Automatic Cough Analysis for Non-Small Cell Lung Cancer Detection.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Early detection of non-small cell lung cancer (NSCLC) is crucial for improving patient survival rates.
• Novel, non-invasive screening tools are needed to facilitate early NSCLC diagnosis.
• Automatic cough analysis presents a potential avenue for pre-screening lung cancer.

Purpose of the Study:

• To investigate the efficacy of automatic cough analysis using machine learning for distinguishing NSCLC patients from healthy controls.
• To compare the performance of various machine learning algorithms, including SVM, XGBoost, and CNNs, for cough-based NSCLC detection.
• To assess model interpretability and fairness across demographic groups.

Main Methods:

• Acquired cough audio recordings from 227 subjects (NSCLC patients and healthy controls).
• Applied machine learning techniques: Support Vector Machine (SVM), XGBoost, Convolutional Neural Networks (CNN), and transfer learning (VGG16).
• Utilized Shapley Additive Explanations (SHAP) for model interpretability and assessed fairness using equalized odds difference across age and gender.

Main Results:

• Convolutional Neural Networks (CNN) achieved the highest accuracy (0.83) on the test set.
• Support Vector Machine (SVM) demonstrated competitive performance (0.78 accuracy) and suitability for low-resource settings.
• SHAP analysis improved the transparency of the SVM model; fairness analysis indicated minor disparities across age and gender.

Conclusions:

• Automatic cough analysis, powered by machine learning, shows significant potential as a non-invasive tool for NSCLC pre-screening.
• CNNs offer superior performance, while SVM provides a viable alternative with enhanced interpretability.
• Further research with larger, diverse datasets is necessary to validate and refine these findings for clinical application.