Prediction of short-term MCI-to-AD conversion using multi-modal neuroimaging features, and demographics

Regina EY Kim, Min-Woo Lee, Yeong Sim Choe, Hyeonsik Yang, Ji Yeon Lee, Jung Hyeon Yong, Donghyeon Kim, Minho Lee, Dong Woo Kang, So Yeon Jeon, Sang Joon Son, Young-Min Lee, Hyun Kook Lim, Hye Weon Kim Alzheimer's Association | 발행연도 2023.12.25


Multi-modality data can achieve better classification and regression performance than the use of only the single modality data. For this reason, we observed the potential of multiple data for early dementia conversion.

In this study, the conversion group was defined as cases where MCI patients converted to dementia between 2 and 4 years, and T1-weighted, T2 FLAIR, and Amyloid PET images were acquired from 4 domestic hospitals and ADNI. 114 volume features, WMH volume, and Fazekas scale were acquired from T1, T2 FLAIR images, and SUVR was calculated from amyloid PET. In addition to image features, age, sex, MMSE, and ApoE genotype were used. The 206 cases dataset was divided in a stratified way into a training set (80%) and testing set (20%), keeping the sample percentage of each class in both sets. We wanted to check the performance of each model on our dataset. The models used for this purpose were Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Linear Regression Classifier (LR), Gradient Boosting Model (GBM), Extreme Gradient Boosting (XGB). We trained and tuned each model, set up a grid search through hyperparameters to select a model that generalized well. The metrics used to evaluate model performance were BA, SE, SP, and AUC.

Table 2 and 3 present the obtained results of all ML models for the two-class classification task using 10-fold cross-validation and testing. In the case of 10-fold cross-validation, the model with highest mean BA was the SV model with 0.927. However, in the case of testing, we can see in Table 3 that the model with the highest BA, SE, and SP obtained was the GBM model, which was 0.917, 0.900, and 0.933, respectively. When ensemble was performed using the top3 models, BA, SE, SP, and AUC of the ensemble model were 0.917, 0.900, 0.933, and 0.963, respectively.

Muti-modal neuroimaging features with minimal demographic information showed reliable results in predicting the converters and non-converters between 2 and 4 years. This study suggests it could be utilized for future clinical trials to provide a high-risk group for dementia conversion in advance.