Improving the Realism of Synthetic Clinical Benchmarks Under Utility Constraints

Synthetic clinical benchmarks for enterprise AI agents can pass existing utility checks and still remain structurally unrealistic, especially in privacy-sensitive healthcare settings where operational data are hard to access. We study how to improve such benchmarks without breaking the downstream utility checks already used in practice. We formulate benchmark revision as utility-constrained realism improvement: dataset changes should increase realism while staying above an operational utility floor. We instantiate this idea on a care-gap benchmark derived from Synthea-generated patients exercised through demonstration electronic health record workflows and then processed by the same downstream pipeline as operational data. Realism is measured through missingness structure, simplicity, structural plausibility, and population alignment. The baseline benchmark is extremely thin: sampled-pair missingness is 79.44%, only 12.75% of rows are actionable, 38.94% of patients have zero actionable measures, and top-three token concentration reaches 100.0%. Two deterministic revisions improve these panels while remaining above the current utility floor, whereas a naive densification control preserves unrealistic templating. We further show that internal benchmark realism and source fidelity to an aggregate operational reference are related but distinct objectives. These results suggest that synthetic benchmark quality should be optimized explicitly, with utility treated as one constraint rather than as sufficient evidence of realism.

CanvOI, an Oncology Intelligence Foundation Model: Scaling FLOPS Differently

The rapidly evolving field of digital oncopathology faces significant challenges, including the need to address diverse and complex clinical questions, often involving rare conditions, with limited availability of labeled data. These limitations hinder the development of robust AI-driven tools in the biomedical space, where accuracy in probabilistic determinations is of utmost importance. To address this, digital pathology foundation models have begun to emerge, typically developed with the size and diversity of the pre-training dataset and model parameters in mind. Here, we present CanvOI, a ViT-g/10-based foundation model designed to enhance the capabilities of digital pathology by addressing these challenges through a different approach. Considering the unique nature of oncologic histopathological images and the requirements from the embeddings to provide meaningful representations for Multiple Instance Learning (MIL) downstream models, we chose to modify the input image characteristics. By introducing larger tile sizes (380 x 380 pixels) and smaller patch sizes (10 x 10 pixels), we were able to optimize the model's performance, pushing computational resources in a new direction and achieving state-of-the-art performance on cancer-related benchmarks. CanvOI demonstrated a 1.5-7.4% improvement in averaged AUC compared to other leading foundation models built for digital pathology. Moreover, our results demonstrate that CanvOI significantly outperformed the other models, with the performance gap widening substantially when trained on just 10% of the initial cohort. This work highlights an alternative approach that, if integrated with traditional development approaches, has the potential to advance Oncology Intelligence (OI), overcome some of the current barriers and ultimately improve the clinical outcome of cancer patients.