Deep learning for objective estimation of Parkinsonian tremor severity

TL;DR

A pixel-based deep learning model designed to analyse postural tremor in Parkinson's disease (PD) from video data is introduced, overcoming the limitations of traditional pose estimation techniques.

Abstract

Accurate assessment of Parkinsonian tremor is vital for monitoring disease progression and evaluating treatment efficacy. We introduce a pixel-based deep learning model designed to analyse postural tremor in Parkinson's disease (PD) from video data, overcoming the limitations of traditional pose estimation techniques. Trained on 2,742 assessments from five specialised movement disorder centres across two continents, the model demonstrated robust concordance with clinical evaluations. It effectively predicted treatment effects for levodopa and deep brain stimulation (DBS), detected lateral asymmetry of symptoms, and differentiated between different tremor severities. Feature space analysis revealed a non-linear, structured distribution of tremor severity, with low-severity scores occupying a larger portion of the feature space. The model also effectively identified outlier videos, suggesting its potential for adaptive learning and quality control in clinical settings. Our approach offers a scalable and objective method for tremor scoring, with potential integration into other MDS-UPDRS motor assessments, including bradykinesia and gait. The system's adaptability and performance underscore its promise for high-frequency, longitudinal monitoring of PD symptoms, complementing clinical expertise and enhancing decision-making in patient management. Future work will extend this pixel-based methodology to other cardinal symptoms of PD, aiming to develop a comprehensive, multi-symptom model for automated Parkinson's disease severity assessment.

Figures (6)

• Figure 1: (a) Marker-less pose estimation confidence of the detected hand key points, for different MDS-UPDRS items. One-sided t-test comparison determined that postural tremor of hands had a significantly lower key point confidence ($\text{p-value}<0.0001$) than finger tapping and hand movements items. A total of 60 assessments were randomly sampled for each item, (12 from each severity score). Confidence from all hand key points ($n=21$) was extracted for a sub sample of 60 frames from each of the sampled assessments, accounting for a total of $n=75600$ confidence sample points for each item. (b) Hand position for postural tremor assessment, as per MDS-UPDRS guideline UPDRS; occlusion of hand key points influences the confidence of the pose estimation model. (c) Hand key point visibility is less affected during other MDS-UPDRS items (finger tapping depicted), increasing the confidence of the predicted poses.
• Figure 2: Schematic of the inference pipeline. 1. Data pre-processing: Crop and resize of videos. 2. Inference model: Data is fed to a 3D CNN and then an LSTM module, finalizing with a fully-connected prediction layer. Refer to Table \ref{['tab:model_params']} model parameter details.
• Figure 3: (a) Performance of the 3D Conv-LSTM model across various binary classification tasks, measured by the area under the ROC curve (AUC). The gray dashed line indicates an AUC of 0.5, representing random behavior equivalent to flipping a fair coin. (b) Model performance in predicting the MDS-UPDRS-III item 3.15, measured using linearly weighted Cohen's Kappa and balanced accuracy on the combined test sets from a 5-fold cross-validation.
• Figure 4: Agreement between clinicians and the model to detect lateral asymmetry of tremor on $n=1340$ assessments with both left and right hand severity scoring. The model scores each hand independently, while clinicians have a access to both hands for a simultaneous comparison with one another. R: Right hand score, L: Left hand score
• Figure 5: Average improvement in MDS-UPDRS-III postural tremor score after treatment relative to the baseline, $n=27$ patients with existent baseline tremor ($\mathrm{score} \geq 1$). Error bars indicate the standard error of the mean. Refer to tables \ref{['tab:wilcoxon_between_treatments']},\ref{['tab:wilcoxon_between_raters']} for statistical analysis of the different treatments. L-Dopa: levodopa, DBS: deep brain stimulation