Preprint · Feb 18, 2021
Machine learning prediction of mild cognitive impairment conversion in a cohort of normal, non-demented elderlies (Preprint)
By Drima EP, Vrabie CD
BACKGROUND In aged subjects, mild cognitive conversion is difficult to assess from the clinical and medico-legal perspective as minimal changes are difficult to interpret. OBJECTIVE To evaluate how a machine learning algorithm (k-nearest neighbors) can predict conversion to mild cognitive impairment in a cohort of aged subjects with normal cognition, based on brain MRI data (hippocampus volume, whole brain volume), demographics and psychometric performance data. METHODS The study used raw imaging, demographics and psychometric tests result downloaded from OASIS (an open database dedicated to brain aging investigation). We used OASIS 1 and 2 datasets (T1-1.5T MRI). Within OASIS 2 (longitudinal) all subjects that converted (normal cognition - mild cognitive impairment) were carefully matched with normal controls and created a training set (28 cases). All normal patients form OASIS1 (81 cases) were included in a testing set. The right hippocampus was manually segmented and measured in all cases. MTA was measured as well. Psychometric test results and brain volume were available and were included as machine learning algorithm instances. After a training phase, normal patients from the test set were evaluated and classified by kNN algorithm. All machine learning files are provided in an annex. RESULTS Based on algorithm classification, 11 subjects from the OASIS 1 dataset (13.6%) were considered at risk for mild cognitive impairment conversion. The predicted time to conversion was calculated, using a linear regression algorithm, under two scenarios to 762.7±106.3 days (based on hippocampus volume) or 863.81±176.4 days (based on brain calculated volume) (p=0.11). Prediction was not confirmed in all cases but 6 subjects (7.5%) were identified between OASIS2 demented subjects based on demographic data (different identification labels). Limitations: low images resolution made the manual hippocampus segmentation time consuming and subjective. The total absence of clinical data was seen as main limitation of this study. CONCLUSIONS After a supervised learning phase, a machine learning algorithm may predict very mild cognitive conversion based on anatomic, psychometric and demographic data analysis. Clinical data can improve prediction, making the machine learning process clinically useful. CLINICALTRIAL None