Implementation of an iPod wireless accelerometer application using machine learning to classify disparity of hemiplegic and healthy patellar tendon reflex pair

Abstract

The characteristics of the patellar tendon reflex provide fundamental insight regarding the diagnosis of neurological status. Based on the features of the tendon reflex response, a clinician may establish preliminary perspective regarding the global condition of the nervous system. Current techniques for quantifying the observations of the reflex response involve the application of ordinal scales, requiring the expertise of a highly skilled clinician. However, the reliability of the ordinal scale approach is debatable. Highly skilled clinicians have even disputed the presence of asymmetric reflex pairs. An alternative strategy was the implementation of an iPod wireless accelerometer application to quantify the reflex response acceleration waveform. An application enabled the recording of the acceleration waveform and later wireless transmission as an email attachment by connectivity to the Internet. A potential energy impact pendulum enabled the patellar tendon reflex to be evoked in a predetermined and targeted manner. Three feature categories of the reflex response acceleration waveform (global parameters, temporal organization, and spectral features) were incorporated into machine learning to distinguish a subject's hemiplegic and healthy reflex pair. Machine learning attained perfect classification of the hemiplegic and healthy reflex pair. The research findings implicate the promise of machine learning for providing increased diagnostic acuity regarding the acceleration waveform of the tendon reflex response.

Figure 1

iPod wireless quantified reflex system. (A) Impact pendulum attached to a reflex hammer. (B) Precise targeting of the reflex hammer and iPod mounted to the lateral malleolus.

Figure 2

Average reflex waveform for the hemiplegic and unaffected leg. Waveforms were aligned so that the first local peak coincided in time. Line thickness reflects the standard error from the mean (SEM).

Figure 3

Univariate differences in spectral features between the hemiplegic and unaffected leg. (A) Sum of spectral energy across 0-50 Hz. (B) Difference in spectral energy at specified frequencies. Error bars reflect SEM.

Figure 4

Univariate differences in global features between the hemiplegic and unaffected leg. The maximum acceleration (A) or change in acceleration (B) observed in the waveform. The average auto correlation of acceleration (C) or change in acceleration (D). Error bars reflect SEM.

Figure 5

Univariate differences in local features between the hemiplegic and unaffected leg. (A) The magnitude of the local peaks in acceleration. (B) The temporal delay between the local peaks in acceleration. Error bars reflect SEM.

Figure 6

Cross-validation accuracy of the SVM when using global features, spectral features, local features or all available features to discriminate between the hemiplegic and unaffected leg. Error bars reflect SEM.