USING TRANSFER LEARNING TO IMPROVE ALGORITHMIC DEMENTIA CLASSIFICATION ACROSS DIFFERENT RACIAL AND ETHNIC GROUPS

Abstract Background In the absence of clinical assessments, algorithmic ‘probable dementia’ estimation aids public health and epidemiological research on determinants and consequences of dementia. Algorithmic performances have differed by social and racial/ethnic groups due to heterogeneous sample sizes. In this study, we present improvements in the accuracy of group-specific ‘probable dementia’ estimation using a transfer learning approach. Transfer learning – a novel approach not previously applied in algorithmic dementia classifications – uses a large “source” dataset with imprecise (non-clinical) outcome assessments to develop an initial algorithm with a large and diverse sample, complemented by a smaller “target” dataset with high-quality (clinical) outcome assessments to develop models with refined estimation of model coefficients. Methods Data came from the Health and Retirement Study (source data: N=6,630) and the Harmonized Cognitive Assessment Protocol (target data: N=2,388). Different algorithms to estimate probable dementia were evaluated through overall accuracy (Brier score), calibration (intercept, slope), and discriminative ability (area under the receiver operating characteristic curve, AUC; area under the precision-recall curve, AUPRC). Results Among non-Hispanic Black participants, the transfer-learned algorithm had higher accuracy than the best previously reported algorithm (Brier 0.049 vs. 0.061; AUC 0.84 vs. 0.81; AUPRC 0.52 vs. 0.39). Among Hispanic participants, the transfer-learned algorithm improved model calibration (Intercept −0.07 vs. −1.19; Slope 0.88 vs. 0.08). Discussion The use of transfer learning can improve algorithms to estimate probable dementia from non-clinical assessments for groups historically underrepresented in research. More accurate classification will lead to more equitable epidemiological and public health research on dementia.