COMPUTERIZING THE CONVERSATION: A CODIFIED APPROACH TO CONSENSUS DIAGNOSIS

摘要

Determining an individual’s level of cognitive impairment is a consensus process requiring two or more clinicians integrating their individual perspectives of each case. The process is subject to influences of individual personalities and biases, time and labor intensive, and varies by site. To ensure high inter-rater reliability and validity, we codified an algorithm integrating the Clinical Dementia Rating Scale Sum of Boxes (CDRsb) and neuropsychological diagnoses. We determined the optimal cut-offs of CDRsb scores and neuropsycholgical diagnoses to distinguish between cognitive categories: normal, impaired not MCI (I-nMCI), MCI, and dementia, while maximizing the sensitivity and specificity of the receiver operating characteristic (ROC) curve. Using Python and decision-making based on Duara (2010), we codified this enhanced algorithm (eAlgDx) using data from 35,183 participants and 118,341 observations in the National Alzheimer’s Coordinating Center Uniform Data Set (2005–2017), resulting in 294 permutations. The optimal CDRsb cut-offs were: normal vs. I-nMCI = 0.25, I-nMCI vs. MCI = 0.75, and MCI vs. dementia = 3.25. Agreement between original diagnoses and the eAlgDx was 81.62% (K=0.72; SE=0.0019). The eAlgDx AUC was 0.88 (normal v. I-nMCI), 0.75 (I-nMCI v. MCI) and 0.76 (MCI vs. Dementia). The eAlgDx classified 72,765 of 78,401 previously unclassified diagnoses, and more precisely diagnosed 7,228 visits initially classified as MCI. By integrating functional (CDRsb) and neuropsychological diagnoses, we used eAlgDx to provide an expedient, reliable, and valid alternative to the classical consensus diagnosis method, which is time-consuming, repetitive, and subject to external influences. Future refinements of eAlgDx should increase its sensitivity and specificity.