How Foundation Models Are Reshaping Healthcare
By Eugene Osei Afriyie
For decades, artificial intelligence in healthcare consisted largely of siloed experiments. An algorithm could detect a tumor in an X-ray, or predict sepsis risk from vital signs, but it could do nothing else. These were “narrow” tools—highly effective at single tasks, but largely non-transferable.
That paradigm is rapidly changing. We have entered the era of foundation models (FMs).
Unlike their predecessors, which were trained on labeled datasets for specific tasks, foundation models are trained on vast amounts of unstructured data—text, images, genomic sequences, and chemical structures. They are adaptable, largely self-supervised systems that can be fine-tuned for a wide range of downstream applications. (Bommasani et al., 2021).
In the technology sector, this shift is viewed as an infrastructure upgrade akin to the arrival of cloud computing. But in healthcare, the implications are more profound: we are moving from algorithms that simply categorizemedical data to systems that can reason with the fundamental languages of biology and medicine.
The Architecture of Intelligence
To understand the impact on healthcare, one must first understand the shift in architecture. Foundation models, often built on Transformer architectures such as those underlying GPT-4, learn statistical regularities within large, structured datasets.
When trained on the internet, they learn human language. When trained on biological data, they learn the statistical structure of biological sequences and signals. As Moor et al. (2023) note, modern medicine is increasingly multimodal; a patient’s profile consists of clinical notes (text), MRI scans (vision), and genetic markers (sequences). Foundation models are currently the most scalable approach for learning shared representations across these disparate data modalities.
The New Clinical Partner
The application of these models is currently bifurcating into two distinct streams: operational efficiency and biological discovery.
1. The Death of “Pajama Time”
The most immediate impact is administrative. Multiple studies suggest that physicians spend close to two hours on electronic health records (EHRs) for every hour of direct patient care, a phenomenon contributing to record levels of burnout.
New systems integrated by major health networks utilize foundation models to “listen” to patient consultations. Unlike old dictation software, these models extract relevant medical history, filter out small talk, and generate structured clinical notes and billing codes automatically. These applications are designed not to replace clinicians, but to reduce administrative burden and reclaim time for patient interaction (Thirunavukarasu et al., 2023).
2. Decoding the Language of Biology
The more radical disruption is occurring in the laboratory. Just as Large Language Models (LLMs) predict the next word in a sentence, biological foundation models predict the next amino acid in a protein or the binding affinity of a molecule.
This approach treats biological sequences as structured symbol systems amenable to statistical modeling. AlphaFold, a system developed by Google DeepMind, demonstrated high-accuracy prediction of static protein structures across a large fraction of known proteins—a feat that would have taken millennia using traditional experimental methods. This capability is compressing the timeline of drug discovery, allowing pharmaceutical companies to generate novel molecular structures that have a higher probability of success before they ever enter a wet lab (Jumper et al., 2021).
The Hallucination Hazard
However, the integration of foundation models into clinical workflows is not without peril. One of the defining limitations of these models is so-called “hallucination”—the generation of plausible-sounding but factually incorrect information.
In a creative writing context, a hallucination is a quirk; in an oncology ward, it is a liability. There is also the risk of ingrained bias. If a model is trained predominantly on data from specific demographics, its outputs may exhibit degraded performance or systematic bias when applied to underrepresented populations. As noted by Singhal et al. (2023) regarding Google’s Med-PaLM, rigorous reinforcement learning, and human oversight are required to align these models with medical safety standards.
The Economic Imperative
Despite the risks, the economic gravity is undeniable. Healthcare systems across the OECD are facing aging populations and shrinking workforces. The “Baumol effect”—where wages rise in sectors with stagnant productivity—has plagued medicine for decades. Foundation models may represent one of the first credible opportunities to alter this cost trajectory.
We are witnessing the transition of AI from experimental diagnostics toward infrastructure-level clinical and operational systems. The question is no longer whether AI will be used in medicine, but how quickly we can build the guardrails to ensure it follows the first rule of the profession.
References
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., … & Liang, P. (2021). On the opportunities and risks of foundation models. Stanford Center for Research on Foundation Models. https://arxiv.org/abs/2108.07258
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., … & Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583-589. https://doi.org/10.1038/s41586-021-03819-2
Moor, M., Banerjee, O., Abad, Z. S. H., Krumholz, H. M., Leskovec, J., Topol, E. J., & Rajpurkar, P. (2023). Foundation models for generalist medical artificial intelligence. Nature, 616(7956), 259-265. https://doi.org/10.1038/s41586-023-05881-4
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., … & Natarajan, V. (2023). Large language models encode clinical knowledge. Nature, 620(7972), 172-180. https://doi.org/10.1038/s41586-023-06291-2
Thirunavukarasu, A. J., Ting, D. S. J., Elangovan, K., Gutierrez, L., Tan, T. F., & Ting, D. S. W. (2023). Large language models in medicine. Nature Medicine, 29(8), 1930-1940. https://doi.org/10.1038/s41591-023-02448-8
