POCA: a CPG signal analysis algorithm using peak-based feature extraction and machine learning

TL;DR

This paper introduces POCA, a new algorithm for analyzing CPG signals using peak-based features and machine learning to classify locomotor activity more accurately.

Contribution

The novel peak-based feature extraction framework improves CPG signal analysis and classification performance compared to traditional epoch-based methods.

Findings

POCA achieved an F1 score of 0.911 and accuracy of 0.957 using peak prominence as a key feature.

Incorporating additional peak features with an SVM improved performance to an F1 score of 0.923 and accuracy of 0.966.

POCA's results aligned closely with human-expert assessments of locomotor rhythms.

Abstract

Our understanding of the central pattern generator (CPG) for locomotion is primarily based on motor output analyses in isolated neonatal rodent preparations. Recent studies show that biomimetic neural modulation protocols, which mimic biological signals, outperform traditional methods in sustaining long-lasting fictive locomotor rhythms. However, fine-tuning such protocols requires extensive experimental trials, highlighting the urgent need for an automated CPG signal analysis tool. This study introduces the Peak-based Oscillation Classification Algorithm (POCA) for analyzing CPG signals using a novel peak-based feature extraction and machine learning. Although epoch-based feature extraction is widely applied in other biological oscillation analyses, they are suboptimal for CPG signals due to issue like challenging annotation and indirect feature representation. POCA addresses these limitations by extracting features directly from individual oscillation peaks, enabling more accurate and interpretable classification of locomotor versus non-locomotor activity. Using datasets from three independent stimulation protocols, a thresholding method using “peak prominence” feature achieved an F1 score of 0.911 and accuracy of 0.957, demonstrating the effectiveness of “peak prominence” as a key discriminative feature. A radial basis function kernel Support Vector Machine, incorporating additional peak features, further improved performance to an F1 score of 0.923 and accuracy of 0.966. The locomotor rhythm characterization results, based on oscillation detection, also aligned closely with human-expert assessments. The proposed POCA algorithm provides a robust, scalable tool for CPG signal analysis, facilitating large-scale evaluation of biomimetic protocols. The novel peak-based feature extraction framework also offers a versatile strategy for broader biological oscillation detection tasks.

Figures