Activity-aware essential tremor evaluation using deep learning method based on acceleration data

Abstract

Background: Essential tremor (ET), one of the most common neurological disorders is typically evaluated with validated rating scales which only provide a subjective assessment during a clinical visit, underestimating the fluctuations tremor during different daily activities. Motion sensors have shown favorable performances in both quantifying tremor and voluntary human activity recognition (HAR).

Objective: To create an automated system of a reference scale using motion sensors supported by deep learning algorithms to accurately rate ET severity during voluntary activities, and to propose an IOTA based blockchain application to share anonymously tremor data.

Method: A smartwatch-based tremor monitoring system was used to collect motion data from 20 subjects while they were doing standard tasks. Two neurologists rated ET by Fahn-Tolosa Marin Tremor Rating Scale (FTMTRS). Supported by deep learning techniques, activity classification models (ACMs) and tremor evaluation models (TEMs) were created and algorithms were implemented, to distinguish voluntary human activities and evaluate tremor severity respectively.

Result: A practical application example showed that the proposed ACMs can classify six typical activities with high accuracy (89.73%-98.84%) and the results produced by the TEMs are significantly correlated with the FTMTRS ratings of two neurologists (r₁ = 0.92, p₁ = 0.008; r₂ = 0.93, p₂ = 0.007).

Conclusion: This study demonstrated that motion sensor data, supported by deep learning algorithms, can be used to classify human activities and evaluate essential tremor severity during different activities.

Keywords: Blockchain; Convolutional neural network; Deep learning; Essential tremor; Human activity recognition; IoTA.