APRIL: Ambiguity-Aware Expert Routing for Respiratory Sound Domain-Incremental Learning

Xinhong Li; Weiping Tu

cs.SDOct 20, 2026

APRIL: Ambiguity-Aware Expert Routing for Respiratory Sound Domain-Incremental Learning

Authors: Xinhong Li, Weiping Tu

Organizations: School of Computer Science, Wuhan University, Wuhan, Hubei, China

Abstract

Automated respiratory sound classification enables continuous, non-invasive respiratory assessment, yet long-term deployment remains challenging due to domain shifts from evolving recording devices and acquisition periods. When privacy and governance regulations prohibit replaying historical data, sequential learning easily overwrites previously acquired knowledge. Preserving dedicated domain experts prevents parameter interference, but selecting the correct expert at inference is difficult under overlapping acoustic distributions and absent domain labels. To resolve this dilemma, we propose APRIL, an exemplar-free domain-incremental framework for RSC. Built on a shared frozen BEATs backbone, APRIL sequentially inherits and isolates low-rank experts to enable new-domain adaptation without altering earlier models. A multi-level prototype router retrieves the top-two candidate experts and dynamically selects the final model by integrating domain-routing posteriors with classification confidence, adaptively weighted by matching ambiguity. We establish standardized device- and temporal-incremental benchmarks on the ICBHI and SPRSound datasets across forward and reversed domain orders. Evaluated over five random seeds, APRIL consistently achieves the highest final-session average Score among compared continual learning approaches, reaching 64.4% and 65.4% on ICBHI, and 83.1% and 82.2% on SPRSound under forward and reversed sequences, respectively. Crucially, average forgetting remains under 1.0% across all protocols, demonstrating robust domain adaptation and long-term knowledge retention without data replay. Overall, APRIL offers a reliable and privacy-preserving paradigm for deploying continuous respiratory sound classification across real-world healthcare settings.

Figures & tables

Panel of Figure 1 — Figure 1: Catastrophic forgetting under sequential fine-tuning. Each curve reports the Score on a previously learned domain after successive domain updates. Lines and shaded regions denote the mean and standard deviation over five seeds, respectively. $\Delta$ denotes the change between the stage at which a domain is first learned and the final stage.

Figure 2: Overview of APRIL. Each new domain expert inherits the preceding expert’s LoRA adapter and classification head, while the shared backbone and earlier experts remain frozen. During inference, multi-level prototype matching retrieves two candidate experts, and routing posterior and classifier confidence jointly determine which candidate’s prediction is returned.

Table 1: Summary of the domain-incremental protocols. Each entry is reported as training/test count. Patient totals count unique subjects, and the global training and test patient sets are disjoint.

Method	Year	Forward order				Reversed order
Method	Year	FSA ( $\uparrow$ )	AF ( $\downarrow$ )	BWT ( $\uparrow$ )	OS ( $\uparrow$ )	FSA ( $\uparrow$ )	AF ( $\downarrow$ )	BWT ( $\uparrow$ )	OS ( $\uparrow$ )
LwF	2016	62.58 $\pm$ 2.13	6.97 $\pm$ 3.69	-6.36 $\pm$ 3.97	59.98 $\pm$ 1.60	65.19 $\pm$ 0.91	2.60 $\pm$ 1.43	-0.30 $\pm$ 2.50	60.21 $\pm$ 0.54
EWC	2017	55.63 $\pm$ 1.54	14.20 $\pm$ 2.25	-14.20 $\pm$ 2.25	57.04 $\pm$ 1.11	64.79 $\pm$ 1.56	4.53 $\pm$ 2.26	-3.69 $\pm$ 3.47	59.43 $\pm$ 0.41
ER	2019	58.96 $\pm$ 2.31	11.16 $\pm$ 2.60	-11.09 $\pm$ 2.49	60.18 $\pm$ 1.06	64.41 $\pm$ 1.18	3.25 $\pm$ 1.53	-2.71 $\pm$ 1.49	61.02 $\pm$ 0.90
DER++	2020	56.68 $\pm$ 2.12	12.60 $\pm$ 2.20	-12.22 $\pm$ 2.68	58.69 $\pm$ 0.61	64.19 $\pm$ 1.89	2.81 $\pm$ 1.69	-2.21 $\pm$ 1.98	60.72 $\pm$ 1.91
InfLoRA	2024	59.39 $\pm$ 2.41	6.12 $\pm$ 4.77	-5.55 $\pm$ 4.46	57.23 $\pm$ 1.70	62.20 $\pm$ 1.49	0.89 $\pm$ 0.58	1.74 $\pm$ 2.43	58.36 $\pm$ 1.71

Table 2: Performance comparison on ICBHI under forward and reversed device orders. Results are mean

\pm

standard deviation over five seeds. Bold and underlined means indicate the best and second-best results among continual-learning methods, respectively. Seq. FT and Joint Train. denote sequential fine-tuning and joint training, respectively.

Figure 3: Forgetting dynamics on ICBHI and SPRSound. Columns (a)–(d) show ICBHI forward/reverse and SPRSound forward/reverse, respectively. Rows report stage-wise AF, BWT, and final domain-wise changes

\Delta\mathrm{Score}_{j}=R_{T-1,j}-R_{j,j}

. Values are five-seed means; shading and error bars show one standard deviation for APRIL.

Method	Forward order				Reversed order
Method	FSA ( $\uparrow$ )	AF ( $\downarrow$ )	BWT ( $\uparrow$ )	OS ( $\uparrow$ )	FSA ( $\uparrow$ )	AF ( $\downarrow$ )	BWT ( $\uparrow$ )	OS ( $\uparrow$ )
ICBHI
w/o isolation	53.74 $\pm$ 2.09	16.50 $\pm$ 3.27	-16.35 $\pm$ 3.17	59.15 $\pm$ 1.15	62.62 $\pm$ 0.80	4.56 $\pm$ 1.71	-3.98 $\pm$ 1.52	56.38 $\pm$ 0.84
Fre. / Fre.	63.03 $\pm$ 0.46	0.97 $\pm$ 0.32	-0.97 $\pm$ 0.32	60.27 $\pm$ 0.71	61.64 $\pm$ 0.63	0.41 $\pm$ 0.19	-0.34 $\pm$ 0.26	59.14 $\pm$ 1.22
Inh. / Fre.	63.08 $\pm$ 0.80	0.96 $\pm$ 0.33	-0.96 $\pm$ 0.33	59.86 $\pm$ 0.81	63.04 $\pm$ 0.61	0.29 $\pm$ 0.09	0.21 $\pm$ 0.27	59.77 $\pm$ 1.39
Inh. / Inh.	64.43 $\pm$ 0.85	0.37 $\pm$ 0.14	-0.32 $\pm$ 0.25	61.39 $\pm$ 0.78	65.41 $\pm$ 0.56	0.33 $\pm$ 0.32	0.02 $\pm$ 0.32	61.27 $\pm$ 1.16

Table 4: Expert isolation and initialization ablation. w/o isolation sequentially updates a shared head and LoRA adapter; other rows use isolated experts. Method names specify head/LoRA initialization: Fre. denotes fresh initialization (random head or standard LoRA initialization), and Inh. denotes inheritance from the preceding expert. Best means are bold.

Dataset	Order	Routing features	Expert selection	FSA ( $\uparrow$ )	AF ( $\downarrow$ )	BWT ( $\uparrow$ )	OS ( $\uparrow$ )
ICBHI	Forward	Final layer	Hard	64.09 $\pm$ 0.92	0.39 $\pm$ 0.20	-0.32 $\pm$ 0.29	61.02 $\pm$ 0.91
		Multi-level	Hard	64.48 $\pm$ 0.83	0.29 $\pm$ 0.11	-0.23 $\pm$ 0.21	61.41 $\pm$ 0.79
		Final layer	Top-2 confidence	64.11 $\pm$ 0.92	0.39 $\pm$ 0.20	-0.32 $\pm$ 0.29	61.09 $\pm$ 0.94
		Multi-level	Top-2 confidence	64.43 $\pm$ 0.85	0.37 $\pm$ 0.14	-0.32 $\pm$ 0.25	61.39 $\pm$ 0.78
	Reverse	Final layer	Hard	65.17 $\pm$ 0.38	0.18 $\pm$ 0.10	0.43 $\pm$ 0.24	61.16 $\pm$ 1.02
		Multi-level	Hard	65.43 $\pm$ 0.55	0.29 $\pm$ 0.31	0.09 $\pm$ 0.36	61.28 $\pm$ 1.16

Table 5: Factorial ablation of routing features and expert selection on ICBHI and SPRSound under forward and reversed domain orders. Hard selects the top-ranked expert. Top-2 confidence selects between the two top-ranked experts using routing and classifier confidence.

Dataset	Order	Top-1 Acc. (%)	Top-2 Recall (%)	Domain-oracle FSA $\uparrow$	Routed FSA $\uparrow$	Gap $\downarrow$
ICBHI	Forward	83.60	93.53	65.16 $\pm$ 0.74	64.43 $\pm$ 0.85	0.72 $\pm$ 0.45
ICBHI	Reverse	83.49	93.27	65.81 $\pm$ 0.50	65.41 $\pm$ 0.56	0.41 $\pm$ 0.23
SPRSound	Forward	55.98	83.66	82.92 $\pm$ 0.39	83.12 $\pm$ 0.32	-0.20 $\pm$ 0.18
SPRSound	Reverse	55.90	83.63	82.13 $\pm$ 0.43	82.23 $\pm$ 0.91	-0.11 $\pm$ 0.79

Table 6: Routing diagnostics. Top-1 accuracy and Top-2 recall indicate whether the true domain is retrieved first or among two candidates. Domain-oracle uses the corresponding domain expert. Gap is domain-oracle FSA minus routed FSA, computed per seed. FSA and Gap are mean

\pm

standard deviation over five seeds. Domain labels are used only for diagnostics.

Figure 4: t-SNE visualization of multi-level domain-matching scores under the forward order. The same recordings are colored by their true domains (left) and selected experts (right). Paired insets magnify the same region; black outlines mark recordings assigned to an expert from a different domain.

Dataset	Bank params. (M)	Storage (MiB)	Latency (ms/sample)
ICBHI	2.372	9.12	4.81 $\pm$ 0.01
SPRSound	2.365	9.09	4.97 $\pm$ 0.04

Table 7: Deployment cost of APRIL using Top-2 confidence selection. Latency is measured with a batch size of 32 and reported over three repetitions. Storage excludes the shared frozen backbone.

Figure 5: Sensitivity to prototype count

K

. Curves and shading show five-seed means and standard deviations.

Respiratory Sound Classification

Respiratory sound classification uses audio and machine learning to identify breathing events and support respiratory disease assessment.

See all 28 papers

Parameter isolation with domain-specific experts for incremental audio classification
Sep 13, 2026Jongyeon Park, Do-Hyeon Lim, Sang-won Park +4Domain-Incremental Learning Audio Classification
Dept. of AI Convergence, Gwangju Institute of Science and Technology (GIST) · Dept. of EECS, Gwangju Institute of Science and Technology (GIST) · AunionAI Co., Ltd., Gwangju, Korea +1
Clinically Informed Acoustic-Aware Attribute Modeling for Respiratory Sound Classification
Oct 16, 2026Xinhong Li, Li Xiao, Ziyang Chen +2Respiratory Sound Classification Classification
School of Computer Science, Wuhan University, Wuhan, China · Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, Wuhan, China
Mitigating Stethoscope-Induced Shortcuts in Respiratory Sound Classification under Federated Domain Generalization with Causality-Inspired Interventions
May 28, 2026Heejoon Koo, Yoon Tae Kim, Miika Toikkanen +1OOD Generalization Federated Domain Generalization
University of Illinois Urbana-Champaign, USA · RSC LAB, MODULABS, Korea · Wonkwang University, Korea

Sep 13, 2026J/K move · Enter open · S save

Parameter isolation with domain-specific experts for incremental audio classification

Jongyeon Park, Do-Hyeon Lim, Sang-won Park, Hong Kook Kim+3

Dept. of AI Convergence, Gwangju Institute of Science and Technology (GIST) · Dept. of EECS, Gwangju Institute of Science and Technology (GIST) · AunionAI Co., Ltd., Gwangju, Korea+1

Domain-Incremental Learning Audio Classification

Open paper AskPDF arXiv

To successfully deploy a model in time-varying environments such as streaming data prediction and sensing control, domain-incremental learning (DIL) has attracted attention since it aims to adapt a previously trained model to newly arriving domains, while reserving knowledge from earlier domains without accessing their data. Incremental learning across domains can be regarded as a recurrent update, in which the current model is obtained by updating the model carried over from previous domains. Conventional DIL approaches that rely on domain-invariant feature learning and weight regularization gradually overwrite or constrain parameters learned in previous domains, leading to catastrophic forgetting. Instead, this paper proposes a new domain-specific parameter-isolation architecture that retains all past domains. The proposed architecture mitigates catastrophic forgetting through a full-order recurrent update, constructing a new expert using domain-specific data conditioned on all previously frozen models. To achieve this, we incorporate data-free generative replay to reconstruct previous-domain data and cross-domain feature generation to recover later expert features missing from earlier domain samples. Finally, we apply the proposed model architecture to domain-agnostic incremental learning for audio classification, as defined in the DCASE 2026 Challenge Task 7. Consequently, we achieve micro and macro accuracies of 78.4% and 78.9%, respectively, representing increases of 33 and 25 percentage points over the Challenge baseline. Ablation studies are conducted to examine the effectiveness of each processing component in terms of classification accuracy.

Abstract

Respiratory Sound Classification

Explore similar work

Parameter isolation with domain-specific experts for incremental audio classification

Clinically Informed Acoustic-Aware Attribute Modeling for Respiratory Sound Classification

Mitigating Stethoscope-Induced Shortcuts in Respiratory Sound Classification under Federated Domain Generalization with Causality-Inspired Interventions