There are existing tools that can calculate a personalized estimate for the probability of developing dementia in the future. These are known as dementia prediction models. Such models could be used to aid doctors when they are making medical decisions. However, it is important to ensure a prediction model is accurate before it is used clinically.
Research thoroughly investigating the accuracy of existing dementia prediction models is limited and this likely hinders the ability to use such models. In this study we take the first step towards thoroughly investigating the dementia models.
The first step to investigate the accuracy of published prediction models is applying the model to new sets of patients. However, this requires being able to replicate the model based on the authors documentation. In this study we investigate how feasible is it to replicate published dementia models. For models with adequate documentation to be replicated, we then apply them on many new sets of patients to determine any additional limitations.
Dementia is an umbrella term to describe various illnesses that affect cognition and may lead to mental degradation. Early diagnosis of individuals at high risk of dementia, e.g. through the use of a clinical prediction model, allows for improved care and risk-factor targeted intervention. This study investigates how well existing dementia prediction model are reported in literature, and to which extend reporting allows for external validation of these models. External validation assesses a model’s reliability for clinical use in external data sources that have not been used for model development. To that end we externally validate three existing prediction models that estimate the risk of dementia or Alzheimer's diseases and related dementias (ADRD) in patients aged 60 and above with type 2 diabetes mellitus and hypertension, patients between 60 and 95 years, and patients aged 45 and above, as defined in their respective research studies.
External validation is performed in a network of observational databases including Clinical Practice Research Datalink (CPRD), IBM MarketScan Medicare Supplemental Databases (MDCR), Iqvia Disease Analyzer (DA) Germany, Optum De-Identified Clinformatics Data Mart Socio-Economic Status, Optum De-Identified Electronic Health Record, and Integrated Primary Care Information (IPCI). These databases have been converted to the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM), which allows for use of standardized cohort and predictor definitions as well as standardized analytics tools developed by the Observational Health Data Science and Informatics (OHDSI) initiative. Given definitions of the study population, index date, time-at-risk/follow-up time, outcome, prediction method, predictors, predictor time window, and other full model specifications e.g. baseline risk, we are replicating two Cox proportional hazard and one logistic regression model from three published research studies. We evaluate discrimination and calibration performance of these models to assess the validity of our model replication process.
Reporting of existing dementia prediction models to allow for replication and external validation; externally validate three existing dementia prediction models, which were found to report sufficient detail for model replication, in a network of observational databases (including CPRD)
Jenna Reps - Chief Investigator - Johnson & Johnson ( JnJ - USA )
Henrik John - Corresponding Applicant - Erasmus University Medical Center ( EMC )
Egill Fridgeirsson - Collaborator - Erasmus University Medical Center ( EMC )
Jan Kors - Collaborator - Erasmus University Medical Center ( EMC )
Peter Rijnbeek - Collaborator - Erasmus University Medical Center ( EMC )
Jenna Reps - Collaborator - Johnson & Johnson ( JnJ - USA )