2.1 Subjects
Three cohorts of patients with PET imaging data were collected, including 355 patients from the Alzheimer Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu, Cohort I), 22 patients from Huashan hospital (Cohort II), and 80 patients from Xuanwu hospital (Cohort III). Of these, subjects with amyloid-PET (Florbetapir F-18 [AV45]) scanning were selected for the Cohort I and III. This study was approved by institutional review boards of ADNI, Huashan, and Xuanwu hospital, and written informed consent was obtained from all participants or authorized representatives.
The inclusion criteria for data collection were as follows: 1) Subjects had both MRI and FDG PET scans and neuropsychological assessments (mini-mental state examination, MMSE) at baseline. 2) For MCI non-converters, patients did not convert to AD during the three-year follow-up period; for MCI converters, who converted to AD within three-year. 3) Participants with a bidirectional change of diagnosis (MCI to AD, and back to MCI) within the follow-up period were excluded. 4) SCD was defined by the research criteria for pre-MCI (SCD) proposed by Jessen et al in 2014 [20]; both MCI and AD dementia patients were included as cognitive impairment (CI) group.
2.3 Data pre-processing
Data preprocessing for both PET and MRI images was done by using Statistical Parametric Mapping 12 (SPM12, the Wellcome Department of Neurology, London U.K.) implemented in Matlab 2016b (Mathworks Inc.). First, original FDG PET image for each subject was registered with corresponding structural MRI image. Then, MRI images were segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) tissue probability maps using the unified segmentation method. The registered PET image was spatially normalized to the MNI space using the transformation parameters. Finally, the normalized PET images were smoothed using an isotropic Gaussian kernel of 8 mm to increase signal-to-noise ratios. Notably, all FDG PET images were underwent count normalization using the global cortical uptake to increase individual metabolic differences, as the mean cortical SUVR representing amyloid β level.
2.4 Radiomics-based predictive modelling analysis
In this study, we implemented RPM method to develop predictive models of brain-behavior relationships from glucose metabolic data. RPM method included the following steps: 1) radiomics feature extraction, 2) radiomics feature summarization, 3) model building and assessment, and 4) validation of crucial features.
Feature extraction
To obtain more detailed features, 80 cortical regions from the automated anatomical labeling (AAL) atlas were used as ROIs. We extracted 430 radiomics features from each ROIs for FDG PET data. In brief, the most basic features in each ROIs include two parts: first-order intensity features (n = 3) and texture features (n = 40). We extracted features under the combination of different wavelet filter weights (5 levels) and quantization of gray levels (2 levels), and the total number of features per ROIs was 430 ((3 + 40) * 5 * 2 = 430). The details of these features could obtain in supplement.
Feature summarization
The 10-fold cross-validation and Z-normalization strategies were implemented in Cohort I. To reduce the dimension of features and solve the over-fitting problem, three feature selection methods were performed separately for training data in Cohort I. The feature selection steps were as follows: 1) Feature stability analysis: the stable features with Cronbach's alpha coefficient greater than 0.75 were selected in longitudinal HC data in Cohort I. 2) Statistical test: t-test and rank sum t-test were used to identify the features with significant differences (P < 0.01) [21]. 3) Least absolute shrinkage and selection operator (LASSO).
Model construction and assessment
The Cox model was constructed using radiomics features from FDG PET. Ten-fold cross-validation was used to evaluate the prediction performance. Cox model was constructed during the training phase while selecting typical features. The clinical outcome was whether MCI subject was converted to AD. Time of outcome appearance was the interval between the baseline time and the endpoint. The Cox model was used in the test dataset to calculate the prognostic index (PI) for each subject. PI was a linear combination of the selected feature and its coefficient. The prediction performance of the model was evaluated using Harrell’s consistency coefficient (C-index). The C-index of the test dataset was calculated by PI. In addition, we also counted the number of times each feature participated in model construction. It is worth noting that the 10-fold cross-validation was also repeated 20 times, and the conserved features were also identified. These conserved features were used for further analysis.
Clinical Cox model was also constructed using available clinical variables (age, gender, education, and MMSE) to compare the predictive performance with RPM method. Cohort II data was used for external validation of the predictive model derived from RPM method.
Validation of crucial features
To further explore the relationship between conserved features and Neuropsychiatric assessments at different cognitive stages, partial correlation coefficients were calculated between the features neuropsychological assessments, and amyloid burden in individual subjects, combining MCI converters and MCI non-converters in Cohort I and combining SCD and CI subjects in Cohort III. The effects of age, gender and education were controlled. We also evaluated the differences in the identified features between HC, MCI-nc and MCI-c in Cohort I and HC, SCD and CI in Cohort III.
To further study the prediction performance of key features of important brain regions, we used features in the brain regions related to MMSE and amyloid β level as predictors to construct Cox models, respectively. The prognostic index of each subject was calculated according to the corresponding model, and then the individuals were stratified into high-risk and low-risk prognostic groups according to the median of prognostic index. Log-rank test was employed to examine the survival difference between different risk groups. We also combined these features to build a comprehensive Cox model and evaluated its performance for disease stratification.
As a comparison, the SUVR values of hippocampus, paracingulate gyrus and whole brain area as the feature were used to construct the SUVR Cox model. The process was the same as described above.
Statistical analysis
The group differences of clinical characteristics were assessed using two-sample t test, χ2 test, one-sample t test, or Tukey’s test. Log-rank test was employed to examine the survival difference between different risk groups. P values were 2-tailed, and P < 0.05 was considered statistically significant. All statistical tests were performed using SPSS 24.0. Cox model were constructed in R (http://www.R-project.org/) employing the ‘glmnet’ and ‘survival’ packages [22–24].