Bang Zheng | Searcharxiv

arXiv2026

MEDSYN: Benchmarking Multi-EviDence SYNthesis in Complex Clinical Cases for Multimodal Large Language Models

Multimodal large language models (MLLMs) have shown great potential in medical applications, yet existing benchmarks inadequately capture real-world clinical complexity. We introduce MEDSYN, a multilingual, multimodal benchmark of highly complex clinical cases with up to 7 distinct visual clinical evidence (CE) types per case. Mirroring clinical workflow, we evaluate 18 MLLMs on differential diagnosis (DDx) generation and final diagnosis (FDx) selection. While top models often match or even outperform human experts on DDx generation, all MLLMs exhibit a much larger DDx--FDx performance gap compared to expert clinicians, indicating a failure mode in synthesis of heterogeneous CE types. Ablations attribute this failure to (i) overreliance on less discriminative textual CE ($\it{e.g.}$, medical history) and (ii) a cross-modal CE utilization gap. We introduce Evidence Sensitivity to quantify the latter and show that a smaller gap correlates with higher diagnostic accuracy. Finally, we demonstrate how it can be used to guide interventions to improve model performance. We will open-source our benchmark and code.

cs.CL↗

arXiv2023

Efficient estimation of weighted cumulative treatment effects by double/debiased machine learning

In empirical studies with time-to-event outcomes, investigators often leverage observational data to conduct causal inference on the effect of exposure when randomized controlled trial data is unavailable. Model misspecification and lack of overlap are common issues in observational studies, and they often lead to inconsistent and inefficient estimators of the average treatment effect. Estimators targeting overlap weighted effects have been proposed to address the challenge of poor overlap, and methods enabling flexible machine learning for nuisance models address model misspecification. However, the approaches that allow machine learning for nuisance models have not been extended to the setting of weighted average treatment effects for time-to-event outcomes when there is poor overlap. In this work, we propose a class of one-step cross-fitted double/debiased machine learning estimators for the weighted cumulative causal effect as a function of restriction time. We prove that the proposed estimators are consistent, asymptotically linear, and reach semiparametric efficiency bounds under regularity conditions. Our simulations show that the proposed estimators using nonparametric machine learning nuisance models perform as well as established methods that require correctly-specified parametric nuisance models, illustrating that our estimators mitigate the need for oracle parametric nuisance models. We apply the proposed methods to real-world observational data from a UK primary care database to compare the effects of anti-diabetic drugs on cancer clinical outcomes.

stat.ME↗