📄 medicine

From Disease Classification to Whole-Patient State Regulation: A Longitudinal Clinical Computation Framework for Whole-Patient State Trajectories

This paper proposes and validates a longitudinal clinical computation framework that shifts the focus from static disease classification to the dynamic regulation of whole-patient state trajectories by establishing them as the core computational object for organizing clinical observations, estimating hierarchical latent states, and driving feedback-based interventions.

Original authors: Bing Yuan, Gang Fan, Qifan

Published 2026-08-06

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Bing Yuan, Gang Fan, Qifan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Doctor's Dashboard: From Snapshots to Movies

Imagine you are trying to understand a complex story, like a favorite TV show. Most medical computers today work like a photographer who takes a single, high-quality snapshot of a character at one specific moment. They are excellent at identifying who the character is right now: "That's a hero," or "That's a villain," or "That's a person with a broken leg." This is the world of disease classification, where the goal is to sort patients into neat boxes based on what is wrong with them at a single point in time. It's like looking at a still image of a car crash and saying, "That is a crash."

However, real life isn't a series of still photos; it's a movie. A patient isn't just a "broken leg"; they are a whole person whose body is constantly shifting, reacting to medicine, fighting off germs, and dealing with stress. This is the concept of the whole-patient state—the idea that a person is a dynamic, moving system where everything is connected. While doctors have always tried to keep this "movie" in their heads, current computer systems struggle to do it. They are great at labeling the frames but bad at understanding the plot. This paper asks a big question: What if we stopped trying to just take snapshots and started building a system that could watch the whole movie, tracking how a patient's state changes over time to help guide their recovery?

From Snapshots to Streaming Movies: A New Way to Track Health

This paper proposes a new way for computers to help doctors, moving away from the old "snapshot" style of medicine toward a "movie" style. The authors, Bing Yuan, Gang Fan, and Qifan Yuan, suggest that instead of just asking, "What disease does this patient have right now?", we should ask, "How is this patient's whole state changing, and how do we steer that change in the right direction?"

To explain how this works, imagine a doctor trying to fix a very complex, old-fashioned clock.

The Old Way (Disease Classification): The computer looks at the clock, sees a broken spring, and says, "Broken Spring Syndrome." It then suggests a generic fix for broken springs. It doesn't care if the gears are rusted or if the clock is ticking too fast; it just sees the one broken part and stops there.
The New Way (Whole-Patient State Regulation): The computer realizes the clock is a whole system. It doesn't just look for a broken spring; it watches the entire clock as it ticks. It notices that the spring is broken, but also that the gears are moving too slowly, and the pendulum is swinging too wide. It sees these as a connected "state" of the clock.

The authors call this new approach a Longitudinal Clinical Computation Framework. "Longitudinal" just means looking at things over a long period, like watching a plant grow from a seed to a flower, rather than just measuring it once.

The Core Idea: The Patient as a "State Space"

The paper suggests we should think of a patient not as a list of diseases, but as a hierarchical latent-state space. That sounds complicated, but think of it like a video game character's dashboard.

The Dashboard: Instead of just showing "Health: 50%," the dashboard has many different sliders and gauges. Some are big, global gauges (like "Overall Energy" or "Stress Level"), and some are smaller, local gauges (like "Lung Function" or "Digestion").
The Sliders: These are called state variables. In the old system, a variable was either "On" (you have the disease) or "Off" (you don't). In this new system, the sliders can be anywhere. They can be "mildly off," "moderately off," or "severely off."
The Movie: As the patient gets new treatments or their body reacts, these sliders move. The computer doesn't just say "You have a cold." It says, "The 'Lung Heat' slider is moving up, but the 'Energy' slider is dropping. We need to adjust the treatment to cool the lungs without draining the energy."

How the System Works: A Dynamic Loop

The paper describes a system that runs in a continuous loop, much like a video game that updates in real-time based on your actions.

Gathering Clues (Structured Observation): Instead of just typing notes into a box, the system organizes information like symptoms, lab results, and even how the patient is sleeping or eating into a structured map. It's like a detective who doesn't just write down clues but connects them: "The patient has a cough and a fever and is sweating, which suggests a specific pattern."
Estimating the State (The Diagnosis): The system looks at all these clues to figure out where the sliders are. It doesn't just pick one label. It might say, "The patient is mostly 'Wind-Heat' (a type of imbalance), but also has a bit of 'Phlegm'." It understands that a patient can have multiple things happening at once.
Picking the Target (Regulation Objectives): This is the smart part. The system looks at all the "broken" sliders and asks, "Which one matters most right now?" Maybe the fever is the biggest problem today, even if the stomach ache is also there. The system picks a target state-variable set—the specific group of sliders it needs to fix first.
Making a Plan (Intervention): Based on that target, it suggests a treatment plan. It's not just "take this pill." It's "do this to lower the fever, but be careful not to make the stomach ache worse."
Watching the Movie (Longitudinal Updating): This is the most important part. The system doesn't stop there. When the patient comes back a week later, the system doesn't start over. It takes the new clues (did the fever go down? did the stomach hurt more?) and updates the movie. The sliders move again. The "Wind-Heat" might be gone, but now "Yin Deficiency" (low fluids) has appeared. The system shifts its goal from "cooling the heat" to "replenishing fluids."

What the Paper Actually Found

The authors built a functional prototype system to test this idea. They didn't just write a theory; they made a working model.

The Simulation: They ran the system through a representative case (a pretend patient scenario) to see if it could handle the complexity.
The Result: The system successfully took a patient's condition, organized it into these "state variables," and then tracked how the condition changed over several visits.
- At first, the system identified a dominant problem (like "Wind-Heat").
- As the patient got treated, the system saw that problem fade away.
- Crucially, the system noticed a new problem appearing (like "Lung Heat" or "Yin Deficiency") and automatically shifted its focus to fix that new problem.
- It successfully updated the treatment plan from "clearing heat" to "nourishing fluids" as the patient's state evolved.

The paper suggests that this approach allows for dynamic regulation. It means the computer isn't just a label-maker; it's a co-pilot that helps the doctor steer the patient's health journey over time.

What This Is NOT

It is important to know what this paper is not claiming.

It is not a magic cure: The authors are very clear that this is a framework and a prototype. They did not test it on thousands of real patients in a hospital to prove it saves lives. They showed that the structure works and makes sense computationally.
It is not replacing doctors: The system is designed to help organize the massive amount of information a doctor deals with, not to replace human judgment.
It is not just about Traditional Chinese Medicine (TCM): The paper uses TCM concepts (like "Qi deficiency" or "Phlegm-heat") as a perfect example of how "state regulation" has worked for centuries. However, the authors are not saying we should switch to TCM. They are saying, "Look, TCM has been doing this 'whole-patient state' tracking for a long time. Let's build a modern computer system that does the same thing, but using our current medical data too."

Why This Matters

The authors argue that modern medicine has gotten really good at identifying diseases (the "snapshots"), but it's getting harder to manage complex, long-term illnesses where everything is connected. By shifting the focus from "What disease do you have?" to "How is your whole state changing?", this framework offers a new way to think about healthcare.

The paper concludes that future medical AI shouldn't just be a classification tool. Instead, it should be a longitudinal regulation system—a tool that helps doctors watch the movie of a patient's life, understand how the plot is changing, and adjust the story to ensure a happy ending. The authors suggest that while we can't always know every tiny mechanical detail of how the body works (the "mechanistic incompleteness"), we can still successfully manage the patient's state by tracking these deviations and adjusting the treatment as the story unfolds.

In short, this paper suggests that the future of medical computers isn't just about taking better photos; it's about learning to direct the movie.

Technical Summary: From Disease Classification to Whole-Patient State Regulation

Problem Statement
Modern clinical medicine and contemporary healthcare AI systems are predominantly organized around disease entities, utilizing single-timepoint classification architectures. While effective for acute, isolated conditions, this paradigm faces significant limitations in complex clinical environments involving multimorbidity, chronic management, and dynamic physiological interactions. Current systems lack an explicit structural representation and formal computational framework for the "whole-patient state." Consequently, the continuous, dynamic process of clinical regulation—which relies on holistic assessment and longitudinal adjustment—remains an implicit, unstructured clinical intuition rather than a formal computational object. Existing AI models typically address the question "What category does the patient belong to?" but fail to support the question "How is the patient's state evolving, and how should regulation be adjusted?"

Methodology
The study proposes a Longitudinal Clinical Computation Framework that reorients clinical organization from disease classification to the regulation of whole-patient state trajectories. The framework is built upon the following core components:

Dynamic Structured Observation Space: Clinical information (symptoms, signs, diagnostics, functional status, lifestyle) is organized into a structured network rather than isolated fields. This space supports dynamic supplementary inquiry based on current observations to refine state discrimination.
Hierarchical Latent-State Representation: Whole-patient state is modeled as a hierarchical latent-state space ( $X = (X_G, X_1, ..., X_n)$ $X = (X_{G}, X_{1}, ..., X_{n})$ ) composed of multiple interacting state variables.
- Global Level ( $X_G$ ): Represents overall functional coordination, resource status, and adaptive capacity.
- Subsystem Level ( $X_i$ ): Represents local functional deviations and regulatory imbalances.
- Nature of Variables: These are operational abstractions of state deviations (e.g., "qi deficiency," "phlegm-heat") rather than static disease labels. They are not mutually exclusive and can coexist, compete, or couple dynamically.
Hierarchical State Estimation: Instead of direct observation-to-label classification, the system employs a multi-layer support structure (Manifest Observation $\to$ Observation Group $\to$ Substrate Layer $\to$ State Variable). This allows for the tracing of how observations contribute to state formation and reduces redundancy from correlated inputs.
Target State-Variable Set Formation: Recognizing that not all deviations require simultaneous regulation, the system prioritizes state variables based on their degree of deviation, influence on stability, and potential for future evolution. This forms a "target state-variable set" that guides current regulatory focus.
Regulation-Oriented Intervention: Interventions are generated based on the composite deviation structure of the target set, rather than single symptoms. This includes forming regulation objectives (e.g., "clearing heat") and organizing intervention strategies.
Longitudinal Trajectory Updating: The system treats clinical practice as a continuous closed-loop process. New observations and treatment feedback dynamically update state-variable values, inter-variable relationships, the target set, and regulation objectives.

Key Contributions

Formalization of Whole-Patient State: The study establishes the whole-patient state trajectory as the primary computational object of clinical organization, moving beyond static disease labels.
Operational State Abstraction: It introduces state variables as operational representations of deviation structures that support dynamic regulation without requiring the complete reconstruction of underlying biological mechanisms (addressing "mechanistic incompleteness").
Dual-Layer Coupling Structure: The framework proposes a coupling of Mechanistic Constraints (disease mechanisms, safety boundaries) and State Regulation (global functional deviations). Mechanisms constrain the boundaries of regulation, while state variables drive the dynamic adjustment of interventions.
Interpretable Longitudinal AI: Unlike black-box classification models, this framework provides an explicit state structure where the entire process—from observation support to trajectory shifts—is traceable and interpretable.
Reinterpretation of TCM: The paper posits that Traditional Chinese Medicine's pattern differentiation and treatment serves as a historical prototype for this dynamic, state-variable-based regulatory system, offering a structural model for whole-patient regulation.

Results
A functional system prototype was developed and demonstrated using representative longitudinal cases. The results indicate the system's ability to:

Continuously integrate new clinical observations and feedback.
Dynamically update state-variable structures, target sets, and regulation objectives.
Track the evolution of a patient's state trajectory (e.g., transitioning from a "wind-heat exterior invasion" state to "lung-heat/phlegm-heat," and finally to "yin-fluid deficiency" during recovery).
Demonstrate that the object of regulation is not a single diagnostic outcome but the evolving trajectory of state deviations.

Significance and Claims
The authors claim that this study provides a computational foundation for trajectory-oriented clinical management and interpretable longitudinal healthcare AI. The significance lies not in the development of a specific disease-predictive model, but in the redefinition of the core organizational objects of clinical computational systems.

The paper argues that future clinical systems must evolve from single-timepoint classification architectures toward longitudinal regulation centered on whole-patient state trajectories. This shift addresses the structural gap in current healthcare AI, which lacks explicit state layers and long-term dynamic regulation capabilities. By formalizing the whole-patient state as a computable, dynamic entity, the framework aims to support dynamic regulation in complex, multimorbid, and chronic clinical environments where complete mechanistic reconstruction is often impossible. The study emphasizes that this is a structural and organizational proposal, demonstrating feasibility through a prototype, rather than a claim of established clinical efficacy through large-scale validation.