Alzheimer’s disease (AD) is a neurodegenerative disease with progressive loss of memory and other functions that can be recognized using neuropsychological evaluation (Schmidtke and Hüll, 2002). Different domains of brain functions can be assessed through neuropsychological tests such as the Clinical Dementia Rating Sum of Boxes (CDRSB), Alzheimer’s Disease Assessment Scale Cognitive13 (ADAS13), Mini-Mental State Exam (MMSE), and Rey Auditory Verbal Learning Test (RAVLT) (Crane et al., 2012). Neuropsychological tests and neuroimaging data have been increasingly used for identifying patients with AD and mild cognitive impairment (MCI). MCI is typically considered a stage between healthy aging and AD (Ramakers et al., 2008; Duc et al., 2020). Some patients with MCI progress to AD, MCI converter (MCI-C), and the rest of them do not progress to AD, MCI non-converter (MCI-NC) (Gomez-Sancho et al., 2018; Shen et al., 2018). Considering that MCI-C and MCI-NC patients are in an intermediate stage between healthy aging and AD, it is challenging to identify robust and reliable biomarkers for prediction of progression of these patients to AD.

AD is a complex brain disease and integration of different neuropsychological tests is needed to adequately identify the evidence of dementia. It has also been shown in previous studies that neuropsychological scores are intrinsically correlated with each other (Van Der Maas et al., 2006; Tosi et al., 2020), and, thus, individuals with a healthy score in one neuropsychological test are more likely to achieve a healthy score in other tests. To efficiently identify the early stage of AD, it is critical to find the relationship and association between different neuropsychological tests.

It has been shown that AD starts years before any appearance of symptoms, and alterations in the brain structural and network characteristics occur while neuropsychological scores are still normal (Sperling et al., 2014). In recent years, several studies focused on diagnostic criteria for the early stage of AD based on neuroimaging biomarkers (Moradi et al., 2015; Hojjati et al., 2018, Hojjati et al., 2019). Neuroimaging-based diagnosis of AD [e.g., using structural magnetic resonance imaging (sMRI) and fluorodeoxyglucose (FDG) from positron emission tomography (PET)] can identify biomarkers that are sensitive to changes in the brain in the early stages of the disease (Eskildsen et al., 2013; Pagani et al., 2015; Hojjati et al., 2019; Tabarestani et al., 2020). Several neuroimaging studies have shown that the neuroimaging biomarkers may be more capable than the neuropsychological tests in identifying the early stage of AD (Leung et al., 2010; Gao et al., 2015; Hojjati et al., 2018; Lu et al., 2018).

Recent advances in the neurobiology of AD suggested that AD is a multifactorial and heterogeneous disease that cannot be explained by a single biomarker (Jack et al., 2018). Over the past decades, several projects have been funded to collect data from large cohorts of older adults. However, less effort has been made to implement integrative methods for aggregating data across modalities and capture the heterogeneity of AD. Previous studies utilized neuroimaging biomarkers, mostly in a single-modality approach, to find a relationship between neuropsychological tests and AD risk factors (Frisoni et al., 2002, 2010). It has been demonstrated that the trajectories of neuroimaging biomarkers for prediction of AD are complex and have a non-linear trend due to the interactions with age (Sabuncu et al., 2011). Hence, a non-linear relationship between neuropsychological scores and neuroimaging features should be considered. In the current study, we used artificial neural networks (ANNs) to investigate a non-linear model for AD progression trend based on neuropsychological scores. ANNs have been successfully utilized for various brain imaging applications (Savioz et al., 2009; Choi et al., 2020; Wen et al., 2020) and have gained increasing interest in diverse applications, such as classification, speech recognition, age modeling, modeling and forecasting extreme events, and even face recognition (Cole et al., 2017; Tuan Tran et al., 2017; Duc et al., 2020; Chowdhury et al., 2021). Moreover, ANN was utilized to provide an effective method for early diagnosis of AD (Wang et al., 2019). In particular, ANNs were shown to be effective in predicting progression from healthy cognition to AD (Albright and Alzheimer’s Disease Neuroimaging Initiative, 2019), and this predictive ability of ANNs outperformed that of a linear model (Grossi et al., 2007).

In this study, we investigated finding neuroimaging biomarkers as inputs features of an ANN to accurately model target neuropsychological scores (e.g., ADAS13) in a wide range of cognitive impairments, from healthy aging to severe AD, in four groups of subjects (healthy controls (HC), MCI-NC, MCI-C, and AD). By including data from MCI-C patients in our study, we investigated the accuracy of this model in a challenging condition by using their data in the baseline (when they showed a MCI) before they progressed to AD in the next several months or years. There are a few studies that proposed methods to predict and model the neuropsychological scores (Apostolova et al., 2006; Ferrarini et al., 2008). Duc et al. (2020) predicted the Mini-Mental State Exam (MMSE) score based on resting-state functional MRI using various regression methods (e.g., support vector regression and bagging-based ensemble regression) and utilized the 3-D convolutional neural network to classify two groups of subjects (HC and AD). In Ferrarini et al. (2008), the authors investigated an association between the atrophy in periventricular structures and cognitive impairment in MCI and AD, as estimated by the MMSE score. Moradi et al. (2017) utilized whole-brain gray matter density maps for predicting the RAVLT Immediate and RAVLT Percent Forgetting scores based on a machine learning approach. They reported R = 0.50 and R = 0.43 correlation between the estimated and observed RAVLT Immediate and RAVLT Percent Forgetting, respectively. Ito et al. (2011) attempted to fit a linear model for the longitudinal ADAS scores by including age, apolipoprotein 34 [APOE 34] genotype, gender, family history of AD, years of education, and baseline severity as inputs of the model. Their results showed that AD progression increased with baseline severity while age, APOE 34 genotype, and gender also influenced this progression. In another study, Liu et al. (2019) proposed supervised densely connected neural network methods to predict neuropsychological scores (i.e., CDRSB and MMSE) in three groups (HC, MCI, and AD). They utilized the landmark detection algorithm and found future multiple clinical scores in four different time-points using single modal neuroimaging of baseline MRI data.

This is the first study to our knowledge that attempted to model the progression trend of neuropsychological scores using a non-linear ANN approach based on bi-modal neuroimaging data in a large sample size (n = 1143) and four groups of subjects (HC, MCI-NC, MCI-C, and AD). In the current study we: (1) applied a machine learning approach based on ANN to predict ADAS13 and CDRSB scores (as the target scores) from normal aging to AD using neuroimaging biomarkers and other neuropsychological scores; (2) compared the performances of sMRI and PET in single- and bi-modal approaches to find the best neuroimaging biomarkers for predicting the target neuropsychological scores; (3) evaluated the ability of the best neuroimaging biomarkers in predicting the ADAS13 score in stable (i.e., MCI-C) and converting (i.e., MCI-NC) stages of AD; and (4) compared the progression trend of sMRI and PET biomarkers based on ADAS13 to find the prediction power of these biomarkers in different stages of AD. It is noteworthy that we chose the CDRSB and ADAS13 as the target tests because of their sensitivity to the assessment of the severity of AD progression and general cognitive functions (Kueper et al., 2018).

We utilized the ANN to investigate the association between the sMRI and PET biomarkers with two target neuropsychological scores (i.e., ADAS13 and CDRSB) from normal aging to AD. Using data in 951 subjects (341 HC, 393 MCI-NC, and 217 AD), our results revealed that the RAVLT Immediate, among five different neuropsychological tests (e.g., MMSE), provided the best prediction for the progression trend of ADAS13. Compared to other neuropsychological tests, however, the ADAS13 was the best predictor for the CDRSB score. Two remarks can be inferred from Figure 4 by exploring the association between ADAS13/CDRSB and other neuropsychological scores. First, the RAVLT Immediate was the best predictor for ADAS13 and the ADAS13 was the best predictor for the CDRSB. In addition, we observed a strong association between ADAS13, CDRSB, RAVLT Immediate, and MMSE. Second, the RAVLT Forgetting score is not correlated with ADAS13 or CDRSB scores, and, thus, the former score may provide independent information about the cognitive declines in AD compared to the latter scores. These observations indicate that the RAVLT tests should be integrated with other tests to assess different domains of cognitive declines in AD (Tosi et al., 2020). Previous studies reported that the RAVLT test can assist in identifying patients with subjective memory complaints who progress to AD, and this test can be used to differentiate MCI from normal aging (Estévez-González et al., 2003).

We found an association between neuroimaging (sMRI and PET) biomarkers and the target cognitive scores (i.e., ADAS13 and CDRSB) across a wide range of cognitive declines from normal aging to AD (Table 2). Uni-modal prediction of the target scores based on six sMRI features provided a competitive performance, but the entorhinal volume (followed by the hippocampus volume) outperformed other features (Figure 5). For PET features, results in Figure 6 show that the FDG outperformed AV45 for association with ADAS13 and CDRSB in three groups of subjects (i.e., HC, MCI, and AD). Our results for association of cognitive declines in AD (measured by the ADAS13 and CDRSB scores) with the volumes of the entorhinal cortex and hippocampus (extracted from sMRI) and the average FDG of the angular gyrus, temporal gyrus, and posterior cingulate are in line with previous studies (Wang et al., 2018; Prosser et al., 2020). The hippocampus is the core of the neural memory system and the entorhinal cortex is the hub for the widespread network for memory, navigation, and perception of time (O’Keefe, 1976; Eichenbaum et al., 2007). Atrophy of the hippocampus and entorhinal cortex have been highly associated with AD progression (Jessen et al., 2006; Devanand et al., 2007, 2012; Stoub et al., 2010). FDG-PET is also a well-known technique to identify the brain glucose metabolism which is mainly determined by synaptic activity in the brain. Previous studies reported capability of the FDG-PET in identification of neurocognitive declines in AD (Liu et al., 2018; Ding et al., 2019). It has been reported that AD patients have significantly reduced glucose metabolism in the angular gyrus, temporal gyrus, and posterior cingulate (Hunt et al., 2007). These areas are involved in memory deficit in the early stages of AD. A severe hypometabolic pattern has been associated with awareness of memory deficit (Nobili et al., 2010). It is well known that the posterior cingulate cortex and entorhinal cortex are involved in memory retrieval and self-referential processes (Izquierdo et al., 1997; Maddock et al., 2001), and a strong relationship between atrophy in these cortices with cognitive declines in AD has been reported (Mosconi et al., 2004; Hirao et al., 2006).

By comparing performances of the unimodal neuroimaging features for prediction of the target scores (i.e., ADAS13 and CDRSB), we found that the FDG-PET outperformed the six sMRI features, including the volumes of the entorhinal cortex and hippocampus. Furthermore, the bimodal prediction results (Table 2 and Figure 6) show that integrating two modalities (i.e., sMRI and PET) outperformed the unimodal approach. A previous study compared performance of FDG-PET and voxel-based morphometry (VBM) on MRI for classification of mild AD in unimodal and bimodal approaches (Kawachi et al., 2006). Results of this study revealed that the combination of two modalities provided a higher diagnostic accuracy (94%) compared to the unimodal approach based on FDG-PET (89%) or VBM-MRI (83%). Results of this study are in line with our findings showing superior performance of the bimodal approach compared to a unimodal approach. However, it is noteworthy that no study, to our knowledge, has yet investigated integrating PET and sMRI for predicting the ADAS13 and CDRSB across a wide range of cognitive decline from normal aging to severe AD. Previous studies investigated the relationship between the neuropsychological assessments and neuroimaging biomarkers (Godbolt et al., 2005; Musicco et al., 2009; Ito et al., 2011), and most of them utilized a single modality (typically sMRI) approach for this investigation (Frisoni et al., 2002, 2010; Apostolova et al., 2006). The structural-based biomarkers, such as gray matter volume and cortical thickness, have been utilized to find the association between neuropsychological scores and brain atrophy in AD (Frisoni et al., 2002; Zhou et al., 2013).

Our results in Figure 8 confirmed that integration of the entorhinal volume with the average FDG of the angular gyrus, temporal gyrus, and posterior cingulate was capable of predicting ADAS13 in the MCI-C, MCI-NC, and AD groups. Interestingly, the predicted trends of the ADAS13 score in the three groups show that the MCI-C group is an intermediate stage between MCI-NC and AD groups (Borroni et al., 2006). Another observation in Figure 8 for MCI-C subjects is that the ANN algorithm predicted a larger value for the ADAS13 score than its real value. This observation indicates that the neuroimaging biomarkers may be more sensitive in identification of the early stage of AD as the predicted value of ADAS13 by these biomarkers is larger than its real value in MCI-C patients. It has been reported that there is a long asymptomatic period (up to two decades) between the onset of brain changes and reaching an endpoint with the earliest development of clinical symptoms of AD (Mesulam, 1999; Perry and Hodges, 1999). In fact, many patients reach their end of life without developing fully characterized AD (Sperling et al., 2011). Therefore, current diagnostic criteria for AD depends significantly on the imaging biomarkers of the AD pathologies (DeKosky and Marek, 2003; Johnson et al., 2012; Bonifacio and Zamboni, 2016; Jagust, 2018) as the biological processes underlying AD may occur while the patient’s cognitive scores are still in MCI stage.

We investigated a relationship between the brain structural and functional neurodegeneration and the ADAS13 score from normal aging to severe AD (Figure 10). The curves related to the atrophy of the hippocampus and entorhinal cortex showed a plateau at the severe stage of AD (i.e., x > 0.75 in Figure 10) while the FDG-PET curve did not show this plateau. This observation indicates that the volumes of the entorhinal cortex and hippocampus have less sensitivity than the FDG-PET to model progression of AD in the severe stage of this disease. By comparing the structural atrophy in healthy aging subjects and patients with a very MCI, we found that the average volumes of hippocampus and entorhinal cortex across MCI-NC patients were 6.1 and 3.7%, respectively, smaller than that across HC subjects (Figure 9). These results are in agreement with previous studies that age-related atrophy in the medial temporal lobes occurs with larger hippocampal decline than the entorhinal cortex (Raz et al., 2004; Rodrigue and Raz, 2004). We also found that the average ADAS13 in MCI-NC patients corresponded with a larger atrophy in the hippocampus compared to the entorhinal cortex (57.32% vs. 46.73%, Figure 10). Considering that the ADAS score measures the severity of cognitive and non-cognitive dysfunction from mild to severe AD (Rosen et al., 1984), our results support a more important role of the hippocampus in this dysfunction compared to that of the entorhinal cortex. Specific to memory decline, Jessen et al. (2006) reported that the significantly affected entorhinal cortex causes slight memory dysfunction at the earliest clinically detectable stage of AD when patients experience worsening memory and the hippocampus is significantly affected with disease progression. Our results revealed that the volumes of hippocampus and entorhinal cortex reduced approximately 25% from a mild to severe stages of AD by comparing the average volumes of these structures in MCI-C and AD patients (Figure 9). Referring to the large atrophy of these structures in the mild to severe stages of AD, these two biomarkers showed a plateau for modeling ADAS13 in these stages in Figure 10. Based on the estimated ADAS13 curves corresponding to the volumes of hippocampus and entorhinal cortex, it can be inferred that these structural biomarkers had a good sensitivity in identification of the early stage of AD and they may not be sensitive for identification of the severe stage of AD. This observation is consistent with a general understanding of AD progression and the different stages of progression of dementia (Kung et al., 2021).

Our results revealed an association between FDG-PET and ADAS13/CDRSB. This observation shows that the glucose metabolism can measure alteration of cognition and functional ability in patients with MCI and AD. We also found that FDG-PET can track the AD progression and has a potential to be used as a clinically helpful measure of cognitive decline, particularly in MCI and AD patients. Previous studies reported that a lower FDG-PET at baseline has an association with greater longitudinal cognitive decline in AD (Alexander et al., 2002; Ottoy et al., 2019) and FDG-PET at a pre-dementia stage (MCI) has higher sensitivity to subsequent decline than neuropsychological tests (Chételat et al., 2005).

In this study, we used ANNs for quantitative analysis of the AD progression. This analysis can be used as a guide to help in the evaluation of different AD study designs, as well as to understand the complex relationship between various factors such as neuropsychological scores and imaging data. Establishing the relationships between neuropsychological scores and imaging data can help identify potential surrogates of clinical outcome and may guide to design future clinical trials.

Publicly available datasets were analyzed in this study. This data can be found here: https://tadpole.grand-challenge.org.

Ethical review and approval was not required for the study on human participants in accordance with the local legislation and institutional requirements. The patients/participants provided their written informed consent to participate in this study.

SH: preparing required data, data analysis, statistical analysis, and drafting and revision of the manuscript. AB-F: study design and conceptualization, data analysis and interpretation of the results, and drafting and revision of the manuscript. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.