What are the most important prognosticators of Alzheimer's disease progression? – Answers from machine learning

Abstract

AbstractBackgroundNumerous variables that influence the risk of decline in Alzheimer’s disease (AD) have been identified. Previous studies have only examined a relatively small number of these variables. To our knowledge, no study has compared all variables in a large dataset to determine which are the most important.We hypothesised that the most important variables are neuropsychological assessment (NPA) scores and neuroimaging findings; moreover, the prognostic value of all variables will vary with disease stage: normal control (NC), mild cognitive impairment (MCI) or dementia. Secondly, we hypothesised that non‐NPA variables, such as imaging, cerebrospinal fluid (CSF) and genetics, will add prognostic value to that provided by NPA variables alone.MethodWe analysed all relevant variables in the Alzheimer’s Disease Neuroimaging Initiative database – the largest open‐access dataset of subjects with or at risk of AD. With machine learning, we identified those variables which are the strongest predictors of the annual rate of change in up to the first 10 years of Alzheimer’s Disease Assessment Scale – Cognitive Subscale, 13‐item version (ADAS‐Cog13) and Clinical Dementia Rating, Sum of Boxes (CDR‐SB) scores. We repeated this analysis for each disease stage subgroup. Secondly, we compared models that only used NPA‐related variables (NPA‐only) with those that used up to 2 non‐NPA modalities (NPA‐plus) and contrasted the findings between the same subgroups.ResultAcross all subjects, the strongest predictors of the rate of change of both ADAS‐Cog13 and CDR‐SB scores are CSF total tau and amyloid levels, fluorodeoxyglucose (FDG) positron emission tomography (PET) findings, and certain NPA measures. At the NC and MCI stages, CSF and imaging variables are strong predictors, but only NPA variables remained useful at the dementia stage. Secondly, NPA‐plus models consistently outperformed NPA‐only models. The non‐NPA modalities which added the most predictive power varied based on a subject’s disease stage in a manner similar to the previous findings.ConclusionWe identified a multi‐modal set of variables that have the strongest impact on predicting decline during different stages of AD. Our findings are clinically relevant as they recommend the investigations that are of most prognostic yield for clinicians to undertake at each stage of AD progression.