Objective: The Clinical Rating Scale for Tremor (CRST) is widely adopted for quantifying upper extremity tremor severity in essential tremor (ET), yet manual scoring remains limited by inter-rater variability and integer-only resolution (0-4), where scores across patients may not reflect equivalent underlying severity. With growing volumes of patients undergoing focused ultrasound thalamotomy, scalable and reproducible scoring has become essential.

Background: We developed a multimodal vision-language model trained on a curated dataset of expert clinician-annotated CRST spiral drawing PDFs to learn the visual features corresponding to each severity band, including line smoothness, tremor amplitude, loop regularity, and overall spiral form integrity. For every score, the system produces a derived Tremor Index (0-100), a confidence estimate (0.0-1.0), and a structured reasoning block grounded exclusively in the actual drawing content.

Method: Ground truth labels were established through independent expert clinician scoring using standard CRST criteria (0-4). The model was trained end-to-end on these annotated drawings and outputs scores on a continuous 0.0-4.0 scale, addressing the known limitation of integer-only bins. Each drawing task (large spiral, small spiral, straight lines, and handwriting) is scored independently. The trained model was evaluated on 480 patients encompassing 2,985 drawing pages collected between 2021 and 2025, with each patient treated as a unique case.

Results: Across the 480-patient evaluation cohort, the model produced a mean CRST score of 1.43 (range 0.37-3.50). The continuous scale resolved within-category variation that integer scoring collapses, distributing patients within each integer band across a full range of single-decimal values. Severity classification yielded 27.5% Normal-Trace, 55.2% Slight, 15.0% Moderate, and 2.3% Marked-Severe, consistent with expected clinical distributions. The model maintained high confidence across variable scan qualities, and the Tremor Index translated small score differences into intuitive severity shifts on a 0-100 scale.

Conclusion: Validated across 480 patients and 2,985 drawing pages, this system delivers clinician-grade automated CRST spiral drawing scoring with continuous-scale precision, per-drawing transparency, and the scalability required for high-volume clinical programs and multi-site trials.

References: [1] Anandapadmanabhan R, et al. Deep Learning–Based Artificial Intelligence Algorithm to Classify Tremors from Hand-Drawn Spirals. Movement Disorders. 2025;40(6):1172–1181. doi:10.1002/mds.30176

[2] Peng Y, et al. A Deep Learning Approach to Remotely Assessing Essential Tremor with Handwritten Images. Scientific Reports. 2025;15:10783. doi:10.1038/s41598-025-94729-0

[3] Fahn S, Tolosa E, Marín C. Clinical Rating Scale for Tremor. In: Jankovic J, Tolosa E, eds. Parkinson’s Disease and Movement Disorders. Baltimore: Williams & Wilkins; 1993:271–280.

[4] Hopfner F, et al. Archimedes Spiral Ratings: Determinants and Population-Based Limits of Normal. Movement Disorders Clinical Practice. 2024;11(10):1257–1265. Doi:10.1002/mdc3.14201

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/automated-crst-spiral-drawing-scoring-using-a-multimodal-vision-language-model-development-and-validation-in-a-large-essential-tremor-cohort/