APRIL: A New Replay-Free Incremental Learning Framework for Respiratory Sound Classification

Title

APRIL: A New Replay-Free Incremental Learning Framework for Respiratory Sound Classification

Meta Description

APRIL proposes replay-free domain incremental learning, solving domain shift and catastrophic forgetting in respiratory sound classification via low-rank expert isolation and ambiguity-aware routing, achieving less than 1% forgetting on ICBHI and SPRSound.

Article Content

The long-term deployment of respiratory sound classification faces domain shift challenges—device updates and changes in the recording environment cause data distribution drift, while privacy regulations prohibit replaying historical data. Traditional continual learning easily forgets old knowledge, and multi-expert systems avoid parameter interference but struggle to select the correct expert at inference time. To address this dilemma, researchers propose the APRIL framework, which shares a frozen BEATs backbone, sequentially inherits and isolates low-rank experts, and adapts to new domains without altering the old models. Its multi-level prototype router retrieves the top-two candidate experts, fuses the domain routing posterior with classification confidence, and selects the final model via adaptive weighting based on matching ambiguity. On device-based and time-based incremental benchmarks built from the ICBHI and SPRSound datasets, under both forward and reverse domain orders, APRIL achieves the highest final-session average scores across five random seeds: 64.4% and 65.4% on ICBHI, and 83.1% and 82.2% on SPRSound, with an average forgetting rate below 1.0%. This work provides a reliable paradigm for continuous respiratory sound classification under privacy constraints.

The Continual Learning Dilemma in Respiratory Sound Classification

Automated respiratory sound classification enables continuous, non-invasive respiratory assessment, but long-term deployment faces domain shift. As recording devices are updated and acquisition periods change, the acoustic distribution drifts and model performance degrades. More challenging still, privacy and governance regulations often prohibit replaying historical data, making sequential learning prone to overwriting prior knowledge. Retaining dedicated domain experts prevents parameter interference, but at inference time it is difficult to select the correct expert when acoustic distributions overlap and no domain labels are available.

Domain Shift refers to the phenomenon where the input data distribution encountered after model deployment differs from the training data distribution. In respiratory sound classification, device replacement (e.g., different microphone sensitivity and frequency response) and changes in the acquisition environment (background noise, patient posture) both cause drift in acoustic feature distributions, significantly degrading model generalization.

Sample-Free Continual Learning is a continual learning setting in which replaying historical data is not allowed. Because privacy regulations (such as HIPAA and GDPR) restrict the storage and reuse of medical audio, models can only be updated sequentially on new-domain data while retaining old-domain knowledge, which greatly increases the risk of catastrophic forgetting.

The APRIL Framework: Expert Isolation and Ambiguity-Aware Routing

APRIL is a sample-free domain incremental learning framework. Built on a shared frozen BEATs backbone, it sequentially inherits and isolates low-rank experts so that new-domain adaptation does not interfere with old models. At its core is a multi-level prototype router: it first retrieves the top-two candidate experts, then fuses the domain routing posterior with classification confidence, and dynamically weights the selection of the final model based on matching ambiguity. This design enables adaptive decisions when experts overlap, avoiding the errors caused by hard switching.

BEATs (Bidirectional Encoder representation from Audio Transformers) is a pretrained audio Transformer backbone that acquires general acoustic representations through self-supervised learning on large-scale audio data. APRIL freezes it as a shared feature extractor, avoiding representation drift caused by fine-tuning while greatly reducing trainable parameters, allowing low-rank expert modules to adapt lightly to different domains.

Low-Rank Expert refers to a lightweight adaptation module constructed via low-rank decomposition (such as LoRA). Each expert contains only a small number of trainable parameters, inserted into specific layers of the frozen backbone to learn domain-specific residual adjustments. During sequential inheritance, new experts are initialized independently while old experts remain frozen, thereby isolating knowledge of different domains at the parameter level and avoiding mutual interference.

Benchmarks and Performance

Researchers built standardized device-incremental and time-incremental benchmarks on the ICBHI and SPRSound datasets, covering both forward and reverse domain orders. Evaluated across five random seeds, APRIL consistently leads the compared continual learning methods in final-session average score: 64.4% forward and 65.4% reverse on ICBHI; 83.1% forward and 82.2% reverse on SPRSound. More critically, the average forgetting rate is below 1.0% under all protocols, demonstrating robust domain adaptation and long-term knowledge retention without data replay.

Final Session Average Score is a commonly used evaluation metric in domain incremental learning, referring to the average of the model's final performance across all test domains, reflecting the model's overall retention of learned domains. Average Forgetting measures the degree to which the model's performance on old domains declines relative to its historical best after learning a new domain; the lower the value, the less forgetting. Together they provide a comprehensive assessment of the model's stability and plasticity in sequential learning.

Privacy Protection and Clinical Deployment Significance

APRIL's replay-free nature aligns with the privacy requirements of medical scenarios, its expert isolation mechanism avoids parameter interference, and ambiguity-aware routing improves inference reliability. This offers a feasible path for deploying continuous respiratory sound classification in real clinical environments, especially in scenarios with frequent device iteration and non-retrievable data. Future work may explore more modality fusion and dynamic expert expansion.

One-line Summary

By freezing the BEATs backbone and isolating low-rank experts, APRIL enables privacy-constrained, replay-free domain incremental learning for respiratory sounds.

Paragraph Summary

To address domain shift and privacy-prohibited replay in long-term deployment of respiratory sound classification, researchers propose the APRIL framework. Built on a shared frozen BEATs backbone, it sequentially inherits and isolates low-rank experts, preventing new-domain adaptation from interfering with old models. Its multi-level prototype router retrieves the top-two candidate experts, fuses the domain routing posterior with classification confidence, and adaptively weights the final model selection based on matching ambiguity. On device-based and time-based incremental benchmarks built from the ICBHI and SPRSound datasets, under both forward and reverse domain orders, APRIL achieves the highest final-session average scores across five random seeds (64.4% and 65.4% on ICBHI, 83.1% and 82.2% on SPRSound), with an average forgetting rate below 1.0%. This work provides a reliable paradigm for continuous respiratory sound classification under privacy constraints.