Proceedings of The Physiological Society

Europhysiology 2018 (London, UK) (2018) Proc Physiol Soc 41, PCB023

Poster Communications

Derivation of heart rate variability (HRV) indicators in older pet cats from stethoscopes: A Deep Learning neural network (AI) approach

F. O'Brien1, N. Hughes1, N. Dowgray1, A. German1, R. Barrett-Jolley1

1. University of Liverpool, Liverpool, United Kingdom.

Stethoscope are still a frontline tool in first opinion practice and give the clinician useful diagnostic data. We believe that artificial intelligence (AI) could increase further the clinical potential of this technique. To investigate this we recorded feline heart rate (HR) with Bluetooth stethoscopes and analysed the data post hoc. Aims: To test (i) whether AI could be used to extract inter-beat intervals from stethoscope data and facilitate heart rate variability analysis (HRV) and (ii) whether age and/or gender influence short range HRV parameters of healthy cats. Methods: Data were collected by Bluetooth stethoscope. Several hundred clearly audible records (from 48 cats) were manually labelled with MIDI software creating labels for each S1 heart sound. Training data were then augmented by 3 methods; (i) scaling and inverting, (ii) superposition of random amplitude, phase and frequency noise (in the range 50Hz- 18kHz) and (iii) superposition of randomly scaled sets of ambient noise. This created, in effect, 10,000+ labelled training records. Training sufficiency was calculated as selectivity and sensitivity of the network to detect beats in a 20% random sample of labelled data kept separate from the training process itself. HRV analysis (Matlab) included SDNN/RR (SDNN normalised by respiration rate) and Lomb-Scargle high frequency, low frequency (LF) and very LF powers. We included 117 patients; 52 female, 65 male age range 7 to 10. Results: Initial training reached 92% sensitivity and 88% selectivity to detect previously unseen, but "labelled" S1 sounds. The overall correlation between expert (veterinarian) estimation of mean HR (by ear) and CNN predicted mean HR gave Spearman's Rho of 0.63, p<0.005. We found no statistically significant correlation between age and any of our selected HR or HRV parameters, but gender was significantly associated with both time domain (SDNN/RR) and frequency domain (LF) components of HRV (higher in female). Conclusion: The Deep Learning AI approach showed great promise; we found significant trends in HRV between genders, but not with age. In terms of age, there was a very limited age range included and we will in future extend this range by many years. Estimation of HR, by ear, when HR exceeds 180 bpm is extremely challenging and so the relatively weak correlation between AI and human does not necessarily imply inadequacies of our numerical methods. Whilst conducting this study we observed that the major weakness of these methods could be the underlying training data rather than the AI architecture itself. Close inspection of prediction outputs suggested that where the network makes "errors", subjectively, it appeared that such "errors" were in the labelling itself rather than the prediction. Full exploitation of this technology will require the generation of more accurate training sets.

Where applicable, experiments conform with Society ethical requirements