Sonus Health: Calibrated Heart-Murmur Detection from Smartphone-Based Veterinary Auscultation

Source: arXiv:2606.24885 · Published 2026-06-23 · By Aswin Jose, Roeland P-J E. Decorte, Laurent Locquet

TL;DR

This paper addresses the challenge of early detection of heart murmurs in companion animals, specifically dogs and cats, using smartphone-based auscultation recordings. Heart murmurs are subtle acoustic signs of heart disease that general-practice veterinarians often miss, while specialist assessment is costly and limited. The authors present Sonus Health, a platform that analyzes 30-second or longer smartphone audio recordings captured by pet owners or veterinary staff to detect murmurs with calibrated confidence tiers. The system defers uncertain cases to experts, prioritizing safe triage over forced classification.

Sonus Health combines two probabilistic classification models — a bilinear neural network analyzing frequency-domain signals sensitive to subtle murmur features, and a LightGBM model using summary acoustic features robust to noise. Calibrated outputs from both models are combined with an agreement rule producing three confidence tiers. Evaluated on 322 real-world smartphone auscultation clips (149 murmur-positive), the system achieves 74.8% overall accuracy and 0.842 AUROC under cross-validation. The high-confidence tier (covering roughly 30% of cases) reaches 95.9% accuracy, 94.0% sensitivity, and 97.9% specificity, closely approximating cardiologist-level performance on this subset. The platform’s design emphasizes calibrated uncertainty and deferral to veterinary review for ambiguous inputs, establishing a new, accessible cardiac screening tier in veterinary care.

Key findings

• On the full evaluation cohort (n=322), the combined Sonus Health system achieves 74.84% accuracy (95% CI [70.2, 79.5]) and AUROC 0.8415 using standard five-fold cross-validation.
• The bilinear network alone achieves 66.15% accuracy, 0.773 AUROC, 0.586 precision, and 0.913 recall, showing recall dominance but lower precision than LightGBM.
• Isotonic calibration applied to the bilinear network output raises AUROC from 0.773 to 0.792 without changing accuracy or F1, improving probability calibration.
• The tabular LightGBM model alone achieves 73.91% accuracy, 0.831 AUROC, 0.692 precision, outperforming the bilinear network on accuracy and precision.
• The combined agreement rule (requiring both models to agree on prediction) yields 74.84% accuracy, 0.842 AUROC, and 0.743 precision, improving precision over LightGBM alone.
• Under group-aware cross-validation (animals not shared between splits), accuracy is 74.53% and AUROC 0.8371, indicating limited data leakage impact.
• The high-confidence tier (30.1% of recordings where models strongly agree) achieves 95.88% accuracy with 94.0% sensitivity and 97.9% specificity, approaching expert-level performance.
• About 30% of recordings fall into the low-confidence tier (model disagreement) and are deferred to veterinary review, yielding only 30% sensitivity by design for deferral.

Threat model

The adversary is not an explicit focus; rather, the system is designed to operate under real-world noise and recording variability from lay users and veterinary staff capturing smartphone audio. It assumes no deliberate attacks on the recordings or model outputs. The key operational assumption is minimizing confidently incorrect automated classifications by deferring uncertain cases to experts.

Methodology — deep read

1. Threat Model & Assumptions: The system assumes an end-user (pet owner or veterinary professional) capturing auscultation recordings with a smartphone microphone. The adversary is not explicitly modeled since focus is on real-world noise and variability rather than active attacks. The platform assumes veterinary cardiologist labels as ground truth and aims to minimize confidently incorrect automated classifications by deferring uncertain cases.

2. Data: The evaluation cohort consists of 322 auscultation recordings collected under realistic conditions (home or clinic) using consumer smartphones. These recordings are approximately 30 seconds or longer, include environmental and handling noise, and cover both dogs and cats. Labels (murmur presence/absence) were assigned independently by veterinary cardiologists blinded to model predictions. 149 recordings are murmur-positive and 173 negative, drawn from 253 distinct animals (mostly one recording per animal). No additional noise filtering beyond routine preprocessing was applied.

3. Architecture / Algorithm: The classification pipeline has two probabilistic models analyzing each audio recording from complementary perspectives. The first model is a bilinear neural network operating on a frequency-domain representation to capture subtle spectro-temporal murmur patterns, especially distributed signals. The second is a LightGBM gradient boosting decision tree consuming tabular summary features that capture session-level acoustic and timing statistics, providing robustness to noise and typical morphologies. Outputs from these models are calibrated onto a common probability scale with isotonic regression applied to the bilinear network’s output. A deterministic agreement rule then assigns a confidence tier: "high confidence" when both models strongly agree, "moderate confidence" when both agree less confidently, and "low confidence" (disagreement) where classification is deferred.

4. Training Regime: The bilinear neural network is trained with standard supervised learning on the labeled data splits, though exact training details (loss, optimizer, epochs) are proprietary and not specified. LightGBM is trained on tabular features extracted from the same recordings. Calibration is fit on out-of-fold bilinear model predictions to avoid optimistic bias. Multiple random seeds (42, 1337, 2026, 7) are used to ensure result stability.

5. Evaluation Protocol: Performance was measured with stratified 5-fold cross-validation (standard and group-aware keyed by animal to prevent leakage) and a held-out 70/15/15 train/validation/test split (test n=46). Metrics reported include accuracy, AUROC, precision, recall (sensitivity), and F1 score, with 95% confidence intervals from bootstrap resampling. Ablations isolated each component’s contribution, and results were stratified by confidence buckets assigned prospectively by the model.

6. Reproducibility: The evaluation pipeline is fully deterministic and regenerates byte-for-byte outputs from a fixed seed. Feature extraction and calibration are deterministic, and the full model artifacts are compact (<4MB). However, dataset and some system details are proprietary and not publicly released, limiting external replication from public data.

Example End-to-End: For a given 30-second smartphone auscultation audio recording captured by a pet owner, the system first preprocesses the audio and extracts frequency-domain spectrogram-like features for input to the bilinear neural network. Simultaneously, summary statistics of the audio signal (e.g., heart rate-related features, timing metrics) are computed for the LightGBM model. The bilinear network outputs a calibrated probability after isotonic regression; the LightGBM outputs another probability. If these two calibrated probabilities both surpass a high-confidence threshold and indicate murmur presence (or absence), the system assigns the recording to the high-confidence tier and returns that classification directly. If the two disagree, the system assigns a low-confidence tier and defers to veterinary review. Otherwise, a moderate-confidence tier classification with a caveat is returned. This triage logic ensures that only high-confidence calls are automated, reducing false positives and negatives in deployment.

Technical innovations

• Combination of two probabilistic models analyzing complementary audio representations (frequency-domain bilinear net plus tabular LightGBM), leveraging their disjoint failure modes.
• Application of a rank-preserving isotonic calibration on the bilinear network output to align probability scales before combination.
• A deterministic agreement-based confidence tiering system that defers ambiguous cases to expert review, trading recall for high precision and clinical safety.
• Use of smartphone microphone data captured by untrained users in realistic home and clinical environments, rather than relying on specialized electronic stethoscopes.

Datasets

• Sonus Health veterinary auscultation recordings — 322 audio files (149 murmur-positive, 173 negative) — proprietary collection from real-world smartphone captures in home and clinical settings

Baselines vs proposed

• Bilinear network (raw): Accuracy = 66.15%, AUROC = 0.773, Precision = 0.586, Recall = 0.913
• Bilinear network + isotonic calibration: Accuracy = 66.15%, AUROC = 0.792, Precision = 0.586
• LightGBM model: Accuracy = 73.91%, AUROC = 0.831, Precision = 0.692
• Combined agreement rule: Accuracy = 74.84%, AUROC = 0.842, Precision = 0.743
• Group-aware cross-validation combined rule: Accuracy = 74.53%, AUROC = 0.837
• High-confidence bucket (30% of data): Accuracy = 95.88%, Sensitivity = 94.0%, Specificity = 97.9%

Limitations

• Relatively modest dataset size (322 recordings), limiting statistical power and generalization assessment.
• Held-out test set (n=46) is small, resulting in wide confidence intervals and limited definitive external validation.
• Labels rely on veterinary cardiologists but no direct comparison with echocardiographic gold standard reported.
• Evaluation is retrospective; prospective multi-site, multi-device validation remains to be done.
• Proprietary calibration parameters and model architecture details limit full reproducibility and external scrutiny.
• System robustness to diverse consumer hardware, varied capture techniques by untrained owners, and longitudinal label drift are not yet studied.

Open questions / follow-ons

• How does the platform perform prospectively across diverse geographic locations, clinical sites, and smartphone hardware variants?
• What is the performance impact and generalizability across breeds, sizes, ages, and health statuses of dogs and cats beyond the current cohort?
• Can the deferral mechanism be further optimized to reduce the proportion of recordings routed to veterinary review while maintaining high confidence?
• How do longitudinal recordings of individual animals track murmur progression, and can the platform support early intervention strategies longitudinally?

Why it matters for bot defense

This study exemplifies the critical importance of calibrated uncertainty and tiered decision-making in automated classification systems—principles that translate well to bot defense and CAPTCHA engineering. The dual-model approach leveraging complementary feature representations to reduce false positives and negatives, combined with an explicit defer-to-human tier, offers a practical blueprint for balancing automation with risk management in security-critical applications.

Practitioners designing bot detection or CAPTCHA filters can draw from this design to implement confidence calibration and disagreement-based deferrals, thereby improving trustworthiness and reducing false positives that degrade user experience. It also highlights challenges of deploying ML systems in noisy, real-world environments with untrained users and diverse hardware, underscoring the value of rigorous cross-validation protocols and prospective confidence stratification.

Cite

@article{arxiv2606_24885,
  title={ Sonus Health: Calibrated Heart-Murmur Detection from Smartphone-Based Veterinary Auscultation },
  author={ Aswin Jose and Roeland P-J E. Decorte and Laurent Locquet },
  journal={arXiv preprint arXiv:2606.24885},
  year={ 2026 },
  url={https://arxiv.org/abs/2606.24885}
}

Read the full paper

• arXiv:2606.24885 (abstract)
• arXiv:2606.24885 (PDF)