What is EHRead? Clinical NLP for regulatory-grade RWE

What is EHRead? Clinical NLP for regulatory-grade RWE

Most real-world evidence derived from electronic health records rests on an incomplete picture of the patient. The reason is not a lack of data – it is a failure to read it.

The clinical record that structured data cannot capture

Clinical NLP (cNLP) is the computational discipline that reads and structures free-text medical language: the kind written by clinicians for clinicians, not generated for machines. Savana cNLP (EHRead) is Savana’s proprietary clinical NLP engine, trained exclusively on real EHRs – more than seven billion clinical documents across seven languages. This distinction matters. Most RWE pipelines extract from structured fields: diagnosis codes, procedure codes, laboratory values. EHRead extracts from the clinical narrative where disease activity, treatment response, adverse events, and comorbidity burden are actually documented.

Clinicians write in language because clinical reality is too complex, too contextual, and too temporal to fit a drop-down menu. The note is not a supplement to the clinical record; it is the clinical record. Discharge summaries, consultation reports, referral letters, radiology findings – these documents are the primary record of a patient’s disease trajectory, their response to treatment, and the clinical reasoning behind every decision. No structured field in the EHR was designed to capture what those notes contain.

The consequence for real-world evidence is profound. Disease activity, treatment rationale, response assessment, adverse event characterisation, comorbidity burden – all of this sits in free text that conventional data pipelines cannot read. The result is systematic incompleteness. No statistical adjustment corrects for information that was never extracted.

Why general-purpose NLP fails clinical evidence

The intuitive response to this problem is to apply a language model. Modern NLP tools can read text, extract entities, and summarise documents. The issue is that clinical language is not general language – and that difference is not superficial.

Negation and context

A clinician writing ‘no evidence of recurrence’ means something precise. A model trained on news articles and encyclopaedias does not reliably understand negation in that clinical context. The abbreviation ‘MS’ means multiple sclerosis in a neurology note and mitral stenosis in a cardiology one – the same characters carry different meanings depending on specialty, institutional convention, and the section of the document in which they appear.

Temporal structure

‘The patient had previously responded to treatment X’ requires understanding chronology. ‘Rule out Y’ requires distinguishing a diagnostic hypothesis from a confirmed finding. These distinctions are not edge cases – they define how clinical reasoning is documented. A general-purpose model produces outputs that appear fluent but are clinically unreliable.

Regulatory consequence

For most applications, unreliability is an inconvenience. For evidence that will support a regulatory submission, a post-market safety signal, or an HTA dossier, it is disqualifying. Regulatory-grade evidence requires that every extraction is traceable, auditable, and reproducible – properties that fluency without clinical accuracy cannot provide.

How EHRead works: extraction built on clinical text

Savana cNLP (EHRead) is trained exclusively on real EHRs – more than seven billion clinical documents across seven languages. EHRead does not apply a general language model to clinical text; it was built from clinical text. Every concept it extracts – diagnoses, disease progression markers, treatment responses, adverse events, temporal relationships, and negations – maps to a standard clinical ontology and links back to the source phrase in the original document.

“Source traceability is not a feature. It is the property that makes EHRead outputs usable as evidence rather than as data. An extraction that cannot demonstrate its clinical origin is, from a regulatory perspective, a claim without provenance.”

EHRead produces that provenance by construction – each output carries the audit trail that regulatory and HTA reviewers require to accept the evidence it supports. Learn more about how Savana cNLP operates within the broader evidence infrastructure on the Savana cNLP product page.

Why training on authentic clinical text is non-negotiable

The distinction between training on authentic clinical text and training on proxies matters more than it might appear. Proxies include annotated public datasets, synthetic EHR records, and academic corpora curated for NLP benchmarks. These do not reproduce the linguistic patterns of clinical documentation under real conditions: the abbreviation conventions that vary by institution, the shorthand that varies by specialty, the documentation styles that reflect how clinicians write under real workload pressures.

A model trained on proxies may perform well on benchmark tests and poorly on the documentation it encounters in actual deployment. That performance gap compounds when evidence generated from those extractions targets regulatory submission – where the cost of an inaccurate extraction is measured not in model accuracy metrics, but in the defensibility of the evidence itself.

EHRead was trained on authentic, longitudinal EHRs written by clinicians across European health systems – in English, Spanish, Portuguese, French, Italian, German, and Dutch. That breadth matters for multi-country evidence generation, where the same evidence objectives must extract the same clinical concepts from documentation written in different languages and under different clinical conventions. Every EHRead extraction is auditable and reproducible: a third party reviewing a dRWE evidence generation activity can trace each analysis variable back to the clinical text that generated it.

That reproducibility is the connection between the clinical NLP layer and the regulatory-grade evidence requirements that FDA, EMA, and HTA bodies apply. For a deeper look at what those requirements mean in practice, see Regulatory-grade RWE: what FDA and EMA actually require.

EHRead and large language models: complementary, not interchangeable

Large language models can read clinical text, summarise patient records, and answer questions about clinical documentation with genuine capability. That capability is real, and it is worth acknowledging before explaining why it does not replace validated clinical NLP at the extraction layer.

LLMs perform well at reasoning tasks: synthesising information across multiple documents, generating structured summaries from unstructured input, and supporting clinical decision-making in contexts where approximate answers have value. Savana’s infrastructure accommodates these capabilities above the extraction layer – LLMs can add reasoning, synthesis, and generation without displacing what sits below them.

LLMs are probabilistic. They produce the most likely output given their training distribution; they do not guarantee that the output reflects what the clinical document actually states. For general-purpose applications, that probabilism is an acceptable property. For evidence that will appear in a regulatory submission or an HTA dossier, it is not.

EHRead provides the stable, deterministic, validated base layer. The same document, run through the same extraction process, must return the same output – every time, with a traceable audit trail. LLMs operate above that layer; they do not replace it. The extraction layer is where regulatory defensibility is won or lost, and that is precisely where probabilism carries the highest cost.

EHRead across the pharma lifecycle

EHRead does not operate as a standalone tool. It is the extraction engine inside the infrastructure that makes regulatory-grade deep real-world evidence possible across the pharma lifecycle.

Clinical development

EHRead enables patient identification and cohort characterisation from real EHR populations – extracting eligibility variables, baseline characteristics, and endpoint measures from the documentation where they actually live, not from structured fields that approximate them.

Post-market safety surveillance

EHRead monitors patient records continuously as clinical documentation is generated, enabling detection of adverse event patterns as they emerge in practice rather than in periodic data pulls.

HEOR and market access

EHRead extracts the comparative effectiveness data that HTA bodies require: treatment sequencing, real-world response rates, discontinuation reasons, and quality-of-life markers that clinicians document in notes rather than in structured outcome fields. The evidence that market access teams need sits in precisely the part of the record that conventional data extraction ignores.

All of this operates through Savana NGR – the living registry infrastructure in which EHRead functions as the extraction engine. NGR is not a database; it is the operating layer for evidence generation that inherits EHRead’s validation. Because the clinical NLP layer has been validated, the evidence it generates carries that validation forward into every downstream analysis.

The Smart Health Alliance is the network at which EHRead operates at scale: more than 280 institutions across 15 countries, connected through a federated architecture that keeps patient data at its source while making the clinical content of that data available for evidence generation. At that scale, EHRead does not process individual documents – it processes the longitudinal clinical reality of entire patient populations.

The question your evidence strategy must answer

When pharma teams evaluate their RWE capabilities, they typically assess dataset size, registry coverage, or analytical pipeline sophistication. These are legitimate questions. However, they all presuppose that the data entering those pipelines is complete, accurate, and clinically representative.

If the extraction layer cannot read clinical narrative, the data is not complete. If it cannot distinguish negation from assertion, or hypothesis from finding, it is not accurate. If it systematically misses the information documented in free text, it is not representative. No downstream analysis compensates for those gaps. No statistical method reconstructs information that was never extracted.

The choice of clinical NLP layer is not a technical procurement decision. It is a decision about the scientific validity of every piece of evidence your organisation generates from EHR data – about whether that evidence will hold up to regulatory scrutiny, reproduce across geographies, and scale as evidence objectives evolve.

“The question is not what EHRead does. The question is what your evidence is built on.”

Sources & further reading

  1. Savana. Savana cNLP (EHRead) – product overview. savanamed.com.
  2. Savana. Savana NGR – next generation registry infrastructure. savanamed.com.
  3. Savana. Smart Health Alliance – federated hospital network. savanamed.com.
  4. Savana. What is deep real-world evidence? Savana blog.
  5. Savana. Regulatory-grade RWE: what FDA and EMA actually require. Savana blog.
  6. Savana. What is a disease registry in clinical research? Savana blog.
  7. Savana. Savana Data Space – federated data governance layer. savanamed.com.

See what EHRead extracts from your clinical records

Savana cNLP (EHRead) is the only clinical NLP engine trained exclusively on more than seven billion real EHR documents across seven languages. It extracts audit-ready, regulatory-grade evidence from the clinical narrative that structured data pipelines miss – with full source traceability built in by design.

Explore Savana cNLP
ABOUT SAVANA

Savana is the only healthtech SaaS company delivering scientifically validated, multilingual Clinical NLP exclusively trained on EHRs, ensuring the deepest dRWE for hospitals, HTA bodies, pharma, and MedTech. It also powers the Smart Health Alliance, a global hospital network collaborating within Savana’s ecosystem to drive the most innovative and disruptive RWE use cases, including the market’s only AI-powered living registries.

Share it here:

Related Posts

The evidence behind the headlines.

Stay ahead of how real-world data is reshaping clinical practice, drug development, and healthcare policy.