Artificial Intelligence based diagnosis support tool improves diagnostic confidence in neurologist assessments

04/07/2024

On 4 July, a team of USA and China based scientists published a paper about an Artificial Intelligence (AI) based diagnosis support framework for dementia in the journal Nature Medicine. The article is focused on the development and validation of the AI-based tool for diagnosing different causes of dementia using multimodal data. The researchers used nine datasets, comprising 51,269 participants altogether. The development of the multimodal Machine Learning (ML) framework was based on data modalities from these data sets including demographics, individual and family medical history, medication use, neuropsychological assessments, functional evaluations, and multimodal neuroimaging. 

The tool classified individuals with normal cognition, mild cognitive impairment, and dementia. Additionally, it could differentiate between Alzheimer's disease (AD), Lewy body dementia (LBD), vascular dementia (VD), prion disease (PRD), frontotemporal lobar degeneration (FTD), normal pressure hydrocephalus (NPH), systemic and environmental factors (SEF), psychiatric conditions (PSY), traumatic brain injury (TBI), and other dementia conditions (ODE). Reporting on the model performance, the team reported that it achieved a microaveraged area under the receiver operating characteristic curve (AUROC) of 0.94, which indicates a high accuracy in classifying individuals with normal cognition, mild cognitive impairment, and dementia. Furthermore, they indicate that it was able to achieve a microaveraged AUROC of 0.96 in differentiating dementia etiologies. With regards to mixed dementia cases, it only mean AUROC of 0.78 for two co-occurring pathologies. 

The researchers were also interested in how well the model could improve neurologist assessments. They therefore investigated its predictions in a subset of 100 cases, and found that it exceeded neurologist-only evaluations by 26.25%. These predictions were aligned with biomarker evidence as well as post mortem findings. The full open access article can be read here:

https://www.nature.com/articles/s41591-024-03118-z