A machine learning model using photoplethysmography signals achieved F1 scores of up to 100% for classifying normal versus prehypertension and 96.0% for detecting type 2 diabetes mellitus.
Cross-Sectional (n=219)
Can machine learning models using photoplethysmography (PPG) signals accurately stratify hypertension risk and detect type-2 diabetes mellitus?
Photoplethysmography (PPG) signals combined with deep learning models can accurately stratify hypertension risk and detect type-2 diabetes mellitus, offering a potential non-invasive screening tool.
Abstract Early diagnosis of prehypertensive patients is crucial in managing and preventing subsequent complications. The major challenge is that there are no alarming symptoms for the prehypertensive patients resulting in delayed diagnosis. Further, patients with hypertension have an increased risk of being diagnosed with type-2 diabetes. The existing systems are not suitable for large-scale screening. Additionally, they lack diagnostic accuracy, which is essential for early risk assessment of hypertension. This article aims to develop a diagnostic expert system for hypertension risk stratification and diabetes mellitus type 2 (DM-II) detection using photoplethysmography (PPG) signals. A total of 156 time-domain features are extracted from the PPG signal and its derivative in terms of time-span, amplitude, area, power and their ratios. ReliefF and minimum redundancy maximum relevance (mRMR) feature selection algorithms are employed to select 20 top optimal features with a correlation to systolic blood pressure (SBP) and Diastolic blood pressure (DBP). Several classification models optimized using Bayesian optimization with 10-fold cross-validation are adopted for comparison. The highest F1 scores for the Normal (NT) versus prehypertension (PHT), NT versus hypertension type 1 (HT-I) and NT versus hypertension type 2 (HT-II) are found to be 100%, 73.9%, 80.7% for SBP and 100%, 72.8%, 81.8% respectively for DBP. The F1 scores achieved by Bi-directional long short-term memory for NT vs. PHT, NT vs. HT-I, and NT vs. HT-II are 95.1%, 97.2% and 100%, respectively. Furthermore, the classification accuracy for NT vs. DM-II achieved an F1 score of 96.0%. Our results indicate that PPG can be successfully used for risk stratification of hypertension and detection of DM-II. Future work is required to prove the efficacy of the proposed technique on a larger dataset. Multi-modal or combination of clinical data with PPG for classification is also considered in future scope.
Khan et al. (Thu,) conducted a cross-sectional in Hypertension and Type 2 Diabetes Mellitus (n=219). Photoplethysmography (PPG) machine learning classification vs. Normal (normotensive/non-diabetic) subjects was evaluated on F1 score for classification of Normal vs Prehypertension and Normal vs Diabetes Mellitus Type 2. A machine learning model using photoplethysmography signals achieved F1 scores of up to 100% for classifying normal versus prehypertension and 96.0% for detecting type 2 diabetes mellitus.