Neurological disorders such as cerebral palsy can lead to lifelong limitations in everyday life. With an early diagnosis within the first year of life, it is possible to initiate targeted therapy and avoid or reduce consequential damage and high rehabilitation costs later on. Although there exist approaches for the evaluation of gross motor development, there is usually no targeted assessment by welltrained clinicians. The aim of the thesis was therefore the systematic development and validation of a user-friendly movement detection system to support the early detection of neurological movement restrictions. After extensive literature research, a recording system was developed to capture infant’s movements using two RGBD-cameras, seven Inertial Measurement Units (IMUs) and a sensor mattress. In a data collection study, 67 data sets from infants in a clinical setting were recorded. Accompanying the measurements, clinical staff and parents were asked to evaluate the system. Based on an abridged version of the AttrakDiff questionnaire, consistently high acceptance of the system was confirmed. Algorithms were trained and tested for the different movement parameters based on the camera data and inertial sensor data. Due to the large class imbalance of most parameters, greater attention was paid to balanced weighting. Five-fold cross-validation was performed with separate test subjects. As a result, models based on artificial intelligence for ten movements were created. These achieved up to 95% balanced accuracy with the camera data (hand-hand contact) and up to 92% balanced accuracy with the IMU data (legs lift). In addition, the “body symmetry” and the movement character were assessed with threshold-based algorithms. Side differences between right and left are taken into account for an indication of unilateral cerebral palsy. With the involvement of therapists and paediatricians, the user interface and the hardware for a demonstrator system were evaluated. Immediately after the measurement, users are shown whether the movements detected indicate typical or delayed development. An observational study with seven infants and three therapists has shown that the reports issued correspond to the therapists’ observations. In an expanded form, incorporating additional movement parameters, the system has great potential for use in paediatric practices. In user tests with paediatricians, interest was shown in integrating the system into everyday practice and cooperating on further research. For a valid system, the robustness of the IMU connection via Bluetooth must be enhanced, or a feasible solution to use only a camera device must be developed (with or without depth information). Furthermore, the safety, private policy and the regulations of the AI act must be more considered. The data set needs to be enlarged and a higher amount of atypical behaviours would be beneficial for the development of enhanced classification models.
Vivian Waldheim (Thu,) studied this question.