgery?

Pulmonary function tests alone fail to flag risk

Repeat pulmonary resection (RPR) means a second lung operation in a patient who has already lost lung tissue. During one-lung ventilation (OLV), the surgeon deliberately collapses the operative lung so the other lung can be worked on; the remaining ventilated lung must then carry the whole gas exchange load. Intractable hypoxemia is the failure of that arrangement: blood oxygen falls and does not recover with routine correction.

The paper's first result is a negative one, and it matters for how the rest should be read. Using five standard preoperative pulmonary function test (PFT) parameters — FVC, FEV1, FEV1/FVC, DLCO and MVV — the authors trained 19 different machine learning learners and evaluated them with 10×10 repeated cross-validation in the training cohort. Every one of those 19 models performed poorly, with AUROC below 0.6 and AUPRC below 0.2.

That comparison sets the baseline the rest of the study must beat. A model that merely reproduces spirometry adds nothing, because the clinical question is precisely whether risk can be seen before the operation in patients whose routine lung function looks acceptable.

Adding CT volumetry and two time points

The authors then rebuilt the models with a wider feature set drawn from two moments: before the first procedure (FP) and before the second procedure (SP). From 282 candidate features, 37 with plausible clinical relevance and acceptable missingness entered a wrapper-based selection step, in which each learner repeatedly proposed feature subsets and kept the one with the best AUPRC under 10-fold cross-validation. AUPRC, the area under the precision-recall curve, was chosen over AUROC because the outcome is imbalanced and the practical question is how well true cases are found.

The distinctive inputs are CT-derived volumetric measures of the ventilated lung (VL): absolute VL volume, the VL volume ratio (VL volume divided by total lung volume), and VL lung function measures that multiply that ratio by FVC or FEV1. These describe how much functional lung will remain to oxygenate blood once the operative side is collapsed, which spirometry alone does not capture.

Among the 19 learners, the Multivariate Adaptive Regression Splines (MARS) model offered the best trade-off between performance and simplicity, using only five predictors: age, history of coronary heart disease, FP operation time, SP VL volume ratio, and SP left lung volume. Variance inflation factors ranged from 1.01 to 1.16, indicating the five were not redundant with each other.

Discrimination holds across three external cohorts

In the training cohort (1,639 patients, 218 events, 13.3% incidence), the final MARS model reached an AUROC of 0.790 (95% CI 0.755–0.824) and an AUPRC of 0.445 (95% CI 0.374–0.514) against a baseline incidence of 0.133. Because the AUPRC baseline equals the event rate, that figure represents a clear lift over chance-level precision.

The three independent validation cohorts differed substantially in how often the outcome occurred, which makes their results worth reading separately rather than as one number. Validation cohort 1 (146 patients, 51 events) had a 34.9% incidence and an AUROC of 0.750 with AUPRC 0.641; cohort 2 (133 patients, 23 events) had 17.3% incidence, AUROC 0.765 and AUPRC 0.529; cohort 3 (122 patients, 22 events) had 18.0% incidence, AUROC 0.782 and AUPRC 0.473. Discrimination was therefore stable in ranking terms, while the precision-recall figures moved with each cohort's underlying event rate.

Calibration was acceptable: the training cohort had a calibration slope of 1.000 and Brier score 0.094, and the external cohorts had slopes of 0.678 to 0.922, intercepts of 0.257 to 1.005, and Brier scores of 0.115 to 0.210. Decision curve analysis showed the lower bound of the 95% CI for net benefit staying above the treat-none line between threshold probabilities of 0.1 and 0.25 in all cohorts.

The five predictors, not the algorithm, do the work

A reader could easily conclude that a machine learning model outperformed conventional statistics here. The paper undercuts that reading itself: a logistic regression model built on the same five predictors performed comparably to MARS across all cohorts, and MARS added only modest gains in discrimination.

The authors therefore describe the machine learning pipeline mainly as a systematic framework for selecting predictors and comparing models, not as a route to large performance gains through model complexity. The practical implication is that the value lies in which variables were chosen — particularly the CT-derived ventilated-lung volumes measured before the second operation — rather than in the nonlinear fitting.

Risk rises as ventilated lung volume falls

Partial dependence plots describe the marginal effect of each retained predictor on predicted risk. Risk increased substantially as SP VL volume ratio and SP left lung volume decreased, and it rose as FP operation time lengthened. Patients without a history of coronary heart disease had a lower predicted probability than those with it.

Age behaved non-linearly, rising to a peak between 60 and 70 years and then falling. The authors interpret the decline after 70 as selection bias rather than protection: older candidates undergo specific preoperative evaluation, so those who reach surgery represent a selected group with cardiopulmonary function sufficient to tolerate the procedure. That explanation is the authors' interpretation of an observed curve, not a demonstrated causal mechanism.

A 12% cutoff favors sensitivity over precision

The authors fixed a probability cutoff of 0.12, chosen by maximizing the Youden Index in the training cohort. Patients above it would be flagged before surgery to prompt planning such as choosing the lung-isolation strategy, preparing CPAP or PEEP, closer intraoperative monitoring, ICU or step-down planning, and more informed consent discussions.

The reasoning is explicitly asymmetric: because intractable hypoxemia is severe and potentially life-threatening, a triage tool should favor sensitivity, accepting more false positives to avoid missing vulnerable patients. F1 scores across validation cohorts ranged from 0.475 to 0.602 at this threshold, which reflects the sensitivity and positive predictive value balance rather than overall accuracy.

The outcome itself was defined strictly to avoid counting transient desaturations: SpO2 below 90% during OLV plus either return to two-lung ventilation, documented difficulty oxygenating, or SpO2 below 92% for at least 10 minutes. This definition excludes brief events from probe displacement or poor peripheral perfusion, but it also means the model predicts this operational definition, not every form of intraoperative desaturation.

Retrospective evidence awaiting prospective testing

All results come from retrospective cohorts at seven Chinese tertiary hospitals between June 2013 and September 2023, and the authors state that the model has so far been evaluated only in retrospective cohorts and requires prospective validation. The cohorts also differed markedly in baseline characteristics — for example, anesthesia type ranged from predominantly inhalational in the training cohort to almost entirely combined intravenous-inhalational in validation cohort 3 — so transportability to other practice settings is untested.

The clinical value of the model therefore rests on a specific comparison rather than on the label "machine learning": a five-variable set that includes CT-derived ventilated-lung volumes measured before the second operation clearly outperformed PFT-only models, and a simple logistic regression on the same five variables would likely perform similarly. The next question worth watching is whether prospective use of the 0.12 threshold actually changes perioperative planning and outcomes.

Sources used in this answer