

Machine Learning for Dementia: Progress and Pitfalls
The review identified 75 studies, with almost 60% of them published in the last two years, reflecting a rapid growth of interest in this field. Results from the meta-analysis show that machine learning models using these brain imaging markers perform strongly, achieving an area under the curve (AUC) of 0.88 when distinguishing Alzheimer’s dementia from healthy controls, and 0.84 for identifying cognitive impairment. These figures highlight the potential of AI to improve early detection and prognosis.
However, the study also points out significant limitations in current research. Of the 75 studies, only 16 met the criteria for inclusion in the meta-analysis due to inconsistent reporting, while just five evaluated generalisability on external datasets and six lacked clear diagnostic criteria. These shortcomings emphasise the need for more rigorous methods, better reporting standards, and stronger validation before such models can be reliably used in clinical settings.
Dr Mazo, Assistant Professor in the School of Computing, contributed her expertise in artificial intelligence and computational methods to this interdisciplinary project, which bridges neuroimaging, AI, and dementia research.
Read more: Machine learning applications in vascular neuroimaging for the diagnosis and prognosis of cognitive impairment and dementia, Alzheimer’s Research & Therapy here.