Development and Validation of a contactless tool for measuring patient tremor

Background Tremor affects one million people in the UK and is a common manifestation of many neurological diseases, drug side effects, and psychological disorders. It is usually assessed via human judgement. However, assessment is time consuming and can differ between assessors.

An accelerometer can be used to provide objective measures of tremor severity, but few hospitals have access to this equipment.

Our aim is to develop and validate software that accurately measures tremor frequency and amplitude using only smartphone or webcam video. In previous work, we have already demonstrated how tremor frequency can be ascertained from smartphone videos.

Therefore, the scientific objectives of the project are (i) to develop and validate a method to measure tremor amplitude from video (ii) to validate the resulting tremor frequency and amplitude software (iii) to assess usability of the software (iv) to conduct exploratory data analysis using deep learning to classify clinical tremor type.

Methods Objective (i) - To develop contactless tremor amplitude measurement, we need to accurately assess object depth. We will investigate three approaches: parallax depth estimation, LIDAR sensor, and Densedepth neural network. We will collect pilot data from healthy volunteers, mimicking tremor, to refine our data collection process.

We will then collect patient hand tremor videos (N=60), recorded on smartphone and webcam.

Approaches will be compared to an objective standard using Bland-Altman analysis to provide 95% limits of agreement for tremor amplitude.

The best performing approach will be used, alongside our previous tremor frequency algorithm, to develop standalone software for smartphone camera and webcam. Objective (ii) - We will collect further tremor videos from patients to assess the software. Amplitude will be assessed as in Objective (i). Frequency will be compared to clinical-grade accelerometer, using a Bland-Altman analysis.

Objective (iii) – We will assess usability via a time-motion study of clinical users of the software. This will be supported by qualitative user surveys and interviews.

Objective (iv) - In secondary exploratory analysis, we will apply 3D convolutional neural networks to the videos collected for (ii) to determine whether subtle clinical features can be extracted from the raw video to automatically distinguish between different types of clinical tremor. We will report cross-validation F1-score, sensitivity and specificity.

Delivery Timeline All objectives will be completed within 24 months.

Objective (i) is scheduled for M1-M9, Objective (ii) is scheduled for M8 to M18, Objective (iii) is scheduled for M13 to M18, and Objective (iv) is scheduled for M18 to M24.

Impact and dissemination This project will deliver a clinically validated tool to quantify hand tremor frequency and amplitude.

We expect it to have high translational impact amongst doctors and other clinicians, especially neurologists, GPs, geriatricians, and researchers.

Results describing the development of methods, development of software, and validation of software, will be written for publication in high-impact clinical and health informatics journals.

Finally, the work will provide a platform for us to conduct larger-scale research into digital biomarkers, by providing an infrastructure for research video data to be collected and securely stored from multiple sites.