Screening hypertension using non-laboratory risk factors with machine learning: a retrospective cross-sectional study in Indonesia.

OBJECTIVE

This study aimed to screen for hypertension in a vast Indonesian population using machine learning (ML) and 11 non-laboratory risk factors, validating the results through internal and external validations.

SETTING AND PARTICIPANTS

From the initial 1 782 365 participants aged 15 and above registered at the Integrated Counseling Post primary care centres across Indonesia from 2014 to 2017, incomplete data and outliers were excluded, and 268 210 participants were included in our analysis. The dataset was split deterministically into a dataset for training using 10-fold internal cross-validation of 204 315 participants and another dataset for external validation of 63 895 participants.

DESIGN

This retrospective cross-sectional study used three ML algorithms, that is, random forest, gradient boosting and extreme gradient boosting (XGBoost), and compared them against logistic regression as a benchmark to screen hypertension based on the WHO and International Society of Hypertension criteria. The importance of the risk factors was ranked. By partly using continuous versus categorical age, waist circumference (WC) and body mass index (BMI) risk factors, we evaluated the screening performance regarding sensitivity and area under the receiver operating characteristic curve (AUC).

RESULTS

The external validations revealed that the XGBoost model performed the best in hypertension screening. The external validation, which partly uses continuous variables, provides 0.97 sensitivity and 0.75 AUC, indicating excellent screening capability. The importance rank of the risk factors was consecutively family history of hypertension (FH-HTN), age, WC, BMI, occupation, education, sex, smoking, low physical activity, lack of fruit or vegetable intake and alcohol consumption.

CONCLUSIONS

By using 11 easy-to-collect non-laboratory risk factors, the ML model successfully screens for hypertension with better performance than the benchmark. Using the numerical variables of age, WC and BMI yields a better discrimination capability than the categorical variables. FH-HTN and age are the two top risk factors for the development of hypertension. This study is a useful academic exercise and shows ML's importance in handling large data sets.