PaperlayerView exact revision on arXiv

Source-linked AI summary

TGAge: A Phenotypic Aging Clock Based on Tabular Foundation Models and the Gompertz Law for Assisted Living Facility Residents

Chengzhuo Li, Rongjia Zhou, Shanshan Wang, Runze Yan, Carl Yang, Xiaobo Sun, Huiwen Xu, Jiaying Lu

arXiv:2610.32994v1stat.APq-bio.QMstat.ML

TL;DR

Assisted living facility residents may lack the epigenetic data used by many aging clocks, and their clinical measurements can be infrequent or incomplete. TGAge estimates risk-equivalent biological age from routine clinical phenotypes using a tabular foundation model and the Gompertz law. It outperforms the compared baselines in survival prediction, and its age acceleration is associated with aging-related comorbidities.

  • Problem

    Existing biological aging clocks are difficult to apply to ALF residents because epigenetic profiling is generally unavailable and clinical measurements may be infrequent or incomplete.

    Source: §I · p. 1

  • Method

    TGAge uses a tabular foundation model to estimate all-cause mortality risk from clinical phenotypes and converts the risk ratio into risk-equivalent biological age through the Gompertz law.

    Source: §I · p. 1, A. Risk-equivalent BA under the… · p. 4, B. Tabular Foundation Model for… · p. 4

  • Results

    TGAge consistently outperforms all baselines in survival prediction, and its age acceleration stratifies risk groups and is associated with aging-related comorbidities.

    Source: §I · p. 2

  • Takeaways & Limitations

    TGAge uses routinely collected EHR variables and is presented as practical for resource-limited long-term-care settings.

    Source: §VI · p. 8

    TGAge has been evaluated only among ALF residents, and its current implementation relies on complete clinical feature availability.

    Source: §VI · p. 8

Abstract

from arXiv · show

Biological age captures overall health status beyond chronological age, yet existing biological aging clocks are difficult to apply to assisted living facility (ALF) residents due to scarce epigenetic data. We propose TGAge, a second-generation phenotypic aging clock that estimates risk-equivalent biological age from routinely collected clinical measurements. TGAge employs a tabular foundation model to predict all-cause mortality risks over discrete time-to-death intervals and converts the resulting mortality-risk ratio into biological age acceleration via the Gompertz law. We develop and evaluate TGAge on a large cohort of ALF residents curated from a proprietary, multi-state long-term care database. TGAge outperforms clinical indices and established biological aging clocks in survival prediction. Moreover, TGAge-derived age acceleration is significantly associated with multiple aging-related comorbidities. These findings support TGAge as a practical, interpretable, and highly predictive biological aging clock for vulnerable ALF populations using only routinely collected clinical measurements.

I. INTRODUCTION

TGAge addresses the difficulty of applying biological aging clocks to assisted living facility residents when epigenetic profiling is generally unavailable and clinical data are less frequent or comprehensive. It estimates risk-equivalent biological age from clinical phenotypes and evaluates the clock in an ALF cohort.

I. INTRODUCTION

  • A non-invasive, accessible, and robust measure of aging pace and health deterioration could be clinically valuable for guiding timely interventions in ALFs.ALFs support people who need help with personal care or daily activities but do not require continuous skilled nursing.Source: §I · p. 1
  • ALF residents’ clinical measurements are often infrequent and incomplete, while epigenetic profiling is generally unavailable because of its cost.These constraints motivate a clock based on routinely collected clinical measurements that can handle irregular sampling.Source: §I · p. 1
  • TGAge uses a tabular foundation model to estimate all-cause mortality risk from clinical phenotypes, then applies the Gompertz law to derive risk-equivalent biological age.The paper frames the output as an interpretable indicator of physiological status and risk of adverse health outcomes.Source: §I · p. 1
  • The study curates an ALF resident cohort of approximately 10,000 subjects from a U.S.-based long-term-care EHR database and evaluates TGAge against clinical indices and established first- and second-generation biological aging clocks.The cohort covers ALFs across all 50 states and 2 U.S. territories.Source: §I · p. 1
  • TGAge provides accurate survival prediction and consistently outperforms all baselines, while its age acceleration stratifies risk groups and is associated with aging-related comorbidities.The study also reports that ablation analyses support the effectiveness of key technical components.Source: §I · p. 2

II. RELATED WORK

Prior phenotypic clocks differ in whether they predict chronological age or future health risks, while tabular foundation models have only begun to be adapted for survival analysis. The paper also describes its EHR data source as a large longitudinal resource spanning long-term and post-acute care settings.

A. Phenotypic Aging Clock Models

  • First-generation phenotypic clocks predict chronological age, whereas second-generation clocks target future health risks; the latter include Gompertz-based risk-equivalent approaches such as Levine’s clock and PCAge.Because chronological age is not a direct measure of health status, first-generation clocks may not fully reflect age-related physiological decline.Source: A. Phenotypic Aging Clock Models · p. 2

B. Tabular Foundation Models for Survival Analysis

  • Tabular foundation models have been adapted to survival analysis through regression, discretized-time classification, and survival prediction heads, but their role in risk-equivalent biological-age estimation remains underexplored.This gap motivates the paper’s use of a tabular foundation model to derive Gompertz-law-based biological age from mortality-risk estimates.Source: B. Tabular Foundation Models for… · p. 2

A. Data Source

  • The proprietary PointClickCare database contains longitudinal EHRs for more than 20 million unique patients across skilled nursing, assisted living, and other long-term and post-acute care settings.Its records include demographics, diagnoses, medications, laboratory results, vital signs, functional and cognitive assessments, and activities of daily living.Source: A. Data Source · pp. 2–3

B. Cohort Construction

The study selects ALF encounters with complete clinical measurements from a heterogeneous care setting and uses a temporal split for model development and evaluation. It defines the encounter-level biological-age task and the clinical features and survival outcomes used in that task.

B. Cohort Construction

  • ALFs provide individualized care with varying practices and documentation standards, while routinely collected clinical and medication data support mortality-risk and biological-age modeling.The source cohort includes 1,236,208 encounters from 1,202,978 residents; requiring all 24 clinical measurements yields 10,516 encounters from 10,467 residents.Source: B. Cohort Construction · p. 3
  • The temporal split uses 2015–2021 encounters for training and 2022–2025 encounters for testing, with a 2015–2021 reference set that is a superset of the training set before the complete-feature filter.Table II is identified as presenting characteristics of the selected study cohort.Source: B. Cohort Construction · p. 3, Table II · p. 3

C. Notation and Problem Setup

  • The BA estimation task uses chronological age, sex, and phenotypic features to estimate a positive biological age for each ALF resident encounter.The task is defined at the encounter level.Source: C. Notation and Problem Setup · p. 3
  • Phenotypic features are tabular clinical measurements, and each survival outcome records follow-up time from admission and whether all-cause death was observed.The paper gives blood pressure and BMI as examples of clinical measurements.Source: C. Notation and Problem Setup · p. 3

IV. METHODS

TGAge converts relative mortality risk into risk-equivalent biological age acceleration using the Gompertz approximation. Sex-specific mortality-rate doubling time translates risk doublings into years.

A. Risk-equivalent BA under the Gompertz Approximation

  • Relative mortality risk versus an age- and sex-matched reference determines risk-equivalent biological age acceleration under the Gompertz approximation.Higher mortality risk corresponds to positive age acceleration.Source: A. Risk-equivalent BA under the… · p. 4
  • The number of mortality-risk doublings is log R_i / log 2, and multiplying it by sex-specific MRDT converts it into years.MRDT denotes mortality-rate doubling time.Source: A. Risk-equivalent BA under the… · p. 4
  • Sex-specific MRDT is estimated with a Cox proportional hazards model using chronological age as its sole predictor.The model is fit on a reference set; its baseline hazard is unspecified and β_CA,sex is estimated by partial likelihood.Source: A. Risk-equivalent BA under the… · p. 4

B. Tabular Foundation Model for Mortality Risk Estimation

TGAge uses TabPFN to estimate phenotypic and reference mortality risks from discrete-time survival outcomes. Cumulative interval probabilities are aggregated for the risk calculation, including the final interval for far-future risk.

B. Tabular Foundation Model for Mortality Risk Estimation

  • TabPFN estimates phenotypic mortality risk and reference mortality risk, whose ratio supplies the relative risk used by TGAge.The phenotypic risk model uses clinical measurements, chronological age, and sex; the reference model uses chronological age and sex.Source: B. Tabular Foundation Model for… · p. 4
  • Follow-up time is partitioned into discrete intervals, making mortality prediction a supervised multiclass classification problem.Subjects are assigned to the interval containing their observed follow-up time; the last interval extends to infinity.Source: B. Tabular Foundation Model for… · p. 4
  • Cumulative summation aggregates interval-specific mortality probabilities into scalar risks for the phenotypic and reference models.The final interval represents far-future mortality risk beyond the typical observation window.Source: B. Tabular Foundation Model for… · p. 4
  • Ablations compare cumulative aggregation with min and max operators to assess the aggregation choice.The stated rationale for cumulative aggregation is evaluated in the ablation study.Source: B. Tabular Foundation Model for… · p. 4

V. EXPERIMENTAL RESULTS

The experimental section frames evaluation around survival prediction, the clinical utility of age acceleration, and ablation-study insights.

V. EXPERIMENTAL RESULTS

  • RQ1 asks how accurately TGAge predicts survival time compared with baseline models.Source: §V · p. 4
  • RQ2 examines the clinical utility of TGAge-derived age acceleration.Source: §V · p. 4
  • RQ3 asks what the ablation studies reveal.Source: §V · p. 4

A. Implementation Details

The implementation uses four follow-up intervals and compares TGAge with clinical indices and biological aging clocks. Evaluation uses sex-stratified survival metrics on the selected cohort's test set.

A. Implementation Details

  • TGAge uses TabPFN-V3 with four intervals: 0–1, 1–2, 2–3, and over 3 years; numeric clinical features are z-score standardized.Predicted interval probabilities are clipped to [10^-6, 1 − 10^-6] before aggregation.Source: A. Implementation Details · p. 5
  • Baselines include chronological age, two clinical indices, first-generation phenotypic clocks, and second-generation PCAge and LinAge.First-generation implementations include linear regression, random forest, support vector regression, XGBoost, TabNN, and ClinicalBERT; PCAge uses 10- or 20-component variants.Source: A. Implementation Details · p. 5
  • Survival metrics are evaluated on the selected cohort's test set using sex-stratified Cox models, C-index, and time-dependent AUC at 1, 2, and 3 years.Second-generation clocks also use the reference set during development; time-dependent AUC applies inverse-probability-of-censoring weighting.Source: A. Implementation Details · p. 5

B. TGAge Accurately Predicts Survival Time (RQ1)

TGAge achieves the strongest overall survival discrimination among compared methods in female and male ALF residents. Its age estimates track chronological age while also distinguishing survival trajectories and greater age-related variation among younger residents.

B. TGAge Accurately Predicts Survival Time (RQ1)

C. TGAge-predicted Age Acceleration Captures Individual Health Risks (RQ2)

TGAge-derived age acceleration is associated with multiple diagnosis-defined comorbidities, including conditions across several physiological systems. These associations persist after adjustment for chronological age.

C. TGAge-predicted Age Acceleration Captures Individual Health Risks (RQ2)

  • Each comorbidity indicator was tested in a separate linear regression of cumulative age acceleration, adjusting for chronological age, sex, and admission date, with Benjamini–Hochberg FDR correction.Source: C. TGAge-predicted Age Acceleration… · p. 6
  • 14 of 108 evaluated condition categories were significantly associated with age acceleration at FDR q < 0.01.The associated conditions span neurofunctional, cardiovascular, pulmonary, infectious or inflammatory, hematologic, and musculoskeletal conditions.Source: C. TGAge-predicted Age Acceleration… · pp. 6–7, Table IV · p. 7
  • Paralysis had the largest adjusted effect among the reported conditions (β = 4.56).Neurofunctional impairment and frailty-related functional loss included several of the largest effect sizes.Source: C. TGAge-predicted Age Acceleration… · p. 6

D. Ablation Analyses (RQ3)

Two ablations examine TGAge’s mortality-risk estimator and its aggregation operator. TabPFN performs best among the tested estimators, while cumulative aggregation performs best among the tested operators and has a 3-year cumulative-mortality interpretation.

D. Ablation Analyses (RQ3)

  • The ablations vary one technical component at a time while keeping the model design and experimental settings otherwise fixed.Source: D. Ablation Analyses (RQ3) · p. 7
  • TabPFN achieves the best observed performance across all metrics among the tested mortality-risk estimators.The comparison includes a linear Cox proportional hazards model, DeepSurv, and Random Survival Forest.Source: D. Ablation Analyses (RQ3) · p. 7
  • Cumulative aggregation achieves the highest observed performance among the three tested operators and estimates cumulative mortality risk at a 3-year horizon.The alternatives are max and min aggregation of the predicted mortality probabilities.Source: D. Ablation Analyses (RQ3) · p. 7

VI. CONCLUSION

TGAge estimates risk-equivalent biological age from routine clinical data by combining a tabular foundation model with a Gompertz-based risk-to-age transformation. It outperforms established clocks and clinical indices, while its evaluation scope and complete-feature requirement remain limitations.

VI. CONCLUSION

  • TGAge combines a tabular foundation model with a Gompertz-based transformation to estimate risk-equivalent age from routinely collected clinical data.Source: §VI · p. 7
  • TGAge’s age acceleration supports risk stratification and is significantly associated with comorbidities, while the clock outperforms established aging clocks and clinical indices.The authors also describe its routine EHR inputs as practical for resource-limited long-term care settings.Source: §VI · pp. 7–8
  • Evaluation has so far been limited to ALF residents, and reliance on complete clinical features may constrain use when measurements are incomplete.Source: §VI · p. 8

Related papers

Loading 2610.32994v1…