Predicting Alzheimer's Disease Via Machine Learning Through ADNI Data

Predicting Alzheimer's Disease Via Machine Learning Through ADNI Data

Alvin Sherman Library

Alzheimer's disease (AD) is a progressive neurodegenerative disease. Since the neurological pathology, including protein deposits and neurodegeneration, starts a decade before cognitive, memory, and behavioral symptoms, AD development should be predictable from physical biomarkers. This study investigates how patients' memory and cortical brain thickness at baseline can predict their likelihood of developing AD through supervised machine learning. From baseline visits, 205 participants' data were extracted from the ADNI database, all designated as healthy control in the first visit, but some converted to AD or Mild cognitive impairment (MCI) in later visits (Group 2), and some remained healthy (Group 1). The dataset was analyzed using machine learning in WEKA. Preprocessing attribute selection with GreedyStepwise coupled with CfsSubsetEval, followed by classification with J48 resulted in a decision tree that predicts AD/MCI conversion based on ADNI's Harmonized memory and executive-function scores (PHC_MEM, PHC_EXF) and MEMSLOPES. Unfortunately, despite the good accuracy and precision (0.790), the model relies on MEMSLOPES (a measure of memory decline from several memory tests). Classification without MEMSLOPS uses PHC MEME, β-amyloid presence, PHC EXF, and the left lateral ventricle volume, with accuracy reducing to 0.605. Random forest analysis increased the accuracy to 0.698. Additionally, MRI-derived cortical brain thickness and brain volumes showed poor accuracy. This is due to Group 2 individuals being mostly MCI-coverted and only a fewbeing AD-converted with likely detectable neurodegeneration. In future studies, we will exclude MCI-converted and only a few being AD-converted with likely detectable neurodegeneration. In future studies, we will exclude MCI-converted patients and investigate the prediction of AD converison from baseline cortical thickness, using additional data from other databases.