PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Neural Network for Respiratory Sound Classification in Wearable Devices Enabled by Patient Specific Model Tuning

Jyotibdha Acharya, Arindam Basu

arXiv:2004.08287v1eess.AScs.LGcs.SDstat.ML

TL;DR

The paper addresses automated detection of wheeze and crackle for respiratory monitoring despite limited patient data and wearable-device memory constraints. It combines a hybrid CNN-RNN model, patient-specific retraining, and logarithmic weight quantization, achieving 71.81% after patient-specific tuning while reducing memory requirements.

  • Problem

    Automated respiratory anomaly detection must handle clinically important wheeze and crackle, sparse medical data, patient variability, and wearable-device memory constraints.

  • Method

    The paper classifies Mel-spectrograms with a hybrid CNN-RNN, screens patients for patient-specific retraining, and applies layerwise logarithmic weight quantization.

  • Results

    71.81% was achieved for patient-specific models with leave-one-out validation, compared with 66.31% for four-class classification on the 80-20 split.

  • Takeaways & Limitations

    Patient-specific retraining and quantized weights support reliable respiratory monitoring on memory-constrained wearable healthcare systems.

Abstract

from arXiv · show

The primary objective of this paper is to build classification models and strategies to identify breathing sound anomalies (wheeze, crackle) for automated diagnosis of respiratory and pulmonary diseases. In this work we propose a deep CNN-RNN model that classifies respiratory sounds based on Mel-spectrograms. We also implement a patient specific model tuning strategy that first screens respiratory patients and then builds patient specific classification models using limited patient data for reliable anomaly detection. Moreover, we devise a local log quantization strategy for model weights to reduce the memory footprint for deployment in memory constrained systems such as wearable devices. The proposed hybrid CNN-RNN model achieves a score of 66.31% on four-class classification of breathing cycles for ICBHI'17 scientific challenge respiratory sound database. When the model is re-trained with patient specific data, it produces a score of 71.81% for leave-one-out validation. The proposed weight quantization technique achieves ~4X reduction in total memory cost without loss of performance. The main contribution of the paper is as follows: Firstly, the proposed model is able to achieve state of the art score on the ICBHI'17 dataset. Secondly, deep learning models are shown to successfully learn domain specific knowledge when pre-trained with breathing data and produce significantly superior performance compared to generalized models. Finally, local log quantization of trained weights is shown to be able to reduce the memory requirement significantly. This type of patient-specific re-training strategy can be very useful in developing reliable long-term automated patient monitoring systems particularly in wearable healthcare solutions.

I. INTRODUCTION

The paper targets automated detection of respiratory sound anomalies while addressing clinician subjectivity, sparse medical data, patient variability, and wearable-device resource constraints. It proposes hybrid deep learning, patient-specific tuning, and weight quantization to support respiratory diagnosis.

  • Clinical motivation: Wheeze and crackle are clinically significant respiratory anomalies associated with airway and interstitial lung diseases.Wheeze is continuous and high-pitched, whereas crackles are explosive, discontinuous, and short-lived.
  • Clinical motivation: Automated analysis can reduce dependence on trained professionals and mitigate subjectivity in interpreting adventitious lung sounds.Auscultation is non-invasive, but limited practitioner availability and differing interpretations hinder scalable assessment.
  • Data constraints: Sparse medical datasets make deep-network training difficult, motivating transfer learning from learned domain representations to smaller task-specific datasets.The paper describes transfer learning as retraining a smaller network section using limited data for a specific task and domain.
  • Patient specificity: Generalized models may be unreliable for unseen patients because of inter-patient variability, whereas patient-specific models can produce more consistent diagnoses despite greater data-collection demands.The paper frames additional patient-specific data collection as particularly relevant for chronic-disease monitoring.
  • Proposed direction: The proposed system combines a hybrid CNN-RNN classifier, patient screening and tuning from limited patient data, and layerwise logarithmic quantization for lower memory use.These components target respiratory sound classification and deployment on memory-constrained wearable or mobile platforms.

II. MATERIALS AND METHODS

The study uses the ICBHI’17 respiratory sound database and a three-stage hybrid CNN-RNN architecture for four-class breathing-cycle classification.

  • A. Dataset: The ICBHI’17 database is described as the largest publicly available respiratory sound database used in this work.It contains recordings from patients and expert-annotated breathing cycles.
  • Model architecture: The hybrid CNN-RNN has three stages: CNN feature extraction from Mel-spectrograms, Bi-LSTM temporal modeling, and fully connected-softmax class prediction.The architecture converts learned spectrotemporal representations into breathing-sound class outputs.
  • A. Dataset: Each breathing cycle is assigned one of four classes: normal, wheeze, crackle, or both wheeze and crackle.The dataset contains 6,898 respiratory cycles from 920 recordings and 126 patients.

B. Evaluation Metrics

The evaluation uses official four-class respiratory metrics and supplements them with macro-averaged precision, recall, and F1-score to reduce class-imbalance effects.

  • B. Evaluation Metrics: The official evaluation concerns four-class classification of normal, crackle, wheeze, and both respiratory cycles.The dataset’s original challenge split contains 539 training recordings and 381 testing recordings with no common patients.
  • B. Evaluation Metrics: The experiments use an 80-20 patient split comprising recordings from 101 training patients and 25 testing patients.The split is motivated by the large data requirements of deep learning models.
  • B. Evaluation Metrics: Precision, recall, and F1-score are computed with macro-averaging because micro-averaged official metrics may let normal samples overshadow other classes.Macro-averaging computes each metric per class and then averages across classes.

C. Related Work

Prior work on ICBHI’17 spans classical models, RNNs, CNNs, residual networks, and patient-level systems, but reported results vary across tasks, splits, and labels. One highlighted concern is the reliability and reproducibility of manually obtained noise labels.

  • Classical and recurrent models: Jakovljevic et al. reported 39.56% on the original train-test split and 49.5% with 10-fold cross-validation using an HMM-GMM pipeline.Their approach used spectral-subtraction noise suppression and MFCC features.
  • Classical and recurrent models: Kochetov et al. reported 65.7% with an attention network for noise filtering followed by an RNN for four-class classification.The paper notes that the method depends on noise labels not included in the ICBHI metadata.
  • Limitations in prior work: The authors identify manually obtained noise labels as a limitation because the dataset lacks them and the labeling procedure is not described.They argue that the absence of an objective labeling method makes the results unreliable and irreproducible.
  • Deep learning approaches: Perna et al. used a deep CNN with MFCC features and reported 83% accuracy for healthy-versus-unhealthy classification and 82% for ternary disease classification.These tasks differ from four-class breathing-cycle classification.
  • Deep learning approaches: Chen et al. reported 98.79% accuracy using S-transform feature maps and ResNets on 489 recordings for three-class recording-level classification.The reported result used a 70-30 train-test split.
  • Patient-level models: Chambres et al. combined handcrafted acoustic features with boosted trees for breathing-cycle classification and subsequent healthy-versus-unhealthy patient classification.Their feature groups included melbands, MFCCs, rhythm, sound-effect, and tonal features.

D. Proposed Method

The method standardizes respiratory audio, augments the limited dataset, and represents each breathing cycle as a Mel-frequency spectrogram for model input.

  • Signals are downsampled to 4kHz because wheeze and crackle typically occur between 0 and 2kHz.The downsampling is intended to preserve relevant anomaly information while harmonizing sampling frequencies.
  • Noise addition, speed variation, random shifting, and pitch shifting augment the relatively small training dataset.These transformations also expose the network to varying recording conditions, equipment, patient age, gender, and inter-patient variability.
  • Each breathing cycle is converted into a 2D Mel-frequency spectrogram using a 60 ms window with 50% overlap.Rows encode Mel-scale frequencies, columns encode time windows, and values represent log amplitude.

2) Hybrid CNN-RNN:

The proposed hybrid CNN-RNN processes breathing-sound spectrograms through convolutional feature extraction, bidirectional temporal modeling, and classification layers, with CNN baselines used for comparison.

  • Hybrid CNN-RNN: Batch normalization, convolution, ReLU activation, and max-pooling form the CNN feature-extraction stage.Max-pooling reduces network parameters while convolutional kernels produce abstract feature maps.
  • Hybrid CNN-RNN: Bi-LSTM outputs depend on current, past, and future sequence values because its two LSTM layers process opposite directions.The recurrent cells use gates to block or pass sequence information according to learned importance.
  • Hybrid CNN-RNN: The final layers output class probabilities and train the four-class model using categorical crossentropy, Adam, and dropout regularization.Dropout is applied in the fully connected layer to reduce overfitting.
  • Comparison models: VGG-16 and Mobilenet are benchmarked using ImageNet-trained weights because the augmented respiratory dataset remains limited.Their dense layers are replaced with an architecture similar to the proposed model’s final classification layers.

3) Patient Specific Model Tuning:

Patient-specific tuning addresses inter-patient variability by adapting a domain-pretrained network with limited patient data, while local logarithmic quantization reduces deployment memory requirements.

  • Patient-specific motivation: Inter-patient variability is reported to exceed intra-patient variability across the examined audio features.The comparison includes duration, RMSE, bandwidth, roll-off, and ZCR.
  • Patient-specific motivation: Patient-specific training can improve consistency, but collecting enough data for a deep model from one patient is difficult.This constraint is especially relevant to chronic respiratory monitoring after initial treatment.
  • Patient-specific tuning: The tuning strategy first trains on a large respiratory database, then retrains a smaller network portion using limited patient-specific data.The approach transfers domain-specific representations into patient-specific classification models.
  • Weight quantization: Layerwise logarithmic quantization maps each layer’s weight magnitudes into the log domain for compact storage.The quantized representation uses the layer’s weight range and preserves a sign bit in addition to N precision bits.
  • Weight quantization: Local normalization is preferred because convolution, batch-normalization, and LSTM layers can have different weight ranges.The extra storage for each layer’s minimum and maximum weights is described as insignificant relative to quantized weights.
  • Weight quantization: Very small weights are rounded to zero before logarithmic quantization to limit the log-domain quantization range.

III. RESULTS AND DISCUSSIONS

The evaluation compares the proposed model with standard CNN architectures on four-class breathing-cycle classification using repeated randomized train-test splits.

  • The proposed model, VGGnet, and Mobilenet are evaluated on four-class breathing-cycle classification.The results are averaged over five randomized train-test sets.

B. Patient Specific Model Tuning Strategy

The strategy screens patients with a pre-trained model, then fine-tunes that model on limited patient-specific data for unhealthy patients. Patient-specific tuning outperforms the compared models and strategies, reaching 71.81% while using approximately 47 breathing cycles on average.

  • Screening and tuning: The strategy first screens patients as healthy or unhealthy using the proportion of breathing cycles predicted as unhealthy.Patients classified as unhealthy then receive patient-specific model retraining for subsequent four-class breathing-cycle prediction.
  • Screening and tuning: Approximately 47 patient breathing cycles are used on average to fine-tune patient-specific models.The dataset contains varying numbers of recordings and breathing cycles across patients.
  • Evaluation design: The evaluation compares patient-specific tuning against a VGG-16 feature extractor with an SVM classifier trained only on patient-specific data.This comparison tests whether the proposed tuning strategy offers an advantage over a classifier relying solely on patient-specific data.
  • Evaluation design: The study also tests whether respiratory-data pre-training transfers domain-specific representations better than pre-training on speech-recognition or urban-sound datasets.Each pre-trained model is subsequently retrained using patient-specific data.
  • Results: 71.81% is the proposed strategy’s score, outperforming the other evaluated models and strategies in patient-specific model comparisons.VGG-16 and MobileNet achieve 68.54% and 67.60%, respectively; ImageNet-trained VGG-16 remains below ICBHI-trained models.

C. Memory and Computational Complexity

The models differ in their memory and computation trade-offs under local log quantization. Hybrid CNN-RNN minimizes memory and supports near-MobileNet computational efficiency, while MobileNet prioritizes efficiency and VGG-16 incurs higher resource costs.

  • Local log quantization: MobileNet shows poor quantization performance, while the hybrid CNN-RNN performs slightly worse than VGG-16 because its LSTM layer requires higher bit precision.MobileNet’s quantization behavior is attributed to its batch-normalization layers and RELU6 activation.
  • Local log quantization: VGG-16 achieves full-precision accuracy at minimum bit precision, whereas MobileNet requires maximum bit precision under local log quantization.
  • Resource comparison: VGG-16 requires significantly more memory despite weight compression, while hybrid CNN-RNN requires the least memory and MobileNet requires the least computational complexity.Computational complexity is measured in GFLOPS per sample, and memory is normalized relative to hybrid CNN-RNN.
  • Resource comparison: MobileNet is more suitable when power is the primary constraint, whereas hybrid CNN-RNN offers better performance at minimal memory footprint.
  • Real-time operation: The proposed system processes each breathing cycle once, and its computational cost is of the same order as MobileNet.Because the minimum breathing cycle duration exceeds one second and MobileNet latency is approximately 100 ms per sample on modern mobile SoCs, the system should support real-time classification.

IV. CONCLUSION

The paper develops a hybrid CNN-RNN model and a patient screening and tuning strategy for respiratory sound classification. It reports strong ICBHI’17 performance and identifies future work toward embedded wearable implementation.

  • Model performance: The hybrid CNN-RNN model achieves a 66.31% score on four-class respiratory cycle classification using an 80-20 split of the ICBHI’17 dataset.
  • Patient-specific tuning: Patient screening and patient-specific retraining achieve a 71.81% score under leave-one-out cross-validation on the original train-test split.
  • Transfer learning: Image-recognition pre-trained models transfer knowledge better than speech-pre-trained models in the reported experiments.The authors suggest larger ImageNet training data may explain this result, but they could not verify the hypothesis because of limited audio-domain models and training costs.
  • Future work: The authors plan an embedded implementation for wearable, at-home patient monitoring and further reductions in computational complexity using neuromorphic spike-based methods.
Loading 2004.08287v1…