Quantitative method comparison in non-targeted analysis using cascade classification.

Non-targeted analysis (NTA) lacks consistent processes for comparing method performance across different analytical conditions. As a result, method comparison between NTA methods is often dependent on the chemicals detected in a particular study, making it difficult to objectively evaluate how instrumental performance influences the overall detectable space. Here, we describe a two-stage cascade classification approach that treats detectability as a binary outcome defined by experimental retention behavior and signal-to-noise criteria and characterizes the boundary between detectable and non-detectable chemical space using receiver operating characteristic (ROC) analysis. In Stage 1, compounds are filtered by estimated retention index to establish chromatographic accessibility. In Stage 2, detectability is classified using the electron impact cross-section (Q), with the classification threshold selected to satisfy a minimum recall constraint (recall ≥ 0.80). Using this treatment for the detectable space, the area under the ROC curve (AUC) emerges as a method-level figure of merit. Applied to two GC-MS systems operating under identical chromatographic conditions but with different mass selective detectors (Agilent 6890/5975B and 8890/5977C), the approach yielded Stage 2 AUC values of 0.659 (95% CI: 0.536-0.788) for Method A and 0.842 (95% CI: 0.702-0.933) for Method B at 10 μg/mL, quantitatively capturing the performance difference between instruments. The minimum-recall criterion selected near-identical classification thresholds for both methods (τ = 19.18 and 19.30, respectively), indicating that Q estimates essentially the same detectable chemical space for both instruments, and that the observed difference in detectable space reflects instrument performance. This approach provides a reproducible and transparent basis for GC-MS method comparison in NTA and refines the statistical basis needed for a generalized analytical evaluation of NTA methods.