Prediction model for postoperative delirium risk in elderly hypertensive patients: machine learning-based development and validation Authors: Kun Wang, Zhengzheng Zhao, Jiayi Chen, Wenjie Kong, Yuanlong Wang, Yizhi Liang, Yanan Lin, Chuan Li, Jiahan Wang, Hongyan Gong, Bin Wang, Xu Lin, Yongxin Liang, Yanlin Bi Published: 2026 ABSTRACT BackgroundPostoperative delirium (POD) is a severe complication in elderly hypertensive patients, associated with poor long-term outcomes. Existing models often rely on intraoperative data, limiting preoperative risk stratification. This study aimed to develop a non-invasive machine learning model to predict POD and investigate its preoperative markers’ impact on three-year mortality.MethodsPreoperative variables were selected using LASSO regression from a cohort of 1,782 patients. Ten machine learning models were trained and validated (7:3 ratio). Model performance was evaluated via AUC-ROC and decision curve analysis (DCA). The optimal model was interpreted using SHAP values. Long-term prognosis within the POD cohort was assessed using Kaplan-Meier curves and multivariable Cox proportional hazards regression.ResultsThe POD incidence was 10.9%. The Gradient Boosting Machine (GBM) demonstrated optimal performance (AUC = 0.868, 95% CI: 0.819–0.917). SHAP analysis identified MMSE score as the most influential predictor, followed by HADS score, age, CFS, frailty, and PSQI score. Multivariable Cox analysis revealed that lower MMSE, alongside elevated HADS, CFS, frailty, and PSQI scores—but not chronological age—were independent predictors of increased three-year mortality in POD patients (all P < 0.05).ConclusionWe developed a robust machine learning tool for individualized POD prediction. Cognitive impairment, psychological distress, frailty, and poor sleep quality serve as critical dual-prognostic markers for both acute POD occurrence and long-term survival. These findings underscore the necessity of routine multidimensional preoperative assessment to facilitate personalized interventions for vulnerable hypertensive populations. Background Postoperative delirium (POD) is one of the most common complications in surgical patients, characterized by acute impairment of attention, thinking, and cognitive function, with symptoms exhibiting diurnal fluctuations (1, 2). Among patients undergoing elective major surgery, the incidence of POD is approximately 15%–25%, while it can reach as high as 50% in high-risk surgeries (3, 4). POD is not only closely associated with prolonged hospital stays, poor functional recovery, and cognitive decline, but also significantly increases healthcare costs, readmission rates, and even postoperative mortality, imposing a heavy burden on patients, families, and the healthcare system (5–7). Therefore, accurately predicting and identifying high-risk patients for POD remains a significant challenge in current clinical practice. The occurrence of POD is associated with multiple factors, including advanced age, hypertension, diabetes, and preoperative cognitive impairment (8). However, complex interactions exist among these factors, and there is currently a lack of simple and effective predictive tools for identifying high-risk patients (9). Clinically, multiple scoring scales (such as MMSE, PSQI, etc.) are commonly used to assess patients’ preoperative status and perform POD risk stratification (10, 11). However, these approaches still have limitations when addressing multidimensional, nonlinear clinical characteristics, such as polypharmacy, frailty, and sensory impairments, in elderly patients (12). In recent years, machine learning methods based on multidimensional data analysis have demonstrated significant potential in disease prediction by effectively integrating complex variable interactions (13, 14). These advanced approaches overcome the limitations of traditional models in capturing nonlinear relationships, thereby providing new directions for constructing highly accurate POD prediction tools. Hypertension, as one of the most common chronic diseases, has a clear pathophysiological link to cognitive impairment (15). Elderly hypertensive patients often exhibit impaired cerebral vasomotor regulation and neurovascular unit dysfunction, which may reduce their tolerance to surgical and anesthetic stress, thereby increasing the risk of postoperative hypotension (16, 17). Although research on POD prediction models has made some progress, current models for predicting POD in elderly hypertensive patients remain under-explored (18). General prediction models may not accurately reflect the risk characteristics of postoperative delirium in this population. To this end, this study aims to identify independent risk factors for POD in elderly hypertensive patients and to develop and validate a POD risk prediction model tailored to this population using machine learning algorithms. By systematically screening key predictive variables and evaluating their predictive value for POD and three-year postoperative mortality, we seek to provide a basis for the early and precise identification of high-risk patients and the formulation of personalized intervention strategies. Method Study population This retrospective cohort study initially evaluated a consecutive series of patients who underwent elective non-cardiac surgery under general anesthesia with endotracheal intubation at Qingdao Municipal Hospital between January 2020 and June 2022, totaling 2,000 candidates. The study protocol was approved by the Institutional Ethics Committee, and the requirement for informed consent was waived due to the retrospective and anonymous nature of the data analysis. Eligibility criteria included patients with a documented diagnosis of hypertension, aged 65 years or older, with preserved preoperative cognitive status and sufficient educational background to independently complete cognitive and psychological assessments. Exclusion criteria were predefined as: (1) emergency surgery or major cardiovascular procedures performed within the preceding month; (2) a preoperative Mini-Mental State Examination (MMSE) score < 23 (this threshold was implemented to ensure patients possessed the cognitive capacity required to reliably respond to subjective questionnaires, such as HADS and PSQI, and to prevent the confounding of acute POD with baseline severe dementia); (3) history of significant neurological disorders, including central nervous system infections, traumatic brain injury, epilepsy, or multiple sclerosis; (4) uncontrolled cardiovascular or cerebrovascular conditions, or severe hepatic and renal insufficiency; (5) pre-existing psychiatric disorders or impaired consciousness; and (6) a history of substance abuse or chronic use of psychotropic or systemic corticosteroid medications. Ethical statement This study adheres to the principles of the Declaration of Helsinki. The research protocol has been approved by the Ethics Committee of Qingdao Municipal Hospital (Approval No.: 2024-KY-094) and formally registered with the China Clinical Trial Registry (Registration No.: ChiCTR2500099473). All patient identities are replaced with hospital identification numbers and anonymized for confidentiality. Therefore, the Ethics Committee has granted exemption from written informed consent. Clinical baseline data collection Qualified anesthesiologists collected clinical variables within 24 hours before surgery. Information gathered included patient gender, age, years of education, smoking status, alcohol consumption history, physical activity levels, and past medical history of hypertension, diabetes, and coronary heart disease. Smoking status was defined as current smoking (at least one cigarette daily for over one year); alcohol consumption history was defined as drinking at least once weekly for over one year. Physical activity was assessed by frequency and duration, with regular exercise defined as at least three sessions per week of moderate-intensity physical activity lasting 30 minutes or longer per session; a history of hypertension was defined as blood pressure ≥140/90 mmHg; Diabetes history was defined as meeting any of the following criteria: fasting venous blood glucose ≥7.0 mmol/L, random venous blood glucose ≥11.1 mmol/L, or 2-hour post-oral glucose tolerance test blood glucose ≥11.1 mmol/L. Coronary heart disease history was defined as coronary artery disease confirmed by coronary angiography or clinical manifestations (e.g., typical angina symptoms, history of myocardial infarction). Questionnaires and assessment tools This study employed a series of validated standardized questionnaires to collect multidimensional health data. Sleep quality was assessed using the Pittsburgh Sleep Quality Index (PSQI; total score 0–21, >5 indicating poor sleep quality) and the Athens Insomnia Scale (AIS; total score 0–24, ≥6 suggesting insomnia), while subjective perceptions of sleep disturbance were quantified using a Numerical Rating Scale (NRS). Cognitive function was assessed using the Mini-Mental State Examination (MMSE; total score, 0–30, typically with a cutoff of <24 for cognitive impairment), and subjective complaints were evaluated using the Subjective Cognitive Decline (SCD) questionnaire. Nutritional status was assessed using the Mini Nutritional Assessment (MNA), categorized as: ≥24 points (nutritionally adequate), 17–23.5 points (at risk of malnutrition), and <17 points (malnourished). Psychological distress symptoms were screened using the Hospital Anxiety and Depression Scale (HADS), with subscale scores ≥8 indicating potential anxiety or depression. Frailty status was assessed using two tools: the Clinical Frailty Scale (CFS), which categorizes frailty severity into 1–9 levels (Level 1 being very healthy, Level 9 being terminal), and the Frail Scale, which performed rapid screening across five dimensions: fatigue, endurance, walking ability, illness, and weight loss. Physical activity levels were assessed using the International Physical Activity Questionnaire (IPAQ), with results expressed in MET-minutes per week. Primary outcome The Confusion Assessment Method (CAM) was used to identify POD. Individuals testing positive for delirium were assigned to the POD group, while those testing negative were assigned to the non-POD (NPOD) group. All subjects were assessed by trained medical personnel unaware of their perioperative status between 1 and 7 days postoperatively. Follow-up Survival time was measured from the date of surgery to either the occurrence of death or the most recent follow-up appointment. Investigators followed patients who developed POD for three years postoperatively. Data preprocessing To ensure the integrity of the predictive model, clinical records with missing values were managed using a complete-case analysis approach. Prior to feature selection and model development, all continuous variables were standardized using Z-score normalization to attain a mean of zero and unit variance, ensuring that variables with different scales contributed proportionately to the models. Feature selection Feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) regression, implemented via the “glmnet” package in R. This method is particularly effective for high-dimensional clinical data as it penalizes regression coefficients, effectively shrinking non-essential predictors to zero and reducing the risk of overfitting. To ensure the robustness of feature selection and prevent data leakage, the LASSO procedure was conducted exclusively within the training set. We employed a 10-fold cross-validation framework to determine the optimal regularization parameter (λ). The λ value that minimized the cross-validated deviance (lambda.min) was selected as the threshold for feature inclusion. Following this rigorous selection process, 14 predictors with non-zero coefficients were identified and subsequently utilized as the finalized input features for the development of all ten machine learning models. Machine learning models for POD prediction The finalized dataset, comprising 1,782 patients, was randomly partitioned into a training set (70%, n = 1,247) for model construction and an independent testing set (30%, n = 535) for performance evaluation. To maintain consistency across the feature space, all machine learning algorithms were developed and validated utilizing only the 14 optimal predictors previously identified through the LASSO regression analysis. Ten distinct machine learning algorithms were implemented and compared: a Neural Network (NN), Logistic Regression (LR), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbours (KNN), AdaBoost, LightGBM, and CatBoost. To optimize model performance and mitigate overfitting, hyperparameter tuning was conducted exclusively within the training set using a grid-search strategy combined with 5-fold cross-validation. The predictive performance of each model was rigorously quantified on the testing set using the Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score. Furthermore, the Brier score was calculated to assess model calibration, supplemented by visual inspection of calibration curves to evaluate the agreement between predicted probabilities and observed outcomes. The clinical utility and net benefit of the models were determined through Decision Curve Analysis (DCA), with confidence intervals and stability ensured via a bootstrap resampling method incorporating 50 iterations. SHAP analysis for model interpretation To enhance the interpretability of the optimal machine learning model, this study employed the Shapley Additive exPlanations (SHAP) framework. This method quantifies the contribution of each feature to individual predictions by calculating its SHAP value, thereby elucidating both the magnitude and direction (positive or negative influence) of its effect on the prediction. The analysis involved calculating SHAP values for each sample using the final predictive model. Global feature importance was then determined by aggregating the absolute SHAP values across all samples to rank the predictors. To visually communicate the complex interactions between risk factors and model outputs, a suite of graphical tools was employed, including summary plots for feature importance, dependence plots to illustrate feature-prediction relationships, and force plots to explain individual predictions. Development of a logistic regression model and clinical tools based on SHAP-identified core variables Leveraging the core predictive variables identified through the SHAP analysis, we constructed a multivariate logistic regression model to estimate the probability of POD occurrence as a binary outcome. To facilitate clinical translation, we developed a practical score-based nomogram derived from the model, which enables a visual assessment of POD risk. Furthermore, the model’s coefficients were implemented into an interactive web-based application, designed to provide clinicians with a tool for real-time, individualized risk calculation. Statistical analysis The normality of continuous variables was assessed via the Kolmogorov-Smirnov test. Data conforming to a Gaussian distribution are expressed as mean ± standard deviation, whereas non-normally distributed data are presented as median (interquartile range). Intergroup comparisons of continuous variables were performed using independent samples t-tests for parametric data or Mann-Whitney U tests for non-parametric data. Categorical variables were compared employing χ² tests or Fisher’s exact tests as appropriate. For the longitudinal prognosis analysis, Kaplan-Meier survival curves and Log-rank tests were initially employed to compare three-year mortality rates among different subgroups (e.g., Age, MMSE, HADS, CFS, Frail, and PSQI) within the POD cohort. To identify independent predictors of mortality while adjusting for potential confounders, a multivariable Cox proportional hazards regression analysis was performed. Variables demonstrated to be clinically significant or showing a p-value < 0.1 in univariate analyses were included in the multivariable model to calculate adjusted Hazard Ratios (HRs) and their corresponding 95% confidence intervals (CIs). Statistical analyses were performed utilizing SPSS (version 25.0; IBM SPSS, Chicago, USA) and R (release 4.3.1; R Statistical Computing Foundation, Vienna, Austria). A significance threshold of P < 0.05 was applied for all tests. A data analysis and statistical plan was written after the data were accessed. Results Patient screening and grouping A total of 2,000 patients were initially screened, of whom 218 (116 ineligible and 102 with missing clinical data) were excluded based on predefined criteria, resulting in 1,782 patients included in the final retrospective analysis (Figure 1; Supplementary Figure 1). Patient screening process and study design diagram. (A) Patient screening process. (B) Data collection and feature selection flowchart. (C) The study cohort was divided into training and validation sets in a 7:3 ratio, followed by the development of ten machine learning models. (D) After model comparison, the best model was selected, visualized, and the risk factors contributing most to the outcome were identified. Based on these risk factors, survival analysis and nomogram model construction were further conducted. Baseline clinical characteristics This study included 1,782 elderly hypertensive patients, among whom 194 (10.9%) developed postoperative delirium (POD). These patients were compared with 1,588 patients in the non-POD group (see Table 1). Patients with POD were older and had higher NRS, HADS, PSQI, frailty, and CFS scores, as well as a higher prevalence of subjective cognitive decline, diabetes, coronary heart disease, and malnutrition. They also had fewer years of education, lower MMSE scores, and a lower proportion of physical activity (all P < 0.05). However, no significant differences were observed between the groups in terms of AIS score, MET score, gender, smoking, alcohol consumption, or history of COVID-19 infection (all P > 0.05). Table 1Demographic and characteristics of participants.VariablesNPODPODP value1588194Age, M (IQR), years70 (67-74)74 (71-81) <0.001 NRS, M (IQR), score2 (1-3.75)3 (2-4) 0.030 AIS, M (IQR), score3 (2-4)3 (1-5)0.085MET, M (IQR), score3.5 (3-4)3 (2-4)0.540Education, M (IQR), years9 (6-10)6 (5-9) <0.001 MMES, M (IQR), score27 (26-28)25 (24-26) <0.001 HADS, M (IQR), score5 (3-6)6.5 (4-13.0) <0.001 PSQI, M (IQR), score6 (4-9)9 (6-12) <0.001 Frail, M (IQR), score1 (0-2)2 (1-3) <0.001 CFS, M (IQR), score2 (1-3)4 (3-5) <0.001 Sex (female/male)850/738106/880.769SCD, yes (%)718 (45.2)123 (63.4) <0.001 Diabetes, yes (%)529 (33.3)91 (46.9) <0.001 CHD, yes (%)846 (53.3)126 (64.9) 0.002 Cigarette smoking, yes (%)324 (20.7)35 (18.0)0.439Drinking, yes (%)271 (17.1)37 (19.1)0.485Physical activity, yes (%)499 (31.4)43 (22.2) 0.008 COVID, yes (%)1515 (95.4)187 (96.4)0.530MNA Category, n (%) <0.001 Well-nourished749 (47.2)65 (33.5)At risk of malnutrition651 (41.0)65 (33.5)Malnourished188 (11.8)64 (33.0)POD, Postoperative delirium; NPOD, No postoperative delirium; PSQI, Pittsburgh Sleep Quality Index; AIS, Athens Insomnia Scaletotal; NRS, Numerical Rating Scale; MMSE, Mini-Mental State Examination; SCD, Subjective Cognitive Decline. MNA, Mini Nutritional Assessment; HADS, Hospital Anxiety and Depression Scale; CFS, Clinical Frailty Scale; CHD, Coronary heart disease; The bold values are all p ≤ 0.05, statistically significant parameters. Demographic and characteristics of participants. POD, Postoperative delirium; NPOD, No postoperative delirium; PSQI, Pittsburgh Sleep Quality Index; AIS, Athens Insomnia Scaletotal; NRS, Numerical Rating Scale; MMSE, Mini-Mental State Examination; SCD, Subjective Cognitive Decline. MNA, Mini Nutritional Assessment; HADS, Hospital Anxiety and Depression Scale; CFS, Clinical Frailty Scale; CHD, Coronary heart disease; The bold values are all p ≤ 0.05, statistically significant parameters. Model construction and variable selection The total cohort of 1,782 elderly hypertensive patients was randomly partitioned into a training set (n = 1,247) and a testing set (n = 535) at a 7:3 ratio. Comparative analysis revealed no statistically significant differences in baseline demographic or clinical characteristics between the two cohorts (Table 2), ensuring the internal validity of the subsequent modeling process. Table 2Comparison of demographic characteristics between the training dataset and the testing dataset.VariablesTraining setTesting setP value1247535Age, M (IQR), years70 (67-75)71 (66-75)0.202NRS, M (IQR), score2 (1-4)2 (2-4)0.435AIS, M (IQR), score3 (2-5)3 (2-4)0.164MET, M (IQR), score3 (3-4)3 (3-4)0.501Education, M (IQR), years9 (5-10)9 (6-10)0.830MMES, score, M (IQR)27 (25-28)27 (25-28)0.064HADS, M (IQR), score5 (3-6)5 (3-6)0.784PSQI, M (IQR), score6 (4-9)6 (4-9)0.267Frail, M (IQR), score1 (0-2)1 (0-2)0.365CSF, M (IQR), score3 (1-3)2 (1-3)0.929POD, yes (%)136 (10.9)58 (10.9)0.982Sex (female/male)568/680258/2760.277SCD, yes (%)609 (48.8)232 (43.4) 0.038 Diabetes, yes (%)435 (34.9)185 (34.6)0.932CHD, yes (%)681 (54.6)291 (54.5)0.977Cigarette smoking, yes (%)267 (21.4)92 (17.2) 0.045 Drinking, yes (%)221 (17.7)87 (16.3)0.469Physical activity, yes (%)373 (29.9)169 (31.6)0.459COVID, yes (%)1202 (96.3)500 (93.6) 0.012 MNA Category, n (%)0.715Well-nourished567 (45.4)247 (46.3)At risk of malnutrition499 (40.0)217 (40.6)Malnourished182 (14.6)70 (13.1)The bold values are all p ≤ 0.05, statistically significant parameters. Comparison of demographic characteristics between the training dataset and the testing dataset. The bold values are all p ≤ 0.05, statistically significant parameters. For feature selection, we utilized LASSO regression analysis within the training set to mitigate potential overfitting and identify the most robust predictors for POD. The optimal regularization parameter (λ) was determined through 10-fold cross-validation, with the λ value (the threshold that minimizes cross-validated binomial deviance) selected to ensure peak predictive performance. Following this procedure, 14 preoperative predictors with non-zero coefficients were identified from the initial variable pool: sex, smoking history, history of COVID-19 infection, age, CHD, diabetes, SCD, NRS, AIS, HADS, MMSE, PSQI, frailty, and CFS score (Figure 2). These 14 variables were consistently utilized as the finalized input feature set for the development and validation of all ten machine learning models. LASSO feature selection for model construction. (A) Binomial deviance as a function of the logarithm of the λ parameter for LASSO regression. (B) Coefficient path of the LASSO regression model as a function of the regularization parameter, λ. LASSO, Least Absolute Shrinkage and Selection Operator. Model performance comparison The ROC analysis revealed that ensemble models generally performed well, with the GBM achieving the highest AUC (0.868, 95% CI: 0.819–0.917), closely followed by the Neural Network (AUC = 0.867). In comparison, the AdaBoost (AUC = 0.736) and LightGBM (AUC = 0.799) models yielded lower predictive accuracy in this context (Figure 3A). Decision curve analysis (DCA) further supported the clinical applicability of the GBM, demonstrating a relatively high net benefit across a range of threshold probabilities (Figure 3B). At an optimal probability threshold of 0.15, the GBM model achieved an accuracy of 0.905, a sensitivity of 0.919, a specificity of 0.904, a precision of 0.539, an F1-score of 0.679, and Brier score of 0.076 (Table 3). Calibration was assessed using Brier score and calibration curves (Table 3; Supplementary Figure 3). The GBM model showed acceptable calibration (Brier score = 0.076), and its calibration curve closely followed the diagonal. Machine learning model evaluation curves. (A) Comparison of ROC curves for machine learning models. Logistic, Logistic Regression; SVM, Support Vector Machine; GBM, Gradient Boosting Machine; Neural Network, Neural Network; Random Forest, Random Forest; XGBoost, Extreme Gradient Boosting; KNN, K-Nearest Neighbours; AdaBoost, Adaptive Boosting; LightGBM, Lightweight Gradient Boosting Machine; CatBoost, Categorical Boosting Algorithm. (B) Clinical Decision Curve Analysis (DCA). The x-axis represents the probability threshold for intervention selected by the clinician or decision maker, and the y-axis represents the net benefit of the decision strategy at that threshold. The higher the curve, the more effective the model. Table 3Comparison of machine learning model performance (test dataset).ModelThresholdAccuracySensitivitySpecificityPrecisionF1Brier scoreLogistic0.040.7170.8620.7000.2590.3980.068SVM0.040.7720.8100.7670.2970.4350.072GBM0.110.8410.7070.8570.3760.4910.076NeuralNetwork0.030.6630.9310.6300.2350.3750.066RandomForest0.050.7320.8790.7140.2730.4160.067Xgboost0.500.8240.6900.8400.3450.4600.247KNN0.010.7770.8100.7730.3030.4410.067Adaboost0.170.6390.7760.6220.2000.3180.082LightGBM0.010.8650.7240.8820.4290.5380.085CatBoost0.520.7720.8100.7670.2970.4350.258 Comparison of machine learning model performance (test dataset). SHAP-based model interpretability analysis To elucidate the decision-making process of the GBM model, SHAP analysis was employed to quantify feature contributions. The summary plot and importance ranking identified MMSE score as the most influential predictor, followed by HADS score, age, and geriatric vulnerabilities including CFS, Frail index, and PSQI (Figures 4A, B). Demographic and clinical feature selection using the LASSO regression. (A) SHAP summary plot of the GBM model. The color represents the value of the variable. (B) SHAP importance for each feature of the GBM model. (C) SHAP force plot of the GBM model. SHAP summary plots revealed that lower MMSE scores and elevated HADS scores exhibited the strongest positive attribution to POD risk. Notably, the model captured non-linear trends, with risk increasing significantly as cognitive reserve diminishes and psychological distress intensifies. Furthermore, individual force plots demonstrate the specific direction and magnitude of feature contributions for representative cases. For instance, in a patient with an MMSE of 30 and HADS of 6, these factors exerted a protective effect (SHAP values < 0), collectively driving the predicted risk below the baseline expectation (Figure 4C). Nomogram development and logistic regression model-based application for POD The discriminative performance of our six-predictor (age, MMSE, HADS, CFS, Frail, PSQI scores) logistic regression model was satisfactory, yielding an AUC of 0.835 (95% CI: 0.776–0.895) (Supplementary Figure 2). For clinical application, we presented the model as a visual nomogram to facilitate individualized POD risk estimation (Figure 5A). Additionally, we created an openly accessible, interactive web platform (https://wangkun.shinyapps.io/make_web/) that implements this model for real-time risk calculation (Figure 5B). Nomogram development and logistic regression model-based application for POD. (A) To estimate the POD risk, the value for each clinical variable is aligned with its corresponding “Points” axis at the top to obtain an individual score. The scores for all six variables (MMSE, HADS, Age, CFS, Frail, PSQI) are summed to yield a total points value. This total is then projected downward to the “Risk of POD” axis at the bottom to read the corresponding predicted probability. (B) The patient’s clinical parameters are entered into an online tool developed using the underlying logistic regression model. The platform automatically processes the inputs and generates a real-time POD risk estimate. Survival analysis To identify determinants of long-term prognosis, we performed a survival analysis within the POD cohort. Initial Kaplan-Meier analysis showed that lower MMSE (P = 0.002), higher HADS (P < 0.001), higher CFS (P = 0.027), higher Frailty scores (P < 0.001), and higher PSQI scores (P < 0.001) were associated with increased three-year mortality (Figure 6). Kaplan-Meier survival curves comparing three-year survival probabilities in patients with POD. (A) Survival probability stratified by MMSE score. (B) Survival probability stratified by HADS score. (C) Survival probability stratified by age. (D) Survival probability stratified by CFS score. (E) Survival probability stratified by Frail Scale score. (F) Survival probability stratified by PSQI score. Red texts indicate P-values from Log-rank tests. Abbreviation: POD, postoperative delirium; MMSE, Mini-Mental State Examination; HADS, Hospital Anxiety and Depression Scale; CFS, Clinical Frailty Scale; PSQI, Pittsburgh Sleep Quality Index. Multivariable Cox regression, adjusted for demographics (sex, smoking), preoperative comorbidities (CHD, diabetes), and clinical status (NRS, COVID-19, AIS, and SCD), identified several independent risk factors. These included cognitive impairment (MMSE: HR = 0.250, 95% CI: 0.083–0.752, P = 0.014), psychological distress (HADS: HR = 16.014, 95% CI: 5.181–49.495, P < 0.001), frailty (CFS: HR = 2.296, 95% CI: 1.088–4.847, P = 0.029; Frail: HR = 2.799, 95% CI: 1.290–6.071, P = 0.009), and poor sleep quality (PSQI: HR = 4.847, 95% CI: 1.913–12.280, P < 0.001). In contrast, advanced age was not significantly associated with three-year mortality in either the Kaplan–Meier analysis (P = 0.13) or the multivariable Cox regression (HR = 2.202, 95% CI: 0.972–4.987, P = 0.059) (Table 4). Table 4Factors associated with three-year mortality in patients with POD.VariablesCategoryHR (95% CI)P valueMMSE score<260.250(0.083-0.752) 0.014 ≥26HADS score≤716.014(5.181-49.495) <0.001 >7Age<75 years2.202(0.972-4.987)0.059≥75 yearsCSF score<52.296(1.088-4.847) 0.029 ≥5Frail score<32.799(1.290-6.071) 0.009 ≥3PSQI score<104.847(1.913-12.280) 0.001 ≥10HR, hazard ratio; CI, confidence interval; The bold values are all p ≤ 0.05, statistically significant parameters. Factors associated with three-year mortality in patients with POD. HR, hazard ratio; CI, confidence interval; The bold values are all p ≤ 0.05, statistically significant parameters. Discussion This study developed and validated machine learning models to predict POD in elderly hypertensive patients undergoing non-cardiac surgery. Among the ten models evaluated, the GBM demonstrated the best predictive performance, achieving an AUC of 0.868 (95% CI: 0.819–0.917) and superior DCA net benefit. The model incorporated 14 routinely accessible preoperative variables, providing a clinically practical tool for individualized preoperative risk stratification. The strong performance of the GBM model can be attributed to its ability to capture complex, nonlinear interactions through sequential ensemble learning (19). While the Neural Network showed a comparable AUC, GBM achieved a superior F1-score and higher net benefit in DCA, highlighting its clinical utility. These results align with our previous findings; however, the current model enhances clinical feasibility by replacing invasive CSF biomarkers with routinely accessible scales (20). Compared to an external study using GBM for POD prediction (AUC = 0.80, n=797) (13), our study utilized a substantially larger cohort (n=1782) and achieved a higher test-set AUC of 0.868 (95% CI: 0.819–0.917). The EPV ratio of 13.9:1 further indicates a low risk of overfitting. This research represents one of the most comprehensive efforts to develop a non-invasive, multidimensional prediction tool specifically for elderly hypertensive patients. The rigorous validation via ROC and DCA reinforces the model’s discriminative capacity and potential to provide meaningful net clinical benefits across diverse risk thresholds. SHAP analysis identified the MMSE score as the most influential predictor, followed by HADS, consistent with the stress–vulnerability model. Low MMSE reflects reduced neural reserve, impairing tolerance to surgical stress (21, 22). Elevated HADS may trigger HPA axis activation and neuroendocrine dysregulation, exacerbating susceptibility to POD (23–26). Furthermore, our multivariable Cox regression analysis—a significant extension requested to strengthen the evaluation of long-term prognosis—demonstrated that preoperative cognitive impairment (MMSE), psychological distress (HADS), frailty (CFS/Frail), and poor sleep quality (PSQI) are robust independent predictors of three-year mortality within the POD population. Notably, consistent with our Kaplan-Meier analysis, chronological age did not emerge as a significant independent risk factor in the multivariable model, suggesting that functional and psychological resilience may be more critical than biological age in determining the long-term survival of POD patients (27–29). Specifically, the association between lower MMSE, higher HADS, and increased mortality suggests that compromised neural and psychological reserve not only predisposes patients to acute delirium but also hampers the physiological recovery essential for long-term survival (30, 31). Psychological distress, as indicated by elevated HADS, may trigger prolonged neuroendocrine-immune dysregulation, creating a state of chronic stress that exacerbates underlying pathologies (32). Furthermore, the independent prognostic value of poor sleep quality (PSQI) and frailty (CFS/Frailty) underscores the role of systemic biological vulnerability. Chronic sleep fragmentation, reflected by high PSQI scores, is known to hinder neural repair and exacerbate systemic inflammation through NF-κB activation, thereby diminishing long-term resilience (33, 34). Similarly, frailty likely reflects underlying mitochondrial dysfunction and reduced biological capacity to withstand the catabolic stress of surgery (35). Collectively, these parameters define a “high-risk phenotype” characterized by multi-system exhaustion, which warrants integrated preoperative assessment and targeted interventions, such as cognitive prehabilitation, psychological counseling, and multimodal frailty programs. However, several limitations must be acknowledged. First, this single-center retrospective study lacked external validation. The internal validation may overestimate performance, and the model’s generalizability to other populations remains unknown. Future multicenter prospective studies are needed to verify the model’s performance across diverse settings. Second, we excluded patients with preoperative MMSE <23, this exclusion may introduce selection bias and potentially overestimate model performance by removing the highest-risk subgroup. We implemented this threshold to ensure participants possessed the cognitive capacity to reliably complete subjective scales (e.g., HADS, PSQI); nevertheless, our findings may not be directly generalizable to patients with moderate-to-severe baseline dementia. Third, the current model relies on preoperative characteristics and does not yet incorporate dynamic intraoperative factors, such as hemodynamic variability or anesthetic depth, which could further refine risk stratification in future iterations. Conclusion We developed a GBM-based machine learning model (AUC = 0.868) that accurately predicts POD in elderly hypertensive patients. MMSE, HADS, frailty, and PSQI were identified as key independent predictors for both POD and three-year mortality. These functional and psychological vulnerabilities—forming a “high-risk phenotype”—hold greater prognostic value for long-term survival than chronological age. Our results emphasize the necessity of multidimensional preoperative assessments for personalized risk management in this population.