A total of 196 subjects were enrolled from four tertiary hospitals and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. Within a four-year follow-up period, we defined the subjects as the dementia conversion group when their global Clinical Dementia Rating (CDR) score reached 1.0 or higher within the follow-up period. Subjects maintaining a global CDR score of 0.5 were defined as the non-conversion group. The collected demographics of all sites are (1) age, (2) sex, (3) Mini-Mental State Examination (MMSE), (4) ApoE 4 Carrier, (5) CDR. Those aged 50-85 years, diagnosed with MCI at the time of initial treatment, and who underwent follow-up diagnostic tests within 2–4 years were included in the eligibility criteria. Approval of the MRI and αPET images used for this study was obtained from the Yeouido St. Mary's Hospital Institutional Review Board (IRB) [2022-1185], the IRB of Chungnam National University Hospital (CNUH-2022-05-020), the IRB of Ajou University Hospital (AJIRB-MED-EXP-22-284) and the IRB of Kyung Hee University Hospital (KNUH-2022-05-012) with a waiver of informed consent. All conformed to the Declaration of Helsinki (https://www.nature.com/srep/journal-policies/editorial-policies#experimental-subjects). Image acquisition methods are described for each site.
Site1 dataset underwent to brain MRI and PET at the Catholic University of Korea, Yeouido St. Mary’s Hospital, Seoul, Republic of Korea. A dataset satisfying the conversion definition was extracted and 44 non-conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. The Site1 dataset was acquired from human subjects on 3.0T a Siemens scanner. T1-weighted MRI images were acquired (TR=1700~1800ms, TE=2.6ms, and flip angle=9 ). T2 FLAIR MRI images were acquired (TR/TI=9000/2500ms, TE=76ms, Flip angle=150 ). αPET images were acquired with 18F-Florbetaben, 18F-Flutemetamol.
Site2 dataset underwent brain MRI and PET at Chungnam National University Hospital, Daejeon, Republic of Korea. A dataset satisfying the conversion definition was extracted, and two non-conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. 3D T1-weighted MRI images were acquired on a 3.0T Siemens (TR=2000ms, TE=2.29ms, flip angle=8 ), 3.0T GE (TR=7.956ms, TE=2.82ms, flip angle=10 ). T2 FLAIR MRI images were acquired on a 3.0T Siemens (TR/TI=9000/2500ms, TE=121ms, Flip angle=121 ), 3.0T GE (TR/TI=11000/2648.61ms, TE=93.544, flip angle=160 ). αPET images were acquired with 18F-Flutemetamol.
Site3 dataset underwent to brain MRI and PET at the Ajou University Hospital, Suwon, Republic of Korea. A dataset satisfying the conversion definition was extracted, and 34 non-conversion and 3 conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. 3D T1-weighted MRI images were acquired on a 3.0T GE (TR=7.1~8.88ms, TE=2.776~3.396ms, Flip angle=8 or 12 ), 3.0T Philips (TR=9.8ms, TE=4.6ms, Flip angle=8 ). T2 FLAIR MRI images were acquired on a 3.0T GE (TR/TI=8800–12000/2450~2709ms, TE=89~128ms, Flip angle=160 ), 3.0T Philips (TR/TI=8000/2500ms, TE=125ms, Flip angle=90 ). αPET images were acquired with 18F-Flutemetamol.
Site4 dataset underwent to brain MRI and PET at Kyung Hee University Medical Center, Seoul, Republic of Korea. A dataset satisfying the conversion definition was extracted, and 29 non-conversion and 14 conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. 3D T1-weighted MRI images were acquired on a 3.0T Philips (TR=9.4ms, TE=4.6ms, Flip angle=8 ), 3.0T Siemens (TR=2000ms, TE=3.05ms, Flip angle=9 ). T2 FLAIR MRI images were acquired using a 3.0T Philips (TR/TI=10000/2800, TE=120 or 125ms, Flip angle=90 ) a 3.0T Siemens (TR/TI=8000~10730/2500~2665.9ms, TE=86~115ms, Flip angle=150 ). αPET images were acquired with 18F-Florbetaben.
For this study, we used the ADNIMERGE subset, in which demographic and clinical test scores and MRI and PET variables were summarized. This subset is part of the official dataset provided by the ADNI. When data satisfying the conversion definition were extracted from the subset, 40 non-conversion and 12 conversion groups were obtained. 3D T1-weighted MRI images were acquired on a 3.0T GE (TR=7.3~7.6ms, TE=3.05~0.12ms, Flip angle=11 ), 3.0T Philips (TR=6.5ms, TE=2.9ms, Flip angle:9 ), 3.0T Siemens (TR=2300ms, TE=2.95~2.98ms, Flip angle=9 ). T2 FLAIR MRI images were acquired on a 3.0T GE (TR/TI=4800/1442~1482ms, TE=115.7~117ms, Flip angle=90 ), 3.0T Philips (TR/TI=4800/1650ms, TE=271~275ms, Flip angle=90 ), 3.0T Siemens (TR/TI=4800 or 9000/1650~2500ms, TE=90~443ms, 120 ). αPET images were acquired using 18F-Florbetapir, 18F-Florbetaben.
The acquired 3D T1 images were pre-processed and segmented into 114 ROIs using AQUA (Neurophet, South Korea). After calculating the volume of the segmented area, intracranial volume (ICV) normalization was performed. In addition, the hippocampal occupancy score (HOC), which is used as an index of neurodegenerative disease biomarkers [13], was calculated and used as an input. The white matter hyper intensities (WMHs), periventricular WMHs, and deep WMHs were calculated from the T2 FLAIR image, where registration was applied to the 3D T1 image, and the Fazekas scale was rated for each region as minimal (0), moderate (1), and severe (2). The acquired αPET images were also registered with 3D T1 images, the voxels in αPET images were scaled using the mean uptake value in the cerebellar gray matter to calculate the standardized uptake value ratio (SUVR). Consequently, the number of features of each modality used as input was 115 for volumetric information, 6 for WMH information, and 144 for SUVR information.
We used the synthetic minority oversampling technique (SMOTE) to remove the possibility of biased prediction by balancing dementia conversion and non-conversion data. Standardization was performed to ensure the same level of importance, and all features were used in the model. For this reason, the z score method was used, where is the original value for feature j, is the normalized value, is the feature’s mean and is the feature’s standard deviation. Consequently, the z-score method produces a new dataset in which all features have zero mean and unit standard deviation. The values for categorical features were also encoded.
The purpose of ICV normalization was to correct for differences in the ROI volume due to the different head sizes of individuals and sexes. This was performed by dividing the total intracranial volume (ICV) by each volumetric feature of the subject. This normalization method is commonly used [14].
The 196 cases dataset was divided in a stratified way into a training set (80%) and a testing set (20%), maintaining the sample percentage of each class in both sets. We evaluated the performance of each model using our dataset. The models used were decision trees (DT), random forests (RF), support vector machines (SVM), linear regression classifiers (LR), gradient boosting models (GBM), and Extreme Gradient Boosting (XGB). We trained each model and set up a grid search using the hyperparameters to select a model that generalized well. In the process of hyperparameter tuning, a 10-fold cross-validation was performed. Three categories were considered for the prediction models. The models were constructed based on demographic characteristics and each modality (T1, T2-FLAIR, and αPET) features. Single-modality models were built using demographic characteristics and modality features. To observe the change in model performance according to the change in input information, we selected a model that was excellent in a single modality among the three models, with high performance for each model. After training the single-modality model, the modalities with relatively low performance were excluded. By adding modality features to the demographic characteristics using the selected model, the change in the performance of the model according to the input information was confirmed. The reason for considering this approach was that it was possible to check the performance change when the demographic characteristics and modality features were combined, and which combination showed good performance.
Because a two-class task model was developed to predict the possibility of dementia conversion in the MCI patient group, four metrics representing classification performance were used as model evaluation indicators. where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives (Eq. 1, 2, 3, and 4):
A total of 196 subjects were enrolled from four tertiary hospitals and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. Within a four-year follow-up period, we defined the subjects as the dementia conversion group when their global Clinical Dementia Rating (CDR) score reached 1.0 or higher within the follow-up period. Subjects maintaining a global CDR score of 0.5 were defined as the non-conversion group. The collected demographics of all sites are (1) age, (2) sex, (3) Mini-Mental State Examination (MMSE), (4) ApoE 4 Carrier, (5) CDR. Those aged 50-85 years, diagnosed with MCI at the time of initial treatment, and who underwent follow-up diagnostic tests within 2–4 years were included in the eligibility criteria. Approval of the MRI and αPET images used for this study was obtained from the Yeouido St. Mary's Hospital Institutional Review Board (IRB) [2022-1185], the IRB of Chungnam National University Hospital (CNUH-2022-05-020), the IRB of Ajou University Hospital (AJIRB-MED-EXP-22-284) and the IRB of Kyung Hee University Hospital (KNUH-2022-05-012) with a waiver of informed consent. All conformed to the Declaration of Helsinki (https://www.nature.com/srep/journal-policies/editorial-policies#experimental-subjects). Image acquisition methods are described for each site.
Site1 dataset underwent to brain MRI and PET at the Catholic University of Korea, Yeouido St. Mary’s Hospital, Seoul, Republic of Korea. A dataset satisfying the conversion definition was extracted and 44 non-conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. The Site1 dataset was acquired from human subjects on 3.0T a Siemens scanner. T1-weighted MRI images were acquired (TR=1700~1800ms, TE=2.6ms, and flip angle=9 ). T2 FLAIR MRI images were acquired (TR/TI=9000/2500ms, TE=76ms, Flip angle=150 ). αPET images were acquired with 18F-Florbetaben, 18F-Flutemetamol.
Site2 dataset underwent brain MRI and PET at Chungnam National University Hospital, Daejeon, Republic of Korea. A dataset satisfying the conversion definition was extracted, and two non-conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. 3D T1-weighted MRI images were acquired on a 3.0T Siemens (TR=2000ms, TE=2.29ms, flip angle=8 ), 3.0T GE (TR=7.956ms, TE=2.82ms, flip angle=10 ). T2 FLAIR MRI images were acquired on a 3.0T Siemens (TR/TI=9000/2500ms, TE=121ms, Flip angle=121 ), 3.0T GE (TR/TI=11000/2648.61ms, TE=93.544, flip angle=160 ). αPET images were acquired with 18F-Flutemetamol.
Site3 dataset underwent to brain MRI and PET at the Ajou University Hospital, Suwon, Republic of Korea. A dataset satisfying the conversion definition was extracted, and 34 non-conversion and 3 conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. 3D T1-weighted MRI images were acquired on a 3.0T GE (TR=7.1~8.88ms, TE=2.776~3.396ms, Flip angle=8 or 12 ), 3.0T Philips (TR=9.8ms, TE=4.6ms, Flip angle=8 ). T2 FLAIR MRI images were acquired on a 3.0T GE (TR/TI=8800–12000/2450~2709ms, TE=89~128ms, Flip angle=160 ), 3.0T Philips (TR/TI=8000/2500ms, TE=125ms, Flip angle=90 ). αPET images were acquired with 18F-Flutemetamol.
Site4 dataset underwent to brain MRI and PET at Kyung Hee University Medical Center, Seoul, Republic of Korea. A dataset satisfying the conversion definition was extracted, and 29 non-conversion and 14 conversion groups were obtained. MRI and PET images were obtained from patients with mild cognitive impairment. 3D T1-weighted MRI images were acquired on a 3.0T Philips (TR=9.4ms, TE=4.6ms, Flip angle=8 ), 3.0T Siemens (TR=2000ms, TE=3.05ms, Flip angle=9 ). T2 FLAIR MRI images were acquired using a 3.0T Philips (TR/TI=10000/2800, TE=120 or 125ms, Flip angle=90 ) a 3.0T Siemens (TR/TI=8000~10730/2500~2665.9ms, TE=86~115ms, Flip angle=150 ). αPET images were acquired with 18F-Florbetaben.
For this study, we used the ADNIMERGE subset, in which demographic and clinical test scores and MRI and PET variables were summarized. This subset is part of the official dataset provided by the ADNI. When data satisfying the conversion definition were extracted from the subset, 40 non-conversion and 12 conversion groups were obtained. 3D T1-weighted MRI images were acquired on a 3.0T GE (TR=7.3~7.6ms, TE=3.05~0.12ms, Flip angle=11 ), 3.0T Philips (TR=6.5ms, TE=2.9ms, Flip angle:9 ), 3.0T Siemens (TR=2300ms, TE=2.95~2.98ms, Flip angle=9 ). T2 FLAIR MRI images were acquired on a 3.0T GE (TR/TI=4800/1442~1482ms, TE=115.7~117ms, Flip angle=90 ), 3.0T Philips (TR/TI=4800/1650ms, TE=271~275ms, Flip angle=90 ), 3.0T Siemens (TR/TI=4800 or 9000/1650~2500ms, TE=90~443ms, 120 ). αPET images were acquired using 18F-Florbetapir, 18F-Florbetaben.
The acquired 3D T1 images were pre-processed and segmented into 114 ROIs using AQUA (Neurophet, South Korea). After calculating the volume of the segmented area, intracranial volume (ICV) normalization was performed. In addition, the hippocampal occupancy score (HOC), which is used as an index of neurodegenerative disease biomarkers [13], was calculated and used as an input. The white matter hyper intensities (WMHs), periventricular WMHs, and deep WMHs were calculated from the T2 FLAIR image, where registration was applied to the 3D T1 image, and the Fazekas scale was rated for each region as minimal (0), moderate (1), and severe (2). The acquired αPET images were also registered with 3D T1 images, the voxels in αPET images were scaled using the mean uptake value in the cerebellar gray matter to calculate the standardized uptake value ratio (SUVR). Consequently, the number of features of each modality used as input was 115 for volumetric information, 6 for WMH information, and 144 for SUVR information.
We used the synthetic minority oversampling technique (SMOTE) to remove the possibility of biased prediction by balancing dementia conversion and non-conversion data. Standardization was performed to ensure the same level of importance, and all features were used in the model. For this reason, the z score method was used, where is the original value for feature j, is the normalized value, is the feature’s mean and is the feature’s standard deviation. Consequently, the z-score method produces a new dataset in which all features have zero mean and unit standard deviation. The values for categorical features were also encoded.
The purpose of ICV normalization was to correct for differences in the ROI volume due to the different head sizes of individuals and sexes. This was performed by dividing the total intracranial volume (ICV) by each volumetric feature of the subject. This normalization method is commonly used [14].
The 196 cases dataset was divided in a stratified way into a training set (80%) and a testing set (20%), maintaining the sample percentage of each class in both sets. We evaluated the performance of each model using our dataset. The models used were decision trees (DT), random forests (RF), support vector machines (SVM), linear regression classifiers (LR), gradient boosting models (GBM), and Extreme Gradient Boosting (XGB). We trained each model and set up a grid search using the hyperparameters to select a model that generalized well. In the process of hyperparameter tuning, a 10-fold cross-validation was performed. Three categories were considered for the prediction models. The models were constructed based on demographic characteristics and each modality (T1, T2-FLAIR, and αPET) features. Single-modality models were built using demographic characteristics and modality features. To observe the change in model performance according to the change in input information, we selected a model that was excellent in a single modality among the three models, with high performance for each model. After training the single-modality model, the modalities with relatively low performance were excluded. By adding modality features to the demographic characteristics using the selected model, the change in the performance of the model according to the input information was confirmed. The reason for considering this approach was that it was possible to check the performance change when the demographic characteristics and modality features were combined, and which combination showed good performance.
Because a two-class task model was developed to predict the possibility of dementia conversion in the MCI patient group, four metrics representing classification performance were used as model evaluation indicators. where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives (Eq. 1, 2, 3, and 4):
$$Sensitivity \left(SE\right)=\frac{TP}{\text{T}\text{P}+\text{F}\text{N}} \left(1\right)$$
$$Specificity \left(SP\right)=\frac{TN}{\text{T}\text{N}+\text{F}\text{P}} \left(2\right)$$
$$Balanced Accuracy \left(BA\right)=\frac{(Sensitivity+Specificity)}{2} \left(3\right)$$
$$AUC=Area under ROC curve \left(4\right)$$
Statistical analysis
The Kruskal-Wallis test was performed to assess the influence of T1 image features and αPET image features on predicting AD conversion in patients with MCI (Figure 2). Statistically significant superiority for a variable was determined by a p-value < 0.05. But adjustment for significance tests for multiple comparisons, such as comparing the 10-fold cross validation results of each model by image features, was made using the Bonferroni correction, which adjusts the significance level to p-value < 0.05/3 (0.017).