Deep learning-based artificial intelligence algorithm to classify tremors from hand-drawn spirals

---

 

In this study, we developed and validated a deep learning (DL)-based artificial intelligence algorithm capable of classifying common tremor syndromes using hand-drawn pen-on-paper spiral images. Despite advancements in neurophysiology and wearable technologies, the diagnosis of tremor syndromes remains primarily clinical and subjective, often lacking objective biomarkers. We aimed to address this gap by designing a scalable, non-invasive diagnostic tool grounded in computer vision and AI.  

We prospectively recruited 521 participants, including patients with dystonic tremor (DT), essential tremor (ET), essential tremor plus (ETP), Parkinson’s disease (PD), cerebellar ataxia (AT), and healthy volunteers (HV), from the All India Institute of Medical Sciences (AIIMS) in New Delhi. Participants provided a total of 2,078 spiral drawings, which were digitized and used to train a DL classifier based on the InceptionResNetV2 architecture with transfer learning, implemented via a Keras sequential model. Rigorous data augmentation, stratified sampling, and oversampling were employed to enhance generalizability and mitigate class imbalance.  

The model achieved an initial multiclass classification accuracy of 81% in the development cohort, with particularly high performance in identifying PD, AT, and HV. To address concerns related to data leakage and digital fingerprinting, we restructured the dataset and redeveloped the model using patient-level grouping. The redesigned model achieved a robust accuracy of 70%.  

For external validation, we tested the model on an independent cohort of 1,535 spiral drawings, where it retained an accuracy of 59–61%, outperforming trained human raters (average accuracy: 46%) who were asked to classify spirals without clinical context. Notably, the model demonstrated superior inter-category discrimination and greater reliability than expert visual assessments, particularly for PD and HV.  

Our findings affirm that hand-drawn spirals — an easily deployable clinical tool — contain rich, quantifiable features that can be leveraged by DL models for diagnostic classification. The model’s limitations in classifying ET and ETP underscore the phenotypic overlap between these categories, aligning with emerging evidence that questions their biological distinctness. The study also highlights the value of external validation in ensuring real-world applicability.  

This work represents one of the largest AI-based studies in tremor classification using pen-and-paper spirals and demonstrates the feasibility of using DL for accurate, objective, and scalable tremor diagnostics. Our approach holds promise for integration into telemedicine workflows, epidemiological screening, and AI-assisted clinical decision support systems. 

View paper