gms | German Medical Science

67. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS), 13. Jahreskongress der Technologie- und Methodenplattform für die vernetzte medizinische Forschung e. V. (TMF)

21.08. - 25.08.2022, online

Reducing a complex two-sided smartwatch examination for Parkinson’s Disease to an efficient one-sided examination preserving machine learning accuracy

Meeting Abstract

Suche in Medline nach

  • Alexander Brenner - Westfälische Wilhelms-Universität Münster, Institut für Medizinische Informatik, Münster, Germany
  • Tobias Warnecke - Klinik für Neurologie und Neurorehabilitation, Klinikum Osnabrück – Akademisches Lehrkrankenhaus der WWU Münster, Osnabrück, Germany
  • Julian Varghese - Westfälische Wilhelms-Universität Münster, Institut für Medizinische Informatik, Münster, Germany

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie. 67. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS), 13. Jahreskongress der Technologie- und Methodenplattform für die vernetzte medizinische Forschung e.V. (TMF). sine loco [digital], 21.-25.08.2022. Düsseldorf: German Medical Science GMS Publishing House; 2022. DocAbstr. 161

doi: 10.3205/22gmds060, urn:nbn:de:0183-22gmds0603

Veröffentlicht: 19. August 2022

© 2022 Brenner et al.
Dieser Artikel ist ein Open-Access-Artikel und steht unter den Lizenzbedingungen der Creative Commons Attribution 4.0 License (Namensnennung). Lizenz-Angaben siehe



Introduction: With the development of a Smart Device System (SDS) we have conducted a prospective study to research Parkinson’s Disease (PD). This study provided a large PD sample size of two-hand synchronous smartwatch measurements for different movements, details are in the original design [1]. Previously we demonstrated the diagnostic potential of the sensor data using Machine Learning (ML) [2]. In this work we systematically reduced the utilized data to evaluate the potential of a light system with only one smartwatch. Thus, the objective was to simplify the system while maintaining classification accuracy.

Methods: The study was approved by the ethical board of the University of Münster and the physician’s chamber of Westphalia-Lippe (Reference number: 2018-328-f-S) and registered ( ID: NCT03638479).

504 participants were recorded and categorized into three groups: PD, differential diagnosis (DD) and healthy controls (HC). Neurologists confirmed all diagnoses. All participants performed 11 assessment steps for 10 to 20 seconds with one smartwatch attached to each wrist respectively. Features were computed on each combination of all 14 time-series (20 seconds split into two parts), 2 hands, 2 sensors (rotation and acceleration) and 3 spatial axes, resulting in 168 time-series channels per participant. For each channel the features were composed of power spectral density and statistical measures.

For ML we used support-vector machines in combination with 5-fold cross-validation in 3 randomized repetitions. Hyperparameters were selected via grid-search for every changing input setting. We tested three classification tasks: 1) PD vs. HC, 2) Movement disorders (PD + DD) vs. HC, 3) PD vs. DD. Performance was compared based on balanced accuracy.

For every classification task a forward and a backward feature selection was performed to evaluate performances on reduced sets of movements. For this process features were grouped movement-wise. The selection of the best recording arm was conducted analogously, both before and after movement subset selection.

Results: Baseline balanced accuracy was: 82.68% for task 1, 82.65% for task 2 and 67.65% for task 3. 3 movements were consistently excluded early in the optimization procedures. With removing these we got 82.25% for task 1, 82.64% for task 2 and 68.95% for task 3. Keeping only data from the right arm on the same set resulted in 83.24% for task 1, 80.99% for task 2 and 68.34% for task 3.

Discussion: The step-wise optimization has shown that a reduced set of movements achieved similar performance compared to the baseline. While we found differences in the importance of certain movements dependent on the classification task, we still observed a subset of relevant features for all classification tasks. Reducing the hardware to a single smartwatch generally showed only a marginal effect on accuracy.

Conclusion: Based on our analysis we identified a reduced assessment setting compared to our original study. This will reduce the time of active assessment while maintaining accuracy. Further, we have validated one-sided smartwatch measures as a valuable option for classifying PD. These changes reduce complexity of the SDS, making it more practical for routine-screening and potential home-based assessments in the future.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.


Varghese J, Niewöhner S, Soto-Rey I, Schipmann-Miletic S, Warneke N, Warnecke T, et al. A smart device system to identify new phenotypical characteristics in movement disorders. Frontiers in neurology. 2019;10:48. DOI: 10.3389/fneur.2019.00048 Externer Link
Varghese J, Alen CM, Fujarski M, Schlake GS, Sucker J, Warnecke T, et al. Sensor Validation and Diagnostic Potential of Smartwatches in Movement Disorders. Sensors. 2021;21(9):3139. DOI: 10.3390/s21093139 Externer Link