Identifying prognostic factors of dementia in individuals with mild cognitive impairment (MCI): Are statistical models adequate?

Abstract

AbstractBackgroundIdentifying prognostic factors of MCI to dementia conversion is crucial. Most risk scores are developed based on selected variables using data‐driven methods from available variables in patient data. However, there is mixed evidence about the reliability of predictors selected based on data alone. Structured expert elicitation (SEE) is one approach to elicit clinician expert knowledge about relevant predictors for dementia risk. This study aims to identify critical predictors of dementia risk in persons with MCI from both SEE and random survival forests (RSF) and compare the differences between the two approaches.MethodClinical expert knowledge was quantified using SEE methodology: 11 experts (6 neurologists, 3 geriatricians, and 2 geriatric psychiatrists) were recruited to complete two rounds of surveys to rank predictors of dementia risk. RSF variable importance was used to assess the rank order of the potential predictors, according to the corresponding reduction of predictive accuracy when the predictor is replaced with its random permutation value risk using data from the Prospective Registry for Persons with Memory Symptoms of the Cognitive Neurosciences Clinic at the University of Calgary (N = 273; 26 potential predictors).ResultAccording to experts, the SEE revealed age and cerebrospinal fluid profile patterns consistent with Alzheimer’s disease (AD) as the two most important predictors of dementia risk in persons with MCI. In contrast, data‐driven RSF approach, using information from routine clinical practice without AD biomarkers, identified the Consortium to Establish a Registry for AD total score and the Montreal cognitive assessment (MoCA) total score as the most important predictors. Both approaches identified four common variables (among the top 10 predictors) included age, signs of parkinsonism, MoCA total score, and behaviour impairment.ConclusionOur findings revealed discrepancies between a data‐driven approach and SEE in identifying significant predictors of dementia in people with MCI. The SEE identified some crucial predictors, such as AD biomarker status, that are not routinely collected in registries and observational studies or available in widespread practice. Most published risk scores are not used in practice. This study offers a possible explanation: what is important to clinicians may be not included in the risk scores.