Biological Age Estimation From the Age Gap Using Deep Learning Integrating Morbidity and Mortality: Model Development and Validation Study.

Background: Biological age (BA) is increasingly recognized as a valuable alternative to chronological age (CA) for assessing an individual's health and aging status. However, existing models are based on limited clinical parameters and have not thoroughly integrated morbidity and mortality data. Objective: This study aimed to develop and validate a novel transformer-based model, referred to as the BA - CA gap model, for BA estimation that incorporates morbidity and mortality information to improve predictive accuracy and enhance clinical use in the early identification of the risk of age-related diseases. Methods: We retrospectively analyzed data from 151,281 adults aged 18 years or older who underwent routine health checkups between 2003 and 2020. Participants were classified into normal, predisease, and disease groups based on comorbidities (diabetes mellitus, hypertension, and dyslipidemia) to evaluate the model's ability to discriminate health status along a clinically relevant spectrum. Variables with less than 50% missingness had missing values imputed using the mean, while features with 50% or more missingness were excluded. We develop a custom transformer architecture that learns multiple objectives simultaneously, including input feature reconstruction, BA and CA alignment, health status discrimination, and mortality prediction. Model training used unsupervised and self-supervised strategies. We compared our model's performance with conventional BA estimation approaches, including Klemera and Doubal's method, a CA cluster-based model, and a deep neural network, by examining BA gap distributions, health status stratification, and mortality prediction. Results: The proposed BA - CA gap model provided a more accurate reflection of health status and superior stratification of mortality risk than existing methods. The model effectively distinguished among normal, predisease, and disease groups, with a clear gradient of BA gap values. Kaplan-Meier analyses demonstrated stronger discrimination of future mortality in men, while a similar but not statistically significant trend was observed in women. Sensitivity analyses across multiple random splits and training subsets confirmed the robustness of the model's performance. Conclusions: By integrating morbidity and mortality information within a transformer-based framework, the BA - CA gap model offers a more granular and clinically meaningful assessment of aging and health status than CA alone. This approach supports the potential for personalized health management and risk stratification, although external validation in diverse populations is warranted to further confirm its generalizability.