Related Experiment Video
Updated: Jun 27, 2026
12:18
A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment
Published on: January 11, 2020
Machine Learning-Based Multiclass Classification of Cognitive Stages Using Plasma Biomarkers, Clinical Assessments,
1Department of Electrical and Electronic Engineering, University of Manchester, Manchester M13 9PL, UK.
Diagnostics (Basel, Switzerland)
|June 26, 2026
Summary
Plasma biomarkers show promise for Alzheimer's disease (AD) staging, but gains from combining them with clinical data are inflated by label circularity. Independent analysis reveals realistic performance for blood-based AD screening.
Area of Science:
- Neuroscience
- Biomarker Discovery
- Clinical Diagnostics
Background:
- Plasma biomarkers are increasingly recognized for Alzheimer's disease (AD) staging.
- Head-to-head comparisons with clinical scales are limited, raising concerns about machine learning fusion gains reflecting label circularity.
- Quantifying this circularity is crucial for accurate biomarker assessment.
Purpose of the Study:
- To evaluate the independent classification power of plasma biomarkers for Alzheimer's disease (AD) staging.
- To quantify the impact of label circularity on machine learning fusion models.
- To assess the generalizability of plasma biomarker panels in an independent cohort.
Main Methods:
- Utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) including plasma biomarkers (pT217, Aβ42/40, NfL, GFAP), clinical scales (MMSE, CDR-SB, FAQ), and APOE genotype.
- Employed repeated, nested cross-validation with logistic regression, SVM, Random Forest, and XGBoost classifiers.
- Validated a reduced feature panel on an independent Center for Neurodegeneration and Translational Neuroscience (CNTN) cohort.
Main Results:
- Clinical scales alone achieved a three-class AUC-OVR of 0.9539; fusion yielded an indistinguishable gain of 0.9559, largely due to label circularity.
- An internal plasma plus demographic-genetic model, independent of circularity, achieved AUC-OVR = 0.7455, with pT217 being the primary contributor.
- A reduced feature panel (excluding pT217 and Aβ42/40) transferred to the CNTN cohort with AUC-OVR = 0.702.
Conclusions:
- Observed fusion gains in ADNI are significantly influenced by label circularity.
- Internal plasma biomarker models offer realistic performance for three-class Alzheimer's disease (AD) screening (AUC ≈ 0.74).
- A reduced panel without pT217 shows moderate transferability to independent cohorts (AUC ≈ 0.70), highlighting MCI detection as a key challenge.