Anandapadmanabhan Reghu, Vishnoi Aayushi, Raman Geetha, Thachan Jeena, Gangaraju Beulah Amulyavathi, Radhakrishnan Divya, Vishnu Venugopalan Yamuna, Kamble Nitish, Holla Vikram, James Praveen, Srivastava Achal, Joshi Deepak, Mahabal Ashish, Krishnan Syam, Pal Pramod, Rajan Roopa
Department of Neurology, All India Institute of Medical Sciences (AIIMS), New Delhi, India.
Comprehensive Care Centre for Movement Disorders, Sree Chitra Tirunal Institute for Medical Sciences and Technology, Thiruvananthapuram, India.
Mov Disord. 2025 Mar 17. doi: 10.1002/mds.30176.
No objective biomarkers exist for diagnosing and classifying tremor syndromes.
The aim was to develop and validate a deep learning (DL) algorithm for classifying tremors from hand-drawn pen-on-paper spirals.
We recruited participants with dystonic tremor (DT), essential tremor (ET), essential tremor plus (ETP), Parkinson's disease (PD), cerebellar ataxia (AT), and healthy volunteers (HV). Participants drew free-hand spirals on paper, which were used to train a DL algorithm based on transfer learning using InceptionResNetV2 and Keras sequential model. We validated the model externally in two independent tremor cohorts, evaluating accuracy and F1 scores, and compared its performance to expert raters.
We recruited 521 participants and obtained 2078 spirals (365 DT, 215 ET, 208 ETP, 212 PD, 78 AT, and 525 HV). Mean age of the participants was 46.1 ± 12.4 years, duration of illness was 8.9 ± 8.3 years, and the mean Fahn-Tolosa-Marin Tremor Rating Scale score in patients was 32.4 ± 14.7. The DL classifier demonstrated an overall accuracy of 81% [95% confidence interval, CI: 0.77-0.85] in distinguishing among tremor syndromes. To mitigate the potential risks of data leakage and digital fingerprinting, the algorithm was redeveloped and reanalyzed, yielding an adjusted accuracy of 70% [95% CI: 0.66-0.74]. External validation on an independent cohort of 1535 spiral drawings resulted in an accuracy of 61% [95% CI: 0.59-0.63], with the adjusted algorithm achieving 59% [95% CI: 0.58-0.60], outperforming human raters (accuracy: 46%).
Supervised DL algorithms can effectively detect tremor syndromes from hand-drawn spirals, offering unbiased, feature-independent classification with higher accuracy than human raters. © 2025 International Parkinson and Movement Disorder Society.
目前尚无用于诊断和分类震颤综合征的客观生物标志物。
旨在开发并验证一种深度学习（DL）算法，用于根据纸上手绘螺旋线对震颤进行分类。
我们招募了患有肌张力障碍性震颤（DT）、特发性震颤（ET）、特发性震颤叠加型（ETP）、帕金森病（PD）、小脑共济失调（AT）的参与者以及健康志愿者（HV）。参与者在纸上徒手绘制螺旋线，这些螺旋线被用于基于迁移学习使用InceptionResNetV2和Keras顺序模型训练DL算法。我们在两个独立的震颤队列中对该模型进行外部验证，评估准确性和F1分数，并将其性能与专家评分者进行比较。
我们招募了521名参与者，获得了2078条螺旋线（365条DT、215条ET、208条ETP、212条PD、78条AT和525条HV）。参与者的平均年龄为46.1±12.4岁，病程为8.9±8.3年，患者的平均法恩 - 托洛萨 - 马林震颤评定量表评分为32.4±14.7。DL分类器在区分震颤综合征方面的总体准确率为81%[95%置信区间，CI：0.77 - 0.85]。为了减轻数据泄露和数字指纹识别的潜在风险，对算法进行了重新开发和重新分析，调整后的准确率为70%[95%CI：0.66 - 0.74]。在一个由1535幅螺旋线图组成的独立队列上进行的外部验证中，准确率为61%[95%CI：0.59 - 0.63]，调整后的算法达到59%[95%CI：0.58 - 0.60]，优于人类评分者（准确率：46%）。
监督式DL算法可以有效地从手绘螺旋线中检测震颤综合征，提供无偏倚、与特征无关的分类，且准确率高于人类评分者。©2025国际帕金森和运动障碍协会。
