(C-159) Description of chronic graft versus host disease using an innovative Artificial Intelligence Natural language processing on electronic medical records from 8 centers in France

Background: Chronic Graft Versus Host Disease (cGVHD) is the most common and serious complication following allogenic hematopoietic stem-cell transplantation (allo-HSCT). In Europe, there is no agreed-upon standard of care or preferred therapeutic agents for patients who failed second line of systemic therapy (LOT). CGVHD treatments beyond the first LOT are not well known and are to be described. But some treatments can’t be captured in claims, and manual chart review used to generate real-world data can be long and expensive. Thus, we used another approach.

Objectives: We aimed to identify cGVHD patient, describe their characteristics and treatment patterns by LOT, using an innovative approach based on an artificial intelligence natural language processing (AI-NLP) method to structure data.

Methods: The AI-NLP scanned all electronic medical records of patients who received an allo-HSCT in 8 participating centers in France, identified patients with a cGVHD and >1 year follow-up after onset of cGVHD (eligibility criteria) and used relevant information to derive and infer the variables.
To ensure completeness and accuracy of the NLP derived data, a 3-step quality check was performed:
1) review by clinical research associate (CRA)
2) comparison with the French Branch of the European Bone Marrow Transplant (EBMT) registry
3) validation by investigator experts.

Results: Among 513 screened allo-HSCT patients’ medical files, 235 developed a cGVHD, whose data were extracted and structured by AI-NLP;
60% were men, mean age at allo-HSCT was 49.6 years, main reason for allo-HSCT was acute myeloid leukaemia (41.3%), acute lymphoblastic leukaemia (14%). Median delay from allo-HSCT to cGVHD was 174 days.
cGVHD was severe in 103 patients, moderate in 52, mild in 72, among 227 with at least one score. Organs involved were skin (68.4%), eye (34.6%), liver (33.3%), mouth (29.8%), gastrointestinal (25.9%), lung (23.7%); 84 patients received a second LOT treated with ruxolitinib 42.9%, tacrolimus 13.1%, ciclosprine 7.1%, belumosudil 8.3%; 35 received a third LOT with ruxolitinib 37.1%, Belumosudil 22.9%, ibrutinib 11.4%, tacrolimus 5.7%.
The completeness of data varied according to the nature of variables: sociodemographic and cGVHD treatments (100%), graft characteristics (>75%), mostly depending on the information present in the medical records.
The comparison with EBMT registry showed high concordance rates (>90%) on most of the variables in common in both data sources.

Conclusions: This is a first insight into the real-world management of cGVHD in France, using AI-NLP to generate a comprehensive, robust and accurate structured dataset, issued from information in electronic medical records of transplanted patients, with minimal burden on the medical teams in the participating centers.