Few-Shot Lung Cancer Classification via Electronic Nose Using Large Language Models: A Multicentre Prospective Study

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Electronic Nose (eNose) breathprints show potential for non-invasive lung cancer diagnosis.
• Cross-site validation and adaptation are key challenges for clinical use of eNose technology.
• The efficacy of large language models (LLMs) for few-shot, site-specific lung cancer classification using eNose data is unexplored.

Purpose of the Study:

• To investigate if a natural language processing-pretrained LLM can enable few-shot, site-specific classification of lung cancer from eNose breathprints.
• To compare the performance of an LLM against traditional Convolutional Neural Networks (CNNs) in this task.

Main Methods:

• Collected 432 eNose breathprints from lung cancer and non-lung cancer patients across two medical centers in Taiwan.
• Utilized a GPT-2-backbone LLM with parameter-efficient adaptation.
• Compared LLM performance with CNNs trained from scratch or pretrained on CIFAR-100, using few-shot (2-6 samples/class) and full-data protocols.

Main Results:

• With 6 shots, the LLM achieved an AUC of 0.79 on Site 1 and 0.76 on Site 2.
• The LLM significantly outperformed both scratch CNNs and CIFAR-100 pretrained CNNs on both sites (p < 0.05).
• Transfer learning to a new site did not improve performance for either LLM or CNN models, with LLM performance even slightly decreasing.

Conclusions:

• Pretrained LLMs demonstrate potential for few-shot lung cancer classification using eNose breathprints in mixed clinical settings.
• This approach can reduce the need for extensive training datasets in developing non-invasive lung cancer diagnostics.
• Further research is needed to address cross-site adaptation challenges for robust clinical deployment.