Aging, a biological process leading to physical and physiological decline, culminates in organ failure and death. The rate and manner of aging vary significantly between individuals, even within inbred populations. The concept of biological age, aiming to better reflect an individual's aging process than chronological age, is gaining traction. Developing accurate biological measures predictive of health and longevity would significantly accelerate research into genetic and pharmacological interventions targeting aging and age-related diseases. A useful biological age metric should correlate with chronological age and serve as a better predictor of remaining lifespan and age-associated outcomes than chronological age alone, especially at ages where most of the population is still alive. Importantly, measurement should be non-invasive, allowing repeated assessments without impacting the subject's health or lifespan. While several biometrics and biomarkers meet some of these criteria in humans (e.g., grip strength, gait, immune system measures, telomere length, advanced glycosylation end products, cellular senescence levels, and DNA methylation clocks), DNA methylation clocks in mice are currently expensive, time-consuming, and invasive. Frailty indices (FIs) in humans strongly predict mortality and morbidity, even outperforming DNA methylation clocks. FIs quantify the accumulation of health deficits, including lab results, symptoms, diseases, and activities of daily living. A higher FI score indicates greater frailty. The FI has been successfully adapted for mice, incorporating 31 non-invasive items across various systems. The mouse FI is strongly linked to chronological age, correlates with mortality and age-related outcomes, and is sensitive to lifespan-altering interventions. However, its potential to model biological age or predict individual life expectancy in mice remained unexplored before this study.

Existing literature demonstrates a strong correlation between frailty indices and mortality in humans, surpassing other biological age indicators like DNA methylation clocks. The mouse frailty index, a reverse-translated version of the human index, shows promise in predicting age-related outcomes and responding to lifespan-extending interventions. However, the ability to predict individual lifespan in mice using the frailty index was previously unknown. While DNA methylation clocks show promise as biomarkers of biological age in both humans and mice, their expense, invasiveness, and time-consuming nature present significant limitations. Other studies have attempted to predict mouse lifespan using various physiological and behavioral parameters, but these often yielded low to moderate correlations with survival or focused on specific physiological conditions or short-lived strains. This study aimed to address the gap in knowledge by using longitudinal data and machine learning to assess the predictive power of the mouse frailty index for lifespan prediction.

This study longitudinally tracked frailty in a cohort of 60 naturally aging male C57BL/6Nia mice from 21 months of age until death. Frailty Index (FI) scores were measured approximately every 6 weeks using a modified version of the existing mouse clinical FI, excluding malocclusions and body temperature. The assessment involved scoring each of the 29 items on a 0, 0.5, or 1 scale, representing the degree of deficit observed. Additional variables were incorporated to represent body weight change (total percent weight change from 21 months, recent percent weight change from the previous month, and a threshold recent weight change score based on weight fluctuation). Machine learning algorithms were employed to build two predictive models: FRIGHT age and the AFRAID clock. For FRIGHT age, four multivariate linear regression models (simple least squares, elastic net, random forest, and Klemera-Doubal) were compared to a univariate regression using FI score alone. Model selection was based on the lowest median error, as determined through bootstrapping on the training dataset. The random forest regression model was chosen due to its ability to handle complex interactions between variables. FRIGHT age, the output of this model, was then evaluated against chronological age using the testing dataset. For the AFRAID clock, a similar approach was used, with the model trained to predict remaining lifespan (in days). Again, four multivariate linear regression models were compared to a univariate model using FI score. The random forest regression model was selected based on bootstrapping results. The AFRAID clock's accuracy in predicting lifespan was assessed using the testing dataset. The predictive power of both models was evaluated using various metrics, including median error, mean error, r-squared values, and p-values. To determine the models' responsiveness to interventions known to extend healthspan or lifespan, data from two previously published studies were reanalyzed: one involving enalapril treatment and another involving dietary methionine restriction. The FI, FRIGHT age, and AFRAID clock scores were calculated for the intervention and control groups in these studies, and differences were assessed using appropriate statistical tests.

Frailty scores were found to significantly correlate with and predict chronological age in mice (Figure 1). A univariate regression model using the FI score as a single predictor showed a median error of 1.8 months in predicting chronological age (r² = 0.642, p ≤ 3.4e-38). The error in the prediction was interpreted as possibly representing biological age, with mice exhibiting healthier profiles showing predicted ages younger than their actual ages. However, the correlation between this age difference (delta age) and survival was not uniform across all age groups. The strongest correlation between delta age and survival was observed in older mice (Table 1). To improve the prediction of chronological age, multivariate regression models were developed incorporating individual FI items. The random forest regression model, termed FRIGHT age, significantly outperformed the univariate model, exhibiting a median error of 1.3 months and an r² of 0.748 (p=1.1e-50) (Figure 3). The FRIGHT age model's most important components were breathing rate, tail stiffening, kyphosis, and total weight change. While more accurate in predicting chronological age, FRIGHT age exhibited weak correlation with mortality at most ages, suggesting that the factors correlating with apparent age may differ from those driving mortality (Table 1). To predict life expectancy more accurately, a second random forest model, the AFRAID clock, was developed using individual FI items and chronological age. The AFRAID clock significantly outperformed a univariate model based on FI score alone, achieving a correlation (r) of 0.505 with survival, resulting in a median error of 1.7 months (p=1.1e-26) (Figure 4). The most important variables included total weight loss, chronological age, and tremor. Unlike FRIGHT age, AFRAID clock showed a stronger correlation with mortality at specific ages (Table 1), demonstrating its potential in predicting lifespan, even up to a year in advance (Figure 4). The impact of interventions on both models was assessed using data from enalapril-treated and methionine-restricted mice. Enalapril treatment, known to enhance healthspan but not lifespan, reduced FRIGHT age but did not significantly affect the AFRAID clock score (Figure 5). Methionine restriction, a known lifespan-extending intervention, resulted in reduced FRIGHT age and increased AFRAID clock scores, further validating the models' sensitivity to lifespan-extending interventions (Figure 5).

This study's results demonstrate the successful application of machine learning to develop two novel clocks based on the mouse frailty index. The FRIGHT age model provides an accurate estimation of chronological age, while the AFRAID clock effectively predicts life expectancy. The non-invasive nature of the frailty index is an important advantage. The finding that the AFRAID clock demonstrates sensitivity to healthspan and lifespan-extending interventions highlights its potential for use as an early biomarker in interventional studies. The capacity to predict lifespan, even up to a year in advance, could significantly accelerate the evaluation of potential longevity interventions by allowing researchers to make informed decisions at earlier time points. The ease of assessment of frailty and the use of these clocks could help improve research efficiency in longevity studies.

This study successfully developed two novel machine learning-based clocks, FRIGHT age and the AFRAID clock, to estimate chronological age and predict life expectancy, respectively, using a mouse frailty index. These clocks were shown to be sensitive to interventions known to extend healthspan and lifespan, and their non-invasive nature makes them valuable tools for future longevity research. Future research should focus on validating these models in larger cohorts, including both male and female mice, and potentially incorporating other biomarkers and longitudinal data analysis techniques to further enhance predictive accuracy and broaden the application of these tools.

The study used a relatively small cohort of male mice. Future studies should include larger, more diverse cohorts, including female mice, to enhance the generalizability of the findings. The current model is a standard fixed-time predictive model, which treats each time point independently. Future work could explore dynamic prediction models that account for longitudinal correlations within individual mice's frailty trajectories. The inclusion of additional, non-invasive biomarkers may further enhance the predictive power of the models.