Generating Counterfactual Patient Timelines from Real-World Data

A research team from Japan has published a groundbreaking paper on arXiv demonstrating that AI can generate realistic 'what-if' scenarios for patient outcomes. The team, led by Yu Akagi, trained an autoregressive generative model on a massive dataset of real-world electronic health records. This dataset encompassed over 300,000 patients and a staggering 400 million individual patient timeline entries, allowing the model to learn the complex, sequential nature of medical events.

To validate the model's clinical plausibility, the researchers applied it to a cohort of patients hospitalized with COVID-19 in 2023. They performed counterfactual simulations by modifying key patient variables like age, serum C-reactive protein (CRP), and serum creatinine to project 7-day outcomes. The model's simulations successfully reproduced established medical knowledge: it showed increased in-hospital mortality for older patients and those with elevated CRP or creatinine, and it adjusted remdesivir prescription rates logically based on kidney function and inflammation levels.

This work is significant because it moves beyond simple prediction to enable causal exploration. The self-supervised training on real-world data, rather than curated trial data, makes the approach highly scalable. The authors posit that this methodology can establish a foundational model for counterfactual clinical simulation, opening the door to applications like in silico trials for drug safety and personalized treatment planning by testing interventions on a digital twin of a patient.