Abstract:
Parkinson’s disease (PD) is a progressive neurological disorder that results in loss of
dopaminergic neurons in brain and has a profound effect in patients, particularly because of the motor symptoms, including bradykinesia, tremor and dyskinesia. The
focus of this thesis is to enhance the prediction and monitoring of such symptoms by
developing an encoder-decoder based transformer model in conjunction with wearable
inertial sensors for multi-label classification. The proposed model adopts an encoderdecoder structure within a transformer framework and employs multiple attention
heads to capture subtle features and temporal relationships from the data gathered by
the smartphones, smartwatches, and wrist-worn devices. Notably, wrist-worn sensors
provide particularly rich data, contributing to exceptional performance with diagnostic
accuracies of 99.11% with phone data, 96.61% with smartwatch data, and 99.34% with
wrist-worn sensor data for multi-label classification. This approach emphasizes on how
the placement of sensors affects the performance, as the wrist-worn sensors perform
better than those placed in other positions. The model also achieves low false positive
and false negative rates, along with strong precision and recall values, and exhibits good
generalizability. Advanced preprocessing techniques, including normalization, feature
scaling, and sequence generation, further enhance model efficiency and address class
imbalance. This innovative solution not only improves the management of PD but also
enhances patients’ quality of life through a targeted and responsive approach to care.
The findings highlight the potential of this technology to transform Parkinson’s disease management, paving the way for early detection and close monitoring of patient
symptoms through a novel multi-label classification architecture. Thus, the incorporation of these technologies improves diagnostic assessment and understanding of motor
symptom evolution, enabling more targeted and efficient treatment plans based on individual patient conditions, ultimately enhancing PD management and patient quality
of life.
