Interpretable Multiclass Parkinson’s Disease Classification from Single-Camera Video Using Kinematic and Nonlinear Dynamic Features

Technical Summary: Interpretable Multiclass Parkinson's Disease Classification from Single-Camera Video

Problem Statement
Parkinson's disease (PD) is a progressive neurological disorder characterized by motor symptoms such as gait impairment, bradykinesia, and rigidity. While clinical scales like the UPDRS are the standard for assessment, they are subjective and often insensitive to subtle gait alterations. Existing objective methods, such as motion capture and force plates, are accurate but expensive and impractical for routine use. Vision-based approaches offer a non-contact alternative, yet current literature faces three persistent gaps:

1. Most studies frame PD detection as a binary task (PD vs. Healthy), neglecting the clinically critical task of severity staging (e.g., distinguishing mild vs. moderate PD).
2. While nonlinear dynamic metrics (e.g., fractal dimension, entropy) have been validated as sensitive biomarkers, their integration into an interpretable, multiclass classification pipeline remains unexplored.
3. Publicly available PD gait datasets are typically small, imbalanced, and lack diversity, leading to a high risk of optimistic performance estimates if evaluation protocols are not rigorously designed to prevent data leakage.

Methodology
The authors propose a single-camera, markerless framework that extracts kinematic and nonlinear dynamic features from video-based pose estimation to classify gait into three categories: Healthy (NM), Mild PD (PD_ML), and Moderate PD (PD_MD).

• Data Acquisition & Preprocessing: The study utilizes a public dataset (Kour et al., 2020, 2022) containing 78 usable video samples (57 NM, 10 PD_ML, 11 PD_MD). Videos were recorded at 25 fps using a fixed RGB camera. Pose estimation was performed using MediaPipe Holistic to extract 33 anatomical landmarks, focusing on upper and lower extremity joints.
• Feature Extraction:
  • Kinematic Features: Six stride-level descriptors including stride time, swing time, stance time, stride length, and stride velocity.
  • Nonlinear Dynamic (NLD) Features: Four metrics computed from 2D joint coordinate time series: Hurst Exponent (H), Lempel–Ziv Complexity (LZC), Detrended Fluctuation Analysis (DFA), and Box-Counting Dimension (D).
  • This resulted in an initial 46-dimensional feature space (40 NLD + 6 Kinematic).
• Feature Selection: To address the small sample size and prevent overfitting, the authors employed a strict nested cross-validation pipeline. Feature selection was performed using Ordinary Least Squares (OLS) criteria (Mallows' Cp, AIC, BIC, Adjusted R²) on a temporarily recoded binary target (NM vs. PD) to identify optimal subsets. This yielded three fixed feature sets:
  • Case 1: 7 features (Cp/BIC)
  • Case 2: 9 features (AIC)
  • Case 3: 19 features (Adjusted R²)
• Model Training & Evaluation: Three machine learning models were evaluated: Random Forest (RF), Support Vector Machine (SVM) with RBF kernel, and a Feed-forward Neural Network (NN).
  • Validation Protocol: A repeated stratified 5-fold cross-validation (10 repeats) was used. Crucially, all preprocessing steps (scaling, SMOTE for class balancing) were performed only within the training folds to prevent information leakage.
  • Interpretability: Feature importance was analyzed using Permutation Importance (for RF) and SHAP values (for SVM and NN).
  • Ablation Study: Models were tested on Kinematic-only, NLD-only, and Combined feature sets.

Key Contributions

1. Multiclass Severity Staging: The work demonstrates the feasibility of distinguishing between healthy, mild, and moderate PD stages using pose-estimated gait features, moving beyond binary detection.
2. Integration of Nonlinear Dynamics: The study identifies that nonlinear dynamic descriptors, particularly those derived from distal joints (wrist, ankle, elbow), constitute the majority of selected features and serve as strong candidate biomarkers for gait variability in PD.
3. Rigorous Evaluation Framework: The authors implement a strictly nested cross-validation pipeline that confines feature selection and resampling within folds, ensuring unbiased performance estimation in a data-scarce environment.

Results

• Performance: On the held-out test set (approx. 16 samples), all models achieved high accuracy.
  • Case 2 (9 features): All three models (RF, SVM, NN) achieved 100% accuracy and perfect precision/recall across all classes on the test set.
  • Case 1 & 3: Models achieved 93.75% accuracy, with errors primarily occurring in distinguishing the minority PD classes (PD_ML or PD_MD).
  • Cross-Validation: The SVM with RBF kernel achieved the highest mean cross-validation accuracy (0.8366), while the Neural Network achieved similar accuracy (0.8338) with lower variance.
• Feature Importance:
  • Nonlinear dynamic features dominated the selected subsets (ranging from 71% to 84% of features across cases).
  • Key contributors included stride length (kinematic) and NLD metrics such as the Hurst exponent (wrist), Lempel-Ziv complexity (ankle), and DFA (knee/shoulder).
  • Ablation studies showed that while kinematic features drove accuracy in smaller sets, the combined feature set consistently provided the most stable and balanced performance (highest Macro-F1 and AUC).
• Computational Efficiency: SVM demonstrated the lowest training and inference times, making it suitable for real-time applications. Neural Networks incurred significantly higher computational costs but remained within acceptable ranges for offline clinical use.

Significance and Claims
The paper claims to provide a foundation for scalable, AI-assisted gait assessment in neuromotor healthcare. Its primary significance lies in demonstrating that interpretable, multiclass classification of Parkinsonian gait is achievable using standard single-camera video without specialized hardware.

The authors modestly note that while the results are promising, the small and imbalanced nature of the dataset limits the strength of generalization claims. They emphasize that the nonlinear metrics function as empirical descriptors of gait variability rather than mechanistic explanations of neurological processes. Consequently, the study advocates for future work involving larger, independent cohorts and the integration of demographic covariates (age, height) to validate these findings and refine the clinical utility of these digital biomarkers. The framework is presented as a versatile platform extensible to other gait pathologies, bridging the gap between complex gait dynamics and actionable clinical insights.