Counterfactual AI for Dynamic Dose Optimization with Side-Effect Constraints

We present DoseAI, an online-updateable causal Artificial Intelligence (AI) framework for outcome prediction and dynamic dose optimization in continuous time. DoseAI models patient-specific disease dynamics using a Neural Controlled Differential Equation (Neural CDE) encoder-decoder architecture and enables counterfactual prediction under continuous treatment trajectories. A central challenge in this setting is time-dependent confounding, where time-varying covariates influence both treatment decisions and outcomes. To address this, we introduce a training objective for continuous treatments under time-dependent confounding that penalizes the dependence between the latent disease state and future dose trajectories via the absolute Spearman correlation, yielding a differentiable surrogate for representation balancing in continuous-dose settings. We evaluate DoseAI in controlled counterfactual scenarios using two complementary setups. First, synthetic lung cancer data generated from a mechanistic tumor-growth model provide ground-truth counterfactuals and enable quantitative assessment of dose optimization under chemotherapy and radiotherapy. Second, we construct semi-synthetic trajectories from the MIMIC-IV database by combining real patient covariates with simulated treatment effects, preserving realistic association structures while allowing controlled analysis of confounding. DoseAI supports both categorical and continuous dosing strategies and enables dynamic optimization under explicit toxicity constraints. Our results demonstrate that DoseAI can perform causal outcome prediction for continuous doses in continuous time and that, in the synthetic cancer setting, it discovers effective dosing strategies that balance therapeutic benefit and toxicity.