The 70% Paradox: Why EPR Data is Insufficient for the AI Era

Nelson Advisors
5 hours ago
9 min read

The "70% Paradox" represents a critical convergence of infrastructure barriers threatening healthcare's AI transformation: 70% of doctors identify poor electronic patient record (EPR) integration as the primary obstacle to AI adoption, while approximately 80% of healthcare data remains locked in unstructured formats that AI systems struggle to leverage effectively.

This paradox exposes a fundamental misalignment between healthcare's data infrastructure and the requirements of artificial intelligence systems, creating what medical informatics experts describe as "broken digital foundations" that undermine the promise of AI-driven.

The Infrastructure Crisis: Physicians' Perspective

A January 2026 survey by the Royal College of Physicians delivered a stark verdict on healthcare's AI readiness. Among 541 physicians surveyed, 68% believe the NHS lacks the digital infrastructure to introduce AI effectively, with 70% specifically citing inability to integrate AI tools with existing digital systems such as EPRs as the main barrier to adoption.

Dr. Anne Kinderlerer, digital health clinical lead at the RCP, articulated the challenge bluntly: "Physicians think the NHS is fundamentally unprepared for AI because its digital foundations are broken. Many systems can't talk to each other; infrastructure is outdated and there is poor standardisation in the function of electronic patient records".

This infrastructure deficit creates cascading problems. The systems cannot communicate effectively, infrastructure remains outdated, and EPR standardization is poor, generating inefficiencies, administrative burden, cognitive load on overstretched workforces, patient care delays, and increased clinical risk.

The RCP warns against "simply chasing emerging innovative technologies like AI at the expense of optimising existing digital systems," emphasisng that AI tools will not deliver meaningful change without "effective implementation on strong foundations of interoperable digital systems and complete datasets".

The situation extends beyond technical integration challenges. Two-thirds of doctors (66%) report having no access to AI training, despite 79% wanting such support. This disconnect between demand and institutional capability reveals a healthcare system simultaneously unprepared for AI technically and educationally, while 69% of UK doctors already use personal access to AI tools like ChatGPT and Microsoft Copilot for clinical questions, a worrying indicator of unregulated AI adoption in clinical contexts.

The Data Completeness Crisis

The infrastructure problems compound a deeper issue: EPR data quality and completeness remain fundamentally inadequate for AI applications. While the exact "70%" figure in completeness varies by definition, research consistently demonstrates that the proportion of "complete" records in clinical databases is far lower than nominal totals, and heavily dependent on how completeness is defined.

A comprehensive study analysing electronic health record completeness identified four prototypical definitions: documentation completeness (whether expected notes and reports exist), breadth completeness (coverage across required data types), density completeness (sufficient volume of data points), and predictive completeness (adequate information for specific analytical tasks).

The findings were sobering: only 48.3% of approximately 3.9 million patients had at least one visit with expected documentation recorded. When completeness was defined as density (at least 15 laboratory results or medication orders adjusted for temporal variance), only 11.8% had complete records. Using breadth criteria (five key data types: date of birth, sex, medication order, laboratory test and diagnosis), only 11.4% of patients had complete records.

The implications for quality measurement are profound. Research examining electronic quality measures calculated using single-site EHR data versus longitudinal data from health information exchanges found that quality measure calculations changed significantly, affecting 19% of patients, when including HIE data sources. For measures including diagnoses as part of calculation (diabetes and hypertension measures), over 24% of measure calculations changed when HIE data were included. This demonstrates that individual EHRs often contain fundamentally incomplete data for clinical decision-making and AI training.

The Unstructured Data Challenge

Compounding the completeness problem is data structure. Approximately 80% of healthcare data exists in unstructured formats, narrative physician notes, clinical summaries, radiology reports, medical images and patient narratives. While structured fields like laboratory values and billing codes can be readily parsed by AI systems, the rich contextual information necessary for nuanced clinical decision-making lives predominantly in unstructured text and images,

This unstructured data contains critical information absent from structured fields. As one healthcare data analytics organisation noted, "approximately 80% of clinical data in electronic health records is found in unstructured physician notes, including data needed to generate insights on clinical outcomes and determine longitudinal trends in patient care". The challenge intensifies because unstructured data resists easy extraction and analysis, physician shorthand varies by institution and provider, abbreviations lack standardisation, and critical context often exists in what clinicians don't document rather than what they do.

Medical imaging data exemplifies the scale challenge. While imaging constitutes 80% of all clinical content, a single chest X-ray may be 15 megabytes, a 3D mammogram can reach 300 megabytes and digital pathology files can be 3 gigabytes, equivalent to a high-definition full-length movie. Healthcare organisations accumulate this unstructured data quickly and haphazardly, creating storage burdens and analytical challenges that AI systems must somehow navigate.

The $70 Billion Evidence Gap

The data infrastructure crisis intersects with what medical informaticist Blackford Middleton terms "The $70 Billion Paradox", a massive AI healthcare market built on alarmingly thin clinical evidence. The healthcare AI market has reached $70 billion, with over 1,200 FDA-cleared AI/ML tools and 350,000+ consumer health apps flooding the market. Yet the evidence base supporting these tools reveals systematic gaps:

Fewer than half of FDA device summaries report their study design
More than half omit sample size entirely
Less than 1% report patient outcomes
95% lack demographic data
91% include no bias assessment whatsoever

This represents an industry where the vast majority of approved tools have never demonstrated they actually help patients and almost no systematic understanding exists of whether they work equitably across populations. Over 95% of FDA cleared AI/ML medical devices used the 510(k) pathway, demonstrating "substantial equivalence" to existing devices rather than proving clinical efficacy through rigorous trials. This creates "equivalency all the way down", systems approved by comparison to other tools that were themselves often approved on thin evidence.

The validation crisis extends to performance in real-world settings. Research demonstrates that 81% of AI models experience performance degradation when deployed in external datasets, with 24% showing substantial decreases and 12% experiencing complete failure. A systematic review of AI algorithms for diagnostic analysis of medical imaging found that only 6% of 516 eligible published studies performed external validation. Without rigorous external validation involving adequately sized datasets from institutions other than those providing training data, AI systems fail to demonstrate they can handle variations in patient demographics and disease states encountered in real clinical settings.

The Real World Performance Gap

The controlled-environment versus real-world performance gap emerges as particularly concerning. While frontier AI models like o1-preview demonstrate superhuman performance on diagnostic reasoning tasks in controlled, text-based environments, often exceeding physician accuracy on management tasks and emergency case diagnosis, their performance degrades precisely when clinical medicine becomes difficult.

When data is complete and problems well-defined, large language models excel. When faced with uncertainty, missing information, or changing clinical context, conditions describing the majority of real clinical encounters, they break down. The models exhibit "miscalibrated confidence," remaining certain even when reasoning has gone awry, unlike human experts who naturally moderate confidence when situations become ambiguous

Controlled testing reveals this brittleness. When researchers introduced "none of the other answers" (NOTA) testing, changing expected patterns of multiple-choice responses, model accuracy dropped from 81% to 43%. This suggests sophisticated pattern recognition rather than genuine clinical reasoning; the models learned what "right" answers typically look like rather than how to reason through clinical problems. Performance degrades significantly when models shift from processing static clinical vignettes to engaging in multi-turn conversation, the format more closely mimicking actual patient encounters.

Data Silos and Interoperability Barriers

The infrastructure supporting EPR systems compounds these challenges through persistent data silos and interoperability failures. Healthcare organisations face fragmented data residing in disparate systems, with different departments using incompatible platforms that store data in varying formats. Even when organisations use a single EHR across multiple locations, individual providers document differently, creating significant variation in data elements and formatting.

Legacy systems present particular obstacles. Many healthcare providers operate on systems implemented decades ago, creating complex technological ecosystems never designed for modern integration. Different departments use specialized systems tailored to specific functions, acquisitions and mergers result in multiple incompatible platforms operating simultaneously, and technical debt accumulates as short-term solutions layer upon one another. The fragmented nature of this data creates significant barriers to AI implementation.

A joint study by Bain & Company and KLAS Research found that approximately 70% of healthcare organizations were directly affected by the Change Healthcare cyberattack in February 2024, highlighting both system interconnectedness and vulnerability when data exchange channels are disrupted. The incident underscored how fragile healthcare data infrastructure remains despite years of interoperability efforts.

The AI Implementation Gap

These convergent challenges produce what researchers term the "AI implementation gap", the distance between what is developed and what is successfully deployed in clinical practice. Healthcare ranks at the bottom for embedding AI into practice at only 12%, compared to the average of 19% or 31% for leading industries. Approximately 80% of healthcare AI projects fail to scale beyond pilot phase.

Three critical barriers drive this gap. First, physicians who would use AI systems often don't see the benefit, whether due to poorly defined problems during development, insufficient communication of model accuracy during implementation, or failure to address actual physician problems. Second, information technology difficulties emerge, translating model information to computers, acquiring real-time patient data, and displaying results to physicians requires significant IT expertise and time. Third, change management issues arise from failure to properly engage users and explain why and how AI models can be utilised.

The data reality gap manifests acutely during implementation. Proof-of-concept projects rely on carefully curated datasets, clean and standardised. Production systems must contend with fragmented data across multiple EHR systems, inconsistent formats, and inevitable gaps in clinical workflows. A diagnostic AI achieving 95% accuracy on laboratory datasets might struggle to maintain 70% accuracy when processing real patient data.

The Path Forward: Building Better Foundations

Addressing the 70% Paradox requires systematic investment in digital infrastructure before attempting widespread AI deployment. The Royal College of Physicians recommends that governments invest in well-functioning digital infrastructure to bring IT systems up to date and establish central banks of NHS-approved algorithms, AI tools, and patient-facing apps meeting national standards. The focus must shift toward EPR content models improving integration with AI tools, rather than pursuing AI innovation while neglecting foundational systems.

Healthcare organisations need integrated data platforms supporting seamless data exchange, standardised data formats, and ensured data security and compliance. Establishing robust data governance frameworks and adhering to industry standards such as HL7 (Health Level Seven) and FHIR (Fast Healthcare Interoperability Resources) facilitates data exchange and interoperability across disparate systems. API-first integration strategies using FHIR standards establish common protocols for data exchange while reducing integration complexity.

For AI systems themselves, the evidence standards must become substantially more rigorous. Healthcare leaders should demand safety evaluations, conversational degradation testing, NOTA testing and validation in conditions resembling actual clinical environments, not just multiple-choice benchmark performance. A clinical trials-informed framework mirroring FDA-regulated trial phases, safety, efficacy, effectiveness, and post-deployment monitoring, ensures AI solutions undergo rigorous validation at each stage.

The near-term AI adoption strategy should prioritise narrow over broad applications. The most immediate value comes from AI tools tightly scoped to specific clinical domains and contexts, which are easier to validate, integrate, and explain to clinicians. Focus should target what clinicians actually want: AI reducing administrative burden through documentation assistance, prior authorisation, inbox management, and care coordination, tasks that burn out clinicians but remain underrepresented in research and commercial products.

Conclusion: Choosing the Right Path

The 70% Paradox illuminates healthcare's current crossroads. One path leads to AI that genuinely augments clinical care, carefully validated and thoughtfully integrated, making medicine better for patients and practitioners. The other leads to premature deployment, preventable harm, loss of clinical trust, and regulatory backlash potentially setting the field back years.

The gap between AI capabilities and healthcare's ability to evaluate, regulate, and integrate these systems safely creates both opportunity and risk in equal measure. After four decades in health IT, Middleton warns: "The AI systems being deployed today are more powerful than anything we've seen before. The gap between controlled-environment performance and real-world readiness is also larger than anything we've seen before".

Addressing this paradox demands recognition that technology alone cannot solve healthcare's challenges. Without complete datasets, interoperable systems, standardised documentation, and rigorous validation frameworks, even the most sophisticated AI systems will fail to deliver their promised benefits. The 70% Paradox serves as both warning and roadmap, identifying precisely where healthcare must invest to build the foundations that will finally make the AI era possible.

Nelson Advisors > European MedTech and HealthTech Investment Banking

Nelson Advisors specialise in Mergers and Acquisitions, Partnerships and Investments for Digital Health, HealthTech, Health IT, Consumer HealthTech, Healthcare Cybersecurity, Healthcare AI companies. www.nelsonadvisors.co.uk

Nelson Advisors regularly publish Thought Leadership articles covering market insights, trends, analysis & predictions @ https://www.healthcare.digital

Nelson Advisors publish Europe’s leading HealthTech and MedTech M&A Newsletter every week, subscribe today! https://lnkd.in/e5hTp_xb

Nelson Advisors pride ourselves on our DNA as ‘Founders advising Founders.’ We partner with entrepreneurs, boards and investors to maximise shareholder value and investment returns. www.nelsonadvisors.co.uk

#NelsonAdvisors #HealthTech #DigitalHealth #HealthIT #Cybersecurity #HealthcareAI #ConsumerHealthTech #Mergers #Acquisitions #Partnerships #Growth #Strategy #NHS #UK #Europe #USA #VentureCapital #PrivateEquity #Founders #SeriesA #SeriesB #Founders #SellSide #TechAssets #Fundraising #BuildBuyPartner #GoToMarket #PharmaTech #BioTech #Genomics #MedTech