Two cohorts were used in this study:

Psychiatric evaluation and neuropsychological assessment were performed three times: at admission to the department of geriatric psychiatry (baseline) and discharge, and after 1-year. At each participating center trained clinicians performed the psychiatric and neuropsychological assessments using a harmonized procedure and the same protocol (Borza et al., 2015).

Clinical assessments included: demographic information, history of depression or other psychiatric condition, family history of psychiatric problems, current psychiatric problems, and treatment(s). Patients were diagnosed based on the criteria of the 10th Revision of the International Classification of Diseases and Health Related Problems (ICD-10; World Health Organization [WHO], 1993). Cognition was measured with the Mini Mental Status Examination, MMSE (Folstein et al., 1975) and Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE-16; Jorm and Jacomb, 1989). At the 1-year follow-up assessment patients were also diagnosed using Winblad’s criteria for MCI (Winblad et al., 2004).

Patients without dementia diagnoses according to ICD-10 criteria at inclusion and a T1W 3D/structural brain MRI were included in this study (n = 95, Figure 1). Not all the patients could be included in the final analyses for a variety of reasons, as specified in Figure 1. Based on the 1-year follow-up diagnosis the patients were divided into one mild cognitive impairment-dementia (MCI-DEM) group including those diagnosed with either MCI (n = 21; Winblad et al., 2004) or dementia (n = 8, including: Alzheimer’s disease n = 5, unspecified dementia n = 3) [World Health Organization (WHO)]. A matched cognitively stable (CS) group was selected from the remaining 51 patient based on age and gender (n = 40) and controlling for the balance in number of CS and MCI at each center of inclusion (Figure 1).

The participating patients and caregivers were given oral and written information, and they subsequently gave their written consent to participate. The PRODE study was approved by Regional Committee of Medical Research Ethics and Privacy and Data Protection Officer at Oslo University Hospital (approval number 2009/1774).

R programming language (R Core Team, 2016), version 3.3.0, was used to compare demographic and clinical data, and to create random-forest algorithm (RF) classification models. For group comparisons of demographic and clinical variables, the chi-squared test was used for categorical variables, and t-test or Mann–Whitney where appropriate for continuous variables. RF classification models were established using R packages “random forest” (Liaw and Wiener, 2002) and “caret” (Kuhn et al., 2016), ROC-curve and area under the ROC-curve (AUC) were estimated and plotted using “pROC” package (Robin et al., 2011). Regions important for the classification were correlated with clinical and demographic variables.

Random forest method allows performing supervised classifications based on an ensemble of classification trees (Breiman, 2001). RF selects a bootstrapped subset of all observations – about 66% per tree and random subset of all predictors/features (here: CTH, SV) at each node of the tree. RF uses the majority vote of its trees terminal nodes to predict the class label of a new observation. Each tree casts a unit vote for the class. Thus, high numbers of decision trees are expected to increase reliability of the results. For each predictor a Gini index is estimated at each node. Overall importance of a predictor for the model is based on the summation of the decreases in the Gini index at each node (Breiman, 2001). The remaining 33% of the data, i.e., out-of-bag (OOB) data, is used to measure the RF performance. The classification error of the OOB observations is referred to as OOB error (Breiman, 1996). Kappa is another measure of performance demonstrating how close the RF classifications were to the actual classes, controlling for the accuracy of a random classifier as measured by the expected accuracy. Kappa is suggested to provide more reliable information regarding the classifier performance than actual accuracy in case of different class distribution in the dataset (Fleiss, 1971).

Clinical and demographic variables in MCI-DEM and CS groups are provided in Table 1. MCI-DEM group had significantly lower MMSE scores at all three time points compared with CS group. Other variables did not differ between the groups. Similarly, the MCI group alone did not differ from the matched CS patients in terms of age (p = 0.21) or gender (p = 0.92). Notably, the standard deviation of the MMSE scores increased in the MCI/dementia group between assessments “at discharge” and “follow-up,” reflecting increased variance in their cognitive statuses.

The most important regions for predicting a diagnosis of MCI or dementia after 1-year follow-up were volumes of R-VD, mid-anterior CC, and R-HC, respectively.

Right ventral diencephalon was demonstrated to be the most relevant structure. The ventral diencephalon in Freesurfer includes several structures: hypothalamus with mammillary body, subthalamic, lateral geniculate, medial geniculate and red nuclei, substantia nigra and surrounding white matter. Some of these structures, i.e., substantia nigra and red nuclei, are not locate to the diencephalon but mesencephalon according to standard anatomical nomenclature. The hypothalamus is known to be strongly involved in depression. As a part of hypothalamic–pituitary–adrenal (HPA) axis, it is crucial for emotional behavior and stress response. There are numerous studies showing dysregulation of the HPA axis in depression, but also in aging and neurodegeneration (Sapolsky et al., 1986; Sapolsky, 2000; Varghese and Brown, 2001; Du and Pang, 2015). It has also been proposed that HPA-axis dysfunction is central to the development of AD (Ishii and Iadecola, 2015). Indeed, hypothalamic dysfunction can explain the overlap in symptoms between depression and AD (mood, appetite, sleep, memory, autonomic). Consistent with our finding, several previous imaging studies have shown structural and functional abnormalities in the hypothalamus in MCI, preclinical AD, and AD compared with control groups (Callen et al., 2001; Nestor et al., 2003; Cross et al., 2013; Brueggen et al., 2015). For instance, Hall et al. (2008) have demonstrated reduced basal forebrain and hypothalamus volumes in preclinical AD, and interestingly the combination of reduced forebrain and hippocampal volumes was associated with more rapid cognitive decline. However, to the best of our knowledge the current study is the first study to assess the role of the entire ventral diencephalon in the context of neurodegeneration. There are two previous studies examining the ventral diencephalon (segmented in Freesurfer) in relation to mood disorders. The first showed that bilateral ventral diencephalon volume obtained from Freesurfer was one of three top-ranked endophenotypes of major depressive disorder in an analysis of a high-dimensional set of over 11,000 traits (Glahn et al., 2012). The other study demonstrated that volume of the ventral diencephalon discriminated patients with major depressive disorder from those with bipolar depression as well as controls (Sacchet et al., 2015). Taken together, these findings suggest that neurodegeneration in ventral diencephalon including the hypothalamus might be a link between depression and cognitive impairment.

After R-VD, CC (mid-anterior CC) and hippocampal (R-HC) volumes were the most relevant structures for predicting MCI-DEM. HPA-axis dysregulation is associated with elevated cortisol levels (O’Brien, 1996; Du and Pang, 2015) which is hypothesized to cause reduced CC and hippocampal volumes due to neurotoxic effects (Bao et al., 2008; Liu et al., 2016). The importance of CC volume for predicting MCI-DEM suggests presence of inter-hemispheric disconnection already in the early stages of the neurodegenerative process. Previously, it has been shown that anterior, middle and posterior portions of the CC have less volume in AD compared to controls, but only the middle part was smaller in amnestic MCI compared with controls (Qiu et al., 2016). Moreover, decreased volume of mid-anterior portion of CC has been linked to the memory loss in MCI and AD (Qiu et al., 2016). The present findings support a connection between R-VD, R-HC, and CC pathology in the development of MCI and dementia in LLD, which may be linked to HPA-axis dysregulation. A previous study examining only CC volume reported that structural changes in CC predicted MCI-to-AD conversion after 2.5 years on average (Lee et al., 2016), similar to the current study where a whole brain approach was used and conversion was to MCI/dementia. Hippocampal abnormalities are one of the most replicable findings in both depression and AD (Kempton et al., 2011; Sabuncu et al., 2011). Hippocampus has large bidirectional connection with the mammillary bodies of the hypothalamus which also might explain the concordant changes in these two structures. The limbic-diencephalic pathways, including the mammillothalamic tract and the mammillary bodies per se are crucial for episodic memory (Vann and Nelson, 2015; Aggleton et al., 2016). Hypometabolism in the mammillary bodies has been shown in both MCI and AD (Nestor et al., 2003). Limbic-diencephalic pathway dysregulation has been shown in the earliest stages of AD (Acosta-Cabronero and Nestor, 2014). Indeed, limbic regions, which are crucial for emotion processing, are also crucial for episodic memory. Thus, abnormal limbic-diencephalic interaction may be a core feature of in MCI-DEM development in LDD.

Of note, only brain structures from the right hemisphere were important for the classifications. This supports the right hemi-aging model, proposing that the right hemisphere shows greater age-related decline than the left hemisphere (Dolcos et al., 2002). However, this model is based on behavioral data rather than neuroimaging and the evidence has been controversial (Daselaar and Cabeza, 2005).

Interestingly, the number of depressive episodes had significant inverse association with the R-VD and mid-anterior CC volumes. Previously, it was reported that the number of depressive episodes is associated with reduced volume of the dentate gyrus in patients with major depressive disorder (Treadway et al., 2015). It is not known whether reductions of R-VD and mid-anterior CC are developmental phenomena, which leads to increase of the number of depressive episodes or degenerative consequences of a larger number of depressive episodes. However, including total ventricular volume to the regression model negated the effect of the number of depressive episodes on the R-VD volume and reduced on mid-anterior CC, providing indirect evidence for the degenerative nature of R-VD and mid-anterior CC reduction. Future studies using longitudinal imaging data could uncover the causality.

MCI-DEM had significantly lower MMSE scores at all three time points compared to the CS group. Regardless of whether cognitive assessment in depressed patients is confounded by depression per se, including MMSE scores at admission improved the model’s prediction of MCI and dementia. Along with a recent study of Heser et al. (2016), the present results suggest that even during the depressive episode cognitive impairment require clinical attention as a possible sign of incipient dementia. Given the highly significant difference in MMSE scores at all three time points between the MCI/DEM and CS group, the probability that there was a true difference in MMSE scores between MCI-DEM and CS groups at all time points is very high, suggesting that MMSE in an appropriate additional predictor of MCI/AD even in LLD groups.

There was no correlation between MMSE scores R-VD and mid-anterior CC volumes suggesting that MMSE scores did not bias classification results. The absence of correlation between MMSE and the two main predictors suggested that MMSE improved classification models performance by explaining additional factor variance.

The SV model trained on MCI vs. CS demonstrated better performance than the model trained on the mixed sample of MCI-DEM dementia vs. CS. The reasons might be a more heterogeneous pattern of structural brain changes across dementias in the MCI/DEM combined group and presence of larger variability in the stages of pathological process. On the other hand, the CTH+SV model performed better on the mixed MCI+DEM group compared with MCI only. One explanation might be that CTH get altered in the later stages and/or more severe cases of cognitive impairment in LLD and that relevance of the structural brain measurement as a biomarker therefore depends on the stage of the disease.

MCI-DEM status was discriminated from the CS group with 67% accuracy using the ADNI model. In other words, patients in the MCI-DEM group were more likely to be classified as cases than those in CS group based on their baseline brain MRI.

Interestingly, the ADNI model trained only on SV had much higher accuracy compared with the one trained on CTH+SV. Taking into account that the most relevant structures for MCI-CS discrimination (PRODE models) were subcortical structures and that the model combining SV with CTH had higher accuracy in MCI-DEM than in MCI alone, the present results may suggest that in LLD patients, who will develop MCI or dementia, neurodegeneration appears to start in subcortical structures and spread up to the neocortex in later stages. In agreement with our findings it has been shown that atrophy in basal forebrain and hypothalamus but not neocortex, precedes clinical symptoms of AD by 4.5–5 years (Brueggen et al., 2015). Several earlier studies have shown hypometabolism restricted to hippocampus and parahippocampal gyrus in MCI, whereas AD had additional temporal neocortical hypometabolism (De Santi et al., 2001; Nestor et al., 2003). Taken together, the current results suggest that classification models for prediction of MCI/preclinical AD should focus on subcortical structures rather than the neocortex.

One of the limitations of the present study is that the time interval between MRI assessment and MCI diagnosis was relatively short, thus it is possible that some CS patients could be diagnosed with MCI/dementia at a later time point. Excluding LLD participants with the lowest MADRS and MMSE scores at discharge (n = 4) did not alter the results notably (no substantial drop in accuracy) which indicate that the possibility that some patients could receive MCI diagnosis earlier did not bias the results strongly. Another issue is that MMSE might not be sensitive enough to detect MCI. In any case, there are no cognitive screening instruments validated for use in a depressed elderly population according to our knowledge. All the models including only brain measurements as predictors had very good specificity and sufficient sensitivity. One of the reasons for this might be the relatively small sample size. In future studies classification models should be trained on larger samples after longer follow-up periods. Inclusion of MRI scans obtained from scanners with different magnetic field strengths might be considered a limitation; however, it was shown that Freesurfer output measurements are consistent between 1.5T and 3T MRI, which was confirmed in our reliability analysis. We have shown that ICC is high across the centers. We also balanced the number of participants from each center and verified the results excluding each center, in order to assess reliability of the results which were consistent.

Our results suggest that LLD with smaller volumes of the R-VD, mid-anterior CC, and R-HC and lower MMSE scores at inclusion have a higher probability of receiving a diagnosis of MCI or dementia the following year. These measurements can provide clinicians with novel evidence for an expected trajectory of cognitive functioning in LLD and help to define a target group for interventions against cognitive decline (Bredesen, 2014).

Future mechanistic studies should verify processes underlying diencephalic neurodegeneration in LLD patients.

Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd. and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.