Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait

Imanne El Maachi, Guillaume-Alexandre Bilodeau, Wassim Bouachir

Abstract

Diagnosing Parkinson's disease (PD) traditionally relies on subjective clinical evaluations of motor symptoms, such as gait abnormalities. This paper introduces a novel deep learning approach using 18 parallel 1D-Convolutional Neural Networks (1D-Convnets) to automate PD detection and severity prediction from raw Vertical Ground Reaction Force (VGRF) signals. By bypassing manual feature extraction, the proposed model achieves a state-of-the-art detection accuracy of 98.7% and a severity prediction accuracy of 85.3% on the Unified Parkinson’s Disease Rating Scale (UPDRS).

Introduction

Parkinson's disease (PD) significantly impacts motor function, often manifesting as distinct gait abnormalities including shortened steps, slower cycles, and flat foot strikes. While automated tools have been developed to assist clinicians, they typically depend on hand-crafted features—such as temporal statistics or frequency analysis—which may fail to capture the complex, non-linear characteristics of a Parkinsonian walk across diverse patient demographics. This research aims to replace these brittle features with a deep learning architecture capable of learning discriminative representations directly from raw sensor data.

Methodology

The core of the proposed system is a Deep Neural Network (DNN) designed to process raw 1D time-series data from 18 foot sensors measuring VGRF (8 sensors per foot plus 2 aggregate total force signals).

Proposed parallel 1D-Convnet architecture for PD detection and severity prediction.

Parallel Architecture and Late Fusion

Unlike traditional models that aggregate signals early, this architecture employs 18 parallel 1D-Convnets. Each branch independently processes a single sensor's signal through four convolutional layers, two max-pooling layers, and a fully connected (FC) layer. This design allows the network to learn specific spatial kernels for different regions of the foot. The deep features from all branches are then concatenated into a single vector (late fusion) and passed through final FC layers for classification.

Inference Strategy

Continuous walking data is segmented into 100-timestep windows with 50% overlap. Subject-level classification is determined via a majority vote across all segments belonging to a single individual, ensuring robust final predictions.

Experimental Setup

The model was evaluated using the public Physionet Gait Database, which contains 305 recorded walks from 93 PD patients and 73 healthy controls. Data was sampled at 100 Hz during natural-paced walks. The study addressed two primary tasks:

  1. Binary Classification: Distinguishing between PD patients and healthy controls.
  2. Severity Prediction: Categorizing patients into five discrete classes based on their continuous UPDRS scores (0-70).

Results

The proposed 1D-Convnet architecture demonstrated superior performance compared to existing methods:

  • Detection Accuracy: The model achieved 98.7% accuracy at the subject level (98.1% sensitivity, 100% specificity), significantly outperforming previous state-of-the-art methods such as MLPs with hand-crafted features (88.9%) and hybrid 2D-Convnet/RNN models (90.3%).
  • Severity Prediction: The model established a strong baseline for automated disease monitoring, achieving 85.3% accuracy across five UPDRS severity classes.
  • Ablation Study: Removing specific sensor pairs (e.g., L2/R2) led to significant performance drops, highlighting that certain pressure points are more critical for capturing Parkinsonian gait characteristics.

Discussion and Future Work

By leveraging raw data and parallel feature extraction, this model successfully captures the nuanced physiological signals of PD. However, limitations remain, particularly regarding dataset imbalance, where over-representation of certain classes affected severity prediction accuracy. Future research will focus on "opening the black box" of the DNN to identify the specific physiological features learned by the network, potentially offering new clinical insights. Additionally, the authors envision deploying this technology in wearable biometric sensors for continuous, non-invasive screening in elderly populations.

Summary updated

BibTeX
@article{maachi2020deep,
  author = {Imanne El Maachi and Guillaume-Alexandre Bilodeau and Wassim Bouachir},
  title = {{Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait}},
  journal = {Expert Systems with Applications},
  year = {2020},
  doi = {10.1016/j.eswa.2019.113075},
}
Figure