| Abstract [eng] |
Hand tremor is a rhythmic shaking of one or both arms that can occur as a symptom of various diseases or as a common occurrence in the general population. It is the most prevalent movement disorder seen in clinical practice. Among the numerous diseases associated with hand tremors, essential tremor (ET) and Parkinson's disease (PD) are the most prevalent. This study examines conventional and contemporary methods for diagnosing and assessing the severity of tremors. Additionally, it proposes an innovative solution that combines accelerometer and touch sensor data from mobile smart devices. The quantification of tremors is approached as a regression problem, with frequency and time-domain features serving as dependent variables and an estimate of tremor strength ranging from 0 (severe tremor) to 10 (no tremor) as the independent variable. A total of 440 sample pairs of accelerometer and touch sensor measurements and corresponding tremor strength estimates were collected using a mobile application. These samples encompassed individuals with ET (241), a control population (CP) (180), and PD (19). The collected data was used to train and evaluate XGBoost and multilayer perceptron (MLP) regression algorithms. The results yielded a mean absolute error (MAE) of 0.58 and a root mean square error (RMSE) of 0.79. Furthermore, the diagnosis problem was approached as a classification problem, with three classes: ET, CP, and PD. Using the MLP method led to nearly perfect outcomes, as evidenced by an area under the true positive rate/false positive rate curve (AUC) of 1.00, a Cohen's kappa coefficient (KAP) of 0.98, and only one mispredicted sample. The study also addresses the challenges associated with processing sensor data from smart devices. Lastly, an evaluation of the potential economic impact of the developed tool is presented. |
