Development of biomarkers of aging and health outcomes in chest X-ray images using deep learning
Our project aims not only to develop and validate a deep learning model for age estimation from chest X-ray images, but also to expand the analytical framework by applying additional pretrained AI models to extract diverse information from chest X-rays. These models will be used to investigate associations with aging, disease onset, prognosis, and multiple demographic and clinical variables.
The study will proceed in two main stages.
Stage 1: Develop and validate an AI model that estimates age using chest X-ray images from multi-institutional health check data, excluding individuals with known diseases.
Stage 2: Apply the developed AI model, together with other pretrained AI models, to patient populations with various diseases. We will then examine the relationships between AI-derived features (including estimated age) and outcomes such as disease occurrence, prognosis, mortality, and broader demographic characteristics.
The goal of the proposed study is to establish chest X-ray–derived biomarkers of aging and health by integrating deep learning–based age estimation with broader AI-driven analyses.
Aim 1: Validate the accuracy of the AI age-estimation model using a multi-ethnic dataset. Since our primary dataset consists mainly of Japanese individuals, we will test whether accuracy is maintained across different populations.
Aim 2: Explore the relationship between estimated age, disease onset, and chronic conditions. Using logistic regression analysis and datasets such as NLST, we will evaluate correlations between estimated age and diseases including hypertension and COPD, and expand this to other chronic illnesses.
Aim 3: Investigate the relationship between estimated age and mortality rates of malignant diseases. Prior work has shown associations between estimated age and all-cause mortality. We hypothesize that AI-estimated age correlates not only with overall survival but also with disease-specific outcomes across malignancies.
Aim 4: Compare the predictive ability of existing prognostic indicators with AI-estimated age. To our knowledge, no prior studies have compared AI-derived age with established prognostic scores. We hypothesize that combining estimated age with conventional indices will improve prognostic accuracy beyond staging alone.
Aim 5: Apply additional pretrained AI models to chest X-rays to extract complementary information beyond age estimation. These models may detect patterns associated with early disease onset, organ health, or other biomarkers. We will analyze the relationships between these AI-derived features, disease occurrence, prognosis, and demographic characteristics.
Aim 6: Incorporate demographic, socioeconomic, and clinical variables (e.g., sex, age, first-time visit status, race, marital status, educational attainment, blood test results, social status) into our analyses. We aim to determine whether AI-derived age and other biomarkers interact with these variables to better explain health disparities and disease outcomes.
If AI-estimated age and related features prove to be useful biomarkers, they could become powerful tools for risk stratification, early detection, and prognostic prediction in future clinical practice.
Daiju Ueda M.D., Ph.D. Osaka Metropolitan University
Toshimasa Matsumoto Ph.D. Osaka Metropolitan University
Shannon L Walston M.S. Osaka Metropolitan University
Hiroyuki Tatekawa M.D., Ph.D. Osaka Metropolitan University
Hirotaka Takita M.D., Ph.D. Osaka Metropolitan University
Akira Yamamoto M.D., Ph.D. Osaka Metropolitan University
Yukio Miki M.D., Ph.D. Osaka Metropolitan University