Subclinical tremor differentiation using long short-term memory networks

• Published in: Australasian physical & engineering sciences in medicine Vol. 48; no. 2; pp. 591 - 602
• Main Authors: Nanayakkara, Gerard Ruchin Randil, Chan, Ping Yi
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.06.2025; Springer Nature B.V
• Subjects: Aged; Amplitudes; Artificial intelligence; Biological and Medical Physics; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Biophysics; Deep Learning; Diagnosis, Differential; Differentiation; Essential Tremor - diagnosis; Feature selection; Female; Fourier transforms; Humans; Inertial sensing devices; Machine learning; Male; Medical and Radiation Physics; Memory, Short-Term; Middle Aged; Neural Networks, Computer; Parkinson Disease - diagnosis; Parkinson's disease; Patients; Physiology; Scientific Paper; Sensors; Statistical analysis; Tremor (Muscular contraction); Tremor - diagnosis; Tremors; Subclinical tremor; LSTM; Essential tremor; Parkinson’s disease
• ISSN: 2662-4729; 0158-9938; 2662-4737; 2662-4737; 1879-5447
• DOI: 10.1007/s13246-025-01526-0

Subclinical amplitudes complicate the differentiation between essential tremor (ET) and Parkinson’s disease (PD) tremor, which is uncertain even when the tremors are apparent. Despite their prevalence—up to 30% of PD cases exhibit subclinical tremors—these tremors remain inadequately studied. Therefore, this study explores the potential of artificial intelligence (AI) to address this differentiation uncertainty. Our objective is to develop a deep learning model that can differentiate among subclinical tremors due to PD, ET, and normal physiological tremors. Subclinical tremor data were obtained from inertial sensors placed on the hands and arms of 51 PD, 15 ET, and 58 normal subjects. The AI architecture used was designed using a long short-term memory network (LSTM) and was trained on the short-time Fourier transformed subclinical tremor data as the input features. The network was trained separately to differentiate firstly between PD and ET tremors and then between PD, ET, and physiological tremors and yielded accuracies of 95% and 93%, respectively. Comparative analysis with existing convolutional LSTM demonstrated the superior performance of our work. The proposed method has 30–50% better accuracies when classifying low amplitude tremors as compared to the reference method. Future enhancements aim to enhance model interpretability and validate on larger, more diverse datasets, including action tremors. The proposed work can potentially serve as a valuable tool for clinicians, aiding in the differentiation of subclinical tremors common in Parkinson’s disease, which in turn enhances diagnostic accuracy and informs treatment decisions.