Causal AI for cancer immunotherapy: a narrative framework review of target trial emulation, treatment-effect learning and clinical translation.

This narrative framework Review examines how artificial intelligence (AI) can move cancer immunotherapy research from outcome prediction toward target-trial-based treatment-effect learning. AI has produced increasingly accurate models for predicting response, survival and immune-related toxicity during cancer immunotherapy. Yet most models estimate outcome risk under observed care rather than the causal effect of choosing one strategy over another. We therefore frame immunotherapy AI as a causal digital-medicine problem: clinically useful AI should begin with a target-trial question that specifies eligibility, time zero, treatment strategies, comparators, outcomes, estimands and bias-control plans before model development. Within this framework, multimodal AI outputs from imaging, digital pathology, omics, microbiome data, electronic health records and clinical text can function as baseline confounders, candidate effect modifiers, longitudinal state variables or outcome-ascertainment tools. We distinguish established causal-inference approaches, such as target trial emulation, propensity-score weighting, g-methods, TMLE, DML and heterogeneous-treatment-effect estimation, from exploratory technologies such as reinforcement learning and digital twins that require prospective safety validation. We close by outlining validation, reporting, workflow, regulatory and lifecycle-monitoring requirements for moving from predictive biomarkers to trustworthy causal learning systems in immuno-oncology.