We analyzed a deidentified dataset from annual health checkups conducted at the Preventive Medicine Center (Ningen Dock Division), Yoshida Hospital, Keiyukai Medical Corporation (Asahikawa, Hokkaido, Japan). The analytic cohorts comprised examinees from 2021 (n=3744), 2022 (n=5153), and 2023 (n=5352), each analyzed as an independent annual cohort.

The study was approved by the Institutional Review Board of Yoshida Hospital, Keiyukai Medical Corporation (Approval No. 20251002001) and conducted in accordance with the Declaration of Helsinki. Data were deidentified prior to analysis. Written informed consent, including consent for secondary use of de-identified data, was obtained at the time of the health checkup. No additional interventions or participant contact occurred as part of this study.

Thirteen routine clinical indicators were included: BMI, waist circumference, systolic blood pressure, diastolic blood pressure, fasting plasma glucose, hemoglobin A1c (HbA_1c), triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, aspartate aminotransferase, alanine aminotransferase, γ-glutamyl transferase, and uric acid. Sex-specific thresholds were applied for waist circumference according to institutional criteria (see Table S2B in Multimedia Appendix 1).

Eight lifestyle questionnaire items were included (eg, smoking, alcohol consumption, breakfast habits, snacking, eating speed, mastication, physical activity/walking, and motivation for health improvement). Lifestyle questionnaire items were coded as binary variables according to the facility codebook. Variable definitions are provided in Table S2A in Multimedia Appendix 1.

Continuous variables were median-imputed and standardized within each training fold of each annual cohort. The imputation and standardization parameters were estimated within the training fold and applied to the corresponding held-out fold to prevent information leakage.

Indicator-level missingness rates are summarized in Table S1 in Multimedia Appendix 1. Missingness in indicators other than HbA_1c was low (<2% in each y).

The primary outcome was defined as a composite abnormality label derived from routine health-check classification rules used in the screening program. For each clinical domain (eg, glucose metabolism, blood pressure, lipids, liver enzymes, and anthropometric indices), examinees were categorized according to prespecified threshold-based grades (A–D).

In the institutional screening system, grade A indicates no abnormality; grade B indicates mild abnormality typically requiring lifestyle guidance; grade C indicates follow-up or re-evaluation; and grade D indicates recommendation for further diagnostic evaluation or treatment.

The composite outcome was binarized using an OR rule: participants were labeled outcome-positive if any domain met or exceeded the predefined abnormality threshold (grade B or higher); otherwise, they were labeled outcome-negative. This definition was applied consistently across all annual cohorts to evaluate structural reproducibility of the operational framework rather than severity-specific prognostic discrimination. Detailed domain-specific thresholds corresponding to grade B or higher are provided in Table S2B in Multimedia Appendix 1.

The grading thresholds (A–D) were defined according to the standardized health-check classification framework established by the Japanese Society of Ningen Dock and Preventive Medicine, which is based on national health screening standards and specialty society guidelines. These classifications are widely used in routine health-check programs across Japan to guide follow-up recommendations (eg, observation, repeat testing, referral, or treatment). This study adopted these externally defined operational criteria without modification and dichotomized grade B or higher to capture any clinically relevant abnormality that warrants structured follow-up under this program. Participants categorized as grade E (under active treatment) were excluded from the outcome classification process.

HbA1c was structurally missing for a subset of participants due to the program’s screening policy-based test selection. Indicator-level missingness rates are summarized in Table S1 in Multimedia Appendix 1.

Because the composite outcome was defined using an OR rule across multiple domains, outcome ascertainment did not depend solely on HbA_1c.

To evaluate robustness to structural HbA1c missingness, we conducted a prespecified sensitivity analysis excluding HbA1c from the predictor set while maintaining an identical evaluation pipeline (HbA_1c included vs excluded).

A composite score was generated from standardized features using a genetic algorithm (GA). All stochastic components were controlled by fixing the random seed (SEED=42) for both Python’s random module and NumPy. Fitness estimation used stratified K-fold cross-validation with shuffling (random_state=42).

The GA-derived composite score was subsequently entered into a logistic regression model to generate calibrated predicted probabilities within a prespecified out-of-fold (OOF) validation framework (Platt scaling [3]).

Bayesian updating was applied post hoc for interpretability purposes to the calibrated predicted probabilities and did not influence GA optimization or probability estimation.

For each fold, the model was trained on the training subset and evaluated on the corresponding held-out subset. OOF predictions were aggregated across folds to obtain a single internally validated prediction for each participant within each annual cohort.

Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC) calculated from OOF predicted probabilities generated within the prespecified cross-validated pipeline.

Overall predictive performance was quantified using the OOF-based Brier score. Calibration was examined descriptively using calibration plots based on OOF predicted probabilities (Figure S1 in Multimedia Appendix 1). No additional recalibration or threshold optimization was performed beyond the prespecified validation framework.

To improve statistical transparency, 95% percentile bootstrap confidence intervals for OOF AUC and Brier score were computed using 2000 participant-level resamples within each annual cohort and HbA_1c condition (ON/OFF). Performance metrics were recalculated from the previously generated OOF predicted probabilities without refitting the model, thereby preserving internal validation and avoiding information leakage while maintaining the integrity of the original cross-validated predictions.

Year-stratified OOF performance estimates and confidence intervals are reported in Table S4 in Multimedia Appendix 1 and prespecified ON–OFF differences are summarized in Table S5 in Multimedia Appendix 1.

All analyses were performed in Python (version 3.13.5) using scikit-learn (version 1.6.1) [4] and DEAP (version 1.4) [5]. Analyses were conducted in a Jupyter-based environment (Anaconda distribution). Detailed GA configuration parameters, including population size, number of generations, weight initialization range, crossover and mutation settings, and random-seed control, are provided in Multimedia Appendix 1 (“Genetic Algorithm Implementation”) to facilitate reproducibility.

Across the three annual cohorts, sample sizes ranged from 3744 to 5352 individuals, with outcome prevalence between 36.7% and 37.9%, indicating comparable class balance across years.

Under the standardized OOF validation framework, discrimination remained stable across cohorts. OOF AUC values were 0.810 (2021), 0.814 (2022), and 0.812 (2023), with overlapping 95 % bootstrap confidence intervals (Table 1). Brier scores ranged from 0.172 to 0.176 across cohorts, indicating stable overall predictive performance across years (Table 1).

Corresponding calibration plots are provided in Figure S1 in Multimedia Appendix 1.

In the prespecified HbA_1c ON/OFF sensitivity analysis, exclusion of HbA1c resulted in minimal changes in discrimination and Brier score (Table S5 in Multimedia Appendix 1), suggesting limited sensitivity of model performance to policy-driven structural HbA1c missingness under the harmonized OOF validation framework.

As the primary objective was to assess temporal reproducibility under standardized analytical procedures, formal hypothesis testing of between-year differences was not performed.

This study evaluated the temporal reproducibility of a genetic algorithm (GA)–derived composite health risk score across three consecutive annual health checkup cohorts under a prespecified OOF validation framework. Cross-validated OOF AUC values ranged from 0.810 to 0.814, with overlapping bootstrap confidence intervals, indicating stable discrimination under standardized analytical procedures. Earlier exploratory analyses yielded higher apparent AUC values; however, these were not derived exclusively from OOF predictions under the same pipeline and are therefore not presented as primary performance estimates. The primary contribution of this study is methodological: it demonstrates that a GA-derived score can achieve stable OOF discrimination across consecutive cohorts when preprocessing, optimization, and evaluation are uniformly applied, rather than emphasizing peak performance under heterogeneous analytical conditions.

Because performance estimates were based exclusively on OOF predicted probabilities, the evaluation preserved internal validation and minimized information leakage. Variability observed in earlier exploratory analyses likely reflected differences in analytical procedures rather than underlying cohort characteristics, highlighting the importance of consistent preprocessing, optimization, and evaluation when assessing artificial intelligence–based risk stratification models.

Importantly, the outcome definition reflects operational screening classification rather than confirmed clinical diagnoses. The composite label was constructed using threshold-based OR combinations across correlated clinical domains, and the dominant abnormality domain contributing to classification may vary across cohorts depending on distributional shifts. In this context, AUC values represent the model’s ability to consistently reconstruct the structured screening framework under harmonized analytical conditions rather than discrimination of severity-specific disease states.

Methodologically, GA optimization produced a composite score from standardized predictors, which was subsequently mapped to calibrated predicted probabilities using logistic regression. Bayesian updating was applied post hoc as an interpretability add-on to the calibrated predicted probabilities and did not influence GA optimization or probability estimation. The explicit specification of evolutionary hyperparameters, weight initialization range, and random-seed control further strengthens the reproducibility of the optimization procedure and reduces the likelihood that the observed discrimination reflects stochastic artifacts.

The prespecified HbA1c ON/OFF sensitivity analysis showed minimal changes in AUC and Brier score, suggesting limited sensitivity to policy-driven structural HbA1c missingness under the standardized evaluation framework; however, re-optimizing GA weights under an altered feature set represents a distinct modeling exercise.

This study has limitations. It was conducted at a single center in Japan using an occupational health checkup population. The outcome was cross-sectional, and prospective or external validation was not performed. In addition, because grade B or higher includes heterogeneous categories with varying clinical implications (ranging from lifestyle guidance to referral for diagnostic evaluation), model discrimination reflects detection of operational abnormality rather than exclusively high-severity disease states. Furthermore, because the outcome was defined using threshold-based criteria that partially overlap with included predictors, discrimination should be interpreted as reconstruction of the operational classification within the same workflow rather than independent prognostic discrimination. This endpoint is an operational classification designed for workflow consistency, not a clinically adjudicated diagnosis. Accordingly, the present OOF-based results provide an internally validated assessment of temporal reproducibility within this screening system.

In conclusion, under a prespecified, harmonized OOF validation framework within a single institutional screening system, the GA-derived composite risk score demonstrated stable temporal discrimination and consistent overall predictive performance across three consecutive annual cohorts. These findings support methodological reproducibility and structural consistency under standardized analytical procedures. However, the present results do not establish external generalizability or clinical effectiveness, and independent external and prospective validation is required before broader clinical implementation can be inferred.