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

doi:10.21203/rs.3.rs-4707454/v1

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 < 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.

Health sciences/Biomarkers/Prognostic markers

Health sciences/Health care/Disease prevention/Preventive medicine

Health sciences/Risk factors

Health sciences/Medical research/Biomarkers/Prognostic markers

The aging process reflects the inexorable structural and functional decline of an organism,¹ although the specific mortality risk over time can widely vary according to a host of genetic and environmental modifiers. Historically, chronological age and sex have driven many healthcare decisions regarding prevention, screening, and intervention. However, chronological age represents an incomplete and fallible measure of health status (or “healthspan”) and longevity, and there is growing public awareness that other contributing factors should be considered.^2–4 Biological age (BA) is a potentially useful construct that attempts to reflect the cumulative physiologic effect of lifestyle habits, genetic predisposition, and superimposed disease processes beyond simply the number of years lived. Attempts at deriving an effective BA date back at least half a century,⁵ but with only limited success. Much of the current geroscience focus to date for attempting to derive an effective BA has centered on various “frailomics” at the cellular and subcellular levels, including (epi)genomics (eg, telomere length and epigenetic clock), proteomics, and metabolomics, as well as various other laboratory and clinical measures.^1,6–9

Imaging biomarkers have generally received less attention for estimating BA,^7,8 but arguably may better reflect the cumulative macroscopic effects of aging at the tissue and organ levels. In particular, abdominal computed tomography (CT) represents an appealing candidate for a more personalized investigation. Specifically, CT can provide an objective, understandable, and reproducible assessment of internal tissue composition, including quantitative measures of skeletal muscle, abdominal fat, vascular calcification, bone density, and other organs.^10,11 When combined, these CT-based cardiometabolic biomarkers may better reflect the phenotypical characteristics that result from the interaction of one’s genotype with environmental factors and lifestyle. In particular, CT can reveal findings of silent, pre-symptomatic disease processes such as osteoporosis, atherosclerosis, sarcopenia, and metabolic syndrome, potentially allowing for earlier preventive action.^12–17 Consequently, these CT-based measures have been shown to correlate with aging and survival.¹³ Furthermore, the once arduous task of manually deriving these CT biomarkers has been replaced by “explainable” artificial intelligence (AI) algorithms that are rapid, understandable, and indefatigable.¹⁸ Unlike other frailomic approaches, these CT-based cardiometabolic biomarkers can be derived retrospectively (or prospectively), regardless of the clinical indication, including scans performed many years earlier to allow for both “snapshots in time” and for built-in longitudinal follow-up for survival analysis.¹⁰ Given that CT scans are the most frequently performed abdominal imaging test in middle-age and older adults,¹⁹ the opportunity already exists to leverage or repurpose this body composition data for general health assessment.^10,18

The purpose of this study was to derive an abdominal CT-based biological age (CTBA) model informed only by an automated pipeline of validated cardiometabolic biomarkers and compare survival prediction over the usual demographic input data of chronological age, sex, and race.

The final study cohort consisted of 123,281 adults (mean age 53.6 years [SD 17.4]; 58,308 [47%] women and 64,973 [53%] men) who underwent abdominal CT scanning over a 20-year time interval. Median clinical post-CT follow-up was 5.3 years (IQR, 1.9–10.4 years), consisting of 811,802 total person-years of follow-up. A total of 26,554 (22%) patients died over the post-CT follow-up interval, whereas the remaining 96,727 (78%) were alive at last verifiable clinical contact. The median post-CT clinical follow-up interval was 6.0 years [IQR 2.6–11.3] for individuals still alive at last contact, and 2.6 years [IQR 0.7–6.8] for those who died. By race, the patient cohort was predominately White (92%), followed by Black (5%), Asian (2%), American Indian (1%), and Hawaiian (< 1%) descent.

From the full panel of potential automated CT measures (Fig. 1), a total of eight biomarkers contributed sufficiently to the CTBA model to warrant inclusion. Of these, muscle density, abdominal aortic calcium score, visceral fat density, and bone density demonstrated the largest IPA-drop, signifying the greatest contribution to the CTBA model. A patient example is shown in Fig. 2. The final panel of CT biomarkers, including their drop in IPA values, are shown in Table 1, along with the results of the demographics model utilizing patient chronological age, sex, and race. As expected, the demographics model was largely driven by chronological age, to which the CTBA model is blinded. For the full CTBA model, the index of prediction accuracy (or IPA) was 29.2, compared with an IPA of 21.7 for the demographics (chronological age/sex/race) model (p < 0.001). A nomogram for predicting survival according to the CTBA model is shown in Fig. 3. Skeletal muscle density was the dominant CT biomarker in the survival model, whereas muscle area played only a minor role. In terms of ROC curve analysis, the 5-year and 10-year AUCs for the CTBA model were 0.890 (95% CI, 0.884–0.896) and 0.880 (95% CI, 0.875–0.885), respectively, compared with 0.784 (95% CI, 0.776–0.792) and 0.782 (95% CI, 0.776–0.788) using patient CA, sex, and race, respectively (p < 0.001). For the eight CT biomarkers used in the CTBA model, substantial differences were observed between patients who died versus survived during their clinical follow-up. For example, Table 2 shows the difference in these biomarker measures for patients who died within 5 years after CT versus those who survived at least 5 years after CT, according to age group and sex. Neither the hepatic nor splenic biomarkers contributed to the final CTBA model, with a drop in IPA values < 0.1.

A Kaplan-Meier plot according to CTBA model output quartiles for the full study cohort is shown in Fig. 4. The age- and sex-corrected survival HR for the highest versus lowest risk CTBA quartile was 8.73 (95% CI, 8.14–9.36). When comparing the highest-risk against the other three quartiles, the age- and sex-corrected HR was 3.13 (95% CI, 3.04–3.23). Kaplan-Meier plots subcategorized by patient sex (males and females), middle age (40–59 years), and older age (60–79 years) are shown in Fig. 5. Table 3 provides AUC and HR results for the total cohort and sub-cohorts. An example case is shown in Fig. 2 for a patient in the highest-risk CTBA quartile. A scatter plot for individual survival prediction by the CTBA model is shown in Fig. 6.

External validation cohort

The final study group for the external validation cohort consisted of 40,718 adults (mean age 53.9 years [SD 16.9]; 22,316 [55%] women and 18,402 [45%] men) who underwent abdominal CT over a 20-year time interval. Median clinical post-CT follow-up was 5.3 years (IQR, 2.9–8.8 years), consisting of 253,298 total person-years of follow-up. A total of 3,718 (9%) patients died over the post-CT follow-up interval, whereas the remaining 37,000 (91%) were alive at last verifiable contact. The median post-CT clinical follow-up interval was 5.4 years [IQR 3.1-9.0] for individuals still alive at last contact, and 3.5 years [IQR 1.4–6.9] for those who died. According to race, the patient cohort was predominately White (89%), followed by Black (5%), Asian (6%), American Indian (< 1%), and Hawaiian (< 1%) descent.

The CTBA model performed similarly well on the external validation cohort, with a full model IPA of 28.6, compared with 18.5 for the demographics model. A calibration plot using the CTBA nomogram on this cohort is shown in Fig. 7. The 5-year and 10-year AUC values were 0.893 (95% CI, 0.867–0.918) and 0.888 (95% CI, 0.869–0.908), respectively. A Kaplan-Meier survival plot for the external validation cohort based on CTBA model quartiles is shown in Fig. 8. The age- and sex-corrected survival HR for the highest-vs-lowest risk CTBA quartile was 5.14 (95% CI, 3.98–6.63). When comparing the highest-risk against the other three quartiles, the age- and sex-corrected HR was 2.45 (95% CI, 2.24–2.69).

We found that a prediction model incorporating only understandable CT-based biomarkers of abdominal tissues and organs can provide a useful assessment of cardiometabolic health and estimation of longevity. This study demonstrates the value of harnessing the rich biometric tissue and organ data embedded within all body CT scans, but which typically go unused in routine practice.^10,13,18 Regardless of clinical indication, these CT scans can be opportunistically leveraged as an objective means for detecting silent or pre-symptomatic cardiometabolic conditions, including cardiovascular disease, osteoporosis, sarcopenia, diabetes, and metabolic syndrome. ^12–17 When previously unsuspected, these CT findings could initiate early preventive measures. For individuals with suspected or known risk factors, the objective and visual nature of the CT biomarker display may nonetheless motivate positive action. The advent of fully automated AI-based algorithms to mimic and replace more arduous manual approaches to these CT-based measurements provides for an efficient, explainable, objective, and reproducible method that is generalizable. Since body CT scans are already performed in such high volumes in middle-aged and older adults for a wide array of reasons,¹⁹ the potential for quasi- population-based opportunistic screening already exists.

The concept of biological aging is not new, but has lately become a topic of keen public interest, as seen in the recent lay press.^2–4 Beyond just the health-conscious “worried well”, there is growing recognition that many health care decisions should not be based solely on chronological age, but rather should account for the cumulative physiologic effects of lifestyle habits, genetic predisposition, and superimposed disease processes. The burgeoning interdisciplinary field of geroscience has largely focused on cellular and subcellular biomarkers, such as mitochondrial dysfunction, proteostasis, stem cell dysfunction, nutrient sensing, genomic instability, telomere dysfunction, cellular senescence, and epigenetic change.⁸ These “frailomics” measures of aging will undoubtedly provide some insight, but are unlikely to fully translate to the overall state of health of tissues, organs, or most importantly, the individual patient at the organism level.

Radiologic imaging biomarkers, whether more straightforward “explainable” measures as we employ or more complex radiomics (that we avoid), have generally received little attention for their potential role in determining an effective biological age.^7,9 In fact, a recent international task force on biological aging enumerated a myriad of potential biomarkers but failed to include imaging biomarkers and radiomics.⁸ However, we believe that imaging features (particularly CT-based) may better reflect the cumulative macroscopic effects of aging at the tissue and organ levels. Although numerous studies have shown a correlation between various imaging findings and patient age, comparatively few have explored the concept of biological aging.^20–22 Furthermore, we are not aware of any prior large-scale population-based studies on the order of 100,000 patients.

Our findings suggest that CT-based cardiometabolic biomarkers can effectively reflect the phenotypic pathologic and senescent changes at the tissue, organ, and organism level that result from the interaction of environmental factors on genetic predisposition. These macroscopic changes may be more relevant than (or at least complementary to) changes observed at the cellular or subcellular level. By utilizing only explainable AI algorithms, as opposed to a more opaque “black box” radiomics methodology, we believe this transparent approach could be more readily understood and accepted by patients and adopted by healthcare providers. The explainable methodology for our CTBA model provides transparency and avoids the “black box” opaqueness of deep learning approaches. Furthermore, our feature selection process using the IPA drop retains only biomarkers that improve predictive accuracy. Our focus on the predictive accuracy of the model largely eliminates concerns over multicollinearity, which may impact other approaches used for biological aging.^7,9

Clinical frailty assessments in current use are generally aimed more at advanced geriatric and acute care settings, and tend to be somewhat onerous to execute.²³ CT-based biological aging could also serve as an objective frailty assessment, and could be further modified in terms of reporting for sarcopenia, myosteatosis, and fracture risk.^12–17 Our CT-based approach could also be used to augment existing clinical risk prediction models, assuming the combination provides complementary information. A number of simple online risk calculators exist, most of which are disease specific in scope (eg, for breast or lung cancer assessment). Broader online risk calculators such as ePrognosis require manual entry of a host of demographic, clinical, and laboratory data (https://eprognosis.ucsf.edu). While these can provide for some level of risk assessment, a single CT likely provides more detailed objective insight into a patient’s actual cardiometabolic status. Of course, these approaches may prove to be complementary in nature with CT-based assessment.

The fact that CT-based biomarkers of muscle density, aortic plaque burden, visceral fat, and bone mineral density contributed the most to our CT biological age model was not unexpected given their established relationship with cardiometabolic disease.^12,13,17 With the exception of visceral fat, these biomarkers have a well-defined relationship with age.^24–26 However, more effective biological aging likely goes beyond simple quantification of the cumulative effects of aging, but also includes inflammation and related metabolic derangements. Skeletal muscle density, which is measured at CT using attenuation values and reflects the degree of myosteatosis, was the dominant biomarker in the CTBA model, whereas muscle cross-sectional area played a very minor role. This is consistent with prior work showing that CT-based measures of muscle quality (sarcopenic myosteatosis) are significantly more predictive of survival than CT-based measures of muscle quantity (myopenia).¹² The prognostic value of coronary calcium scoring at CT is also well established, and we have found that quantifying calcific plaque of the abdominal aorta is also a powerful biomarker for risk prediction.^13,15 Our automated aortic plaque tool also has the additional advantage that it can be applied to CT scans with IV contrast.²⁷ The opportunity for incidental osteoporosis screening at CT has also been recognized for over a decade.²⁸ However, manual case-by-case assessment in the course of routine CT interpretation has failed to move the needle like a more programmatic, automated approach would. There is evidence that the opportunistic reporting of automated quantification of atherosclerotic plaque and bone mineral density at abdominal CT would be a cost saving measure.²⁹ By systematically leveraging or repurposing these incidental tissue and organ measures on CT scans, there could be substantial implications for more intelligent utilization of limited healthcare resources.

We acknowledge limitations to our investigation. Due to the need for a large patient cohort with built-in long-term survival outcomes, this was by necessity a retrospective study. The indications for CT imaging varied widely – both a methodological strength and a weakness. The primary patient cohort and the external validation cohort lacked substantial racial or ethnic diversity, with both comprising Midwestern U.S. populations that were approximately 90% White. We plan to address this limitation with a multicenter trial consisting of broad national and international participation. We did not consider socioeconomic factors in the demographic-based model, but we also plan to investigate this utilizing the area deprivation index (ADI), a validated measure of socioeconomic disadvantage.³⁰ The automated AI pipeline used to obtain the CT cardiometabolic biomarkers is a research tool that is not yet commercially available. Finally, our CT-based biological age model is based on measurable cardiometabolic and senescent factors and cannot presently account for other co-existing maladies that may impact survival, such as trauma, cancer, infection, and dementia, among others. However, despite ignoring this potentially confounding clinical overlay, the CT-based model proved to be robust.

In summary, we have shown that a CT-based biological age (CTBA) model informed only by a panel of explainable AI-derived biomarkers provides a phenotypic cardiometabolic assessment for improved and personalized prediction of remaining life expectancy over usual demographic inputs. These CT measures reflect the cumulative impact of lifestyle, genetic predisposition, and chronological aging. In addition, these objective body composition findings may reflect an early pre-symptomatic phase of disease, prior to the development of clinically recognizable findings. This valuable imaging data can be opportunistically derived from nearly any abdominal CT, whether retrospectively or prospectively and regardless of the clinical indication. Incorporating this objective biological information into the full clinical assessment might better inform downstream healthcare decisions and resource allocation.

Study design and patient cohort

This retrospective cohort study was HIPAA compliant and IRB-approved. The need for signed informed consent was waived for this large retrospective cohort. The patient inclusion criteria were kept intentionally broad, consisting of any adult aged 18 years or older with an abdominal CT scan available in the PACS at the University of Wisconsin Hospital and Clinics (Madison, WI, USA) performed over a 20-year period. A similar process was followed to curate an external validation cohort of outpatients from Duly Health and Care (Downers Grove, IL, USA), with CT scans also performed over a 20-year period.

For the purpose of this study, the earliest available abdominal CT scan for each patient was used, both to ensure the longest possible clinical follow-up and to minimize the impact of any subsequent treatment or interventions. Given the broad inclusion criteria, a wide variety of clinical indications for scanning at the outpatient, inpatient, and emergency department settings was observed. The specific make and model of multi-detector CT scanners also widely varied, but these CT biomarkers algorithms have proven robust to all encountered vendors.³¹ Additional CT scanning details are included in the methodology supplement.

The main clinical outcome measure was patient death (all-cause mortality) or, if not deceased as of their verifiable clinical encounter, the date of last reliable contact. Acceptable encounters with medical staff included clinic visit, procedure, hospitalization, laboratory testing, and consultation. Confirmation of death for the main study cohort was periodically updated using internal and external sources, included the electronic medical record and the Social Security Death Master File.

Automated CT biomarker panel

A pipeline of mature, validated, and explainable CT-based AI algorithms automatically quantified skeletal muscle, abdominal fat, aortic calcification, bone density, liver, spleen, and kidneys, as described below and in the supplemental methods.

The panel of fully automated deep learning AI CT-based body composition algorithms used in this investigation have been integrated into a single portable Docker container at the University of Wisconsin. The individual CT body composition tools were developed, trained, and tested in separate cohorts at the NIH Clinical Center and University of Wisconsin (see supplemental methods for details). The tools have been subsequently modified with deep-learning improvements and validated at the University of Wisconsin. Source CT data from patient scans were preprocessed and reformatted into 3x3-mm series, upon which the AI tools were applied to create the body composition measures.

The first step for the AI toolkit is automatic vertebral body localization using a convolutional neural network (CNN) based on the unsupervised body part regression algorithm and applied in Caffe (Fig. 1). This process is used to identify the T12 through L4 vertebral bodies levels, from which the various segmented tissue and organ composition measures (density, area, volume) are subsequently derived. For muscle measures, a U-Net model architecture was used to identify the body wall, paraspinal, and psoas musculature, including intermuscular adipose tissue. Fat and bone measures were obtained using a U-Net like architecture with VGG11 encoder to segment visceral fat, subcutaneous fat, and trabecular bone. Aortic calcification was quantified using a modified 3D U-Net architecture to determine calcified aortic plaque from the diaphragm to the aortic bifurcation. Calcification is reported as an Agatston score. Volumetric organ segmentation for the liver, spleen, and kidneys each entailed a modified 3D U-Net and CycleGAN for CNN segmentation. The specific CT body composition biomarkers included in this study are listed in Supplemental Table 1. Figure 1 depicts a schematic flowchart. In addition to the numerical output for the various CT biomarkers, QA images (Fig. 2) are derived that depicts the tissue and organ segmentation at the L1 and L3 vertebral levels, as well as a coronal maximum intensity projection (MIP) to allow for rapid visual confirmation for individual cases.

CT biological age model methodology and statistical analysis

Multivariable survival analysis was modeled using Cox proportional hazards regression. A standard demographics assessment used patient chronological age, sex, and race as the predictors, whereas the CTBA model used the CT parameters exclusively, without any demographic input. Biomarker selection for the CTBA model was performed based on the index of prediction accuracy (IPA).³² Individual CT biomarkers were assessed according to their “IPA drop” – the higher the drop, the more important the biomarker contribution to the CTBA model. More specifically, predictors were ranked by their smallest contribution towards the IPA and eliminated unless they contributed an IPA drop value of at least 0.1. Linearity assumptions for continuous predictors were relaxed by using restricted cubic splines with 3 knots. Predictive abilities were assessed by calculating the time-dependent areas under the receiver operating characteristic curve (AUC), as well as calibration curves.³³ These assessments of discrimination and calibration were done with use of bootstrapping (200 resamples).

Five- and 10-year AUC values for the final CTBA and demographics models were derived. Survival curves were plotted using the Kaplan-Meier estimator, splitting the CTBA model results into quartiles. Univariable and multivariable analyses of the cardiometabolic CT biomarkers were performed, with patient chronological age, sex, and race considered as potential confounders. Age- and sex-corrected hazard ratios (HRs) with 95% CIs were computed for the CTBA model, comparing the highest-risk quartile with both the lowest-risk quartile, and with the other three quartiles. We compiled and compared summary statistics for patients who died versus survived over the course of available clinical surveillance. All analyses were performed using R version 4.3.1.

The CTBA model was not informed by demographic factors (chronological age, sex, and race), or by any acute or chronic medical conditions, such as known cardiovascular disease, diabetes, or cancer, even though this clinical information would have improved prediction of life expectancy. Consequently, this model was based solely on the CT biomarkers. A nomogram for this standalone model of individual CT biomarkers contributing to the CTBA model was generated.

Table 1

Model results according to drop in index of predictive accuracy (IPA) CTBA Model:
Full Model IPA = 29.2
Automated CT Biomarker*	Drop in IPA**
Skeletal muscle density	5.1
Aortic calcium score	2.0
Visceral adipose tissue density	1.5
L1 trabecular bone density	1.1
Visceral-to-subcutaneous fat ratio (VSR)	0.4
Kidney volume	0.3
Subcutaneous adipose tissue area	0.2
Skeletal muscle area	0.1

Demographics Model:

Full Model IPA = 21.7
Demographic measure	Drop in IPA**
Chronological age	21.1
Sex	0.3
Race	0.1
*CT biomarkers with an IPA-drop < 0.1 were excluded from the final model
**A larger IPA-drop signifies a greater contribution to the model prediction results

Table 2

CT body composition biomarker metrics according to 5-year post-CT survival horizon*
Sex	Age* (years)	Muscle Density (HU)		Abdominal Aortic Calcium Score (Agatston)		Visceral Fat Density (HU)		Trabecular Bone Density (HU)		Visceral-to-Subcutaneous Fat Ratio (VSR)		Kidney Volume (mL)		Subcutaneous Fat Area (cm²)		Muscle Area (cm²)		Patient N
Sex	Age* (years)	Alive	Dead	Alive	Dead	Alive	Dead	Alive	Dead	Alive	Dead	Alive	Dead	Alive	Dead	Alive	Dead	Alive	Dead
Men	18–39	48 (41–53)	44 (36–49)	0 (0–0)	0 (0–0)	-90 (-97- -82)	-89 (-96- -82)	167 (145–192)	166 (146–189)	0.63 (0.42–0.90)	0.70 (0.51–0.98)	424 (374–485)	419 (338–489)	150 (83–237)	155 (80–241)	181 (160–203)	182 (152–203)	6169	292
	40–59	40 (32–45)	34 (25–41)	0 (0-117)	65 (0-885)	-95 (-100- -88)	-90 (-98- -82)	144 (124–165)	138 (115–161)	1.02 (0.72–1.43)	1.02 (0.69–1.46)	445 (386–514)	442 (368–528)	179 (129–250)	176 (120–252)	188 (169–210)	180 (160–204)	11215	1514
	60–79	31 (22–38)	26 (17–35)	275 (0-1645)	942 (54-3510)	-95 (-100- -88)	-93 (-98- -86)	123 (101–144)	116 (94–139)	1.27 (0.89–1.74)	1.32 (0.92–1.63)	444 (382–520)	433 (363–509)	182 (138–241)	182 (133–242)	180 (161–201)	177 (158–198)	6468	2232
	80+	21 (12–30)	19 (10–28)	2057 (84-5239)	2645 (240–7342)	-93 (-97- -87)	-91 (-97- -85)	103 (82–128)	98 (78–124)	1.48 (1.03–1.86)	1.45 (1.09–1.94)	398 (336–478)	384 (333–459)	166 (138–219)	153 (119–202)	163 (147–178)	158 (144–175)	319	501
Women	18–39	44 (36–50)	38 (30–45)	0 (0–0)	0 (0–2)	-86 (-92- -80)	-85 (-92- -79)	181 (160–204)	171 (148–195)	0.24 (0.16–0.36)	0.31 (0.20–0.52)	352 (304–405)	346 (264–419)	219 (129–348)	213 (117–310)	127 (113–144)	125 (109–143)	8812	265
	40–59	34 (25–42)	28 (17–38)	0 (0–23)	12 (0-745)	-91 (-97- -84)	-89 (-95- -81)	155 (133–178)	144 (122–166)	0.38 (0.26–0.56)	0.48 (0.31–0.72)	346 (297–401)	349 (291–411)	246 (164–351)	234 (141–351)	130 (115–148)	132 (115–153)	14440	1324
	60–79	23 (12–32)	17 (6–27)	100 (0-1122)	757 (17-3512)	-92 (-98- -86)	-91 (-97- -84)	122 (103–143)	116 (95–137)	0.50 (0.35–0.71)	0.60 (0.42–0.85)	324 (276–379)	313 (262–371)	245 (172–335)	226 (152–316)	124 (110–140)	123 (108–143)	7284	2080
	80+	11 (1–20)	9 (-1-19)	1914 (127–6039)	2614 (548–6666)	-90 (-96- -84)	-88 (-95- -81)	99 (77–118)	93 (74–116)	0.60 (0.43–0.84)	0.62 (0.41–0.88)	294 (252–339)	271 (227–316)	205 (141–276)	193 (131–257)	116 (104–130)	116 (102–130)	534	694

Numbers are median values with interquartile ranges (IQR) in parentheses.

*Refers to chronological age of patients (not CT biological age)

Table 3

Results of the CT biological age (CTBA) model
Cohort	N	5-year AUC	10-year AUC	HR*	HR**
Total Cohort	123,281	0.890 (0.884–0.896)	0.880 (0.875–0.885)	8.73 (8.14–9.36)	3.13 (3.04–3.23)
Females	58,308	0.905 (0.898–0.913)	0.889 (0.882–0.896)	8.40 (7.58–9.31)	3.08 (2.95–3.23)
Males	64,973	0.874 (0.865–0.883)	0/871 (0.863–0.878)	8.82 (8.04–9.68)	3.20 (3.07–3.33)
40–59 year-olds	47,651	0.857 (0.841–0.872)	0.842 (0.829–0.856)	7.10 (6.57–7.68)	4.33 (4.13–4.53)
60–79 year-olds	38,973	0.834 (0.824–0.844	0.816 (0.806–0.825)	5.13 (4.83–5.45)	3.26 (3.15–3.38)
External Cohort	40,718	0.893 (0.867–0.918)	0.888 (0.869–0.908)	5.14 (3.98–6.63)	2.45 (2.24–2.69)
Note: the CTBA model was constructed from CT-biomarkers only, without any input regarding chronological age, sex, or race; HRs are age- and sex-corrected for total and external validation cohorts; HRs for sex cohorts were corrected for age, whereas HRs for age cohorts were corrected for sex
*Comparing the highest-risk CTBA quartile vs the lowest-risk quartile
**Comparing the highest-risk CTBA quartile vs the other three quartiles

Declaration of Interests: P.J.P.: advisor to Bracco Diagnostics, GE HealthCare, and Nanox-AI; J.W.G.: Advisor to RadUnity, Shareholder in NVIDIA; R.M.S.: Royalties from iCAD, ScanMed, Philips, Translation Holdings, PingAn, MGB; research support through a CRADA with PingAn.

Funding source:

Supported in part by the Intramural Research Program of the National Institutes of Health Clinical Center.

Levine ME (2013) Modeling the rate of senescence: can estimated biological age predict mortality more accurately than chronological age? journals Gerontol Ser Biol Sci Med Sci 68(6):667–674
Attia P, Gifford B (2023) Outlive: the Science & Art of Longevity. Harmony Books
Smith DG (2023) What’s Your ‘Biological Age’? New tests promise to tell you if you have the cells of a 30-year-old or a 60-year-old. New York Times, Dec. 19, (accessed 3/23/24).
Janin A (2024) To Get Ahead of Diseases, It May Help to Find Your Organ Age. Wall Str J, Mar 9, (accessed 3/23/24).
Comfort A (1969) Test-battery to measure ageing-rate in man. Lancet (London England) 2(7635):1411–1414
Oh HS, Rutledge J, Nachun D et al (2023) Organ aging signatures in the plasma proteome track health and disease. Nature 624(7990):164–172
Li Z, Zhang W, Duan Y et al (2023) Progress in biological age research. Front Public Health 11:1074274
LeBrasseur NK, de Cabo R, Fielding R et al (2021) Identifying Biomarkers for Biological Age: Geroscience and the ICFSR Task Force. J frailty aging 10(3):196–201
Bafei SEC, Shen C (2023) Biomarkers selection and mathematical modeling in biological age estimation. NPJ Aging 9(1):13
Pickhardt PJ (2022) Value-added Opportunistic CT Screening: State of the Art. Radiology 303(2):241–254
Pickhardt PJ, Graffy PM, Perez AA, Lubner MG, Elton DC, Summers RM (2021) Opportunistic Screening at Abdominal CT: Use of Automated Body Composition Biomarkers for Added Cardiometabolic Value. Radiographics 41(2):524–542
Nachit M, Horsmans Y, Summers RM, Leclercq IA, Pickhardt PJ (2023) AI-based CT Body Composition Identifies Myosteatosis as Key Mortality Predictor in Asymptomatic Adults. Radiology 307(5):e222008
Pickhardt PJ, Graffy PM, Zea R et al (2020) Automated CT biomarkers for opportunistic prediction of future cardiovascular events and mortality in an asymptomatic screening population: a retrospective cohort study. Lancet Digit Health 2(4):E192–E200
Pickhardt PJ, Graffy PM, Zea R et al (2021) Utilizing Fully Automated Abdominal CT-Based Biomarkers for Opportunistic Screening for Metabolic Syndrome in Adults Without Symptoms. Am J Roentgenol 216(1):85–92
O'Connor SD, Graffy PM, Zea R, Pickhardt PJ (2019) Does Nonenhanced CT-based Quantification of Abdominal Aortic Calcification Outperform the Framingham Risk Score in Predicting Cardiovascular Events in Asymptomatic Adults? Radiology 290(1):108–115
Liu D, Garrett JW, Perez AA et al (2024) Fully automated CT imaging biomarkers for opportunistic prediction of future hip fractures. The British journal of radiology
Pickhardt PJ, Graffy PM, Zea R et al (2020) Automated Abdominal CT Imaging Biomarkers for Opportunistic Prediction of Future Major Osteoporotic Fractures in Asymptomatic Adults. Radiology 297(1):64–72
Pickhardt PJ, Summers RM, Garrett JW et al (2023) Opportunistic Screening: Radiology Scientific Expert Panel. Radiology 307(5):e222044
Moreno CC, Hemingway J, Johnson AC, Hughes DR, Mittal PK, Duszak R (2016) Jr. Changing Abdominal Imaging Utilization Patterns: Perspectives From Medicare Beneficiaries Over Two Decades. J Am Coll Radiol 13(8):894–903
Jonsson BA, Bjornsdottir G, Thorgeirsson TE et al (2019) Brain age prediction using deep learning uncovers associated sequence variants. Nat Commun 10(1):5409
Rule AD, Grossardt BR, Weston AD et al (2024) Older Tissue Age Derived From Abdominal Computed Tomography Biomarkers of Muscle, Fat, and Bone Is Associated With Chronic Conditions and Higher Mortality. Mayo Clin Proc
Raghu VK, Weiss J, Hoffmann U, Aerts H, Lu MT (2021) Deep Learning to Estimate Biological Age From Chest Radiographs. JACC Cardiovasc imaging 14(11):2226–2236
Church S, Rogers E, Rockwood K, Theou O (2020) A scoping review of the Clinical Frailty Scale. BMC Geriatr 20(1):393
Jang S, Graffy PM, Ziemlewicz TJ, Lee SJ, Summers RM, Pickhardt PJ (2019) Opportunistic Osteoporosis Screening at Routine Abdominal and Thoracic CT: Normative L1 Trabecular Attenuation Values in More than 20 000 Adults. Radiology 291(2):360–367
Graffy PM, Liu J, Pickhardt PJ, Burns JE, Yao J, Summers RM (2019) Deep learning-based muscle segmentation and quantification at abdominal CT: application to a longitudinal adult screening cohort for sarcopenia assessment. Br J Radiol : 20190327
Graffy PM, Liu J, O'Connor S, Summers RM, Pickhardt PJ (2019) Automated segmentation and quantification of aortic calcification at abdominal CT: application of a deep learning-based algorithm to a longitudinal screening cohort. Abdom Radiol 44(8):2921–2929
Summers RM, Elton DC, Lee S et al (2021) Atherosclerotic Plaque Burden on Abdominal CT: Automated Assessment With Deep Learning on Noncontrast and Contrast-enhanced Scans. Acad Radiol 28(11):1491–1499
Pickhardt PJ, Pooler BD, Lauder T, del Rio AM, Bruce RJ, Binkley N (2013) Opportunistic Screening for Osteoporosis Using Abdominal Computed Tomography Scans Obtained for Other Indications. Ann Intern Med 158(8):588–595
Pickhardt PJ, Correale L, Hassan C (2023) AI-based opportunistic CT screening of incidental cardiovascular disease, osteoporosis, and sarcopenia: cost-effectiveness analysis. Abdom Radiol 48(3):1181–1198
Lee MH, Zea R, Garrett JW, Summers RM, Pickhardt PJ (2024) AI-generated CT body composition biomarkers associated with increased mortality risk in socioeconomically disadvantaged individuals. Abdom Radiol (New York)
Pooler BD, Garrett JW, Southard AM, Summers RM, Pickhardt PJ (2023) Technical Adequacy of Fully Automated Artificial Intelligence Body Composition Tools: Assessment in a Heterogeneous Sample of External CT Examinations. AJR Am J Roentgenol 221(1):124–134
Kattan MW, Gerds TA (2018) The index of prediction accuracy: an intuitive measure useful for evaluating risk prediction models. Diagn Progn Res 2:7
Harrell FE Jr., Lee KL, Mark DB (1996) Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 15(4):361–387

Yes there is potential Competing Interest. P.J.P.: advisor to Bracco Diagnostics, GE HealthCare, and Nanox-AI; J.W.G.: Advisor to RadUnity, Shareholder in NVIDIA; R.M.S.: Royalties from iCAD, ScanMed, Philips, Translation Holdings, PingAn, MGB; research support through a CRADA with PingAn.

SupplementalMethodsCTBA7724.pdf

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

Status:

Version 1

Abstract

Figures

Introduction

Results

External validation cohort

Discussion

Online Methods

Study design and patient cohort

Automated CT biomarker panel

CT biological age model methodology and statistical analysis

Declarations

Funding source:

References

Additional Declarations

Supplementary Files

Status:

Version 1