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 in-cipientAlzheimer'sDisease(AD)dementiainmildcognitiveimpairment(MCI).Focusedonprovidinganearlierand more accurate diagnosis, sophisticated MRI machine learning algorithms have been developed over therecentyears,mostofthemlearningtheirnon-diseasepatternsfromMCIthatremainedstableover2–3years.Inthis work, we analyzed whether these stable MCI over short-term periods are actually appropriate trainingexamples of non-disease patterns. To this aim, we compared the diagnosis of MCI patients at 2 and 5years offollow-up and investigated its impact on the predictive performance of baseline volumetric MRI measures pri-marily involved in AD, i.e., hippocampal and entorhinal cortex volumes. Predictive power was evaluated interms ofthe areaunder the ROCcurve(AUC), sensitivity,andspecificity inatrialsample of248 MCIpatientsfollowed-up over 5years. We further compared the sensitivity in those MCI that converted before 2years andthose that converted after 2years. Our results indicate that 23% of the stable MCI at 2years progressed in thenextthreeyearsandthatMRIvolumetricmeasuresaregoodpredictorsofconversiontoADdementiaevenatthemid-term, showing a better specificity and AUC as follow-up time increases. The combination of hippocampusand entorhinal cortex yielded an AUC that was significantly higher for the 5-year follow-up (AUC=73% at2yearsvs.AUC=84%at5years),aswellasforspecificity(56%vs.71%).Sensitivityshowedanon-significantslightdecrease(81%vs.78%).Remarkably,theperformanceofthismodelwascomparabletomachinelearningmodels at the same follow-up times. MRI correctly identified most of the patients that converted after 2years(with sensitivity>60%), and these patients showed a similar degree of abnormalities to those that convertedbefore 2years. This implies that most of the MCI patients that remained stable over short periods and subse-quentlyprogressedtoADdementiahadevidentatrophiesatbaseline.Therefore,machinelearningmodelsthatuse these patients to learn non-disease patterns are including an important fraction of patients with evidentpathological changes related to the disease, something that might result in reduced performance and lack ofbiological interpretability.