A random forest machine learning model accurately predicted post-operative hypotension in patients with type 2 diabetes mellitus (T2DM) undergoing non-cardiac surgery and maintained performance across internal, prospective temporal, and external validation cohorts, according to a study published in Cardiovascular Diabetology. 

“The tool provided by this study has the potential to assist clinical practice on multiple levels,” wrote first author Yu Gao, MD, of the General Hospital of Central Theater Command of the People's Liberation Army in Wuhan, China, and colleagues. “For anesthetists and surgeons, it could help preoperatively identify T2DM patients at high risk who require special attention; intraoperatively, it could provide dynamic risk assessment; and post-operatively, it could enable early warning.” 

Researchers analyzed routinely collected perioperative data from 44,540 patients with T2DM who underwent non-cardiac surgery between 2012 and 2025. The retrospective development cohort included 34,012 patients, while a prospectively collected temporal validation cohort comprised of 10,528 patients. An additional external validation cohort included 2,156 patients treated at a second hospital.
  
Post-operative hypotension was defined as systolic blood pressure below 90 mm Hg during the post-anesthesia care unit (PACU) stay. After evaluating 113 candidate variables through a four-step feature-selection process, the researchers identified 13 predictors for the final model. These included obstructive sleep apnea, anticoagulant use, intraoperative blood loss, history of cancer, crystalloid administration, cirrhosis, calcium channel blocker use, prior stroke, post-operative respiratory rate at PACU admission, body mass index, steroid use, congestive heart failure, and age. 

The researchers compared 14 machine learning approaches based on discrimination, calibration, and clinical utility before using SHapley Additive exPlanations (SHAP) to interpret the predictions of the best-performing model.

Random forest model outperformed the other algorithms, achieving an area under the receiver operating characteristic curve (AUROC) of 0.843 in the training cohort, 0.854 in the internal validation cohort, and 0.847 in the prospective validation cohort.
  
External validation at an independent hospital yielded an AUROC of 0.822, with sensitivity of 88% and specificity of 62%, demonstrating preserved performance in a separate clinical setting. Decision curve analysis showed the model demonstrated positive net benefit across risk thresholds, indicating clinical utility, according to the researchers.

SHAP analysis ranked intraoperative blood loss as the strongest predictor of post-operative hypotension, followed by age, heart failure, obstructive sleep apnea, and body mass index. Global and patient-level SHAP analyses illustrated how these variables contributed to each prediction, allowing clinicians to better understand the factors underlying individual risk estimates. The dependence plots also demonstrated interactions among several predictors, including body mass index and calcium channel blocker use. 

The researchers also examined model performance according to surgical urgency. The random forest model maintained good discrimination for both elective and emergency procedures, although performance was modestly lower in emergency surgery, with an AUROC of 0.858 and 0.825, respectively. Sensitivity remained above 90% in both groups, and feature importance rankings were highly consistent, suggesting stable predictive mechanisms across surgical settings. A sensitivity analysis excluding patients with preoperative isolated hypotension produced nearly identical performance metrics, supporting model robustness. 

The researchers noted several limitations. The model was developed primarily from data at a single institution, with external validation limited to one additional hospital. The study also did not include race, ethnicity, insurance status, or socioeconomic variables, limiting assessment of fairness and generalizability. In addition, the model predicts hypotension only during the PACU stay and does not address delayed hypotension after transfer to the surgical ward. Finally, as an observational study, it cannot determine whether implementation of the model improves patient outcomes. 

"While further multi-center validation is warranted, our findings suggest that the model offers a promising analytical framework for understanding post-operative hemodynamic risk in T2DM patients, with the potential to support individualized perioperative management," wrote Dr. Gao and colleagues. 

The researchers reported no conflicts of interest. 
 
(Editor’s note: The researchers noted that the study manuscript will undergo further editing before final publication, and there may be errors present that affect the content. All legal disclaimers apply.)