Mild cognitive impairment (MCI) is constructed as an intermediate stage between normal aging and Alzheimer disease (AD). Various clinical criteria have been developed to quantify the risk of MCI patients converting to AD. One risk assessment criterion in assisting clinical decision is based on the amount of cerebral amyloid measured with florbetapir-fluorine-18 positron emission tomography (18F-AV45-PET) imaging. However, PET imaging is not usually readily available. As a result, the advantages of these important imaging based biomarkers may not be fully utilized clinically. To tackle the problem where patients have these biomarkers missing, we propose to develop ensemble regression tree to estimate the biomarkers based on clinical and demographic features (Age, APOE status, cognitive test, etc.) and other imaging biomarkers such as MRI. The makeup dataset filled with these estimates are then used to develop a classification model to assess the risk of MCI patients converting to AD. Using dataset of 146 MCI patients from Alzheimer's disease neuroimaging initiative (ANDI), we conduct 16 sets of experiments with the missing ratios changing from 0.05 to 0.80 to test the performance of our proposed approach. The advantages of our model show well when the missing ratio ranges 0.2 to 0.6 with average 7.1% higher accuracy and 7.4% higher sensitivity comparing to the model without using the estimated fill-ins. This advantage diminishes as the missing ratio increases to 80% as expected.