Prediction of Alzheimer's disease dementia with MRI beyond the short-term: Implications for the design of predictive models.

Abstract

Magnetic resonance imaging (MRI) volumetric measures have become a standard tool for the detection of incipient Alzheimer's Disease (AD) dementia in mild cognitive impairment (MCI). Focused on providing an earlier and more accurate diagnosis, sophisticated MRI machine learning algorithms have been developed over the recent years, most of them learning their non-disease patterns from MCI that remained stable over 2–3 years. In this work, we analyzed whether these stable MCI over short-term periods are actually appropriate training examples of non-disease patterns. To this aim, we compared the diagnosis of MCI patients at 2 and 5 years of follow-up and investigated its impact on the predictive performance of baseline volumetric MRI measures primarily involved in AD, i.e., hippocampal and entorhinal cortex volumes. Predictive power was evaluated in terms of the area under the ROC curve (AUC), sensitivity, and specificity in a trial sample of 248 MCI patients followed-up over 5 years. We further compared the sensitivity in those MCI that converted before 2 years and those that converted after 2 years. Our results indicate that 23% of the stable MCI at 2 years progressed in the next three years and that MRI volumetric measures are good predictors of conversion to AD dementia even at the mid-term, showing a better specificity and AUC as follow-up time increases. The combination of hippocampus and entorhinal cortex yielded an AUC that was significantly higher for the 5-year follow-up (AUC = 73% at 2 years vs. AUC = 84% at 5 years), as well as for specificity (56% vs. 71%). Sensitivity showed a non-significant slight decrease (81% vs. 78%). Remarkably, the performance of this model was comparable to machine learning models at the same follow-up times. MRI correctly identified most of the patients that converted after 2 years (with sensitivity >60%), and these patients showed a similar degree of abnormalities to those that converted before 2 years. This implies that most of the MCI patients that remained stable over short periods and subsequently progressed to AD dementia had evident atrophies at baseline. Therefore, machine learning models that use these patients to learn non-disease patterns are including an important fraction of patients with evident pathological changes related to the disease, something that might result in reduced performance and lack of biological interpretability.