Overcoming methodological barriers in electronic nose clinical studies, a simulation data-based approach.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Oncology

Background:

• Non-invasive screening for neoplasia using volatile organic compounds (VOCs) analyzed by electronic nose (e-nose) technology shows promise.
• Limited guidance exists for clinical study design, particularly concerning machine learning's impact on sample size requirements.

Purpose of the Study:

• To evaluate how neoplasia prevalence, data augmentation, and the number of e-nose devices influence sample size needs for accurate prediction models.
• To provide data-driven insights for optimizing e-nose study designs in clinical settings.

Main Methods:

• Simulated e-nose breath test data based on real-world data.
• Analysis of varying neoplasia prevalence (5%-50%) and data augmentation effects on model performance.
• Development of prediction models using single value decomposition, random forest, and convolutional neural networks.

Main Results:

• Lower neoplasia prevalence drastically increases sample size requirements; 5% prevalence necessitates up to five times more data than 50% prevalence for stable model performance.
• Model performance varied across different e-nose devices, with multi-device integration requiring larger sample sizes.
• Approximately 400 data points per device at 50% prevalence and 2100 data points at 5% prevalence were needed for stable performance.

Conclusions:

• Neoplasia prevalence and the number of e-nose devices are critical factors determining sample size requirements.
• Minimizing device variability and ensuring adequate case/control samples per device are essential for reliable predictive performance in e-nose studies.