The data that biomedical AI still needs

Key points

• Across every layer of biomedical AI, the limited availability of longitudinal, multimodal, representative health data poses the greatest constraint on discovery and robust, clinically reliable performance.
• Much of the most valuable data already exists but is discarded or never captured; the first opportunity, therefore, is to unlock and reuse what health systems already produce.
• Routine data are intermittent and coarse; the next gains will come from data collected by design, such as those from wearables, remote monitors, and ambient sensing. These data capture the continuous, pre-symptomatic signal that prediction and prevention require.

In their lead article, Guni et al. present an ambitious, comprehensive map of artificial intelligence (AI) in medicine (1). They trace the field through successive epochs, from symbolic systems to deep learning, foundation models, and emerging agentic AI; propose frameworks for AI’s clinical functions and degrees of autonomy; and survey regulator-approved devices and the rise of generative and multimodal models. Furthermore, they chart a parallel revolution in the life sciences, spanning protein-structure prediction, genomics, and drug discovery, and set out a future of multimodal, precision, and ultimately pre-disease medicine.

Beneath this breadth lies a recurring dependency. Across every layer they describe, including diagnostic devices, multimodal “navigator” models, precision therapeutics, and pre-disease prediction, one indispensable input remains the same: large-scale, longitudinal, multimodal health data. The authors rightly list datasets as one of six fundamental domains requiring sustained effort going forward. Of these, one could argue that data constitute the factor most likely to limit novel discovery and the robustness and clinical reliability of medical AI. One could further argue that the health system is the only engine capable of producing such data at the scale, breadth, and quality required. This viewpoint focuses on this domain, not because the other domains matter less, but because even the most capable models and infrastructure deliver little without data that are rich, abundant, representative, and trustworthy. Recognizing this clarifies both where the promise of medical AI lies and where, today, it is most constrained.

Unlocking the data we already generate

Consider what Guni et al.’s most advanced examples have in common: diagnostic and predictive devices learn from labeled images and outcomes; multimodal generalist models integrate text, imaging, and signals; and agentic systems coordinate tasks across a patient’s record. Few of these applications are currently limited by model architecture; rather, they are limited by access to data that are representative, linked across modalities, and tied to real outcomes (2). Many of these data already exist. Some are generated as part of routine care; however, these are rarely reused in their richest form. Examples include the raw electrocardiogram (ECG) waveform rather than its one-line interpretation, or the whole-slide image behind a pathology report. Other data streams are produced by equipment already in use but are seldom retained at all: the synchronized video and electroencephalograms (EEGs) recorded during epilepsy monitoring; endoscopic and in-room camera feeds from operating rooms; and high-frequency telemetry of robotic surgical systems, including instrument kinematics, forces, and event logs. Each is generated in real time, used for the task at hand, and then often discarded, although it encodes far more than the summary that ultimately reaches the record. The first opportunity is therefore less about generating new data than it is about unlocking and linking what the health system already produces but discards, silos within departments, or holds in systems that do not interoperate. When this is achieved, the self-reinforcing cycle of a learning health data system (3) becomes more feasible: care generates data, data train models, models improve care, and improved care generates better data.

These same data also reach upstream into the life sciences, where routine and population-scale records are beginning to reshape drug discovery. Biobanks that link genomic, multiomic, and biomarker data to clinical records have powered case-control genetic discoveries that can nominate new therapeutic targets (4), while models trained on linked clinical and molecular data can suggest drug-repurposing candidates for thousands of diseases, including those with few existing treatments (5). The molecular substrate itself is also deepening: genomic sequencing is spreading well beyond oncology into rare-disease diagnosis, pharmacogenomics, and inherited-risk prediction, and it is increasingly joined by other layers, such as transcriptomic, proteomic, and metabolomic data, that are becoming faster and cheaper to generate. Each layer adds a complementary view of disease biology; therefore, linking these richer molecular readouts to clinical phenotypes could push target discovery and mechanistic understanding beyond what genomics alone can achieve.

Not yet the data we need

Unlocking existing data is necessary, but it is not sufficient because the data we generate are not yet the data we need. They are collected unevenly across individuals and places, and models trained on these data inherit and amplify those gaps; AI applied to underserved populations can systematically underdiagnose them (6). Much of the literature still relies on single-center development, which inflates apparent performance and obscures how a tool behaves elsewhere. Representativeness, moreover, begins upstream of any dataset: populations that face barriers to care are underrepresented in the data precisely because they are underrepresented in the system that generates them. Thus, widening access to care is itself a precondition for better data. Bias, in turn, cannot be corrected if it cannot be seen, which argues for routinely capturing social determinants of health alongside clinical variables and for adopting emerging standards for dataset diversity and transparency (7). This would make disparities visible and measurable rather than latent. The goal, therefore, is to generate more representative data that reflect the populations medicine actually serves.

Proving value in deployment

Deployment is where this cycle either turns or stalls and where the field’s persistent gap between technical capability and demonstrated clinical benefit comes into focus. The majority of AI systems are still evaluated retrospectively, on curated benchmarks, which neither prove clinical value nor produce reusable evidence. The metrics themselves are also too narrow: a high area under the receiver operating characteristic (ROC) curve says nothing about whether a tool changes decisions, improves outcomes, reduces time or cost, or is even accepted by the clinicians and patients meant to use it, which is why involving those end-users in how a tool is deployed and judged matters as much as its discrimination. Evaluation should, therefore, reach beyond discrimination to the endpoints that determine real benefit: clinical outcomes, cost-effectiveness, broader health economic impact, workflow efficiency, and clinician and patient experience. Deployment can be designed to generate exactly this type of evidence. Silent or shadow deployment lets a model run alongside care and be monitored without taking action (8), and prospective, multisite studies anchored to these endpoints turn everyday practice into a source of trustworthy data (9). Because the substrate and the populations it represents keep changing, these evaluations cannot be one-time events; models must be monitored and periodically revalidated as the underlying data drift.

The limits of routine data

There is, however, a ceiling to what routine care data can deliver. In discussing processing power, Guni et al. (1) observe that text-only training is “intrinsically low-density,” that further scaling of text yields diminishing returns, and that progress will require shifting toward higher-bandwidth data such as continuous physiological monitoring, surgical video, and longitudinal imaging, sources they suggest may yield more clinically relevant models than text-based corpora. However, this barely scratches the surface of the need. The limitation is not only one of modality bandwidth but also of resolution and coverage. Much of what we record is compressed into coarse, ordinal scores, such as pain scores, which flatten the rich, continuous signals underlying them. These signals include skin tone, facial movement, and gait kinematics, which clinical practitioners summarize as pallor, affect, or mobility. An audiovisual modality would capture these signals directly. Furthermore, patients are observed only intermittently during routine care: most clinically important changes in health happen between visits, in the long, unmonitored intervals when deterioration and recovery actually unfold. Data generated passively, as a byproduct of documentation and billing, can capture neither the resolution nor the continuity that prediction requires. Raising the “intelligence per byte” of medical AI, therefore, depends less on continuing to scale models using already saturated health-record data, such as codes—a process that brings rising computational and energy costs for diminishing returns—and more on higher-signal data. Some of that signal can be recovered from what is already collected but discarded or summarized away, such as when longitudinal imaging or raw waveforms are used in their full form; the rest must be gathered by design. The most efficient route to better models is therefore better data that carry more clinical value per unit of energy.

Collecting what we are missing

This points to a conclusion that Guni et al. gesture toward: routine multimodal data are the right place to start but are not sufficient for biomedical AI. The next increment of clinical value will come from adjunctive data that the health system is deliberately instrumented to collect. The authors’ own clinical decision-support vision is full of candidates: implantable cardiac monitors, continuous glucose monitors, and other biosensors; wearables that track vital signs and sleep; in-home and remote monitoring systems (10); and smartphone or in-home cameras that could enable the noninvasive “video biopsy” they envisage. In this approach, optical and motion-derived digital biomarkers support assessment at a distance. These data come with new challenges: they are harder and costlier to collect and raise real privacy and safety concerns; however, they are precisely the kind of data that carry the pre-symptomatic, longitudinal signal on which the authors’ aspiration, to shift care “upstream toward prediction and prevention,” ultimately depends. Closing the distance between the capability and benefit of biomedical AI will require building the means to gather the data we do not yet have, alongside making fuller and more effective use of the data we already do.

Generative AI statement

The author declared that generative AI was used in the creation of this manuscript. An AI assistant (Claude, Anthropic) was used to help draft and edit the text. All content, claims, and references were reviewed, verified, and approved by the author, who takes full responsibility for the work.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited and reviewed by: Karin Verspoor, RMIT University, Australia