Source-linked AI summary
Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait
Imanne El Maachi, Guillaume-Alexandre Bilodeau, Wassim Bouachir
TL;DR
Gait evaluation can be challenging because it is affected by intrinsic factors, motivating an objective analysis tool. The proposed system uses parallel 1D-ConvNets and a fully connected network, achieving 98.7% accuracy for Parkinson’s detection and 85.3% for UPDRS severity prediction.
Problem
Gait evaluation can be challenging because it is affected by factors such as age and health condition, motivating an objective tool to assist physicians.
Method
The model uses parallel 1D-ConvNets processing VGRF signals, then concatenates their outputs for a fully connected network producing the final decision.
Results
98.7% accuracy was achieved for Parkinson’s gait recognition, while UPDRS severity prediction achieved 85.3% accuracy.
Takeaways & Limitations
The work presents an algorithm for recognizing Parkinsonian gait and predicting UPDRS severity, including a first reported algorithm for UPDRS severity prediction.
Takeaways & Limitations
Gait evaluation is constrained by factors such as age and health condition.
Abstract
from arXiv · showhide
Diagnosing Parkinson's disease is a complex task that requires the evaluation of several motor and non-motor symptoms. During diagnosis, gait abnormalities are among the important symptoms that physicians should consider. However, gait evaluation is challenging and relies on the expertise and subjectivity of clinicians. In this context, the use of an intelligent gait analysis algorithm may assist physicians in order to facilitate the diagnosis process. This paper proposes a novel intelligent Parkinson detection system based on deep learning techniques to analyze gait information. We used 1D convolutional neural network (1D-Convnet) to build a Deep Neural Network (DNN) classifier. The proposed model processes 18 1D-signals coming from foot sensors measuring the vertical ground reaction force (VGRF). The first part of the network consists of 18 parallel 1D-Convnet corresponding to system inputs. The second part is a fully connected network that connects the concatenated outputs of the 1D-Convnets to obtain a final classification. We tested our algorithm in Parkinson's detection and in the prediction of the severity of the disease with the Unified Parkinson's Disease Rating Scale (UPDRS). Our experiments demonstrate the high efficiency of the proposed method in the detection of Parkinson disease based on gait data. The proposed algorithm achieved an accuracy of 98.7 %. To our knowledge, this is the state-of-the-start performance in Parkinson's gait recognition. Furthermore, we achieved an accuracy of 85.3 % in Parkinson's severity prediction. To the best of our knowledge, this is the first algorithm to perform a severity prediction based on the UPDRS. Our results show that the model is able to learn intrinsic characteristics from gait data and to generalize to unseen subjects, which could be helpful in a clinical diagnosis.
1. Introduction
Parkinson’s gait is clinically relevant but difficult to evaluate objectively because gait varies across individuals and clinician assessment is challenging. The paper proposes a deep-learning gait classifier that detects Parkinson’s symptoms and predicts UPDRS-based severity without explicit manual feature extraction.
- Motivation: Gait abnormalities can appear early in Parkinson’s disease and are important for diagnosis, but clinicians face challenges evaluating them objectively.Gait assessment is affected by factors including age and health condition.
- Contribution: The proposed system uses deep learning to detect Parkinson’s disease symptoms and predict disease severity from gait data using the UPDRS.The approach is designed as an intelligent tool for clinical decision support.
- Research gap: Existing temporal and frequency-based feature extraction methods are limited because gait differs according to age, health, and other intrinsic factors.The paper therefore targets automated learning from gait signals rather than relying on hand-crafted preprocessing.
- Approach: The DNN processes 18 VGRF signals with parallel 1D-Convnets, concatenates their learned features, and uses a fully connected network for final classification.Each signal is processed independently before the learned representations are combined.
- Results: 98.7% accuracy was achieved for Parkinson’s detection, while 85.3% accuracy was achieved for UPDRS severity prediction.The paper reports the detection result as state-of-the-art and presents the severity-prediction algorithm as novel.
2. Related work and background
Prior gait-classification studies extracted temporal, frequency, or spatiotemporal features and applied conventional classifiers, while later work also combined convolutional and recurrent networks. The paper introduces background on deep neural networks and their convolutional, pooling, and fully connected layers.
- Related work: Earlier approaches combined engineered features with classifiers including Random Forest, SVM, MLP, KNN, and HMM-based representations.Feature selection or transformation was often applied before final classification.
- Related work: Previous gait studies used temporal, frequency, or spatiotemporal features to distinguish Parkinsonian gait from control gait.Reported features include statistics such as entropy, skewness, kurtosis, mean frequency, and frequency variance.
- Related work: One prior deep-learning system used parallel 2D-Convnet and recurrent neural networks to model spatial and temporal force distributions.Its final classification was based on averaging the network outputs.
- Background concepts: Deep neural networks learn progressively higher-level abstractions through hidden layers and optimize weights to reduce loss during training.The background describes nonlinear neuron outputs and stochastic optimization as components of DNN learning.
- Background concepts: The model uses convolutional layers for filtering, max-pooling for dimensionality reduction, and fully connected layers to connect learned features to the final output.Dropout is described as a regularization technique that randomly shuts off neurons during training.
3. Methods
The method divides foot-sensor VGRF recordings into labeled segments and processes the 18 signals through parallel 1D-Convnets before combining their learned features for classification. Separate output designs support Parkinson’s detection and five-class severity prediction.
- Input data: Each recorded walk contains 18 VGRF time series from foot sensors, including eight sensors under each foot and two per-foot summed signals.The signals measure vertical ground reaction force in Newtons over time.
- Input data: The DNN receives labeled VGRF segments as samples and classifies walks into Parkinson and control categories.Walk-level classification is determined by the majority classification across segments.
- Feature extraction: Each of the 18 parallel 1D-Convnets processes one VGRF signal independently through convolutional layers and a fully connected layer.The parallelization preserves signal-specific deep features from sensors recording different locations.
- Feature integration: The outputs of the 18 subnetworks are concatenated into one deep feature vector and passed through two fully connected layers to the output layer.The architecture combines learned features after multiple convolution steps rather than merging raw sensor data at the input.
- Output tasks: For Parkinson’s detection, the output layer predicts classification probability with one neuron; severity prediction uses five neurons for five classes.The architecture is intended to learn relationships between extracted spatial features and the final decision.
4. Experiments
The experiments evaluate Parkinson’s detection and severity prediction using gait recordings from foot-based VGRF sensors. They use subject-level cross-validation, segment-level training with subject-preserving splits, and compare the proposed network with prior approaches.
- Dataset: Each walk provides 18 VGRF time-series signals, including eight sensors per foot and two total under-foot VGRFs.The dataset also includes double-task walking recordings.
- Tasks: Severity prediction uses UPDRS scores segmented into five levels, with the first class defined as UPDRS < 5.The database reports total UPDRS scores from 0 to 70.
- Training and evaluation: 10-fold cross-validation is performed with subject-level folds while preserving the 70% Parkinson and 30% control dataset balance.Evaluation metrics are reported over the combined validation predictions from 300 walks.
- Models: The proposed architecture uses 18 parallel 1D-Convnets followed by a fully connected network, with five softmax outputs for severity prediction.For severity prediction, the architecture and training hyperparameters remain unchanged except for the final layer.
5. Results & Discussion
The 1D-Convnet achieved high accuracy for Parkinson’s detection and severity prediction, while its separate processing of input signals supported feature learning and adaptation across gait studies. Ablation results identified influential VGRF inputs, and severity prediction remained affected by class imbalance and confusion with class 1.
- Parkinson’s detection: 98.7% subject-level accuracy was achieved for Parkinson’s detection, exceeding the reported segment-level accuracy of 98.3%.Segment-level specificity was 99.2% and sensitivity was 97.8%.
- Model behavior and adaptability: The model processes multiple input signals independently, enabling signal-specific feature extraction and adaptation to different gait clinical studies.The authors attribute this flexibility to the independence between input signals and use it for VGRF ablation experiments.
- Evaluation considerations: A discrepancy was observed between the reproduced Zhao et al. result and the accuracy reported in their paper, potentially due to different evaluation procedures.The reported paper accuracy was 98.6%, while the authors attribute the discrepancy to cross-validation versus a single test fold.
- Model behavior and adaptability: Deep learning was considered better suited than classical machine learning for nonlinear, noisy gait signals and for learning discriminative features from data.The proposed model learns spatial kernels for distinguishing Parkinsonian and control gait rather than relying on hand-crafted features.
- VGRF signal selection: The total VGRF was the most important input feature, whereas removing L3 and R3 did not affect performance and suggested limited relevance for classification.Removing L2 and R2 or L4 and R4 had the largest performance impact in the segment-level ablation results.
- Parkinson’s severity prediction: 85.3% global accuracy was obtained for Parkinson’s severity prediction despite an unbalanced dataset.Most subjects in each class were correctly classified, but the predominant errors involved confusion with class 1, associated with the amount of available data.
6. Conclusions
The paper presents deep-learning gait analysis for Parkinson’s gait recognition and severity prediction, addressing the difficulty of diagnosis and gait-based assessment. The proposed approach achieves high reported accuracy and may support clinical screening and future monitoring applications.
- The algorithm recognizes Parkinsonian gait and predicts disease severity from gait data using deep-learning techniques.The approach avoids hand-crafted feature extraction.
- 98.7% accuracy was achieved for Parkinson’s gait recognition, reported as state-of-the-art performance.
- 85.3% accuracy was achieved for predicting UPDRS severity, described as the first such algorithm to perform this prediction.
- The method is proposed as a practical tool for screening populations to detect potential Parkinson patients in a clinical context.
- The authors propose longer-term gait monitoring and analysis during daily activities, including for elderly people, using biometric sensors.
- Future work includes examining learned DNN representations to deepen understanding of parkinsonian gait characteristics.