Stanford University, working with Harvard, the University of Washington and other institutions, recently unveiled HealthFlux, the world's first pan-modal human health world model, a development widely seen across the industry as a landmark signal of the medical AI sector's next "AlphaFold moment."
This industrial revolution, moving from "single-disease prediction" toward "whole-health-state inference," strikes at a core pain point the field has long carried: medical AI can recognize more and more diseases, yet it has never established a holistic understanding of the "person." The key may lie not in the algorithms, but in the form of the data the models rely on.
A person's health is preserved in the manner of "item-by-item records": blood glucose, blood lipids, an image of one body part, a single diagnosis 鈥?each piece of data is generated to answer one specific question, and the pieces are not connected to one another. Their timing is also unevenly spaced; some indicators are tested every few months, while some images are captured only once every few years. Yet in the intervals with no records, the body does not stop changing. The overall state that truly determines a person's future disease risk always exists, but it is very hard to observe directly. Most existing models can only advance along these fragments, either using data from a single point in time or building a separate model for one predefined disease. Such models can make predictions for a specific disease accurate enough, but they struggle to answer a more fundamental question 鈥?what overall state is this person in, and where is that state heading?
The research led by the Stanford University team attempts to turn that hard-to-observe object into a computable state. Their HealthFlux seeks to learn a unified, latent human health state and to model how that state continuously evolves with age, disease and biological change. (The related paper was published on the preprint site medRxiv.)
When a model can predict health and disease risk with extreme accuracy, predict the changes after drug administration with extreme accuracy, and generalize well, it means the model may understand health to some degree. This is the judgment offered by the study's authors, and it is what the research calls the next "AlphaFold moment." It turns "understanding" from a philosophical question into a testable standard: predictions are accurate, post-administration changes can be calculated, and it still works in a different setting. Once this judgment is put forward, the question is no longer merely "can we predict a certain disease," but "is there a sufficiently general foundation that lets this predictive capability transfer to tasks it has never seen."
At the scale of imaging, the solution offered by DIAGENS-B (02526.HK) comes in two layers 鈥?first use a general foundation to fit images from different sources into one shared feature system, then use a data platform to feed clinical application feedback back into training.
Breaking imaging data fragmentation: a cross-modal foundation restructures the R&D efficiency system
Imaging faces another kind of fragmentation. CT reflects differences in how tissue absorbs X-rays, MRI comes from the signal of hydrogen protons under a magnetic field, ultrasound relies on sound wave reflection, and chromosome karyotyping comes from morphology and banding under microscopic imaging. The imaging principles differ, and the features they generate each form their own system. They record the body from different dimensions, but for a long time there has been no model able to fit these dimensions into a single representation. Every extension to a new modality or a new disease type requires going through data, annotation, training, evaluation and deployment all over again; previously accumulated capabilities are hard to reuse, and costs stay high.
The general medical imaging foundation model iMedImage built by DIAGENS-B aims to solve this reuse problem. It is not trained separately for any single type of image; instead, through cross-modal pretraining, it enables the model to form a shared feature representation, and subsequent specialty tasks start from this feature system. This feature system is not tied to any one disease, nor limited to any one modality; it describes the common structure of the images themselves. Specialty tasks perform discrimination on top of it and do not need to learn "how to read images" from scratch.
iMedImage is built around a 104-billion-parameter architecture and was pretrained on 8 million medical imaging records, covering 19 imaging modalities, 26 clinical specialties and more than 90% of imaging diagnostic scenarios, and supporting model development for over 400 diseases. The direct change this scale brings lands on the development of disease-specific models. On top of iMedImage, developing a vertical disease-specific model can use as few as about 200 imaging records, with a development cycle of 2 to 3 months. Compared with the traditional "one disease, one model" development approach, the demand for annotated data drops to one two-hundredth of the original, the development cycle shortens to one-twelfth, and both development costs and computing expenses fall to one-tenth. Disease-specific model development that previously only large top-tier hospitals and leading AI companies could afford is now becoming accessible to small and medium-sized medical institutions. To date, DIAGENS-B has reached cooperation with 99 hospitals and has landed 158 disease-specific model projects in total, covering 61 disease directions.
Positioning in the world model race: an imaging-side prototype has already landed first
DIAGENS-B's research in imaging is extending from judging the current state to judging future risk. A thymus study based on iMedImage and combined with radiomics features used preoperative CT to predict staging in 323 cases of thymic epithelial tumors, of which 258 were used for training and 65 for internal testing; the model achieved 95.38% test accuracy and an AUC of 0.9328 in distinguishing the early-stage and advanced-stage groups, with the results published in the European Journal of Radiology Open in July 2026 鈥?a judgment that originally required postoperative pathology was moved forward to before surgery.
Information from DIAGENS-B's interim results roadshow shows that the company's 673-case breast cancer recurrence prediction study incorporated pathology slides, ultrasound and clinical information into a single framework, and the optimized model achieved a C-index of 0.76 in five-fold cross-validation. The thymus study is learning "what disease state an image corresponds to," while the breast cancer study is exploring "what future risk the current features point to." The two studies happen to cover the two time points of clinical decision-making: before treatment begins and after treatment ends. They point to the same capability 鈥?turning information in images that has not yet been read out into usable grounds for judgment.
For these judgments to be continuously tested, they rely on DIAGENS-B's self-developed iMedLoop platform. The platform has brought together more than 300 professional annotators and 28.95 million imaging records, integrating data, tools, models and applications into a loop of "data production 鈥?model training 鈥?clinical application 鈥?feedback iteration": feedback generated in clinical use returns to the model and drives the next round of iteration.
From molecular structure to cellular state, and then to imaging and individual health, biology is undergoing a common shift: objects that once relied on experiments to observe are becoming systems that can be modeled, predicted and repeatedly simulated. AlphaFold let protein structure take the first step, virtual cells and virtual tissues continue at the scales that follow; medical imaging sits in the middle of this chain, facing the state of the body at the macroscopic level, and it too is moving from "being read by radiologists" to "being computed."
Along this path, one can look one step further: the prototype of a medical world model. A world model learns how the environment changes and how actions affect subsequent states; placed in medicine, it can be understood as a computational environment for studying disease 鈥?based on a patient's current state, it infers how the course of disease will develop, and then explores what results different interventions may bring. Diagnosis provides the current state, follow-up tests previous predictions, and real outcomes drive model improvement; if these links can be connected with physician participation, a continuous loop of "observation, prediction, simulation, real feedback" may form. If it reaches that stage in the future, more problems that are hard to test repeatedly in reality will have the chance to be explored first in computation.
From this perspective, the imaging understanding, risk prediction and multi-specialty R&D capabilities DIAGENS-B has already demonstrated are forming a key piece of the prototype of a medical world model. Its scarcity as an AI for Science target may lie here 鈥?the real value may not be in how far a single model can go, but in whether, when a doctor raises the next question, the accumulated groundwork can be reorganized.
Comments