Foundation Models for Multimodal Medical Imaging Triage: A Simulated Cross-Site Generalization and Calibration Study

This article presents a clearly labeled simulated empirical study spanning Computer Science and Engineering / Biomedical Engineering. A simulated multi-site imaging benchmark represented chest radiography, CT-derived summary features, and short clinical text from four hypothetical hospitals. A task-specific CNN, image foundation model, and multimodal vision-language foundation model were compared under internal validation and cross-site shift. Calibration, macro-F1, abstention accuracy, inference time, and subgroup stability were modeled. The multimodal foundation-model configuration achieved the strongest simulated cross-site macro-F1 and the lowest calibration error, but its computational burden was substantially higher than the task-specific baseline. Confidence- based abstention improved safety under the most severe site shift by routing uncertain cases for specialist review. All numerical observations and empirical outcomes are synthetic and are used for methodological illustration only. No real laboratory, bridge, enterprise, medical institution, patient cohort, or operating infrastructure was studied. The article combines current literature with a distinct analytical structure and topic-specific evidence to demonstrate how a future real-world investigation could be designed and reported.