Alzheimer’s disease (AD) is characterized by a prolonged prodromal phase, spanning from the preclinical stage to the prodromal stage before progressing to full-blown AD (Liu et al., 2021). Amnestic mild cognitive impairment (aMCI) represents a level of memory loss and cognitive decline in elderly individuals that falls short of dementia criteria and is considered a risk factor for Alzheimer’s disease (Oltra-Cucarella et al., 2018). Currently, clinical and research settings employ positron emission computed tomography (PET) imaging and the measurement of amyloid-β 42 (Aβ42) and Tau protein concentrations in cerebrospinal fluid as diagnostic biomarkers for AD. However, the high cost and radiation associated with PET scans limit their clinical utility. Cerebrospinal fluid examination via lumbar puncture, while providing valuable insights, carries the risks of local anesthetic allergies and intracranial infections. In contrast, structural magnetic resonance imaging (MRI) offers a non-invasive and highly repeatable alternative.

The hippocampus is an important brain region in the central nervous system involved in learning and memory storage. Hippocampal atrophy is one of the most effective, convenient, and widely used biomarkers for the clinical diagnosis of AD (Flores et al., 2015). Recent advancements in neuroimaging have revealed abnormal intracranial structures in AD patients, notably the reduction in hippocampal volume. Our previous research has identified disparities in the imaging characteristics of the left and right hippocampi in aMCI patients compared to normal controls, suggesting that hippocampal imaging features hold promise as potential biomarkers for aMCI and AD diagnosis (Feng et al., 2019). The hippocampus is an important part of the memory system, which is composed of different subfields with different functional characteristics, and each subregion is affected differently in the course of AD (Tamnes et al., 2014). West et al. (1994) reported a substantial neuronal loss, reaching up to 68%, in the hippocampal CA1 region of AD patients. Accurate measurement of hippocampal subfield volumes is of paramount importance for early diagnosis, and further advancements in this domain, including optimization, standardization, and automation of techniques, are imperative (Chételat, 2018).

Numerous methods have been employed to investigate hippocampal subfields in the context of AD spectrum diseases, including stereoscopic brain imaging, radial atrophy assessment, and the voxel-based morphometry (VBM) technique. One study examining hippocampal subfields in AD revealed that as the disease advances, the volume of both the hippocampus and its subfields (e.g., CA1, subiculum, presubiculum, molecular layer, and fimbria) gradually diminishes, with the left hippocampus exhibiting particularly pronounced changes (Zhao et al., 2019). Hari et al. (2022) found that compared with the MCI group and the subjective cognitive impairment (SCI) group, the volume of bilateral CA1, dentate gyrus DG, and subiculum decreased in the AD group; no significant volume change was found between the SCI and MCI groups. Boccardi et al. (2018) identified significant volume differences in 11 left hippocampal subfields (excluding the parasubiculum) among the AD, MCI, and NC groups, as well as in all 12 right hippocampal subfields among these groups. Moreover, Huang et al. (2022) found atrophy in CA1, molecular layer, subiculum, GC-ML-DG, CA4, and CA3 in the aMCI group, but the left and right hemispheres were not studied separately. However, studies have indicated that the asymmetry of hippocampal subfield atrophy in AD and MCI exists (Alessia et al., 2018; Jahanshahi et al., 2023). There is no consensus on how the hippocampal subfield atrophies.

There are relatively few studies based on hippocampal subfield volume modeling to distinguish aMCI from NC. Guo et al. (2020) applied FreeSurfer software version 5.0.0 to divide the bilateral hippocampus into 16 subregions. The achieved accuracy rates were 64.62%, 78.96%, and 70.33% for the support vector machine (SVM) classifier based on hippocampal subfield volumes in distinguishing stable MCI-NC, converted MCI-NC, and converted MCI-stable MCI cases, respectively. However, their study employed only a single machine learning method. Qu et al. (2023) found the volumes of the right subiculum, CA1, molecular layer, whole hippocampus, whole amygdala, basal, and accessory basal were significantly larger in NC compared to the MCI group. They achieved an area under the curve (AUC) of 0.79 in the integrated receiver operating characteristic (ROC) analysis for discriminating between MCI and NC; however, cross-validation was not conducted in their study.

Several studies have made attempts to uncover the intrinsic connection between the hippocampus or its regions and memory, suggesting their potential involvement in early changes or progression of AD. In one correlational analysis, it was found that within the aMCI group, volumes of the subiculum, presubiculum, and CA4/dentate gyrus were associated with delayed recall scores, and volumes of the subiculum and CA4/dentate gyrus correlated with informant-reported memory difficulties (O’Shea et al., 2022). Huang et al. (2022) found that in the aMCI group, visual memory scores were positively correlated with CA1, molecular layer, subiculum, and GC-ML-DG volumes. Zhao et al. (2019) found that the left subiculum showed a positive correlation with memory scores. However, employing a combination of multiple memory tests can be burdensome for elderly individuals and may not be suitable for routine clinical studies. Meanwhile, the Montreal Cognitive Assessment Scale (MoCA) offers a quick means of cognitive assessment that encompasses executive function, visuospatial recognition, memory, and other cognitive domains.

To address these considerations, this study pursued three primary objectives: (1) investigate changes in hippocampal subfield volumes in aMCI patients utilizing 3D T1 structural MRI and FreeSurfer 6.03 software’s detailed partitioning method; (2) develop machine learning classification models grounded in hippocampal subfield volumes; and (3) analyze the correlation between hippocampal subfield volumes and cognitive function scores, with a specific focus on MoCA.

This study investigated the change in volume in aMCI patients and established machine learning classification models based on the hippocampal subfield volume. The volumes of hippocampal tail, subiculum, CA1, presubiculum, molecular layer, GC-ML-DG, CA3, CA4, and fimbria were significantly different between the two groups, especially in the left hippocampus. The highest area under curve (AUC) value in the left and right hippocampal machine learning classifiers was 0.678 and 0.701, respectively. The volume of the left subiculum, left molecular layer, right subiculum, right CA1, right molecular layer, right GC-ML-DG, and right CA4 was most strongly and actively correlated with MoCA scores.

Hippocampal atrophy is widely recognized as an early and specific marker in AD patients. Neuroimaging studies have consistently reported varying degrees of involvement of the hippocampus and its subfields in AD and early AD (Chételat, 2018). In the volume studies of hippocampal subregions in AD patients, the atrophy of the CA1 region is a relatively consistent result (Flores et al., 2015). There have been many studies on hippocampal subfield volume using FreeSurfer software. It is commonly found that the volume of these subfields of the CA1, subiculum, presubiculum, molecular layer, GC-ML-DG, CA4, CA3, and fimbria decreases in MCI (Boccardi et al., 2018; Zhao et al., 2019; Hari et al., 2022; Huang et al., 2022). In our study, significant differences in the volumes of the hippocampal tail, subiculum, CA1, presubiculum, molecular layer, GC-ML-DG, CA3, CA4, and fimbria between the aMCI and NC groups were observed, with a pronounced effect noted in the left hippocampus. In terms of neuron loss, one study reported that the main subfields of loss in AD patients were CA1 (68%) and subiculum (47%) compared to controls (25%) (West et al., 1994). The presubiculum plays a unique role in AD; these studies showed that “lake-like” amyloid beta (Aβ) deposits appear under the presubiculum early in the disease process (Murray et al., 2018). Interneurons and synaptic connections in the molecular layer are involved in the regulation of activity in the hippocampus. Volumetric studies showed that the volume reduction mainly occurred in CA1, but also involved CA3/DG (Yassa et al., 2010), CA4/DG (Pluta et al., 2012). Fimbria plays an important role in executive function (Seiger et al., 2021). These results are consistent with most of the aforementioned findings, and a novel subregion, the hippocampal tail, is introduced, which appears to play a distinctive role in aMCI patients (Peter et al., 2018). The hippocampal tail is often associated with verbal memory. Parker et al. reported significant volume declines in CA1 and hippocampal tail in AD patients (Parker et al., 2018). Additionally, previous studies have indicated the importance of the right hippocampal tail in SCD patients, with significant reductions in gray matter volume (Liang et al., 2020).

There are relatively few studies based on hippocampal subfield volume modeling to distinguish aMCI from NC. One study showed that the accuracy of an SVM classifier based on hippocampal subfield volumes was 64.62% for stable MCI-NC, 78.96% for converted MCI-NC, and 70.33% for converted MCI-stable MCI (Guo et al., 2020). However, it used only one machine learning method. Qu et al. (2023) found the AUC was 0.79 for the identification of MCI and NC, but no cross-validation was performed. In the present study, the highest AUC values achieved for the left and right hippocampal machine learning classifiers were 0.678 and 0.701, respectively. Although the classification accuracy of the model is not very high, it also provides meaningful information, which can only say that the trend of hippocampal subfield volume change has diagnostic significance for aMCI, but cannot be said to have clinical value. The clinical value needs to be further studied.

Although there have been many studies showing that hippocampal subfield volume is associated with various memory scores (Zammit et al., 2017; Zhao et al., 2019; Huang et al., 2022; O’Shea et al., 2022). MoCA is a cognitive assessment that can measure executive function, visuospatial recognition, memory, and other cognitive functions, which is very important for the cognitive assessment of AD and aMCI patients. In our investigation, we identified that the volumes of the left subiculum, left molecular layer, right subiculum, right CA1, right molecular layer, right GC-ML-DG, and right CA4 were the most strongly and significantly correlated with MoCA scores. Liang et al. (2018) found that in MCI patients, left subiculum and presubiculum volumes were positively correlated with MoCA scores, but statistical p-values were not corrected. A study of AD and MCI in Japan Ogawa et al. (2019) found that the Japanese version of the MoCA had a higher correlation with the subfield volume of CA1, DG, subiculum, and entorhinal cortex than the MMSE. However, the sample size was relatively small. One study has found that some hippocampal subfield volumes were significantly correlated with the MMSE scores in AD patients (Chu et al., 2021). Nevertheless, in our study, there was no significant correlation observed between MMSE scores and the volumes of bilateral hippocampal subfields, potentially indicating the MMSE’s reduced sensitivity in identifying MCI.

The present study has several limitations. First, the sample size was relatively small, and the study was conducted at a single center. Expanding the sample size and incorporating data from multiple centers would allow for the creation of independent validation sets. Second, the study exclusively utilized structural MRI data; a better classification accuracy of approximately 0.9 could be achieved if other imaging metrics such as abnormal functional connection mode of hippocampus subfields, default mode network (DMN) modulation, and quantitative analysis of amyloid and tau using PET molecular tracers were included and more classifiers were implemented (Zhou, 2021). Future research should consider the integration of multimodal imaging data and other domain metrics, such as multiomic biospecimens and genetic data. Third, this study was cross-sectional in nature; further longitudinal studies are needed in future to observe the dynamic changes in hippocampal atrophy patterns in AD spectrum diseases.

Our findings indicate that multiple hippocampal subfield volumes undergo changes in individuals with aMCI. These alterations in hippocampal subfield volumes may serve as predictors of aMCI and appear to play a pivotal role in overall cognitive performance. Furthermore, ongoing refinement in the pattern of subfield atrophy holds promise for advancing precision medicine for MCI and AD patients, facilitating the identification of those MCI individuals at the highest risk of progressing to dementia.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by the Ethics Committee of Zhejiang Provincial People’s Hospital (2012KY002). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

QF: Formal analysis, Investigation, Methodology, Writing – original draft, Writing – review & editing. LW: Methodology, Validation, Writing – original draft, Writing – review & editing. XT: Methodology, Software, Writing – original draft. XG: Methodology, Validation, Writing – original draft. HH: Software, Visualization, Writing – original draft. ZL: Conceptualization, Validation, Writing – review & editing. ZD: Conceptualization, Validation, Writing – review & editing.