Machine learning enhances tracking and prediction of Parkinson’s disease symptoms

A groundbreaking study published in the IEEE Transactions on Neural Systems and Rehabilitation Engineering has introduced a novel automated system employing machine learning (ML) techniques to assess and predict the progression of Parkinson’s disease (PD). This system offers a promising enhancement in the evaluation of motor symptoms, potentially transforming how this neurodegenerative disorder is managed.

Parkinson’s disease, a condition with no known cure, is primarily managed through symptomatic treatments focusing on alleviating tremors, mood disturbances, bradykinesia (slowness of movement), and postural instability. Traditionally, clinicians have relied on the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) to gauge disease progression through stages such as mild, moderate, or advanced, and to assess patient responses to treatments. While the MDS-UPDRS Part III scoring method is deemed sensitive and reliable, it is not without its shortcomings, including its subjective nature and limited capacity to detect early-stage or prodromal PD symptoms.

Recognising these limitations, researchers have explored a digital method that utilises ML algorithms to extract movement markers from MDS-UPDRS Part III video recordings, such as those from the finger-tapping test, which evaluates limb bradykinesia. This approach has proven to yield a higher diagnostic and prognostic accuracy regarding PD severity.

Existing digital detection models often assume a uniform set of kinematic features across all stages of PD, varying only with disease severity. However, this assumption is increasingly challenged by findings suggesting that motor symptoms evolve non-uniformly throughout the disease’s progression. The recent study hypothesised that a more nuanced consideration of varying kinematic features could enhance the detection and accurate classification of PD severity at different stages.

The study analysed video data from 66 PD patients and 24 age-matched healthy controls, excluding any individuals with a history of brain tumours, strokes, or neurostimulatory implants. All PD diagnoses were validated using the United Kingdom PD Brain Bank criteria. Data collection occurred at baseline and was repeated a year later, with participants abstaining from anti-Parkinsonian medications overnight prior to the recordings. These videos not only captured motor tasks but also cognitive evaluations under the MDS-UPDRS III.

Researchers compared three classification models: a multiclass classification model that uses consistent features across all severity levels; an ordinal binary classification model that accounts for the progressive nature of the disease; and a novel tiered binary classification approach that adjusts the kinematic features considered based on the severity of symptoms.

In total, 180 videos were analysed, including 123 from PD patients. These were categorised by motor symptom severity scores of zero to three. The videos demonstrated significant variations in kinematic features, such as movement amplitude and sequence effect, which is the diminishing amplitude during repeated movements. Notably, the study found that kinematic features associated with lower severity scores differed distinctly from those linked with higher scores, supporting the hypothesis that PD-related motor impairments evolve uniquely over time.

The study identified several innovative kinematic features, including amplitude decay and variations in the speed of opening and closing movements. These features, quantifiable through video analysis, showcased differences between various severity levels with higher precision than traditional methods. The tiered binary classification model, in particular, demonstrated superior effectiveness in predicting PD severity, suggesting that a multi-stage model or a combination of models that consider different features at various disease stages could significantly enhance PD management and treatment efficacy assessments.

This machine learning-based approach to video analysis holds great potential to revolutionise PD management by enabling more accurate monitoring and quantification of motor symptoms, thus paving the way for more tailored and effective treatment strategies.