PASA: Progressive-Adaptive Spectral Augmentation for Automated Auscultation in Data-Scarce Environments (AAAI 2026)

Wang, Y., G. Huang, X. Gong, X. Wang, and X. Yuan. “PASA: Progressive-Adaptive Spectral Augmentation for Automated Auscultation in Data-Scarce Environments”. Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 2, Mar. 2026, pp. 1213-21, doi:10.1609/aaai.v40i2.37093

PASA Paradigm

Five-Step Workflow

1. State Construction: Extract features and construct MDP states
2. Action Sampling: Use SAC to sample augmentation actions
3. HBS Strategy: Execute Hybrid Batch-Sample augmentation
4. Reward Computation: Calculate performance-based rewards
5. Network Optimization: Joint optimization of all networks

---

Installation

# Clone repository
git clone https://github.com/wangying1586/PASA.git
cd PASA

# Create conda environment
conda create -n pasa python=3.8
conda activate pasa

# Install dependencies
pip install -r requirements.txt

---

Dataset Preparation

Download Datasets

• SPRSound: https://github.com/SJTU-YONGFU-RESEARCH-GRP/SPRSound
• ICBHI 2017: https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge
• CirCor DigiScope: https://physionet.org/content/circor-heart-sound/1.0.3/

Directory Structure

Place downloaded datasets in the following structure:

data/
├── SPRSound/
│   ├── Classification/
│   │   ├── restore_train_classification_wav/
│   │   ├── train_classification_json/
│   │   ├── restore_valid_classification_wav/
│   │   └── valid_classification_json/
├── ICBHI/
│   ├── ICBHI_final_database/
│   └── ICBHI_challenge_train_test.txt
└── CirCor/
    └── the-circor-digiscope-phonocardiogram-dataset-1.0.3/

Preprocessing

Run preprocessing scripts to generate .npy files [Update dataset paths in preprocessing scripts before running]:

# SPRSound (year:22, 23 | all tasks: 11, 12, 21, 22)
python data_preprocess/SPRSound2022_2023_preprocessor.py

# ICBHI (tasks: binary & multiclass)
python data_preprocess/ICBHI_preprocessor.py

# CirCor DigiScope (heart murmur detection)
python data_preprocess/CirCor_preprocessor.py

After preprocessing, verify generated files:

datasets/
├── SPRSound_2022+2023_processed_data/
│   ├── task11/
│   │   ├── train_specs.npy
│   │   ├── train_labels.npy
│   │   └── test_*.npy
│   ├── task12/
│   ├── task21/
│   └── task22/
├── ICBHI2017_processed_data/
│   ├── binary/
│   └── multiclass/
└── CirCor_DigiScope_2022_processed_data/
    ├── train_specs.npy
    └── train_labels.npy

---

Training

Basic Usage

./train.sh      

Examples

# SPRSound Task 22 (5-class recording classification)
./train.sh SPRSound 22 True 20 0.01 0.95

# CirCor DigiScope (3-class heart murmur detection)
./train.sh CirCor heart_murmur True 20 0.01 0.95

# Baseline without augmentation
./train.sh SPRSound 22 False

Model Checkpoints

After training step, the models are saved in experiments/ directory:

experiments/
└── SPRSound_task22_EfficientNet-B4_log-mel_TwoPhaseSAC_P20_T0.01_YYYYMMDD_HHMMSS/
    ├── models/
    │   ├── best_fold_1.pth
    │   ├── best_fold_2.pth
    │   ├── best_fold_3.pth
    │   ├── best_fold_4.pth
    │   ├── best_fold_5.pth
    │   └── final_model.pth
    ├── config.txt
    └── cv_results_summary.json

---

Testing and Evaluation

Run Testing

python test.py \
    --dataset SPRSound \
    --task_type 22 \
    --data_dir ./datasets \
    --exp_dir experiments/SPRSound_task22_EfficientNet-B4_log-mel_TwoPhaseSAC_P20_T0.01_YYYYMMDD_HHMMSS \
    --batch_size 32

Replace YYYYMMDD_HHMMSS with your actual experiment timestamp.

Test Results

Results are saved in test_results/ directory:

test_results/
└── <DATASET>/task_<TASK_TYPE>/YYYYMMDD_HHMMSS/
    ├── confusion_matrix_fold_*.png
    ├── confusion_matrix_ensemble_*.png
    ├── confusion_matrix_final_*.png
    ├── roc_curve_*.png
    ├── pr_curve_*.png
    ├── results_summary.txt
    ├── <DATASET>_task_<TASK_TYPE>_test_results.json
    ├── test_config.json
    └── detailed_predictions/
        ├── test_*_predictions.csv
        └── ensemble_predictions.csv

Visualizations

The testing script generates:

1. Confusion Matrices: Shows prediction distribution across classes

   • Per-fold models (best_fold_1.pth to best_fold_5.pth)
   • Ensemble model (average of all folds)
   • Final model (final_model.pth)

2. ROC Curves: Receiver Operating Characteristic with AUC scores per class

3. Precision-Recall Curves: PR curves with Average Precision per class

4. Performance Summary Table: Comprehensive metrics comparison including:

   • Accuracy, F1-score
   • Sensitivity (Recall), Specificity
   • Overall Score (harmonic and arithmetic mean)
   • For CirCor: W.acc (Weighted Accuracy) and UAR (Unweighted Average Recall)

Understanding CirCor Metrics

For CirCor DigiScope heart murmur detection:

• W.acc: Weighted accuracy based on challenge scoring (5×Present + 3×Unknown + Absent)
• UAR: Unweighted Average Recall across three classes
• Class-specific Recalls: Present, Absent, Unknown

[THE VISUALIZATION OF SPRSound 2022 and SPRSound 2023 test datasets for task 1-1 and task 1-2 are in the task_11/20250702_030839 and task_12/20250702_030405.]

Adding Custom Augmentation Operations

Edit PASA.py in the TwoPhaseSACStrategy class:

1. Add operation to _init_magnitude_ranges():

def _init_magnitude_ranges(self):
    configs = {
        "your_augmentation": (min_val, max_val),
        "time_mask": (0.05, 0.3),
        # ... existing operations
    }

2. Implement the augmentation function:

def _apply_your_augmentation(self, x, y, magnitude):
    # Your augmentation logic here
    augmented_x = x  # Apply transformation
    return augmented_x, y

3. Add to apply_augmentation() method:

def apply_augmentation(self, x, y, operation_idx, magnitude_idx):
    operation_name = self.aug_operations[operation_idx]
    
    if operation_name == "your_augmentation":
        return self._apply_your_augmentation(x, y, magnitude)
    # ... existing conditions

---

Citation

If you find this work useful, please cite:

@inproceedings{pasa2026,
  title={PASA: Progressive-Adaptive Spectral Augmentation for Automated Auscultation in Data-Scarce Environments},
  author={Ying Wang, Guoheng Huang, Xueyuan Gong, Xinxin Wang, Xiaochen Yuan*},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

---

Contact

For questions or issues, please:

• Open an issue on GitHub
• Contact: ying.wang@mpu.edu.mo

---