To investigate whether CXR can be implemented as a biomarker for aging
1) Collect CXR images and data
2) Biomarker modeling
Label the collected CXR images with the diagnosed disease.
Training dataset: Used to train the AI model.
Tuning dataset: Used to internally tune the AI model.
Internal test dataset: A dataset collected from the same organization as the training and tuning datasets and not used for training and tuning.
3) Building and training AI models
Establish an AI model consisting of a deep neural network to estimate patient age. First, the AI model is trained using only CXR images as training data. After training, fine-tune the hyper-parameter variables through validation, and verify whether the AI model predicts the patient's age as intended through test data.
4) Model evaluation and statistical analysis of prediction results
We then check the error between the actual patient's age and the model's predicted age. Apply various loss functions such as MSE, RMSE, and MAE and select the best loss function for the AI model. Update the parameters of the model through backpropagation with the adopted loss function to gradually reduce the error of the AI model.
To investigate the degree of relationship between each disease and age difference, we perform logistic regression analysis with each disease as the target variable and age difference and actual age as the predictors. To address the interaction with disease, we also perform a multivariate analysis technique that includes chronic disease as a predictor. In sub-analyses, we incorporate gender as a predictor and calculate odds ratios (ORs) for malignancy in each organ.

Estimating age and diagnosing cardiovascular prognosis with AI models from CXRs.
1) Minimizing the deviation between estimated age and actual age using AI models
Establish a deep neural network AI model that learns based on the collected data. The deviation between the actual age of the patient and the age estimated by the model is checked, and the internal variables of the model are modified several times to minimize the deviation. By reducing the deviation, the accuracy of the AI model's patient age estimation is improved.
2) Statistical analysis of test results and correlation between aging and cardiovascular disease
After finding no significant difference in the prediction performance of the AI model and CXR data, we investigated the reproducibility of the model by extracting images of the same patient taken multiple times. The results show that age can be accurately estimated from CXRs in a variety of populations and cohorts.
We produce a heatmap-like morphology of previously reported CXR-based age prediction models, demonstrating that tortuosity and calcification of the aorta are hallmarks of atherosclerosis and are associated with old age, consistent with the literature, and suggesting that our AI model captures age-related changes.
The results extracted from the test demonstrate that the difference between the estimated age and the actual age can be an outcome indicator of CXR, which indicates the presence of disease, and thus reveals the relationship between aging and cardiovascular disease.

Daegu Hanny University: Choi wojin
Yeungnam University: Park sebin
Yeungnam University: Kim Namgyu
Keimyung University: Lee jinho