What evidence gaps are holding back medical foundation models?

Why do medical foundation models make dangerous mistakes?

The most critical evidence gap is the high rate of medical hallucinations — outputs that are factually incorrect, logically inconsistent, or unsupported by clinical evidence. A 2025 evaluation of 11 foundation models (7 general-purpose, 4 medical-specialized) across seven medical tasks found that general-purpose models produced hallucination-free responses a median of 76.6% of the time, while medical-specialized models managed only 51.3% — a 25-percentage-point gap [1]. Even more concerning, physician audits revealed that 64-72% of residual hallucinations stemmed from failures in causal or temporal reasoning, not from missing knowledge [1]. This means the models often get the facts right but connect them incorrectly, which is especially dangerous in clinical decision-making.

The real-world impact is substantial: in a global survey of 70 clinicians, 91.8% had encountered medical hallucinations, and 84.7% believed these errors could cause patient harm [1]. However, the study also found a promising mitigation: chain-of-thought prompting (where the model shows its reasoning step-by-step) reduced hallucinations in 86.4% of comparisons, and the best model (Gemini-2.5 Pro) reached 97% accuracy with this technique [1]. This suggests that the gap is not just about more data, but about building models that can reason explicitly and verify their own outputs.

Can we train good models without massive medical datasets?

A second major gap is the lack of large, high-quality, annotated medical datasets — collecting them is expensive, time-consuming, and raises privacy concerns. Two studies here tackle this from different angles. One approach used synthetic data: researchers pretrained a retinal foundation model on about 1 million synthetic retinal images, then fine-tuned it with only 16.7% of the real-world images required by a previous state-of-the-art model (RETFound). The resulting model performed as well or better across nine public datasets and four diagnostic tasks, and for diabetic-retinopathy grading it needed only 40% of the expert-annotated training data [4]. This suggests that synthetic data can partially bridge the data gap.

Another study harnessed publicly shared medical images from social media (medical Twitter) to create OpenPath, a dataset of 208,414 pathology images with natural language descriptions. Training a vision-language model on this data achieved state-of-the-art zero-shot classification, with F1 scores of 0.565-0.832 across four external datasets, compared to 0.030-0.481 for previous models [2]. Both approaches show that creative data sourcing and synthetic generation can reduce dependence on traditional annotated datasets, but they also highlight a persistent gap: these methods still require careful validation to ensure the data is representative and free from biases inherent in their sources.

How well do generalist models work on medical images without retraining?

A third evidence gap is the poor zero-shot performance of generalist vision foundation models on medical images. A 2023 study tested the Segment Anything Model (SAM) on nine medical image segmentation benchmarks covering modalities like OCT, MRI, and CT, as well as applications in dermatology, ophthalmology, and radiology. The results showed that while SAM performed well on general-domain images, its zero-shot segmentation was severely limited on medical images — and for certain structured targets like blood vessels, it failed completely [5]. However, fine-tuning with even a small amount of medical data led to remarkable improvements, indicating that the models have potential but require domain-specific adaptation that is not yet well understood or standardized [5].

This finding aligns with the broader challenge noted in a 2023 review: medical foundation models span a 'spectrum' from general imaging models to organ- or task-specific models, and each has different trade-offs in accuracy, data efficiency, and generalizability [3]. The evidence suggests that no single model architecture or training strategy currently works across all medical tasks, and the field lacks systematic benchmarks to compare approaches fairly.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2023 to 2025, 2 from 2024 or later, 3 in Q1 journals, collectively cited 879 times — selected as the most relevant from 7 studies that passed quality screening, drawn from 65 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions