Model for the detection of elevated blood pressure based on machine learning and signal processing techniques

Abstract

Hypertension or high blood pressure is referred to as the increase of blood pressure in the body. It is considered a global public health issue whose late diagnosis could lead to heart attack, heart failure, coronary heart diseases, diabetes, or strokes. Nevertheless, its diagnosis is complicated since this health condition often does not produce notable symptoms that could alert from its presence in an early stage. Due to the above, hypertension is also known as the silent killer. Hypertension is generally diagnosed with a sphygmomanometer, the gold standard for blood pressure measurements. However, the clinical guidelines are not precise about how many consultations or ambulatory measurements are needed to perform the final diagnosis. Therefore, the clinical practice suggests the constant monitoring of blood pressure for several days and weeks. As a consequence, continuous monitoring of blood pressure is required. Still, cuff-based methods only take a snapshot of blood pressure and thus require several readings to obtain a more reliable value. In addition, they perform arterial compression, which can affect the patient's quality of life. Self-measurements could provide wrong measurements that, in the long-term, could generate an incorrect diagnosis.

On the other hand, the literature has developed alternative methods to keep track of blood pressure non-invasively during long periods. One of the approaches is to fit either regression or classification models employing machine learning techniques that can potentially monitor blood pressure using as input physiological signals such as electrocardiography and photoplethysmography for blood pressure estimation or clinical and socio-demographic information employing variables such as age, gender, body mass index, and heart rate for high blood pressure detection. In addition, a popular method is to use artificial neural networks, since they produced good performance due to their ability to map arbitrary relationships between a set of input and output data without relying on assumptions related to the distribution of the given data or domain knowledge about the data. Nonetheless, this method requires a large sample size to fit the data correctly. They lacked interpretability, which is a critical factor in applying machine learning in a medical context. Besides, some of the datasets available in the current literature only provide a small set of samples, which difficulties the training of deep learning techniques. Moreover, there is a lack of homogeneity between the data sets that are studied in the current state of the art that difficulties a comparison between the proposals and the reproduction of the studies.

This work presents a multi-modal data approach for elevated blood pressure detection. First, photoplethysmography waveforms were classified according to the hypertension stage using a transfer learning approach and pre-trained convolution neural networks. Then, this method was compared with the wavelet scattering transform for feature extraction and fitting the data with classical machine learning techniques such as logistic regression, support vector machine, linear discriminant analysis, decision trees, and k-nearest neighbor. Consequently, the socio-demographic and clinical variables like age, gender, body mass index, and heart rate were studied separately to analyze which of them can be helpful to classify subjects in different hypertension stages. Furthermore, other techniques for this feature selection process were considered, such as the statistical significance and relative feature importance through the Gini impurity index. Besides, classical machine learning techniques were chosen to provide a final model that can be more interpretable, avoiding black-box approaches based on deep learning. Finally, after analyzing each data modality separately, a final classifier was designed using photoplethysmography waveforms and clinical variables in which two fusion approaches were considered, early and late fusion.

Based on the results present in this analysis, it was observed that the wavelet scattering transform achieves a better performance than transfer learning in terms of accuracy, sensitivity, and F1-score for the prehypertension class by considering a support vector machine. Moreover, the use of clinical and socio-demographic variables to detect high blood pressure showed lower performance than the model trained using the wavelet scattering transform and photoplethysmography data with and without considering feature selection. However, the reduced model that considered the clinical data achieved better performance compared to fitting the models without feature selection by training a support vector machine. Finally, comparing the performance of early and late fusion to consider the effect of both clinical and socio-demographic data did not improve the performance of the classification task between normotension and elevated blood pressure subjects in terms of accuracy and F1-score. The two fusion approaches tested show that late fusion provides better performance than early fusion. Nevertheless, the reported results showed comparable performance with the current state of the art in terms of F1-score for the prehypertension class.