This study employed a retrospective cross-sectional design. Clinical data were collected from patients with diabetes who underwent elective non-cardiac, non-neurological surgeries at our institution between January 2020 and December 2023.

Patients or the public WERE NOT involved in the design, or conduct, or reporting, or dissemination plans of our research.

Inclusion criteria were: 1. Aged 18–75 years, regardless of gender. 2. Undergoing elective non-cardiac and non-neurological surgery. 3. Diagnosed with Type II diabetes. Exclusion criteria were: 1. Preoperative pulmonary infection. 2. Body mass index (BMI) ≥ 35 kg/m² 3. Significant abnormalities in cardiac, pulmonary, hepatic, or renal function. 4. Physical status classification ≥ IV. 5. Missing Hemoglobin A1c (HbA1c) data. 6. Excessive missing data.

Relevant patient information was extracted from the hospital’s electronic medical record system. Specific data sources included: Inpatient medical records; Surgical reports; Anesthesia records; Laboratory test reports; Imaging examination reports; Nursing records.

The collected data primarily encompassed: 1. Demographic characteristics: Age, gender, height, weight, etc; 2. Diabetes-related information: Medication adherence, insulin therapy usage; 3. Comorbidities: Hypertension, coronary heart disease, COPD; 4. Laboratory tests: HbA1c, average blood glucose levels within 3 days postoperatively; 5. Surgical information: Surgical department, duration of surgery, type of anesthesia; 6. Preoperative assessment: ASA classification, New York Heart Association (NYHA) classification, preoperative oral carbohydrate, baseline blood pressure, baseline heart rate; 7. Postoperative management: Use of PCA, transfer to the ICU after surgery; 8. Other: Blood transfusion during perioperative period.

Data Cleaning: 1. Identification and correction of obvious data entry errors; 2. Handling of duplicate records.

Missing Value Handling: 1. For variables with a missing rate <5%, median or mean imputation was applied; 2. For variables with a missing rate between 5% and 10%, the decision to impute or exclude was based on the variable’s importance and the pattern of missingness; 3. Variables with a missing rate >10% were excluded.

Outlier Handling: All outlier data points were removed to minimize their impact on the analysis.

Feature Engineering: 1. Standardization of continuous variables: Z-score standardization method was used; 2. Encoding of categorical variables: One-hot encoding was applied; 3. Creation of composite features: BMI based on height and weight, Mean Arterial Pressure (MAP) derived from systolic and diastolic blood pressures.

The primary outcome of this study was the incidence of pulmonary infections diagnosed by the operating surgeon before patient discharge. PPI were defined as new-onset pulmonary infections diagnosed postoperatively.

We selected nine commonly used machine learning algorithms to build our predictive models: K-Nearest Neighbors (KNN): An instance-based learning algorithm. Support Vector Machine (SVM) with Linear Kernel: A supervised learning algorithm for classification. Random Forest (RF): An ensemble learning method based on decision trees. Decision Tree (DT): A tree-based classification algorithm. Light Gradient Boosting Machine (LightGBM): A gradient boosting framework. Ada Boost (ADA): An ensemble boosting algorithm. Naive Bayes (NB): A probabilistic classifier based on Bayes’ theorem. Logistic Regression (LR): A statistical method for binary outcome prediction. Linear Discriminant Analysis (LDA): A method for finding linear combinations of features that separate classes.

Then, LASSO logistic regression was used to select PPI features (Mullah et al., 2021). The dataset was randomly divided into a training set and a testing set at a ratio of 3:1. The model was trained and optimized using 10-fold cross-validation on the training set.

The performance of the models was evaluated using the following metrics: 1. AUC; 2. Accuracy; 3. Precision; 4. Specificity; 5. PPV; 6. NPV.

To enhance the interpretability of the models, SHAP (SHapley Additive exPlanations) values were utilized to analyze the contribution of each feature to the prediction results.

Based on the predictions from the machine learning models and the analysis of feature importance, we aimed to: 1. Identify high-risk patient groups. 2. Quantify the impact of each risk factor. 3. Develop targeted preventive measures and intervention plans.

Continuous variables were summarized using means ± standard deviation (SD) for normally distributed data or medians with interquartile ranges (IQR) for non-normally distributed data. Categorical variables were described using frequencies and percentages. Machine learning models were built and evaluated using R version 4.2.2 with the caret package and the pROC package. A p-value <0.05 was considered statistically significant.

This study was approved by the Ethics Committee of the First Affiliated Hospital of Chengdu Medical College (2023CYFYIRB-BA-May03) and registered in the Chinese Clinical Trial Registry (ChiCTR2400080801, www.chictr.org.cn). The requirement for informed consent was waived due to the retrospective nature of the study. All procedures were conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines (von et al., 2007).

Between January 2020 and December 2023, a total of 19,690 patients with diabetes underwent surgical treatment at our institution. After screening, 1,269 patients were ultimately included for analysis. The patient enrollment flowchart is shown in Figure 1, and the baseline characteristics are presented in Table 1.

Of the 1,269 patients included in the study, 187 (14.7%) developed PPI. Patients who developed PPI were significantly older (64.24 ± 8.37 vs. 60.43 ± 9.96 years, P < 0.001), had a higher ASA score (P < 0.001), and were more likely to have COPD (9.62% vs. 2.87%, P < 0.001) compared to those who did not develop PPI. The duration of surgery was also significantly longer in the PPI group (359.90 ± 171.12 vs. 237.86 ± 125.61 min, P < 0.001).

We used LASSO (Least Absolute Shrinkage and Selection Operator) logistic regression for feature selection, implemented using the glmnet package in R. The optimal regularization parameter λ was selected through cross-validation using cv. glmnet function. Feature importance was ranked based on the absolute values of their non-zero coefficients.

Due to the potential negative impact of correlated or less important features on the performance of machine learning models (Figure 2), we conducted feature selection and ranked the importance of the features.

Using the LASSO-Logistic feature selection method, we identified 12 optimal features from the initial 28. These features performed exceptionally well in predicting the risk of PPI in patients with diabetes: Sex; Postoperative ICU admission; MAP; Systolic Blood Pressure (SBP); ASA classification; COPD; NYHA classification; BMI; Preoperative oral carbohydrate loading; Surgical departments; Age; Duration of surgery.

To further simplify the model and enhance its interpretability, we selected six of the most important features. These features exhibit the highest importance in predicting PPI: Postoperative ICU admission; Age; ASA classification; COPD; Surgical department; Duration of surgery.

Using these recommended features for model training not only ensures model performance but also simplifies the model structure, enhancing its interpretability. These results provide strong support for subsequent clinical decision-making and the development of personalized preventive strategies.

Random Forest analysis of the six selected features revealed their relative importance through Mean Decrease in Gini index (Figure 3). Surgical department emerged as the most influential predictor, followed by surgery duration and age, while ICU admission, COPD status, and ASA classification showed relatively lower contributions. This ranking suggests that surgical type and procedural factors have stronger predictive power than patient baseline characteristics.

The performance metrics of each model on the train and test set are summarized in Figure 4. The Ada Boost (ADA) model performed the best across all metrics, with its ROC curve illustrated in Figure 4.

Our LASSO-Logistic model identified six critical predictors of PPI risk in diabetic patients: 1. Surgical Department: Thoracic surgery carried the highest risk (72.81% vs 5.33% in other departments, P<0.001), followed by gastrointestinal surgery (31.76%) and orthopedics (12.93%), likely due to direct respiratory system involvement and procedure complexity. 2. Postoperative ICU Admission: Strongly indicative of case complexity and patient vulnerability (11.23% in PPI group vs 1.48% in non-PPI group, P<0.001). These patients typically require more intensive monitoring and have higher severity of underlying conditions. 3. Age: Significantly higher in PPI group (64.24 ± 8.37 vs 60.43 ± 9.96 years, P<0.001), associated with immunosenescence, pulmonary function decline, reduced recovery capacity, and increased comorbidities. Advanced age particularly impacts respiratory defense mechanisms and wound healing. 4. ASA Classification: Higher proportion of Grade III in PPI group (55.61% vs 31.15%, P<0.001), reflecting poorer overall health status, multiple systemic comorbidities, and increased complication susceptibility. 5. COPD Status: Markedly higher prevalence in PPI group (9.62% vs 2.87%, P<0.001), significantly elevating PPI risk due to impaired lung function, compromised respiratory defense mechanisms, and chronic airway inflammation. 6. Surgery Duration: Substantially longer in PPI group (359.90 ± 171.12 vs 237.86 ± 125.61 min, P<0.001). Longer procedures increase pathogen exposure, tissue trauma, anesthesia-related risks, and physiological stress response.

To ensure data consistency and representativeness, we analyzed the distribution of patient characteristics across different surgical departments and risk groups. The 1,269 patients were distributed across thoracic surgery (8.98%, n = 114), orthopedics (20.72%, n = 263), gastrointestinal surgery (6.70%, n = 85), and other departments (63.60%, n = 807). The incidence of PPI varied significantly among departments: thoracic surgery (72.81%), orthopedics (12.93%), gastrointestinal surgery (31.76%), and other departments (5.33%). This variation reflects the inherent risks associated with different surgical procedures and was successfully captured by our model, with the surgical department being identified as the most important predictor (Figure 3). Key clinical characteristics also showed consistent patterns between PPI and non-PPI groups. Age (64.24 ± 8.37 vs 60.43 ± 9.96 years, P<0.001), ASA classification (Grade III: 55.61% vs 31.15%, P<0.001), and surgery duration (359.90 ± 171.12 vs 237.86 ± 125.61 min, P<0.001) were significantly different between groups, indicating robust discriminative features for our prediction model.

These findings offer valuable insights for clinical practice: 1. Risk Stratification: Enables targeted preventive measures for high-risk patients. 2. Age-Adapted Care: Tailored strategies including enhanced respiratory care, early mobilization, and optimized medication and nutrition for older patients. 3. Preoperative Optimization: Thorough evaluation and management, especially for patients with high ASA scores or COPD. 4. Surgical Planning: Where clinically appropriate, consider minimizing surgery duration and optimizing technique. 5. Specialized Protocols: Implement enhanced postoperative care and monitoring, particularly for thoracic surgery patients.

This study demonstrates significant methodological and clinical strengths by rigorously comparing nine advanced machine learning algorithms and using a large, diverse sample across multiple surgical specialties. LASSO feature selection and SHAP value analysis improve model efficiency and interpretability. Translating findings into actionable preventive strategies bridges the gap between advanced analytics and clinical practice, highlighting the study’s innovative approach and practical relevance to postoperative care for diabetic patients.

However, our study has several limitations. First, this is a retrospective single-center study without external validation. Our patient population and institutional-specific practice patterns (such as surgical protocols and diagnostic criteria) may affect model generalizability. Second, the model’s performance in real-world clinical settings remains untested. Its applicability may vary across different healthcare settings, geographic regions, and institutions. These limitations underscore the need for future external validation studies. Future research should focus on: (1) external validation through multi-center studies, (2) prospective evaluation of clinical effectiveness, and (3) assessment of model performance in different healthcare settings. These steps are crucial for validating the model’s clinical utility (Ramspek et al., 2021).