Causality-inspired AI removes stethoscope bias in respiratory diagnosis

A team of researchers led by Heejoon Koo has introduced a novel approach to make AI-driven respiratory sound classification (RSC) robust across different stethoscope brands. The problem: each stethoscope imparts a unique acoustic signature, creating unintended shortcuts that models learn, causing them to fail on unseen devices. The paper, published on arXiv, proposes a causality-inspired multimodal federated domain generalization (FedDG) framework. It combines three key interventions: a device style intervention network that perturbs audio in a content-preserving way, counterfactual text augmentation to neutralize metadata like stethoscope model, and gradient alignment to enforce device-invariant representations across hospitals. By building on a multimodal language-audio pretraining model, the system learns to focus on disease-specific sounds rather than device artifacts. The authors validate their method on two public datasets, ICBHI and SPRSound, using a leave-one-device-out protocol—simulating real-world conditions where a model trained on data from certain stethoscopes must perform on an unseen one. Their framework consistently outperforms conventional data augmentation and federated learning baselines, with relative improvements in F1-score of roughly 10-15% in the most challenging cross-device scenarios.

This work addresses a critical bottleneck in deploying AI for respiratory diagnostics across hospitals: each site uses different stethoscopes, and models trained on one device fail on another. Federated learning alone does not solve this because device-specific shortcuts persist. The causality-inspired approach reframes the problem as a confounding variable—the device—and uses interventions akin to randomized experiments to break spurious correlations. The counterfactual text augmentation is particularly clever: it generates alternative textual descriptions (e.g., “stethoscope model: unknown”) during training to reduce reliance on metadata shortcuts. The framework is privacy-preserving by design, as it only shares gradient updates, not raw data. This makes it suitable for multi-institutional collaborations without violating data regulations. The authors plan to release the code upon publication, which could accelerate adoption in clinical AI pipelines. For healthcare AI teams, this paper offers a practical blueprint for building generalizable models that work across different hardware, potentially enabling nationwide or global deployment of automated pulmonary disease screening tools.