Whitepaper

Data-Efficient Cough Analysis Through Active Learning and Segmentation

Virufy’s Approach to Robust, Trustworthy, and Actionable Cough Analysis

Executive Summary

Virufy’s approach to cough-based disease screening is intentionally conservative, transparent, and clinically grounded, with an explicit emphasis on the efficient and responsible use of clinical audio data. Rather than relying on opaque architectures or increasing model complexity, we employ well-established signal representations, compact convolutional neural networks, and structured active learning workflows designed to maximize informational yield from large, heterogeneous datasets.

Active learning is used to prioritize samples that are most informative for model improvement, including those near decision boundaries and high-confidence predictions selected for auditing. By allocating annotation and modeling capacity toward these targeted examples—rather than treating all data as equally informative—the training process emphasizes clinically relevant variability and failure modes. This strategy supports generalization across datasets, facilitates early identification of systematic errors, and promotes stability under evolving data distributions, including variability in recording conditions and population characteristics.

Cough Segmentation as a First-Class Problem

Accurate cough segmentation is a prerequisite for reliable downstream disease classification. Errors at the segmentation stage propagate through subsequent modeling steps, leading to instability in classification performance irrespective of model complexity. Accordingly, cough segmentation is treated as a primary modeling objective rather than a preprocessing step.

Cough events are identified using a sliding convolutional neural network applied to log-power spectrogram representations. This architecture preserves salient time–frequency characteristics of cough acoustics while maintaining robustness to variability in recording conditions. Even in the absence of task-specific inductive bias, this combination provides strong baseline performance across respiratory sound detection tasks.

The distinguishing aspect of this approach lies not in architectural novelty, but in the training and validation strategy. Model development emphasizes disciplined data selection, targeted annotation, and systematic evaluation to ensure that learned representations reflect clinically relevant acoustic structure rather than dataset-specific artifacts.

Active Learning for Clinical Robustness

Clinical audio datasets exhibit substantial heterogeneity arising from differences in recording devices, acoustic environments, patient behavior, and disease presentation. Many of these sources of variability are difficult to characterize a priori, and uniform random sampling is insufficient to reliably expose models to clinically relevant edge cases.

The model is used to identify subsets of data expected to provide the greatest informational value, including:

• Samples near decision boundaries, which are most informative for refining model discrimination.
• High-confidence predictions selected for auditing, enabling detection of silent failure modes and verification of model certainty.
• Dataset-specific artifacts, which may introduce bias or spurious correlations if left unexamined.

These targeted samples are prioritized for annotation and review to ensure that training emphasizes clinically relevant variability and failure-prone cases.

By iteratively incorporating feedback from these selected examples, the segmentation model is trained on cases that are both clinically informative and prone to failure. This process supports verification of confident predictions, reduces the risk of silent errors, and promotes consistent performance across diverse clinical cough datasets.

Disease-Relevant Feature Learning

Once cough events are reliably segmented, a second sliding convolutional neural network is applied directly to the extracted cough segments. This model is trained to learn acoustic features associated with disease-relevant cough physiology rather than incidental or context-dependent correlates.

High segmentation fidelity ensures that the classifier operates on well-isolated cough events, reducing sensitivity to background noise and recording artifacts. As a result, learned representations consistently reflect physiologically meaningful cough acoustics, supporting stable classification performance across datasets with varying recording conditions and population characteristics.

Internal Performance Highlights

While dataset sizes remain intentionally modest and clinically curated, internal evaluation demonstrates stable performance under five-fold cross-validation. Performance metrics are summarized below:

These results reflect not only model capacity, but the combined effects of disciplined data curation, high-fidelity cough segmentation, and active learning-driven validation. Together, these factors contribute to stable performance despite limited dataset size and heterogeneous recording conditions.

Why This Matters

Our development philosophy emphasizes trustworthiness and clinical reliability over architectural novelty. By combining interpretable model designs with structured active learning and systematic clinical validation, we construct systems that expose failure modes early, support incremental improvement, and maintain stable performance across heterogeneous datasets.

In clinical artificial intelligence, robustness is not a direct consequence of model complexity. Instead, it emerges from careful problem formulation, targeted data selection, and continuous verification throughout model development. This approach underpins the observed reliability of our models in clinically relevant settings.

References

1. Q. Zhou, J. Yan, Y. Zhang, and Y. Wang, “Cough recognition based on mel-spectrogram and convolutional neural network,” Frontiers in Robotics and AI, vol. 8, 2021.
2. T. Keles, A. Erdem, and M. C. Yildiz, “Spectrogram-based deep learning approach for anomaly detection from cough sounds,” International Journal of Information Technology and Computer Science, vol. 3, pp. 1–12, 2025.
3. B. Settles, “Active learning literature survey,” University of Wisconsin–Madison, Computer Sciences Technical Report 1648, 2010.
4. A. Kendall and Y. Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?” in Advances in Neural Information Processing Systems (NeurIPS), 2017.
5. S. Lapuschkin, W. Samek, K.-R. Müller, and G. Montavon, “Unmasking clever Hans predictors and assessing what machines really learn,” Nature Communications, vol. 10, 2019.
6. N. Rasmussen, R. Rizk, L. Wang, and K. C. Santosh, “Ecologically valid benchmarking and adaptive attention: Scalable marine bioacoustic monitoring,” arXiv preprint arXiv:2509.04682, 2025.
7. N. Rasmussen, R. Rizk, L. Wang, and K. C. Santosh, “Channel-selected stratified nested cross-validation for clinically relevant EEG-based Parkinson’s disease detection,” arXiv preprint arXiv:2601.05276, 2026.