Beyond Evidence Hierarchies Toward I-OM Systems Medicine
Richard Z. Cheng, MD, PhD, and Eric K. Cheng
Orthomolecular Medicine News Service
Artificial intelligence is moving rapidly into medicine. It can search enormous literatures, summarize medical records, interpret laboratory trends, generate differential diagnoses, and assist clinicians and patients in ways that were difficult to imagine only a few years ago.
But a fundamental question has received much less attention:
What model of medicine is medical AI actually learning?
A highly capable AI trained within an incomplete medical framework does not automatically correct that framework. It may simply reproduce its assumptions faster, more consistently, and on a much larger scale.
This is the central concern behind our newly published paper, Beyond Evidence Hierarchies: An Intervention-Aware, Contextual Systems Framework for Medical Artificial Intelligence.[1]
Our argument is not that evidence-based medicine should be discarded. Nor is it that conventional medicine lacks enormous value. Modern medicine is extraordinarily effective in trauma, emergency care, surgery, acute infection, imaging, intensive care, and many other fields.
The problem is different.
The dominant medical framework is often better at identifying and treating established disease than at integrating the multiple upstream processes that produce chronic disease over time.
If artificial intelligence inherits that limitation, we may create increasingly sophisticated AI that becomes very good at practicing an incomplete model of medicine.
A diagnosis is not necessarily an explanation
Consider type 2 diabetes.
A diagnosis identifies a clinically important state, and glucose-lowering medications may reduce hyperglycemia and improve outcomes in appropriate patients.
But the diagnosis itself does not explain why the patient developed diabetes.
Important questions may include:
-
Why did insulin resistance develop?
-
What happened to metabolic flexibility?
-
What was the role of diet and chronic excess energy exposure?
-
Was visceral adiposity present?
-
How much skeletal-muscle activity and muscle mass did the patient have?
-
What roles were played by sleep, stress, cortisol, nutrient status, inflammation, medications, environmental exposures, or other contributors?
Similar questions arise in cardiovascular disease, obesity, fatty liver disease, neurodegeneration, frailty, many cancers, and other chronic disorders.
A disease label is important.
But it is not a complete causal model.
When medicine is organized primarily around diagnostic categories, specialties, and disease-specific treatment pathways, interconnected biological processes can become fragmented across clinical silos.
AI can inherit exactly the same fragmentation.
The danger of disease-label matching
Much of contemporary medical AI can be conceptualized, in simplified form, as moving from:
clinical question → information retrieval → evidence ranking → recommendation
That approach is useful.
But it is not sufficient for complex biological systems.
Human biology is dynamic, adaptive, heterogeneous, and only partially observed. The same diagnosis can arise through different combinations of contributing processes, and the same intervention can produce different effects depending on baseline physiology, dose, timing, route, nutritional status, phenotype, adherence, and other therapies.
Medical reasoning therefore requires more than asking:
“What treatment is recommended for this diagnosis?”
It also requires asking:
“What biological processes are operating in this particular person, what may be driving them, which of those processes are modifiable, and how can we determine whether our causal model is actually correct?”
That is a systems-medicine problem.
Evidence quality is not the same as evidence fit
A second limitation concerns how medical evidence is evaluated.
Randomized controlled trials remain indispensable tools for causal inference. The issue is not whether RCTs are valuable. They are.
The problem arises when a single evidence hierarchy is applied too mechanically across fundamentally different types of interventions.
A novel pharmaceutical, an essential nutrient, a dietary pattern, exercise, sleep modification, fasting, psychotherapy, surgery, environmental change, and a multi-component metabolic program do not have identical causal structures.
Consider a nutrient.
Unlike a novel drug, the nutrient is usually already present in every participant before the trial begins. Response may depend upon:
-
baseline nutrient status;
-
background intake;
-
absorption;
-
dose;
-
achieved biological concentration;
-
threshold and saturation effects;
-
co-nutrients;
-
phenotype;
-
disease state;
-
duration of intervention.
Robert Heaney highlighted this problem years ago, proposing that nutrient trials should account for baseline nutrient status, achieved changes in nutrient exposure, and the nonlinear biological responses characteristic of nutrients.[2]
A rigorously randomized trial can therefore still be poorly matched to the biological question it is being asked to answer.
This does not make randomized trials unimportant. Rather, it means that methodological rigor and methodological fit are separate questions.
We should ask both:
Was the study conducted rigorously?
and
Was the study designed in a way capable of testing the biological effect in question?
The distinction is particularly important for nutrition, lifestyle, behavior, exercise, complex interventions, and personalized treatment.
From an evidence hierarchy to an evidence map
Medical AI could make this problem worse—or help solve it.
If an AI system simply ranks evidence according to a conventional hierarchy and retrieves the highest-ranked publication, it may appear rigorous while overlooking crucial biological context.
A better approach is to construct an evidence map.
Different questions may require different combinations of evidence:
-
randomized clinical trials;
-
prospective and observational studies;
-
mechanistic studies;
-
pharmacokinetic or nutrient-status data;
-
dose-response relationships;
-
physiological measurements;
-
clinical experience;
-
safety data;
-
repeated longitudinal observations in the individual patient.
These forms of evidence should not all be treated as equivalent.
But neither should important information automatically disappear merely because it occupies a lower position on a generic evidence ladder.
The objective is not to lower evidentiary standards.
It is to match the method of evaluation to the intervention, the biological question, and the clinical context.
Generic AI may inherit the blind spots of generic medicine
General-purpose AI systems are extraordinarily powerful. They can synthesize more information than any individual physician could read in a lifetime.
But they largely work with the medical literature, guidelines, terminology, and conceptual structures humans have created.
Consequently, they may reproduce familiar patterns:
-
disease-centered rather than process-centered reasoning;
-
fragmentation of interconnected physiology into specialties;
-
preference for well-standardized interventions over complex or individualized ones;
-
insufficient distinction between nutrient biology and pharmaceutical pharmacology;
-
inadequate attention to dose, route, baseline status, and achieved biological effect;
-
difficulty integrating mechanistic, clinical, metabolic, and longitudinal evidence;
-
confusion between absence of evidence and evidence of absence;
-
recommendations based on population averages with insufficient attention to individual trajectory.
AI does not automatically escape our intellectual framework simply because it can process more data.
A powerful AI operating inside a narrow model can become a very powerful narrow-model practitioner.
Integrative Orthomolecular Systems Medicine
We believe medical AI needs a broader clinical framework.
We call this Integrative Orthomolecular Systems Medicine — I-OM Systems Medicine.
The central premise is straightforward:
Chronic disease should be understood not only by its diagnostic label, but also through the interacting upstream drivers and downstream biological processes that shape the patient’s disease trajectory.
Potential upstream contributors may include:
diet and metabolic dysfunction, nutritional inadequacy, environmental exposures, chronic infections, oral pathology, hormonal disturbances, sleep disruption, physical inactivity, psychological stress, medication effects, early-life influences, and genetic or epigenetic susceptibility.
These can converge on recurring biological processes such as:
oxidative stress, mitochondrial dysfunction, metabolic inflexibility, chronic inflammation, immune dysregulation, endothelial and microvascular dysfunction, barrier dysfunction, abnormal signaling, and impaired repair and recovery.
The purpose is not to declare that every possible driver is causal in every patient.
Quite the opposite.
A systems approach should force us to distinguish among:
established causes, probable contributors, plausible hypotheses, and unsupported speculation.
Then we test those hypotheses against measurable biology and longitudinal response.
What orthomolecular medicine contributes
Orthomolecular medicine adds another dimension.
Its objective is not merely to prevent classical deficiency diseases. It asks whether the molecular environment in which cells operate can be optimized to support physiological function.
This requires attention to individual context and dose.
For example, “vitamin C” is not a single intervention.
Dietary vitamin C, oral supplementation, and high-dose intravenous vitamin C involve profoundly different exposures, pharmacokinetics, biological concentrations, purposes, and safety considerations.
An AI system that places all three under the single heading “vitamin C supplementation” has not adequately characterized the intervention.
The same principle applies broadly.
Dose, route, timing, baseline state, biological target, and patient phenotype matter.
From I-OM Systems Medicine to I-OM.ai
These ideas form part of the conceptual foundation for I-OM.ai, which we are developing as an AI-enabled clinical reasoning and education platform.
The objective is not to build another general medical chatbot.
It is to explore whether AI can support a more integrated form of reasoning across:
metabolism, nutrition, lifestyle, biomarkers, conventional therapy, medications, safety, patient goals, environmental factors, and longitudinal change.
A useful system should be able to ask:
What is changing?
What processes might explain it?
What is modifiable?
What evidence supports each hypothesis?
What intervention is being considered?
At what dose and by what route?
What are the risks and interactions?
What measurements could confirm or refute the hypothesis?
And after intervention:
Did the predicted biological response actually occur?
That last question is crucial.
Medicine should not end with a recommendation.
It should proceed through an iterative cycle:
observe → hypothesize → intervene → measure → learn → update.
AI is particularly well suited to helping humans manage that longitudinal complexity.
AI should augment—not replace—clinical judgment
None of this means that AI should independently practice medicine.
Medical decisions involve uncertainty, values, risks, tradeoffs, feasibility, and human circumstances that cannot simply be reduced to an algorithm.
AI systems also remain capable of error.
They can hallucinate, misinterpret evidence, overlook contraindications, and express unjustified certainty.
The purpose of systems-oriented medical AI should therefore be to augment clinical reasoning, not replace physicians or eliminate professional judgment.
It should make reasoning more transparent.
It should identify uncertainty.
It should show competing hypotheses.
It should distinguish what is well established from what is emerging or speculative.
And it should remain bounded by safety.
Our proposed framework itself requires prospective validation before systems based upon it should be relied upon for clinical care.[1]
The larger opportunity
Artificial intelligence presents medicine with an unusual opportunity.
We could use AI primarily to automate the existing medical system.
It could summarize charts faster, generate notes faster, find guidelines faster, code diagnoses faster, and match diseases to established therapies faster.
Those applications are useful.
But they leave a much larger opportunity unexplored.
AI can potentially integrate biological information at a scale that exceeds ordinary human cognitive capacity:
diet, metabolism, laboratory trajectories, medications, nutrients, physical activity, sleep, symptoms, environmental exposures, medical history, imaging, wearable data, and scientific literature.
But integration requires a conceptual framework.
Without one, more information can simply produce more noise.
The next breakthrough in medical AI may require a better model of medicine
The race to build more powerful medical AI understandably focuses on larger models, more data, better retrieval, improved agents, and greater computational capacity.
Those advances matter.
But perhaps one of the most important questions is more fundamental:
What are we asking the AI to understand about health and disease?
If disease is primarily represented as a label connected to a treatment pathway, AI will become increasingly efficient at label-to-treatment reasoning.
If instead disease is also understood as the evolving expression of interacting metabolic, nutritional, environmental, physiological, behavioral, and genetic factors, AI can assist with a much richer form of clinical reasoning.
This is why medical AI needs more than artificial intelligence.
It needs a sufficiently comprehensive model of medicine.
Integrative Orthomolecular Systems Medicine is one attempt to build that model.
I-OM.ai is our effort to explore what becomes possible when artificial intelligence is built upon it.
The future of medical AI should not simply automate what medicine already does.
It should help medicine reason better.
References
About the authors
Richard Z. Cheng, MD, PhD, is Editor-in-Chief of the Orthomolecular Medicine News Service (OMNS) and works in integrative orthomolecular and systems medicine.
Eric K. Cheng works in artificial intelligence and is co-developing the medical-AI reasoning framework described in the authors’ recent paper.
Disclosure: Richard Z. Cheng is associated with I-OM.ai. I-OM.ai is under development. The discussion above describes a conceptual framework and does not constitute validation of an AI system for independent clinical decision-making.