HealthFlux Ignites the Medical World Model Wave: DIAGENS-B (02526) Imaging-Side Prototype Takes Shape

Title context: HealthFlux Ignites the Medical World Model Wave: DIAGENS-B (02526) Imaging-Side Prototype Takes Shape Text: Recently, Stanford University, together with Harvard, the University of Washington, and other institutions, launched HealthFlux, the world's first pan-modal human health world model, widely regarded by the industry as a landmark signal of the next "AlphaFold moment" in medical AI. This industrial revolutionmoving from "single-disease prediction" to "whole-health-state inference"has struck at the industry's long-standing core pain point: medical AI can recognize an increasing number of diseases, yet it has never established a holistic understanding of the "person." The key may lie not in the algorithm, but in the form of the data the model relies on. A person's health is preserved in the manner of "item-by-item records": blood glucose, blood lipids, an image of a certain body part, a certain diagnosiseach piece of data is generated to answer a specific question, and they are not connected to one another. Their timing is also unevenly spaced: some indicators are tested once every few months, while some images leave only one record every few years. And during the periods with no records, the body does not stop changing. The overall state that truly determines a person's future disease risk always exists, yet it is very difficult to observe directly. Most existing models can only advance along these fragments, either using data from a single point in time or modeling a predefined disease separately. Such models can make predictions for a specific disease accurate enough, but they struggle to answer a more fundamental questionwhat overall state is this person in, and where is that state evolving toward? The study led by the Stanford University research team attempts to transform that hard-to-directly-observe object into a computable state. The HealthFlux they developed attempts to learn a unified, implicit human health state and to model how this state continuously evolves with age, disease, and biological changes. (The related paper was published on the preprint website medRxiv.) When a model can predict health and disease risk with extreme accuracy, can predict post-administration changes with extreme accuracy, and can generalize well, it means the model may understand health to some degree. This is the judgment given by the authors of the study, and it is also what the study calls the next "AlphaFold moment." It turns "understanding" from a philosophical inquiry into a testable standard: predictions are accurate, post-administration changes can be calculated, and it still works in a different scenario. Once this judgment is put forward, the question is no longer just "can it predict a certain disease," but "is there a sufficiently general foundation that can transfer this predictive capability to tasks it has never seen." At the scale of imaging, the solution offered by DIAGENS-B (02526.HK) has two layersfirst use a general foundation to fit images from different sources into the same feature system, then connect feedback from clinical applications back into training through a data platform. Breaking the Fragmentation of Imaging Data: A Cross-Modal Foundation Restructures the R&D Efficiency System Imaging faces another kind of fragmentation. CT reflects differences in tissue absorption of X-rays, MRI comes from the signals of hydrogen protons under a magnetic field, ultrasound relies on sound wave reflection, and chromosome karyotyping comes from morphology and banding patterns under microscopic imaging. The imaging principles differ, and the features they generate each form their own systems. They record the body from different dimensions, but for a long time there has been a lack of a model that can fit these dimensions into the same representation. Each extension to a new modality or a new disease type requires going through data, annotation, training, evaluation, and deployment all over again. The capabilities accumulated previously are difficult to reuse, and costs have remained hard to push down. The general medical imaging foundation model iMedImage built by Diagens aims to solve this reuse problem. It is not trained separately for any single type of imaging. Instead, through cross-modal pretraining, it enables the model to form a shared feature representation, and subsequent specialty tasks all start from this feature system. This feature system is not bound to any single disease, nor is it limited to any single modality; it describes the common structure of imaging itself. Specialty tasks perform discrimination on top of it and do not need to learn "how to read images" from scratch. iMedImage is centered on a 104-billion-parameter architecture and has completed pretraining based on 8 million medical imaging data entries, covering 19 imaging modalities, 26 clinical specialties, and more than 90% of imaging diagnostic scenarios, supporting model development for more than 400 diseases. The direct change brought by this scale falls 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 data entries, with a development cycle of 2 to 3 months. Compared with the traditional "one disease, one model" development model, 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 drop to one-tenth. Disease-specific model development that originally only large top-tier hospitals and leading AI companies could afford is now beginning to become accessible to small and medium-sized medical institutions as well. As of now, Diagens has reached cooperation with 99 hospitals, has cumulatively implemented 158 disease-specific model projects, and covers 61 disease directions. Positioning in the World Model Track: The Imaging-Side Prototype Has Already Landed First Diagens' research in imaging is extending from discrimination of the current state to judgment of 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's test accuracy in distinguishing the early-stage and advanced-stage groups was 95.38%, with an AUC of 0.9328. The results were published in July 2026 in the European Journal of Radiology Opena judgment that originally required postoperative pathology was moved forward to the preoperative stage. Information from Diagens' interim results roadshow shows that the company's study of 673 breast cancer recurrence prediction cases incorporated pathological slides, ultrasound, and clinical information into the same 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 two time points in clinical decision-making: before treatment begins and after treatment ends. They point to the same capabilityturning information in images that has not yet been read out into a basis for judgment that can be used. For these judgments to be continuously tested, they rely on Diagens' self-developed iMedLoop platform. The platform has gathered more than 300 professional annotators and 28.95 million imaging data entries, integrating data, tools, models, and applications to form a cycle of "data productionmodel trainingclinical applicationfeedback 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 change: objects that used to rely on experiments for observation are becoming systems that can be modeled, predicted, and repeatedly inferred. AlphaFold let protein structure take the first step, and virtual cells and virtual tissues continue at subsequent scales; medical imaging sits in the middle of this chain. It faces the body's state at the macroscopic level and is likewise moving from "being read" to "being computed." Along this path, one can see 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 diseasebased on a patient's current state, infer how the disease course will develop, and then explore what different interventions may bring. Diagnosis provides the current state, follow-up tests verify previous predictions, and real outcomes drive model improvement; if these links can be connected with physician participation, a continuous cycle of "observation, prediction, simulation, real feedback" may form. If it reaches that point in the future, more problems that are difficult to test repeatedly in reality will have the opportunity to be explored first in computation. From this perspective, the imaging understanding, risk prediction, and multi-specialty R&D capabilities that Diagens 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 precisely herethe true value may not lie in how far a certain model can go, but in whether, when a physician raises the next question, the accumulated foundation can be reorganized.