A machine learning approach to predict hypertension using cross-sectional & two years follow up data from a health & demographic cohort of Assam, North East India

Source of data

The HDSS, originally called a population laboratory, is one of the prominent sources of socio-demographic and public health data within a community. HDSS, Dibrugarh (Dibrugarh-HDSS), launched in the year 2019, included a total of 1,06,769 participants from 60 villages and 20 tea estates in Dibrugarh District, Assam, a State in north-east India⁷. It collects information on a wide range of characteristics of the participants (>18 yr), including socio-demographic, anthropometric, behavioural, and clinical profiles such as blood pressure (BP).To facilitate data collection, 30 well-trained field assistants were appointed. The quality of collected data, including BP measurements, was assessed during the data collection time by the respective supervisors. The collection of data was done using a mobile application specially developed for the study.

After removing the outliers and participants <18 yr, 53,301 participants were identified based on completeness of relevant information, of which 14,701 were from the hypertensive group, and 38,600 were from the normotensive group. The blood pressure categories were based on the ESC-ESH Guidelines⁸. An imbalanced dataset can lead to poor model performance and affect the evaluation of the model accuracy because the model may become biased towards the majority class. To handle this problem, one of the simplest methods ‘oversampling and under sampling’ technique, was used⁹ in the present study. The basic idea behind this strategy was to resample the original dataset randomly, either by over-sampling the smallest class or under-sampling the largest class until the sizes of the classes were approximately the same. Using under-sampling data balancing technique⁹, the data set was balanced, and finally, 29,402 participants (14,701 cases from the hypertensive group and 14,701 cases from the normotensive group) were used to train and evaluate the models for identification of hypertension. Again, normotensive (n=38,600) individuals were followed up after two years, of which 1,437 were found to be hypertensive. Using over-sampling and under-sampling techniques, the dataset was balanced, and finally, 4,400 cases were used for training and evaluation of the models for predicting hypertension after two years. For internal validation, the Train-Test Split method was utilised. Seventy per cent of the data set was used for model training, and the remaining 30 per cent was used for testing and evaluating the model performance for both the data sets (baseline and 2-yr follow up). Details are highlighted in figure 1.

Close

Fig. 1.

Flow diagram showing sample sizes in different stages.

Export to PPT

Different statistical analyses were performed using the IBM-SPSS software for Windows, Version 26.0 (IBM Corp., New York, USA). The models were developed using Python (version 3.11.5) and Scikit-learn library (version 1.4.1) within the Jupyter Notebook environment. Additional libraries used include NumPy (version 1.24.3) for numerical operations and pandas (version 2.0.3) for data manipulation, and Yellow brick for data visualisation. As our primary objective of this study was to check the applicability of the ML approach in the prediction of hypertension based on some readily available features, we utilised the default hyperparameters setting as a baseline to assess the model’s performance and generate a benchmark.

Statistical test and classification models used for prediction

To identify the significant predictors (scale variables), an independent t-test was performed between hypertensive and normotensive groups. On the other hand, for categorical predictors, the chi-square test was used to determine the degree of association between these two groups. For both identification and prediction of the risk of hypertension after two years, seven classification algorithms, namely, DTC, RFC, SVM, LDA, LR, Ada-Boost classifier, and XG boost classifier were used.

Classification models used for prediction

Evaluation Measures used to evaluate the performance of the models

ROC curve (receiver operating characteristic curve), precision and recall

A receiver operating characteristic curve (ROC) is a two-dimensional graph showing the performance of a classification model at all classification thresholds. In this plot True Positive Rate (TPR) is plotted on the Y axis and False Positive Rate (FPR) is plotted on the X axis where- $\text{TPR}=\text{TP}/\left(\text{TP}+\text{FN}\right),\text{\&nbsp;FPR}=\text{FP}/\left(\text{FP}+\text{TN}\right)$

With the help of the predicted outcomes of the fitted models, the precision and recall value is calculated for each instrument, where, $\text{Precision}=\text{TP}/\left(\text{TP}+\text{FP}\right)\text{\&nbsp;Recall}=\text{TP}/\left(\text{TP}+\text{FN}\right)$

In machine learning, precision (also called positive predictive value) gives the value of the fraction of relevant instances among the retrieved instances, and recall (also known as sensitivity) gives the value of the fraction of relevant instances that were retrieved¹¹. Where both false positive and false negative are equally serious, F-1 score is an effective model evaluation measure, which is the harmonic mean of precision and recall.

Since the splitting of the dataset for training and testing is purely random, the process is repeated 100 times so that the average value and the 95% confidence intervals (CI) of the average accuracy score can be calculated using the formula $\text{CI}=\overline{x}\pm Z.\frac{s}{\sqrt{n}};$

where $\overline{x}$=average accuracy score, Z=1.96, s=sample standard deviation, n=sample size.

Variables considered as predictors

Of the wide range of data collected in Dibrugarh-HDSS, we included features that were relevant in the context of hypertension prediction. An independent t-test was performed to identify the variables having a significant difference in mean between hypertensive and normotensive participants. Similarly, we carried out a Chi-square test for categorical variables. Features having statistically significant differences in both the groups were finally used for training of the models (Table I and II). Following these basic analyses, we selected 19 variables in our study for prediction of hypertension at base line. For the prediction of the risk of hypertension after two yr a set of 19 features were considered, of which some features were common for constructing a prediction model at baseline.

Predictors of hypertension

BMI, waist circumference, and waist-hip ratio were significantly higher in hypertensive than normotensive participants. The Chi-square test revealed significantly (P<0.05) higher values for the number of variables among the hypertensive participants (Table I). The variables having statistically insignificant associations were excluded from fitting the models.

Average accuracy scores

For 100 random test samples, the model accuracy scores along with the AUC values were recorded, and their average value, as well as 95 per cent confidence interval, was determined (Table III). The Adaptive Boosting Classifier performed the best with a maximum average accuracy score of 87.02 per cent (CI: 86.01-88.03) followed by XG- boosting (86.83%) and RFC (86.08%). The performance of the other models was significantly lower than these three models in terms of their average accuracy score. However, the maximum average AUC score of 98.37 per cent (CI: 97.36-99.38) was obtained under RFC followed by XG boost (96.43%) and AdaBoost classifier (94.05%).

For a specific randomly selected test sample, the performances of the models were visualised in the confusion matrices (Supplementary Fig. 1). In confusion matrices, generated for a particular randomly selected test sample of size 8,821, it was found that the total number of false positive and negative cases was lowest (1,128) in the case of the AdaBoost classifier followed by XG-boost (1,148) and RFC (1,224). The class prediction errors were visualised diagrammatically (Supplementary Fig. 2). In the case of RFC, XG boost classifier, and Ada-boost classifier, the class prediction error was comparatively lesser than the other models. For this random test sample, the lowest number of false positive cases were observed in case of Ada-boost classifier, while the lowest number of false negative cases were observed in case of XG boost classifier.

ROC curve (receiver operating characteristic curve)

For this test sample, the ROC curve was drawn for each of the models (Supplementary Fig. 3). Here, the Ada-boost classifier had the maximum AUC of 0.9351 followed by the XG-boost classifier and RFC with AUC of 0.8344 and 0.9206 respectively. However, this AUC score was based on a particular test sample. For better comparison, average AUC scores were calculated for 100 randomly selected test samples for each of the models (Table III). It was observed that the maximum average AUC of 0.9837 was observed in the case of RFC, followed by the XG boost classifier and Ada-boost classifier.

The precision and recall value, along with the F-1 are shown in figure 2 for each of the models. It is clear from the figure that the precision and recall values for the Ada-boost classifier and RFC were significantly higher compared to those in the other models.

Close

Fig. 2.

Classification report for all the fitted models.

Export to PPT

Relative importance of the features under consideration

Features used in developing the models, relative importances were determined using random forest (visualised diagrammatically in Supplementary Fig. 4). It reveals that the waist circumference of the participants had the maximum relative importance in the identification of hypertension followed by BMI, age group, and the waist-hip ratio. The other features were showing very low relative importance compared to the above three.

Prediction of risk of hypertension after two years

All the normotensive participants were followed for two years, and 1,437 were found to be hypertensive. Using the over-sampling and under-sampling technique, the dataset was balanced and 4,400 cases were used for the training and evaluation of the models for predicting hypertension after two years. Of the 22 relevant features considered initially, we included 19 statistically significant features for developing the models to predict hypertension after two years (Table II).

Average accuracy score

For 100 randomly selected test samples, the model’s predictability in finding risk of hypertension after two years was assessed using average accuracy score and average AUC score. Table IV reveals that the XG Boost classifier performed the best with a maximum average accuracy score of 77.57 per cent (CI: 76.45-78.11) and a maximum average AUC score of 0.8508. The performance of RFC and DTC was also good than the other models. The average accuracy score of RFC and DTC was 77.2 and 75.6 per cent, respectively. It is observed that both XG-boost classifier and RFC were performing consistently better than DTC, SVM, LDA and LR in both identification as well as prediction of hypertension after two years.

For a specific randomly selected test sample of size 1,320 (consisting of 659 hypertensive participants and 661 normotensive participants), the confusion matrices for all seven classification algorithms are shown in figure 3. In the prediction of hypertension after two years, the XG-boost classifier showed the lowest number of false positive and negative cases (287) followed by RFC (301). The class prediction error for the specific test sample are shown diagrammatically in supplementary figure 5.

Close

Fig. 3.

Confusion matrices of each of the trained models for a specific randomly selected test sample of size 1,320 (consisting 659 hypertensive participants and 661 normotensive participants).

Export to PPT

ROC curves, precision and recall

ROC curves for each of the models are shown in supplementary figure 6, where a maximum AUC of 0.8508 was observed under the XG boost classifier and lowest AUC of 0.6880 was observed under SVM. The precision, recall, and F-1 score for the models are shown in the classification report of the models in figure 4.

Close

Fig. 4.

Classification reports of the fitted models in two-year risk prediction.

Export to PPT

Relative importance of the features

The relative importance of all the 19 statistically significant features was determined using RFC, (results visualised diacritically in Supplementary Fig. 7). It reveals that the systolic blood pressure showed the maximum relative importance followed by BMI, waist circumference, diastolic blood pressure, and age group, etc. On the other hand, smoking, religion, and diabetes showed the lowest contribution in predicting hypertension after two years.