Machine learning approaches for bio-signal analytics to estimate blood pressure

According to the World Health Organization (WHO), the disease that causes most deaths worldwide is cardiovascular disease (CVD). The primary causes of CVD are hypertension, atrial fibrillation, and high cholesterol. Hypertension can be predicted by blood pressure (BP) estimation. Several machine learning approaches that use biosignals for estimating cardiovascular health indicators have been studied in the literature and have shown promising results. There are still some challenges that need to be addressed to improve the performance of machine learning models for blood pressure estimation using biosensors. Firstly, it is imperative to consider the temporal behaviour of biosignals, which leads to non-linearity and distinct statistical properties. Furthermore, the learning model's performance may respond incorrectly to noise while acquiring the data from biosensors, including environmental confounders or patient movements. Understanding the relationship between the information from these temporal sensors and feature extraction methods is still necessary for the learning algorithms to perform robustly. Moreover, the existing literature has focused on specific traditional feature extraction techniques that may neglect some competent features, resulting in low performance.
To address the research gaps mentioned above, this study examines and provides analysis related to biosensors and machine learning techniques with a measure of their accuracy to analyse this technology better. To comprehend traditional feature engineering, we comprehensively investigate the association between traditional feature extraction methods and their machine learning algorithm's performance. Therefore, to mitigate the impact of artefacts caused by environmental confounders and analyse multiple input modalities with temporal behaviour, we propose a novel deep learning model to gain optimal feature selection from biosensor signals based on a temporal attention mechanism. Importantly, this model can learn traditional features of the signal data provided to examine the sequential analysis and identify the significance of the given features. Thus, with this proposed model, we can improve performance to reduce noise or confound input signal data by allocating the weight impact on the signal standard features.
This research thesis focuses on the following concrete contributions: providing an in-depth state-of-the-art survey that analyses the research taxonomy of machine learning algorithms for blood pressure estimation using PPG signals, current tools and techniques, existing feature extraction techniques, available datasets, and current research challenges. Furthermore, this research analyses and investigates the association of PPG signal features with early cardiovascular diseases using several machine learning algorithms. The results show minimal differences between features engineering methods' accuracy. Even though these differences are minimal, we cannot ignore them because of the sensitivity of this research domain. This research further extends the analysis and provides a benchmark study for traditional feature extraction techniques for blood pressure estimation using statistical and deep machine learning algorithms. The experimental results on two datasets show that time domain-based features achieve better accuracy than other feature extraction techniques. More importantly, deep learning algorithms performed better to achieve high performance with less error for blood pressure estimation. This thesis proposes a novel attention-based approach for continuous and cuffless blood pressure estimation from PPG signals. The investigation further combines features from different feature extraction techniques and selects the competitive features for the best analysis of the biosensor signal. The experimental results outperformed the state-of?the-art techniques using two datasets. To the best of our knowledge, this research study provides the first and most extensive survey in this research area and a benchmark study for several feature extraction techniques from PPG signals.