10121113An NLP development platform for scalable, AI-enabled structuring of psychiatric clinical textClosedCollaborative R&DInnovate UKAkrivia curates a database of 4.6 million+ patients' anonymised psychiatric electronic health records (EHRs) on behalf of 19 NHS healthcare organisations (HCOs). Using this privileged access, Akrivia aims to transform mental health and dementia research, where current treatments are often ineffective, or simply do not exist. Akrivia faces a challenge that ~85% of our EHRs' research-relevant data is stored in 'free-text' fields (e.g., clinical notes), which are difficult to analyse at scale and inaccessible outside the NHS due to their sensitivity. 'Natural language processing' (NLP) can help solve this. Akrivia has developed an AI-based NLP system to extract structured data on medications, symptoms, diagnoses etc from clinical notes. Akrivia's NLP is now enabling numerous NHS research projects, and driving revenue via our commercial research services. Thanks to this success, Akrivia has an opportunity to apply our NLP to new datasets, including NHS GP and US psychiatric data. This could dramatically improve our research impact, providing more complete patient journeys and increasing the diversity of our population. However, to support this, Akrivia's NLP solution needs to become more scalable. Some of our NLP models are based on older technology which takes a long time to update/expand, and our deployment processes rely heavily on manual steps. In this project, Akrivia will industrialise a prototype NLP solution we have developed, leveraging recent advances in AI research to improve scalability, accuracy, and human-interpretability. We will redevelop our technical infrastructure for increased automation, and expand our inhouse NLP tools so those without technical AI expertise (e.g., clinicians) can apply their specialist domain knowledge directly in model development. These innovations will enable us port our NLP approach to new datasets, helping provide researchers in mental health and dementia with the best possible data to address these complex, costly, and historically unaddressed conditions.