163、PL-NUT:从营养—代谢状态到肌肉功能结局——元律医学多维计算与可证伪预测(PL-NUT: From Nutritional–Metabolic States to Muscle Function Outcomes—Multidimensional Computation and Falsifiable Prediction in Prim-Lex Medicine)

The clinical predicament in nutritional assessment is not a shortage of indicators but a conjunction of construct misalignment and integration deficit. The GLIM criteria demonstrate acceptable diagnostic specificity in hospitalized older adults (pooled specificity 0.88, 95% CI 0.81–0.93), and GLIM-defined malnutrition is associated with higher in-hospital and post-discharge mortality, reduced activities of daily living, and prolonged length of stay; however, validation evidence in community-dwelling and nursing-home populations remains insufficient, and the choice of first-step screening tool systematically alters GLIM-based sensitivity and agreement. Serum albumin, historically treated as a nutrition marker, has been explicitly characterized by an ASPEN position paper as an inflammation- and disease-severity-related indicator rather than a valid proxy for total body protein or muscle mass. Meanwhile, the dimension of nutritional status most directly linked to functional outcomes—muscle function—typically appears in routine nutritional assessment only as a coarse “sarcopenia yes/no” classification, lacking a predictive structure that takes continuous functional change as the outcome, routinely available variables as inputs, and improvement in prediction error (rather than statistical significance alone) as the criterion. Prim-Lex Medicine posits the Animal-Life Prim-Fire—the NADH/NAD⁺ redox couple and glycolysis–oxidative phosphorylation coupling—as its theoretical anchor, but its direct measurements (NAD(P)H FLIM and , lactate/pyruvate ratio, mitochondrial oxygen consumption) are inaccessible in general hospitals due to equipment, reagent standardization, and cross-centre comparability barriers. The methodological shift of the PL-NUT framework is explicit: rather than “measuring the Prim-Fire” in general hospitals, it tests whether the Prim-Fire level has left reproducible week-12 handgrip-strength predictive traces—independent of conventional nutritional variables—in routinely available metabolic and inflammatory variables. This shift does not deny the fundamental role of NADH/NAD⁺ in skeletal muscle metabolism, but it refuses the untested leap from “mechanistic importance” to “routine proxies must carry clinically detectable incremental information”. The detection layer comprises six groups of routine information: MNA-SF and GLIM components, anthropometry and handgrip strength, fasting plasma glucose (FPG) and HbA1c, SIRI derived from complete blood count, hs-CRP, and standard clinical data. The computation layer compresses the Five-Element structure into two fixed equal-weight modules: Fire () , an equal-weight aggregation of standardized FPG and HbA1c, positioned as a joint candidate proxy for host glucose-metabolic burden; and Wood () , an equal-weight aggregation of standardized and , positioned as a joint candidate proxy for systemic inflammatory burden. The interaction is the sole coupling structure entering formal testing. All standardization parameters are estimated in the training set and frozen; the validation set applies the training-set parameters without re-estimation. Four core methodological choices define this protocol. First, the primary outcome is frozen as week-12 handgrip strength (kg) , with baseline handgrip as a mandatory covariate in the baseline model and change score as sensitivity analysis. Handgrip strength is chosen over GLIM diagnosis to avoid circularity and information leakage inherent in using GLIM’s own components (weight, muscle mass, inflammation) to predict the GLIM label; handgrip provides a functional outcome that does not fully overlap with GLIM rules yet holds independent standing in clinical nutritional functional assessment. Second, a four-tier model comparison hierarchy is established: (conventional clinical baseline: baseline handgrip, age, sex, MNA-SF; slope degrees of freedom ), (; ), (; ), and (baseline + four standardized raw variables: FPG, HbA1c, , ; ). The maximum candidate model is (), not () ; sample size planning and model complexity auditing must use as the benchmark, because overfitting risk is driven by the most parameterized model. The comparison of versus is the primary test of equal-weight aggregation parsimony; the comparison of versus is the primary test of interaction increment. Third, the primary comparison quantity for H1 is frozen as on the validation set (a positive value indicates that has lower prediction error than , i.e., the extended model improves). MAE, predictive (allowed to be negative, not truncated), and continuous-outcome calibration (intercept and slope ) serve as supporting evidence. Probability-scale DCA is not applied to the continuous primary outcome; DCA is reserved for the binary secondary outcome and only under a pre-specified, guideline-concordant clinical decision scenario. Fourth, sample size follows Riley et al. (2019) four criteria for continuous outcomes, with maximum candidate slope degrees of freedom , target shrinkage , estimation error of residual standard deviation and outcome mean each not exceeding 10% of the true value, taking the maximum of the four criteria, and dividing by the expected proportion of valid week-12 handgrip data to obtain the planned recruitment number. The framework explicitly states that FPG, HbA1c, SIRI, and hs-CRP are not direct measures of NADH/NAD⁺ or Prim-Fire coherence; that a statistically significant Fire–Wood interaction does not constitute proof of Five-Element generative or restrictive biological mechanisms; and that PL-NUT scores must not replace GLIM diagnosis, refeeding-syndrome risk screening, nutrition-support indication judgment, or clinical decision-making. Death, disability, and loss to follow-up are treated as three mechanistically distinct forms of outcome unavailability and must not be uniformly imputed by MICE. Its anticipated contribution is strictly limited: a minimum viable clinical validation protocol executable in geriatric, rehabilitation, or nutrition clinics, with actually collectable data and honestly falsifiable hypotheses. No patient data have yet been obtained; this is a clinical prediction research protocol, and it does not claim that PL-NUT has established incremental predictive value, nor does it claim that any simulation output constitutes clinical evidence.