JMIR Publications

ABSTRACT

Background:

Photoplethysmography (PPG) is an optical technique for monitoring cardiovascular activity by measuring blood volume variations in tissue beds. PPG signals contain valuable information about heart rate variability, respiratory rate, among other vital signs. Many of those metrics are derived from the interbeat interval (IBI), which is the time-difference between consecutive systolic peaks. Hence, distinguishing whether a detected PPG peak is a valid systolic peak is an essential task. Kolmogorov-Arnold Networks (KANs) are a recently proposed class of deep neural networks capable of extracting complex patterns and features from time-series data, which have not yet been evaluated for this kind of problem.

Objective:

This paper aims to apply KANs to the classification problem of systolic peaks in PPG signals. The goal is to find the architecture that minimizes the cross-entropy in this application and compare it to benchmarks such as multilayer perceptrons (MLPs). The scenarios focus on model restrictions common in wearable technologies.

Methods:

Five datasets, consisting of four pulse-PPG and one finger-PPG, were used for model training and testing. Random hyperparameter search was applied to the models, establishing the range of 1-3 hidden layers and 2-50 neurons per layer, to consider wearable limitations. A comparison with classical classifiers such as Linear Discriminant, Gradient Boosting and K-Neighbors is performed. Finally, the KAN method is compared to state-of-art rule-based PPG peak detectors: Elgendi, HeartPy, Li, and Seongsil.

Results:

Across all datasets, KANs generally outperformed MLPs for number of hidden layers n=1, exhibiting higher mean accuracies (range: 0.9185-0.9806 vs. 0.9177-0.9799), F1 scores (0.9451-0.9899 vs. 0.9447-0.9896), and AUC (0.8013-0.9635 vs. 0.7984-0.9621). At n=2, minor discrepancies appeared mainly at the fourth decimal point. Meanwhile, MLPs regained advantage at n=3, showing enhanced mean accuracies surpassing KANs in four out of five datasets (accuracy range: 0.9188-0.9808 vs. 0.9090-0.9803). Three datasets witnessed KAN outperforming classic classifiers, while MLP excelled in four instances. Evaluating against rule-based methods confirmed improvements when adopting KAN, showing best performance in three datasets, while HeartPy showed top-tier results in the remaining two.

Conclusions:

The obtained results indicate the effectiveness of KANs in accurately identifying and classifying systolic peaks, with performance superior to MLPs in scenarios of few number of layers, a restriction often imposed to wearable solutions. It also shows the best performance compared to traditional classifiers and rule-based peak detectors for most tested devices. Clinical Trial: CAAE 07327119.8.0000.5599