A Smart Tendon Hammer System for Remote Neurological Examination

The deep tendon reflex exam is an important part of neurological assessment of patients consisting of two components, reflex elicitation and reflex grading. While this exam has traditionally been performed in person, with trained clinicians both eliciting and grading the reflex, this work seeks to enable the exam by novices. The COVID-19 pandemic has motivated greater utilization of telemedicine and other remote healthcare delivery tools. A smart tendon hammer capable of streaming acceleration measurements wirelessly allows differentiation of correct and incorrect tapping locations with 91.5% accuracy to provide feedback to users about the appropriateness of stimulation, enabling reflex elicitation by laypeople, while survey results demonstrate that novices are reasonably able to grade reflex responses. Novice reflex grading demonstrates adequate performance with a mean error of 0.2 points on a five point scale. This work shows that by assisting in the reflex elicitation component of the reflex exam via a smart hammer and feedback application, novices should be able to complete the reflex exam remotely, filling a critical gap in neurological care during the COVID-19 pandemic.

The hammer is manufactured by removing material from the silicone hammer head and inserting the plastic case containing the IMU. This material removal can be done by conventional milling, or through a stamping process with a thin metal tube of the appropriate diameter. Because of the elastic nature of the silicone, dimensional tolerances during machining are unimportant.

Steel handle IMU Plastic case
Mounting screw Silicone head LED indicator Figure S1. Picture of the smart tendon hammer used throughout this work.  Figure S2. Tapping feedback and data collection application, A) Result after successful tap, B) Result after an incorrect tap, the red and green lines represent the acceleration of the most recent incorrect and correct taps respectively.

ACCELERATION MEASUREMENTS
Acceleration measurements in this work were made using the IMU shown in Figure S1. The sensor can be used to log data at up to 800Hz, or stream data directly from the sensor at up to 200Hz. Logged data must later be transmitted to the mobile device for offline processing, while streamed data can be processed online, as it is collected. Regardless of sample rate, data has a similar structure to the representative plots shown in Figure S3. The largest peak corresponds to the deceleration of the hammer during impact. The classification and frequency spectra results both come from these type of acceleration measurements. Tapping locations for the human Achilles experiments are shown in Figure S4. The same two locations were used in both the automated and manual tapping experiments. Results from the tapping experiments on human Achilles tendons are shown in Figure S5 and Figure S6. The frequency spectrum in Figure S5 provides a justification for the use of a 200 Hz accelerometer, as the majority of the signal power is in the lower frequency region. Additionally, Figure S6 shows the results of a χ 2 test comparing each 100 sample group for each frequency. Even with a conservatively corrected α, there are significant differences between the two locations at 4 frequencies, all below 100 Hz.

TAPPING VARIABILITY
Tapping variability can be quantified in a number of ways. This paper is concerned only with measuring variability in the intensity of stimulation, as the reported classifier tracks location variability where it matters. In comparing tapping variability, the results in Figure S7 show comparable distributions between expert and novice accelerations, both before and during impact with the rubber tendon analog.

SVM BASED CLASSIFICATION
An example of the feature vector used for the SVM based classifier is shown in Figure S8. The feature vector consists of 51 acceleration points, 2 prior to peak deceleration, and 49 after the peak deceleration. Prior to training each of the features is standardized, other than this, the feature vectors consist of the raw acceleration data, without filtering or other feature extraction. Each tap collected from the human subjects corresponds with a feature vector used for training or testing. A total of 800 feature vectors were used, with 100 held out for testing each of the 8 classifiers. Standardized acceleration Figure S8. Representative feature vector used as training data for the SVM classifier.
An example of the survey questions presented to research participants is shown in Figure S9. Both the training video with labeled responses and each of the unlabelled responses are shown in the supplementary video file. The survey consisted of a training video, in which subjects watched 3 labelled videos for each of the 5 scores, and a response section, in which they scored 25 unlabelled videos. Each survey question provided the 0-4 scale with descriptions from the NINDS scale. Figure S9. Example survey question with obscured hammer impact and 0-4 grading choices. Frontiers