0482 Novel Breath Detection Algorithm to Assess Sleep-Disordered Breathing in a Mouse Model

Abstract Introduction Mouse sleep studies are commonly used to observe the effects of trial drugs on sleep and breathing. However, the current manner of detecting breaths is manual, laborious, and often inconsistent, especially over long periods of recordings. To address these limitations, an automated breath detection algorithm was developed. Methods Existing gold-standard manually detected and analyzed samples of respiratory recordings from adult male C57BL/6J mice were reanalyzed during NREM (n=16) and REM (n=13) sleep episodes with the algorithm. The algorithm used two third-order Savtizky-Golay filters to remove artifact noise, and calculated baselines using a second-order Butterworth low-pass filter with a cutoff frequency responsive to sample length and normalized by flow signal sampling rate. Flow signal zero-crossings were identified as inspiration starts and ends. Instant respiratory rate (RR), maximal inspiratory flow (Vimax), and instant minute ventilation (VE) were calculated and averaged using both manually and automatically marked NREM breaths, REM breaths, and all breaths. Bland-Altman plots were constructed between manual and automatic breathing metrics to determine bias and limits of agreement (LoA) between automatic and manual detection. Results The derived data from the algorithm showed high accuracy compared to manual scoring for VE in all, NREM, and REM breaths: 0.67% bias with [-1%, 2.3%] LoA, 0.26% bias with [-2.1%, 2.6%] LoA, and 0.74% bias with [-1%, 2.5%] LoA, respectively. Similar accuracy was achieved for RR: 2.1% bias with [-1%, 5.2%] LoA, 0.25% bias with [-1.2%, 1.7%] LoA, and 5.9% bias with [-0.65%, 5.3%] LoA, respectively. In contrast, there was a positive bias for Vimax: 7.1% bias with [2.1%, 12.1%] LoA, 4.2% bias with [0.99%, 7.4%] LoA, and 8.2% bias with [3.6%, 12.8%] LoA, respectively. Subsequent analysis revealed that manual scoring failed to account for baseline drift, leading to later inspiration onsets and Vimax discrepancy. For an average of 15.3 minutes of signal per study, the algorithm detected breaths in an average of 0.22 seconds. Conclusion The automatic breath detection algorithm showed excellent performance compared to manual analysis, significantly simplifying and expediting sleep study analysis. Support (if any) R01HL133100, R01HL174409