Respiratory diseases are a major global health challenge. However, identification of respiratory diseases is often limited by subjectivity, environmental noise and inter-clinician variability. This study presents an explainable multimodal deep learning framework for recording-level multiclass classification of respiratory audio signals. The proposed system integrates two complementary representations—a spectro-temporal encoder based on a CNN–BiLSTM-attention architecture and a handcrafted acoustic-feature encoder capturing acoustic descriptors commonly used in respiratory-audio analysis, including MFCCs, zero-crossing rate, spectral centroid, spectral bandwidth, chroma, RMS energy, and spectral rolloff features. These branches are combined through late-stage fusion to leverage both data-driven representation learning and domain-informed acoustic cues. The proposed model was trained and internally evaluated on the Asthma Detection Dataset Version 2, comprising five respiratory categories: bronchial disease, asthma, COPD, healthy, and pneumonia. Mono conversion, resampling to 16 kHz, 100–2000 Hz band-pass filtering, amplitude normalisation, fixed 4 s trimming or zero-padding, training-only augmentation, handcrafted-feature extraction, mel-spectrogram generation, quality control auditing, and stratified recording-level partitioning have been applied in the pre-processing steps. Across five repeated experiments with different random seeds, the proposed hybrid model achieved a mean held-out recording-level test accuracy of $0.9099\pm 0.0163$, balanced accuracy of $0.8936\pm 0.0152$, macro F1-score of $0.8937\pm 0.0177$, macro ROC–AUC of $0.9867\pm 0.0010$, and macro PR–AUC of $0.9489\pm 0.0044$. Conventional machine learning baseline comparisons showed that the proposed model achieved stronger internal accuracy, balanced accuracy, macro recall, macro F1-score, and macro ROC–AUC than classical machine learning algorithms trained on handcrafted acoustic features, although Random Forest remained competitive in macro PR–AUC. Ablation analysis shows that the deep spectro-temporal branch was the primary contributor to predictive performance, while the handcrafted branch provided complementary interpretable acoustic information rather than consistently improving all classification metrics. Explainability was incorporated using Grad-CAM and Integrated Gradients for spectrogram-based interpretation and SHAP for handcrafted-feature attribution. Domain-shift evaluation on the ICBHI Respiratory Sound Database and a COPD-focused cohort revealed substantial dataset shift effects, including poor healthy-case recognition on ICBHI and seed-dependent COPD recognition in the COPD-focused cohort. Identifier-aware sensitivity analyses showed lower performance than the main recording-level split, suggesting that subject-like or source-level overlap may inflate internal performance estimates. The findings should be interpreted as promising internal held-out recording-level algorithmic performance with limited external transfer, rather than evidence of readiness for clinical use.

Background/Objectives: Estrogen receptor immunohistochemistry (ER-IHC) exhibits heterogeneous staining and substantial class imbalance, complicating segmentation of ordered expression categories. We evaluated whether class weighting, minority-focused crop sampling, adaptive minority curriculum (AMC), and Focal–Tversky optimization improved a common ResUNet-DS backbone for segmenting background/non-target pixels and C1 ER-negative, C2 weak-positive, C3 moderate-positive, and C4 strong-positive foreground categories. Methods: The dataset comprised 220 paired $512\times 512$ image–mask patches organized into 44 recovered five-patch source groups. A source-group-preserving nested five-fold design used outer folds of 45, 45, 45, 45, and 40 patches. Six controlled training conditions were compared, with four independently selected inner models ensembled for each condition and outer fold. Results: Unweighted random training achieved the best numerical mean for all five primary endpoints: foreground Dice ($0.7479\pm 0.0123$), C2–C4 Dice ($0.7098\pm 0.0128$), foreground IoU ($0.6074\pm 0.0147$), foreground quadratically weighted kappa ($0.9761\pm 0.0040$), and foreground-ordinal MAE ($0.0446\pm 0.0062$). Weighted random training was the strongest weighted/minority-sensitive condition but did not exceed the unweighted reference. None of 25 paired outer-fold comparisons reached $p<0.05$; the minimum raw p-value was $0.0625$, and all Holm-adjusted p-values were $0.3125$. A 20,000-replicate paired source-group bootstrap preserved the same overall ordering. Conclusions: More complex imbalance-handling strategies did not improve aggregate segmentation over unweighted training. Post hoc analyses showed scope-dependent probability quality and error ranking. IHC4BC provided expression-ordering consistency but not direct external segmentation validation. The findings support preliminary source-group-preserving methodological evidence rather than patient-level, whole-slide, or clinical validity.