Blood Pressure Estimation Through Photoplethysmography Using Deep Learning in Clinical Setting: Critical Survey and Solutions

Abstract

   Current solutions for blood pressure monitoring can be classified as invasive or non-invasive, both with drawbacks. Invasive blood pressure monitoring can lead to complications. Non-invasive blood pressure monitoring is intermittent which leads to missed episodes of hypertension and hypotension, also leading to complications. The state of the art for blood pressure monitoring through machine learning methods usually requires personalization, which is prohibitive in a clinical application. These proposed methods are generally not evaluated for clinical application. Datasets are usually split randomly, while a patient-wise split is required.

We first start by performing a survey of the literature to find candidate models for evaluation. These models are reproduced for evaluation alongside our proposed models. Popular input modalities from the literature are also reproduced with our proposed input modality. All combinations of models and input modalities are then evaluated against a patient-wise and random split. We perform a learning curve analysis to estimate how much data would be required to pass the AAMI standard.

The performance results establish that no model can provide calibration-free, non-invasive blood pressure monitoring using a single PPG site. The performance metrics show that our models and input modalities outperform the state of the art for random and patient-wise splits. Comparison against the models demonstrates that model complexity is insufficient to achieve better performance and that better preprocessing is a more efficient way to improve performance. The learning-curve analysis estimates that additional data could help achieve a model that passes the AAMI standard.