Accurate and Robust Prediction of Amyloid-B Brain Deposition from Plasma Biomarkers and Clinical Information Using Machine Learning

Jiayuan Xu^1,2Andrew J Doig^1,3Sofia Michopoulou^4,5,6Petroula Proitsi^7,8,9Fumie Costen^1,2*

• ¹The University of Manchester, Manchester, United Kingdom
• ²Department of Electrical and Electronic Engineering, Faculty of Science and Engineering, The University of Manchester, Manchester, England, United Kingdom
• ³School of Biological Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Manchester, England, United Kingdom
• ⁴University Hospital Southampton NHS Foundation Trust, Southampton, United Kingdom
• ⁵School of Clinical and Experimental Sciences, Faculty of Medicine, University of Southampton, Southampton, Hampshire, United Kingdom
• ⁶Faculty of Medicine, Southampton General Hospital, Southampton, Southampton, United Kingdom
• ⁷Department of Basic & Clinical Neuroscience, School of Neuroscience, King's College London, London, England, United Kingdom
• ⁸Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, England, United Kingdom
• ⁹Wolfson Institute of Population Health, Queen Mary University of London, London, London, United Kingdom

Background: Alzheimer's disease (AD) greatly affects the daily functioning and life quality of patients and is prevalent in the elderly population. Amyloid-B (AB) accumulation in the brain is the main hallmark of AD pathophysiology. Positron Emission Tomography (PET) imaging is the most accurate method to identify AB deposits in the brain, but it is expensive and not widely available. This study aims to develop and validate machine learning algorithms for accurately predicting brain amyloid-B (AB) positivity using plasma biomarkers, genetic information, and clinical data as a cost-effective alternative to PET imaging. Methods: We analyzed 1043 patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and validated our models on 127 patients from the Center for Neurodegeneration and Translational Neuroscience (CNTN) dataset. Brain AB status was determined using plasma biomarkers (AB42, AB40, Phosphorylated tau (pTau) 181, Neurofilament light chain (NfL)), Apolipoprotein E (APOE) genotype, and clinical information (Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA),age, education year, and gender). Decision tree, random forest, support vector machine and multilayer perceptron (MLP) machine learning methods were used to combine all this information. We introduced a feature selection method to balance the performance and the number of features. We conducted a feature matching technique to enable our model to be tested on the external dataset without retraining.Our system achieved a value of 0.95 for the Area Under the ROC curve (AUC) using the ADNI dataset (n=340) and the full set of 11 features. Our architecture was also tested on an external dataset (CNTN, n=127) and achieved an AUC of 0.90. When using only five features (pTau 181, AB42/40, AB42, APOE E4 count, and MMSE) on 341 ADNI patients, we achieved an AUC of 0.87 with the MLP method.The random forest, support vector machine and multilayer perceptron methods can accurately predict brain AB status using plasma biomarkers, genotype, and clinical information. The method generalizes well to an independent dataset and can be reduced to using only five features without losing much accuracy, thus providing an inexpensive alternative to PET imaging.