What is the clinical evidence from this study?

Study design: Cohort. Population: Diabetes mellitus in patients undergoing non-cardiac surgery (n=4293). Intervention: AdaBoost machine-learning model vs. LASSO logistic regression. Primary outcome: Impaired outcome (composite of major adverse cardiovascular and cerebrovascular events) (AUC 0.82, 95% CI 0.78-0.86).

What question did this study set out to answer?

The aim is to develop machine-learning models to predict impaired outcomes in diabetic patients undergoing non-cardiac surgery.

June 24, 2026Open Access

Machine-learning prediction of impaired outcome in diabetic patients undergoing non-cardiac surgery

Key Result

The AdaBoost machine-learning model predicted impaired postoperative outcomes in diabetic patients undergoing non-cardiac surgery with an AUC of 0.82, outperforming the penalized linear baseline.

Key Points

The aim is to develop machine-learning models to predict impaired outcomes in diabetic patients undergoing non-cardiac surgery.
Developed five machine-learning models including LASSO logistic regression, AdaBoost, LightGBM, XGBoost, and CatBoost.
Used a dataset of 1,117 diabetic patients and performed internal model comparisons with receiver operating characteristic analysis.
Utilized Boruta and SHAP analysis to identify key predictors influencing risk.
AdaBoost achieved an AUC of 0.82 (95% CI, 0.78-0.86) with specificity 0.76 and sensitivity 0.72.
Key predictors included prior ischemic stroke, myocardial infarction, preoperative creatinine, and inflammatory markers.
Impaired outcomes were associated with older age, higher ASA class, comorbid cardiovascular issues, and intraoperative hypotension.

Study Design

Type

Cohort (n=4,293)

Structured PICO

Do machine-learning models improve prediction of impaired outcomes in diabetic patients undergoing non-cardiac surgery?

Population

1,117 diabetic patients undergoing non-cardiac surgery

Intervention

Machine-learning models (AdaBoost, LightGBM, XGBoost, CatBoost)

Comparator

LASSO logistic regression (sparse penalized baseline)

Outcome

Database-defined composite impaired outcomecomposite

Machine-learning models, particularly AdaBoost, offer modest improvements in predicting perioperative composite impaired outcomes in diabetic patients undergoing non-cardiac surgery.

Main Result

Effect estimate: AUC 0.82 (95% CI 0.78-0.86)

Limitations

Retrospective design, meaning associations cannot be interpreted as causal effects
Database-derived composite endpoint with unrecoverable exact follow-up duration
Lack of a fully reproducible record of cohort assembly, missing-data handling, and hyperparameter tuning
Absence of external validation and calibration statistics
Medication variables and coded procedure indicators may encode institutional practice patterns that shift across settings

Abstract

Background: Diabetic patients undergoing non-cardiac surgery are vulnerable to postoperative cardiovascular and cerebrovascular complications, yet risk stratification tools tailored to this population remain limited. This study aimed to develop interpretable machine-learning models for predicting impaired outcome in diabetic patients undergoing non-cardiac surgery. Methods: = 1,117) according to a database-defined composite impaired outcome recorded during the available follow-up of the source dataset; the exact follow-up duration for the composite endpoint could not be recovered from the finalized analytic extract. Five models were internally compared, including LASSO logistic regression, AdaBoost, LightGBM, XGBoost, and CatBoost. LASSO tuning used cross-validation. Model performance was assessed using receiver operating characteristic analysis and decision-curve analysis, while Boruta and SHAP were used to identify and interpret key predictors. Results: Patients with impaired outcome were older, had lower body weight, higher ASA class, heavier cardiovascular comorbidity burden, and greater intraoperative hypotension exposure. Among the five models, AdaBoost showed the best discrimination with an AUC of 0.82 (95% CI, 0.78-0.86), specificity of 0.76, sensitivity of 0.72, and PPV of 0.69. LightGBM, XGBoost, and CatBoost also outperformed the sparse penalized baseline. Boruta and SHAP consistently highlighted prior ischemic stroke, prior myocardial infarction, preoperative creatinine, albumin, inflammatory markers, age, ASA class, heart-rate and systolic blood-pressure summaries, and BUN as major contributors to risk. Conclusion: Interpretable ensemble learning provided modest internal improvement in perioperative risk stratification in diabetic patients undergoing non-cardiac surgery. Impaired outcome appeared to be driven by the combined burden of established vascular disease, reduced renal and nutritional reserve, inflammatory activation, and intraoperative hemodynamic instability. Further calibration testing, external validation, and fully reproducible pipeline reporting are needed before clinical deployment.

Demander à l'IA

Bookmark

View Full Paper