A preoperative pulmonary risk score reaches international validation
A new prognostic tool for elective surgery places postoperative pulmonary complications on a more measurable footing before a patient reaches the operating room. The GSU-Pulmonary Score was developed to estimate the 30-day risk of a composite outcome that included pneumonia, acute respiratory distress syndrome, and unexpected mechanical ventilation, using variables available at the time a surgical decision is made. In a setting where elective backlogs have stretched operating capacity and intensified triage decisions, a score that can separate low-risk from higher-risk patients before incision carries obvious operational value.
Pulmonary complications remain one of the major drivers of postoperative death, and they are clinically awkward in a way many surgical complications are not: they can emerge suddenly, escalate resource use fast, and force escalation to intensive care even after an apparently routine procedure. A model that captures that risk before surgery can inform consent, bed planning, and whether a case should proceed in a lower-acuity setting or with additional postoperative monitoring.
A global derivation cohort built from more than 86,000 patients
Model development used data from the GlobalSurg-CovidSurg Week dataset, assembled in October 2020, with candidate predictors drawn from 86,231 patients across 1,158 hospitals in 114 countries. Thirteen candidate predictor variables entered the development process. Two machine learning strategies were explored: XGBoost and the least absolute shrinkage and selection operator, or LASSO. The LASSO model ultimately offered the best balance of performance, interpretability, and parsimony, and became the final GSU-Pulmonary Score.
The choice matters clinically. A model may perform well on paper yet fail at the bedside if it requires variables that are hard to gather, hard to standardize, or unavailable at the point of preoperative review. Here, the investigators deliberately favored a score that could be deployed using simple predictor variables already in hand when a surgeon, anesthetist, or perioperative team is deciding whether elective surgery should proceed.
The overall rate of pulmonary complications in the derivation cohort was 2.0%. Even at that relatively low absolute rate, the consequences are severe enough to justify formal risk estimation, because the affected subgroup is the one most likely to consume postoperative respiratory support and critical care resources.
Discrimination held up across pandemic and pre-pandemic validation cohorts
External validation was performed in two further prospective cohorts. The first, CovidSurg-Cancer, included 30,492 patients from 726 hospitals in 75 countries during the COVID-19 pandemic. The second, RECON, included 6,789 patients from 150 hospitals in three countries before the pandemic.
The score’s performance remained acceptable in both settings, though not identically so. In CovidSurg-Cancer, the AUROC was 0.746, with a Brier score of 0.036, CITL of 0.109, and calibration slope of 1.056. In RECON, discrimination was lower, with an AUROC of 0.716, Brier score of 0.045, CITL of 1.040, and calibration slope of 1.009. Those values show a model that retained meaningful separation between higher- and lower-risk patients, but one that did not map perfectly onto every population.
The divergence is clinically important. A score can preserve discrimination while drifting in calibration, and calibration is what tells a clinician whether the absolute risk estimate is close enough to reality to guide a concrete decision such as postoperative bed allocation or escalation planning. In this study, the score remained useful, but the RECON findings show that local recalibration may be needed before the tool is used as a hard threshold for triage.
LASSO offered a simpler model without sacrificing much performance
The comparison between LASSO and XGBoost is one of the most practical parts of the analysis. XGBoost and LASSO produced very similar discrimination in the derivation data: AUROC 0.786 for LASSO and 0.785 for XGBoost. Yet LASSO was more explainable and required fewer covariables, which favors real-world adoption in surgical clinics and preassessment pathways where transparency matters.
The final score included ten predictor variables. Its internal validation remained strong, with AUROC 0.773, Brier score 0.020, calibration in the large of 0.034, and calibration slope of 0.954. Those metrics suggest that the model was not merely fitting noise from a single dataset. It retained discrimination, stayed reasonably calibrated, and avoided the opacity that can make more complex machine learning methods difficult to trust in high-stakes preoperative decision-making.
This is exactly where pulmonary risk prediction can become useful. Patients with elevated pulmonary vulnerability often need tailored anesthetic planning, more cautious fluid management, intensified postoperative surveillance, and clearer discussion of the possibility of respiratory support. Even when surgery still proceeds, the conversation changes because the expected postoperative course is no longer being guessed at from intuition alone.
Why a pulmonary score matters during surgical recovery backlogs
The score was deployed as an online web application, which gives it a direct route into workflow if hospitals choose to use it. The intended use is straightforward: preoperative estimation of the risk of 30-day pulmonary complications in adult patients undergoing elective surgery across operation types and disease categories. Because the predictor set is simple, the score can be applied before the day of surgery rather than after the patient has already entered the perioperative pathway.
That timing is particularly relevant during periods of backlog recovery. Hospitals expanding elective volume often need to distinguish patients who can safely proceed in standard pathways from those who may require more resources, different site selection, or delayed scheduling until the system can provide adequate postoperative support. A preoperative model does not make the decision for the team, but it gives the team a common numerical reference point.
The development and validation design also strengthens confidence in the tool’s generalizability. The cohorts spanned global, pandemic-era, and pre-pandemic settings, which gives the score a broader evidentiary base than a single-country or single-procedure model could offer. Still, the variation in calibration across validation cohorts shows that model transportability is not automatic. A score built for wide use still has to earn its place in a local environment.
The bedside question now is implementation, not invention
The score’s main contribution is not simply that it predicts postoperative pulmonary complications. It does so using variables available before elective surgery, with discrimination that held up across very large international cohorts and with enough simplicity to be plausible at scale. The study also shows the limits of any one-size-fits-all risk tool: calibration drift appeared in RECON, meaning sites may need to recalibrate the model before relying on absolute probabilities.
Surgical teams now have a tool that can support consent, resource allocation, and prioritisation, including during SARS-CoV-2 outbreaks. The harder work is local: deciding how the score will be embedded into preassessment, whether it changes booking behavior, and whether it improves patient flow without delaying operations that remain safe to perform. Those are the next questions for implementation, not for derivation.