Newsletters

‹ Back

From Rare Disease Diagnosis to Learning Health Systems: Generative AI Enters Clinical Practice

ELHS Monthly Newsletter (June 2026)

Paving the Way for Global Health Equity with Generative AI and Learning Health Systems

Subscribe | All Issues

 

 

Dear Friends,

 

June marked an important milestone for healthcare AI. Rather than simply improving benchmark scores, several landmark studies demonstrated how generative AI is beginning to address one of medicine's greatest challenges—earlier diagnosis of rare and complex diseases—while building the technical foundation for evidence-based, continuously learning clinical care.

 

 

Earlier Diagnosis of Rare Diseases

The most significant advance came from the OpenAI–Harvard collaboration, which showed that LLM-assisted genomic reanalysis increased the diagnostic yield of previously unsolved rare disease cases across multiple cohorts. By combining genomic sequencing, longitudinal clinical records, and continuously evolving medical knowledge, AI can revisit difficult cases that previously remained undiagnosed, illustrating a shift from one-time diagnosis toward continuous diagnostic learning.

LLMs as Diagnostic Decision Support

A real-world study of complex polyneuropathies demonstrated that ChatGPT-4o achieved diagnostic performance comparable to non-specialist neurologists while generating stronger differential diagnoses and more appropriate confirmatory testing recommendations. Importantly, physicians frequently improved their own diagnostic decisions after reviewing AI suggestions, highlighting AI's potential to extend specialist-level expertise into routine clinical practice.

From Diagnosis to Disease Management

Clinical AI is rapidly expanding beyond diagnosis. Google's latest AMIE system demonstrated that conversational AI can support longitudinal disease management across multiple visits while remaining grounded in clinical practice guidelines and drug formularies, achieving performance comparable to primary care physicians in management reasoning and treatment planning.

Building Knowledge-Grounded AI

As LLMs enter clinical workflows, grounding AI recommendations in trusted medical knowledge has become increasingly important. New studies introduced ClinVec, a comprehensive embedding framework for standardized clinical concepts, while AMIE demonstrated retrieval from continuously updated clinical guidelines and drug formularies. Together, these advances improve the reliability, consistency, and trustworthiness of AI-assisted clinical reasoning.

Fine-Tuning Models for Clinical Reasoning

Several studies showed that specialized fine-tuning can substantially improve clinical reasoning. By combining physician expertise, electronic medical records, curriculum learning, reinforcement learning, and multimodal medical imaging, researchers developed models that more closely emulate how experienced clinicians learn, reason, and make diagnostic decisions in real-world practice.

From AI Models to Learning Health Systems

A comprehensive review of 140 studies concluded that successful healthcare AI depends not only on model performance but also on thoughtful workflow integration, appropriate clinician trust, and continuous evaluation in routine care. These findings reinforce the central vision of Learning Health Systems: embedding AI into everyday clinical workflows so that every patient encounter generates new evidence to improve care for future patients.

Looking Forward

Together, this month's publications suggest that healthcare AI is entering a new phase—from retrospective demonstrations of diagnostic capability toward continuously learning clinical systems that support earlier diagnosis, evidence-based decision making, and longitudinal patient care. We believe the next frontier is not simply developing better LLMs, but generating prospective real-world evidence that demonstrates how AI can safely help ordinary physicians diagnose rare diseases earlier at scale.

Below, I discuss with my ChatGPT copilot the major bottlenecks in rare disease diagnosis, the capabilities and limitations of today's frontier LLMs, and where AI can be most effectively integrated into routine clinical workflows. Please read on.

 

Best regards,

AJ

AJ Chen, PhD
Founder & PI, ELHS Institute
Silicon Valley, USA
https://elhsi.org/Newsletters
https://elhsi.com

 

~

 

From Page Mill

(Recent papers, news, and events showcasing the progress of GenAI and LHS) 

 

A. Jaech and Others. LLM-Assisted Reanalysis of Unsolved Rare Disease Genomes Increases Diagnostic Yield. NEJM AI 2026;3(7).

[2026/6] Across four cohorts, new local diagnoses were made in 10 of 100 rare disease neurodevelopmental cases (10.0%, [exact binomial: 95% confidence interval (CI), 4.9 to 17.6]), 4 of 61 neuromuscular cases (6.6%, [CI, 1.8 to 16.0]), 2 of 200 cases of sudden unexpected death in pediatrics (1.0% [CI, 0.1 to 3.6]), and 2 of 15 early psychosis cases (13.3% [CI, 1.7 to 40.5]) for an overall diagnostic yield of 18 of 376 (4.8%, [CI, 2.9 to 7.5]). We identified seven rediscoveries in which pathogenic or likely pathogenic findings had been established externally but were not available in the local research record at the time of review.

See OpenAI blog for more details on this Harvard collaboration study on GenAI for rare disease.

 

De Lorenzo, A., Moretti, G.S., Bertini, A. et al. ChatGPT in the diagnosis and management of complex polyneuropathies: comparative analysis with neurologists using real-world cases. npj Digit. Med. (2026).

[2026/6] Polyneuropathies are common and often require specialist expertise for accurate diagnosis. This study evaluated the diagnostic performance of ChatGPT-4o on real-world polyneuropathy cases, comparing it to peripheral neuropathy specialists and non-specialist neurologists. One hundred cases were selected from two tertiary centers in Milan, Italy. Standardized summaries included clinical, laboratory, and electrophysiological data. ChatGPT-4o was prompted to provide a leading diagnosis, two differentials, and a confirmatory test. Neurologists reviewed the same cases and generated comparable outputs, then could revise their responses after viewing ChatGPT-4o’s suggestions. ChatGPT-4o achieved 65.5% leading diagnosis accuracy, comparable to non-specialists (63.0%) but lower than specialists (74.0%, p = 0.002). For differential diagnoses, it outperformed non-specialists (82.0% vs. 77.5%, p = 0.043) and recommended more appropriate tests (68.0% vs. 53.0%, p < 0.001). After reviewing ChatGPT-4o outputs, non-specialists revised their assessments in 21.8% of cases, improving accuracy. ChatGPT-4o shows potential as a diagnostic aid, particularly in non-specialist or resource-limited settings.

 

Liévin, V., Palepu, A., Weng, WH. et al.  Towards Conversational AI for Disease Management.  Nature (2026).

[2026/6] We advance the previously demonstrated diagnostic capabilities of the Articulate Medical Intelligence Explorer (AMIE)1−3 through a new LLM-based agentic system optimized for multi-visit clinical management and dialogue. To ground its reasoning in authoritative clinical knowledge, AMIE leverages Gemini’s long-context capabilities4, combining in-context retrieval with structured reasoning to align its output with up-to-date clinical practice guidelines and drug formularies. In a randomized, blinded virtual Objective Structured Clinical Examination (OSCE) study, AMIE was compared to 21 primary care physicians (PCPs) across 100 multi-visit case scenarios designed to reflect UK NICE Guidance and BMJ Best Practice guidelines. AMIE was non-inferior to PCPs in management reasoning as assessed by specialists and scored better in both preciseness of treatments and investigations, and in its alignment with and grounding in clinical guidelines. To benchmark medication reasoning, we developed RxQA, a multiple-choice question benchmark derived from two national drug formularies (US, UK) and validated by board-certified pharmacists. Though AMIE and PCPs both benefited from the ability to access external drug information, AMIE outperformed PCPs on higher difficulty questions. While further research would be needed before real-world translation, AMIE’s strong performance across evaluations marks a significant step towards conversational AI as a tool in disease management.

 

Johnson, R., Gottlieb, U., Shaham, G. et al. Embeddings of clinical codes enable knowledge-grounded AI in medicine. npj Digit. Med. (2026).

[2026/6] Standardization of electronic health records (EHRs) has enabled the use of clinical codes in AI. We introduce ClinVec, an embedding store that provides embeddings for 153,166 clinical codes and concepts across eight vocabularies. ClinVec embeds ClinGraph, a knowledge graph with over 2 million edges tailored to clinical vocabularies used in EHRs. We validate the embeddings using an inter-institutional clinician panel and N = 3767 clinical term pairs spanning 11 disease areas, and we find that embedding similarity reflects clinical relatedness. We use ClinVec for knowledge injection in large language model medical question answering and for unsupervised patient stratification and risk prediction. By providing a shared representation of clinical concepts, ClinVec supports knowledge-grounded AI systems for modeling patients and populations.

 

Zhuang, Y., Wang, B., Yin, C. et al. Clinical large language model centered on electronic medical records. npj Digit. Med. (2026).

[2026/6] Motivated by the need to create a realistic medical practice environment, we propose that a combination of direct knowledge transfer from seasoned doctors and the strategic use of EMR can augment the abilities of LLM, enabling it to more closely mimic the clinical acumen of healthcare practitioners. To navigate the complexities of merging diverse instructional sources, we employ a curriculum learning approach during the fine-tuning process. Moreover, we advance our model’s performance by developing a reward system that incentivizes the alignment of the LLM’s outputs with the valuable attributes inherent in both doctors’ expertise, including diagnostic priors, risk thresholds, and heuristic saliencies accumulated from practice and EMR data. This is achieved through a novel reinforcement-learning approach. Besides, we introduce a new benchmark involving a comparative evaluation. We utilize a subjective evaluation system wherein experts critically assess the responses from a professional perspective as well. Our research underscores the potential of this hybrid model to serve as a robust tool in medical consultations, bridging the gap between artificial intelligence and real-world clinical practice.

 

Tan, M., Wu, Q., Zhang, B. et al. Learning like a radiologist: a medical vision-language model for radiological image analysis via curriculum learning. npj Digit. Med. (2026).

[2026/6] We propose RadiSim-CL, an MVLM trained via curriculum learning by simulating the three-phase pathway of a radiologist: foundational knowledge understanding, anatomical knowledge, and advanced diagnostic reasoning. To support this, we curate RadiSim, a 12-million image-text pair dataset aligned to these phases. We evaluate the model using a five-stage coarse-to-fine validation framework: (1) modality recognition, (2) anatomical recognition, (3) anatomical localization, (4) abnormality and disease diagnosis, and (5) disease differentiation and grading. This framework spans 24 zero-shot subtasks across MR, CT, and DR imaging. RadiSim-CL achieves comparable performance to state-of-the-art baselines in both foundational and anatomical tasks, and demonstrates superior capabilities in complex reasoning (e.g., an AUC of 0.953 for brain tumor diagnosis and an accuracy of 0.764 for meningioma grading). Ablation studies further confirm the curriculum’s effectiveness. RadiSim-CL thus offers a scalable, clinically aligned solution to enhance diagnostic precision.

 

Strong, J., Rogers, H., Sun, E. et al. Human-AI Collaboration in Healthcare: A Scoping Review. npj Digit. Med. (2026).

[2026/6] Artificial intelligence is increasingly embedded in clinical pathways, making effective human-AI collaboration (HAIC) a practical and policy priority in healthcare. We conducted a scoping review of empirical studies of HAIC in healthcare published from January 2015 to October 2025, using Joanna Briggs Institute methodology and PRISMA-ScR reporting. Of 17,463 records identified, 140 studies were included. Evidence was concentrated in diagnostic interpretation, with fewer studies in screening and triage, therapeutic decision-making, and administrative workflows. Effectiveness was defined inconsistently across task contexts and was usually assessed using short-term task-level metrics rather than patient or system outcomes. Most studies, particularly in diagnostic interpretation, reported benefits for human-AI teams. These depended on task fit, workflow integration, training, and appropriately calibrated trust. 

 

~

 

 

In the Eyes of GPTs

(AJ in conversation with ChatGPT to understand how machines view human healthcare)

 

 Let's discuss how to generate evidence for GenAI-assisted rare disease diagnosis. Could you briefly summarize the major challenges facing rare disease diagnosis today?

 

 ChatGPT

Rare disease diagnosis remains one of the most difficult problems in medicine despite advances in genomics and AI. The challenges can be grouped into several categories.

1. Clinical rarity and limited physician experience

Most physicians encounter only a small fraction of the approximately 7,000 known rare diseases during their careers. Many rare diseases present with nonspecific symptoms that resemble common conditions, making misdiagnosis or delayed diagnosis common. Patients often see multiple specialists over many years before receiving a correct diagnosis—the well-known "diagnostic odyssey."

2. Fragmented and longitudinal patient information

The diagnostic clues for rare diseases are often scattered across years of medical records, laboratory results, imaging studies, pathology reports, family history, and genetic testing performed at different institutions. No single clinical encounter contains all of the necessary evidence, making comprehensive synthesis difficult.

3. Rapidly expanding medical knowledge

Hundreds of new disease-gene associations and thousands of research papers are published each year. New pathogenic variants and treatment options continue to emerge. It is unrealistic for individual clinicians to remain current across the entire rare disease literature.

4. Genetic findings are difficult to interpret

Although sequencing technologies have become widely available, many patients receive inconclusive results, including variants of uncertain significance (VUS). Determining whether a genetic variant explains a patient's phenotype requires integrating genomic evidence with detailed clinical presentation and continuously updated scientific knowledge.

5. Phenotypic heterogeneity

The same genetic disorder may present differently among patients, while different rare diseases may produce highly similar symptoms. Variable age of onset, incomplete penetrance, and atypical presentations further complicate diagnosis.

6. Lack of systematic re-evaluation

Diagnosis is often treated as a one-time event. However, patients whose initial workup is inconclusive frequently benefit from periodic reanalysis as new clinical information accumulates and medical knowledge evolves. Most healthcare systems currently lack scalable mechanisms for continuous diagnostic reassessment.

7. Limited specialist access

Expert centers for rare diseases are concentrated in large academic medical centers. Many patients, especially those in rural or underserved regions, have limited access to clinicians with deep expertise in specific rare disorders.

8. Limited evidence for AI implementation

Recent studies—including the Harvard/OpenAI collaboration published in NEJM AI—demonstrate that LLMs can increase diagnostic yield by reanalyzing previously unsolved cases. However, most published evidence remains retrospective. Key questions remain unanswered:

• Which patients benefit most from AI assistance?
• At what point in the diagnostic journey should AI be introduced?
• How much diagnostic delay can AI reduce?
• Does earlier diagnosis improve patient outcomes and healthcare utilization?
• How should AI be integrated safely into routine clinical workflows?

 

Implications for Learning Health Systems

From an ELHS perspective, these challenges suggest that the primary bottleneck is no longer AI's ability to generate hypotheses. Rather, it is the lack of infrastructure to continuously generate real-world evidence demonstrating that AI can improve diagnosis safely, earlier, and at scale.

This shifts the research question from:

"Can AI diagnose rare diseases?"

to

"How can health systems continuously measure, validate, and improve AI-assisted diagnosis for every rare disease?"

That distinction is important because it frames rare disease diagnosis not as a one-time AI prediction problem, but as a continuous learning problem, which aligns naturally with the Learning Health System paradigm. This perspective also differentiates ELHS from most current studies, which focus on retrospective accuracy rather than prospective, continuously improving clinical implementation.

 

 The major bottleneck is the lack of rare disease expertise in ordinary clinical settings. In addition, many patients likely have disorders that do not yet fit any recognized disease entity, which further amplifies the difficulty for most physicians. To address this problem, every physician should have access to GenAI that can help detect rare diseases as early as possible. Do you think current frontier LLMs, such as ChatGPT and Gemini, already have this capability? If not, how soon do you think the best LLMs will be able to achieve it?

 

 ChatGPT

 

I think your first point is exactly right:

The bottleneck is not the lack of AI. The bottleneck is the lack of rare disease expertise at the point of care.

Most patients first present to primary care physicians, emergency physicians, hospitalists, pediatricians, or general specialists. These clinicians simply cannot remember thousands of rare diseases. Even specialists usually know only their own subspecialty. Consequently, many rare diseases are never considered in the differential diagnosis.

Can today's best LLMs detect all rare diseases?

I would answer:

No—not yet.

But the reasons are interesting.

What they already do surprisingly well

Today's frontier LLMs (ChatGPT, Gemini, Claude, etc.) already possess several capabilities that were unavailable even two years ago.

They can:

• recognize many thousands of diseases
• connect symptoms across multiple organ systems
• interpret long longitudinal histories
• integrate literature
• generate broad differential diagnoses
• explain their reasoning
• assist with genomic interpretation
• retrieve guideline information

The recent Harvard/OpenAI paper is a very good example. AI helped uncover diagnoses that had been missed despite years of evaluation.

So the capability is already clinically meaningful.

 

Why they still cannot solve the whole problem

The biggest limitation is not medical knowledge.

It is clinical evidence.

Diagnosis is fundamentally an inference problem under uncertainty.

The model never observes:

• the patient's future
• treatment response
• pathology obtained later
• evolution over months or years

A physician also doesn't know these things prospectively.

Therefore the model is estimating probabilities from incomplete evidence.

For common diseases this often works well.

For ultra-rare diseases, uncertainty becomes enormous.

 

Three major limitations remain

1. Missing patient information

The EMR rarely contains everything.

Symptoms may never have been documented.

Family history is incomplete.

Important physical findings were never examined.

AI cannot infer information that simply does not exist.

2. Unknown diseases

This is perhaps the biggest scientific limitation.

Many patients likely have disorders that have never been described.

No LLM can identify something that medicine itself has not yet defined.

The best it can do is recognize:

"This patient does not fit any known syndrome."

Ironically, that may become one of AI's most valuable contributions.

3. Lack of prospective learning

Current LLMs are essentially static.

Even retrieval-augmented systems only retrieve published knowledge.

They do not automatically learn from every patient encountered in routine care.

That is exactly the problem Learning Health Systems were designed to solve.

 

When will they be good enough?

This is obviously speculative.

My personal estimate is something like this.

Today (2026)

Frontier LLMs are already useful as differential diagnosis assistants.

They probably should be available to physicians managing difficult diagnostic cases.

Next 2–5 years

I expect models will become excellent at suggesting nearly all known rare diseases, provided they receive:

• longitudinal records
• laboratory data
• imaging
• genetics
• family history

This improvement will likely come less from larger models than from better integration of structured clinical data, retrieval, multimodal reasoning, and specialized fine-tuning.

Beyond that

The larger challenge is discovering diseases we do not yet recognize.

This requires something beyond a better chatbot.

It requires a continuously learning healthcare system.

 

Where I think ELHS becomes interesting

This is where I think your work differs from most current AI research.

Most people ask:

Can AI diagnose rare diseases?

I think the more important question is:

How can AI help every ordinary physician diagnose every rare disease earlier?

Those are fundamentally different problems.

The first is a benchmark problem.

The second is a health system problem.

To solve the second problem, an AI system needs more than high diagnostic accuracy. It must:

• continuously screen every patient rather than only selected difficult cases,
• identify patients whose presentations are inconsistent with common diseases,
• update diagnostic probabilities after every new clinical encounter,
• recommend appropriate confirmatory testing,
• learn from confirmed diagnoses across health systems, and
• generate real-world evidence demonstrating that earlier AI-assisted diagnosis improves patient outcomes.

 

This is why I think Learning Health Systems and frontier LLMs are complementary technologies rather than competing ones. LLMs provide the reasoning engine, while a Learning Health System provides the continuous feedback loop that allows AI-assisted diagnosis to be evaluated, improved, and safely deployed at scale.

In that sense, the long-term goal is probably not an AI that "knows every rare disease." It is an AI-enabled learning system that continuously improves its ability to recognize known diseases, identifies patients who may have previously unrecognized disorders, and generates the evidence needed to make earlier diagnosis routine rather than exceptional. That vision aligns well with the direction suggested by the recent NEJM AI rare disease study and other emerging work, while extending it from retrospective analyses to prospective clinical practice.

 

 You're right. What are the best integration points for GenAI within routine clinical workflows?

 

 ChatGPT

 

I think this is actually one of the most important research questions in healthcare AI.

Most people assume AI should be consulted only when physicians are stuck. I think that's too late. By then, the patient may already have spent years in the diagnostic odyssey.

Instead, the integration point should be determined by where the probability of preventing diagnostic delay is highest while minimizing unnecessary work.

I would organize routine care into five possible integration points.

Workflow stage	AI role	Potential impact	Feasibility
1. Before physician sees patient	Pre-visit screening	Moderate	High
2. During physician encounter	Real-time differential diagnosis	High	Moderate
3. After encounter	Silent quality assurance	Very High	Very High
4. Between visits	Continuous monitoring	Very High	Moderate
5. Population level	Registry surveillance	Extremely High	High

I actually think Stages 3 and 4 are the sweet spots.

 

1. Pre-visit AI screening

Patient completes:

• symptoms
• family history
• previous diagnoses
• uploaded records

AI prepares

• summarized history
• diagnostic hypotheses
• missing information
• suggested questions

The physician starts the visit already aware that "This patient has features compatible with Fabry disease." instead of discovering it 30 minutes later—or never. This saves physician time.

 

2. During the encounter

This is what most companies focus on today.

Physician enters findings.

AI continuously updates

• Differential diagnosis
• Recommended tests
• Rare disease alerts
• Drug cautions

This is useful. But physicians already have enormous cognitive load during visits.

Alert fatigue is a real concern.

 

3. Post-visit diagnostic QA (my favorite)

After physician signs the note,

AI quietly reviews

• Entire history
• Labs
• Imaging
• Medications
• Family history
• Previous visits

If nothing unusual: No notification.

If concern exists: Consider Gaucher disease.

Confidence: High.

Suggested next test: β-glucocerebrosidase activity.

The physician reviews later. No interruption. This resembles radiology double reading.

I think this workflow is much easier to deploy.

 

4. Continuous longitudinal monitoring

This is where ELHS becomes different. Every time new information arrives, such as lab, MRI, consultation, pathology, hospitalization, genetic report, etc, AI automatically recomputes Diagnostic probabilities.

 

For example:

2025     Probability of hereditary amyloidosis      2%

2026     Neuropathy appears     7%

2027     Cardiomyopathy            31%

2028     Family history updated                67%

AI sends notification. No physician could realistically recompute thousands of rare diseases after every new event. This is exactly what computers are good at.

 

5. Population surveillance

Imagine running AI every night on every patient in a health system.

Not for diagnosis. For prioritization.

Patients receive Rare Disease Risk Score. Only the top 0.1% are reviewed.

This resembles Diabetic retinopathy screening, Cancer screening.

Instead of waiting for referrals, AI actively finds patients who may benefit from specialist evaluation.

 

Which workflow will have the greatest impact?

I think the answer depends on the goal.

If the goal is physician assistance

Stage 2

If the goal is reducing diagnostic delay

Stage 4

If the goal is health system implementation

Stage 3

If the goal is population health

Stage 5

 

I think there is a missing workflow

One idea that I don't see discussed much in the literature is what I would call AI-assisted diagnostic surveillance.

Today, diagnosis is generally considered complete after each visit. In reality, many rare diseases evolve slowly over years, and the diagnostic evidence accumulates incrementally. Rather than treating diagnosis as a series of isolated decisions, AI could continuously monitor the patient's longitudinal record, updating diagnostic probabilities whenever new information becomes available. The physician would only be notified when the probability of a clinically important condition exceeds a predefined threshold or when a new combination of findings suggests that a previously overlooked rare disease should be reconsidered.

This shifts AI from being an on-demand consultant to becoming a continuous diagnostic safety net.

 

An even larger opportunity: integrating with Learning Health Systems

I think we can extend this one step further, and this may become the conceptual contribution of ELHS.

Current AI research focuses on helping physicians diagnose the current patient.

A Learning Health System asks a different question:

How does every diagnosed patient improve AI for the next patient?

That suggests a continuous learning loop:

1. Every patient is automatically screened after each encounter.
2. AI updates probabilities for both common and rare diseases.
3. Physicians review only high-confidence alerts.
4. Confirmed diagnoses become new real-world evidence.
5. The health system analyzes diagnostic delays, false positives, and patient outcomes.
6. AI models, prompting strategies, and diagnostic pathways are refined based on these real-world results.

In this framework, every confirmed rare disease not only benefits that individual patient but also strengthens the system's ability to recognize the next patient with the same—or even a related—condition. This transforms rare disease diagnosis from isolated clinical successes into a continuously improving, evidence-generating process, which is the essence of an AI-native Learning Health System. I think this is the direction where ELHS can make a unique contribution beyond current LLM-based diagnostic assistants.

 

~

 

🔹 ELHS Institute Resources 🔹

🤖 ELHS GenAI Platform

🌐 ELHS Institute Website

📰 ELHS Monthly Newsletters

▶️ ELHS Videos

👉 For Clinical AI technology support, contact us at support@elhsi.org 📩

 

~ the end ~