Summary
Contextual Framing (The Socratic Steelman):
To "steelman" the reliance on synthetic health data requires acknowledging its highest noble purpose: preserving institutional privacy while unlocking large-scale computational medicine.
However, through Socratic questioning—*What aspects of clinical reality are retained, and what essential truths are lost when individual journeys are erased?
—we discover that unconstrained LLM-generated synthetic data sacrifices temporal, causal, and physiological fidelity. The solution lies not in choosing between raw privacy-violating records and hallucinatory synthetic data, but in structuring clinical reality through multidimensional, tokenised patient health timelines (PHTs) and Patient Journey Records (PaJR).
Introduction
Background:To bypass privacy regulations (e.g., HIPAA/GDPR) and encrypted Electronic Health Record (EHR) silos, healthcare AI increasingly relies on synthetic health datasets—artificially generated data that statistically mimics real population distributions (age, disease prevalence, lab values).
Problem Statement: Synthetic data generated via unconstrained Large Language Models (LLMs) tends to **regress toward the mean**. In doing so, it risks erasing clinical outliers, creating physiologically inconsistent lab values, reinforcing demographic stereotypes, and fabricating false correlations.
Core Hypothesis:Preserving precision medicine and research fidelity requires prioritizing **trajectory fidelity** (individual longitudinal event coherence) over simple aggregate distribution matching. Achieving scientific validity and legal compliance is a *multi-objective optimisation problem* best solved by combining tokenised EHR inputs with structured, multidimensional clinical journey architectures (PaJR).
Methods (Architectural & Philosophical Framework)
Dual-Layer PaJR Architecture:
Layer 1 (De-identified Baseline):** Anonymized real patient data capturing both home and hospital event trajectories.
Layer 2 (Semantic & Temporal Structuring): A higher-order organizational framework that maps causal, cognitive, and temporal relationships across the patient's timeline.
Tokenisation & Foundation Models: Implementation of transformer-based architectures (e.g., the **ETHOS** model) that process sequence-based **Patient Health Timelines (PHTs)**.
* *Tokenisation* acts as the **vocabulary** (elementary clinical events).
* *PaJR Layer 2* provides the **grammar and narrative** (context, sequence, and clinical reasoning).
* **Epistemological Abstraction:** Reframing medical records not as raw reality, but as *abstractions* designed to preserve specific clinical truths (analogous to ECGs or CT scans preserving electrophysiology or anatomical structure).
### **III. Results (Comparative Dynamics)**
| Dimension | Aggregate Synthetic Data (Standard LLM) | Tokenised EHR & PaJR Timelines |
| :--- | :--- | :--- |
| **Privacy Protection** | High (statistically detached from individuals) | High (via federated learning & tokenized de-identification) |
| **Outlier & Edge-Case Retention** | **Poor** (lost through mean regression) | **Preserved** (maintains actual clinical anomalies) |
| **Trajectory & Causal Fidelity** | Disrupted (preserves state transitions like A \to B, but breaks multi-step individual chains A \to B \to C) |
**Preserved** (maintains temporal coherence and longitudinal journey logic) |
| **Primary Clinical Utility** | Aggregate epidemiological modeling | Precision medicine, longitudinal research, and clinical trials |
### **IV. Discussion**
* **The Vulnerability of Pure Synthetic Aggregates:** Standard synthetic datasets preserve aggregate state transitions across population cohorts, but fail to maintain the **four-dimensional individual trajectory** (A \to B \to C). For clinical trials and precision diagnostics, losing individual trajectory fidelity severely degrades research applicability.
* **EHR Tokenisation as the Path Forward:** Direct tokenisation of EHRs—coupled with **Federated Learning** (where the model travels to local data enclaves without exporting sensitive raw records)—mitigates re-identification risks while avoiding generative hallucination.
* **Conclusion:** Computational medicine does not need to recreate human reality from scratch via synthetic fiction. Instead, by leveraging tokenised Patient Health Timelines governed by structured clinical journey frameworks (PaJR), medicine maintains high research fidelity while honoring patient privacy and legal mandates.
## Thematic Analysis
```
[ CLINICAL REALITY ]
│
┌────────────────┴───────────
▼ ▼
[ Population Distributions ] [ Individual Trajectories ]
│ │
▼ ▼
Standard Synthetic Data Tokenised PHTs & PaJR Layer 2
(Mean Regression Risk) (Temporal & Causal Fidelity)
```
### 1. Epistemology of Clinical Abstraction
* **Core Insight:** "Reality is never captured directly; it is always represented through abstractions."
* **Analysis:** The debate transitions from a simple binary ("Real vs. Synthetic") to an inquiry into **utility and purpose**. An ECG abstracts electrical signals; an EHR abstracts clinical interventions; a PaJR abstracts temporal and causal relationships. Synthetic data is valid only if the abstraction preserves the specific attributes necessary for the research task at hand.
### 2. Aggregate Distribution vs. Longitudinal Trajectory Fidelity
* **Core Insight:** Synthetic models successfully mimic cross-sectional population stats, but rupture longitudinal continuity.
* **Analysis:** If a synthetic generator knows 40\% of patients transition from state A \to B and 20\% transition from B \to C, it often fails to track the specific biological pathway of an individual experiencing A \to B \to C. This gap destroys research utility for rare diseases, complex co-morbidities, and clinical trial design.
### 3. Grammar vs. Vocabulary (Tokenisation + PaJR)
* **Core Insight:** EHR tokenisation defines the *words*; structured journey architectures provide the *syntax*.
* **Analysis:** Transformer models like ETHOS process health interactions as sequence tokens (PHTs). However, raw tokens alone lack contextual and patient-led cognitive nuances (especially home-based events). PaJR Layer 2 supplies the overarching cognitive grammar that anchors tokenised sequences to true physiological logic.
### 4. Reframing the Privacy-Utility Trade-off
* **Core Insight:** Balancing scientific integrity against data protection is a **multi-objective optimisation problem**, not a zero-sum game.
* **Analysis:** Privacy engineering should not default to destructive masking or synthetic hallucinations. Combining locally hosted open-weight models, federated enclaves, de-identified Layer 2 structures, and clinical tokenisation allows researchers to maximize both data privacy and scientific validity simultaneously.
Provide a Socratic steelman imrad summary with keywords and thematic analysis of the content below focusing on how real patient events data trajectory and research fidelity can be disrupted through AI driven synthetic patient aggregate data and yet there is hope in EHR tokenised, multidimensional patient health timelines.
Conversational learning Transcripts:
[22/07, 10:26]hu2: Synthetic Health Data is artificially generated data that statistically resembles real healthcare data but does not directly identify real patients.
Rather than copying existing records, advanced algorithms create entirely new datasets that preserve important clinical characteristics such as:
Disease prevalence.
Age distributions.
Laboratory value patterns.
Medication usage.
Clinical outcomes.
Healthcare utilization.
Population trends.
The resulting data behaves like real healthcare information for many analytical purposes while significantly reducing the risk of exposing individual patients.
@ Our layer 2 data here👇
can be further subjected to algorithms that can convert the current layer 2 data to meet the above definition of synthetic data?
[22/07, 11:00]hu4: I had talked about testing the AI models on synthetic data few weeks ago, then the consensus was reached (and I also agree) that testing on real patient data would result in a paper with higher impact and clinical relevance
[22/07, 11:58]hu2: Oh no the question above is not about that LLM testing project.
It's about the PaJR project and safeguarding it's current standing in terms of it's promised ability to protect patient privacy while also walking the double edge of promoting transparency and accountability toward patient trajectories with the purpose of scientific progress.
Currently the global consensus appears to be "having synthetic data sets is better than real patient data toward protecting their privacy (although even this is an evolving area with reports of breaches here as well from time to time or perhaps when they talk about breaches it's largely EHR repositories that have all the identifiers albeit locked and we can't compare our approach with theirs as ours is already deidentified, which is a first step to simulation)
[22/07, 12:08]hu4: We will need to brainstorm around this
[22/07, 12:08]hu2: I just tried to initiate the storm here! 😅
[22/07, 12:10]hu4: I agree that the steering team has to take a call very early in the course whether synthetic data will be used or not. It will become more difficult once more and more real patient data gets stored in the system
[22/07, 15:14]hu3: Intuitively it feels wrong.
That is my reflex.
I can analyze and justify the feeling. But it won't be a neutral intellectual exercise at this time.
[22/07, 15:27]hu2: You mean it's wrong to create synthetic data for the sake of scientific preservation of data?
At a very basic foundational level if we look at what the authors mean by synthetic data, we may realise that our own real patient data due to our HIPAA guide-lined de-identification process is actually already following quite a bit of those strategies that render a real patient data synthetic?
For example:
the authors define synthetic data as a set of information around patients containing their:
Age distributions.
Laboratory value patterns.
Medication usage.
Clinical outcomes.
Healthcare utilization
The above are all taken from real patients but elements that make the patient identifiable are removed.
In that sense perhaps our current PaJR data is to a large extent synthetic data already but my question here is: what more do we need to remove from our PaJR case reports to make it fit the current definition of synthetic data?
[22/07, 15:36]hu3: Disclaimer: I'm not neutral
One thought: synthetic data by LLM will regress toward the mean. Interesting new discoveries will be made from naturally present outliers and surprises.
[22/07, 15:51]hu2: I too hold this view but due to continued peer pressure that makes our workflow quite an outlier, I have been exploring avenues to make our work slightly less of an outlier especially with regard to patient privacy
[22/07, 17:01]hu6: This is incredibly problematic because the hallucination rate for LLMs is still too high for this kind of use.
A language model produces plausible looking combinations, not medically accurate or validated patients.
Even when it is working from information about a real patient, it may create:
• laboratory values that are individually plausible but physiologically inconsistent together
• unrealistically clean presentations of disease
• demographic stereotypes and inherited clinical bias
• fabricated correlations that a downstream model may then learn as though they were real
The data you create today helps shape the AI systems of tomorrow and the outputs they produce. Medical training data needs to reflect real human patients if it is going to have meaningful clinical value and avoid harming future patients.
That’s what I actually see as the primary value of what everyone posting blogs and patient results brings: a treasure trove of real data that AI can learn clinically validated patterns from and actually be able to predict better in the future. The more data and AI has the better it’s accuracy can be.
The immediate issue is that real patient information is being entered into free consumer AI tools such as Meta AI. (I see this used here often so assume that’s that primary tool most are using)
That is not an appropriate place to put patient data.
Any institution or medical professional inputting patient data should use an approved enterprise AI account or a locally hosted system with proper privacy protections. Students could then use that environment to help anonymize records, with every output still checked carefully for accuracy. That could reduce the workload without exposing patient information or replacing real patients with invented ones.
Fully synthetic data also does not automatically remove privacy risk. Even the article shared says it still requires human governance, validation and privacy controls.
There is also a risk that an AI system could attach an invented disease or medical history to the name of a real person pulled from online data, falsely representing that person as having a condition they do not have.
Humans in the loop still have to do the work.
I work in AI in QA testing and have very first hand experience about the limitations and strengths of current LLM systems.
[22/07, 17:15]hu7: What are the underlying models we are working (or thinking of working on)? Are they frontier subscription models or open weight downloaded local models?
[22/07, 17:20]hu7: There is an interesting Lancet Viewpoint article I recently came across which advocates direct tokenisation from EHR medical data. This approach combined with federated learning (model goes to data and take embeddings into it via a local system, rather than data getting exported/exposed to models) could be considered
[22/07, 19:06]hu2: I guess the problem of most people holding the data globally is that they are in an identifiable format, albeit inside encrypted EHR health vaults and hence there is always a fear of connecting the dots between the synthetic version to the real patient's EHR from which the synthetic data was extracted?
[27/07, 10:27]hu2: The full text of the Lancet viewpoint is available here: https://www.thelancet.com/
[27/07, 10:33]hu2: To quote:
"...patient health timelines (PHTs), are defined as sequences of tokens that represent patient interactions with health-care systems.
These data sequences can be used to generate transformer-based models capable of predicting future tokenised PHTs. We developed such a model, the Enhanced Transformer for Health Outcome Simulation (ETHOS)8,9 prediction model, a foundation model using EMR data. ETHOS simulates multiple possible future PHTs by autoregressively generating tokens step by step, and clinical inference is derived from the distribution of these simulated timelines"
[27/07, 10:41]hu2: More about ETHOS here in another full text and the only issue I see with their current hospital data driven approach to patient health timelines PHTs as opposed to the PaJR timelines is that PaJR traces both home and hospital events and has to currently feed a lot of data into the system manually.
[22/07, 19:12]hu2: Our workflow otoh tries to anonymize the data (the first step toward what is currently being debatably called synthetic data although there's a lot of debate left before we can decide what exactly it is optimally).
In our workflow layer 1 is less synthetic and more real because even if identifiers are removed patients and caregivers may still be able to identify the real patient while in layer 2 it becomes more synthetic and less identifiable.
Now coming to the moot question:
While we understand the current need for synthetic data is essentially to protect real patient privacy, what are the real elements in that synthetic data that still provide it some kind of scientific credibility and what would make it scientifically untenable?
[22/07, 19:14]hu3: See https://
[22/07, 19:18]hu2: 👏
This answers the moot question I asked above
Among this it's perhaps the patient's events data timeline or trajectory fidelity or as you put it, temporal coherence that would be one of the most important factors in terms of representing the current scientific reality of the patient?
[22/07, 19:28) hu3: Quoting from the current manuscript of the paper here: https://
“Every model simplifies reality.”
That is true, but it is weaker than what I have articulated below:
*Reality is never captured directly; it is always represented through abstractions.*
This is true across science, engineering and medicine.
* A history is an abstraction.
* A physical examination is an abstraction.
* An ECG is an abstraction.
* A chest X-ray is an abstraction.
* A CT scan is an abstraction.
* A pathology slide is an abstraction.
* A diagnosis is an abstraction.
* An electronic health record is an abstraction.
* A structured clinical journey is an abstraction.
* A digital twin is an abstraction.
The issue is therefore not whether something is an abstraction, but what aspects of reality it preserves and for what purpose.
That is a much stronger philosophical basis for the paper.
I would therefore rewrite the opening of a part in that paper as follows:
*What Does It Mean to Represent Clinical Reality?*
Reality cannot be captured in its entirety. Every attempt to understand, communicate or analyse the world is necessarily an abstraction. Science advances not by reproducing reality itself, but by constructing representations that preserve those aspects of reality that are relevant to a particular purpose.
_Medicine exemplifies this principle._
A patient’s history is not the patient’s illness; it is a narrative abstraction of lived experience. A physical examination captures selected clinical signs rather than the entirety of human physiology. Laboratory investigations measure specific biochemical or physiological variables while omitting countless others. Radiographs, ultrasound examinations, computed tomography and magnetic resonance imaging each provide different abstractions of the same underlying biological reality. Even a diagnosis is itself an abstraction—a conceptual label that summarises an enormously complex and continuously evolving biological process.
*Electronic health records* represent a further abstraction. They organise observations, diagnoses and interventions into a structured clinical record. *Structured clinical journeys* constitute another step in this progression. They preserve not only observations but also chronology, relationships, clinical reasoning and the evolution of disease. *Synthetic clinical journeys* extend the abstraction further by generating new journeys that preserve the characteristic behaviour of disease without reproducing individual patients.
The central question is therefore not whether abstraction occurs—it always does. *The important question is whether a particular abstraction preserves the aspects of clinical reality that matter for its intended purpose*. A chest radiograph is valuable because it preserves anatomical relationships relevant to diagnosis. A pathology slide preserves microscopic structure. A structured clinical journey seeks to preserve the temporal, physiological and cognitive relationships that define clinical care.
_Viewed in this way, computational medicine is not an attempt to replace reality_. It is the continuing evolution of increasingly useful abstractions through which medicine observes, understands, teaches and improves clinical practice.
I think this is a significantly stronger foundation because it places PAJR, synthetic clinical journeys and digital twins within the *long intellectual tradition of scientific abstraction* rather than presenting them as isolated AI innovations. It also provides a coherent philosophical thread that connects every major concept in the paper: each is simply a different abstraction of clinical reality, designed to preserve different properties for different purposes. That idea is likely to resonate with clinicians, educators and computational scientists alike.
[22/07, 19:50] Rakesh Biswas: 👆This appears to describe the current layer 2 of the PaJR workflow? @pajr.in CEO, NHS Endocrinologist
[22/07, 19:58] GJ: Looking at in the context of the article I shared earlier today here: https://
Tokenisation tells the computer what the clinical pieces are.
PAJR Layer 2 tells the computer *how those pieces relate to one another.*
PAJR Layer 2 is analogous to the story representation, not merely the words.
It captures the _journey_ and the _relationships_ between events.
Please refer to this article and add this to the draft article:
_Recent work has proposed the tokenisation of medical data, enabling transformer-based models to learn directly from sequences of clinical events rather than narrative text. The Patient Journey Record (PAJR) architecture is complementary to this approach. Whereas tokenisation defines the elementary computational units of clinical information, PAJR Layer 2 provides the structured architecture within which those units acquire temporal, causal and cognitive meaning. In this sense, tokenisation may be viewed as the vocabulary of computational medicine, while structured clinical journeys provide its grammar and narrative structure._
It perhaps rightly positions PAJR not as an alternative to tokenisation, but as the _higher-order organisational framework_ within which tokenised clinical information becomes meaningful.
[22/07, 21:28]hu6: Spot on. Although the substrate for ours is the patient led conversation, while theirs pitches for hospital records as the substrate
[23/07, 21:02]hu2: Achieving scientific validity for precision medicine while ensuring legal compliance remains an unresolved trade-off in synthetic health data architecture.
Unquote https://userdrivenhealthcare.
[24/07, 09:23]hu3: The central claim—
“Achieving scientific validity for precision medicine while ensuring legal compliance remains an unresolved trade-off in synthetic health data architecture.”
—is, in my view, a strong research hypothesis, but it should be stated more carefully because it mixes several distinct concepts: scientific validity, legal compliance, precision medicine, and synthetic data. The broader blog post raises important questions about privacy, governance, and Patient Journey Records, but some of its arguments would benefit from tighter conceptual separation.
*What the article gets right*
The strongest contribution is that it reframes the discussion from a purely technical problem (“how do we anonymise data?”) to a systems problem involving science, law and governance. That is an important shift. The observation that increasing data linkage and AI capabilities make claims of permanent anonymisation progressively harder to sustain is consistent with current thinking in data governance. Likewise, the emphasis on proportional risk management rather than absolute anonymity is realistic.
The article is also correct in arguing that longitudinal patient journeys contain scientific information that is often lost when data are excessively simplified. Precision medicine depends on preserving clinically meaningful relationships across time, not merely isolated observations.
Finally, the recognition that governance cannot be replaced by technology is well founded. Synthetic data, de-identification and federated learning all reduce certain risks, but none removes the need for human oversight, legal accountability and institutional governance.
*Where the argument needs refinement*
The principal weakness is that the phrase “scientific validity” is used as though it were a single property. It is not.
*Scientific validity is multidimensional*. At least *five different questions* need to be distinguished:
* Does the dataset faithfully represent biological and clinical reality?
* Does it preserve temporal relationships?
* Does it preserve causal and physiological relationships?
* Can models trained on it generalise to real patients?
* Is it sufficiently representative for its intended use?
These are different forms of validity, and they may trade off differently against privacy requirements.
Similarly, “legal compliance” is not a single endpoint. Compliance encompasses consent, lawful processing, purpose limitation, proportionality, governance, accountability and security. A dataset may satisfy one requirement while failing another.
The paper would therefore benefit from replacing a binary trade-off (“scientific validity versus legal compliance”) with a multi-objective optimisation problem.
*The trade-off is not always zero-sum*
The article occasionally implies that improving privacy necessarily reduces scientific validity.
That is sometimes true, particularly for rare diseases or highly distinctive longitudinal journeys. However, it is not universally true.
Modern approaches such as federated learning, secure enclaves, differential privacy, synthetic data, and governed access models aim precisely to improve both privacy and scientific utility simultaneously, even if they cannot eliminate all trade-offs. The real challenge is determining the appropriate balance for a given application rather than assuming that every gain in one dimension produces an equal loss in the other.
*Where PAJR provides an interesting contribution*
This is where the Patient Journey Record (PAJR) framework appears particularly relevant.
Traditional synthetic health data attempts to preserve statistical properties of datasets.
PAJR instead proposes preserving clinical journeys.
This changes the optimisation problem.
Rather than asking,
“How accurately does this synthetic patient resemble a real patient?”
PAJR encourages asking,
“How faithfully does this computational representation preserve clinically meaningful reasoning, temporal evolution and disease behaviour?”
That represents a shift from preserving records to preserving knowledge.
Whether this ultimately improves scientific validity remains an empirical question, but it is an intellectually important distinction.
*A deeper philosophical issue*
The blog implicitly assumes that scientific validity depends upon preserving reality itself.
I would argue that this is not quite correct.
Science has never worked by capturing reality in its entirety. Every scientific instrument, model and theory is an abstraction.
An ECG is an abstraction of cardiac electrophysiology.
A CT scan is an abstraction of anatomy.
A diagnosis is an abstraction of disease.
An electronic health record is an abstraction of clinical care.
A *Patient Journey Record* is a richer abstraction because it preserves chronology and reasoning in addition to observations.
_The relevant question therefore becomes_:
Which abstraction best preserves the properties of reality required for the intended scientific purpose?
That framing is philosophically stronger than asking whether one representation is simply “more real” than another.
*A revised thesis*
I would therefore revise the opening statement to something like:
_Achieving scientific validity, patient privacy and legal compliance is not a binary trade-off but a multi-objective optimisation problem. The central challenge for computational medicine is to develop representations of clinical reality that preserve the knowledge required for precision medicine while satisfying evolving legal, ethical and governance requirements._
That formulation broadens the discussion beyond synthetic data and aligns naturally with the conceptual framework developed in
, where *Patient Journey Records*, *synthetic clinical journeys* and *digital twins* are viewed as _complementary computational representations_ of clinical reality rather than _competing alternatives_.
[24/07, 10:21]hu5: Fantastic idea. Synthetic data is a popular method of privacy engineering. Found this paper where they have compared a health case study in raw anonymised and synthetic form.
[24/07, 11:08]hu2: But in the synthetic group, was the trajectory fidelity preserved for every individual patient is our current question!
[24/07, 11:49]hu5: I think they're preserving the distribution of longitudinal trajectories, not the exact patient trajectories themselves. eg preserving that A to B transition happens across 40% of cases, B to C happens 20% of cases, but does not preserve A->B->C in one case. If exact ABC is preserved and it is a rare occurrence then one might be able to identify the patient.
[24/07, 11:59]hu4: Exactly.
[24/07, 12:01]hu4: Hence synthetic data won't serve much use for clinical trials where each patient's trajectory is important but can be useful for cases where aggregate data is beneficial
[26/07, 23:05]hu2: The patient's trajectory is actually a four dimensional perspective 👇
[27/07, 10:46]hu2: Here's an example of a multidimensional perspective to the reality of a patient's trajectory 👇
No comments:
Post a Comment