Longitudinal treatment selection under irreversible discontinuation: History-aware prediction and trajectory-aware learning in peanut-allergy prevention

Abstract

Clinical decision-support systems that guide longitudinal treatment must determine not only which treatment is preferred, but also when the available evidence is sufficient to change course. This challenge is especially important when discontinuation is irreversible, because stopping too early prevents later information from informing care. Food allergy is a growing public-health concern that can cause severe reactions and imposes substantial medical, dietary, and psychosocial burdens. Prevention is therefore a major clinical priority. In peanut-allergy prevention, this longitudinal decision is particularly consequential because resuming peanut consumption after discontinuation may be unsafe without renewed clinical assessment. In this work, we introduce a framework for preferred-treatment prediction and evaluate it using data from 628 children assigned to early peanut consumption or avoidance in the Learning Early About Peanut Allergy trial. We first introduced visit-wise and history-aware models that update recommendations as clinical and immunological measurements accumulate. Preferred treatment was poorly discriminated at baseline but became substantially more predictable during follow-up, with history-aware models achieving area-under-the-curve values of about 0.9 at later visits. However, imposing irreversible discontinuation after training significantly reduce the performance of these models. We therefore developed TreatNet, which jointly optimizes visit-specific predictors through a trajectory-level objective representing the downstream consequences of early discontinuation. Under the same irreversible constraint, TreatNet achieved accuracies of 0.805-0.839 and area-under-the-curve values of 0.819-0.869 across follow-up visits and two alternative preferred-treatment definitions. TreatNet generally preserved consumption while baseline evidence was limited and selectively recommended discontinuation after more informative measurements became available. These findings show that longitudinal information is most useful when model training reflects how current actions restrict future options. Although evaluated in peanut-allergy prevention, TreatNet provides a general framework for decision-support problems involving irreversible or difficult-to-reverse treatment changes.