Photoplethysmography (PPG) is a widely recognized non-invasive optical technique for monitoring blood volume changes. Recently, PPG signals have gained prominence in healthcare applications, including the detection of cardiac arrhythmias. Cardiac arrhythmias represent a significant global health challenge, with particular focus on identifying atrial fibrillation (AF), the most prevalent type. Accurate detection of systolic peaks in PPG signals is crucial for arrhythmia detection and other applications such as heart rate estimation and heart rate variability analysis. Despite the high accuracy of existing beat detection methods in healthy subjects, the performance in the presence of cardiac arrhythmias is lower. This study employs a deep learning method to enhance the detection of systolic peaks in PPG signals, even in the presence of AF. The model was trained on a dataset comprising 2,477 10-second PPG segments with over 37,000 annotated PPG peaks, including data from patients with AF. Our model achieved an F1 score of 97.3% on the test dataset and F1 score of 94.8% on the test dataset when considering only patients with AF.
This paper focuses on non-invasive blood glucose determination using photoplethysmographic (PPG) signals, which is crucial for managing diabetes. Diabetes stands as one of the world’s major chronic diseases. Untreated diabetes frequently leads to fatalities. Current self-monitoring techniques for measuring diabetes require invasive procedures such as blood or bodily fluid sampling, which may be very uncomfortable. Hence, there is an opportunity for non-invasive blood glucose monitoring through smart devices capable of measuring PPG signals. The primary goal of this research was to propose methods for glycemic classification into two groups (low and high glycemia) and to predict specific glycemia values using machine learning techniques. Two datasets were created by measuring PPG signals from 16 individuals using two different smart devices – a smart wristband and a smartphone. Simultaneously, the reference blood glucose levels were invasively measured using a glucometer. The PPG signals were preprocessed, and 27 different features were extracted. With the use of feature selection, only 10 relevant features were chosen. Numerous machine learning models were developed. Random Forest (RF) and Support Vector Machine (SVM) with the radial basis function (RBF) kernel performed best in classifying PPG signals into two groups. These models achieved an accuracy of 76% (SVM) and 75% (RF) on the smart wristband test dataset. The functionality of the proposed models was then verified on the smartphone test dataset, where both models achieved similar accuracy: 74% (SVM) and 75% (RF). For predicting specific glycemia values, RF performed best. Mean Absolute Error (MAE) was 1.25 mmol/l on the smart wristband test dataset and 1.37 mmol/l on the smartphone test dataset.