Background

Parkinson's Disease (PD) is a neurodegenerative disorder that affects millions worldwide, profoundly altering motor skills and overall quality of life. Despite advancements in medical science, early and accurate diagnosis of PD remains a challenge. This work explores how Artificial Intelligence (AI) can enhance clinical assessments of PD through gait analysis, contributing to improved diagnostics and personalized treatment strategies [4].

Data on gait movements offer a wealth of information about PD progression. Patients with PD exhibit distinct gait disturbances such as bradykinesia, rigidity, and postural instability. Monitoring these changes can provide insights into disease stages and response to therapies. My research leverages the LuxPARK dataset [1], which contains gait data collected using digital sensors and uses this data to monitor patients’ gait. For the data collection, patients do certain exercises, which are recorded by the sensor attached to the patient’s foot. Here, we deal with two subsets of the data separately: One with PD patients and one with healthy controls without PD.

Methods

The primary objective of my work was to determine whether machine learning (ML) models trained on longitudinal gait data outperform those relying on aggregated step features. By capturing temporal dynamics, the aim was to extract more nuanced patterns indicative of PD progression.

To achieve this, I evaluated various Machine Learning (ML) techniques, including classical models like Random Forest (RF) and Support Vector Machines (SVM), as well as advanced neural networks such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). I also integrated Amazon’s Chronos pre-trained transformer to push the boundaries of AI-driven gait analysis.

For this, a certain workflow was followed, which is showcased in Figure 1. It begins with data extraction and preprocessing, which combines clinical and gait data while addressing outliers and normalizing data. The data is then categorized into three output classes: gait subscore, UPDRS (Unified Parkinson’s Disease Rating Scale) III, and axial score for model training using both classical (e.g., Random Forest, XGBoost) and non-classical (e.g., CNN, LSTM) machine learning methods. The pretrained transformer model Chronos [3] was also applied on PD data and the control patient data. The final steps involve model evaluation and explainability using techniques like SHAP (SHapely Addictive exPlanations) to ensure interpretability. A statistical validation was done with the Wilcoxon rank-sum test [2] in order to compare model predictions across control and PD groups to validate the results.

Results and Conclusion

With Mean Absolute Error (MAE) serving as the evaluation metric, the findings revealed that a classical machine learning model (RF) outperformed others in most cases for PD data. In contrast, for the control patient data, the pre-trained transformer model Chronos achieved the best performance for the output class gait subscore but performed the worst on PD data. This is because if the dataset contains classes with very different sample sizes, Chronos may have difficulty learning features that distinguish between classes effectively, leading to poor performance. When the output classes are almost the same, this challenge is eliminated, allowing the model to focus solely on feature extraction for that one class.

For predicting UPDRS III across certain exercises, CNN emerged as the top performer on PD data, leveraging its strength in time-series analysis. Models like CNN and LSTM, tailored for temporal data, were critical for capturing intricate gait patterns.

Explainability is crucial in understanding the decision-making process of machine learning models, especially in clinical applications. Here, SHAP is used to quantify and visualize the impact of individual features on model predictions, enhancing interpretability [5]. SHAP plots highlight the relative importance of features in model predictions. RF provides sharper insights into individual gait features, while CNN captures complex relationships across input data, with key features like Max. Toe Clearance (The maximum elevation of the toe from the ground during the swing phase) shows a significant influence on predictions.

However, these results indicate that further improvements have to be made to the more reliable monitoring of PD predictions, taking a lot of factors into consideration. These include implementing the whole workflow on different gait datasets, exploring other pre-trained models, etc.

Citations

[1] Hannink J., et al., "Sensor-Based Gait Parameter Extraction With Deep Convolutional Neural Networks”. IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 1, pp. 85-93, Jan. 2017, doi: 10.1109/JBHI.2016.2636456.

[2] Goetz, C.G., et al. (2008), “Movement Disorder Society-sponsored revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS): Scale presentation and clinimetric testing results”. Mov. Disord., 23: 2129-2170. https://doi.org/10.1002/mds.22340

[3] Ansari,  A.F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., et al. "Chronos: Learning the language of time series" 2024. Preprint available at: https://arxiv.org/abs/2403.07815 Code available at https://github.com/amazon-science/chronos-forecasting

[4] Clarke, C.E. "Parkinson's disease." Bmj 335.7617 (2007): 441-445. https://doi.org/10.1136/bmj.39289.437454.AD

[5] Datacamp. “An Introduction to SHAP Values and Machine Learning Interpretability.” https://www.datacamp.com/tutorial/introduction-to-shap-values-machine-learning-interpretability