TL;DR
This paper introduces a novel approach to cough event classification by leveraging pretrained deep neural networks (DNNs) combined with Hidden Markov Models (HMMs). By using a pretrained DNN, the model achieves significantly improved performance in distinguishing coughs from non-coughs compared to traditional GMM-HMM methods, offering a more objective and accurate cough monitoring solution.
What Problem Does This Paper Address?
Cough is a critical symptom in respiratory diseases, and accurately measuring its severity is crucial for monitoring disease progression and guiding clinical practice. Current methods rely heavily on subjective self-report scales, which are prone to bias and lack objectivity. While automatic cough detection systems have been developed, they often struggle with satisfactory detection precision and require extensive human labeling. This paper aims to address the need for a more accurate and objective cough classification system, a key component of a cough monitor, by exploring the potential of deep learning.
Methods
The core of the proposed method involves a pretrained Deep Neural Network (DNN) integrated with a Hidden Markov Model (HMM) decoder. The process consists of two main stages:
- Pretraining: A Deep Belief Network (DBN), composed of stacked Restricted Boltzmann Machines (RBMs), is trained in a greedy, layer-wise unsupervised manner. This pretraining step learns a good initial set of weights for the DNN, forcing it to map raw audio features to higher-level abstractions and acting as a data-induced regularizer.
- Fine-tuning: The pretrained DBN's parameters are then used to initialize a DNN. This DNN is subsequently fine-tuned using backpropagation. In this stage, the DNN learns to predict the observation probabilities for each HMM state.
- HMM Decoding: The HMM states are initially derived from a force-alignment process using a Gaussian Mixture Model-HMM (GMM-HMM) on the training data. The system uses three HMMs for coughs and one for non-coughs. A Viterbi decoding algorithm is then applied to the DNN's output probabilities to generate the most likely HMM sequence for each audio sample, thereby classifying it as cough or non-cough.
This approach combines the powerful feature learning capabilities of DNNs with the temporal modeling strengths of HMMs.
Datasets
The study utilized a dataset collected from 22 patients in China suffering from various respiratory diseases, including:
- 12 patients with community-acquired pneumonia (CAP)
- 2 patients with bronchial asthma (BA)
- 8 patients with chronic obstructive pulmonary disease (COPD)
The recordings were made using a portable digital audio recorder attached to patients' collars, capturing approximately 24 hours of audio per patient. The data was segmented into 10-minute clips, then labeled by two annotators using the Pratt platform. These clips were further segmented into 10-second clips for analysis. The dataset was split into patient-dependent (PD) and patient-independent (PI) test sets for evaluation.
Key Insights and Ideas
- Pretraining is Crucial for DNNs: The study strongly suggests that unsupervised pretraining of DNNs using a DBN significantly improves performance, especially when training data is limited. Without pretraining, DNNs sometimes performed worse than baseline models.
- DNNs Excel at Feature Abstraction: The pretraining process forces DNNs to learn more abstract representations of the audio features, which is beneficial for classification tasks.
- HMM-DNN Synergy: Combining DNNs for feature representation and probability estimation with HMMs for temporal modeling yields superior results compared to traditional GMM-HMMs.
- Specificity Advantage: The pretrained DNN-HMM framework demonstrated a notable advantage in specificity, achieving much higher values than GMM-HMM.
- Sensitivity Trade-off: While DNNs excelled in specificity, the GMM-HMM model showed higher sensitivity in some cases, suggesting a potential trade-off that might be addressed by ensemble methods.
- Generalization Ability: The models showed good generalization capabilities, performing similarly on both patient-dependent and patient-independent test sets, indicating they learn general cough characteristics rather than just overfitting to specific patient data.
Key Results
The HMM-DNN framework consistently outperformed the conventional GMM-HMM model across various evaluation metrics.
- Error Reduction: The DNN-based methods achieved maximal error reductions of 14% for F1 and 11% for micro average on PD, and 7% for F1 and 10% for micro average on PI compared to the baseline GMM-HMM.
- Specificity Improvement: DNNs surpassed GMM-HMM on specificity with maximal 14% error reduction on both PD and PI.
- Overall Performance: The paper concludes that the HMM-DNN framework provides better overall performance for cough classification.
- Specificity vs. Sensitivity: Pretrained DNNs achieved over 90% specificity, while GMM-HMM achieved around 77%. However, GMM-HMM showed higher sensitivity (94-97%) compared to DNNs (under 85%).
Limitations and Future Work
- Data Volume: The authors acknowledge that the amount of training data might be insufficient for very deep DNN models, which could explain why the most complex model performed best in some scenarios. More data is needed to train deeper models effectively.
- Sensitivity Gap: The pretrained DNN models did not reach the high sensitivity levels of the GMM-HMM model. This might be due to the variance in cough samples and the DNN's tendency to learn non-cough patterns more robustly.
- Future Directions:
- Focus on building more powerful tools for cough detection (distinct from classification).
- Develop more stable audio recording systems for cleaner data.
- Address the significant workload of labeling cough data, potentially through efficient labeling workflows or by exploring methods that reduce reliance on extensive manual labeling.
- Investigate ensemble methods that combine the strengths of DNNs and GMM-HMMs for potentially even higher classification accuracy.
Summary updated