Newsletters

Your E-mail *

The Hidden Gap in Healthcare AI Adoption

 
ELHS Newsletter 2025-12-15

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

Subscribe | All Issues

 

Dear Friends,

 

A recent JAMA report calculated biweekly trends in the percentage of firms responding “yes” to using AI in healthcare, based on the U.S. Census Bureau’s Business Trends and Outlook Survey (BTOS), as a proxy for AI adoption in the healthcare sector. Between September 2023 and May 2025, the mean adoption rate was 5.9% and increased steadily over time, reaching 8.3% in 2025. Plotting this adoption trend alongside those of other industries provides valuable context for discussion (see graph below).

Recognizing the urgency of accelerating AI adoption, I provided a comment at the end of the JAMA report.

First, traditional clinical AI tools trained on EHR data have often failed during deployment, resulting in limited real-world adoption. The steady uptake of image-based clinical AI tools and non-clinical applications has contributed to the persistently single-digit adoption rate. What is promising, however, is that large language model–based generative AI (GenAI) has become widely available and easy to use, enabling broad democratization in health care. I believe GenAI has likely contributed to the steeper adoption trajectory observed by the authors.

Second, the reported trend likely masks two distinct adoption patterns because the survey data do not distinguish between clinical AI and non-clinical AI. Non-clinical AI can be adopted rapidly once technically validated, given minimal regulatory constraints. In contrast, clinical AI requires additional clinical evidence—often through clinical trials—even after technical performance has been demonstrated, due to stringent regulatory and ethical standards. To date, only a small number of real-world clinical studies of GenAI have been published, indicating that clinical evidence remains sparse. Consequently, the adoption of clinical GenAI is expected to remain near zero. The reported overall adoption trend therefore likely reflects primarily non-clinical AI applications. Dedicated, disaggregated adoption data for clinical AI are needed to accurately depict AI use in health care.

Third, appreciating the masked near-zero adoption of clinical GenAI due to the current evidence gap has important implications. Given existing inefficiencies in evidence generation, accumulating sufficient clinical evidence to transform health care may take decades—far longer than the widely anticipated 5–10-year timeframe. The recent JAMA Summit Report on AI clearly identifies this evidence-generation bottleneck as the most significant barrier to clinical AI adoption. Meeting this challenge requires accelerating evidence generation at scale through systematic approaches. Technically, responsible GenAI for all diseases should be freely accessible to low-resource frontline physicians so that GenAI clinical evaluation studies can be initiated in parallel across diverse settings worldwide. At the system level, more efficient clinical study designs that leverage real-world data can provide timely early-stage evidence, followed by continuous validation within the Learning Health System framework outlined by the US National Academy of Medicine. In addition, aligning GenAI implementation in routine care with physician incentives will be crucial to ensuring broad, sustainable adoption.

Below, I continue the discussion with Copilot ChatGPT on GenAI adoption in healthcare. Please read on.

Warm regards,
AJ

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

 

~

(source: JAMA network article)

 

From Page Mill

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

Nguyen TD, Whaley CM, Simon K, et al. Adoption of Artificial Intelligence in the Health Care Sector. JAMA Health Forum. 2025;6(11):e255029. doi:10.1001/jamahealthforum.2025.5029
[2025/11] Between September 2023 and May 2025, the mean (SD) AI use in health care by firms was 5.9% (1.6%; 119 300 firm-level responses) and increased over time. In 2025, AI use in health care (8.3%) was still lower than in other sectors, such as finance and insurance (11.6%); education (15.1%); professional, scientific, and technical services (19.2%); and information services (23.2%). The estimated break point in AI use trends in the health care sector was December 30, 2024, to January 12, 2025. 

Jain SS, Sun T, Pierson E, et al. Detecting Transthyretin Cardiac Amyloidosis With Artificial Intelligence: A Nonrandomized Clinical Trial. JAMA Cardiol. Published online November 10, 2025. doi:10.1001/jamacardio.2025.4591
[2025/11] How well does an artificial intelligence (AI)–augmented screening program detect transthyretin cardiac amyloidosis (ATTR-CM)? In this nonrandomized clinical trial conducted at multiple sites in a large health system, an AI model identified 1471 patients as positive for potential ATTR-CM, with 256 patients meeting study criteria and eligible for trial inclusion. On physician notification, 50 patients underwent further diagnostic testing, and ATTR-CM was diagnosed in 24 (48%), with treatment initiated in 21 patients (88%). Meaning: AI-augmented clinical programs may identify undiagnosed ATTR-CM and reduce diagnostic delays, but prospective randomized trials are needed to determine if outcomes are improved.

Boustani MA, Ben Miled Z, Owora AH, et al. Digital Detection of Dementia in Primary Care: A Randomized Clinical Trial. JAMA Netw Open. 2025;8(11):e2542222. doi:10.1001/jamanetworkopen.2025.42222
[2025/11] To evaluate the effect of the combined approach of QDRS plus PDM in improving the annual rate of new documented ADRD diagnoses in primary care practices. This randomized clinical trial including 5325 patients found that the odds of new ADRD diagnoses was 31% higher in the clinics randomized to the combined QDRS and PDM approach compared with usual care clinics. This randomized clinical trial found that the combined approach was effective at scale for the early detection of ADRD in primary care settings. 

Laura L. Adams, et al. Advancing the National Academy of Medicine Artificial Intelligence Code of Conduct for Health and Medicine. NEJM AI 2025;2(11). DOI: 10.1056/AIp2500651
[2025/11] In 2024, the U.S. National Academy of Medicine published, for public comment, a draft AI Code of Conduct framework, with Code Principles and Code Commitments, synthesizing the field’s best thinking to align industry players regarding the responsible application of health AI. National experts and patient advocates were then convened to gather perspectives on real-world applications to address the Code Commitments: advance humanity, ensure equity, engage impacted individuals, improve workforce well-being, monitor performance, and innovate and learn. The National Academy of Medicine AI Code of Conduct special publication is now available.

Li, S., Fang, X., Jin, Y. et al. Improving diagnostic accuracy in preoperative glioma classification: performance of knowledge-enhanced large language models compared with radiologists. npj Precis. Onc. 9, 383 (2025). https://doi.org/10.1038/s41698-025-01171-6
[2025/11] Three radiologists extracted VASARI features, while three LLMs (GPT-4, Claude3.5-Sonnet, Claude3.0-Opus) analyzed these features using standard input-output or knowledge-enhanced prompting incorporating diagnostic guidelines. Knowledge-enhanced prompting consistently outperformed standard prompting, improving diagnostic consistency (intra-model agreement: Sonnet κ = 0.91, Opus κ = 0.92, GPT-4 κ = 0.72). For diffuse versus circumscribed classification, senior radiologists (AUC = 0.88) and Claude3.5-Sonnet with knowledge-enhanced prompting (AUC = 0.84) performed similarly (p > 0.05). LLM assistance significantly improved junior radiologists’ performance, with AUC increases from 0.77 to 0.83 (p = 0.026). Knowledge-enhanced LLMs demonstrate diagnostic performance comparable to experienced radiologists and improve junior accuracy, suggesting potential as decision-support tools requiring radiologist oversight.

Albaqshi, A., Ko, J.S., Suh, C.H. et al. Evaluating diagnostic accuracy of large language models in neuroradiology cases using image inputs from JAMA neurology and JAMA clinical challenges. Sci Rep 15, 43027 (2025). https://doi.org/10.1038/s41598-025-06458-z
[2025/11] In image-only tasks, LLM accuracy in identifying pathologic locations ranged from 21.5% (28/130) to 63.1% (82/130). LLMs exhibit strong diagnostic performance with clinical text, yet their ability to interpret complex radiologic images independently is limited. 

Zhou, S., Wang, J., Xu, Z. et al. Uncertainty-aware large language models for explainable disease diagnosis. npj Digit. Med. 8, 690 (2025). https://doi.org/10.1038/s41746-025-02071-6
[2025/11] To fill this gap, we introduce ConfiDx, an uncertainty-aware large language model fine-tuned with diagnostic criteria. We formalized the task of uncertainty-aware diagnosis and curated richly annotated datasets that reflect varying degrees of diagnostic ambiguity. Evaluating on real-world datasets demonstrated that ConfiDx excelled in identifying diagnostic uncertainties, achieving superior diagnostic performance, and generating trustworthy explanations for diagnoses and uncertainties. Moreover, ConfiDx-assisted experts outperformed standalone experts by 10.7% in uncertainty recognition and 26% in uncertainty explanation, underscoring its substantial potential to improve clinical decision-making.

Badani, A., de Moraes, F.Y., Vollmuth, P. et al. AI and innovation in clinical trials. npj Digit. Med. 8, 683 (2025). https://doi.org/10.1038/s41746-025-02048-5
[2025/11] Clinical trials face persistent challenges in cost, enrollment, and generalizability. This perspective examines how artificial intelligence (AI), large language models (LLMs), adaptive trial designs, and digital twins (DTs) can modernize trial design and execution. We detail AI-driven eligibility optimization, reinforcement learning for real-time adaptation, and in silico DT modeling. Methodological, regulatory, and ethical hurdles are addressed, emphasizing the need for validated, scalable frameworks to enable responsible and widespread integration.

Mahajan, A., Bates, D.W., Liao, K.P. et al. Advancing rheumatic disease care through generative artificial intelligence. Nat Rev Rheumatol (2025). https://doi.org/10.1038/s41584-025-01310-0
[2025/11 comment] Generative artificial intelligence promises to reshape clinical care in rheumatology by supporting diagnostic reasoning, treatment planning and patient communication. Yet its potential rests on careful validation, transparent integration and thoughtful collaboration that strengthens, rather than substitutes, the human expertise central to patient care.

Koul, Arman et al. Synthetic data, synthetic trust: navigating data challenges in the digital revolution. The Lancet Digital Health, Published November 30, 2025. DOI: 10.1016/j.landig.2025.100924
[2021/11] Although synthetic data address crucial shortages of real-world training data, their overuse might propagate biases, accelerate model degradation, and compromise generalizability across populations. A concerning consequence of the rapid adoption of synthetic data in medical AI is the emergence of synthetic trust—an unwarranted confidence in models trained on artificially generated datasets that fail to preserve clinical validity or demographic realities. In this Viewpoint, we advocate for caution in using synthetic data to train clinical algorithms. We propose actionable safeguards for synthetic medical AI, including standards for training data, fragility testing during development, and deployment disclosures for synthetic origins to ensure end-to-end accountability. These safeguards uphold data integrity and fairness in clinical applications using synthetic data, offering new standards for responsible and equitable use of synthetic data in health care.

 

~

 

In the Eyes of GPTs

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

 

 What is the estimated adoption rate of AI use in the health care sector?

  ChatGPT

Here’s a snapshot of the estimated adoption rate of AI in the healthcare sector based on recent surveys and reports (as of 2024–2025):

📊 Current Adoption Estimates

1. Organization-level adoption

  • ~22% of healthcare organizations worldwide have implemented domain-specific AI tools in clinical or administrative workflows — a large increase compared with prior years (≈7× over 2024, 10× over 2023). Forbes
  • Adoption rates vary by provider type: health systems ~27%, outpatient providers ~18%, payers ~14%. Forbes

2. Individual clinician usage

  • 66% of U.S. physicians reported using healthcare AI tools in practice by 2024 (up from ~38% in 2023). American Medical Association
  • Surveys indicate widespread use of AI for tasks like documentation, predictive analytics, and routine clinical workflows.

3. Broader organizational sentiment

  • Some reports show very high awareness and engagement: 86% of healthcare organizations surveyed already leverage AI in some capacity (including exploratory or pilot use).

4. Historical and sector comparison

  • Sector-wide AI use in health care (measured as firms reporting meaningful AI use) has been measured around 8–9% in 2025 — lower than many other industries but rising.

📈 Future Adoption Projections

Growth outlook

  • AI adoption in healthcare is expected to grow rapidly with increasing investment, digital transformation priorities, and generative AI applications.
  • One industry analysis estimates overall AI adoption in health care may rise from ~10–30% today to ~30–45% by 2030 as technologies mature and regulatory pathways clear.

🧠 Takeaways

✅ Adoption varies by definition:

  • Formal, full-scale deployment of AI solutions in clinical workflows is still in early stages (~20–30% of organizations).
  • Individual clinician usage and pilot/experimental use are much higher (often >60%).
  • Overall awareness and strategic prioritization is near universal, with many organizations planning or expanding AI use.

📌 In summary: AI adoption in health care is rapidly accelerating — though institution-wide deployment remains in the early-to-mid adoption phase (~20–30%), clinician use and exploratory adoption are much higher (often ~60–85%) and expected to grow meaningfully by 2030.

 

 There appears to be a disconnect. In the medical literature, there are very few non-imaging AI studies that use real patient data in real clinical settings, indicating that there is limited clinical evidence to support the use of EHR-based AI in clinical decision-making. Without such evidence, physicians cannot reasonably use EHR-based AI in routine care. Why, then, are the reported AI adoption rates much higher than the plausible rate of AI adoption in actual clinical practice?

  ChatGPT

You are exactly right — and this “disconnect” is real, structural, and widely misunderstood. The short answer is:

Most reported “AI adoption” in health care does not mean evidence-based, EHR-driven clinical decision support used in routine patient care.

Below is a clear breakdown of why adoption rates appear high while real clinical use remains rare, aligned with what you are observing in the medical literature.

1. “AI adoption” is defined very broadly — and often misleadingly

Most surveys and industry reports count any of the following as AI adoption:

  • Voice dictation / ambient clinical documentation
  • Scheduling, billing, coding, claims optimization
  • Imaging AI (radiology, pathology)
  • Chatbots for patient messaging
  • Administrative workflow automation
  • Experimental pilots or sandbox testing
  • Physician personal use of GenAI (e.g., ChatGPT for summaries)

👉 None of these require clinical-grade evidence from EHR-based outcome studies.

EHR-based AI for diagnosis, prognosis, or treatment decisions is only a tiny subset of what is being labeled “AI adoption.”

2. Clinical evidence ≠ usability ≠ availability

You are pointing to the correct bottleneck:

Without real-world clinical evidence, physicians cannot responsibly use EHR-based AI in routine care.

In practice, however, adoption reports often measure:

  • Availability (“Is an AI tool present?”)
  • Awareness (“Have you tried AI?”)
  • Intent (“Are you planning to adopt AI?”)

They do not measure:

  • Prospective clinical validation
  • Impact on patient outcomes
  • Safety across populations
  • Workflow-embedded decision authority

This explains why the literature is sparse while surveys look optimistic.

3. Imaging AI is the exception — and skews perception

Imaging AI does have:

  • Clear labels
  • Standardized inputs
  • Narrow tasks
  • FDA pathways
  • Strong retrospective and prospective evidence

As a result:

  • Imaging AI is used in real clinical workflows
  • It dominates “successful AI” narratives
  • It inflates perceived overall AI adoption

📌 Non-imaging, EHR-based clinical AI does not share these advantages, and its evidence base is far thinner.

4. Physicians are users, not adopters, in many surveys

When surveys say:

“60–70% of physicians use AI”

They often mean:

  • Physicians personally use GenAI to summarize notes
  • Draft patient letters
  • Look up differential diagnoses
  • Assist with education or research

This is not the same as:

  • AI generating recommendations inside the EHR
  • AI influencing orders, diagnoses, or treatment plans
  • AI being accountable for clinical decisions

👉 Personal cognitive assistance ≠ institutional clinical adoption

5. Regulatory and liability reality blocks true adoption

For EHR-based AI in decision-making, physicians face:

  • Unclear liability if AI is wrong
  • Lack of FDA-cleared indications
  • No reimbursement mechanisms
  • No professional society guidelines
  • No standard evaluation metrics
  • No longitudinal safety data

So even if AI is technically available, it is clinically unusable.

6. The literature reflects reality; surveys reflect aspiration

Your observation aligns with reality:

Dimension

Reality

Published non-imaging AI studies using real EHR data in routine care

Very few

Prospective outcome-based evidence

Rare

AI embedded in real clinical decision workflows

Minimal

Responsible physician adoption for EHR-based decisions

Low single digits

Meanwhile:

Survey metric

What it really captures

“AI adoption”

Any AI exposure or pilot

“Physician use”

Personal GenAI usage

“Organizational AI”

Admin + imaging + pilots

Bottom line (one-sentence summary)

Reported AI adoption rates are high because they count administrative tools, imaging AI, pilots, and personal GenAI use — while true, evidence-based EHR-driven clinical decision support remains rare due to the absence of scalable real-world clinical evidence.

 

 It seems necessary to report AI adoption separately for non-clinical versus clinical use, and for imaging versus non-imaging AI applications. Tracking published AI studies across these categories could serve as a useful proxy for better understanding true trends in AI use and adoption in health care.

  ChatGPT

You’re making a methodologically sound and publishable point. What you are proposing is essentially a redefinition of “AI adoption” as an evidence-tracked phenomenon rather than a self-reported one. Below I formalize this into a clear framework + proxy-based measurement model that you can directly use in a manuscript.

1. Why adoption must be stratified by AI function

“AI adoption” in health care is not a single construct. It spans domains with fundamentally different:

  • Risk profiles
  • Evidence requirements
  • Regulatory pathways
  • Clinical accountability

Pooling them together guarantees systematic overestimation of clinical adoption.

Proposed stratification (minimum viable)

Category

Core purpose

Evidence required for routine use

Non-clinical AI

Operational efficiency

Low

Clinical imaging AI

Pattern recognition

Moderate–high (task-specific)

Clinical non-imaging AI (EHR-based)

Decision-making

Very high

2. A publication-based proxy for AI adoption

You are correct: published studies using real patient data in real clinical settings are the best available proxy for true adoption—especially for clinical AI.

Why publication tracking works as a proxy

  • Requires real data access
  • Requires institutional approval
  • Requires workflow embedding (even minimally)
  • Requires outcome definition
  • Requires reproducibility and peer scrutiny

This aligns with how evidence-based medicine defines readiness for use.

One-sentence takeaway

Stratifying AI adoption by function and tracking published real-world clinical studies provides a more accurate proxy for true AI adoption in health care than self-reported surveys, revealing that evidence-based use of non-imaging, EHR-driven clinical AI remains rare.

 

~

 

🚀 Healthcare GenAI Toolbox

 

🤖 ELHS GenAI Solution

Empowering physicians to accelerate the responsible use of GenAI in healthcare through three growth stages:

🚀 1. GenAI Quick Start – Access GenAI agents preclinically validated with synthetic data to enhance routine clinical decision-making.

🛠️ 2. LLM Customization – Fine-tune LLMs with your own data within ELHS units to adapt GenAI to your clinical setting.

🌍 3. Broader Impact – Scale your ELHS units into research networks so more physicians can use your LLMs, collaboratively improving care quality and outcomes for larger populations.

🧠 Meta Open-Source AI

Creating a positive future by making innovation available to all and enabling advancements that can enhance our lives and spur economic growth. Open-source AI provides tools for more people to build awesome things and expand our horizons.

 

~

 

🔹 ELHS Institute Relevant Resources 🔹

 

🤖 ELHS GenAI Platform

🌐 ELHS Institute Website

📰 ELHS Newsletters

▶️ ELHS Videos

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

 

~ the end ~