Biomarkers Associated with Physical Resilience After Hip Fracture

Abstract

Background

Clinically similar older adults demonstrate variable responses to health stressors, heterogeneity attributable to differences in physical resilience. However, molecular mechanisms underlying physical resilience are unknown. We previously derived a measure of physical resilience after hip fracture—the expected recovery differential (ERD)—that captures the difference between actual recovery and predicted recovery. Starting with biomarkers associated with physical performance, morbidity, mortality, and hip fracture, we evaluated associations with the ERD to identify biomarkers of physical resilience after hip fracture.

Methods

In the Baltimore Hip Studies (N = 304) sera, we quantified biomarkers of inflammation (TNFR-I, TNFR-II, sVCAM-1, and IL-6), metabolic and mitochondrial function (non-esterified fatty acids, lactate, ketones, acylcarnitines, free amino acids, and IGF-1), and epigenetic dysregulation (circulating microRNAs). We used principal component analysis, canonical correlation, and least absolute shrinkage and selection operator regression (LASSO) to identify biomarker associations with better-than-expected recovery (greater ERD) after hip fracture.

Results

Participants with greater ERD were more likely to be women and less disabled at baseline. The complete biomarker set explained 37% of the variance in ERD (p < .001) by canonical correlation. LASSO regression identified a biomarker subset that accounted for 27% of the total variance in the ERD and included a metabolic factor (aspartate/asparagine, C22, C5:1, lactate, glutamate/mine), TNFR-I, miR-376a-3p, and miR-16-5p.

Conclusions

We identified a set of biomarkers that explained 27% of the variance in ERD—a measure of physical resilience after hip fracture. These ERD-associated biomarkers may be useful in predicting physical resilience in older adults facing hip fracture and other acute health stressors.

It has become increasingly clear that, to improve the overall health and quality of life of our aging population, we need to better understand the mechanisms underlying physical resilience. There are multiple definitions of physical resilience; here, we define physical resilience as an individual’s ability to resist decline or to recover function following a physical stressor. Physical resilience is especially relevant to the care of older adults because it partially accounts for the heterogeneity of recovery trajectories in individuals who may otherwise appear clinically similar (1,2). Currently, defined measures of physical resilience require multiple measures of function or well-being over time to reflect the dynamic response to an acute health stressor. Identifying biomarkers of physical resilience would be useful to (i) understand the biological mechanisms underlying the heterogeneity in recovery from acute health stressors; (ii) identify older adults at high risk of disability after an acute health stressor to create realistic care plans and best allocate resources, such as for enhanced rehabilitation; and (iii) develop resilience-building interventions.

Several conceptual frameworks have been proposed to operationalize measures of physical resilience (2,3). We previously defined one such approach—the expected recovery differential (ERD) —in a cohort of older adults with hip fracture (4). The ERD uses predictive modeling of longitudinal data to incorporate baseline clinical, functional, and stressor characteristics with pre- and post-stressor outcomes over time to define how an individual’s actual recovery trajectory differs from their expected recovery (Figure 1). By incorporating data on recovery trajectories, the ERD produces a continuous measure that captures the dynamic nature of physical resilience. This approach facilitates the identification of molecular biomarkers relevant to physical resilience beyond age, demographic characteristics, and pre-stressor function.

Figure 1.

Description of the expected recovery differential (ERD) as a measure of physical resilience. To derive the ERD, we first build a model to predict functional recovery using age, body mass index, education, gender, race, depression, self-reported health, number of comorbidities, fracture site, delirium, dementia, and discharge destination. The ERD is defined as the difference between the predicted and actual functional recovery after 12 months. The ERD for lower extremity physical activities of daily living (LPADL) is shown. LPADL is one of the 10 outcome variables used to derive the total ERD.

Current conceptual models of physical resilience incorporate demographic and psychosocial characteristics, environmental factors, and physiologic reserve. Physiologic reserve reflects molecular processes that contribute to cellular-, organ-, and organism-level homeostasis. The age-related decline in physiologic reserve, and thus physical resilience, is hypothesized to be driven by biological aging (5). We evaluated biomarkers of biological aging representing inflammation (tumor necrosis factor-α receptor [TNFR]-I, TNFR-II, sVCAM-1, and interleukin-6 [IL-6]), metabolic and mitochondrial function (conventional fatty acids, lactate, ketones, acylcarnitines, free amino acids, and insulin-like growth factor 1), and epigenetic dysregulation (circulating microRNAs [miRNAs]). We selected these biomarkers based on previous work by our group and others identifying associations of these biomarkers with physical performance, morbidity (including the risk of hip fracture and post-hip fracture outcomes), and mortality in older adults (6–12).

The objective of this work was to (i) determine whether a panel of biomarkers previously associated with morbidity and mortality in older adults is associated with the ERD—a measure of physical resilience—after hip fracture and (ii) identify a parsimonious set of biomarkers that could be useful for predicting physical resilience.

Methods

Baltimore Hip Studies Cohort

We quantified biomarkers in a pooled cohort of participants from Baltimore Hip Studies (BHS)-4 (N = 180) and BHS-7 (N = 339). Participants from the two cohorts were included based on the availability of sufficient data to calculate the ERD and sufficient sera for biomarker quantification (Figure 2; N = 304). The BHS cohorts are described in detail elsewhere (13,14). Briefly, the BHS comprise a series of prospective studies that enrolled hip fracture patients who were ambulatory and living in the community prior to fracture, English-speaking, aged 65 and older who lived within 70 miles of the study center, and were admitted to study hospitals within the BHS hospital network in the greater Baltimore area for surgical repair of a non-pathological hip fracture. BHS-4 enrolled only women and excluded patients with a Mini-Mental Status Exam score ≤ 20, end-stage renal disease, cirrhosis, or metastatic cancer. Both of the studies excluded patients with hardware in the contralateral hip or weight > 300 lbs.

Figure 2.

Study sample consort diagram. Participants were included in the analysis if they had enough longitudinal follow-up data to calculate the expected recovery differential (ERD) and adequate sera available for biomarker quantification.

Recovery outcomes used to calculate the ERD were obtained from three BHS cohorts, the same two used for this analysis and a third cohort that did not collect serum (Figure 2). Outcomes data were collected at the index hospitalization and at 2, 6, and 12 months in each participant’s place of residence. Recovery measures included 10 self-reported and physical performance measures including ambulatory status, lower extremity physical activities of daily living specifically adapted for hip fracture patients from the Functional Status Index (15), modified Instrumental Activities of Daily Living (16), Yale Physical Activity Scale (17), 2 days of actigraphy, grip strength using a hand dynamometer, a balance test (Short Physical Performance Battery (18) or a modified Tinetti Gait and Balance Test (19)), Lower Extremity Gain Scale (20), 3-m gait speed, and timed single chair stand. For each of these 10 measures, the ERD was defined as the difference between actual and predicted recovery trajectories over the 12 months following hospitalization for hip fracture repair and was developed using a mixed model with random and fixed effects (4). Models accounted for study cohort; demographic factors (age, race, and gender); psychosocial factors (depression, socioeconomic status, and education); environmental factors (place of residence and duration of rehabilitation); baseline self-reported functional status; comorbidities (BMI, diabetes, chronic lung disease, cardiovascular disease, chronic kidney disease, cancer, and cerebrovascular disease); and stressor characteristics (type of fracture, surgery, anesthesia, and length of stay). Sera were obtained at one timepoint, within 22 days (M = 11.4 ± 2.8) of hip fracture or hospitalization, and all but one serum sample was obtained after surgical repair. Serum samples were stored at −80°C and had not undergone a freeze/thaw cycle until analysis.

Molecular Assays

We selected biomarkers previously reported to be associated with physical performance, morbidity (including the risk of hip fracture and post-hip fracture outcomes), and mortality in older adults (6–12). Sera were used for the quantitative determination of all biomarkers. TNFR-I, TNFR-II and IL-6 were quantified by sandwich immunoassay using electrochemiluminescence detection (Mesoscale Discovery Systems, Rockville, MD). Vascular cell adhesion molecule 1 (VCAM-1) and insulin-like growth factor 1 were quantified by ELISA using colorimetric detection (R&D Systems, Minneapolis, MN). The lower limit of detection and intra- and interassay coefficients of variation for the peptide assays are included in Supplementary Methods—Tables 1 and 2. All protein concentrations measured were within the detectable range of the assays; none had concentrations below the lower limit of detection. Forty-five acylcarnitines, 15 free amino acids, and 4 conventional metabolites (non-esterified fatty acids, beta-hydroxybutyrate, ketones, and lactate) were quantified using mass spectrometry by the Sarah W. Stedman Nutrition and Metabolism metabolomics core laboratory of the Duke Molecular Physiology Institute, as described previously (21).

We evaluated 32 miRNAs associated with longevity in a cohort of older adults (unpublished data Kraus, VB). Circulating miRNAs were isolated from 200 µL of sera using the miRNeasy serum/plasma advanced kit (Qiagen, Germany). Reverse transcription was performed to generate cDNAs from miRNAs using the miRCURY LNA RT kit (Qiagen). Expression of 32 miRNAs was quantified by qPCR (miRNAs and primers utilized listed in Supplementary Methods—Table 3) using the miRCURY LNA SYBR Green PCR Kit (Qiagen). All immunoassays and miRNA analyses were performed by the Biomarkers Shared Resource at the Duke Molecular Physiology Institute.

Statistical Analysis

Metabolite and protein biomarker concentrations were not normally distributed and were log-transformed. The inverse of the miRNA C_t values was calculated so that greater values indicated greater miRNA concentrations. Because the rich dataset included more than 100 intercorrelated biomarkers for each participant, statistical techniques were employed to reduce the number of predictor variables. Based on previous work by our group and others, we used principal component analysis (PCA) to reduce the conventional metabolites, acylcarnitines, and free amino acids to four factors and canonical correlation to evaluate the association of the complete biomarker set, including the PCA-derived metabolite factors, with the ERD (21). PCA extracts a set of orthogonal (uncorrelated) linear predictors from multiple predictors that maximize the explained variance. Canonical correlation analysis constructs a set of canonical variates, which are orthogonal linear combinations of variables that best account for the variability within the biomarkers and between the biomarkers and the outcome variable, in this case, the ERD (22). To facilitate visualization of related biomarkers in the canonical correlation, we used hierarchical clustering to cluster biomarkers based on their Euclidean distance (23). Given the variability in the timing of sera collection, we included the timing of sera collection as a covariate in our preliminary analysis.

Many circulating miRNAs decrease with age (24). To determine whether mean miRNA concentrations were associated with the ERD, we used ordinary least-squares linear regression and included age as a covariate. RNU6-1 was initially included as a housekeeping gene for the normalization of the miRNA data, but in our preliminary analysis, RNU6-1 was itself correlated with the ERD, so inverse, nonnormalized miRNA C_t values were analyzed.

To identify the most parsimonious set of biomarkers that predicted the ERD, we employed least absolute shrinkage and selection operator (LASSO) regression. LASSO regression performs feature selection and regularization and is well suited to situations where there are high levels of multicollinearity (25). LASSO regression was performed with p = .1 for each biomarker to enter the model and p < .20 as the threshold for removal from the model. To understand the associations of the individual biomarkers retained in the final LASSO model with the ERD, we conducted additional sensitivity analyses using ordinary least squares regression. To understand the association of the individual biomarkers retained in the final LASSO model with the individual outcomes used to derive the ERD, we evaluated biomarker associations with trajectories (slopes) of individual outcomes over 12 months using Pearson correlation.

All statistical analyses were conducted in SAS version 9.2 or higher and R.

Results

Cohort descriptive statistics are provided in Table 1 for the cohort as a whole and by quartile of ERD. Participants in higher ERD quartiles (reflecting better-than-expected recovery) were predominantly women, had lesser BMIs, fewer disabilities, and were less likely to have cognitive impairment at baseline. Consistent with this, participants in higher ERD quartiles were more likely to come from BHS-4 than BHS-7, given that BHS-4 was composed exclusively of women and had more restrictive inclusion criteria (data not shown). There were no differences in age, race, self-rated health, number of surgical complications, or discharge destination. Given the variability in the timing of sera collection, we explored controlling for the timing of the blood draw, but we did not find any association with the ERD, so this was not retained in the final analysis (data not shown). The inverse mean C_t value of the 32 miRNAs, indicating their overall higher expression, was positively associated with ERD (β = 29.31 95% CI: 14.79, 43.83; p < .001), controlling for age.

Table 1.

Cohort Descriptive Statistics by Expected Recovery Differential (ERD) Quartile

		worse-than-expected			better-than-expected
	BHS (N = 304)	ERD Q1 (N = 76)	ERD Q2 (N = 76)	ERD Q3 (N = 76)	ERD Q4 (N = 76)	p
Age	80.8 (6.76)	81.6 (7.01)	81.8 (6.89)	80.7 (6.32)	79.1 (6.60)	NS
Sex
Women	202 (66.4%)	7 (9.21%)	47 (61.8%)	72 (94.7%)	76 (100%)	<.001
Men	102 (33.6%)	69 (90.8%)	29 (38.2%)	4 (5.26%)	0 (0.00%)
Race/Ethnicity
Non-white	16 (5.26%)	4 (5.26%)	7 (9.21%)	3 (3.95%)	2 (2.63%)	NS
White	288 (94.7%)	72 (94.7%)	69 (90.8%)	73 (96.1%)	74 (97.4%)
BMI	24.9 (4.77)	25.6 (4.31)	25.4 (5.71)	25.3 (4.74)	23.4 (3.85)	.012
Fracture site
Intracapsular	147 (48.4%)	33 (43.4%)	30 (39.5%)	41 (53.9%)	43 (56.6%)	NS
Trochanteric	131 (43.1%)	37 (48.7%)	37 (48.7%)	27 (35.5%)	30 (39.5%)
Subtrochanteric	26 (8.55%)	6 (7.89%)	9 (11.8%)	8 (10.5%)	3 (3.95%)
Self-rated health
Excellent	47 (15.5%)	7 (9.21%)	12 (15.8%)	11 (14.5%)	17 (22.4%)	NS
Very good	105 (34.5%)	18 (23.7%)	25 (32.9%)	30 (39.5%)	32 (42.1%)
Good	104 (34.2%)	33 (43.4%)	26 (34.2%)	25 (32.9%)	20 (26.3%)
Fair	36 (11.8%)	12 (15.8%)	11 (14.5%)	8 (10.5%)	5 (6.58%)
Poor	12 (3.95%)	6 (7.89%)	2 (2.63%)	2 (2.63%)	2 (2.63%)
Presurgical function
Lower body ADL disabilities	2.37 (2.36)	3.39 (2.79)	2.91 (2.72)	2.12 (1.49)	1.05 (1.40)	<.001
IADL disabilities	1.60 (1.49)	2.55 (1.47)	2.01 (1.43)	1.15 (1.22)	0.69 (1.06)	<.001
Cognitive impairment	45 (14.8%)	28 (36.8%)	15 (19.7%)	1 (1.32%)	1 (1.32%)	<.001
Surgical complications	7 (2.30%)	1 (1.32%)	3 (3.95%)	2 (2.63%)	1 (1.32%)	NS
Discharge destination
Inpatient rehabilitation/SNF	280 (92.1%)	72 (94.7%)	67 (88.2%)	74 (97.4%)	67 (88.2%)	NS
Other	24 (7.89%)	4 (5.26%)	9 (11.8%)	2 (2.63%)	9 (11.8%)

Notes: Cohort descriptive statistics are shown for the cohort as a whole (BHS column) as well as by quartile of the ERD. ERD Q1 to ERD Q4 refer to BHS participants in the lowest to highest quartiles of ERD, indicating worse- to better-than-expected recovery after hip fracture. ADL = activities of daily living; BHS = Baltimore Hip Studies; IADL = instrumental activities of daily living; SNF, skilled nursing facility.

We applied PCA to the conventional metabolites, acylcarnitines, and amino acids, identifying four factors that accounted for 63% of the total metabolite variance (Figure 3A). Of the four factors, three were significantly associated with ERD (Factor 2: r = −0.15, p = .011; Factor 3: r = −0.13, p = .025; Factor 4: r = 0.42, p < .001; Figure 3B). Of these, Factor 4 was most strongly associated with greater ERD and contained high loadings (>|0.4|) for aspartate/asparagine, C22, C5:1, lactate (inverse), and glutamate/mine (inverse); inverse loadings indicate that lower concentrations of lactate and glutamate/mine were associated with greater concentrations of the other constituents.

Figure 3.

Principal components analysis (PCA) of conventional metabolites, free amino acids, and acylcarnitines. PCA was used to reduce amino acid and acylcarnitine metabolites into four factors. (A) It shows a table of each PCA-derived metabolite factor, the individual metabolites that loaded onto it, a description of the metabolites included in the factor, and the cumulative variance of the all of the metabolites explained by each factor. Individual metabolites with loadings at least 0.40 are shown in the Factor Loadings column. (B) Scatter plot showing the association of the ERD (x-axis) with the PCA-derived metabolite factor (y-axis) and Pearson correlation (r). AC = acylcarnitines; NEFA = non-esterified fatty acids; BCAA = branched chain amino acids; LNAA = long neutral amino acids; AA = amino acids.

Given the high degree of collinearity in our biomarker panel, we used canonical correlation to determine the variance in the ERD explained by the biomarker set as a whole. We obtained a squared canonical correlation value—analogous to an r²—of 0.37 (p < .001; Figure 4), indicating that the biomarker panel explained 37% of the variance in ERD. The canonical variate was most strongly correlated with Factor 4 (r = 0.70) and was moderately correlated with several miRNAs (r ≥ 0.40): miR-28-3p, miR-23b-3p, and miR-376a-3p. Furthermore, the canonical variate was inversely correlated (r ≥ −0.40) with sVCAM-1, TNFR-I, TNFR-II, and RNU6-1. RNU6-1 is typically used as a housekeeping gene for miRNA normalization.

Figure 4.

Association of the biomarker panel with the expected recovery differential (ERD) using canonical correlation (CC) analysis. We used CC to evaluate the association of the principal components analysis (PCA)-derived metabolite factors, inflammatory cytokines, IGF-1, and longevity-associated microRNAs with the ERD. The ERD column shows the Pearson correlation between each biomarker and the ERD. We calculated a canonical variate, which is an orthogonal linear combination of the biomarker variables that best explain the variability in the ERD. The CC column shows the correlations of each biomarker with the canonical variate and can be interpreted similarly to a PCA factor loading. The strength of the correlation is denoted by darker shading corresponding to a stronger correlation and lighter shading corresponding to a weaker correlation. Positive values or red color indicate a positive correlation. Negative values or blue color indicate an inverse correlation. The squared CC—analogous to an r²—for the association of the biomarkers with the ERD is 0.37. The dendrogram shows the results of the hierarchical clustering of the biomarkers based on their Euclidean distance.

We next sought to identify the most parsimonious set of biomarkers that explained the greatest amount of variance in the ERD. Our final LASSO model included Factor 4 (aspartate/asparagine, C22, C5:1, lactate [inverse], glutamate/mine [inverse]), TNFR-I, miR-376a-3p, and miR-16-5p. In total, these biomarkers explained 27% of the variance in the ERD (Figure 5). Of note, the biomarkers selected by the final LASSO model explained 73% of the complete biomarker variance (LASSO r² = 0.27/canonical correlation r² = 0.37). Although the inclusion of TNFR-I, miR-376a-3p, and miR-16-5p in the final LASSO model only marginally increased the explained variance, in a sensitivity analysis, TNFR-I alone explained 14% and the two miRNAs explained 18% of the variance in the ERD. Excluding Factor 4, TNFR-I and the two miRNAs explained 25% of the variance in the ERD, similar to the 27% obtained in the final LASSO model (data not shown). We attribute this to the high degree of colinearity in the biomarkers retained in the final LASSO model.

Figure 5.

Least absolute shrinkage and selection operator (LASSO) regression of the biomarker panel and the expected recovery differential (ERD). LASSO regression was used to identify the most parsimonious set of biomarkers that explained the ERD. (A) It shows the change in the standardized coefficient with the introduction of additional independent variables into the model. The number followed by the variable name refers to the previous variables in the model (eg, “1 + Factor 4” refers to the intercept + Factor 4). (B) It shows the p-value for each independent variable as it is introduced into the model. The addition of VCAM-1 yielded a p-value of more than .10, so it was removed, resulting in the final model. (C) It shows the final model. The total variance explained by the model was r² = 0.27.

To better understand the association of the biomarkers with the outcomes used to derive the ERD, we evaluated biomarker associations with trajectories of recovery in each outcome. Biomarkers retained in the final LASSO model were significantly associated with trajectories of outcomes used to derive the ERD (Supplementary Figures 1–4).

Discussion

Herein, we determined that a panel of biomarkers selected based on previously reported associations with physical performance, morbidity (including the risk of hip fracture and post-hip fracture outcomes), and mortality are associated with the ERD—a measure of physical resilience after hip fracture. We sought to obtain a parsimonious set of biomarkers that could be used to predict physical resilience. Using LASSO regression, we identified a subset of biomarkers—a PCA-derived metabolic factor (Factor 4: aspartate/asparagine, C22, C5:1, lactate [inverse], glutamate/mine [inverse]), TNFR-I, miR-376a-3p, and mir-16-5p—that captured 73% of the variance in the entire biomarker set and explained 27% of the variance in the ERD.

The PCA-derived metabolic factor (Factor 4) contained high loadings (>|0.4|) for concentrations of aspartate/asparagine, C5:1, C22, lactate (inverse), and glutamate/glutamine (inverse). These ERD-associated metabolites are constituents of metabolic pathways important for ammonia recycling, aspartate metabolism, and the urea cycle and may be indicative of better protein homeostasis (26,27), mitochondrial function (28), and renal function (29,30). Taken together, Factor 4 may indicate a pro-anabolic, pro-oxidative metabolic state poised for injury repair. Consistent with this, greater Factor 4 scores were positively associated with trajectories of recovery over 12 months in grip strength, balance, ambulatory category, Lower Extremity Gain Scale score, and in lower extremity activities of daily living disability (Supplementary Figure 1).

In addition to the PCA-derived metabolic factor (Factor 4), our final LASSO model included the inflammatory biomarker TNFR-I and two miRNAs (miR-376-3p and miR-16-5p). Excluding Factor 4, TNFR-I and the two miRNAs explained 25% of the variance in the ERD. We found that greater baseline TNFR-I concentrations were associated with worse-than-expected recovery after hip fracture. Our findings are consistent with previous BHS studies showing that in the year following hip fracture, greater IL-6 and TNFR-I concentrations predict worse lower extremity function and accelerated bone mineral density decline, respectively (6 (p. 6),7). In addition, baseline concentrations of TNFR-I were significantly associated with trajectories of outcomes used to derive the ERD (Supplementary Figure 2).

MiRNAs are important for coordinating gene regulation across tissues and generally act to inhibit translation of messenger RNA into protein (24). Concentrations of circulating miRNAs decline with age (24). In our analysis, greater mean miRNA concentrations were associated with better-than-expected recovery after hip fracture and two specific miRNAs were retained in our LASSO model: miR-376-3p and miR-16-5p. Concentrations of miR-376a-3p and miR-16-5p were also significantly associated with trajectories of outcomes used to derive the ERD (Supplementary Figures 3 and 4).

RNA, U6 small nuclear 1 (RNU6-1) is a long noncoding RNA routinely used as an internal control for normalizing expression levels in miRNA experiments. We found that concentrations of RNU6-1 were negatively associated with ERD. Additionally, RNU6-1 was moderately correlated with concentrations of TNFR-I (Pearson r = 0.33; p < .001) and Factor 4 (Pearson r = −0.51; p < .001). RNU6-1 concentrations are also greater in tissues and exosomes of patients with cancer (31 (p. 6),32). RNU6-1 may therefore be unsuitable as an internal control for normalization in human miRNA studies and may instead represent a biomarker of metabolism and inflammation.

Our results require additional validation in other cohorts, but biomarker-guided risk stratification could support individualized rehabilitation in older adults predicted to have poor physical resilience. Whether the biomarker signatures we observed are responsive to interventions that improve metabolic health and inflammation—for example, diet modification, exercise training, or pharmacologic modulation—in older adults like those in the BHS is unknown. However, our findings offer preliminary support for the importance of metabolic and inflammatory signaling pathways in physical resilience and a preliminary rationale for resilience-building interventions targeted to these pathways to enhance physical resilience in older adults prior to surgery.

Our study has several limitations. The BHS cohort used for this analysis was two-thirds women, almost entirely Caucasian, and drawn exclusively from the mid-Atlantic region of the United States. Therefore, generalizability to other populations is limited. Sera were collected at various times within the first 22 days after hip fracture and biomarker concentrations may therefore differ from pre-fracture values. The ERD as a measure of resilience is highly dependent on modeling and the variables included; therefore, our findings may not be applicable to other cohorts. However, the associations we identified with the included biomarkers and individual functional outcomes replicate previously reported findings by our group and others. Therefore, we hypothesize that these biomarkers may be predictive of physical resilience in response to stressors other than hip fracture.

Furthermore, it is important to note that our objective was to identify a parsimonious set of independent predictors of the ERD. Using this strategy, we expected that a number of recognized predictors of resilience (eg, IL-6) may not be represented based on colinearity with others highlighted here (eg, TNFR-I) or because of associations with expected recovery trajectory components (eg, obesity and branched chain amino acids).

In conclusion, we identified a set of biomarkers that explain 27% of better-than-expected hip fracture recovery, that is, better physical resilience. These biomarkers highlight pathways relevant to biological aging and potential targets for resilience-building interventions. Furthermore, these biomarkers have the potential to improve predictions of recovery and personalized treatment approaches following acute health stressors in older adults.

Funding

This work was supported by the National Institutes of Health UH2 AG056925-02 mechanism awarded to C.C.-E and H.E.W. This work was supported by the Claude D. Pepper Older Americans Independence Center supported by the National Institute on Aging at the National Institutes of Health (grant number 2P30AG028716-11). The research for the Baltimore Hip Studies datasets provided was supported by grants from the National Institute on Aging (R37 AG009901, R01 AG18668, R01 AG029315, P30 AG028747, and T32 AG000262). The research reported in this publication was supported by the National Center For Advancing Translational Sciences of the National Institutes of Health under award number UL1TR002553. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Author Contributions

D.C.P. contributed to conceptual design, the statistical analysis, participated in the interpretation of the results, produced the initial draft manuscript and figures, and contributed to draft revisions. C.C.-E., V.B.K., H.E.W., D.O., D.M.C., J.M., and K.M.H. participated in the conceptual design, interpretation of the results, and contributed to draft revisions. J.H., O.R.I., M.J.H., and C.H. performed the assays described in the manuscript and contributed to draft revisions. C.F.P. and R.S. conducted the statistical analyses described in the manuscript, participated in the interpretation of the results, and contributed to draft revisions.

Conflict of Interest

During the past year, Cathleen Colón-Emeric has consulted or served on advisory boards for: Novartis, Amgen, and Biscardia with no relevant conflicts of interest with this work. During the past year, Jay Magaziner has consulted or served on advisory boards for: American Orthopaedic Association, Fragility Fracture Network, Novartis, Pluristem, and UCB. The remaining authors have declared no conflicts of interest for this article.