Identification of Parkinson’s Disease Utilizing a Single Self-recorded 20-step Walking Test Acquired by Smartphone’s Inertial Measurement Unit

Parkinson’s disease detection from 20-step walking tests using inertial sensors of a smartphone: Machine learning approach based on an observational case-control study

“…These accuracies are comparable to earlier studies found in the literature. The earlier literature has shown slightly higher accuracies for measurements performed at clinics or research laboratories [22], and smaller datasets [6], but a lower accuracy for completely independent home measurements [28]. Even though RF has been found very efficient in earlier studies [6], it did not perform particularly well in our analysis.…”

“…The best performing classifier in the overall classification was kNN achieving accuracy, sensitivity and specificity of 84.5%, 88.5%, and 81.3%, respectively. The other classifiers from the best five all obtained better results than what was achieved in an earlier study by Mehrang et al [28]. Although our walking tests were based on the same mPower protocol, in our case, the physiotherapist controlled the sensors and calculated the steps instead of the subject themselves as in [28].…”

“…The other classifiers from the best five all obtained better results than what was achieved in an earlier study by Mehrang et al [28]. Although our walking tests were based on the same mPower protocol, in our case, the physiotherapist controlled the sensors and calculated the steps instead of the subject themselves as in [28]. Also, excluding the mild stages of the disease (UPDRS score 1 and some of 1.5) might have an effect on the results.…”

“…Features comparing the recorded signals, such as correlation between two axes, were calculated between ACC and GYRO signals on different axes (for example, ACC-X and GYRO-Y), or different axes of the same sensor (for example, ACC-X and ACC-Z). Similar features have been used in previous PD detection studies as well, for example by Mehrang et al [ 28 ] and Arora et al [ 6 ]. All the calculated features are listed in the S1 Table .…”

“…This implies that large datasets can be collected via smartphone with relatively small resources. The selection of the dataset depends on the research question, for example Nguyen et al [27] used the data of 6805 subjects with less balanced groups, whereas Mehrang et al [28] used only 1237 subjects of the same dataset but with age-matched PD patients and controls. They discriminated PD patients from healthy controls with RF receiving accuracy, sensitivity and specificity of 70% each.…”