A vaping classifier can decline to label a case: read coverage with accuracy
Keep withheld predictions separate from negative results when a model reports performance only for accepted cases.
No Vape · Updated

Identify whether the system has a third operational state
A binary label system can still have an additional decision about whether to return any label. Look for terms such as abstain, reject or defer, and read what they mean in the report. A label that was withheld differs from a negative label, a missing input or a failed download. Do not collapse those states into one empty cell.
Pugnana and Ruggieri’s original 2023 methods abstract describes selective classification as pairing prediction with a selection function and trading coverage against performance. That abstract supports the distinction used here. Its algorithm, ranking results and other application examples do not validate any vaping classifier or quitting app.
Keep two denominators on the page
Write all eligible evaluated inputs and the subset that received a prediction. Coverage describes the accepted share under the report’s rule. Then identify which performance measure applies to that subset. Accuracy, AUC and other metrics answer different questions, so copy the reported measure rather than renaming every favorable value accuracy.
For fictional arithmetic, a system receives 100 eligible windows, labels 60 and withholds labels for 40. If 54 of the 60 labels match the reference, accepted-case accuracy is 90 percent and coverage is 60 percent. Calling all 100 windows 90-percent accurately labeled would ignore the 40 without labels. The numbers are invented and do not describe an actual device.
Find what happens after a withheld prediction
A paper may evaluate abstention without providing a practical follow-up process. Do not infer that a human reviews the case, that another model takes over or that the case can safely be ignored. If a product claim promises a complete log, ask how that promise accounts for unclassified inputs using functions actually documented.
Changing the selection rule can change which cases remain evaluated, so compare performance at the stated coverage and conditions. A stronger-looking accepted-case metric may accompany fewer answered cases. That is a tradeoff to describe, not proof of deception or a rule that maximizing either number is always best. This guide does not choose a threshold.
Preserve uncertainty outside the accepted subset
Your reader note can state the original unit and reference, acceptance rule, coverage and reported performance among accepted cases. Add the stated handling of withheld cases, or mark it unverified. Avoid assigning those cases an outcome merely to complete a chart or calculate a preferred overall score.
Classification abstention is unrelated to a person’s decision to abstain from vaping. It is a software evaluation term, not a quitting milestone. This exercise performs no model audit, clinical interpretation or app endorsement. A research system’s decision to withhold labels cannot establish that personal records are complete, exposure is absent or using the system improves health.
Sources and scope
Sources inform the guidance. The worksheets and examples are No Vape editorial suggestions, not tested treatments.
Written with AI assistance and checked against the cited sources. No clinical review is claimed. This guide covers practical planning, not individual medical treatment. How we research and write