Aims
Despite growing interest in leveraging Electronic Health Record (EHR) data for dementia detection, the clinical relevance of existing prediction models remains uncertain. This systematic review evaluates the performance and quality of EHR-based dementia prediction models to identify key gaps and guide future development and clinical translation.
Methods
We searched Medline, EMBASE, Scopus, IEEE Xplore, and ACM from inception until July 2024 for studies and grey literature describing the development or validation of probabilistic EHR-based dementia detection models. Risk of bias was assessed using PROBAST.
Results
We screened 8312 studies and included 56 studies describing 461 prediction models or validations. Of these, 326 were prognostic, predicting dementia risk up to 20 years in the future, although majority were <5 years. The remaining were either diagnostic or intended to identify people with known dementia, though reporting was often unclear. Only 7 models (1.5%) were externally validated. Unstructured data was used in 20 studies. Just 5/54 development studies used gold-standard (clinical diagnostic) criteria for case ascertainment, with most others (49/54) reliant on diagnostic codes. Discriminative metrics like AUROC were frequently reported, with calibration measures almost never reported. Modelling techniques ranged from regression-based methods to machine learning/artificial intelligence techniques such as transformers. Most models had ‘high’/‘unclear’ risk of bias due to inadequate reporting.
Conclusions
Current EHR-based dementia prediction models face major limitations, including flawed case ascertainment, inadequate reporting, and limited external validation. Future efforts could prioritise adherence to reporting standards (eg, TRIPOD) and integrate clinical expertise during model development to enhance clinical relevance.