Probabilistic computer‐aided diagnosis of Alzheimer’s disease and frontotemporal dementia based on MRI and biochemical markers

Abstract

AbstractBackgroundOne of the clinical problems for biomarkers' clinical use is the ability to differentiate between Alzheimer’s disease (AD), frontotemporal dementia (FTD), and healthy subjects (CTR). This clearly challenges diagnosis and prognosis. We implemented a ML algorithm that provides individual probabilistic diagnoses for these dementias based on magnetic resonance imaging (MRI) and we correlated the results with biochemical markers.MethodWe studied 3T‐T1w MRI of 432 subjects. A subset of these participants had cerebrospinal fluid (CSF) and plasma biomarkers (Table 1). We obtained regional subcortical gray matter volumes and cortical thickness measures using Freesurfer. We implemented a calibrated classifier with a Support Vector Machine with only the MRI data. We tested paired‐wise classification and classification across the 3 groups. We obtained individual probabilities associated with group correspondence. We studied the correlation between these probabilities and CSF and plasma biomarkers. For this, we subdivided the groups into true‐group (subjects with classification according to clinical diagnosis) and false‐group (subjects which did not coincide with clinical diagnosis). Finally, we implemented a permutation test to find the importance of each region in the classification.ResultWe obtained accuracies of 90.7 ± 6.7% in the CTR vs AD classification, 88.6 ± 4.5% for CTR vs FTD, 79.3 ± 8.8% for AD vs FTD, and 79.9 ± 5.1% when discriminating the 3 groups. We obtained a significant positive correlation for plasma p‐tau181 for the false‐AD in the comparison AD vs CTR (Figure 1). The correlation of the false‐CTR was significantly positive for the CSF and plasma NfL. Finally, in the AD vs FTD, the true‐FTD had a significant negative correlation with CSF NfL. The other biochemical biomarkers did not provide additional information. The most important regions for classification are shown in Figure 2.ConclusionThe ML algorithm gave high accuracies. Within wrongly classified AD patients, probabilities correlated positively with plasma p‐tau181, suggesting hidden pathological processes associated with subjects clinically classified as CTR. Finally, the group probability within well‐classified FTD patients in comparison with AD negatively correlated with CSF NfL. We suggest that this approach can be used as a tool to try to develop personalized diagnoses.