Application of unsupervised and supervised machine learning approaches to blood-based biomarkers of Alzheimer's disease in a memory clinic cohort.

INTRODUCTION

Blood-based biomarkers (BBMs) are promising tools to facilitate biological diagnosis of Alzheimer's Disease (AD). Establishing reliable cutoffs for BBMs represents a major issue for real-world implementation and the investigation of combined BBM approaches is limited. Our aim is to test different machine learning (ML) models for BBMs interpretation with respect to established gold-standard in vivo biomarkers, i.e., cerebrospinal fluid (CSF) biomarkers, to enhance their standalone and combined diagnostic value for AD.

METHODS

We selected 434 patients who referred to Perugia University Hospital for perceived cognitive decline with available CSF A β 42/40, pTau181 and tTau and plasma Aβ42, Aβ40, pTau217, pTau181 and NfL. Multiple unsupervised clustering algorithms for cutoff determination and supervised classification models were compared for the prediction of AD defined by a CSF A+/T + profile.

RESULTS

Different unsupervised clustering models obtained cutoffs with high performances in predicting an AD CSF profile (accuracy > 0.85). Each combination of cutoffs was tested, the best ones achieving A+/T+ CSF profile prevalences > 0.9 in BBM-positive subgroups. The best supervised ML models achieved accuracies up to 0.88 in 10-fold cross-validation.

DISCUSSION: Our results prove that clustering approaches can help in identifying reliable BBMs cutoffs for AD diagnosis even in the absence of CSF gold standard reference and that ML approaches combining multiple BBMs can strongly enhance their diagnostic accuracy.