Healthflux detonates the medical world model wave Deshit-b (02526) image terminal prototype already exists

The Zhitong Finance App learned that recently, Stanford University, in collaboration with Harvard, Washington University and other institutions, launched the world's first pan-modal human health world model HealthFlux, which is widely regarded by the industry as a landmark signal for the next “AlphaFold moment” in the medical AI field. This industrial revolution from “predicting a single disease type” to “deducing the full state of health” has hit the core pain point of the industry for a long time: medical AI can identify more and more diseases, yet it has never been able to establish an overall understanding of “people.”

Perhaps the key is not the algorithm, but the form of data that the model depends on. Human health is stored in the form of “itemized records”: blood sugar, blood lipids, images of a certain part, a certain diagnosis — each data is generated to answer a specific question and is not connected to each other. The timing of their appearance is also uneven. Some indicators are tested every few months, and some images are only left over a few years. And in an unrecorded period of time, the body did not stop changing.

The overall state that actually determines a person's future disease risk always exists, yet it is difficult to directly observe it. Most existing models can only advance along these fragments, or use data from a specific point in time, or model a pre-defined disease individually. Such models can be accurate enough to predict specific diseases, but it is difficult to answer a more basic question — what state this person is in as a whole, and where is this state evolving.

The study, led by a Stanford University research team, is trying to transform an object that is difficult to directly observe into a computable state. They developed HealthFlux in an attempt to learn a unified, implicit state of human health and model how this state continues to evolve with age, disease, and biological changes.

(Related papers are published on the pre-print website medRXiv)

When the model can predict the risk of health diseases with extreme accuracy, can predict changes after administration with extreme accuracy, and can be generalized well, it means that the model may understand health to some extent. This is the judgment given by the study's authors, and it is also the next “AlphaFold moment” according to the study. It turned “understanding” from a kind of philosophical inquiry into a testable standard: accurate prediction, calculation of changes after administration, and application in different scenarios.

After this judgment was put forward, the question was no longer just “can we predict a certain disease,” but “whether there is a sufficiently general foundation that can transfer this predictive ability to a task it has not seen before.” On the imaging scale, the solution given by Deshai-B (02526.HK) has two layers — first using a generic base to load images from different sources into the same characteristic system, and then feedback from clinical applications back into training through a data platform.

Breaking the game of image data fragmentation: reconstructing R&D efficiency systems across modal bases

The image faces a different kind of fragmentation. CT reflects differences in tissue absorption of X-rays. MRI comes from hydrogen protons in a magnetic field, ultrasound relies on sound wave reflection, and karyotypes come from morphology and band patterns under microscopic imaging. The imaging principles are different, and the characteristics generated are also separate systems. They record the body from different dimensions, but for a long time, there was no model that could fit these dimensions into the same set of expressions. Every time it extends to a new model or new disease type, data, labeling, training, evaluation, and deployment must be repeated. Previously accumulated capabilities are difficult to reuse, and costs cannot be suppressed.

iMediimage, a universal medical imaging base model built by Desch, aims to solve this kind of reuse problem. It does not train a specific type of image separately, but rather uses cross-modal pre-training to form a set of shared feature expressions. Subsequent specialist tasks all start from this feature system. This characteristic system is not bound to any disease, nor is it limited to any kind of modality; it describes the common structure of the image itself. Specialized tasks make judgments on top of it; there is no need to learn “how to view images” from scratch.

iMediimage uses a 104 billion parameter architecture as the core, completing pre-training based on 8 million medical imaging data, covering 19 imaging modes, 26 clinical specialties, and more than 90% of imaging diagnosis scenarios, and supporting model development for more than 400 diseases.

The direct change brought about by this size is due to the development of specialized disease models. On iMediimage, a minimum of about 200 images can be used to develop a vertical specialty model, and the development period is 2 to 3 months. Compared with the traditional “one disease, one model” development model, the demand for labeling data was reduced to 20% of the original, the development cycle was shortened to one-twelfth, and development costs and computing power expenses were reduced to one-tenth. Originally, only the top three hospitals and leading AI companies were able to develop specialized medical models, and small and medium-sized medical institutions are beginning to be able to participate.

Up to now, Deshi has reached cooperation with 99 hospitals, and has implemented a total of 158 specialty model projects, covering 61 disease directions.

Card World Model Circuit: The image terminal prototype has been the first to land

Deshi's research on images is extending from judging the current state to judging future risks.

A thymic study based on iMediimage, combined with imaging characteristics, predicted stages with pre-operative CT in 323 cases of thymic epithelial tumors, of which 258 were used for training and 65 were used for internal testing; the test accuracy of the model distinguishing between early and advanced groups was 95.38%, and the AUC was 0.9328. The results were published in the “European Journal of Radiology Open” in July 2026 -- originally requiring post-operative pathology The judgment was brought forward to before surgery. According to Deshi's mid-term performance roadshow information, the 673 breast cancer recurrence prediction studies carried out by the company incorporated pathological slices, ultrasound, and clinical information into the same framework. The optimized model had a c-index of 0.76 in the 50% cross-validation.

Thymus research is studying “what disease state does imaging correspond to”, and breast cancer research is exploring “what future risks current characteristics point to”. The two studies covered exactly two points in clinical decision making: before treatment begins and after treatment ends. They point to the same ability — turning information that hasn't been read in images into a basis for judgment that can be used.

To be able to continuously test these judgments, we rely on the iMedLoop platform developed in-house by Deshi. The platform has gathered more than 300 professional labeling personnel and 28.95 million pieces of image data, integrating data, tools, models and applications to form a cycle of “data production — model training — clinical application — feedback iteration”: feedback generated during clinical use returns to the model to drive the next round of iteration.

From molecular structure to cellular state, to imaging and individual health, biology is undergoing a common change: objects that in the past relied on experiments to observe are becoming systems that can be modeled, predicted, and deduced over and over again. AlphaFold takes the protein structure one step further, and the virtual cell continues on a subsequent scale; medical imaging is in the middle of this chain. It faces the state of the body at the macro level, and is also moving from “being read” to “being calculated.”

Following this path, you can also take a step further: the prototype of the medical world model.

The world model learns how the environment changes and how actions affect subsequent states; in medicine, it can be understood as a computational environment for studying diseases — deducing how the disease progresses based on the patient's current state, and then exploring what results different interventions may bring. Diagnosis provides the current state, follow-up tests previous predictions, and real results drive model improvement; if these steps can be connected with the doctor's participation, it may form a continuous cycle of “observation, prediction, simulation, and real feedback.” If we get to that point in the future, there will be an opportunity to explore more problems in computation first if they are difficult to test over and over again in reality.

Seen from this perspective, the image understanding, risk prediction, and multi-specialty research and development capabilities already demonstrated by Deshi are forming a key part of the prototype of the medical world model. Its scarcity as an AI for Science target is probably hidden here — the real value may not be how far a model can go, but whether the existing accumulation can be reorganized when the doctor asks the next question.