Development of a machine learning model to predict the probability of health checkup participation in Japan.

Objectives: Enhancing health checkup participation is crucial for early detection and treatment of noncommunicable diseases and for improving public health. Effectively increasing health checkup rates requires identifying and encouraging individuals likely to adopt health-oriented behaviours. We aimed to develop a machine learning model to predict the participation probability in a specific health checkup in the following year.

Study design: Retrospective cohort study.

Methods: We analysed data from 58,863 National Health Insurance-insured individuals in Kochi Prefecture, Japan, who underwent specific health checkups during the fiscal years (FYs) 2013-2017. The dataset includes physical measurements, blood pressure measurements, blood and urine tests, and self-reported questionnaires. Predictive models for FY2018 participation were developed using LightGBM and evaluated using the area under the receiver operating characteristic curve (AUC) and reliability curves. SHAP was used to assess the feature's importance. External validation for FY2019 and FY2020 assessed temporal robustness.

Results: Predictive accuracy for FY2018 was high, with AUCs of 0.824 (95 % confidence interval [95 % CI]: 0.813-0.835) for men and 0.820 (95 % CI: 0.810-0.830) for women. External validation of FY2019 showed AUCs of 0.821 and 0.807 for men and women, respectively. In FY2020, prediction accuracy declined, with AUCs of 0.798 and 0.794 for men and women, respectively. Key predictive features included years since the last checkup, past checkup frequency, age, systolic blood pressure, and lifestyle factors.

Conclusions: By developing an accurate model to predict future health checkup participation, we identified a novel indicator that enables efficient, optimized recommendations and may help improve participation rates.