Autocorrection if→of function words in reading aloud: A novel marker of Alzheimer's risk

Abstract

Objective: The present study investigated cognitive mechanisms underlying the ability to stop "autocorrect" errors elicited by unexpected words in a read-aloud task, and the utility of autocorrection for predicting Alzheimer's disease (AD) biomarkers.

Method: Cognitively normal participants (total n = 85; n = 64 with cerebrospinal fluid [CSF] biomarkers) read aloud six short paragraphs in which 10 critical target words were replaced with autocorrect targets, for example, The player who scored that final [paint] for the local team reported [him] experience. Autocorrect targets either replaced the most expected/dominant completion (i.e., point) or a less expected/nondominant completion (i.e., basket), and within each paragraph half of the autocorrect targets were content words (e.g., point/paint) and half were function words (e.g., his/him). Participants were instructed to avoid autocorrecting.

Results: Participants produced more autocorrect errors in paragraphs with dominant than with nondominant targets, and with function than with content targets. Cognitively normal participants with high CSF Tau/Aβ42 (i.e., an AD-like biomarker profile) produced more autocorrect total errors than those below the Tau/Aβ42 threshold, an effect also significant with dominant-function targets alone (e.g., saying his instead of him). A logistic regression model with dominant-function errors and age showed errors as the stronger predictor of biomarker status (sensitivity 83%; specificity 85%).

Conclusions: Difficulty stopping autocorrect errors is associated with biomarkers indicating preclinical AD, and reveals promise as a diagnostic tool. Greater vulnerability of function over content words to autocorrection in individuals with AD-like biomarkers implicates monitoring and attention (rather than semantic processing) in the earliest of cognitive changes associated with AD risk. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

Figure 1a.

Percent of autocorrect targets that elicited autocorrect errors by part of speech, paragraph type, and Tau/Aβ42 ratio for cognitively normal participants who were biomarker negative (n=50), cognitively normal participants who were biomarker positive (n=14), and for illustrative purposes in a small group of participants with probable AD (n=8) who were not included in the main analysis (note: all but one participant with Probable AD was biomarker positive). Error bars refer to 95% confidence intervals.

Figure 1b.

Percent of autocorrect errors for target words that differed by 1 to 5 letters from their expected completions for cognitively normal participants who were biomarker negative (n=50), cognitively normal participants who were biomarker positive (n=14), and for illustrative purposes in a small group of participants with probable AD (n=8) who were not included in the main analysis. Error bars refer to 95% confidence intervals around the best fitting regression line.

Figure 1c.

Percent of autocorrect errors for function word targets of different lengths in dominant paragraphs for cognitively normal participants who were biomarker negative (n=50), cognitively normal participants who were biomarker positive (n=14), and for illustrative purposes in a small group of participants with probable AD (n=8) who were not included in the main analysis. Error bars refer to 95% confidence intervals around the best fitting regression line.

Figure 2.

Receiver operating characteristic (ROC) curves classifying biomarker status (Tau/AB42 threshold) for 64 cognitively normal participants based on A) function errors, B) content errors, and C) selected neuropsychological tests Note. ** p < .01; significant effects are bolded. AUC = Area Under the Curve; CI = confidence interval; CWIT = Color Word Interference Test; CVLT 1-5 = California Verbal Learning Test - Trials 1-5; NP = neuropsychological