Blood pressure estimation from photoplethysmography by considering
intra- and inter-subject variabilities: guidelines for a fair assessment
Abstract
Cardiovascular diseases are the leading causes of death, and blood
pressure (BP) monitoring is essential for prevention, diagnosis,
assessment, and treatment. Photoplethysmography (PPG) is a low-cost
opto-electronic technique for BP measurement that allows the acquisition
of a modulated light signal highly correlated with BP. There are several
reports of methods to estimate BP from PPG with impressive results; in
this study, we demonstrate that the previous results are excessively
optimistic because of their train/test split configuration. To manage
this limitation, we considered intra- and inter-subject data
arrangements and demonstrated how they affect the results of
feature-based BP estimation algorithms (i.e., XGBoost, LightGBM, and
CatBoost) and signal-based algorithms (i.e., Residual U-Net, ResNet-18,
and ResNet-LSTM). Inter-subject configuration performance is inferior to
intra-subject configuration performance, regardless of the model. We
also showed that, using only demographic attributes (i.e., age, sex,
weight, and subject index number), a regression model achieved results
comparable to those obtained in an intra-subject scenario. Although
limited to a public clinical database, our findings suggest that
algorithms that use an intra-subject setting without a calibration
strategy may be learning to identify patients and not predict BP.