Trajectory based predictive modeling of conversion from mild cognitive impairment to Alzheimer's disease

Accurate prediction of clinical changes of Mild Cognitive Impairment (MCI) patients at future time points is important for early diagnosis and possible prevention of Alzheimer's disease (AD). In this paper, future clinical changes in Neuropsychological Measures (NM) of MCI patients are estimated via three different predictive models employing linear regression and extrapolation. The completed time domain trajectories are processed in the Euclidean space to extract features encompassing clinical change in the biomarker values. These features are then fed to an optimized SVM classifier to predict conversion of an MCI patient to AD. A wrapper based biomarker subset selection is adopted to analyze the effect of single and combined NM biomarkers. The proposed predictive modelling techniques were validated on 186 MCI subjects with the selected NM readings available at baseline and 3 annual follow-up visits. 1 and 2 year ahead conversion prediction was made using the current and at least 1 year old biomarker reading. Maximum accuracy of 77.87% and 73.24% is achieved for 1 and 2 year ahead conversion prediction respectively.