Limited generalizability and high risk of bias in multivariable models predicting conversion risk from mild cognitive impairment to dementia: A systematic review

Abstract

Prediction models have been developed to identify mild cognitive impairment (MCI) cases likely to convert to dementia. This systematic review summarizes multi-source prediction models for MCI to dementia conversion. PubMed and Embase were searched for model development and validation studies from inception up to January 18 2024. Models were assessed for included predictors, predictive performance, risk of bias, and generalizability. 62 studies were included: 41 machine learning models, 11 regression models, and 5 disease state indexes. The number of predictors in the models ranged from 2 to 60; magnetic resonance imaging (MRI) and cognitive scores were the most common sources. Performance measures indicate reasonable predictive capabilities (area under the curve [AUC] range: 0.58-0.98, accuracy range: 66.1-96.3%); however, most studies are at high risk of bias and 47 studies lack external validation. Currently, no highly valid prediction model is available for MCI to dementia conversion risk due to limited generalizability and high risk of bias in most studies. HIGHLIGHTS: Numerous models have been developed to predict the likelihood of conversion to dementia in individuals with MCI. Prediction models seem to have a reasonably good performance in predicting conversion to dementia, however, external validation and generalizability is often lacking. There is no prediction model available with a low risk for bias and that has been externally validated to accurately predict the risk of MCI to dementia conversion. For MCI to dementia conversion prediction models, more emphasis should be directed towards external validation, generalizability, and clinical applicability.

Keywords: Alzheimer's disease; dementia; mild cognitive impairment; prediction model; risk prediction; systematic review.