Comparison of classification methods in breath analysis by electronic nose

Area of Science:

• Biomedical Engineering
• Data Science
• Medical Diagnostics

Background:

• Electronic nose (eNose) technology utilizes breath analysis for medical diagnostics.
• Numerous pre-processing, statistical analysis, and validation methods exist for eNose data.
• A comprehensive comparison of these methods' impact on diagnostic performance is lacking.

Purpose of the Study:

• To empirically evaluate and compare the influence of various dimension reduction, classification, and validation methods on eNose diagnostic performance.
• To guide the selection of appropriate statistical methods in eNose research.
• To aid reviewers in assessing the methodological quality of eNose studies.

Main Methods:

• Literature review of human studies using eNose (PubMed, up to 2014).
• Methodological quality assessment using a tailored QUADAS-2 tool.
• Re-analysis of four independent eNose datasets using published statistical methods.
• Comparison of diagnostic performance (ROC-AUC) using in-set, internal, and external validation.

Main Results:

• High risk of bias was observed in most reviewed studies, often due to non-random patient selection.
• Internal validation consistently decreased performance (ROC-AUC) compared to in-set calculations.
• External validation revealed performance decreases in two datasets and slight increases in two others, highlighting variability.
• No single combination of methods consistently performed well across both internal and external validation sets.

Conclusions:

• Estimating diagnostic performance solely on training data, even with internal validation, is unreliable for eNose studies.
• External validation using independent datasets is critical for robustly assessing the true diagnostic performance of eNose technology.
• Future eNose projects in medicine must incorporate external validation to ensure reliable and generalizable results.