Machine learning revolutionizes tracking symptoms and predicting progression of Parkinson’s disease

Recent IEEE Neural Systems and Rehabilitation Engineering Paper In this study, researchers discuss the development of an automated system using machine learning (ML) algorithms that can quantify motor symptoms of Parkinson’s disease (PD) and predict disease progression.

study: Characterizing disease progression in Parkinson’s disease from videotaped finger tapping tests. Image credit: sruilk / Shutterstock.com

Parkinson’s Disease: Diagnosis and Treatment

To date, there is no cure for PD, therefore management of PD is primarily supportive and focuses on alleviating tremor, mood disorders, bradykinesia, and postural instability.

Physicians often use the Movement Disorder Society Unified PD Rating Scale (MDS-UPDRS) to measure disease progression as mild, moderate, or late PD and to evaluate patients’ response to treatment. Although the MDS-UPDRS part III scoring method is a reliable and sensitive approach, it has limitations in characterizing motor symptoms. In addition, the MDS-UPDRS III relies on subjective interpretation and has limited sensitivity to detect prodromal or early stages of PD.

Therefore, there is a need for better PD assessment systems that are sensitive to small changes in motor function. This approach would allow us to detect different stages of PD and develop effective treatment strategies to slow its progression.

Recently, researchers have investigated the potential of new digital approaches that use ML algorithms to determine motor markers of PD from MDS-UPDRS III video recordings. For example, the MDS-UPDRS III finger tapping test, which is used to assess limb bradykinesia, can be leveraged with a digital approach. In one study, ML algorithms assessed the severity of motor symptoms in finger tapping test videos, resulting in improved accuracy in diagnosing and predicting severity of PD.

Currently, digital methods for detecting PD assume that a common set of kinematic features exists across disease severity that varies consistently with disease severity, however this assumption may not be valid as motor symptoms do not change uniformly as PD progresses.

About the Research

The current study investigators hypothesized that by considering different kinematic features, they could more reliably detect PD and more accurately predict the severity of motor symptoms at different stages of the disease.

For this purpose, video data from 66 PD patients and 24 age-matched healthy controls were used. All PD diagnoses were confirmed by a movement disorder specialist using the UK PD Brain Bank diagnostic criteria. Participants with a history of brain tumour, stroke or implanted devices were excluded.

Data were collected from eligible participants at baseline and 1 year later. MDS-UPDRS III assessments, including both motor and cognitive assessments, were videotaped.

Prior to data collection, study participants were asked to discontinue taking antiparkinsonian medication overnight. ML algorithms were used to assess hand posture and identify video-based kinematic features associated with bradykinesia.

Data analysis compared a multiclass classification model, an ordinal binary classification model, and a newly developed hierarchical binary classification approach. The multiclass classification model uses consistent features for all severity levels, while the ordinal binary classification approach takes into account the ordinal nature of disease severity scores. The new hierarchical binary classification approach takes into account different behavioral features depending on disease severity.

Recording setup and environment. Subjects sit in front of a standard video camera and perform a finger tapping task. The task is videotaped and the video is stored for processing. Task performance is guided by an expert clinician who provides a clinical score.

Hand tracking results provided by the video processing pipeline. We calculate the angular distance between two vectors formed by joining the base of the hand and the tips of the index finger and thumb, localized by Google’s MediaPipe, in each video frame. The angular distance is tracked throughout the video to estimate the angular displacement signal. Kinematic features related to bradykinesia are then calculated from the peaks and valleys of the angular displacement signal (green and red dots).

research result

A total of 180 videos were analyzed, including 123 videos from PD patients and 44 videos from healthy controls. Based on the severity of motor symptoms, 42, 20, 62, and 56 videos were scored as 0, 1, 2, and 3, respectively.

Each participant provided two finger tapping test videos, one on each hand. Some patients showed increased movement variability and a gradual decrease in the sequence effect, which is the amplitude during repetitive tapping movements.

Although most of the video-based kinematic features differed significantly between groups, when analyzing the differences in severity scores, features that differed between the lower-scoring groups were not features that differed between the higher-scoring groups, supporting the research hypothesis that the kinematic features that determine disease severity change as the disease progresses.

Several non-traditional kinematic features were identified, including measures of amplitude decay, velocity of opening and closing movements, and variability in movement and timing that can be quantified from video.Compared to existing methods, the novel hierarchical binary classification approach predicted PD severity and differentiated between different severity levels with greater accuracy.

Conclusion

Automated severity prediction from video has the potential to revolutionize PD management, potentially facilitating the monitoring and quantification of motor symptom severity through video analysis alone.

The novel hierarchical binary classification approach used in this study proved to be efficient in determining the severity of PD. Therefore, this method may effectively improve PD management and evaluation of treatment efficacy. Rather than relying on a single multiclass model, it may be more efficient to use a multistage modeling approach or a combination of models that consider multiple features at different severity levels.