SincPD: An Explainable Method based on Sinc Filters to Diagnose Parkinson's Disease Severity by Gait Cycle Analysis

TL;DR

Parkinson's disease diagnosis and severity assessment from gait signals is challenged by variability and the need for interpretability. The authors propose SincPD, an explainable CNN that uses eight SincConv1D layers to learn adaptive bandpass filters from raw vGRF data collected with in-shoe wearables. They prune the large filter bank by clustering the (center frequency, bandwidth) parameters and replacing them with cluster centroids, yielding a compact model with around 30 filters per sensor. A transfer-learning severity model leverages the frozen clustered features to classify Hoehn and Yahr levels, achieving 97.22% accuracy, while the binary PD diagnosis reaches 98.77% before pruning and 98.15% after pruning. The approach offers strong diagnostic performance with interpretable frequency-domain features and sensor-level insights, advancing practical use of wearables for PD assessment.

Abstract

In this paper, an explainable deep learning-based classifier based on adaptive sinc filters for Parkinson's Disease diagnosis (PD) along with determining its severity, based on analyzing the gait cycle (SincPD) is presented. Considering the effects of PD on the gait cycle of patients, the proposed method utilizes raw data in the form of vertical Ground Reaction Force (vGRF) measured by wearable sensors placed in soles of subjects' shoes. The proposed method consists of Sinc layers that model adaptive bandpass filters to extract important frequency-bands in gait cycle of patients along with healthy subjects. Therefore, by considering these frequencies, the reasons behind the classification a person as a patient or healthy can be explained. In this method, after applying some preprocessing processes, a large model equipped with many filters is first trained. Next, to prune the extra units and reach a more explainable and parsimonious structure, the extracted filters are clusters based on their cut-off frequencies using a centroid-based clustering approach. Afterward, the medoids of the extracted clusters are considered as the final filters. Therefore, only 15 bandpass filters for each sensor are derived to classify patients and healthy subjects. Finally, the most effective filters along with the sensors are determined by comparing the energy of each filter encountering patients and healthy subjects.

Figures (16)

• Figure 1: The impulse-response along with the frequency-response of a bandpass filter that allows selective frequency passage from $f_1$ to $f_2$.
• Figure 2: Overview of the preprocessing pipeline.
• Figure 3: Architecture of the SincPD. The model begins with an input layer followed by SincConv1D layers, specialized for handling 1D signals. These layers are normalized and activated using the Leaky ReLU function. Following the SincConv1D layers, traditional convolutional layers with batch normalization extract higher-level features. Finally, dense layers are utilized for classification.
• Figure 4: Illustration of the gait cycle phases. The gait cycle consists of the stance phase (heel strike, loading response, mid-stance, terminal stance, pre-swing includes $60\%$ of the cycle) and the swing phase (toe-off, mid-swing, terminal swing include $40\%$ of the cycle).
• Figure 5: Training and validation accuracy progression over 1000 epochs