Serum elemental profile-based machine learning models for Alzheimer's disease identification and cognitive score
Haotian Liu1,2,3, Xinnan Liu2,3, Yashuang Chen1
1Department of Immunology, State Key Laboratory of Complex, Severe, and Rare Diseases, Institute of Basic Medical Sciences & School of Basic Medicine Chinese Academy of Medical Sciences & Peking Union Medical College Beijing China.
Background:
Current diagnostic approaches for Alzheimer's disease (AD) largely rely on cerebrospinal fluid biomarkers and neuroimaging, which may be invasive, costly, and not readily accessible in routine clinical settings. We investigated whether serum elemental profiling combined with machine learning could provide complementary information for AD identification and exploratory cognitive score prediction.
Methods:
This retrospective cross-sectional study included 874 participants enrolled between 2017 and 2023 from the Brain Aging National Cohort-Peking Union Medical College cohort, comprising 427 cognitively normal controls (NCs) and 447 patients with clinically defined AD. Serum concentrations of 20 elements were quantified by inductively coupled plasma mass spectrometry. Associations between serum element concentrations and AD status were evaluated using age- and sex-adjusted logistic regression models with false discovery rate (FDR) correction. Associations with Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) scores were assessed using linear regression models with FDR correction. Machine learning classification models were developed for AD identification, whereas regression models were developed for exploratory MMSE and MoCA score prediction. Model performance was evaluated in an internal hold-out test set.
Results:
Age did not differ significantly between NC and AD participants (66.4 ± 9.9 vs. 67.1 ± 9.0 years; t(872) = -1.13, p = 0.260), whereas the proportion of women was higher in the AD group than in the NC group (255/447 [57.0%] vs. 211/427 [49.4%]; χ 2(1) = 5.11, p = 0.024). MMSE scores were significantly lower in patients with AD than in NC participants (AD: n = 277, median [interquartile range (IQR)] = 24 [19-27]; NC: n = 427, median [IQR] = 30 [30-30]; Mann-Whitney U = 3043.5, p < 0.001), as were MoCA scores (AD: n = 240, median [IQR] = 18.00 [14.00-20.25]; NC: n = 427, median [IQR] = 30 [30.00-30.00]; Mann-Whitney U = 13.5, p < 0.001). After adjustment for age and sex, higher serum lead (Pb) and tin (Sn) levels were associated with increased odds of AD (Pb: odds ratio [OR] = 4.95, 95% confidence interval [CI]: 3.42-7.37; Sn: OR = 1.45, 95% CI: 1.20-1.78), whereas higher serum antimony, nickel, manganese, cobalt, selenium, and calcium levels were associated with lower odds of AD (OR range: 0.50-0.81; all FDR-adjusted p-values < 0.05). Among the eight classification models evaluated in the internal hold-out test set, the random forest model showed the highest apparent performance for distinguishing AD from NC, achieving an accuracy of 88% and an area under the receiver operating (AUC) (acharacteristic curv of 0.94. Among the three regression models evaluated, the random forest regression showed the highest apparent performance for cognitive score prediction, yielding the strongest correlations between predicted and observed scores for both MMSE (r = 0.48, p < 0.001) and MoCA (r = 0.62, p < 0.001).
Conclusion:
Serum elemental profiles combined with machine learning may provide a minimally invasive and accessible complementary approach for AD identification and cognitive assessment.
