(190) Investigation of inclusion and exclusion criteria in Alzheimer’s disease clinical trials on patient sample representation through natural language processing

Background: Significant mismatch exists between the sex distribution in Alzheimer’s disease (AD) clinical trials (CT, 58% women) and their estimated representation in the global dementia population (64%). Although clinical trials gear towards patient inclusivity and representation, the effect of common inclusion and exclusion criteria in AD trials on sample diversity and representation remains largely unknown.

Objectives: To systematically identify all interventional phase II and III AD clinical trials and to build a structured dataset containing all inclusion and exclusion criteria using natural language processing (NLP).

Methods: Clinical study registration was downloaded from clinicaltrials.gov (2023-09-08). Relevant studies were identified by querying condition field for `Alzheimer’s disease`, `mild cognitive impairment`, and intervention field for `drug`, `dietary_supplement`, or `biological`. 20% were randomly sampled for manual annotation (10% machine learning (ML) training, 10% for validation) stratified by phase II and III trials. Three annotators independently annotated using annotation tool LANN, in line with Observational Health Data Sciences and Informatics Common Data Model health standards. Finalized annotations were used as gold standard for model training, based on five-fold cross validation and conditional random fields algorithm from the CRFSuite library (CLAMP). Trained model was used to extract inclusion and exclusion criteria from remaining 80% CT. Structured eligibility criteria data were analysed in R.

Results: 1810 CTs were extracted from ClinicalTrials.gov, 1622 (69.7%) were AD, 1477 (81.6%) were interventional. 593 (32.8%) were phase II and 316 (17.5%) were phase III, total 909 (50.2%). After applying criteria the dataset consisted of 563 studies (phase II: 359, III: 204). Overall inter-annotator agreement was good (F1 score >0.7). While 44/72 (61.1%) phase II trials included MMSE test, 70.7% of phase III trials included this (29/41). As well, MMSE range was median 16 – 26 [4-30] for phase II vs. 13 – 26 [1-30] for phase III. 30.6% of phase II trials (22/72) included a female-specific inclusion criteria vs. only 17.1% for phase III, while no male-specific inclusion criteria were found.

Conclusions: While publicly accessible databases such as clinicaltrials.gov contain all CTs, aggregating eligibility criteria remains challenging and additional processing is required to ensure a high quality dataset meeting selection criteria. We have created a dataset consisting of all interventional phase II and III CTs for AD. Analysis of the annotation dataset finds differences in inclusion criteria specific to women. CLAMP model training on this dataset will provide comprehensive results of all AD CTs.