Source-linked AI summary
Deep auscultation: Predicting respiratory anomalies and diseases via recurrent neural networks
Diego Perna, Andrea Tagarelli
TL;DR
Respiratory auscultation can support the detection of disease-related anomalies, but RNN-based approaches for respiratory prediction at both anomaly and pathology levels had not been adequately studied. This paper combines MFCC-based preprocessing with RNN models, reporting that its methods outperform competitors on both tasks on the ICBHI benchmark.
Problem
RNN-based learning frameworks for respiratory prediction at both anomaly and pathology levels had not been adequately studied, although auscultation sound analysis can support disease detection.
Source: §1 · p. 2Method
Results
On the ICBHI benchmark, the authors report that their RNN methods outperform selected competitors on both anomaly- and pathology-level prediction tasks.
Source: §5.3 · p. 12, §6 · p. 13, §5.4 · p. 13Takeaways & Limitations
The empirical findings suggest the framework advances the state of the art in respiratory prediction at anomaly and pathology levels.
Source: §6 · p. 13
Abstract
from arXiv · showhide
Respiratory diseases are among the most common causes of severe illness and death worldwide. Prevention and early diagnosis are essential to limit or even reverse the trend that characterizes the diffusion of such diseases. In this regard, the development of advanced computational tools for the analysis of respiratory auscultation sounds can become a game changer for detecting disease-related anomalies, or diseases themselves. In this work, we propose a novel learning framework for respiratory auscultation sound data. Our approach combines state-of-the-art feature extraction techniques and advanced deep-neural-network architectures. Remarkably, to the best of our knowledge, we are the first to model a recurrent-neural-network based learning framework to support the clinician in detecting respiratory diseases, at either level of abnormal sounds or pathology classes. Results obtained on the ICBHI benchmark dataset show that our approach outperforms competing methods on both anomaly-driven and pathology-driven prediction tasks, thus advancing the state-of-the-art in respiratory disease analysis.
1 Introduction
Respiratory diseases impose a substantial global health burden, making prevention and early diagnosis important. The paper proposes an RNN-based framework for predicting respiratory anomalies and diseases from auscultation sounds, evaluated on the ICBHI benchmark.
1 Introduction
- COPD, asthma, lower respiratory tract infection, tuberculosis, and lung cancer are among the respiratory diseases identified as major causes of severe illness and death worldwide.COPD alone affects 65 million people and causes about 3 million deaths annually.Source: §1 · pp. 1–2
- Auscultation can reveal crackles and wheezes, and computational sound analysis could support physicians in detecting respiratory disease.The authors emphasize prevention and early diagnosis as important to limiting disease spread and effects on length and quality of life.Source: §1 · p. 2
- The paper proposes an RNN-based framework for anomaly- and pathology-level prediction, with flexible extraction of cepstral features and evaluation on the ICBHI benchmark.The authors report that their methods outperform selected challenge competitors and a deep-learning competitor.Source: §1 · p. 2
2 The ICBHI Challenge
The ICBHI Challenge provides respiratory recordings with cycle-level sound annotations and patient-level pathology annotations. The section describes abnormal sounds and the dataset’s respiratory cycles and recording variability.
2 The ICBHI Challenge
- The ICBHI dataset combines recordings collected by two teams in different countries, with variability in equipment, microphone position, and environmental noise.Its annotations identify crackles and/or wheezes for each respiratory cycle and predetermined pathologies for each patient.Source: §2 · p. 3
2.1 Abnormal sounds
- Crackles are discontinuous sounds, while wheezes are continuous, musical sounds with a dominant frequency typically at or above 400 Hz.Crackles are classified by duration: coarse crackles exceed 10 ms and fine crackles are shorter than 10 ms.Source: §2.1 · p. 3
2.2 Respiratory data
- The dataset contains 5.5 hours of recordings and 6,898 respiratory cycles from 920 audio samples and 126 subjects.Of the cycles, 1,864 contain crackles, 886 wheezes, and 506 both; recordings used heterogeneous equipment and lasted 10–90 seconds.Source: §2.2 · p. 4
- A respiratory cycle typically moves from a lower-amplitude inspiratory phase to an expiratory phase with peaks, decreasing amplitude, and higher average energy.Figure 1 shows an example waveform from a healthy patient.Source: §2.2 · p. 4, Figure 1 · p. 4
3 Related Work
Prior work addresses respiratory prediction through signal-processing, machine-learning, and deep-learning approaches, with much of the cited work focused on anomaly detection. This paper also evaluates pathology prediction using MFCC features and RNNs that model time-dependent patterns.
3 Related Work
- Prior anomaly-detection methods include HMM/GMM classifiers, signal-processing features with an SVM, MNRNN, and boosted decision trees.The MNRNN includes noise and anomaly classifiers with a mask for selecting noiseless frames.Source: §3 · p. 5
- An earlier pathology-driven approach used CNNs and MFCCs on whole recordings and applied SMOTE to address class imbalance.Source: §3 · p. 5
- The paper tackles both anomaly-driven and the more challenging pathology-driven prediction tasks.Source: §3 · p. 5
- Its MFCC-based RNN approach uses the whole ICBHI dataset, including high-noise frames, and models time-dependent patterns that other architectures may ignore.The authors contrast this with approaches that omit noisy frames and note that RNNs can discover time-dependent patterns in sound data.Source: §3 · p. 6
4 Our Proposed Learning Framework
The framework applies recurrent neural networks to respiratory prediction at both anomaly and pathology levels. Its pipeline prepares cycle-level inputs with MFCC features and normalization, then evaluates predictions using challenge-specific and macro-averaged criteria.
4 Our Proposed Learning Framework
- The framework uses RNNs for anomaly-level and pathology-level respiratory prediction, including two-class and multi-class problems.Source: §4 · p. 6
4.1 Recurrent Neural Networks
- LSTM and GRU models, in unidirectional and bidirectional forms, capture sequence information from respiratory sounds.Source: §4.1 · p. 7
4.2 Preprocessing
- Each respiratory cycle is segmented with sliding windows; extracted MFCC coefficients are concatenated into frames fed to the RNN.Source: §4.2.1 · p. 8, Table 1 · p. 8
4.2.1 Frame composition
- Seven frame configurations vary window size, step, and the number of concatenated windows; settings use either no overlap or 50% overlap.Source: §4.2.1 · p. 8, Table 1 · p. 8
4.2.2 Feature extraction
- MFCC feature extraction represents sound using cepstral features derived from the amplitude spectrum.The pipeline frames the input, applies a window and DFT, logs the amplitude spectrum, aggregates frequency bins, then applies DCT.Source: §4.2.2 · pp. 8–9
4.2.3 Feature normalization
- The framework considers Min-Max normalization, which scales values to [0,1], and Z-score standardization, which centers and scales values using the population mean and standard deviation.Source: §4.2.3 · p. 9
4.3 Evaluation and assessment criteria
- Evaluation uses an 80/20 ICBHI train-test split, challenge-specific micro-averaged sensitivity and specificity with their ICBHI-score average, and macro-averaged accuracy, precision, recall, and F1-score.Macro-averaged scores are averages over classes; the passages also define correctly recognized and total instances by class.Source: §4.3 · pp. 9–10
5 Experimental Results
The experiments evaluate feature normalization and RNN architectures, then compare the framework with competitors on anomaly- and pathology-driven prediction tasks.
5 Experimental Results
- The experiments assess normalization, compare LSTM and GRU variants, evaluate anomaly prediction against challenge competitors, and test pathology prediction against a CNN method.The evaluation plan assigns these goals to Sections 5.1–5.4.Source: §5 · p. 10
5.1 Impact of feature normalization on RNN performance
- Z-score normalization generally yields higher accuracy than min-max normalization or no normalization, particularly on the four-class anomaly testbed.Table 2 reports LSTM accuracy for binary and four-class anomaly prediction across frame-composition settings.Source: §5.1 · p. 10, Table 2 · p. 11
5.2 Comparison of RNN models
- The four RNN architectures perform within 0.70–0.74; LSTMs are generally stronger but take longer to train than GRUs.For 100 epochs in binary anomaly prediction, training took about 13 minutes for LSTM, 11 for GRU, 26 for BiLSTM, and 22 for BiGRU.Source: §5.2 · p. 11
5.3 Comparison with the ICBHI Challenge competitors
- LSTM models outperform the ICBHI Challenge competitors on all three criteria, with only 1–2% ICBHI-score differences across frame-composition settings.The settings with the best challenge score (73%) are S2, S3, and S4, using different frame sizes and MFCC counts.Source: Table 3 · p. 12, §5.3 · p. 12
5.4 Performance on the pathology-driven prediction tasks
- LSTM methods, especially settings S4 and S7, outperform the CNN competitor on binary pathology prediction; ternary results favor LSTMs in all cases.The reported binary-testbed gains reach up to 16% accuracy, 9% recall, 6% F1-score, 4% specificity, 3% sensitivity, and 3% ICBHI-score.Source: Table 4 · p. 13, §5.4 · p. 13
6 Conclusion and Future Work
The conclusion describes an MFCC-based RNN framework for respiratory-sound prediction and reports results advancing the state of the art on two ICBHI tasks. Future work would explore other architectures and sound representations.
6 Conclusion and Future Work
- The framework combines MFCC-based sound preprocessing with RNNs to detect crackles, wheezes, and chronic or non-chronic diseases.Source: §6 · p. 13
- The ICBHI evaluation suggests the framework advances the state of the art in anomaly-level and pathology-level prediction.Source: §6 · p. 13
- Future research would investigate alternative deep-learning architectures and sound representations, including hybrids combining time- or frequency-domain series with MFCCs.Source: §6 · pp. 13–14