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

Clinical decision-support systems for managing long-term treatment must decide not just which treatment is most suitable, but also when new evidence is sufficient to alter the course of care. This is particularly challenging when discontinuation is irreversible, as early cessation can prevent future information from influencing treatment decisions.

Food allergies, including peanut allergies, are a growing public health concern that can lead to severe reactions and significant medical, dietary, and psychological burdens. Preventing peanut allergies is a major clinical priority. In the case of peanut allergies, once consumption is resumed after a break, it may be unsafe without a new clinical assessment.

This study introduces a framework for predicting the preferred treatment and evaluates it using data from 628 children who were either given early peanut consumption or avoided peanuts in the Learning Early About Peanut Allergy trial. Initially, the system struggled to distinguish between the two preferred treatments, but as more clinical and immunological measurements became available, the predictions became more accurate, with history-aware models achieving an area under the curve of approximately 0.9 later in the study.

However, if an irreversible discontinuation is imposed after training, the models' performance drops significantly. To address this issue, the researchers developed TreatNet, a method that jointly optimizes visit-specific predictors through a trajectory-level objective that considers the downstream effects of early discontinuation.

When constrained by the irreversible discontinuation, TreatNet achieved accuracies between 0.805 and 0.839 and area under the curve values between 0.819 and 0.869 across different follow-up visits and two alternative definitions of the preferred treatment. TreatNet generally recommended continued consumption when baseline evidence was limited and only suggested discontinuation when more informative measurements were available.

These results demonstrate the importance of incorporating longitudinal information into decision-support systems, especially when the decision-making process involves irreversible or difficult-to-reverse treatment changes. Although this study focused on peanut allergy prevention, the framework developed could be applied to various clinical scenarios where treatment decisions have long-term consequences.