Novel biological age model using explainable automated CT-based cardiometabolic biomarkers for phenotypic prediction of longevity

Perry J. Pickhardt, Michael W. Kattan, Matthew Lee, B. Dustin Pooler, Ayis Pyrros, Daniel Liu, Ryan Zea, Ronald M. Summers, John W. GarrettResearch Square 2024 · DOI

Abstract

We derived and tested a CT-based biological age (CTBA) model for predicting longevity, using an automated pipeline of explainable deep learning AI algorithms that quantify skeletal muscle, abdominal fat, aortic calcification, bone density, and solid abdominal organs. These AI tools were applied to abdominal CT scans from 123,281 adults (mean age, 53.6 years; 47% women; median clinical follow-up, 5.3 years). Final weighted CT biomarker selection was based on the Index of Prediction Accuracy (IPA). The CTBA model significantly outperformed standard demographic data for predicting longevity (IPA = 29.2 vs. 21.7; 10-year AUC = 0.880 vs. 0.779; p < 0.001), despite having no knowledge of the latter. The age- and sex-corrected survival hazard ratio (HR) for the highest- vs. lowest-risk CTBA quartile was 8.73 (95% CI, 8.14–9.36). Muscle density, aortic plaque burden, visceral fat density, and bone density contributed most to the model. Unlike (epi)genetic and metabolomic approaches, this personalized phenotypic CTBA model can be opportunistically derived from routine scans, regardless of clinical indication, to better inform risk assessment.

Introduction

Chronological age is often a fallible measure of an individual's true health status. Two 60-year-olds can have vastly different "biological ages" based on genetics, lifestyle, and disease progression. While the field of "geroscience" has explored cellular biomarkers like epigenetic clocks or telomere length, these tests are often expensive and not integrated into routine clinical care.

Medical imaging, specifically Computed Tomography (CT), contains a wealth of phenotypic data about a patient’s tissues and organs. However, this information is largely ignored unless it relates to the specific clinical indication for the scan (e.g., appendicitis). This study addresses the challenge of transforming these "incidental" imaging findings into a standardized, automated, and explainable metric for predicting a patient's remaining lifespan.

Methods

AI Pipeline and Biomarker Extraction

The researchers developed a fully automated AI pipeline using deep learning models (CNNs, U-Nets, and CycleGANs) to localize vertebral levels (T12–L4) and segment various tissues. Unlike "black box" radiomics that use abstract mathematical features, this model relies on human-understandable, explainable biomarkers: skeletal muscle density and area, visceral and subcutaneous fat, aortic calcium (Agatston score), bone mineral density, and organ volumes.

Statistical Modeling and Feature Selection

Survival was modeled using Cox proportional hazards regression. Feature selection was driven by the Index of Prediction Accuracy (IPA); biomarkers were only retained if their removal caused a significant "IPA drop" in the model’s performance. Crucially, the CTBA model was blinded to the patient’s actual chronological age, sex, and race during its derivation to ensure the signal was purely phenotypic.

Contribution of individual CT biomarkers to the model's predictive accuracy, measured by the drop in Index of Prediction Accuracy (IPA).

Datasets

The study utilized two massive longitudinal datasets: a primary cohort of 123,281 adults from the University of Wisconsin and an external validation cohort of 40,718 adults from Duly Health and Care (Illinois). Both cohorts were predominantly White (~90%), representing a limitation in ethnic and geographic diversity.

Key Results

The CTBA model demonstrated exceptional predictive power, achieving an IPA of 29.2, which was significantly higher than the 21.7 achieved by standard demographic measures (p < 0.001). In ROC analysis for 10-year survival, the CTBA model reached an AUC of 0.880, compared to 0.779 for the demographic model. Furthermore, patients in the highest-risk CTBA quartile were 8.73 times more likely to die than those in the lowest-risk quartile, even after correcting for chronological age and sex.

Predictive performance (AUC) and hazard ratios (HR) for the CTBA model across various cohorts and age groups.

Discussion

Quality Over Quantity

A key insight from the model is that muscle density (a measure of fat infiltration or myosteatosis) was a far more dominant predictor of survival than muscle area. This suggests that the metabolic health and "fitness" of the muscle matters more for longevity than its absolute size.

Opportunistic Screening and Explainability

Because millions of abdominal CTs are performed annually for various reasons, this model allows clinicians to generate a "health snapshot" without requiring new tests or additional radiation exposure. By using discrete, measurable biomarkers (like Hounsfield units for bone density), the results are transparent and actionable. A physician can see exactly why a patient has a high biological age, making the AI's output more useful in a clinical setting.

Limitations and Future Directions

The study populations were ~90% White and Midwestern, necessitating future multicenter trials with broader racial and ethnic participation. Additionally, the model currently focuses on cardiometabolic and senescent factors and does not account for other major causes of death like cancer, trauma, or dementia. Future iterations may also incorporate social determinants of health, such as the Area Deprivation Index (ADI).

Summary updated

Abstract

<title>Abstract</title> We derived and tested a CT-based biological age (CTBA) model for predicting longevity, using an automated pipeline of explainable deep learning AI algorithms that quantify skeletal muscle, abdominal fat, aortic calcification, bone density, and solid abdominal organs. These AI tool were applied to abdominal CT scans from 123,281 adults (mean age, 53.6 years; 47% women; median clinical follow-up, 5.3 years). Final weighted CT biomarker selection was based on index of prediction accuracy (IPA). The CTBA model significantly outperformed standard demographic data for predicting longevity (IPA = 29.2 vs. 21.7; 10-year AUC = 0.880 vs. 0.779; p &lt; 0.001), despite any knowledge of the latter. Age- and sex-corrected survival hazard ratio (HR) for the highest-vs-lowest risk CTBA quartile was 8.73 (95% CI,8.14–9.36). Muscle density, aortic plaque burden, visceral fat density, and bone density contributed most. Unlike (epi)genetic and metabolomic approaches, this personalized phenotypic CTBA model can be opportunistically-derived, regardless of clinical indication, to better inform risk assessment.
BibTeX
@inproceedings{pickhardt2024novel,
  author = {Perry J. Pickhardt and Michael W. Kattan and Matthew Lee and B. Dustin Pooler and Ayis Pyrros and Daniel Liu and Ryan Zea and Ronald M. Summers and John W. Garrett},
  title = {{Novel biological age model using explainable automated CT-based cardiometabolic biomarkers for phenotypic prediction of longevity}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-4707454/v1},
}
Figure