For a medicine made from a fixed combination of herbs, listing the ingredients can be enough to suggest which formula it is. If a model trained on those lists performs well, has it learned information useful for other formulas, or recognized formulas it has already seen?
I examined this question using Japanese Kampo formulas and adverse-event reports. Kampo is a system of traditional medicine used in Japanese health care. I organized the modelling and validation to compare information about formula identity with information about its component herbs.
When ingredients become a name tag
Encoding each herb as a variable lets a model use composition without being given the formula’s name. Yet a distinctive combination may still identify a particular formula. Removing the name does not necessarily remove its identity from the data.
I therefore evaluated ingredient-based models against information identifying the formula itself. I compared the inputs on the same reporting data and distinguished handling known formulas from handling formulas absent from training. The point was to establish the scope in which the component representation could support a claim.
What reporting data can answer
Pharmacovigilance examines adverse events reported after medicines are used. The spontaneous-report data in this study do not cover every person who used a medicine. A reported association also does not establish that the medicine caused the event.
I kept discriminating among submitted reports and predicting the incidence of an adverse reaction as separate questions. Rather than ranking herbs or formulas by danger, I focused on what information the model used and prepared the reproducibility material and manuscript within that scope. This is not a validated safety-prediction tool for individual patients.
A technically amended manuscript was resubmitted on 1 September 2026. I received a rejection decision on 6 September (Korea Standard Time). Unpublished signals and performance figures remain private. Once there is a live version of record and an active DOI, I plan to explain the comparison design and its interpretive limits alongside the findings.