Automated acoustic detection of mouse scratching

Abstract

Itch is an aversive somatic sense that elicits the desire to scratch. In animal models of itch, scratching behavior is frequently used as a proxy for itch, and this behavior is typically assessed through visual quantification. However, manual scoring of videos has numerous limitations, underscoring the need for an automated approach. Here, we propose a novel automated method for acoustic detection of mouse scratching. Using this approach, we show that chloroquine-induced scratching behavior in C57BL/6 mice can be quantified with reasonable accuracy (85% sensitivity, 75% positive predictive value). This report is the first method to apply supervised learning techniques to automate acoustic scratch detection.

Fig 1. A 700 ms recorded sound wave including a scratch bout.

The peaks correspond to individual scratches. This bout lasts approximately 400 ms.

Fig 2. A spectrogram showing the short-time Fourier transform of the scratch bout from Fig 1.

Vertical bands in the graph correspond to individual scratches. These bands are more distinct above 10 kHz.

Fig 3. Spectrograms with the filtered signal.

(A,B) Spectrograms of a scratch bout recorded on two channels. (C,D) Corresponding candidate islands. We observe variation between peak heights and number of peaks. The red dots indicate locations that were marked as peaks using the method described in Sec 3.1.2.

Fig 4. Two local maxima in the filtered signal.

The top of each dashed box corresponds to the height of the local maximum, and the vertical dashed lines show the 25 ms band around the local maximum. The bottom of the dashed box shows the height of the local maximum minus the threshold h. On the right, the minimum in the interval is much smaller than the threshold, so the local maximum is considered a peak. On the left, the minimum in the interval is above the dashed line, so that local maximum is not considered a peak.

Fig 5. Classification accuracy for each recording.

We compare the tradeoff between sensitivity and false discovery rate for out-of-bag predictions for the first 5 recordings. We also include results for recording 6 when using recording 5 to train the random forest. Performance is best for recordings 1, 2, 4 and 5. The neighborhood adjustment method (Sec 3.2.2) gives an improvement over the unadjusted cross prediction.

Fig 6. Performance response to class balance.

We compare the results from recordings 3 and 4 to results for stratified subsamples of candidate islands from recording 4. The subsamples draw separately from true scratch bouts and false candidate islands to match the class balance for recording 3. The tradeoffs for recordings 3 and 4 are plotted alongside an interval covering 90% of 30 stratified subsamples from recording 4. The performance under subsampling is comparable to that of recording 3.

Fig 7. Performance response to training sample size.

We compare the results from recordings 3 and 4 to results for subsamples of candidate islands from recording 4. The FDR-sensitivity tradeoffs for recordings 3 and 4 are plotted alongside the average tradeoffs across 10 subsamples from recording 4 of size 2,000 and 500.

Fig 8. Actual vs predicted scratching rates over time.

We compare the true rates of scratching over time estimated using kernel smoothing to rates based on out-of-bag predictions from our method. We see that important trends are picked up, e.g. the decline in scratching over time in recording 4 and the three large groups of scratching in recording 3.