‹ Back

ELHS Newsletter | October 2026

From LHS Vision to Reality: GenAI Enabling Continuous Learning in Routine Clinical Care

Subscribe

 

 

Dear Friends,

 

Where Does the Learning Health System Vision Stand Today?

Nearly 20 years after the Learning Health System (LHS) vision emerged, are we finally ready to make it a reality?

In a New England Journal of Medicine perspective, The Learning Health System, Made Computable, Kenneth Mandl and colleagues revisit the fundamental LHS vision: a health system that continuously learns from patient care, identifies what works, detects harms, updates clinical guidance, and applies accumulated knowledge to improve care for the next patient.

Their central message is compelling: What could not be operationalized at scale for nearly two decades can now become computable.

I strongly agree. Advances in generative AI (GenAI) are creating unprecedented opportunities to operationalize task-oriented LHS units directly within routine clinical workflows.

The next challenge is no longer simply demonstrating AI capabilities. It is enabling ordinary clinical teams to continuously generate, evaluate, and apply clinical evidence to improve early diagnosis and personalized care.

This requires AI scientists and clinical teams to work together. Our ELHS team is pursuing this vision by democratizing GenAI-powered LHS capabilities for everyday clinical practice.

October Research Highlights

🔬 1. GenAI Improving Diagnosis

• Microsoft's sequential diagnosis system achieved 83.2% diagnostic accuracy on NEJM clinicopathological conference cases and 80.3% on emergency department cases, demonstrating the potential for more accurate, cost-conscious diagnostic reasoning. Science
• Google's AMIE conversational AI demonstrated prospective feasibility in primary care. Among physicians who reviewed AMIE transcripts, 33 of 44 found them helpful for visit preparation. The Lancet
• EpiScreen, an LLM-based tool analyzing electronic health records, achieved AUCs of 0.875–0.980 for early epilepsy detection and improved neurologist performance in clinician-AI collaboration. npj Digital Medicine

🏥 2. Clinical AI Deployment: Evidence Gaps and Opportunities

• A scoping review of healthcare AI identified major translation barriers, including limited prospective validation, workflow integration, equity assessment, and economic evidence. Only 4.7% of reviewed records measured workflow or process outcomes. JMIR
• On-premises medical AI agents demonstrated diagnostic accuracies of 83.8%–90.04% on benchmark tasks, supporting the feasibility of locally controlled clinical AI systems. Nature Medicine
• A study of OpenEvidence usage found that clinicians primarily used AI-enabled clinical information retrieval for treatment and management (33.1% of queries), compared with diagnosis (13.9%). JAMA

🧬 3. Synthetic Patient Data for Clinical AI

A systematic review of 87 studies highlighted the growing use of LLM-generated synthetic clinical text for data augmentation, clinical documentation, and privacy-preserving medical AI development. Evaluation of utility and privacy remains essential. npj Digital Medicine

🤖 4. Healthcare AI Foundation Models

Google introduced MedGemma, an open medical vision-language model family supporting diverse text and imaging applications. Its performance and adaptability could lower barriers to developing specialized medical AI tools. Nature Medicine

🎓 5. Rethinking Medical Education in the AI Era

UCSF authors proposed an educational triage for physician training: strengthen skills AI makes more valuable, develop new competencies for physician-AI collaboration, and reconsider skills AI makes less essential. JAMA

Looking Ahead: From AI Performance to Continuous Clinical Learning

Collectively, these studies signal a major transition: medical AI is advancing from isolated demonstrations of model performance toward integration into clinical workflows.

Yet the evidence gap remains substantial. High benchmark accuracy does not automatically translate into earlier diagnoses, better treatment decisions, or improved patient outcomes.

The next frontier is to establish scalable, task-oriented GenAI-LHS units that enable ordinary clinical teams to:

• Detect disease signals earlier and support personalized treatment.
• Continuously evaluate AI performance and physician decisions.
• Learn from each patient to improve care for the next.

The LHS vision is becoming technically achievable. Our shared responsibility is to make it clinically effective, accessible, and equitable.

After exploring the latest research highlights in the “From Page Mill” section, I invite you to follow my conversation with ChatGPT Copilot in “In the Eyes of GPTs” for a deeper discussion of how GenAI can help turn the Learning Health System (LHS) vision into reality.

 

Best regards,

AJ

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

Democratizing GenAI and Learning Health Systems for Global Health Equity.

 

~

 

From Page Mill

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

K.D. Mandl, J.M. McGinnis, and M.M. Bertagnolli. The Learning Health System, Made Computable. N Engl J Med 2026;395:1775-1777. DOI: 10.1056/NEJMp2610734.

[2026/10] A learning health system identifies what works, detects the harms it causes, updates guidance as evidence accrues, applies what it learns to benefit people throughout the system, and can demonstrate that patient n+1 was treated using knowledge gained from the experience of treating the first n patients. For nearly 20 years, these functional requirements couldn’t be satisfied at scale. This definition can now be operationalized.

Nori H, Daswani M, Kelly C, Lundberg S, et al. Sequential diagnosis with language models. Science. 2026 Oct 22;394(6822):423-428. doi:10.1126/science.aeh7930.

[2026/10] We explored the capabilities of frontier language models in sequential diagnosis, a diagnostic paradigm that captures the iterative nature of clinical reasoning. Given a brief clinical presentation, clinicians selectively acquired information to reach a diagnosis, enabling assessment of diagnostic accuracy and workup cost. We developed an interactive benchmark and an artificial intelligence (AI) system that simulates a panel of physicians, continually revising the differential diagnosis and selecting cost-effective tests. Using this benchmark, the system achieved 83.2% accuracy on New England Journal of Medicine clinicopathological conference cases and 80.3% on MIMIC-IV-ED emergency department admissions, outperforming generalist physicians (36.1 and 46.8%, respectively). Although these findings were obtained in a controlled benchmark setting, they highlight the promise of language models for supporting more accurate, lower-cost diagnoses.

Brodeur PG, Koshy JM, Palepu A, et al. Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study. Lancet. 2026 Oct 24;408(10565):1732-1744. doi:10.1016/S0140-6736(26)01535-7.

[2026/10] In this prospective, single-centre, single-arm feasibility study, English-speaking patients aged at least 18 years interacted with the Articulate Medical Intelligence Explorer (AMIE) up to 5 days before a single-complaint urgent primary care appointment. Findings: From April to November, 2025, 114 patients were enrolled with 98 completing both the AMIE interaction and the PCP appointment. Zero conversation safety stops were required on the basis of prespecified criteria. Safety supervisors noted one hallucination and added clinical information in five interactions. AMIE's conversations were rated favourably in 87-100% of cases (17 criteria) by clinical evaluators, and 48-96% (16 criteria) by patients. Patient attitudes towards AI improved after interacting with AMIE and remained elevated after the patient's visit with their physician. PCPs completed post-surveys in 60 of 98 cases, including 44 cases in which they reviewed the AMIE transcript before the visit. PCPs found AMIE helpful for visit preparation in 33 of 44 cases and reported that it might have changed their behaviour in 25 of 44 cases. Interpretation: Although further research is needed, this study shows the initial feasibility of conversational AI in a real-world setting-assessed via conversation safety and quality, as well as user acceptance-and represents a crucial step towards clinical translation.

Zhou, S., Yu, K., Zhan, Z. et al. Early epilepsy detection from electronic health records with large language models. npj Digit. Med. (2026).

[2026/10] Epilepsy and psychogenic non-epileptic seizures often present with similar seizure-like manifestations but require fundamentally different management strategies. Misdiagnosis is common and can lead to prolonged diagnostic delays, unnecessary treatments, and substantial patient morbidity. Although prolonged video-electroencephalography is the diagnostic gold standard, its high cost and limited accessibility hinder timely diagnosis. Here, we developed a low-cost, effective approach, EpiScreen, for early epilepsy detection by utilizing routinely collected clinical notes from electronic health records. Through fine-tuning large language models on labeled notes, EpiScreen achieved an AUC of up to 0.875 on the MIMIC-IV dataset and 0.980 on a private cohort of the University of Minnesota. In a clinician-AI collaboration setting, EpiScreen-assisted neurologists outperformed unaided experts by up to 14.3%. Overall, this study demonstrates that EpiScreen supports early epilepsy detection, facilitating timely and cost-effective screening that may reduce diagnostic delays and avoid unnecessary interventions, particularly in resource-limited regions.

Xu Y, Veillard J, Kong JD. AI for Health Care Quality and Patient Safety: Scoping Review of Diagnostic, Predictive, and Decision Support Applications. J Med Internet Res. 2026 Sep 21;28:e95157. doi:10.2196/95157.

[2026/9] We identified 5 recurring barriers to translating health care AI into clinical value: limited prospective external validation, workflow-mediated effectiveness, infrastructure fragmentation, equity and generalizability deficits, and economic and governance uncertainty. English-language records published from January 1, 2017, to April 30, 2026. Prospective and external validation were uncommon, and retrospective results often lacked evidence of improved care. Workflow or process measures were recorded in only 13 (4.7%) records, and explicit equity or subgroup analyses were recorded in 20 (7.3%) records, indicating limited measurement of real-world impact. Economic and governance evidence was also thin, with formal economic evaluations found in 4 (1.5%) records and little attention to monitoring models after deployment. These gaps indicate limited evidence on how AI performs in clinical practice.

Zhang L, Wölflein G, Ferber D, et al. On-premise medical AI agents for reliable clinical decision-making. Nat Med. 2026 Oct;32(10):3705-3716. doi:10.1038/s41591-026-04609-x.

[2026/10] Here we developed and evaluated a fully on-premise clinical agent that couples local operational control with a multi-perspective reliability framework to support selective autonomy. Across two Medical Information Mart for Intensive Care IV (MIMIC-IV)-derived benchmarks, the agent achieved 90.04% accuracy on a seven-disease task and 83.8% accuracy on a four-disease task, approaching a cloud baseline on the primary benchmark. 

Alshaikhdeeb, B., Lorant, A., Hemedan, A.A. et al. Generation of synthetic clinical text: a systematic review. npj Digit. Med. (2026).

[2026/10] Generating medical synthetic text represents an effective solution for clinical NLP issues like sparsity and privacy. This paper conducts a systematic review on synthetic clinical text generation by formulating quantitative analysis to three research questions concerning the generation’s purpose, generation techniques, and the evaluation methods. We searched four databases including PubMed, Scopus, ScienceDirect and Web of Science for publications associated with the topic and identified 87 relevant articles out of 1148. Results showed that the generation purposes were dedicated to Trustworthy, Medical Assistance, and Language Resource themes. Prompting LLMs was the main source of synthetic text (40 articles). GPTs were the main predominant technique used to generate the text (63 articles), especially ChatGPT (47 articles). Additionally, there were four main aspects of evaluation, including similarity, privacy, structure, and utility, where utility was the prevailing method for assessment (52 articles). The synthetic medical text empowered data augmentation and clinical documentation. Using text alteration, controlling real text portions, and federated learning, generative models can produce privacy-preserving medical text. This accelerates the development of new digital health applications.

Patel VR, Finlayson SG, Liu-Hass M, et al. Clinician Use of an Artificial Intelligence-Enabled Clinical Information Platform. JAMA. 2026 Oct 15:e2621868. doi:10.1001/jama.2026.21868.

[2026/10] This cross-sectional study analyzed clinicians with National Provider Identifier-verified accounts of a commonly used AI-enabled clinical information platform, OpenEvidence, from January 1, 2024, through June 30, 2026. Most queries addressed clinical decision-making (51.9%), particularly treatment and management (33.1%), whereas diagnosis accounted for 13.9% of queries. 

Sellergren, A., Kazemzadeh, S., Mahvar, F. et al. An open vision-language model for diverse medical applications. Nat Med 32, 3695–3704 (2026).

[2026/10] We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3. MedGemma demonstrates advanced medical understanding and reasoning across images and text and multiple medical imaging domains, exceeding the performance of similarly sized generative models while maintaining the general capabilities of the Gemma base models. For out-of-distribution tasks, MedGemma achieves improvements of 2.6–10% in medical image question answering, 15.5–18.1% in chest X-ray finding classification and 10.8% in agentic evaluations compared with the base models. Our results show that fine-tuning MedGemma can be more effective than fine-tuning the base Gemma 3 model for medical tasks, particularly in the setting of limited training data. We additionally introduce MedSigLIP, a medically tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and, as an encoder, achieves performance comparable to or better than that of many specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with the potential to accelerate medical research and the development of downstream applications.

Burk-Rafel J, Wachter RM. Physician Training for the AI Era: Rethinking a Century-Old Wager. JAMA. Published online October 14, 2026. doi:10.1001/jama.2026.20340

[2026/10] Given finite curricular time, we propose an educational triage: emphasize those skills AI makes more valuable, cultivate those it makes newly necessary, and carefully de-emphasize those it makes less important, using recently published educational competencies as a scaffold. Translating this triage into practice depends on 4 conditions. First, training for physician-AI collaboration belongs in a longitudinal thread, not a standalone course. Second, attendings need curiosity and empathy more than technical expertise. Third, assessments must isolate what the learner contributed. Fourth, curricula must be revised at the pace clinical medicine is changing; piloting, measuring, and adjusting cannot wait for a once-a-decade redesign.

 

~

 

 

 

In the Eyes of GPTs

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

 Is GenAI Finally Making the 20-Year Learning Health System Vision Achievable?

 

 ChatGPT

Yes. Generative AI is bringing the Learning Health System (LHS) vision closer to practical realization than ever before. The fundamental breakthrough is not simply that AI can answer medical questions, but that it can now help transform routine clinical data into actionable knowledge, support clinical decisions, and enable continuous learning within everyday care workflows.

For nearly two decades, the LHS vision has been constrained by three major barriers: fragmented clinical data, the high cost of generating clinical evidence, and the difficulty of translating new knowledge into routine practice. GenAI is beginning to overcome all three.

 

1. What technological breakthroughs have changed LHS feasibility?

Five advances are particularly important.

1. Clinical understanding of unstructured data. Large language models (LLMs) can interpret clinical notes, laboratory results, imaging reports, and longitudinal patient histories. This makes it increasingly feasible to extract useful clinical knowledge from electronic health records without requiring every clinical observation to be manually structured.
2. Advanced clinical reasoning and prediction. Frontier LLMs can generate differential diagnoses, identify disease signals, recommend diagnostic investigations, and support treatment decisions. Microsoft's recent sequential diagnosis study, achieving 83.2% accuracy on challenging NEJM cases, illustrates this emerging capability, although benchmark performance does not establish clinical effectiveness.
3. AI agents that integrate multiple clinical tasks. Rather than performing isolated predictions, GenAI agents can coordinate information extraction, risk assessment, evidence retrieval, clinical reasoning, and report generation. Such agents provide a foundation for automating components of the LHS learning cycle.
4. Accessible and customizable medical AI models. Open models, including MedGemma, and locally deployed AI agents create opportunities for healthcare organizations to develop specialized clinical applications without relying exclusively on expensive centralized infrastructure.
5. Rapid, scalable evidence generation. GenAI can help clinical teams analyze accumulated patient experiences, identify diagnostic or treatment gaps, generate hypotheses, and evaluate decision-support performance. This can substantially reduce the technical barriers to conducting iterative clinical studies, although rigorous validation and appropriate study designs remain essential.

 

2. How can GenAI transform the traditional LHS into task-oriented LHS units?

Traditional LHS initiatives often require institution-wide data infrastructure, specialized research teams, and substantial organizational resources. These requirements have limited adoption, particularly in ordinary hospitals and resource-constrained settings.

GenAI creates an opportunity to shift from large, institution-centered LHS programs toward smaller, task-oriented LHS units (LHSUs) embedded directly in routine clinical workflows.

Each LHSU focuses on a specific clinical problem, such as accelerating diagnosis of difficult diseases, predicting rheumatoid arthritis complications, or optimizing medication selection.

The essential transformation is that clinical care and clinical learning become connected activities, rather than separate processes.

Each patient encounter contributes new observations, physician feedback, and outcomes. These data can be used to evaluate and improve the AI agent and clinical workflow. Validated improvements can then inform subsequent patient care.

Importantly, continuous learning does not mean uncontrolled automatic model updates. Changes must be evaluated for accuracy, safety, clinical benefit, and generalizability before deployment.

 

3. What remains necessary to realize the LHS vision?

GenAI makes the LHS vision increasingly technically feasible, but it does not automatically make it clinically effective.

Three challenges remain critical:

• Prospective clinical evidence: Demonstrating that GenAI-assisted decisions improve diagnostic timeliness, treatment appropriateness, patient outcomes, or other meaningful measures.
• Reliable learning mechanisms: Establishing trustworthy reference standards, clinician feedback, longitudinal outcome measurement, and safeguards against errors and bias.
• Democratized implementation: Enabling ordinary clinical teams to deploy, evaluate, and improve task-specific AI agents with affordable infrastructure and appropriate governance.

The recent literature reinforces this distinction. AI capabilities are advancing rapidly, but prospective validation, workflow integration, and evidence of real-world clinical benefits continue to lag.

 

The Outlook: From Learning Health Systems to Learning Clinical Teams

The most consequential impact of GenAI may not be the creation of a single, highly intelligent medical AI system. Instead, it may be empowering thousands of ordinary clinical teams to become continuously learning teams, each addressing the diagnostic and treatment challenges encountered in its own practice.

This represents a potential democratization of clinical evidence generation: moving from research concentrated in a relatively small number of major institutions toward continuous, evidence-driven improvement across everyday healthcare settings.

The 20-year LHS vision is becoming technically achievable. The next breakthrough will be demonstrating that GenAI-powered learning cycles can consistently improve care for patient n+1 using validated knowledge gained from patients 1 through n.

That is the defining promise—and the next major test—of GenAI-powered Learning Health Systems.

 

 Why Hasn't Impressive GenAI Diagnostic Performance Translated into Widespread Improvements in Patient Care?

 

 ChatGPT

The central challenge is that high AI diagnostic accuracy does not automatically translate into better clinical decisions or patient outcomes. GenAI has demonstrated remarkable capabilities in medical reasoning, but most evaluations still measure what AI can predict, rather than what AI actually improves in routine clinical care.

For example, Microsoft's recent study in Science reported 83.2% diagnostic accuracy on challenging NEJM clinicopathological conference cases. However, this performance was measured in a controlled benchmark, not through prospective evaluation of patient outcomes.

Similarly, a 2026 scoping review in JMIR identified persistent gaps in prospective validation, workflow integration, equity assessment, and economic evidence. Only 4.7% of reviewed records measured workflow or process outcomes.

These findings highlight a fundamental gap between AI performance, clinical adoption, and demonstrated patient benefit.

 

1. What are the major barriers?

Five barriers are particularly important.

1. Retrospective accuracy does not establish clinical effectiveness. Most diagnostic benchmarks evaluate AI using completed patient cases containing information accumulated over the entire diagnostic journey. In real clinical practice, physicians must make decisions with incomplete information. An AI system that correctly identifies a final diagnosis retrospectively may not recognize the disease early enough to change care.
2. AI performance is rarely evaluated against diagnostic delays. Conventional accuracy metrics do not measure when AI could identify a disease. For patients with difficult-to-diagnose conditions, recognizing the correct diagnosis months earlier may be more consequential than simply achieving a correct prediction after all diagnostic evidence becomes available.
3. Clinical workflow integration remains difficult. AI recommendations must be available at the appropriate decision point, supported by patient-specific evidence, and presented in a format physicians can efficiently evaluate. Standalone AI demonstrations rarely address these operational requirements.
4. Physician adoption and trust are essential. Even accurate AI predictions cannot improve care unless physicians appropriately incorporate them into clinical decisions. Clinical teams need reliable evidence, interpretable recommendations, and mechanisms to accept, modify, or reject AI suggestions.
5. Continuous learning and evidence generation are missing. Most clinical AI systems operate as prediction tools rather than learning systems. They lack systematic mechanisms to capture physician feedback, evaluate downstream outcomes, detect errors, and improve subsequent decisions.

 

2. How can GenAI-powered LHS units bridge this gap?

The GenAI-powered Learning Health System Unit (LHSU) offers a practical framework for connecting AI capabilities with clinical evidence generation and continuous improvement.

Instead of evaluating AI as an isolated diagnostic tool, an LHSU integrates four interconnected activities within routine clinical care.

First, focus on specific clinical problems. Each LHSU addresses a measurable clinical objective, such as accelerating diagnosis of rare diseases, predicting rheumatoid arthritis complications, or improving disease-risk assessment and treatment selection.

Second, embed AI into physician decision-making. GenAI agents analyze routinely collected clinical information and generate actionable recommendations. Treating physicians review these recommendations, make independent clinical decisions, and document their acceptance, modification, or rejection.

Third, generate prospective clinical evidence. The LHSU systematically measures whether AI-assisted care improves clinically meaningful outcomes. Examples include diagnostic accuracy at earlier time points, time to diagnosis, agreement with expert reference standards, treatment appropriateness, and patient outcomes.

Fourth, continuously improve through validated learning cycles. Physician feedback, expert adjudication, and patient outcomes identify opportunities to improve AI agents and clinical workflows. Updated systems are evaluated before deployment, allowing evidence from previous patients to inform care for subsequent patients.

 

3. What should we measure beyond diagnostic accuracy?

The next generation of clinical AI studies should move beyond conventional model-performance benchmarks.

Conventional AI evaluation	GenAI-LHS clinical evaluation
Final diagnostic accuracy	Earlier recognition of disease signals and confirmed diagnosis
Retrospective test performance	Prospective impact on clinical decisions
Agreement with reference diagnoses	Changes in physician decisions and appropriateness
Model sensitivity and specificity	Patient safety, treatment outcomes, and clinical benefit
One-time model evaluation	Performance across validated learning cycles

 

4. From AI benchmarks to real-world clinical evidence

Recent advances suggest that GenAI has reached a level of capability that justifies broader prospective evaluation. However, the next phase requires shifting research priorities from developing increasingly powerful models toward demonstrating measurable clinical benefits.

Task-oriented LHSUs could make this transition more accessible by enabling ordinary clinical teams to evaluate GenAI within their existing workflows, rather than depending entirely on large, resource-intensive clinical trials or specialized AI research centers.

Importantly, observational learning cycles and single-arm feasibility studies can establish adoptability, workflow integration, and preliminary evidence of clinical impact. Stronger causal claims about patient benefit require appropriately designed comparative studies, with randomized trials when feasible.

 

The Outlook: Clinical Evidence Is the Missing Link

The most important question in medical AI is changing from "How accurately can AI diagnose?" to "How much can AI improve patient care?"

GenAI-powered LHSUs provide a promising pathway to answer this question by connecting AI predictions, physician decisions, clinical outcomes, and continuous learning within routine care.

The opportunity is particularly significant for early diagnosis and personalized treatment, where even modest improvements in decision timing and appropriateness could have substantial clinical value.

Ultimately, the success of GenAI in healthcare should be measured not by how often AI agrees with a final diagnosis, but by how reliably it helps clinicians make better decisions earlier—and demonstrates improved outcomes for patients.

This is how impressive AI performance can become meaningful clinical evidence, and how the Learning Health System vision can begin delivering measurable benefits in everyday patient care.

 

 Can Ordinary Clinical Teams Build and Operate Their Own GenAI-Powered Learning Health System Units?

 

 ChatGPT

Yes. GenAI is creating an unprecedented opportunity to democratize Learning Health Systems (LHS), enabling ordinary clinical teams—not just major academic medical centers—to continuously generate clinical evidence and improve patient care.

Historically, implementing an LHS required substantial investments in data infrastructure, specialized informatics expertise, clinical research personnel, and institutional coordination. These requirements have limited the adoption of LHS approaches in routine healthcare settings.

Today, large language models (LLMs), customizable AI agents, and increasingly accessible computing infrastructure can substantially lower these barriers.

The key innovation is to shift from building large, institution-wide LHS platforms toward establishing small, task-oriented GenAI-powered LHS units (LHSUs) that address specific clinical problems within existing care workflows.

 

1. What are the minimum requirements for a GenAI-powered LHSU?

An ordinary clinical team does not need to develop its own foundation model or establish a sophisticated AI research center.

A practical LHSU can begin with five essential components.

Component	Minimum practical requirement
Clinical expertise	A clinical champion and participating physicians who identify a specific clinical problem, review AI recommendations, and evaluate outcomes.
GenAI capabilities	Access to a reliable LLM or customizable AI agent capable of analyzing clinical records, generating predictions, and providing evidence-based recommendations.
Data infrastructure	Secure access to relevant patient information, structured AI reports, physician feedback, and clinical outcomes.
Evidence generation	Defined clinical endpoints, reference standards, prospective data collection, and basic statistical evaluation.
Governance and safety	Patient privacy protections, appropriate ethics oversight, human supervision, AI performance monitoring, and procedures for managing errors and model updates.

These requirements can be implemented incrementally, beginning with a single clinical task and a small participating team.

The initial objective would not be to automate clinical decisions, but to support physicians while systematically generating evidence about whether AI improves those decisions.

 

2. How can ordinary clinical teams operate a continuous learning cycle?

A GenAI-powered LHSU can integrate four activities into routine patient care.

The GenAI-LHSU Continuous Learning Cycle

(1). Clinical Data

Collect relevant patient information during routine care

(2). GenAI Decision Support

Generate disease predictions, risk assessments, and recommendations

(3). Physician Decisions & Feedback

Review AI reports, make clinical decisions, document adoption and outcomes

(4). Evidence Generation & Improvement

Evaluate performance, detect errors, validate improvements

Apply validated learning to subsequent patients

This approach transforms routine clinical encounters into opportunities for systematic learning without requiring physicians to conduct separate, labor-intensive research activities for every patient.

For example, a physician reviewing an AI-generated diagnostic report could document whether the report introduced a previously unconsidered diagnosis, changed a risk assessment, or influenced a treatment decision.

When these observations are collected systematically, they become valuable real-world evidence about clinical AI performance and adoption.

However, physician acceptance alone does not establish clinical benefit. Independent reference standards, appropriate comparisons, and patient outcomes are necessary to determine whether AI-assisted decisions are actually better.

 

3. How can AI scientists democratize LHS implementation?

The most important contribution of AI scientists may be to make advanced GenAI capabilities accessible, customizable, and evaluable by ordinary clinical teams.

Rather than developing separate AI systems for every disease or institution, AI scientists can create reusable GenAI-LHS frameworks that clinical teams adapt to their own problems.

Four priorities are especially important.

1. Develop reusable clinical AI agents. Provide configurable agents for common tasks such as diagnostic reasoning, risk prediction, treatment recommendation, clinical documentation, and longitudinal outcome analysis. Clinical teams should be able to customize these agents without extensive programming expertise.
2. Simplify clinical workflow integration. Build lightweight interfaces that accept routine clinical information and return structured AI reports similar to familiar laboratory or diagnostic reports. Integration should minimize additional physician workload rather than create new administrative burdens.
3. Automate clinical evidence generation. Provide standardized tools for collecting physician feedback, measuring diagnostic accuracy, comparing decisions with reference standards, evaluating patient outcomes, and monitoring performance across learning cycles.
4. Make responsible deployment affordable and accessible. Support secure cloud-based and on-premises options, open medical AI models, transparent evaluation methods, and practical governance templates. This is particularly important for smaller hospitals and resource-constrained healthcare systems.

The goal is not to eliminate the need for AI scientists. It is to shift their role from building isolated AI applications toward enabling many clinical teams to develop and evaluate their own task-specific AI solutions.

 

4. What would democratized GenAI-LHS implementation look like?

Consider three examples:

• Early diagnosis LHSU.
• Predictive care LHSU.
• Treatment optimization LHSU.

These examples illustrate how LHS implementation can become problem-driven, clinician-led, and supported by reusable AI technology, rather than dependent on large institutional research programs.

 

5. What challenges remain?

Although GenAI can substantially lower technical barriers, successful LHS implementation still requires clinical leadership, organizational commitment, and appropriate oversight.

Three safeguards are essential.

First, clinical accountability must remain with qualified healthcare professionals. GenAI should support clinical reasoning and decision-making, not independently replace physician judgment.

Second, continuous learning must be evidence-based. AI agents should not automatically change clinical recommendations based on unverified feedback. Updates require appropriate validation, safety assessment, and performance monitoring.

Third, democratization must not compromise patient privacy, safety, or scientific rigor. Smaller clinical teams need access to practical ethics, governance, and evaluation resources alongside AI technology.

Importantly, the ability to operate a GenAI-LHSU is not equivalent to proving that it improves patient outcomes. Feasibility and adoption can be evaluated in initial studies, while stronger evidence of effectiveness requires appropriately designed prospective comparisons.

 

The Outlook: Every Clinical Team Can Become a Learning Team

The traditional LHS vision focused on transforming healthcare organizations into systems that continuously learn from patient experiences.

GenAI introduces a potentially more scalable pathway: empowering individual clinical teams to become continuously learning units, each focused on solving the problems encountered in everyday patient care.

AI scientists can accelerate this transformation by providing reusable agents, accessible infrastructure, standardized evidence-generation tools, and practical governance frameworks.

Clinical teams contribute the indispensable expertise: identifying meaningful problems, interpreting AI recommendations, making treatment decisions, and evaluating patient outcomes.

Together, these capabilities could enable thousands of clinical teams to generate evidence and improve care continuously.

The ultimate promise of democratized GenAI-LHS is not that every physician becomes an AI scientist, but that every clinical team gains the capability to learn systematically from its patients and apply validated knowledge to improve care for the next patient.

This is how the Learning Health System vision can move beyond major academic medical centers and become a practical foundation for more equitable, evidence-driven healthcare.

 

~

 

 

🔹 ELHS Institute🔹

Democratizing GenAI and LHS to Advance Global Health Equity

🤖 ELHS GenAI Platform

🌐 ELHS Institute Website

📰 ELHS Monthly Newsletters

▶️ ELHS Videos

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

 

~ the end ~