Structural nested mean models (SNMMs) are a principled approach to estimate the treatment effects over time. A particular strength of SNMMs is to break the joint effect of treatment sequences over time into localized, time-specific ``blip effects''. This decomposition promotes interpretability through the incremental effects and enables the efficient offline evaluation of optimal treatment policies without re-computation. However, neural frameworks for SNMMs are lacking, as their inherently sequential g-estimation scheme prevents end-to-end, gradient-based training. Here, we propose DeepBlip, the first neural framework for SNMMs, which overcomes this limitation with a novel double optimization trick to enable simultaneous learning of all blip functions. Our DeepBlip seamlessly integrates sequential neural networks like LSTMs or transformers to capture complex temporal dependencies. By design, our method correctly adjusts for time-varying confounding to produce unbiased estimates, and its Neyman-orthogonal loss function ensures robustness to nuisance model misspecification. Finally, we evaluate our DeepBlip across various clinical datasets, where it achieves state-of-the-art performance.

Lay Summary Problem. Doctors caring for a patient over days or weeks face a sequence of treatment decisions, and the best choice depends on the patient's history. We'd like AI to learn from past medical records what works for whom — but there's a trap. In real data, sicker patients get more aggressive treatment, and earlier treatments change how sick a patient becomes later. This creates the confounding phenomenon, and misleads naive AI into mistaking correlation for cause. Solution. Statisticians long ago proposed a fix: instead of predicting the outcome of a whole treatment plan at once, break it into small incremental "blip" effects — what each single decision contributes on its own. We invented a training trick that lets a neural network learn all the blips at once, in a single pass, while still respecting the original method's logic. Hence, we substantially scale the statistical method. Impact. Our model, DeepBlip, makes more reliable predictions further into the future than existing approaches. Once trained, it can instantly compare many possible treatment plans for an individual patient — opening the door to faster, more personalized decision support in medicine.