Advanced breath analysis through hierarchical deep convolutional neural network for multi-cancer screening

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Breath analysis offers a noninvasive method for early cancer detection by identifying volatile organic compound (VOC) signatures.
• Existing methods require further development for accurate multi-cancer classification.

Purpose of the Study:

• To develop and validate a hierarchical deep convolutional neural network (HD-CNN) platform for dual-cancer classification using breath analysis.
• To assess the performance of multimodal gas sensor arrays in distinguishing between healthy controls, lung cancer, and gastric cancer patients.

Main Methods:

• Collected breath samples from 206 participants (67 healthy controls, 78 lung cancer, 61 gastric cancer).
• Utilized a multimodal gas sensor array (SMO, EC, PID) to generate 2D response maps from time-resolved signals.
• Developed a two-stage HD-CNN model for coarse classification (healthy vs. cancer) and fine classification (lung vs. gastric cancer).

Main Results:

• The HD-CNN achieved high classification accuracies: 82.1% (healthy), 84.0% (lung cancer), and 88.1% (gastric cancer).
• Average AUCs were 0.89 (healthy), 0.92 (lung cancer), and 0.89 (gastric cancer).
• The HD-CNN outperformed a 1D CNN, showing improved class separability and prediction confidence.

Conclusions:

• Hierarchical learning and multimodal sensing are effective for robust breath-based multi-cancer screening.
• The developed HD-CNN platform demonstrates significant potential for early and accurate cancer detection through breath analysis.
• Optimizing the hierarchical structure, particularly by first isolating healthy controls, enhances overall diagnostic performance.