Effective and accurate diagnosis of AD is essential to initiate effective treatment. Particularly, early diagnosis of AD is pivotal in therapeutic development and ultimately for effective patient care [2]. ML methods can be used to identify clinical features and characteristic MR images and patterns for diagnostic predictions [59, 60]. This approach can potentially help with dementia predisposition identification in those who develop cognitive complaints [61–64], as this is a common question faced by clinicians. In this study, results showed that the ML model that utilized combined sMRI and rs-fMRI features was more accurate in predicting the diagnostic status of MCI than the ML models using brain morphological features only.
In this study, sMRI classification features were first extracted to identify atrophy patterns for the three groups (MCI, AD, HC). The findings were generally consistent with previous research, in which one of the earliest imaging marker was atrophy in the temporal lobe, and specifically in the parahippocampus and entorhinal cortex, which has also been previously found to predict MCI to AD progression [65, 66]. The left temporal CT features had the highest classification performance in AD vs. HC groups with an ACC of 92.05%, SEN of 95.35% and SPE of 88.89%. Similarly, for MCI vs. HC groups, the highest atrophy was observed in the left temporal cortex, with diagnostic accuracy of 88.64%, sensitivity of 85.00% and specificity of 91.11%. Finally, for MCI vs. AD groups, in the modeling using sMRI data of the frontal and temporal cortex, an ACC of 83.13% was obtained, which also verifies that the frontal and temporal lobe is involved in the early stage. These areas have been shown to be involved in complex cognitive behavior, decision making and personality expression [50], showing potential for the early identification of AD progression. Some individuals with MCI present with minimal visible structural brain changes. Some MCI patients have structural atrophy patterns similar to AD, suggesting that MCI may be more similar to AD than HC. The subtle changes between MCI and HC or the similarity between MCI and AD presents the ML models with a higher difficulty challenge to discriminate between the groups, and therefore requires more features. Additionally, clinical experience reflects this challenge as the differentiation between MCI and HC, as well as MCI and AD is based on functional decline, which is occasionally difficult to identify or quantify, causing overlap among groups. Thus, the results identifying features able to distinguish MCI from the other two groups are more clinically meaningful. Combining different types of data could help to further improve the accuracy of the ML model to distinguish between clinically relevant groups.
The classification results obtained by combining sMRI and rs-fMRI features in the present study are better than those obtained by the unimodal (sMRI\rs-fMRI) approach, including those of previous research [11, 67], for distinguishing MCI from AD and HC (Table 3). Most previous studies that constructed brain networks only considered cerebral structural or functional features, while ignoring the cerebellum, and obtained an accuracy lower than that of the present study [68–72]. In this work, the diagnostic performance of combined features of CT with FC from the right cerebellum to the left frontal lobe reached 94.43%. In addition, the mixed-features model reached an accuracy of 90.01% and 89.16% for AD vs. HC and MCI vs. AD, respectively. Previous work from our group revealed that the cerebellum showed increased activity of the frontal and temporal lobes in the pre-dementia stage of AD, reflecting a compensatory function that could mitigate early AD symptoms [35]. The current study demonstrates that the compensatory regulation of the cerebellum to the frontal lobe is the most significant, which strongly agrees with the findings relating to the enhanced FC between cerebellum and the frontal regions by cerebellum repetitive transcranial magnetic stimulation intervention presented in another work [73]. Therefore, it is postulated that when early cognitive impairment occurs, the frontal lobe receives increased assistance from the cerebellum. Furthermore, based on the classification model of CT combined with FC from right cerebellum to bilateral temporal and bilateral parietal lobes, respectively, high ACC of 88.24% and 90.59% were obtained for the classification of MCI vs. HC. This result outperformed other biomarkers for early MCI classification, indicating that the cerebellum has a high classification accuracy rate. Consequently, network-derived regulation is a highly effective biomarker. The crucial role of the cerebellum in brain regulation during the development of early cognitive impairment diseases was confirmed in this study. This conclusion may provide help for the diagnosis and treatment of preclinical AD in the future.
In this article, the potential of the cerebellum-oriented functional-coefficient network of the brain to be used as a diagnostic biomarker for subjects with early cognitive impairment was explored. The frontal lobe functioned as a “rich club” structure with highly interconnected nodes resulting in a large number of connections between networks [74]. These interconnections are involved in episodic memory, reasoning, and executive functions such as working memory and cognitive flexibility [75, 76]. The classification indicates that, as the brain function declines, in the early stage of cognitive impairment, the tightness of the cerebellum and some brain regions will increase. This may be a compensatory regulation mechanism, which only appears in certain brain regions.
However, ML studies have focused on only including clinical, cognitive and structural neuroimaging variables. A previous systematic review highlighted the need to explore additional types of data to improve the performance of ML models [77]. The results of the present study suggest that including cerebello-cerebral FC in ML models has the potential to increase the diagnostic utility and accuracy. These phenomena were hereby proved to be among the most important predictors, along with features of CT and FC from the right cerebellum to the left frontal lobe.
Limitations
The proposed method effectively improves the accuracy of preclinical diagnosis of AD; however, some limitations need to be highlighted. Future work will focus on the following improvements. To improve the effectiveness of the proposed approach, the data set will be expanded in the following aspects: First, extending the longitudinal data set to better understand the progression of MCI and inclusion of multi-modal data, such as gene and PET data, to investigate different insights into AD characteristics. Second, the parameter acquisition process will be optimized to achieve a higher diagnostic accuracy.