Impact of predicting health-guidance candidates using massive health check-up data: A data-driven analysis

Starting in 2008, specific health checkups and health guidance to prevent non-communicable diseases have been provided in Japan, which has the highest proportion of elderly citizens in the world. The attendance rate for health guidance appointments is 17.7%, which is far from the national goal of the system (45%). To improve the attendance rate, we present a model for predicting whether an examinee is a candidate for health guidance; this model was based on a machine learning method and a restricted but massive amount of health checkup information.

Using machine learning methods, we developed the following five prediction models for identifying health-guidance candidates: baseline: this model included sex and age; model 1: this model included variables that can be measured in person+information on whether the examinee was a candidate in the past year; model 2: model 1+systolic blood pressure+diastolic blood pressure; model 3: model 2+all health checkup results from the past year; and model 4: model 3 using the training dataset excluding cases with missing data.

The performance levels of the five prediction models (the AUC values of the models for the test dataset) were as follows: 0.592 [95% CI: 0.586-0.596] for the baseline model, 0.855 [95% CI: 0.851-0.858] for model 1, 0.985 [95% CI: 0.984-0.985] for model 2, 0.993 [95% CI: 0.993-0.993] for model 3, and 0.943 [95% CI: 0.941-0.945] for model 4.

We studied five models for identifying health-guidance candidates. The model that used all health checkup results from the past year had the highest predictive power. Application of the prediction model developed in the present study to the selection of health-guidance candidates could reduce the cost of guidance.