Sophie Fleischmann (Erlangen), Vivien Holzwarth Correa (Erlangen), Birte Coppers (Erlangen), Misha Sadeghi (Erlangen), Robert Richer (Erlangen), Dr. Arnd Kleyer (Erlangen), Dr. David Simon (Erlangen), Johanna Bräunig (Erlangen), Prof. Martin Vossiek (Erlangen), Dr. Verena Schönau (Erlangen), Prof. Georg Schett (Erlangen), Prof. Anne D. Koelewijn (Erlangen), Prof. Sigrid Leyendecker (Erlangen), Prof. Björn M. Eskofier (Neuherberg; Erlangen), PD Dr. Anna-Maria Liphardt (Erlangen)
Abstract-Text (inkl. Referenzen und Bildunterschriften)
Intro
Classifying movement restrictions in patients with rheumatoid arthritis (RA) is challenging. Machine learning (ML) can identify patterns in the data that cannot be captured with statistical tools.
Aim
To test the feasibility of classifying RA patients and controls from hand motion capture data using automatic feature extraction (AFE) and ML and to evaluate if a larger feature set improves the performance.
Methods
24 RA patients (ACR/EULAR 2010 criteria) and 23 controls performed tipping and flexion of the hand [1], which was captured by an optoelectronic measurement system at 100 Hz with 29 reflective markers placed on the hand dorsum. We extracted the 3D trajectories, expressed relative to the central wrist marker and normalized to the hand length, of 24 markers on one hand (excluding thumb markers). After linearly interpolating missing values we filtered the data using a 4th-order Butterworth filter at a 6 Hz cut-off frequency and extracted a minimal and extensive feature set from the hand trajectories using the AFE package tsfresh [2]. We compared 4 classifiers in combination with scaling and dimensionality reduction (Fig 1). For every combination of algorithms, we trained a model using a stratified 10-fold nested cross-validation (CV). Hyperparameters were found from an inner five-fold CV on the respective train set using grid search. We report the accuracy of the best model averaged over the 10 folds.
Results
The best accuracy was 68.2% for flexion and 62.7% for tipping (Fig 2). Using the extensive instead of the minimal feature set led to no major change in accuracy.
Summary
The classification performance based on AFE was below the anticipated level of accuracy (>70%), likely due to the small dataset with limited data quality. We currently evaluate the approach on a more consistent, larger dataset and include expert features in addition to AFE.
References
[1] Phutane et al., Sensors, 21(4), 1208, 2021
[2] Christ et al., Neurocomputing, 307, 72-77, 2018