Calibrated Heart-Murmur Detection from Smartphone-Based Veterinary Auscultation
Aswin Jose, Dr. Roeland P-J E. Decorte, Laurent Locquet*
*Decorte Future Industries Ltd / Sonus Health
A technical account of how Sonus Health screens canine and feline heart sounds — a calibrated, two-model pipeline that returns a confident result on the cases it can stand behind, and routes the rest to a vet.
- 95.9%
- Accuracy, high-confidence tier
- 94.0%
- Sensitivity
- 97.9%
- Specificity
Measured on the high-confidence tier (30% of recordings, n = 97) under out-of-fold cross-validation. The platform defers the cases it cannot confidently call.
Read the whitepaper
Enter your email and we'll send the PDF straight to your inbox.
Abstract
Abstract — Heart disease is among the most common serious conditions in dogs and cats, and a murmur heard on auscultation is one of its earliest signs. Sonus Health is a smartphone-based screening system that analyses an auscultation recording of roughly thirty seconds — captured by an owner at home or by a vet in clinic — and returns a tiered result within moments. Evaluated on 322 veterinary-labelled recordings under standard out-of-fold cross-validation, the high-confidence tier (30% of cases) reaches 95.9% accuracy, with 94.0% sensitivity and 97.9% specificity. Uncertain cases are prospectively routed to veterinary review rather than guessed at. Results are stable across cross-validation protocols, a held-out test split, and multiple random seeds.
Calibrated confidence
It commits where it's sure, and defers the rest.
Every recording is sorted into one of three confidence tiers at the moment of analysis, from the agreement of two independently calibrated models. The platform only returns a call on the cases it can stand behind.
95.9%
Confident result, returned to the owner
70.3%
Returned with a caveat to confirm clinically
Deferred
Routed to a vet for review
Inside the paper
How the result is evaluated
- 322 real-world recordings
- Dogs and cats, captured on consumer smartphones at home and in clinic, each labelled through a layered veterinary review — no idealised lab re-captures.
- Validated four ways
- Standard and group-aware (animal-level) cross-validation, a held-out test split, and a seed-stability analysis, all reported with bootstrap confidence intervals.
- A calibrated two-model pipeline
- A frequency-domain model sensitive to subtle murmur structure, combined with a summary-statistic model robust to noise, placed on a common probability scale before they vote.
- Benchmarked against the literature
- A qualitative side-by-side with published canine and feline screening systems, with the differences in species, device, and labelling spelled out.
Read the full evaluation
Methods, results, tables, and references — sent to your inbox as a PDF.