It is more than 10 years since the inception of CogStack, initially deployed to identify cohorts for Genomics England’s 100,000 Genomes Project from clinical text within electronic health records. This year CogStack announced its spin-out from our founding partner institutions, opening up opportunities for further partnership and innovation. Today, applications of the CogStack platform now extend across the globe, supporting the interface between clinician and patient data well beyond research and integrated into real-time operations.
As another year draws to a close, with numerous papers published showcasing its impact, we look back at 2025 and handpick some cases where CogStack was used to improve the visibility of clinical information and pave the way for better healthcare delivery.
Cohort identification
One of the core challenges in modern clinical research remains identifying the high-value clinical information trapped in narrative: discharge summaries, progress notes, and referral letters. The difficulty in identifying this data at scale often slows cohort identification, increases data extraction costs, and can introduce bias by favouring easily coded fields.
This study of heart failure with preserved ejection fraction across multiple sites within a hospital network identified 2,223 patients meeting HFpEF diagnostic criteria, where an existing clinician recorded HFpEF diagnosis was present for just over 10%.
The use of CogStack’s natural language processing (NLP) capabilities in this study directly supports the proactive identification of patients who may be missed by standard structured coding. By extracting and analysing both structured and unstructured data from electronic health records, including narrative notes, the system was able to surface a large cohort of patients meeting diagnostic criteria but lacking a formal, coded diagnosis.
Crucially, this AI-enabled phenotyping identified an “Early-Stage” subgroup with low comorbidity and a high likelihood of transitioning phenotypes, which the authors suggest represents a critical opportunity for earlier diagnosis. This automated detection of underrecognized, high-risk patients allows clinicians to proactively initiate monitoring and targeted interventions at a stage where the disease trajectory may still be modified, moving from reactive care to preventative clinical action.
Integrated and real-time
It is also possible to use CogStack to support clinical coding as the data is entered at the point of care. This not only improves data accuracy but can also save clinicians considerable amounts of time.
A publication in 2025 detailed the deployment of the Medical information AI Data Extractor (MiADE) at University College London Hospitals. This point-of-care NLP system, integrated with the Epic EHR, is designed to process a clinician’s note as it is drafted, extract entities like diagnoses and medications using CogStack NLP, and link them to standard terminologies such as SNOMED CT. The system provides real-time suggestions to the clinician for verification before the structured data is committed to the patient record.
This approach has implications not just for retrospective research, but for improving data fidelity in real-time at the very point of creation. Moreover, in simulation, it was found that MiADE reduced the time clinicians spent completing structured problem lists by 89%.
Precision and recall
In a separate 2025 study, researchers detailed the use of NLP to analyse clinical correspondence for vestibular schwannoma surgery patients. The project successfully processed 1,148 clinical letters to extract a specialised symptom dataset for downstream analysis, dramatically accelerating a process that is typically manual and time-intensive.
The study was able to demonstrate detailed extraction of symptoms and concepts from unstructured letters, with CogStack NLP’s contextual capabilities successfully differentiating between current and historical status of a condition with full transparency to the source material. The paper also reported strong performance metrics (F1 results) for key symptom extraction, demonstrating how automated, structured extraction can support the large-scale characterisation of patient cohorts.
When critical symptoms and outcomes are locked in narrative text, service teams struggle to perform basic audits and answer questions about patient outcomes. This type of extraction pipeline provides a functional stepping-stone to faster, more consistent pathway evaluation, particularly in complex or rare conditions where structured coding is often sparse.

Dashboards to surface complex data
Mental healthcare records, which are notoriously text-heavy, often bury critical information deep within narrative notes. The Visual & Interactive Engagement With Electronic Records (VIEWER) system at South London & Maudsley (SLaM) was developed to address this by combining search, data accessibility, and interactive visual analytics. This design allows clinicians to fluidly navigate between individual patient views, caseloads, and population-level data.
The system provided an information retrieval and analytics layer, making unstructured clinical narrative accessible within operational workflows for direct care, not just specialized research datasets. The VIEWER system, which provides interactive longitudinal views across over 400,000 patients, was found to significantly improve performance and task completion speed when compared to the standard clinical information system. Moreover, by identifying target populations, care givers could be proactive in their outreach and were able to monitor disparities in service access across demographic groups for better care coordination.
Predictive modeling at a national scale
One of the most extensive efforts described in 2025 was Foresight, a generative AI model trained within the NHS England Secure Data Environment (SDE) for the purposes of COVID-19 related research. The project utilised de-identified national-scale health data from 57 million individuals (representing more than 10 billion healthcare interactions), representing what is thought to be the first national scale foundation model of EHRs. The work built upon modelling methodologies developed through earlier CogStack academic research, with technology partners, AWS and Databricks, providing computational resources and modelling advice. A crucial element was that the industry partners operated without ever accessing the NHS data, the model, or its outputs.
The effort demonstrated how industry expertise can support innovation within the Secure Data Environment, maintaining strict separation between external partners and identifiable patient data. Utilising data at this national scale allows research to better represent minority groups and rare diseases—cohorts that are often underpowered in smaller datasets—supporting both efficiency and equity in evidence generation.
Recent trends and looking to the future
The pattern across these 2025 projects suggests several maturing trends in the use of health data:
- Real-World Evidence Capability: RWE generation is increasingly moving beyond structured codes. NLP is now reliably turning the narrative clinical record into analyzable signals—including subtle phenotypes, treatment response, and safety events—at scale.
- Research Acceleration: The ability to search and extract across both text and structured fields enables feasibility studies, site selection, and recruitment support to move from being month-long processes to being completed in a matter of days.
- Templated Deployment: The standardisation of processes for de-identifying and utilising clinical text for research, service improvement and direct care which can be deployed into new settings to help healthcare providers and research sites act in confidence.
For patients and the public, the continued focus is on safe, proportionate, and governed data use. The direction of travel over the course of 2025 reinforces the move towards the use of de-identified data in controlled environments, purpose-built tools that minimize data exposure by bringing the analysis to the data, and the adoption of tooling that measures and addresses equity as a concrete outcome.
The new year will inevitably present many challenges to the global healthcare industry, not least a continued focus on the use of AI to achieve real-world impact and drive productivity. In the UK, the NHS’s ten year Fit for the Future programme will approach its early delivery window and new standards will come into play with the new financial year. All of this will place a growing reliance on data, with the right AI tools needed to:
- Bring together more data than ever before
- Improve the depth, reliability and security of that data
- Use that data to make predictions about the future.
With new healthcare delivery partners coming on board, continued development of our platform and scaling across the NHS CogStack will continue to tackle these challenges, turning complex, real-world data into insights that have a real world impact. CogStack is dedicated to making sure our partners have the advanced AI tools they need to thrive in an increasingly complex data-driven healthcare environment.



