A theory-constrained digital framework of the traditional Chinese medicine Jing-Shen-Mei theory for mammalian lifespan modeling

Objective To construct a theory-constrained mathematical framework for the traditional Chinese medicine (TCM) Jing (精, essence)-Shen (神, spirit)-Mei (寐, sleep) theory and to evaluate the applicability of a pre-specified rational closure to cross-species mammalian lifespan prediction.

Methods Jing, Shen, and Mei were operationalized as a finite life-budget reservoir, effective homeostatic capacity, and sleep-mediated daily burden reduction, respectively. A harmonized complete-case dataset was constructed by integrating maximum lifespan, sleep duration, and cortical/pallial neuronal abundance from three sources: the AnAge database within the Human Ageing Genomic Resources (HAGR), the Boston University mammalian sleep records, and curated comparative neuroanatomical compilations. The primary dataset included 53 species, comprising 48 directly measured or curated neuronal records and 5 explicitly labelled records derived from brain-mass-estimated. The rational exponents (α = 2/5, β = 1/4, and η = 1/5) were pre-specified prior to any empirical fitting, leaving only the global coefficient C to be estimated within training folds. Model performance was evaluated using repeated five-fold cross-validation with 100 repeats, alongside sensitivity analyses on the 48-species direct-measurement subset and leave-one-out cross-validation (LOOCV). Comparators included an unconstrained power-law model and representative machine learning (ML) regressors. Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) served as primary metrics, and a 729-combination parameter-grid analysis was conducted to assess parameter identifiability.

Results On the 53-species curated dataset, the theory-selected rational closure achieved cross-validated RMSE = 13.73 years, MAE = 9.84 years, and R² = 0.660. Compared with representative ML and algebraic benchmark models under identical out-of-fold evaluation, the rational closure showed competitive predictive performance while maintaining a lower-dimensional interpretable structure. Performance remained robust after excluding the five brain-mass-estimated neuronal records: repeated five-fold cross-validation on the 48 directly measured species achieved RMSE = 14.27 years, MAE = 10.14 years, and R² = 0.647, and LOOCV achieved RMSE = 14.16 years, MAE = 10.08 years, and R² = 0.652. The grid-search analysis identified a minimum RMSE of 13.03 years, while the pre-specified rational closure delivered an RMSE of 13.73 years. Notably, 11, 27, and 69 parameter combinations fell within 1%, 5%, and 10% of the minimum, respectively, indicating that the rational closure is near-optimal yet non-unique.

Conclusion This study provides a preliminary, computationally grounded formalization of the TCM Jing-Shen-Mei framework and demonstrates that a pre-specified, one-coefficient rational closure can maintain empirical utility in a sparse-data, cross-species predictive protocol.