No labs, no needles: AI listens to your voice and detects Asthma

This study offers a major step forward in deploying AI for accessible healthcare solutions. By utilizing only speech recordings, the system eliminates the need for specialized equipment, reducing costs and making asthma screening more feasible in remote or under-resourced areas. Additionally, the explainability features built into the platform address one of the key concerns in clinical AI adoption - transparency of decision-making.

Asthma is a widespread chronic respiratory disorder that currently relies on invasive diagnostic methods, such as antibody testing and spirometry, which are not only time-consuming but often inaccessible in low-resource settings. A new study published in the journal Healthcare shows the effectiveness of artificial intelligence in accurately predicting asthma based on vocal pitch variations, offering a fast, non-invasive alternative to conventional diagnostic techniques. Researchers from National Dong Hwa University and the National Taipei University of Nursing and Health Sciences developed a machine learning-based system that identifies respiratory irregularities linked to asthma through voice analysis, reaching predictive accuracies as high as 98.7%.

This study "AI-Driven Data Analysis for Asthma Risk Prediction" proposes a streamlined solution using AI to analyze subtle changes in voice caused by laryngeal swelling and airway restriction, conditions that frequently accompany asthma. By leveraging acoustic biomarkers, the researchers built a system that processes and classifies voice data with high precision, delivering faster results with fewer clinical demands.

The research team collected 1500 vocal samples from healthy individuals and asthma patients, analyzing high-pitch, normal-pitch, and low-pitch recitations of phonemes [i, a, u]. These samples were sourced from the Saarbruecken Voice Database and converted into frequency spectra using a short-time Fourier transform. The study extracted relevant features using Long-Term Average Spectrum (LTAS) and Mel-Frequency Cepstral Coefficients (MFCCs), standard techniques in audio signal processing.

To ensure a rigorous analysis, seven supervised machine learning models were employed, including Decision Tree, Random Forest, Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). Each model was tested on its ability to classify the presence or absence of asthma based on acoustic input, with training performed on 80% of the dataset and the remaining 20% reserved for validation.

Results showed that model performance varied by vocal pitch. The Decision Tree model achieved a peak accuracy of 98.7% using high-pitch samples, while the LSTM model performed best on normal and low-pitch inputs, achieving accuracies of 76.9% and 85.6%, respectively. In terms of sensitivity and specificity, the Random Forest model excelled in high-pitch sensitivity (0.85), CNN led in normal-pitch sensitivity (0.85), and ANN produced the best results for low-pitch sensitivity (0.69). For F1-score, CNN outperformed other models in low-pitch scenarios (0.86), confirming its ability to balance precision and recall effectively.

To enhance the interpretability of the predictions, the study implemented Shapley Additive Explanations (SHAP) for feature importance analysis. This method enabled the identification of critical acoustic frequency bands contributing to the predictions. Notably, the frequency range between 450 and 500 Hz was found to significantly differentiate healthy individuals from asthma patients. These findings were cross-validated with the average spectral analysis of voice samples, supporting their clinical relevance and grounding the AI model's insights in otolaryngological evidence.

Advanced signal processing techniques were applied to denoise the samples before classification. Algorithms such as CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise), wavelet denoising, and Tunable Q-Factor Wavelet Transform (TQWT) were used to isolate and preserve informative acoustic features while suppressing background noise. This process ensured that only high-quality signals with a signal-to-noise ratio (SNR) above 15 were included in the model training.

The system also accounted for demographic and physiological variation by categorizing samples by age and gender. Participants were grouped into seven categories, ranging from children to older adults, enabling the machine learning models to generalize better across populations. Despite potential limitations due to variability in articulation and recording conditions, the model demonstrated robust generalization, highlighting its real-world applicability.

This study offers a major step forward in deploying AI for accessible healthcare solutions. By utilizing only speech recordings, the system eliminates the need for specialized equipment, reducing costs and making asthma screening more feasible in remote or under-resourced areas. Additionally, the explainability features built into the platform address one of the key concerns in clinical AI adoption - transparency of decision-making.

While these results are promising, further validation, as the researchers stress, is needed across larger and more diverse datasets to ensure broader applicability. They also propose integrating the system into mobile health applications, allowing patients to self-assess their risk using smartphone microphones and receiving instant feedback. Such developments could contribute significantly to early detection, timely treatment, and improved management of asthma, particularly among children and aging populations.