The training and test data used in this study were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database¹. The ADNI was launched in 2003 as a public-private partnership led by the National Institute on Aging (NIA), the Food and Drug Administration (FDA), and National Institute of Biomedical Imaging and Bioengineering (NIBIB). The ADNI aims to aid researchers and clinicians in developing new treatments and monitoring their effectiveness as well as to lessen the time and cost of clinical trials. Up-to-date information can be found at www.adni-info.org. The use of the ADNI data was approved by the institutional review board at each site, and all the participants provided their written consent.

A total of 144 ADNI participants were included in this study 36 EOAD, 36 LOAD, 36 young control (YC), and 36 old control (OC) participants from the ADNI1, ADNI2/GO, and ADNI3 databases. Scans were collected at either screening or baseline visits. First, 36 patients diagnosed with AD onset before the age of 65 years (EOAD) who were enrolled in the ADNI database were eligible for this study. Next, we included 36 patients who were 65 years or older at disease onset (LOAD) and who were 1:1 matched to the EOAD patients by the Clinical Dementia Rating (CDR) Scale. Accordingly, we selected two control groups for each patient group. The controls were matched 1:1 to AD patients for age and sex, thus obtaining a YC group for EOAD (n = 36; YC) and an OC group for LOAD (n = 36; OC). Furthermore, demographic information, medical history, baseline symptoms, and assessment scale scores were included. The MRI and clinical data were downloaded in June 2021.

Independent external validation data were acquired from the Memory Clinic of Shanghai Mental Health Center (SMHC) between July 2017 and May 2021, and normal control subjects were recruited from the community. A total of 60 participants including 15 EOAD, 15 LOAD, 15 YCs, and 15 OCs were enrolled in this study. Similarly, 15 LOAD patients were also 1:1 matched to the EOAD patients using the CDR Scale, and the controls were matched to AD patients for age and sex. EOAD and LOAD patients were diagnosed by two experienced geriatric psychiatrists. The exclusion criteria included the following: (1) other psychiatric disorders comorbidities; (2) a history of major physical illness, cardiovascular disease, or neurological disorder; (3) substance abuse or dependence; (4) pregnancy. Neuropsychological tests and brain imaging scans were performed in all subjects. The retrospective study was approved by the ethics committee of the Shanghai Mental Health Centre of Shanghai Jiao Tong University School of Medicine, and all the participants provided written informed consent after they were given a description of this study.

Regarding ADNI data, T1-weighted structural imaging was collected using a 3D MPRAGE (magnetization prepared rapid gradient-echo imaging) sequence with slightly different MR parameters among participants. The MR images acquired using Siemens scanner were scanned with the parameters as follows: repetition time (TR) = 2,300 ms, matrix = 240 × 256 × 176, slice thickness = 1.2 mm, and those parameters in General Electric scanner were as follows: TR = 7 ms, matrix = 256 × 256 × 166, slice thickness = 1.2 mm and those parameters in Philips scanner were as follows: TR = 6.8 ms, matrix = 256 × 256 × 170, slice thickness = 1.2 mm, respectively. More detailed information about the image acquisition procedures is available on the ADNI website². Additionally, the MR data of Shanghai Mental Health Center were acquired using a Siemens Magnetom Verio 3.0 T scanner, and high-resolution T1-weighted structural images with 176 sagittal slices were collected using a MPRAGE sequence (TR = 2,530 ms, TE = 3.5 ms, flip angle = 9°, FOV = 256 mm × 256 mm, voxel size = 1.0 × 1.0 × 1.2 mm³).

Standardized preprocessing was necessary to improve discrimination between textural features and was performed using Statistical Parametric Mapping (SPM12) software³ implemented in MATLAB R2017a (The MathWorks, Natick, MA, USA). Firstly, each T1-weighted Digital Imaging and Communications in Medicine (DICOM) image was converted to Neuroimaging Informatics Technology Initiative (NIFTI) data. Secondly, correction for bias field inhomogeneities and intensity normalization of images were performed in the VBM12 toolbox. The corrected images were normalized to the Montreal Neurological Institute (MNI) standard T1 template (standard space 181 × 217 × 181 with a resolution of 1 mm × 1 mm × 1 mm) using DARTEL normalization. Then, the obtained images were spatially normalized to ensure that a given voxel corresponded to the same anatomical position in different subjects. Finally, we resliced those images to the standard MNI space with a resolution of 1 mm × 1 mm × 1 mm.

Segmentation of the hippocampus was required to describe the texture characteristics of the region of interest (ROI). First, the bilateral hippocampus from the Anatomical Automatic Labeling (AAL) template provided by the MNI was chosen as the ROI mask. Then, the open-source software 3D-slicer⁴ was applied for medical image visualization and segmentation (Fedorov et al., 2012). Specifically, the viewer window of 3D-Slicer was used to select image “layers”, including “background” image and “label” image. Then, the standardized preprocessing image of each subject was loaded as the “background” image, and the left and right hippocampus mask was loaded as the“label” image, respectively. Next, two expert radiologists worked together to check the segmentation of the hippocampus for each subject and manually modified the unsatisfied image in the “Segment Editor” window of 3D slicer after reaching a consensus. In fact, a previous study has shown that the dice similarity coefficient (DSC) between the manual segmentation and atlas-based methods in brain structure segmentation are 0.79 (Ourselin et al., 2013). In our study, the combination of atlas-based segmentation and manual inspections could assure the segmentation quality and improve the time consumption.

First, we loaded the standardized 3D T1-MPRAGE data for the EOAD, LOAD, YC, and OC subjects into 3D-slicer software, and then we imported the segmented left and right hippocampus. Massive features were selected using the “pyradiomics” package of the software⁵, including the histogram-based matrix (HISTO), GLCM, gray-level dependence matrix (GLDM), gray-level size zone matrix (GLSZM), GLRLM, and neighboring gray-tone difference matrix (NGTDM) in the “feature classes” window.

HISTO is a statistical description of discrete units, while the GLCM using second-order statistics reflects the spatial relationship of pixel gray-level values in the image (Dhruv et al., 2019). The GLDM is also based on the gray-level relationship to acquire the first-order statistics of local property values, and the GLRLM estimates the spatial relationships between groups of pixels with similar gray-level values (Araujo et al., 2018). The GLSZM can be used to compute different pixel distances, whereas the NGTDM measures the total differences in the gray level of a pixel (Thibault et al., 2014).

Before feature selection, preprocessing was required for accurate and valid selection. First, we checked the extracted data and replaced the abnormal values that deviated more than three standard deviations from the mean by the mean. Considering that deleting the abnormal values may cause loss of information and the lack of processing may affect the model construction, combined with the normal distribution of the data, we decided to replace outliers with the mean. Next, the subjects from ADNI data were randomly divided into training and test datasets at proportions of 0.7 and 0.3 for EOAD-YC groups, LOAD-OC groups, and EOAD-LOAD groups, respectively. Then, every extracted feature was standardized by the function of sklearn.preprocessing.scale based on Python programming to achieve Z-score normalization to remove the dimensional constraint.

We used Python programming to accomplish feature selection. First, t-test and Mann-Whitney U test were used to select the features with significant differences (p < 0.05). Next, correlation analysis was performed to further reduce the dimensionality. If the correlation coefficient of two feature columns exceeded 0.8, we removed one of them randomly. Finally, the least absolute shrinkage and selection operator (LASSO) regression analysis method with 10-fold cross validation was applied to determine the most valid features in the training data, and the corresponding lambda value was selected with minimum mean-squared error (MSE) values. The mechanism of LASSO, combining the penalty function and linear regression, makes some regression coefficients become zero and achieve dimension reduction (Tibshirani, 2013).

Support vector machine (SVM) algorithms were used to construct radiomic models for EOAD and YCs, LOAD and OCs, and EOAD and LOAD. SVM is one of the most popular and mature machine learning algorithms based on the neuroimaging literature (Orru et al., 2012). The SVM model employs a radial basis function kernel using LIBSVM⁶ to implement nonlinear mapping from the input space to the feature space (Chang and Lin, 2011). Accordingly, the SVM models were used to construct the prediction models of the EOAD-YC groups, LOAD-OC groups, and EOAD-LOAD groups based on the selected prediction features in training sets, and then the test sets were used to calculate the predictive efficiency based on the predictive models, respectively (Nalepa and Kawulok, 2019). Then, all subjects from the data from Shanghai Mental Health Center were used as independent external validation sets to verify the reliability and robustness of the corresponding models. Additionally, receiver operating characteristic (ROC) curves and the corresponding areas under the curve (AUCs) were used to evaluate the diagnostic capabilities of the radiomic features.

Statistical analyses were performed using SPSS software 22.0 (IBM Corporation, Armonk, NY). The demographic information of the participants was collected as numbers or means ± SD for categorical and continuous variables. The comparisons between the EOAD and OC (EOAD-OC), LOAD and YC (LOAD-YC), and EOAD and LOAD (EOAD-LOAD) subjects were performed using χ² test for categorical variables and Student’s t-test for continuous variables (two-tailed) to evaluate the differences between groups. A p < 0.05 was considered statistically significant.

The demographic and clinical characteristics of the four groups are presented in Table 1. No difference was found in age or sex between the EOAD patients and YCs (EOAD-YC) or between the LOAD patients and OCs (LOAD-OC) in the ADNI and SMHC data. The Mini-Mental State Examination (MMSE) scores were significantly different in the EOAD-YC and LOAD-OC groups (p < 0.001). No significant differences were found in the clinical dementia rating (CDR) scores and MMSE scores between the EOAD and LOAD patients (EOAD-LOAD).

A total of 214 features were extracted from the bilateral hippocampus. After t-test and Mann-Whitney U test, 99, 102, and 37 features were preserved in the EOAD-YC, LOAD-OC, and EOAD-LOAD groups, respectively. After correlation analysis, 51, 73, and 24 (Figure 1) features remained. Finally, the LASSO regression model identified three, four, and four features for the EOAD-YC, LOAD-OC, and EOAD-LOAD groups (Table 2). Meanwhile, the values of the coefficients and the corresponding lambda values, and the MSE values and the corresponding lambda values for the EOAD-YC, LOAD-OC, and EOAD-LOAD groups are shown in Figure 2.

The accuracy (ACC), sensitivity (SEN), specificity (SPE), and AUC were used to evaluate the classification performance. Figure 3 and Table 3 show the final classification performance on the test set and validation set. In the analysis between the EOAD patients and YCs, the ACC, SEN, SPE, and AUC were 0.77, 0.91, 0.64, and 0.90 in the test set and 0.87, 0.87, 0.87, and 0.91 in the validation set, respectively (Figure 3A). By contrast, in the LOAD patients and OCs, the ACC, SEN, SPE, and AUC were 0.86, 0.87, 0.86, and 0.94 in the test set and 0.78, 0.85, 0.70, and 0.92 in the validation set, respectively (Figure 3B). Finally, in the analysis between the EOAD and LOAD patients, the ACC, SEN, SPE, and AUC were 0.79, 0.67, 0.93, and 0.87 in the test set and 0.77, 0.60, 0.93, and 0.86 in the validation set, respectively (Figure 3C). Similar classification performance was found in the test and validation datasets, indicating that our models may have relatively good robustness.