This retrospective cohort study enrolled 1,023 gastric cancer patients who underwent radical gastrectomy at The First Affiliated Hospital of Anhui Medical University (n=815) and Chuzhou First People’s Hospital (n=208) from January 2020 to January 2025. Ethical approval (No. 83220465) was granted by the Institutional Review Board of The First Affiliated Hospital of Anhui Medical University, with waived informed consent in accordance with the Declaration of Helsinki. Patients were selected based on strict criteria:

Inclusion criteria: (1) Histopathologically confirmed non-metastatic gastric adenocarcinoma; (2) Radical gastrectomy performed; (3) No neoadjuvant radiotherapy or chemotherapy; (4) Complete clinical and follow-up records.

Exclusion criteria: (1) Non-gastric malignancy or distant metastasis; (2) Incomplete body composition data [BFR, BMI, Visceral Fat Density (VFD), Visceral Fat Area (VFA)]; (3) Concurrent malignancies.

Data were extracted from electronic medical records using structured templates and verified by two certified researchers. Collected features included: (1) Demographics: Age, sex. (2) Clinicopathological characteristics: AJCC 8th edition TNM stage (8), tumor location (cardia/body/antrum), histological grade (well-moderate/poor-undifferentiated), surgical approach (laparoscopic/open), lymphadenectomy extent (D1/D2). (3) Body composition metrics: BMI (kg/m²) calculated from preoperative height/weight. VFD (g/cm³), BFR (%), VFA (cm²), Subcutaneous Fat Area (SFA, cm²), and Skeletal Muscle Area (SMA, cm²) derived from preoperative abdominal CT scans at L3 level using Slice-O-Matic software. (4) Preoperative laboratory markers (within 1 week pre-surgery): Serum albumin (ALB, g/L), prealbumin (PA, mg/L), hemoglobin (Hb, g/L), white blood cell count (WBC, ×10⁹/L), C-reactive protein (CRP, mg/L). (5) Outcome: Postoperative complications within 30 days (Clavien-Dindo grade ≥II defining clinically significant events) (9).

This study minimized data loss at the source through stringent inclusion and exclusion criteria. In the final analytical sample, only a few variables exhibited minimal random missing values (all missing rates ≤6.0%). To ensure sample completeness and statistical power for subsequent analyzes and predictive models, different single imputation strategies were adopted based on variable types and distribution characteristics: mean imputation was used for normally distributed continuous variables (e.g., SMA, ALB, Hb), while median imputation was applied to non-normally distributed continuous variables (e.g., BFR, VFA, VFD, WBC, CRP). All categorical variables were complete in this dataset; however, planned mode imputation would address minor missingness if present, as detailed in Table 1.

815 patients from The First Affiliated Hospital of Anhui Medical University served as the training set, while 208 patients from Chuzhou First People’s Hospital constituted the testing set. All models were implemented in MATLAB 2024b. Five-fold cross-validation was applied to the training set to mitigate overfitting, and Synthetic Minority Over-sampling Technique (SMOTE) was executed on the training set to address class imbalance.

We propose an AutoML framework driven by the Improved Hike Optimization Algorithm (IHOA) to jointly optimize feature selection and hyperparameter tuning for predicting postoperative complications. The original Hike Optimization Algorithm (HOA) is a novel metaheuristic algorithm inspired by natural aurora phenomena. Key improvements were implemented: (1) Population initialization using Logistic chaotic mapping, mathematically defined as xn+1=μxn(1−xn) (where μ∈(3.57,4] denotes the chaotic parameter), which significantly enhances initial traversal in the solution space and facilitates escaping local optima; (2) A dynamic Lévy flight strategy to adjust search step sizes, computed as step=α⊕Le'vy(β) (where α is the scale parameter, ⊕ represents the dot product, and Le'vy(β) follows a Lévy distribution with exponent β). This strategy employs larger steps for global exploration in early iterations and adaptively reduces step sizes for local exploitation in later stages, achieving dynamic exploration-exploitation balance. The enhanced IHOA was first evaluated on the CEC2022 benchmark test functions to validate its superiority over the original HOA and other mainstream metaheuristic algorithms. At the empirical level, dual validation was performed through clinical prediction models: robustness testing via artificial injection of progressive data perturbations (0%–15% noise and 0%–30% missing values), alongside a feature selection module assessing model complexity control capabilities.

The core of this AutoML framework is a two-stage optimization workflow (pseudocode in Appendix A). In Stage 1 (discrete space optimization), IHOA is applied to feature selection, where each candidate solution (i.e., individual position) is encoded as a binary vector representing a feature subset. Model-based feature importance weights (e.g., XGBoost) serve as prior knowledge to guide the IHOA search, with the fitness function being a weighted combination of performance metrics (e.g., AUC) evaluated on the validation set using a base classifier (e.g., logistic regression). This selects high-weight feature subsets with strong discriminative power and low redundancy. In Stage 2 (continuous space optimization), IHOA optimizes hyperparameters of the target prediction model—integrating classical algorithms including support vector machines, random forests, and gradient boosting trees—for the optimal subset identified in Stage 1. Hyperparameter combinations are encoded as continuous vectors to maximize prediction performance on the validation set. Finally, for comprehensive evaluation, the proposed framework was compared against six widely-used machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and LightGBM. All models were evaluated using identical training-testing datasets to ensure fairness. The technical workflow is shown in Figure 1.

A multidimensional evaluation system was established: (1) Classification performance: For the predictive model, six core metrics were adopted to systematically assess discriminative ability and stability in class-imbalanced scenarios: Accuracy (ACC), Sensitivity (SEN), Specificity (SPE), Precision (PRE), F1 score (harmonic mean of precision and recall), Area Under the ROC Curve (ROC-AUC), and Area Under the Precision-Recall Curve (PR-AUC). (2) Calibration performance: Calibration curves combined with the Brier score (lower scores indicate higher prediction accuracy) were used to evaluate probability prediction precision; the Hosmer–Lemeshow test (χ² statistic) and Spiegelhalter test (z statistic) quantified calibration deviation (P > 0.05 indicated good calibration). (3) Clinical application: Decision Curve Analysis (DCA) was applied to quantify clinical utility by calculating the net benefit (NB) at different threshold probabilities:

where TP is true positives, FP is false positives, N is total sample size, and pt is the risk threshold. By comparing NB with reference lines for traditional intervention strategies, the effective interval for model-assisted decision-making was validated.

Post AutoML feature screening, robustness was verified via LASSO regression, followed by SHapley Additive exPlanations (SHAP) for clinical interpretability: AutoML feature screening: Automatically identified prognosis-associated features. LASSO regression: Validated feature sparsity and stability using regularization. SHAP analysis: By calculating Shapley values for each feature, quantitative attribution of feature contributions to individual predictions was achieved; summary plots were comprehensively used to display global feature importance rankings, while waterfall plots, decision path plots, and force plots were employed to visually illustrate the interpretation process of specific prediction cases, thereby systematically revealing the model’s intrinsic decision-making logic.

A proof-of-concept clinical decision support tool was developed using MATLAB 2024a App Designer due to its accelerated prototyping workflow for model-visualization integration and seamless compatibility with existing statistical preprocessing pipelines (Statistical Package for the Social Sciences, SPSS), integrating the optimal machine learning model.

All study data were uniformly imported into the SPSS 26.0 statistical analysis platform for standardized processing. Continuous variables with normal distribution were expressed as mean ± standard deviation (x̄± s), while those with non-normal distribution were denoted as median (interquartile range) (M [IQR]); categorical variables were expressed as frequency and percentage (n(%)). For inter-group comparisons, continuous variables first underwent normality testing: if both groups conformed to normal distribution, univariate t-test was applied; otherwise, Mann-Whitney U test was used. Categorical variables were compared using Pearson’s chi-square test. Statistical significance was determined based on p-values (two-tailed test, significance threshold α=0.05), and results were presented in tabular formats.

A total of 1,023 subjects were included, with a mean age of 62.65 ± 10.16 years. Among them, 686 (67.06%) were male and 337 (32.94%) were female. The comparison of general characteristics between the training set and testing set is presented in Table 2. The results demonstrate balanced distributions across most demographic and clinical baseline features (P > 0.05), with statistically significant differences observed only in isolated variables. Such limited discrepancies following random partitioning more closely reflect the natural variability inherent in real-world data, avoiding idealization biases that may arise from artificial overmatching. This consequently enhances the extrapolation potential and reliability of the models developed in this study for broader clinical scenarios.

The IHOA demonstrated significant optimization advantages through standardized testing (see Appendix B). By artificially injecting progressive data perturbations, this study evaluated the performance of the prognosis prediction model based on real-world clinical data. Experimental results revealed that all intelligent algorithms exhibited declining performance trends as perturbation intensity increased (Figure 2). Under original data conditions, the five algorithms achieved an average prediction accuracy of AUC = 0.84; however, under extreme perturbations (15% noise + 30% missing values), model performance substantially decreased to an average AUC = 0.46—a relative decline of 45.4%. Critically, the IHOA algorithm exhibited optimal stability: its prediction accuracy decreased from AUC = 0.91 under original conditions to AUC = 0.73 in high-perturbation scenarios, with a relative decrease rate (19.7%) markedly lower than other algorithm groups (all >36%).

At the key feature identification level, the IHOA algorithm demonstrated superior concise characteristics (Figure 3). It consistently screened 8 highly correlated predictors (average correlation coefficient r=0.67, SD = 0.08)—a 26% reduction compared to the average 10.8 features selected by other algorithms. This feature refinement mechanism effectively reduced model complexity while maintaining predictive accuracy.

Note: This figure demonstrates the stability performance of five intelligent algorithms (IHOA/HOA/ALO/HHO/WOA) under increasing data interference scenarios. The boxplots represent the distribution of cross-validated AUC values from 10 repeated experiments, while the scatter points indicate single experimental results. Algorithm color coding: IHOA (forest green), HOA (blue), ALO (red), HHO (purple), WOA (cyan). The red dashed line indicates the clinically acceptable threshold (AUC = 0.7), values below which are considered insufficient for clinical reference.

Note: This study compared the correlation strength between 18 clinical features and postoperative complication risk, and evaluated the feature selection performance of the IHOA algorithm against other optimization algorithms (HOA, ALO, HHO, WOA). The left panel displays the correlation coefficients (r) between each feature and the outcomes, with dark blue indicating highly correlated features (r>0.5), light blue indicating weakly correlated features (r ≤ 0.5), and the gray dashed line representing the correlation threshold of 0.5. The right panel shows the feature selection results of the algorithms, where dark green squares represent features selected by IHSO, and light blue squares represent features selected by other algorithms.

Six machine learning models were systematically evaluated on the training set for accuracy, sensitivity, specificity, F1-score, and area-under-curve metrics. The AutoML model demonstrated comprehensive superiority, achieving ROC-AUC of 0.9580 and PR-AUC of 0.9577 (Table 3; Figures 4A, B). Notably, AutoML exhibited exceptional F1-score (0.8855), indicating strong clinical utility in balancing precision-recall tradeoffs. The algorithm ultimately selected eight key features: BFR, VFA, VFD, SMA, CRP, lymphadenectomy extent, BMI and Age.

In independent testing, AutoML maintained robust performance (ROC-AUC = 0.9380, PR-AUC = 0.9262; Figures 4C, D). Decision curve analysis (Figure 4E) revealed greater net clinical benefit across risk thresholds (1%–93%) compared to conventional methods, with sustained high-level stability confirming superior generalizability. Calibration curves (Figure 4B) demonstrated optimal probabilistic prediction (lowest test-set Brier score = 0.111).

Calibration statistics for all models are presented in Table 4. The Brier score demonstrated that the AutoML model (0.111) achieved the highest prediction accuracy, significantly outperforming other models (reduction of approximately 32–49%). Calibration intercept analysis revealed that all models exhibited a mild overestimation tendency (intercepts ranged from –0.151 to –0.052). The AutoML’s calibration intercept was closest to the ideal value of 0 (–0.052), with minimal calibration bias, demonstrating significant superiority over other models. Calibration slope analysis indicated that AutoML’s slope was closest to the ideal value of 1 (1.082), confirming its optimal balance between discrimination and calibration.

The Hosmer-Lemeshow test showed that AutoML (χ²=8.450, p=0.391) passed the test with the highest p-value, indicating optimal calibration goodness-of-fit; LightGBM (χ²=10.820, p=0.212) and XGBoost (χ²=11.400, p=0.180) also passed the test, suggesting good calibration, while other models exhibited calibration bias (p<0.05). Collectively, across all calibration metrics, the AutoML model demonstrated optimal performance in every dimension, validating its excellence in both prediction accuracy and calibration capability.

A proof-of-concept prototype visualization tool was developed that integrates the seven core predictors, providing support for analyzing the model’s decision logic in a research context. As demonstrated in Figure 7, clinicians input feature values on the dashboard to instantly compute postoperative complication risk.