BACKGROUND: Early detection of individuals at risk for converting to Alzheimer’s disease (AD) can potentially lead to more efficient treatment and better disease management. A well-known approach has aimed at identifying individuals at the prodromal stage of dementia; namely, Mild Cognitive Impairment (MCI). Past studies showed that MCI subjects often have accelerated rates of conversion to AD, or to other types of dementia compared to healthy controls (HCs). However, with more investigations of the MCI population, it became evident that a high level of heterogeneity exists within this group: many remain clinically stable even after 10 years. MCI subtypes defined by the conventional classification criteria showed inconsistent results for determining an individual's risk of AD. As another approach, neuroimaging techniques such as magnetic resonance imaging (MRI) are able to successfully identify neurological changes during early AD. MRI markers including morphological, connectional and abnormal signal patterns in the brain have been shown to have good sensitivity for classifying AD. Based on these findings, recent studies started implementing these imaging markers to create computer-aided classification models for predicting the risk of conversion to AD. Most of these studies enrolled MCI subjects who remained stable or converted to AD within 3 years, and generated computer-aided classification models to predict conversion using various imaging markers and clinical data. To our knowledge, no classification models proposed achieved an accuracy of higher than 80% for predicting MCI-AD conversion earlier than 3 years with only using structural MRI features. In this paper, we tested the prediction range beyond 3 years, and suggested new candidate imaging measures for earlier prediction.
METHODS: The subjects included in the current study are n=51 MCI non-converter, n=157 MCI converter (115 fast converters and 42 slow converters) and n=38 AD, selected from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Using subjects' baseline T1-weighted MRI scans, we combined conventional morphometric measures (e.g. cortical thickness, surface area, volume, etc.) with novel intensity measures to differentiate MCI converters from non-converters. We additionally applied a machine learning approach to classify MCI subgroups by combining features in multiple measurement domains.
RESULTS: Based on group comparison using independent t-test, we found that while MCI fast converters (conversion within 0-2 years) were highly distinct from MCI non-converters across many cortical and subcortical regions, MCI slow converters (conversion within 3-5 years) demonstrated more focal differences from MCI non-converters mainly in the temporal regions and hippocampal subfields. We identified unique imaging features associated with each converter group and had improved classification performance on both MCI converter groups by adding those markers. The best performing classifiers combined conventional imaging features, novel intensity features and neuropsychological features. For our best performing classification models, we were able to classify MCI fast converters (0-2 years) from non-converter with an average accuracy of 86.1%, sensitivity of 85.5%, and specificity of 89.8%, and to classify MCI slow converters (3-5 years) from non-converters with an accuracy of 80.5%, sensitivity of 75.7%, and specificity of 82.3%.
CONCLUSION: Our results demonstrated the potential of the suggested approach for predicting the conversion from MCI to AD at an even earlier time point (3-5 years) before the onset of AD. The combination of standard morphometric features and proposed novel intensity features improved the sensitivity of using T1-weighted MRI for describing the heterogeneity between MCI subgroups.
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/41246 |
| Date | 19 June 2020 |
| Creators | Guan, Yi |
| Contributors | Koo, Bang-Bon |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.002 seconds