Causal counterfactual simulation for treatment decisions in multimodal lung disease data

While deep learning has advanced the automated diagnosis of pulmonary diseases from chest X-rays, transitioning these models to treatment optimisation is fundamentally hindered by the reliance on biased observational data. In datasets, such as MIMIC, confounding factors create a "Treatment Paradox", where active interventions are statistically correlated with higher mortality simply because they are preferentially assigned to the most severely ill patients. Consequently, standard models learn clinical assignment patterns rather than true physiological efficacy. To address this, we propose a causal synthesis strategy that generates de-confounded, high-dimensional treatment counterfactuals. Guided by our experts' clinical priors, we construct a clinical transition matrix to simulate genuine physiological treatment effects. Using patient white blood cells as measure of severity, we applied a disease counterfactual model to generate the corresponding post-treatment images based on this matrix. Our longitudinal generative architecture uses both pre- and post-treatment chest X-rays to capture disease evolution over time. Furthermore, to resolve spatial misalignments caused by varying patient postures across temporal scans, we introduce an explicit lung mask constraint. This anatomy-aware mechanism confines counterfactual modifications to pulmonary regions, preventing clinically implausible alterations to background anatomy. By decoupling true treatment effects from baseline severity and posture variations, our framework successfully resolves the treatment paradox. Ultimately, this approach generates realistic treatment counterfactuals, providing clinicians with a trustworthy visual tool to evaluate patient-specific outcomes under alternative treatment scenarios and support medical decision-making.