Clinical prediction models using artificial intelligence approaches in dementia

Background

While nearly half of all dementia cases are potentially preventable, early detection and targeted interventions are critical. Artificial intelligence (AI)-enhanced clinical prediction models offer promising tools to improve diagnostic and prognostic accuracy by leveraging machine learning (ML) to integrate diverse data sources. This systematic review evaluates the development, performance, and clinical applicability of AI-based prediction models in dementia.

Methods

Searches of PubMed, Embase, and Web of Science identified peer-reviewed studies up to October 2024, focusing on AI-based models predicting dementia onset. Included studies were assessed for model accuracy, bias, and generalizability using the PROBAST tool. Data extraction adhered to the TRIPOD and CHARMS frameworks, capturing study design, participant demographics, predictor variables, and performance metrics.

Results

Among 2699 articles initially screened, 21 studies were included, encompassing over 1 million participants. AI models, extremely heterogenous for their nature, demonstrated good predictive accuracy, with a mean area under the curve of 0.845. While internal validation was conducted in all studies, external validation was limited. Models incorporating ML methods like random forests and support vector machines outperformed traditional approaches. The most used parameters were clinical and cognitive data, whilst data about biomarkers were the less used. Risk of bias was generally low, though calibration and generalizability remained challenges.

Conclusions

AI-based prediction models show strong potential for early dementia detection and personalized care. However, their integration into clinical practice requires addressing issues of external validation, data representativeness, and model interpretability. Further research should focus on robust validation and ethical implementation to optimize their utility in dementia care.