Artificial intelligence-based video monitoring of movement disorders in the elderly: a review on current and future landscapes

Abstract

Due to global ageing, the burden of chronic movement and neurological disorders (Parkinson’s disease and essential tremor) is rapidly increasing. Current diagnosis and monitoring of these disorders rely largely on face-to-face assessments utilising clinical rating scales, which are semi-subjective and time-consuming. To address these challenges, the utilisation of artificial intelligence (AI) has emerged. This review explores the advantages and challenges associated with using AI-driven video monitoring to care for elderly patients with movement disorders. The AI-based video monitoring systems offer improved efficiency and objectivity in remote patient monitoring, enabling real-time analysis of data, more uniform outcomes and augmented support for clinical trials. However, challenges, such as video quality, privacy compliance and noisy training labels, during development need to be addressed. Ultimately, the advancement of video monitoring for movement disorders is expected to evolve towards discreet, home-based evaluations during routine daily activities. This progression must incorporate data security, ethical considerations and adherence to regulatory standards.

INTRODUCTION

The rapid growth of the elderly population worldwide presents an unprecedented challenge to healthcare systems. As people age, they are increasingly at risk for various health issues, including chronic neurological disorders that can severely impact their mobility and quality of life.[1] Notably, chronic movement disorders such as Parkinson’s disease (PD) and essential tremor (ET) impose significant global social and economic burden because their prevalence rises dramatically in the geriatric population. Parkinson’s disease is the most common neurodegenerative movement disorder, distinguished by motor symptoms encompassing resting tremor, bradykinesia (slowness of movement), rigidity and postural instability.[2] The prevalence of PD increases by up to 1% for individuals aged over 65 years.[3] Essential tremor is the most prevalent movement disorder. It is characterised by bilateral action and postural tremors of the upper extremities. Its age of onset exhibits a bimodal distribution, with one peak around puberty and another in the senior years.[4]

Although these debilitating conditions often necessitate accurate diagnosis and continuous monitoring for effective management, the clinical diagnosis and assessments remain surprisingly archaic.[5] While the clinical diagnostic process of most medical disorders is established by objective laboratory or imaging tests, the clinical diagnosis of both PD and ET hinges solely on the medical history and neurological examination.[6] The definitive diagnosis of PD necessitates postmortem identification of Lewy bodies in the substantia nigra, which is hardly practical to guide management premortem. As a result, premortem diagnostic accuracy, even by experienced neurologists, hovers around 80%.[7,8]

The gold standard for assessing the severity of movement disorders is the clinical rating scales, consisting of numerous items designed to assess key disease symptoms. The Unified Parkinson’s Disease Rating Scale (UPDRS) part III and the Tremor Research Group Essential Tremor Rating Scale (TETRAS) are the most commonly used scales for PD and ET, respectively.[9,10] These scales assign an ordinal score to each item, ranging from none (0) to severe (4). Although the Movement Disorder Society’s (MDS) revision (MDS-UPDRS) has mitigated the instruction ambiguities of the original UPDRS,[11] the inherent limitations of the manually administered scale persist. The scale remains coarse-grained; it is time demanding, costly, susceptible to significant inter- and intra-rater variability, and accessible only during in-person patient–clinician interactions. The limited accessibility issue has gained prominence, particularly with the evolving need for telemedicine during the coronavirus disease 2019 (COVID-19) pandemic.[12,13] Moreover, a substantial portion of PD patients experience fluctuating motor symptoms aligned with their daily medication dosing schedules, rendering single clinic visits mere snapshots of an ever-changing condition, and this is hardly reflective of disease severity as a whole.[14]

To address such limitations, artificial intelligence (AI) has emerged as a revolutionary approach to facilitate a more objective assessment of disease status.[15,16] Thus far, wearable sensor-based technologies have played a leading role in this endeavour.[17,18,19] Recently, the guidelines from the United Kingdom National Institute for Health and Care Excellence have officially recommended several wearable sensors to collect real-world evidence for their impact on the health care of PD patients.[20] This recommendation underscores the growing recognition of such technologies by regulatory bodies and relevant stakeholders.

Nevertheless, wearable technology has its limitations, notably in terms of compliance, which has led to the exploration of alternative sources of data. Video-based assessment utilising computer vision technologies has emerged as a promising alternative, highlighted by its non-intrusiveness. In this context, we aimed to report a most up-to-date and comprehensive review of the current landscape of video-based monitoring of movement disorders, along with its potentials, challenges and future prospects.

LITERATURE REVIEW

We summarise the literature on video-based monitoring of movement disorders. Considering the wide heterogeneity and ambiguity in movement disorder phenomena and its classification, we narrowed the search to the upper limb bradykinesia of PD, the most representative disease symptom, and ET. We searched the EMBASE and PubMed databases with the following search terms in the title or abstract (‘artificial intelligence’ OR ‘machine learning’ OR ‘automatic’ OR ‘automated’) AND (‘video’ OR ‘vision’) AND appropriate search terms for the target disease and symptom, that is, (‘parkinson’ OR ‘parkinsonian’ OR ‘parkinsonism’) AND (‘finger tapping’ OR ‘hand movement’ OR ‘hand’ OR ‘finger’ OR ‘bradykinesia’ OR ‘motor’) for PD and (‘essential tremor’ OR ‘tremor’) for ET. The search was conducted on 12 August 2023. Publications from database inception within 5 years before the search date were considered for inclusion.

In the literature review, we included studies that utilised marker-less video-based assessments processed through AI/machine learning techniques for the two diseases. After initial search and removal of duplicates, the abstract of each study was screened for eligibility. Studies that are not original articles (e.g. review articles, conference abstracts), animal studies, articles on wearable sensors or marker-based video systems, articles that do not cover our target diseases and symptoms, and non-English articles were excluded. The flow diagram of article selection is depicted in Figure 1. Ten articles were finally selected for review [Table 1].[21,22,23,24,25,26,27,28,29,30] The selected studies report the accuracy and agreement with the clinician rating of the predicted UPDRS score, binary classification between PD and healthy subjects, or tremor detection.[29]

Figure 1.

Flow chart shows the literature search flow. ET: essential tremor, PD: Parkinson’s disease

Table 1.

Summary of studies that used marker-less consumer-level videos with artificial intelligence models for PD assessments.

Study	Hardware	Experiment settings	Pipeline summary	Extracted kinetic features	Reported results
Sarapata et al., 2023[21] Morinan et al., 2023[22]	Mobile device (KELVIN™ by Machine Medicine Technologies)	949 MDS-UPDRS ratings from PD patients, which include all five bradykinesia items from five centres	2D pose estimation → Action recognition → Region of interest extraction → Signal extraction → Kinetic features extraction → RF classifier	11 features each grouped into speed, amplitude, hesitations, decrementing signal categories	High agreement with the clinician rating for composite bradykinesia score (ICC=0.74)
Guo et al., 2022[23]	RGB-D camera	112 finger tapping videos from 48 PD patients	3D hand pose estimation → Kinetic feature extraction → SVM classifier	Nine features (variance, mean_abs_change, autocorrelation, fft_coefficient, linear trend of distance and variance, abs_energe, quantile, linear_trend of velocity)	81.2% classification accuracy
Guo et al., 2022[24]	Consumer-level cameras	753 hand movement videos from 174 PD patients	Pose estimation → Tree-structure-guided graph convolutional network	NA	73.71% accuracy
Li et al., 2021[25]	Clinical video data	744 finger tapping videos from 157 PD patients	Pose estimation → Three-stream (pose, motion, geometric) feature design → Markov chain fusion model	NA	73.5% accuracy
Williams et al., 2020[26]	Smartphone video	133 finger tapping videos from 39 PD patients and 30 HS	Pose estimation → Kinetic feature extraction → Correlation analysis with MBRS and MDS-UPDRS	Speed, amplitude variability, rhythm regularity	Spearman coefficients −0.74 speed, 0.66 amplitude, −0.65 rhythm for MBRS; −0.56 speed, 0.61 amplitude, −0.50 rhythm for MDS-UPDRS; −0.69 combined for MDS-UPDRS. All P <.001
Liu et al., 2019[27]	Clinical video data	360 videos of finger tapping, hand movement, pronation–supination from 60 PD patients	Pose estimation → Kinetic feature extraction → RF, SVM, K-nearest neighbour	Amplitude, velocity and their variabilities	Overall, 89.7% accuracy
Park et al., 2021[28]	Clinical video data	110 videos of finger tapping and 110 videos of resting tremor from 55 PD patients	Pose estimation → Kinetic feature extraction → SVM classifier	Amplitude, velocity, decremental responses for finger tapping, tremor amplitude for resting tremor	Good to excellent reliability with clinician rating (ICC=0.793, 0.927 for finger tapping and resting tremor, respectively)
Monje et al., 2021[29]	Computer webcam	Finger tapping, hand movement, pronation/supination videos of 22 PD patients and 20 healthy subjects	Single-Shot MultiBox Detector for hands → Pose estimation → Kinetic feature extraction → Logistic regression, Gaussian naïve-Bayse and RF models	Mean amplitude, standard deviation of the peaks, speed, decremental responses	Ranging from 31% to 73% sensitivity for PD versus HS binary classification
Wang et al., 2021[30]	RGB camera	55 participants with tremor	Hand pose estimation → Kinetic feature extraction → SVM, CNN-LSTM classifier	Frequency of motion direction changes, changes in distance of hand movement	81% accuracy for tremor detection

CNN-LSTM: convolutional neural network-long short-term memory, ICC: interclass correlation, MBRS: modified bradykinesia rating scale, MDS-UPDRS: Movement Disorder Society sponsored revision of Unified Parkinson’s Disease Rating Scale, NA: not applicable, PD: Parkinson’s disease, RF: random forest, RGB-D: red, green, blue and depth, SVM: support vector machine

By applying these inclusion and exclusion criteria, we focused on describing the up-to-date advances in the field by limiting the search to articles that utilised videos from commercially available consumer-level cameras, webcam or mobile devices. Traditionally, marker-based optical motion capture systems like the VICON system are considered to be the gold standard for modern clinical gait analysis, including that for PD.[31] Marker-based optical motion capture requires infrared markers to be attached to the key joints of the body and corresponding detection devices. Such a system is costly and applicable to only limited clinical laboratory facilities and resources; thus, it is not universally applicable. Although recently, there are options for home-based camera systems with wearable skin-based markers,[32,33] similar limitations remain: the necessity to equip a specific device and learn how to operate the system, which can be challenging and cumbersome, particularly for geriatric populations with movement disorders, who frequently have associative cognitive decline.

Although the technical details of the studies may vary, it is notable that most of them adopted the following three steps for their analyses: pose estimation, kinetic feature extraction and classification [Figure 2]. In some areas of AI in medical imaging, the so-called end-to-end models, which use an ‘image in, diagnosis out’ scheme, are favoured due to their simplicity, high performance, capacity for handling large datasets and privacy compliance.[34] An example of this scheme is one that takes chest radiographs as input and provides a diagnostic label of the thoracic disease as output empowered by state-of-the-art deep learning techniques. These models are often considered ‘black-box’ models, meaning they operate with complex processing steps hidden between the input and output layers, making it nearly impossible to determine the specific rationale on which the model had based its diagnostic decision. However, in terms of PD and movement disorders, particularly for evaluating severity assessment with clinical scales like UPDRS and TETRAS, such an end-to-end or black-box model as ‘video in, score out’ scheme may not be optimal. This is primarily due to the complex evaluation requirements for movement disorders, which necessitate the observation and scoring of specific features as outlined in the instructions, established by decades of empirical experience from domain experts. For Parkinsonian gait assessment, these features include stride amplitude, stride speed, height of foot lift, heel strike during walking, turning and arm swing. For tremor, they include amplitude, frequency, duration of tremor and the positions that aggravate the symptom. For bradykinesia, they include amplitude, velocity, decrement of amplitude and velocity during repeated movements, and halts or hesitations during the movement.

Figure 2.

The three-step scheme of pipelines for the video-based movement disorder monitoring system.

In light of the uniqueness of movement disorder assessment that places emphasis on ‘explainability’, most of our selected studies incorporated a middle stage of extracting the kinetic features that are representative of the clinical characteristics of bradykinesia or tremor, as depicted in Figure 2 [Table 1]. For example, Park et al.[28] and Monje et al.[29] each computed the kinetic features for amplitude, velocity and decrement of amplitude and velocity during finger tapping, and fed them into their subsequent classification models. It is notable that out of the four key elements of bradykinesia, studies have started to include the other well-known clinical features of halts or hesitations of upper limb movement only recently.[21,22] This may be because so far, there has been no clear mathematical description or definition by MDS to describe the paroxysmal freezing phenomena of the upper limb. However, such recent trend reflects that researchers are now placing stronger emphasis on the comprehensiveness of their models to better align conceptually with UPDRS guidelines.

To extract the data-driven kinetic features from the input videos, studies first require a preceding stage of 2D- or 3D-human pose estimation and subsequent processing of the time series data of the estimated target joint coordinates. Human pose estimation refers to a deep learning computer vision process that captures a set of coordinates for key joints from an image that can describe a person’s pose, resulting in a skeleton-like structure when graphically formatted.[35] An exemplary image is presented in the first step of Figure 2. Studies have commonly performed this stage with existing marker-less open-source 2D- or 3D-human pose estimation libraries, such as OpenPose and MediaPipe, that are frequently utilised in the medical context.[36,37] Among our selected studies, about half opted for OpenPose,[22,24,25,28,29] while Wang et al.[30] used MediaPipe and some others have used more customised human pose estimation algorithms.[23,26] Using existing libraries for human pose estimation offers several benefits, as they are often well optimised to run efficiently on a variety of hardware, have undergone extensive training on large datasets and provide fast inference times enabling real-time analysis. However, these libraries are typically developed and trained on healthy individuals; therefore, when applying these tools to specific elderly patient groups, accuracy and reliability issues may arise. Fine-tuning the model with additional disease-specific data or augmenting the dataset with synthetic or additional real-world data may help to make the models more robust. Finally, after the human pose estimation and subsequent kinetic feature extractions, the extracted features are fed into machine learning classification models to predict the clinical score or disease status [final step depicted in Figure 2].

In summary, this section reviews the current landscape of the video-based monitoring of motor symptoms in movement disorders. Most of the reviewed studies follow a pipeline where they acquire videos using consumer-level cameras, then apply human pose estimation to dynamically process the positional information of joints of interest, extract relevant kinetic features and employ them to a machine learning algorithm to assess the disease status. Recent studies have shown a growing trend towards not only emphasising the model’s performance, but also prioritising its explainability, given the unique characteristics of movement disorders.

POTENTIAL BENEFITS OF VIDEO-BASED MONITORING FOR ELDERLY PATIENTS

The AI technologies for monitoring of movement disorders using not only videos but also wearable sensors and other technologies hold the potential to overcome the limitations of traditional scales such as UPDRS and TETRAS measurements. While researchers currently employ different signals and methods in these early stages, their ultimate goal converges on the same path: to achieve objective remote, automated and continuous monitoring.[16]

Patients can collect real-time data in their own environments without the need for frequent hospital visits or in-person interaction with medical staff, enabling significantly more efficient data collection. The AI models objectively analyse the data, eliminating the possibility of inter-or intra-rater reliability, resulting in a more consistent outcome. Clinical trials to discover novel treatment modalities can greatly benefit from the enhanced objectivity and statistical rigour of such an AI technology-assisted monitoring system by facilitating more accurate assessments of treatment efficacy.[38] As the data can be continuously collected, the ongoing tracking of a patient’s fluctuating condition provides richer information compared to a single-point measure.

In addition, AI-driven monitoring reduces the burden of manual data collection and analysis on medical staff, allowing them to allocate more time to direct patient care and complex medical tasks. On the patients’ side, they can undergo continuous monitoring from the comfort of their homes, eliminating the need for frequent clinic visits and reducing travel-related stress and expenses, and thus they can be more engaged for data collection.[39] Accessing regular medical care can be especially challenging for those residing in rural areas or nursing facilities. This aspect holds particular significance for the elderly population with movement disorders, the majority of whom face varying degrees of physical or cognitive limitations that hinder their capacity to travel to the clinic. In aggregate, AI-powered monitoring for movement disorders will lead to a more patient-centric approach to their clinical care, enhancing their overall satisfaction and quality of life.[40]

Particularly for elderly populations, video-based monitoring offers several advantages over wearable sensor-based monitoring. First, video models collect data without attaching devices to patients, allowing for purely passive, non-intrusive monitoring of daily activities. One major challenge encountered by currently available sensor-based technologies is the limited and highly heterogeneous digital literacy among the geriatric population.[41] Consequently, bias emerges, where only a select few with advanced digital skills can fully access the medical advantages offered by the latest technological advancements. Another major challenge is the long-term compliance issue caused by the intrusive nature of wearable sensors.[42] Artificial intelligence-powered video monitoring using consumer-level cameras makes it easier for the geriatric population to adapt, as they do not need to wear sensors and have special digital knowledge, thus maintaining their comfort during real-time tracking. Also, the system visually observes a patient’s movements and postures and provides critical visual data to healthcare professionals, accurately documenting the movement phenomenon. Meticulous visual observation is often considered the fundamental and the ‘art’ of movement disorder clinical practice. Clinical scales such as UPDRS or TETRAS instruct the raters to evaluate the score based on ‘what they see’.[11] To predict these scores, video-based visual inputs via computer vision analysis conceptually align better with the essence of UPDRS instructions to rate ‘what you see’ rather than the abstract parameters obtained from inertial sensors.

POTENTIAL PITFALLS OF VIDEO-BASED MONITORING

Despite the promise, monitoring movement phenomena via video-based technology poses several potential pitfalls that need to be addressed to enable real-world utilisation in a clinical setting. When examining this matter from a systemic perspective, we observe the following elements: (a) the input comprising the video data; (b) the desired outcome or gold standard information for the AI module to be trained on (represented by UPDRS); and (c) the central component, i.e. an AI/machine learning model responsible for translating inputs into the desired output. In the subsequent discussion, we delve into the potential pitfalls associated with each of these three crucial elements.

Input-related challenges

First, the video’s quality may affect an AI model’s performance.[43] We can categorise the potential factors that impede video quality into three groups: video-related, content-related and network-related factors. Video-related factors encompass not only basic video settings such as temporal and spatial resolutions, but also advanced aspects like lighting conditions in the recording space, motion blurring and focus. Content-related factors involve considerations like the distance between the target body part and the camera, as well as camera angles. In addition, unexpected motion blocks and random frame drops can occur due to network glitches, which can be ultimately misinterpreted as halts or hesitations. At the same time, erratic frame loss can result in an overestimation of movement speed, all of which are crucial components when assessing bradykinesia in PD. To ensure the models can withstand these potential challenges, we must identify and evaluate every factor influencing video quality, pose estimation and subsequent assessment models for the initial stage. Then, clear thresholds for each factor should be determined, and an automated, comprehensive quality control stage for input videos should be integrated into the model training process.

Privacy protection is another significant ethical concern relevant to video-based monitoring. Video recordings inherently capture more identifiable and private information, including the face, voice and surrounding environment of the person. One straightforward solution to address this concern at the camera level is to implement automated camera-end preprocessing of the recording. This process extracts only the key body part information at the camera end, creating a ‘skeletal’ or ‘avatar’ view of the body while discarding the rest of the video data sent to the central server. However, implementing this solution is not as straightforward as it may seem. To continuously track fluctuating symptoms, a home-based video monitoring system must ultimately be deployed. Consequently, the system needs to accurately detect only the target patient while excluding other people, for example, family members, who may also be in the frame. In fact, privacy compliance has been a longstanding debate in medical AI beyond the field of movement disorders.[44] Experts have suggested that integrating high-end AI models that are more privacy compliant, such as federated learning,[45] may be an appropriate solution for video-based monitoring. Federated learning, alternatively termed “collaborative learning”, is a methodology in machine learning, where the model is dispersed across local devices. These devices independently train the model using their unique data, which is not transferred or shared. This contrasts with conventional AI models that centralise data collection on a single server. Federated learning prioritises privacy by keeping individual data confined to the local device, and only the model’s parameters are conveyed to a central server for the purpose of enhancing the model. On a positive note, the perspective of elderly PD patients (mean age 61 years) on continuous, home-based video recording has been surprisingly positive, provided certain requirements are met, including secured privacy and the ability to turn off the camera at their discretion.[46] This finding underscores the unmet needs of end users, specifically patients with PD, for non-intrusive yet remote and automated disease monitoring.

Output-related challenges

One major pitfall pertains to the inherent limitation of using UPDRS and other clinical scales as labels for machine learning. In conventional supervised machine learning, models learn patterns in training data with ground-truth labels, which are then validated using a separate test dataset. Thus, label quality is crucial for model performance. Obtaining high-quality labels is especially vital in medical AI, a high-stakes field, but it is also challenging to acquire them. Labels for medical AI models require expertise from a limited number of medical professionals, making it costly and slow to obtain labels as compared to the rate of new data generation. In the context of PD, many previous studies have used clinical PD diagnoses or UPDRS Part III scores for labelling. However, clinical diagnostic criteria for PD are not definitive for pathological diagnosis. Moreover, the subjectively rated UPDRS Part III exhibits substantial intra-and inter-rater variability,[8,47] indicating a lack of a ‘ground truth’ to train models. Consequently, the ‘noisy’ nature of these labels must be considered when developing AI models for PD monitoring. There are strategies to address the ‘noisy label’ issue, such as using multiple annotators, in this case, multiple movement disorder experts, to establish a consensus UPDRS score. Weak supervision represents a paradigm in machine learning that trains models with imprecise or noisy labels, contrasting with the more conventional supervised learning that depends on precisely annotated data.[48] In weakly supervised models, heuristic rules are employed to automatically label a large training dataset, which are particularly useful when only a limited quantity of manually labelled or inherently noisy data is available. This approach has been effectively utilised in recent studies focusing on tremor detection and gait abnormalities related to PD, helping to address issues arising from noisy labels.[49,50] Furthermore, the generation of synthetic data has proven to be a significant adjunct to medical AI training, particularly in movement disorder contexts where real-world patient data acquisition is hindered by privacy issues or scarcity.[51] When combined with weak supervision techniques, these synthetic data generation methods facilitate the creation of diverse, cost-effective training datasets for computer vision models. However, it is crucial to use these techniques judiciously and in harmony with real patient data to ensure the resulting AI models are accurate, reliable and ethically sound.

However, beyond the noisy label issue, a fundamental question arises — is UPDRS Part III the ideal metric for monitoring patients’ disease status? While UPDRS Part III has been the gold standard for quantifying the severity of PD’s motor symptoms, it may not directly reflect a patient’s daily motor abilities. Rather, UPDRS Part II, a self-administered questionnaire on daily motor experiences like speech, eating and dressing, may provide a more direct insight into everyday motor abilities than UPDRS Part III, which comprises artificially induced movements less relevant to daily life (e.g. finger tapping). This concern is exemplified by the recent rejection of a smartwatch application designed to track PD progression based on UPDRS Part III-based tasks by the US Food and Drug Administration, which argue that “a change in rigidity or finger tapping in the MDS-UPDRS part III cannot be directly interpreted as being meaningful to patients”.[52]

AI/machine learning system-related challenges

First, the challenge of model generalisation stands out.[53] The AI models that are trained exclusively on specific patient populations run the risk of lacking adaptability when applied to different demographics. This concern necessitates implementing transfer learning techniques, where models initially trained on extensive datasets can be tailored to account for the intricacies of smaller, disease-specific data subsets. Using pretrained human pose estimation libraries that have been trained on healthy individuals and applying them to patients with movement disorder is also an example of transfer learning. This approach fosters enhanced versatility, enabling AI to transcend demographic boundaries.

Second, we have the potential pitfall of a lack of real-world validation.[54] The AI models refined through controlled research data might exhibit diminished performance within the dynamic environment of real-world clinical settings. To bridge this gap, collaboration with healthcare practitioners becomes imperative. Through close cooperation, AI algorithms can be rigorously validated under genuine clinical conditions, facilitating continuous adjustments to optimise their efficacy and reliability. A related challenge that emerges is the interpretability of AI.[55] The utilisation of complex black-box models can inadvertently erode clinicians’ trust and comprehension of AI-generated insights. A solution lies in prioritising interpretable AI models, such as decision trees and other rule-based mechanisms. Adopting these models makes the rationale underlying predictions more transparent, empowering clinicians to make informed decisions based on comprehensible insights.

Lastly, the problem of bias in data accentuates the complexity of this endeavour. The AI models trained on biased datasets might excel in specific demographic subsets while faltering in others. A comprehensive solution entails comprehensive data collection strategies embracing diverse ethnicities, genders and socioeconomic backgrounds. Moreover, consistent auditing and retraining of models are essential to actively counteract and mitigate bias, ensuring equitable performance across the entirety of the patient population.

FUTURE PERSPECTIVES

The ultimate form of video monitoring of PD should take the form of home monitoring of unconstrained daily activities. This approach would allow patients to carry on with unconstrained daily lives in their own homes while the system automatically modularises their movement patterns and assesses the degree of impairment in each movement. To achieve this, there are many hurdles to overcome with staged approaches. Currently, in terms of PD, most models predict clinicians’ UPDRS ratings using videos captured under an artificially refined environment according to the UPDRS instructions. Moving forward, the first step is to acquire data in unrefined, free-living conditions with complex movements of daily activities. Second, these complex movements need to be automatically classified into specific modules (e.g. walking, eating, dressing, etc.), and further, the degree of impairment for each action should be automatically measured. Third, this degree of impairment should strongly correlate with how the individuals actually feel about their impairments. Most importantly, all these endeavours should be carried out within secure ethical and regulatory boundaries.

CONCLUSION

Advancements in AI technology have significantly enhanced the role of telemedicine in the diagnosis and treatment of movement disorders. This review focused on the current state of video-based monitoring, particularly for PD and ET. Common approaches in recent studies have involved using consumer-level cameras, human pose estimation, kinetic feature extraction and machine learning algorithms for disease assessment. A notable shift towards emphasising both the performance and the explainability of models is seen, which is particularly relevant for movement disorder evaluation.

The technology holds great promise in improving clinical care for movement disorders, especially for the elderly population. It offers the advantages of remote, automated and continuous monitoring while ensuring non-intrusive data collection. Future research should aim to develop robust models that address challenges such as video quality, the inherent noisy nature of clinical scales and the maintenance of privacy compliance. These advancements will pave the way for accurate home-based monitoring through everyday activities, which would represent a significant milestone in the field.

Financial support and sponsorship

This study was supported by a Canadian Institutes of Health Research (CIHR)/National Science and Engineering Research Council (NSERC) Collaborative Health Research Project (grant no.: CPG-163986, awarded to MJM).

Conflicts of interest

There are no conflicts of interest.