Development and assessment of algorithms for predicting brain amyloid positivity in a population without dementia

Abstract

Background: The accumulation of amyloid-β (Aβ) peptide in the brain is a hallmark of Alzheimer's disease (AD), occurring years before symptom onset. Current methods for quantifying in vivo amyloid load involve invasive or costly procedures, limiting accessibility. Early detection of amyloid positivity in non-demented individuals is crucial for aiding early AD diagnosis and for initiating anti-amyloid immunotherapies at early stages. This study aimed to develop and validate predictive models to identify brain amyloid positivity in non-demented patients, using routinely collected clinical data.

Methods: Predictive models for amyloid positivity were developed using data from 853 non-demented participants in the MEMENTO cohort. Amyloid levels were measured potentially repeatedly during study course through Positron Emision Tomography or CerebroSpinal Fluid analysis. The probability of amyloid positivity was modelled using mixed-effects logistic regression. Predictors included demographic information, cognitive assessments, visual brain MRI evaluations of hippocampal atrophy and lobar microbleeds, AD-related blood biomarkers (Aβ42/40 and P-tau181), and ApoE4 status. Models were subjected to internal cross-validation and external validation using data from the Amsterdam Dementia Cohort. Performance also was evaluated in a subsample that met the main criteria of the Appropriate Use Recommendations (AUR) for lecanemab.

Results: The most effective model incorporated demographic data, cognitive assessments, ApoE status, and AD-related blood biomarkers, achieving AUCs of 0.82 [95%CI 0.81-0.82] in MEMENTO sample and 0.90 [95%CI 0.86-0.94] in the external validation sample. This model significantly outperformed a reference model based solely on demographic and cognitive data, with an AUC difference in MEMENTO of 0.10 [95%CI 0.10-0.11]. A similar model without ApoE genotype achieved comparable discriminatory performance. MRI markers did not improve model performance. Performances in AUR of lecanemab subsample were comparable.

Conclusion: A predictive model integrating demographic, cognitive, and blood biomarker data offers a promising method to help identify amyloid status in non-demented patients. ApoE genotype and brain MRI data were not necessary for strong discriminatory ability, suggesting that ApoE genotyping may be deferred when assessing the risk-benefit ratio of immunotherapies in amyloid-positive patients who desire treatment. The integration of this model into clinical practice could reduce the need for lumbar puncture or PET examinations to confirm amyloid status.

Keywords: Alzheimer's disease; Amyloid; Biomarker; Immunotherapy; Prediction.

Fig. 1

Selection and distribution of amyloid assessments. The MEMENTO cohort, France, 2011–2014. A Flowchart of the study. B Number of amyloid assessments available per participant. C Amyloid assessments at each annual follow-up visit. CDR: Clinical Dementia Rating scale, ApoE: apolipoprotein E, Aβ: amyloid beta, amyloid PET: amyloid positron emission tomography, LP: lumbar puncture, SCD: subjective cognitive decline. An individual could meet more than one exclusion criterion

Fig. 2

Distribution of predictors according to the six amyloid-status prediction models. The MEMENTO cohort, France, 2011–2014. Aβ: amyloid beta, BMI: body mass index, memory test: Free and Cued Selective Reminding Test (FCSRT) used in the MEMENTO cohort and the Rey Auditory Verbal Learning Test (RAVLT) in the external validation cohort ADC, TMT B: Trail-Making Test B, MMSE: Mini-Mental State Examination, CDR: Clinical Dementia Rating scale, ApoE: apolipoprotein E, MRI: magnetic resonance imaging

Fig. 3

Development and validation of six Aβ-positivity prediction models. The MEMENTO and ADC studies. A Discrimination, error predictions, and performance. B ROC curves in a development (MEMENTO) and training set (ADC). C Calibration plots in the training set. BS: Brier score, AUC: area under the curve, Se/Sp: sensitivity/specificity, ApoE: apolipoprotein E, YI: Youden Index

Fig. 4

Discrimination performance of six Aβ prediction models overall and in subsamples following the AUR criteria. The MEMENTO Cohort, France, 2011–2014. Model 1: demographic and cognitive predictors, Model 2: Model 1 and ApoE4 status, Model 3: Model 1 and blood biomarkers, Model 4: Model 1 and MRI, Model 5: Model 1, ApoE4 status, and blood biomarkers, Model 6: Model 1, ApoE4 status, blood biomarkers, and MRI. AUC: area under the curve, AUR: Appropriate Use Recommendations, SCD: subjective cognitive decline, ε4 homozygous: homozygous for ApoE ε4. AUR population for lecanemab: population without SCD participants, without anticoagulant treatment, without siderosis, without severe subcortical hyperintensities consistent with a Fazekas score of 3, and with fewer than five cerebral microbleeds. Total population: population used to develop the predictive algorithm