A preliminary feasibility study on screening cognitive impairment based on multi-modal biomarkers and stacking ensemble AI approach

Abstract

Background: Early screening of cognitive impairment is crucial for patients demanding timely treatment. The need for cost‐effective, easily accessible, and accurate tools to detect cognitive decline rapidly progressed with the breakout of the COVID‐19 pandemic. In this work, we proposed a mobile‐app based cognitive assessment tool, “Alzguard‐D”, equipped with two digital biomarkers along with a cognitive task. Next, we investigated the efficacy of “Alzguard‐D” in early screening of cognitive impairment using a machine learning approach. Method: Nine “Alzguard‐D” tasks were designed based on three biomarkers: Keystroke, speech and eye movement (Table 1; Figure 1). Total of 289 participants were recruited from the national institute of dementia, welfare center, and nursing homes. The participants first completed the Korean Mini‐Mental State Examination (K‐MMSE; 2nd Edition), then performed all tasks in “Alzguard‐D”. “Alzguard‐D” took approximately 20 to 30 minutes. The participants were then categorized to healthy control (HC; n = 241) and cognitively impaired (CI; n = 48), defined as when the scores of K‐MMSE were one standard deviation below the age, education and gender‐matched norm. Given the data, we investigated the extent in which the combination of digital biomarkers and cognitive tasks improves the screening accuracy. We used a stacking ensemble model where the first stack analyzed speech and eye movement, and the second performed the classification. Result: The screening result confirmed that the addition of digital biomarkers collected by “Alzguard‐D” significantly improved the classification performance (p < 0.0001; sample paired t‐test). CatBoost algorithm statistically outperformed all other candidate algorithms (Bagging, Logistic Regression, LGBM, Naïve Bayes, XGBoost, Random Forest, SVM, Gradient Boosting). We then conducted an ablation study with three classifiers trained with (i) Cog+Bio, (ii) Bio and (iii) Cog. Cog+Bio achieved the highest AUC score, 0.876 (max: 0.942), followed by Bio with 0.783 (max: 0.845), then Cog with 0.677 (max: 0.726) (Figure 2). Conclusion: A combination of cognitive and digital biomarkers significantly increased the screening performance of cognitive impairment, and employing an intermediate feature‐level ensemble approach to effectively analyze the collected multi‐modal data increased the AUC levels.