Blood pressure estimation from photoplethysmography by considering intra- and inter-subject variabilities: guidelines for a fair assessment

Cardiovascular diseases are the leading causes of death in humans, and blood pressure (BP) monitoring is an important procedure to prevent, diagnose, assess, and treat these diseases, thereby avoiding more severe conditions. The most reliable and common techniques of measuring BP use sphygmomanometers, but they are more suited to single measurements than continuous ones. A more promising approach is the use of photoplethysmography (PPG), a low-cost, opto-electronic technique that, in the form of wearable devices, allows the acquisition of a modulated light signal highly correlated to BP. In this context, many works have reported methods to estimate BP from PPG, achieving impressive results. In order to prevent overestimation of these results, we investigate---by considering two different data arrangements---how intra- and inter-subject variabilities in BP can affect the results of machine learning algorithms. We also compare the outcomes of these algorithms with the outcome of a regression using age, sex, weight and subject index number as attributes. Our general conclusion is that those algorithms might actually be learning to identify persons instead of predicting BP, showing that the split of data is a very important step.