Cough Detection ML Model

• Cough detection machine learning models are automated systems that identify, classify, and segment cough events from audio signals for health monitoring.
• They employ advanced feature extraction methods such as energy measures, MFCCs, and spectrograms to enhance discrimination between cough and non-cough sounds.
• The integration of HMM-based segmentation and deep learning techniques achieves robust performance, with AUC up to 0.92, making them suitable for resource-constrained deployments.

A cough detection machine learning model is an automated system designed to identify, classify, and segment cough events from continuous audio signals or sensor data, often for the purposes of respiratory disease monitoring, epidemiological research, or assistive healthcare diagnostics. Such models leverage supervised learning paradigms, ranging from hidden Markov models and classical statistical classifiers to deep neural architectures, frequently incorporating domain-specific feature extraction, sequence modeling, and context-aware decision post-processing.

1. Historical and Methodological Foundations

Early approaches to cough detection predominantly utilized statistical pattern recognition with hand-crafted features and probabilistic sequence models. One prototypical example employs a hidden Markov model (HMM) framework with energy-based acoustic features extracted from short (25 ms) bins of continuous audio recordings (Teyhouee et al., 2019). The model is designed around five annotated states corresponding to different phases of a cough and silences, with transitions governed by empirically derived state durations. Emission probabilities are modeled using (potentially independent) density functions over either univariate (single energy band) or multivariate (multiple band energies) feature spaces.

Transition probabilities between states are formulated as:

$P(s_t = y\ |\ s_{t-1} = x) = P_{xy}$

where $P_{xy}$ reflects constraints such as disallowed transitions (e.g., skipping cough phases) and is parameterized using reciprocal mean residence times and empirical exit fractions.

Observation emission modeling in the multivariate case factorizes the likelihood as:

$$P(\mathbf{x}_t | s_t = i) = f_{i,\text{low}}(x_{t,\text{low}}) \cdot f_{i,\text{mid}}(x_{t,\text{mid}}) \cdot f_{i,\text{high}}(x_{t,\text{high}})$$

where each $f_{i,k}$ is a density function over the $k$th energy band.

The training and inference involve maximizing the joint state-observation likelihood, typically using annotated transition matrices derived from manually labeled data. Modern deep learning–based methods now dominate the field but HMMs remain influential for tasks emphasizing temporal sequence structure or explainability.

2. Feature Engineering and Representation

Feature extraction is central to cough detection model performance. Traditional methods focus on time- and frequency-domain energy measures; for instance, energy densities are computed over sliding windows, with separation into frequency bands (<2 kHz, 2–4 kHz, 4–22 kHz) capturing cough-specific spectral power dynamics (Teyhouee et al., 2019). This multivariate energy approach provides phase-resolved acoustic discrimination between cough and non-cough (silence/environmental noise) states, increasing AUC from ~0.74 (univariate) to ~0.79 (multivariate) in multi-class settings and to 0.92 (binary cough/non-cough) in the referenced paper.

Feature engineering strategies also employ Mel-frequency cepstral coefficients (MFCCs), linear predictive coding–derived features, and time-frequency representations (e.g., spectrograms, log-Mel spectrograms), chosen for their efficacy in representing the nonstationary, bursty, and frequency-rich structure of coughs. The feature sets are typically standardized across training corpora and validated empirically for discriminative power.

3. Model Training, Segmentation, and Classification

A cough detection HMM is trained by maximizing empirical likelihoods against state-annotated sequences using expectation-maximization or maximum likelihood routines, as implemented in toolkits such as the “mhsmm” R package (Teyhouee et al., 2019). The model calculates, for each time bin $t$ and HMM state $i$:

$$P(s_t = i | x_t) \propto P(x_t | s_t = i) \cdot P(s_t = i | s_{t-1})$$

where $x_t$ is the feature vector for bin $t$. Inference (decoding) proceeds via the Viterbi algorithm or similar approaches, resulting in a state sequence that segments the audio signal.

Cough event detection is operationalized by mapping runs of cough-phase states (A, B, C) onto cough events, while silence-phase states (D, E) correspond to intervening non-cough intervals. The explicit Markovian structure enforces physiologically plausible event transitions and durations.

Binary or multiclass classification performances are evaluated via confusion matrices, with metrics including sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC). Youden's index is employed for threshold optimization, balancing sensitivity and specificity as determined by validation data.

4. Performance, Evaluation Metrics, and Comparison

The referenced HMM methodology achieves robust performance, particularly in challenging noisy environments (Teyhouee et al., 2019). Specifically, the multivariate HMM achieves an AUC of 0.920 (92% AUR) for binary grouping (coughing vs. non-coughing), compared to lower values for the univariate energy approach. Multi-class AUC improvement from 0.744 to 0.789 further demonstrates the informativeness of multiband energy features. Sensitivity, specificity, and accuracy are all reported in the high 80s to low 90s, confirming the method's reliability.

Grouping the five-state model output into practical binary classes increases clinical utility while reflecting the natural ambiguity in fine-grained cough phase labeling. These results support the deployment of HMM-based segmentation in real-world, continuous monitoring applications.

5. Limitations, Trade-offs, and Model Selection

The main trade-off in HMM-based cough detection models is between interpretability, computational efficiency, and discriminative performance. The HMM's explicit temporal structure and emission probabilities offer interpretable, phase-aware cough segmentation, but such models are limited by their parametric (often unimodal or independent) emission assumptions, and may underperform compared to deep learning approaches on diverse, unlabeled, or large-scale datasets.

For deployment on embedded or resource-constrained platforms, HMMs and energy-based systems offer a lightweight yet accurate alternative to more computationally intensive deep neural networks. However, HMMs require carefully curated, manually labeled training data for all states, and their performance degrades if the real-world acoustic environment deviates substantially from training conditions.

Practical selection between univariate and multivariate feature models rests on the available spectral resolution, memory, and annotation resources. The referenced paper clearly demonstrates the superiority of multivariate approaches when computational resources permit.

6. Real-World Applications and Deployment Considerations

HMM-based cough detection models are well-suited for clinical monitoring, epidemiological surveillance, and health research applications requiring temporal precision (e.g., cough frequency quantification, bout analytics). Their relatively low computational footprint facilitates integration with wearable devices, bedside monitors, or continuous digital health recording systems.

Deployment involves standardizing microphone placement and calibration, real-time feature extraction pipelines, and inference routines capable of streaming segmentation outputs over continuous multi-hour recordings. Such systems are extensible to other symptom detection tasks where event-based segmentation and phase discrimination are required.

Strong numerical performance results (AUC up to 0.92 in binary discrimination) suggest that, even in modern contexts dominated by deep learning, HMMs remain relevant for efficient, robust, and interpretable cough event detection in noisy real-world environments.