Counterfactual AI for Dynamic Dose Optimization with Side-Effect Constraints.

Area of Science:

• Artificial Intelligence in Medicine
• Causal Inference
• Machine Learning for Healthcare

Background:

• Dynamic treatment regimes require accurate patient-specific outcome prediction.
• Time-dependent confounding poses a significant challenge in continuous treatment settings.
• Existing AI frameworks struggle with continuous dosing and time-varying covariates.

Purpose of the Study:

• To introduce DoseAI, an online-updateable causal AI framework for continuous-time outcome prediction and dynamic dose optimization.
• To address time-dependent confounding in continuous treatment scenarios.
• To enable counterfactual prediction and optimize treatment trajectories under toxicity constraints.

Main Methods:

• Utilized a Neural Controlled Differential Equation (Neural CDE) encoder-decoder architecture for modeling patient-specific disease dynamics.
• Developed a novel training objective using Spearman correlation to penalize dependence between latent states and future doses, mitigating confounding.
• Evaluated the framework using synthetic lung cancer data and semi-synthetic trajectories from the MIMIC-IV database.

Main Results:

• DoseAI demonstrated accurate causal outcome prediction for continuous doses in continuous time.
• The framework successfully identified effective dosing strategies balancing therapeutic benefit and toxicity in synthetic cancer data.
• Validated the ability to handle both categorical and continuous dosing strategies under toxicity constraints.

Conclusions:

• DoseAI provides a robust causal AI framework for dynamic treatment optimization in continuous time.
• The proposed method effectively handles time-dependent confounding in complex healthcare scenarios.
• This AI approach holds promise for personalized medicine and improved clinical decision-making.