Classification of AD/MCI/HC based on amyloid‐PET using Random Forest Ensemble

Background Positron emission topography (PET) and magnetic resonance imaging (MRI) are two common in‐vivo techniques for supporting clinical diagnosis of dementia. The correlated neuropathological changes including amyloid‐beta (Aβ) deposition, and cortical atrophy may have the potential to predict conversion from cognitively normal elderly adults (HC), to mild cognitive impairment (MCI) and eventually Alzheimer’s disease (AD). In order to overcome the complex interactions among these biomarker datasets, random forest (RF) as one of the machine learning algorithms is usually used in data classification. In current study, we aim at evaluating the accuracy of RF model in distinguishing HC, MCI from AD and the importance of various neuropathological features in selection. Method Three cohorts were included in our study. We recruited 94 AD, 82 MCI and 85 HC from GAAIN (The Global Alzheimer’s Association Interactive Network) database, AIBL (Australian imaging, biomarkers and lifestyle) database, and our memory clinic database. Based on Centiloid pipeline, we processed the co‐registered MRI and PET images of each subject. Additionally, to compare quantitative amyloid load across different tracers, we converted SUVR (Standardized Uptake Value Ratio) values into standard Centiloid units. RF algorithm was performed on Python via scikit‐learn package. Finally, 122 regional volumes, 68 regional cortical thicknesses, 16 small regional amyloid distribution and 1 global amyloid load were input as features. Result In Table 1, the AUC value was highest (AUC=0.82) in the classification between HC and AD, intermediate (AUC=0.78) between HC and MCI and lowest (AUC=0.65) between AD and MCI. In each binary classification, sensitivity‐71%, specificity‐85%, accuracy‐78%; sensitivity‐88%, specificity‐76%, accuracy‐81% and sensitivity‐86%, specificity‐44%, accuracy‐66% were achieved in differentiating MCI from HC, AD from HC and AD from MCI respectively. For importance ranking of features within the model, 6 features of regional cortical thickness and 4 features of regional volume were found in the classification between HC and MCI, while all 10 essential features belong to regional Aβ ‐ROI in classification between HC and AD as well as MCI and AD (Table 2). Conclusion Random forest model using regional volume, regional cortical thickness and amyloid load (in Centiloid unit) had moderate to high accuracy in differentiating AD from HC and MCI.