Age and life expectancy clocks based on machine learning analysis of mouse frailty

Aging manifests as progressive accumulation of physiological deficits leading to morbidity and death, with substantial inter-individual variability even within inbred animal populations. Biological age aims to capture functional aging better than chronological age and should correlate with age, predict remaining longevity and health outcomes, and be non-invasive for repeated measures. Existing human biomarkers (e.g., grip strength, immune metrics, telomere length, glycation end-products, cellular senescence, DNA methylation clocks) have limitations for preclinical mouse studies due to cost, invasiveness, or practicality. The mouse clinical Frailty Index (FI), a non-invasive composite of 31 health-related items, correlates with chronological age and mortality and responds to lifespan-altering interventions, but its capacity to model biological age or predict individual life expectancy had not been established. This study longitudinally scores frailty in aging male mice and applies machine learning to develop FRIGHT age to model apparent chronological age and the AFRAID clock to predict life expectancy, then tests these clocks in intervention cohorts.

Frailty indices in humans predict morbidity and mortality and often outperform other biological aging measures, including DNA methylation clocks. The mouse FI, reverse-translated from human assessments, correlates with chronological age and mortality and is sensitive to interventions. DNA methylation clocks correlate with age in both humans and mice and can reflect intervention effects but are invasive and costly, and their association with mouse mortality has yet to be fully explored. Prior mouse mortality prediction efforts focused on acute settings, limited biomarker panels, or short-lived strains, yielding modest predictive performance. Machine learning approaches, including random forests, elastic net, and deep learning, have improved aging biomarker development in humans; however, comprehensive non-invasive mouse predictors for individual life expectancy had not been demonstrated.

Animals and longitudinal assessment: Sixty male C57BL/6Nia mice were obtained aged 19–21 months and housed under standard conditions. Frailty was assessed approximately every 6 weeks from 21 months of age until natural death (euthanasia only when moribund). A control subset received AAV-GFP at 21 months for an unrelated study; lifespan and frailty were unaffected compared to untreated mice. FI scoring followed the modified mouse clinical FI with 29 non-invasive items scored 0/0.5/1. Additional predictors encoded weight dynamics: total percent weight change from 21 months (twc), recent percent weight change over ~1 month (rwc), and a threshold recent weight change indicator. Data handling and modeling: Each assessment was treated as independent. Data were randomly split 50:50 into training (n=106 assessments across 30 mice) and testing (n=165 across 30 mice) sets, stratified by mouse to avoid leakage. Missing item values (18 of 7859) were imputed with age-group medians. Modeling was performed in Python (scikit-learn), with model selection via bootstrapping (100 resamples) on the training set, comparing median/mean absolute errors, r², and p values using one-way ANOVA with Dunnett’s post hoc tests. Age prediction (FRIGHT age): Predictors were FI items significantly correlated with chronological age (p<0.05; 21 items). Four approaches were compared: simple least squares, elastic net, Klemera-Doubal method (KDM), and random forest regression (RFR). The RFR (1000 trees; min_samples_leaf=9) was selected for lowest median error and strong performance. FRIGHT age residuals (delta age = predicted − true age) were evaluated against survival at specific ages. Lifespan prediction (AFRAID clock): Predictors included all FI items plus chronological age (32 variables). Least squares, elastic net, and RFR were compared; RFR (1000 trees; min_samples_leaf=6) was selected for superior performance. Feature importance was computed by permutation importance. Model performance was evaluated on the held-out test set. Survival associations at specific chronological ages were analyzed, and Kaplan–Meier curves compared top vs bottom AFRAID quartiles via log-rank tests. Interventions: External enalapril dataset (male C57BL/6, 16–23 months on enalapril 30 mg/kg/day vs control) was reanalyzed to compute FI, FRIGHT, and AFRAID at 23 months. A methionine restriction (MetR) cohort (0.1% Met vs 0.45% control from 21 to 27 months) was assessed at 27 months. Group comparisons used two-sided t-tests. Statistical significance threshold was p<0.05.

- FI correlates with age and modestly predicts chronological age: linear regression yielded median error 1.8 months, mean error 1.9 months, r²=0.642 (p≤3.4e−38). Delta age (predicted−true) correlated with survival at some ages (e.g., 24 months r²=0.390, p=0.023; 34.5 months r²=0.686, p=0.021; 36 months r²=0.881, p=0.018). - Individual FI items varied in age-correlation; strongest were tail stiffening (r²=0.58), breathing rate/depth (0.50), gait disorders (0.42), hearing loss (0.38), kyphosis (0.38), tremor (0.38), all p<0.001. - FRIGHT age (random forest using 21 age-correlated items) improved age prediction: test r²=0.748 (p=1.1e−50), median error 1.3 months, mean error 1.6 months. Top contributors included breathing rate, tail stiffening, kyphosis, and total weight change. FRIGHT delta age generally did not predict survival at most ages (r²<0.1 for most age groups), with stronger associations only in the oldest mice (e.g., 34.5 months r²=0.661, p=0.026). - AFRAID clock (random forest using all FI items plus age) predicted life expectancy: test correlation with survival r=0.505 (p=1.1e−26), median error 1.7 months (~53 days), mean error 2.3 months. Important predictors included total weight loss, chronological age, tremor, distended abdomen, recent weight loss, and menace reflex. AFRAID was significantly associated with survival within age strata (e.g., 24 months r²=0.447, p=0.012; 30 months r²=0.363, p=0.004; 34.5 months r²=0.653, p=0.028). Kaplan–Meier analyses showed significant separation between top and bottom AFRAID quartiles at 24–26, 27–29, 30–32, and 33–35 months (log-rank p=0.032, 0.015, 0.026, 0.034). - Interventions: Enalapril at 23 months reduced FI and FRIGHT age by ~1 month (control 27.8±1.1 vs enalapril 26.8±1.4 months, p=0.046) but did not increase AFRAID-predicted survival (control 5.9±0.7 vs enalapril 6.2±0.9 months, p=0.29), consistent with improved healthspan without lifespan extension. Methionine restriction at 27 months reduced FI (0.37±0.03 control vs 0.30±0.04 MetR, p=0.0009), reduced FRIGHT age by 0.7 months (29.8±0.9 vs 29.1±0.6, p=0.039), and increased AFRAID-predicted survival by 1.3 months (3.0±1.0 vs 4.3±1.0, p=0.006). - Life expectancy can also be modeled from FI items alone by replacing chronological age with FRIGHT age in the AFRAID framework, with similar accuracy (Supplementary analyses).

Longitudinal frailty assessments combined with machine learning yielded two practical, non-invasive clocks. FRIGHT age closely models apparent chronological age from FI components, whereas AFRAID directly predicts remaining lifespan, including within age groups where chronological age is held constant. This addresses the need for early, repeatable biomarkers that capture biological aging and prospective life expectancy in preclinical models. The clocks differentiated healthspan versus lifespan effects in interventions: enalapril improved frailty/FRIGHT without extending predicted lifespan, while methionine restriction improved both. Compared with invasive and costly biomarkers such as DNA methylation clocks, frailty-derived models are rapid, inexpensive, and scalable, enabling interim decision-making in longevity studies and potentially accelerating discovery. Translationally, similar approaches could be adapted to human FI datasets, pending availability of large longitudinal cohorts with mortality outcomes. Methodologically, random forests offered superior performance by modeling non-linearities and interactions among frailty items, though with reduced interpretability. Together, FI, FRIGHT, and AFRAID provide complementary perspectives on biological age (appearance vs prognosis), supporting their combined use in aging research.

This study establishes that non-invasive frailty assessments can be leveraged with machine learning to produce: (1) FRIGHT age, a more accurate estimator of apparent chronological age than FI alone, and (2) the AFRAID clock, a robust predictor of individual life expectancy. Both measures respond to interventions, distinguishing healthspan-only effects (enalapril) from combined healthspan and lifespan extension (methionine restriction). These tools can expedite preclinical aging research by enabling early, repeated, low-cost evaluation of interventions. Future work should: increase sample sizes and age coverage (including younger and older mice), model longitudinal dependence (e.g., joint modeling), incorporate additional minimally invasive biomarkers, validate in female mice and other strains, and develop composite outcomes beyond mortality and age. Adaptation to human FI datasets may allow individualized life expectancy predictions in clinical contexts.

- Study used only male C57BL/6Nia mice; sex differences in frailty and intervention responses may limit generalizability. - Modest sample size (n=60) with fewer assessments at the oldest ages may reduce precision, particularly for age-stratified analyses. - Longitudinal assessments were treated as independent observations; more advanced longitudinal or joint modeling could improve predictions. - Random forest models, while accurate, are less interpretable regarding variable interactions. - Facility-specific baseline frailty differences can shift absolute predicted ages, suggesting the need for within-study controls and calibration. - Models primarily used clinical frailty items; adding molecular/physiological biomarkers may enhance early-life prediction. - External intervention data (enalapril) came from a different facility/strain source, potentially introducing cohort effects.