Can medical foundation models improve patient outcomes in real-world care?

Where do foundation models actually help patients?

The strongest evidence for improved outcomes comes from models trained on structured medical data—like electronic health records (EHRs) and medical images—for specific, well-defined prediction tasks. For instance, a foundation model that combined pathology slides and clinical data to predict breast cancer response to chemotherapy achieved 99.4% accuracy in identifying which patients would have a complete tumor response, and it could flag high-risk patients who might need more aggressive treatment [2]. Similarly, a lung cancer screening model that processed CT scans, text, and tables improved cancer risk prediction by up to 20% and cardiovascular mortality risk by 10% compared to existing state-of-the-art methods [3]. These models don't just guess—they provide interpretable heatmaps linking predictions to specific features in the tumor environment, giving doctors actionable insights.

Another promising area is time-to-event prediction from EHRs. The MOTOR model, pretrained on 55 million patient records (9 billion clinical events), improved prediction of when events like hospital readmission would occur by 4.6% over existing methods and required up to 95% less labeled data to achieve the same accuracy [4]. This means hospitals could deploy accurate risk models without needing massive, expensive datasets. The ARES system, built on a similar transformer architecture, outperformed traditional early warning scores (like NEWS) in predicting ICU admissions and prolonged hospital stays, and it provided personalized explanations for each risk estimate [5]. These tools could help doctors intervene earlier and more precisely.

What's the catch? Hallucinations and safety risks

The biggest barrier to real-world use is the problem of medical hallucinations—outputs that are factually wrong, logically inconsistent, or unsupported by evidence. A comprehensive evaluation of 11 foundation models (7 general-purpose, 4 medical-specialized) found that medical-specialized models actually performed worse, with only 28.6–61.9% of their responses being hallucination-free, compared to a median of 76.6% for general-purpose models [1]. Even the best general-purpose model, Gemini-2.5 Pro, needed chain-of-thought prompting to reach 97% accuracy; without it, it hallucinated 12.4% of the time. Physician audits revealed that 64–72% of residual hallucinations stemmed from errors in causal or temporal reasoning—not simple knowledge gaps—meaning the model gets the facts right but connects them incorrectly.

The real-world stakes are high: a global survey of 70 clinicians found that 91.8% had encountered medical hallucinations from AI, and 84.7% believed these errors could cause patient harm [1]. This means that while a model might ace a benchmark test, using it to guide treatment decisions without human verification could lead to dangerous mistakes. The irony is that models explicitly trained on medical data (like MedGemma) underperformed general-purpose models, suggesting that safety comes from broad reasoning capabilities developed during large-scale pre-training, not from narrow domain optimization [1]. For now, these models are best used as assistive tools—flagging risks or suggesting options—with a human doctor making the final call.

Who benefits most, and under what conditions?

The patients most likely to benefit are those with conditions where structured, high-quality data is available and the prediction task is clear-cut. For example, breast cancer patients undergoing chemotherapy can get personalized risk assessments from a model that integrates their pathology slides and clinical history [2]. Lung cancer screening participants can receive more accurate risk scores for both cancer and heart disease from a single model that analyzes their CT scans and medical records [3]. In hospital settings, patients could benefit from real-time risk monitoring systems like ARES, which dynamically updates risk estimates as new data comes in [5].

However, these benefits depend on several conditions. First, the model must be trained on data from a population similar to the patient—cross-site portability is still being tested, though MOTOR showed it could transfer across different hospital systems [4]. Second, the model must be used for its intended task: a model that excels at predicting chemotherapy response should not be asked to diagnose a rare disease. Third, human oversight is non-negotiable. The studies consistently show that even the best models make errors, and the consequences of those errors can be severe [1]. Finally, the model's predictions must be explainable—doctors need to understand why a model flagged a patient as high-risk to trust and act on that information [2][5]. Under these conditions, foundation models can be powerful tools for improving outcomes, but they are not yet ready to replace clinical judgment.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2023 to 2026, 4 from 2024 or later, collectively cited 106 times — selected as the most relevant from 6 studies that passed quality screening, drawn from 64 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions