Bioscience Methods 2024, Vol.15, No.3, 124-138 http://bioscipublisher.com/index.php/bm 130 electrocardiograms (ECGs), and demographic information, which are crucial for accurate diagnosis and prediction (Kwon et al., 2020; Khan et al., 2021). Figure 3 Block diagram of the blood pressure estimation process using ML techniques (Adopted from Visco et al., 2023) Image caption: The raw signals are prepared through normalization, the correction of baseline wandering due to respiration, and finally, signal filtration. Specifically, to construct a dataset for BP estimation models, it is necessary to accurately extract the features of the original waveform (and underlying demographic and statistical data) and select effective features, improving the generalization and reducing the risk of overfitting the algorithms. PPG: photoplethysmograph; ML: machine learning (Adopted from Visco et al., 2023) Preprocessing is equally important as it involves cleaning and transforming raw data into a format suitable for analysis. Techniques such as empirical mode decomposition (EMD) are used to preprocess PPG signals by decomposing them into their constituent components, which helps in extracting meaningful features (Khan et al., 2021). Additionally, normalization and centering of data are essential steps to ensure that the AI models perform optimally (Judge et al., 2023). Preprocessing also involves handling missing data, removing noise, and ensuring data privacy and security, which are critical for maintaining the integrity and reliability of the diagnostic system (Visco et al., 2023). 5.2 Feature extraction and selection Feature extraction and selection are pivotal in enhancing the performance of AI models. Feature extraction involves identifying and isolating relevant characteristics from the preprocessed data that can be used to train the AI model. For example, in the case of ECG data, features such as the S-wave, P-wave, and T-wave are extracted as they have significant effects on the diagnosis of conditions like pulmonary hypertension (Kwon et al., 2020; Kwon and Kim, 2020). Similarly, multi-domain features are extracted from PPG signals to categorize normal and hypertensive states (Khan et al., 2021). Feature selection, on the other hand, involves choosing the most relevant features from the extracted set to improve model accuracy and reduce computational complexity. Techniques such as the chi-squared statistical model and hybrid feature selection and reduction (HFSR) schemes are employed to eliminate irrelevant features and avoid issues like overfitting and underfitting (Ali et al., 2019; Khan et al., 2021). Advanced methods like Relief, Minimal Redundancy Maximal Relevance (mRMR), and Least Absolute Shrinkage and Selection Operator (LASSO) are also used to enhance the feature selection process (Li et al., 2020). 5.3 Model training and validation Model training and validation are critical steps in developing robust AI-based diagnostic systems. During the training phase, the AI model learns from the preprocessed and feature-selected data. Various machine learning
RkJQdWJsaXNoZXIy MjQ4ODY0NQ==