Machine learning algorithms for plasma AD biomarkers implementation in two real-world cohorts

Background. Blood-based biomarkers (BBMs) have shown high accuracy for Alzheimer’s Disease (AD) diagnosis in highly selected cohorts. In view of the introduction of plasma biomarkers in larger cohorts in real-world clinical settings, we wanted to assess whether machine learning (ML) approaches could generate reliable cutoffs for BBMs, independently from the availability of CSF or Aβ-PET gold standard reference. Methods. Supervised and unsupervised ML models were applied to combined plasma pTau181 and Aβ42/40 and to plasma pTau217 alone for AD classification in a real-world cohort from Perugia University Hospital (UniPg cohort), including patients with AD (n = 285), patients with other neurodegenerative disorders (n = 144) and neurological controls (n = 172) for which paired CSF AD-biomarkers were available. Cutoffs obtained for plasma pTau217 were also tested in an independent cohort from Palermo University Hospital (UniPa cohort), which included CSF confirmed AD (n = 89) and non-AD patients (n = 29). Results. Cutoffs obtained from unsupervised GMM clustering achieved a comparable diagnostic accuracy with respect to traditional CSF-based ROC analysis both for plasma pTau181 (~ 78%) and Aβ42/40 (~ 76%) combined and for plasma pTau217 alone (~ 86%) in UniPg cohort. The combined use of clustering-obtained cutoffs for plasma pTau181 and Aβ42/40 reached an accuracy comparable to the one of pTau217 alone for AD diagnosis in UniPg cohort. Supervised ML models for AD classification (Support Vector Machines and XGBoost) did not outperform unsupervised GMM clustering in UniPg cohort. CSF and clustering-based cutoffs for plasma pTau217 in UniPg cohort reached a comparable accuracy also when applied to UniPa independent cohort (~ 85%). Conclusions. Our unsupervised clustering algorithm enables the definition of reliable cutoffs for blood-based AD biomarkers independently from CSF data in a large, heterogeneous real-world cohort. Cutoffs obtained from unsupervised clustering in UniPg cohort showed a comparable accuracy when applied to an independent validation cohort (~ 85% in UniPa cohort). In UniPg cohort, the combined use of plasma pTau181 and Aβ42/40 increased their diagnostic accuracy, supporting the utility of multiple BBMs testing.