Machine Learning for Differentiating Essential Tremor: A Scoping Review

Methods

This scoping review was conducted in accordance with the 2020 PRISMA-ScR guidelines [2]. All studies identified for potential inclusion were indexed from PubMed, Cochrane Library, or Scopus. For each database, all articles published up to the date of data curation (April 2025) were included. The search strategy consisted of two parts: tremor and AI. Given the wide range of possible patient data types, the search was intentionally inclusive and did not restrict the type of data collected from the patient. Briefly, the types of tremor search consisted of medical subject headings (MeSH) of “Tremor” and “Essential Tremor.” The AI string contained terms related to machine learning and artificial intelligence including: “Machine Learning,” “Supervised Machine Learning,” “Unsupervised Learning,” and “Deep Learning.” These two strings were then combined using the AND boolean operator, which retrieves records containing both terms, according to the specific syntax requirements of each database. All search strategies used, along with subsequent article hits, can be found in the supplemental data S1.

Search strings yielded 548 studies, which were then uploaded to Covidence (Melbourne, Australia). Covidence is an online software that facilitates the screening of potential articles for scoping or systematic reviews; the software contains an artificial intelligence-based learning system that automatically removes duplicates and learns from the reviewers’ choices to remove irrelevant articles.

Eligible articles marked for inclusion in this scoping review included all published articles that used AI to distinguish ET from other tremors. Abstracts, commentaries, or posters were not included for analysis. Similarly, any article that only included animal studies, was not published (peer-reviewed), or was not written in English was excluded. Articles that used AI for different purposes were excluded. Additionally, studies that only compared ET to healthy controls or focused primarily on differentiating non-ET tremor types were excluded, as our scope was intentionally limited to studies distinguishing ET from at least one other tremor disorder. Review articles were not included in the final study count; however, the team reviewed them to identify any additional articles not detected by the search terms. No such additional articles were identified.

To assess article eligibility, two authors (DMF and KEH) initially screened each article based on the title and abstract. This screening used defined PICO criteria described briefly above and expanded in supplemental data S2. Both authors had to agree on article inclusion in order for it to proceed to a second review. If there was a conflict, a third, neutral author (SBR) made the final decision on inclusion. After the first review, included articles were reviewed for a second time, where the first two authors read the full text and decided on inclusion based on the PICO criteria. Again, the third author determined inclusion for any articles with disagreements between the original two authors.

If an article passed two rounds of screening, the authors then extracted data. Data extracted from each article included: study design, total number of participants (for each tremor type), the data recorded, the AI method implemented, and the resulting accuracy, sensitivity, and specificity.

The Joanna Briggs Institute (JBI) Critical Appraisal Tool was used to report bias for each paper [19]. The JBI Critical Appraisal Tool classifies papers based on their design (i.e., case report, clinical trial, cohort study, etc.), and reviewers score the risk of bias using a series of 8–11 questions [19]. Two reviewers (DMF and KEH) independently scored each paper. If there were conflicts, a third reviewer (SBR) broke the tie. The scores were then used to develop a bias assessment, which consisted of a scale with options of “yes, no, or unclear” for each specific question. The overall bias assessment for each study was an average of the questions that scored a “yes.”

Results

The process for manuscript selection is displayed in Figure 1. A total of 721 articles were identified in the literature review. Covidence automatically removed 161 duplicates, and 12 additional duplicates were identified manually, leaving 548 articles for screening. Two reviewers independently assessed the deduplicated titles and abstracts from the literature search for relevance, which resulted in the exclusion of 451 additional articles. Subsequently, the 97 articles whose titles and abstracts met the initial inclusion criteria underwent a full-text review to evaluate study design quality and determine which met the final inclusion criteria, resulting in 46 included articles [11, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64]. This process, including the rationale for article exclusion post-full-text review, is detailed in Figure 1. The final set of studies included articles published between 1999 and 2025, reflecting a quarter of a century of using AI to differentiate ET from other types of tremors. Supplemental data S3 presents each article included in this scoping review, grouped by its study design. Table 1 contains the information below in one central format.

Figure 1

Types of Data Acquired and AI Models

From each included article, both the type of data acquired from participants and the AI models employed were recorded. Common data types included accelerometer data (26 articles), EMG (9 articles), Archimedes spirals (6 articles), gyroscope data (5 articles), and video recordings (5 articles). A comprehensive list of all recorded patient data types is provided in supplemental data S6. The AI models used varied across studies, with the most frequently applied being support vector machines (18 articles), k-nearest neighbors (9 articles), and convolutional neural networks (8 articles) (Figure 2). A comprehensive list of AI models used, along with their respective classifications, is presented in supplemental data S7.

Figure 2

Discussion

This scoping review evaluated 46 studies that utilized ML models with electrophysiological signals, kinematic measurements, drawing tasks, functional assessments, clinical rating scales, and video recordings to distinguish ET from other tremor types across 6,501 patients. All studies met the JBI classification for diagnostic test accuracy studies, with an average quality score of 83.91%. The included studies employed a variety of ML approaches, with reported classification accuracies ranging from 60% to 100% and an overall average of 89%. However, there was substantial heterogeneity in patient demographic reporting, tremor data types, AI models used, and data reporting methods. This variability prevented direct comparison across studies and precluded systematic review or meta-analysis techniques. Importantly, given the clinical overlap between tremor syndromes, these findings highlight the potential of ML models as pattern recognition tools that can support and complement neurologists’ clinical expertise in the assessment of tremor disorders.

Examination of the patient populations and study designs revealed limitations in demographic reporting and study methodology. Across the 46 studies, reporting of basic patient demographic information such as race, gender, and age was inconsistent, limiting comparisons across cohorts. Furthermore, no study included case-matched controls, nor were any randomized controlled trials identified. These design elements are important for objectively evaluating AI performance in distinguishing ET from other tremor types. The emergence of the ET-plus classification introduces additional complexity for ML-based tremor studies. Because ET-plus is variably defined, inconsistently applied across centers, and often underreported, its presence likely contributes to diagnostic heterogeneity and label noise in existing datasets [3, 6]. Addressing these limitations in future research would allow more robust comparisons across AI models and study settings.

Analysis of data acquisition methods across the 46 included studies revealed clear trends in the modalities used for ML model input. Accelerometer data was the most common modality, appearing in 26 studies, followed by EMG (9 articles), Archimedes spirals (6 articles), gyroscope data (5 articles), and video recordings (5 articles). From a practical standpoint, however, not all data acquisition methods translate easily into clinical practice. For example, while EMG provides rich electrophysiological detail and often incorporates accelerometer sensors, full EMG-based assessment workflows may be difficult to implement routinely in outpatient settings due to time requirements, technical expertise, reimbursement constraints, and patient tolerance [65, 66]. In contrast, Archimedes’ spiral drawings, which were used in only six studies, are widely adopted in clinical assessment because of their simplicity, minimal equipment requirements, and ability to visually capture tremor amplitude and variability [67, 68, 69]. Although accelerometer and EMG-based approaches dominate the research space, they may be less feasible in typical outpatient neurology settings, highlighting a disconnect between research priorities and clinical practice. A potential middle ground may lie in smartphone-based assessments that integrate accelerometer and gyroscope data. Although only two studies utilized smartphones, they reported accuracies of 77.90% and 88.00%, comparable to more complex sensor configurations but with far greater theoretical ease of deployment [26, 29].

In addition to the range of data types collected, the ML techniques used to distinguish ET from other tremor types also varied considerably. The most common approach was support vector machines (SVM), used in 18 studies, followed by k-nearest neighbors (KNN) in nine studies and convolutional neural networks (CNN) in eight studies. These are all supervised learning models that rely on labeled examples during training to classify new data. For instance, if the algorithm is trained using tremor patterns labeled as PD or ET, it learns decision boundaries that allow it to distinguish between these categories [18]. However, model training depends heavily on the availability and quality of labeled datasets, an area where the field still faces major limitations. Most studies relied on small, institution-specific datasets, underscoring the need for large, standardized, publicly available datasets. A consortium-led initiative would be well positioned to coordinate this effort and develop equitable AI models that could aid diagnosis or potentially streamline clinical visits for future patients [70].

While standardizing and simplifying data inputs is essential for clinical implementation, evaluating how different ML models perform is equally important, though not straightforward. Comparing diagnostic accuracy across studies proved challenging because of substantial methodological variability. Each study used its own combination of data types, preprocessing steps, model architectures, and evaluation criteria. In addition, many studies employed ensemble or hybrid approaches, making it difficult to isolate the contribution of individual ML methods. Reporting of evaluation metrics also varied; although most studies reported overall accuracy, fewer included sensitivity, specificity, or area under the curve, which are important for assessing clinical utility. Moreover, almost all models were trained and validated on internal datasets from single institutions, with little use of external validation cohorts. This limits understanding of model generalizability across diverse populations and clinical settings.

Among commonly used models, SVMs appeared in 18 studies, with reported accuracies ranging from 83% to 100%. KNN was used in nine studies, with accuracies ranging from 60% to 94.68%, while CNNs appeared in eight studies with accuracies between 70% and 95.40%. Although these ranges demonstrate generally strong performance, they also indicate that no single method consistently outperforms others across all scenarios. Differences in data types, preprocessing approaches, model architectures, and evaluation strategies preclude definitive conclusions regarding the best-performing AI approach. Furthermore, no studies directly compared model performance against clinicians, since clinician diagnosis typically served as the reference standard for labeling training and test data. Consequently, it remains unclear how these models perform relative to expert clinical judgment. For these reasons, a meta-analysis was not performed, and accuracy values should be interpreted within the context of each study’s design rather than through direct comparisons across models.

Future research should prioritize improved demographic reporting, larger datasets, and more consistent study methodologies. Coordinated multi-center efforts could establish shared protocols for collecting and reporting tremor data while enabling the development of large, publicly available datasets that allow fair comparison of AI models. These efforts should also prioritize tools that integrate into routine clinical practice. Nearly all included studies relied on device-based data such as accelerometers, EMG, and video recordings, whereas AI applications based solely on clinical history or physical examination data were notably absent. Given the practicality and lower resource demands of clinical data, exploring AI methods that incorporate or rely on clinical assessments represents an important future direction.

Although EMG provides detailed electrophysiological information, its routine clinical use can be limited by equipment requirements, time demands, and the need for specialized expertise [65, 66]. In contrast, approaches such as Archimedes’ spirals, accelerometers, and gyroscopes are more accessible, particularly when integrated into smartphone-based platforms. One potential approach could involve a smartphone application developed collaboratively by movement disorder specialists and computer scientists. However, smartphone- and wearable-based assessments should currently be considered exploratory rather than validated clinical solutions. Although smartphones can capture accelerometer and gyroscope data, their size, weight, sensor placement, and the need for active holding may alter tremor biomechanics and signal characteristics, effects that wearable devices might mitigate. These factors have not been systematically studied and may introduce confounding effects. Consequently, both smartphone- and wearable-based approaches should currently be viewed as hypothesis-generating tools rather than ready-to-deploy clinical instruments.

Financial Disclosures for the previous 12 months

KEH is supported by the National Science Foundation Research Traineeship (NRT) Program under Grant No. 2125872. AWA is a member of the faculty of the University of Colorado and is supported by university funds. She has served as a site investigator for studies sponsored by Michael J Fox Foundation for Parkinson’s Research, Parkinson Study Group, Aligning Science Across Parkinson’s (ASAP) Initiative, Koneksa Health, Biohaven Pharmaceuticals, Inc., Bial, and the NIH NINDS. She receives grant funding from NIH NICHD. She has received honoraria for consultancy to Aevum through Parkinson Study Group. She is a consultant for PhotoPharmics, Inc. The other authors declare that there are no additional disclosures to report.

Additional Files

The additional files for this article can be found as follows:

Ethics and Consent

The authors confirm that the approval of an institutional review board and informed patient consent was not required for this work. We confirm that we have read the Journal’s position on issues involved in ethical publication and affirm that this work is consistent with those guidelines.

Author Contributions

(1) Research project: A. Conception, B. Organization, C. Execution; (2) Statistical Analysis: A. Design, B. Execution, C. Review and Critique; (3) Manuscript Preparation: A. Writing of the first draft, B. Review and Critique

DMF: 1B, 1C, 2A, 2B, 3A

KEH: 1B, 1C, 2A, 2B, 3A

SBR: 1C, 3A

AM: 1C, 3A

JCM: 1C, 3B

AWA: 3B

AAM: 1A, 1B, 2C, 3B

David M. Fletcher and Kaitlyn E. Heintzelman contributed equally to this work.