Ensemble Learning for Bio-Age: Integrating Epigenetic Clocks and Clinical Biomarker

Abstract

The accurate quantification of biological aging is depen-dent on multimodal integration, as individual biomarker domainscapturedisparateaspectsofphysiologicalde-cay.Whileepigeneticclocks(e.g.,DNAmethylationar-rays)provided epcellularresolution,theylackthesys-temic, real-time metabolic responsiveness characteristic ofclinicalbloodphenotyping.Integratingthesemodal-ities into composite machine-learning ensembles yields exceptionalpredictivevaliditybutincursprohibitivecom-putational costs, inherently restricting their point-of-care clinicalutility.Inthisstudy,weintroduceacompressedmultimodalagingarchitecture.Byoriginallystacking353genomicmethylationtargetswith9systemichematologi-calmarkers,ourfoundationalensembleachievedaMean Absolute Error(MAE) of 2.67years (R2= 0.949).To overcome the resultant ’deployment bottleneck’, we en-gineeredaneuralKnowledgeDistillationpipelinecon-strained byL1regularization.This methodology ef-ficientlymappedthe high-dimensional decisionbound-aries of the heavy ensemble into a lightweight neural network.The distilled architecture successfully shed 97.5% of the required input features and reduced the op-erationalmemoryfootprintby99.89%(34.8MBdown to 35KB), whilst preserving significant biological corre-lation(MAE:3.98years).Ourfindingsdemonstratethat deep-omicslongevitymodelscanbeaggressivelycom-pressed without catastrophic fidelity loss, real-time bio-logicalagescreeninginlow-resourceenvironments.

Introduction

This study presents a multi-modal AI-based Biological Age Prediction System that combines DNA methylation (epigenetic) data and clinical blood biomarkers to estimate biological age more accurately than traditional methods. While epigenetic clocks provide detailed genomic aging information and phenotypic clocks capture real-time physiological health, each has limitations when used independently. To leverage the strengths of both, the researchers developed a large ensemble “Teacher” model that integrates blood and DNA features. However, such models are computationally expensive and unsuitable for mobile or point-of-care healthcare applications.

To overcome this challenge, the study applies Knowledge Distillation (KD) and LASSO-based feature selection. The Teacher model, built from 359 features, transfers its learned aging patterns to a lightweight neural network called the Student model. LASSO regularization removes more than 95% of irrelevant variables, reducing the input space to only 9 key biological features while preserving essential aging information. The Student network learns to mimic the Teacher’s predictions rather than directly predicting chronological age, enabling efficient and accurate biological age estimation with minimal computational requirements.

Experimental results demonstrate that the Teacher model achieved the highest performance (MAE = 2.67 years, R² = 0.949), while the compressed Student model maintained strong predictive capability (MAE = 3.98 years, R² = 0.884) despite reducing model size from 34.82 MB to only 0.035 MB, representing a 99.89% reduction in memory footprint. Correlation analysis showed strong agreement between Teacher and Student predictions, indicating that the compressed model successfully preserved the underlying biological aging patterns.

The explainability analysis identified important aging biomarkers, including C-Reactive Protein (CRP), immune-cell-related indicators, and key CpG methylation sites from established epigenetic clocks. Based on these findings, the authors propose the Biological Information Density Theory, suggesting that human aging may be governed by a low-dimensional biological manifold rather than requiring hundreds of parameters. Although promising, the study acknowledges limitations due to reliance on cross-sectional datasets and emphasizes the need for longitudinal clinical validation. Overall, the proposed framework demonstrates that highly accurate biological age prediction can be achieved using compact, explainable, and deployment-friendly AI models suitable for future digital health applications.

Conclusion

Thisframeworkprovesthatultra-denseomicsanaly-ses do not have to remain computationally stranded in high-performance labs.By executing topological trans-ferthroughsoft-labelKnowledgeDistillation,wemathe-maticallyforcedthe”ManifoldCollapse”ofmulti-modal epigenetic parameters into a radically lightweight neural architecture.This 35 KB distilled network theoretically supportstheexistenceofanultra-densebiologicalmaster clock, and practically enables rapid, real-time biological agecalculationsdeployablenativelyonedge-deviceclin-ical health applications.

References

[1] S.Horvath,“DNAmethylationageofhumantissuesandcelltypes,”GenomeBiology,vol.14,no.10,p.R115,2013.https://doi.org/10.1186/gb-2013-14-10-r1153 [2] G. Hannum et al., “Genome-wide methylation pro-files reveal quantitative views of human aging rates,”MolecularCell,vol.49,no.2,pp.359–367,2013.https://doi.org/10.1016/j.molcel.2012.10.0163 [3] M. E. Levine et al., “An epigenetic biomarker of aging for lifespan and healthspan,” Aging (Albany NY), vol. 10,no. 4,pp. 573–591,2018.https://doi.org/10.18632/aging.1014143 [4] A.T.Luetal.,“DNAmethylationGrimAgestrongly predicts lifespan and healthspan,” Aging (Albany NY), vol. 11,no. 2,pp. 303–327,2019.https://doi.org/10.18632/aging.101684 [5] C. G. Bell et al., “DNA methylation aging clocks: challenges and recommendations,” Genome Biol-ogy,vol.20,no.1,p.249,2019.https://doi.org/10.1186/s13059-019-1824-y [6] P. Klemera and S. Doubal, “A new approach to the conceptandcomputationofbiologicalage,”Mecha-nisms of Ageing and Development, vol. 127, no. 3,pp.240–248,2006. https://doi.org/10.1016/j.mad.2005.10.004 [7] E. Putin et al., “Deep biomarkers of human aging: application of deep neural networks to biomarker development,”Aging(AlbanyNY),vol.8,no.5,pp.1021–1033,2016.https://doi.org/10.18632/aging.100968 [8] G.Hinton,O.Vinyals,andJ.Dean,“Distill-ing the knowledge in a neural network,” arXiv preprint arXiv:1503.02531, 2015.https://arxiv.org/abs/1503.02531 [9] C. Bucilua, R. Caruana, and A. Niculescu-Mizil, “Modelcompression,”inProc.12thACMSIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2006, pp. 535–541.https://doi.org/10.1145/1150402.1150464 [10] J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation:A survey,”Int. JournalofComputerVision,vol.129,no.6,pp.1789–1819,2021. https://doi.org/10.1007/s11263-021-01453-z [11] R.Tibshirani,“Regressionshrinkageandselec-tionviathelasso,”JournaloftheRoyalStatisti-calSociety:SeriesB,vol.58,no.1,pp.267–288, 1996. https://doi.org/10.1111/j.2517-6161.1996.tb02080.x [12] J.Friedman,T.Hastie,andR.Tibshirani,“Regular-ization paths for generalized linear models via co-ordinate descent,” Journal of Statistical Software, vol. 33, no. 1, pp. 1–22, 2010.https://doi.org/10.18637/jss.v033.i01 [13] D.H.Wolpert,“Stackedgeneralization,”Neu-ralNetworks,vol.5,no.2,pp.241–259,1992. https://doi.org/10.1016/S0893-6080(05)80023-1 [14] L.Breiman,“Stackedregressions,”MachineLearn-ing,vol.24,no.1,pp.49–64,1996.https://doi.org/10.1007/BF00117832 [15] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001.https://doi.org/10.1023/A:1010933404324 [16] T.ChenandC.Guestrin,“XGBoost:Ascalabletree boostingsystem,”inProc.22ndACMSIGKDDInt.Conf.KnowledgeDiscoveryandDataMining,2016,pp.785–794.https://doi.org/10.1145/2939672.2939785 [17] G.Hannumetal., “GSE40279: Genome-wide DNA methylation profiles of whole blood,”NCBIGeneExpressionOmnibus,2013. https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE40279 [18] CentersforDiseaseControlandPrevention(CDC), “National Health and Nutrition Examination Sur-vey (NHANES),” National Center for Health Statistics. https://www.cdc.gov/nchs/nhanes/index.htm [19] E. J. Topol, “High-performance medicine:the convergence of human and artificial intelligence,” NatureMedicine,vol.25,no.1,pp.44–56,2019. https://doi.org/10.1038/s41591-018-0300-7 [20] N. Rieke et al., “The future of digital health with federatedlearning,”npjDigitalMedicine,vol.3,p.119,2020.https://doi.org/10.1038/s41746-020-00323-1

Copyright

Copyright © 2026 Hema M S, Adarsh Narain Shukla, Darshana Sethia, Arshad Ali Athani, Angshuman B.. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.