Synthesizing multidimensional clinical profiles from published Kaplan-Meier images

Clinical decision-making relies on understanding intersectional treatment effects across multiple patient characteristics. However, randomized controlled trials typically report one-dimensional marginal summaries, obscuring the underlying joint distributions of these characteristics. To address this, we developed MD-JoPiGo, a computational framework that reconstructs multidimensional clinical profiles from published 1D Kaplan-Meier curves. The approach utilizes the maximum entropy principle to estimate joint stratum frequencies and applies simulated annealing to generate individual-level data. We show that reconstruction fidelity depends on the underlying causal topology. Parallel predictors are resolved unconditionally, whereas interdependent structures require minimal structural priors to resolve unidentifiability. In evaluations using simulated data and empirical cohorts (lung cancer, n = 228; colon cancer, N = 929), the framework accurately recovered unobserved multivariable dynamics. Applied to fragmented and temporally misaligned reports from the CheckMate 227 trial, MD-JoPiGo reconstructed latent intersectional efficacy consistent with the clinical ground truth. By synthesizing multivariable evidence from 1D margins, this framework enables the secondary analysis of historical RCTs, supporting IPD meta-analyses and synthetic trial emulations.