DeepLabCut Applied to the assessment of spiral and sinusoidal patterns in individuals with Parkinson's disease

Abstract

Parkinson’s disease (PD) is a neurodegenerative disorder that affects both motor and non-motor functions. It is caused by dopamine deficiency in the brain, particularly in the substantia nigra. This shortage leads to symptoms such as bradykinesia, tremor, and rigidity, as well as non-motor symptoms like memory loss, anxiety, and depression. One of the major challenges in diagnosing PD is the lack of a precise, objective, and early-stage diagnostic method. Most current diagnostic techniques are subjective and expensive, and imaging methods such as magnetic resonance imaging only become informative at later disease stages. This study proposes a low-cost and replicable protocol supported by Machine Learning (ML) to aid in the diagnosis of PD. The method involves drawing tasks (spirals and sinusoidal waves) recorded via smartphone video. These recordings were processed using DeepLabCut (DLC) (Deep learning tool to train a specific neural network to extract positional data), which was then analyzed using PyCaret to classify individuals with different ML algorithms. Multiple tests were conducted, varying the number of folds and combining feature selection (FS) and tune function (TF) techniques. The protocol successfully simulated symptoms relevant to PD diagnosis in a simple, accessible format. DeepLabCut, when trained, achieved an average detection precision of 99%, struggling only with abrupt movements due to motion blur. The ML models performed well, with the best results occurring in 3, 7, and 9-fold trials. The 7-fold setup was found to be the most balanced. The top three classifiers — Extra Trees, Gradient Boosting Classifier, and K-Nearest Neighbors — each achieved accuracy rates above 90%. FS and tf generally improved model performance, although some trade-offs were observed. Extra Trees and Gradient Boosting Classifier experienced overfitting due to the small dataset, whereas K Nearest Neighbors, despite having slightly lower performance, remained more balanced. Overall, the study presents a promising, low-cost, and reproducible approach to assist PD diagnosis using simple tools and low-code ML pipelines.