Аңдатпа
In this paper, we tackle non-reproducible and not fairly evaluated machine learning (ML) challenges in healthcare monitoring domains by using closed datasets and ML codes, unlike reproducible and fairly evaluable computer vision (CV) domains. To do this, we propose an open-source and pipelined selection-based optimization framework (PiSO) for preprocessed data and a deep learning (DL) model for arterial blood pressure (ABP) estimation, focusing on relatively underexplored Electrocardiogram (ECG)-only data (as well as Photoplethysmogram (PPG)-only and PPG+ECG data) and the recently released Medical Information Mart for Intensive Care (MIMIC)-IV dataset. Our end-to-end framework comprises a pipelined two selection phases: 1) First, for the preprocessed data selection (DS) phase, we evaluate open-sourced preprocessing (PP) techniques and toolkit(s) and select the best PP technique with the corresponding preprocessed datasets. 2) Second, in terms of DL model selection (MS) phase, we evaluate DL models of Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), and Transformers, using the preprocessed data without an exhaustive feature engineering process. By combining the pipelined two selection phases in our reproducible and fairly evaluable framework (PiSO) using the MIMIC-IV database, we achieve the best mean absolute error (MAE) results of 4.46mmHg(4.03mmHg) for systolic blood pressure (SBP) and 1.09mmHg(1.09mmHg) for diastolic blood pressure (DBP), both meeting Grade A standards according to the British Hypertension Society (BHS) standards, using ECG-only data (compared to PPG-only and PPG+ECG datasets) for intra-subject evaluation. For inter-subject evaluation, the best results are 13.65mmHgfor SBP and 4.17mmHgfor DBP, using ECG-only data, meeting Grade A only for DBP estimation.