This analysis used data from the Toledo Study for Healthy Aging (TSHA), a prospective, population-based cohort study investigating the determinants and consequences of frailty syndrome in older adults living in Toledo, Spain [14]. The study included 1972 participants who completed a home-based interview and physical evaluation, with biological samples provided by 1752 individuals for laboratory analysis. Baseline data were collected between June 2006 and September 2009, and participants were followed for an average of 3.5 (SD 1.19; range 0.30‐4.79) years to evaluate study outcomes [11].

The FTS-12 served as the dependent variable in statistical analyses. This scale includes 12 items across 7 dimensions. To calculate the FTS-12 score for each participant, the TSHA collected data in the following areas: (1) self-reported comorbidities, (2) basic and instrumental activities of daily living, (3) depressive symptoms, and (4) cognitive function. Trained nurses conducted physical examinations and clinical tests at participants’ homes to supplement these data. Detailed assessment methods are provided in García-García et al [11].

Participants also provided 12-hour fasting blood samples, from which two primary categories of variables were derived: (1) 15 variables assessing blood characteristics (hemogram) and (2) 19 variables measuring biochemical characteristics (biochemical profile), along with an added 12 unique biochemical traits not included in the standard profile. In total, 46 blood-related variables, combined with participants’ age, were included in the original database.

The TSHA was approved by the clinical research ethics committee of the Complejo Hospitalario de Toledo, Toledo, Spain. The original cohort study (“Estudio Toledo de Envejecimiento”) received ethics approval on March 30, 2005 (application 22). In addition, the specific research project entitled “Predictors of frailty, disability, cognitive impairment, vascular events, hospitalization, and mortality in a cohort of adults older than 64 years. Toledo Study for Healthy Aging” was approved on July 14, 2010 (application 93). All participants provided written informed consent before enrollment. This study is a secondary analysis of previously collected cohort data.

The mean age of participants included in the selection process was 75.04 (SD 5.84; range 65‐95) years for female participants and 74.95 (SD 5.63; range 65‐99) years for male participants. The mean FTS-12 scores observed in the study were 46.68 (SD 15.99; range 6‐92.11) for female participants and 39.62 (SD 15.43; range 0‐84) for male participants. The original database comprised 847 (57%) female participants and 637 (43%) male participants.

Tables S1 and S2 in Multimedia Appendix 3 summarize the variables retained in the regression models for female participants and male participants, respectively. In both sexes, the selected models combined individual laboratory markers with composite functions derived from clinically relevant variables. Table 1 presents the diagnostic statistics for the models stratified by sex. The F-statistics for both models were highly significant (P<.001), and the root mean square error ranged from 12 to 12.45. These findings highlight the relevance of both individual risk factors and their combined effects in modeling FTS-12 scores. Overall model performance was moderate, with R² values of approximately 0.39 in female participants and 0.38 in male participants. The parameter estimates and associated statistics for the selected multiple regression models in female and male participants are presented in Tables 2 and 3, respectively. The definitions and mathematical expressions of the composite risk factors used across the regression models are provided in Table 4.

Figures 2 and 3 further illustrate the modeling quality of the models. A moderately strong Pearson correlation was observed between the observed and modeled FTS-12 scores for both sexes, with correlations of 0.621 for female participants and 0.615 for male participants. The predicted FTS-12 scores were reasonably aligned with the observed scores in both sexes, supporting the feasibility of this approach while also indicating that substantial prediction uncertainty remains. The scatter plots visually represent this positive relationship, with a colored ellipsoid enclosing 90% of the data points. Higher correlations indicate better agreement between observed and predicted FTS-12 values.

Tables 2 and 3 present the regression parameter estimates and related statistics for the 2 selected models. The composite risk functions were more strongly related to the FTS-12 score than most single laboratory variables. This suggests that frailty is better captured by combinations of biologically related markers than by isolated biomarkers alone. In female participants, the strongest association was observed for the function combining red cell distribution width (RDW), age, and albumin, whereas in male participants, the strongest association involved albumin- and age-related composite terms. These findings support the idea that frailty risk is influenced by interacting physiological processes rather than by a single laboratory abnormality. To enhance visualization, colored ellipsoids enclosing 90% of the data points are included in both figures. These ellipsoids highlight the trends that informed the selection of these functions for the models. Notably, the relationship is not strong enough to approximate a straight line. If they were, each value of the function composite risk factor (CRF) 1 (Figure S1 in Multimedia Appendix 4) or CRF3 (Figure S2 in Multimedia Appendix 4) would correspond to only 1 or a few closely related FTS-12 values. Instead, the ellipsoids suggest a broader range of FTS-12 scores for a specific value of each function. Nonetheless, a positive trend is evident for CRF1 (Figure S1 in Multimedia Appendix 4), while a negative trend is observed for CRF3 (Figure S2 in Multimedia Appendix 4). Comparatively, correlations with single risk factors were moderate. For female participants, the highest correlation was 0.187 (P<.001) for uric acid, with a positive correlation observed for gamma-glutamyl transferase (GGT; 0.120; P=.001). In contrast, GGT showed a negative correlation in male participants (−0.128; P=.001).

Correlations between FTS-12 and key variables such as age, RDW, albumin, and alkaline phosphatase were also examined. Age and albumin were of particular interest, as they were included in all CRF functions, regardless of sex. For age, the correlation was 0.515 (P<.001) in female participants and 0.476 (P<.001) in male participants. For albumin, the correlation was −0.389 (P<.0001) in female participants and −0.427 (P<.001) in male participants. Despite its moderately high correlation, albumin was not selected as an “individual variable.” Plausible reasons for this exclusion are discussed later in the text.

Tables S3 and S4 in Multimedia Appendix 3 summarize the statistics for the predicted probabilities (hereafter referred to as “the probabilities”) and the independent variables, including individual risk factors and composite functions, incorporated into the selected logistic models for female participants and male participants, respectively. These tables also include information on “age” and several other risk factors. Although these variables were not selected as individual predictors, they are components of the CRF functions named through the selection process.

Table 5 provides descriptive statistics and diagnostic parameters for the selected logistic models. Among the participants, 200 of 735 (27%) in the female participants and 100 of 564 (18%) in the male participants had FTS-12 scores in the fourth quarter (above the third quartile, showing a high probability of frailty). The remaining participants in each group were distributed across the lower 3 quarters.

Tests of the global null hypothesis were highly significant (P<.001), showing that at least 1 independent variable—whether a single risk factor or a composite function of risk factors—was significantly associated with the probability of FTS-12 falling into the uppermost range.

Tables 6 and 7 summarize the estimated parameters and associated statistics for the selected models applied to female and male participants, respectively. The model for female participants includes 5 single variables and 1 composite function of risk factors. In contrast, the model for male participants comprises 7 single variables and 1 composite function of risk factors.

Considering the Wald chi-square statistics, the presence of CRF3, granulocytes, and GGT is highly significant and essential in the model for female participants. For male participants, CRF4, albumin, hemoglobin, triglycerides, and phosphorus are identified as highly significant and critical associated factors. Tables 6 and 7 also present Pearson correlations between each selected independent single variable or composite function of risk factors and the modeled probability. Additionally, these tables include correlations between the modeled probability and both age and FTS-12 scores within the working datasets. In absolute terms, correlations involving composite functions of risk factors exceeded 0.64 (P<.001). Notably, the function CRF3 in female participants showed particularly strong correlations. Among the single risk factors, albumin showed a highly negative correlation in the selected model for male participants.

Figures S3 and S4 in Multimedia Appendix 4 illustrate the functional relationship between the modeled probability and the composite functions of risk factors with the highest correlations. However, these associations were nonlinear. For the logistic model applied to female participants, the relationship between the observed values of the function and the predicted probability followed an inverted logistic S-curve (Figure S3 in Multimedia Appendix 4), with a simple correlation of −0.871 (P<.001). When the value of the function exceeded 2.20, the probability fell below .20 (odds of 1 to 4). Conversely, when values of the function were less than 1.55, the probability exceeded .70 (odds of 7 to 3).

Similarly, for the logistic model in male participants, the relationship between function CRF4 and the modeled probability also followed an inverted logistic S-curve (Figure S4 in Multimedia Appendix 4), with a simple correlation of −0.641 (P<.001). Here, values of the function above 4.75 corresponded to probabilities lower than .20 (odds of 1 to 4), whereas values below 1.50 were associated with probabilities exceeding .70 (odds of 7 to 3). We also estimated the simple correlations between the modeled probability and both patient age and observed FTS-12 values in the original datasets. Age was consistently included in all composite functions of risk factors selected by the selection process. All correlations were highly significant (P<.001), irrespective of sex. Notably, correlations with age were stronger than those with FTS-12 (Tables 6 and 7).

Multimedia Appendix 4 also illustrates the functional relationship between the modeled probability and age in the selected models. For female participants, the scatter plot in Figure S5 in Multimedia Appendix 4 demonstrates that the distribution of values aligns closely with a logistic S-curve. Similarly, for male participants, Figure S6 in Multimedia Appendix 4 exhibits a positive trend, also resembling a logistic S-curve.

Positive correlations between the observed FTS-12 values and the modeled probabilities are clearer in the scatter plots presented in Figures 4 and 5 for male and female participants, respectively. While the dispersion of points in both plots shows a moderate correlation, a clear trend appears, highlighting the prediction capability of the selected models. This consistency in predictive performance is clear across both sexes.

Using the AIC and −2Log L (minus 2 log-likelihood) to assess goodness-of-fit, the model for male participants proved a better fit compared to the model for female participants. Additionally, chi-square statistics derived from the likelihood ratio test provided strong evidence that the independent variables significantly influenced both selected models. The intercept alone was insufficient to explain the observed log-odds—the logarithm of the ratio between the probability of a FTS-12 value being larger than the third quartile of the distribution, and the probability of it not doing so. In addition to goodness-of-fit metrics, ROC curve analysis was used to evaluate the discriminatory performance of the selected logistic models. Tables S5 and S6 in Multimedia Appendix 3 provide statistics from the ROC analysis of the selected logistic models for both sexes. The ROC curves align closely with the goodness-of-fit statistics. For female participants, the AUC was approximately 0.799. Table S5 in Multimedia Appendix 3 and Figure 6 show that the function CRF3 was the most influential risk factor, accounting for 76% of the total AUC and closely approximating the model’s overall AUC. For male participants, the AUC was approximately 0.824. Table S6 in Multimedia Appendix 3 and Figure 7 reveal that CRF4 and albumin were the most influential risk factors, contributing 73% and 71% of the total AUC, respectively.

This exploratory study aimed to identify and refine sex-specific predictive equations for the FTS-12 score using routinely available biochemical and hematological data, including an analytical strategy to estimate the likelihood of belonging to the highest frailty range of the cohort. In line with these aims, the study showed that sex-specific multiple regression models could estimate the FTS-12 score with moderate accuracy, and that sex-specific logistic models could identify predictors associated with being in the upper range of the FTS-12 distribution. Across both approaches, age, albumin, GGT, and several composite variables emerged as recurrent contributors, suggesting that frailty-related laboratory signals are better captured through combinations of markers than through most single biomarkers alone. Taken together, these findings support the feasibility of a low-cost, laboratory-based approach for exploratory frailty risk stratification, but the present models do not constitute validated clinical prediction tools and would require external validation in independent cohorts before any clinical application can be considered.

Our findings suggest that routinely available biochemical and hematological data can provide a meaningful, although still imperfect, estimate of frailty burden as measured by the FTS-12 score, as well as of the likelihood of being in the highest frailty range of the cohort. Some uncertainty in prediction is expected, as the FTS-12 reflects a multidimensional construct integrating 7 dimensions, including energy balance, nutrition, and vascular system components [11], rather than a single biological process. In this context, combinations of laboratory variables performed better than most isolated biomarkers, consistent with the stronger associations observed for the selected composite risk functions in both sexes (Tables 2 and 3 and Figures S3 and S4 in Multimedia Appendix 4). The logarithmic transformations applied to variables such as albumin, age, and hemoglobin should be understood as a way to capture nonlinear clinical relationships and improve model fit, rather than as an end in themselves [21]. For example, combinations involving age and albumin may reflect the joint contribution of aging-related vulnerability and reduced nutritional or systemic reserve, in line with literature linking low albumin and hypoproteinemia to frailty, inflammation, and nutritional vulnerability in older adults [22,23]. Likewise, functions including RDW, hemoglobin, or albumin may capture overlapping signals related to erythrocyte heterogeneity, oxygen-carrying capacity, anemia-related vulnerability, and physiological resilience, consistent with evidence associating RDW and lower hemoglobin status with frailty in older populations [24,25]. In the logistic models, composite risk functions showed the strongest associations with higher frailty probability; and in male participants, albumin also emerged as an important contributor, supporting the interpretation of frailty as a multidimensional state of vulnerability shaped by interacting biological processes.

The models explained about 40% of the variability in FTS-12 scores, indicating moderate predictive performance. This level of explanation is clinically meaningful for an initial screening-oriented approach, while also showing that a substantial part of frailty remains influenced by factors not captured in laboratory-based models—a multidimensional syndrome reflecting biological, functional, cognitive, psychosocial, and environmental factors. The unexplained variance is therefore expected and reinforces that these models should be viewed as complementary tools rather than substitutes for comprehensive geriatric assessment, with future improvements depending on incorporating additional clinical, behavioral, and social determinants of frailty. Notably, albumin and age emerged as key associated factors consistently included across all composite functions, with logarithmic transformations capturing nonlinear relationships more effectively than models based solely on individual risk factors.

Recent literature includes only a few studies that model frailty risk using biochemical traits in older adults [21,23,26]. Several recent works have developed statistical models for frailty and frailty-associated outcomes such as mortality, most of them using logistic regression [27-30]. For example, Bencivenga et al [31] examined a FI composed of 38 multidimensional health deficits—including comorbidities, laboratory and diagnostic data, and symptoms and signs of disease—using procedures outlined by Searle et al [32]; the independent variables in their multiple regression model included sex (binary), age, BMI, and muscle thickness of the vastus intermedius and rectus femoris. Logistic regression is a widely used technique for modeling binary dependent variables in this setting [33-35]. Wang et al [36] provide a recent example using logistic regression and ROC curve analysis to compare 2 FIs in classifying binary mortality outcomes: the first based on routine physical assessments and laboratory tests, the second on self-reported data encompassing 30 health problems (“health deficits”) such as diseases, symptoms, disabilities, and psychological issues. In their study, FI values were used as independent variables to model the log-likelihood of dying. The clinical relevance of laboratory-based frailty stratification was further supported by Pasternak et al [37], who demonstrated that a laboratory-based FI was significantly associated with all-cause mortality in hospitalized older adults.

In contrast, our study used FTS-12 measurements to generate a new binary dependent variable, categorizing scores as exceeding the third quartile of the distribution or not. This allowed us to model the likelihood that an observed FTS-12 value would fall within the uppermost range, with logistic regression then applied to assess the log-likelihood of these outcomes. Liu et al [38](2021) also used logistic regression, but with a different approach: a binary outcome representing the presence or absence of frailty in older patients with nutritional risks, with frailty status determined from previous research. However, Liu et al [38] focused on exogenous (environmental) predictors such as age, sex, ethnicity, educational level, BMI, walking ability, activities of daily living, and cognitive functioning, whereas our study prioritized “endogenous” biochemical variables to identify both the “most reasonable” model and the “most influential” risk factors.

Álvarez-Sánchez et al [39] used logistic regression to evaluate the association between 2 inflammatory biomarkers—high-sensitivity C-reactive protein and homocysteine—and prefrailty in Spanish adults aged 65 years and older [3], underscoring the role of inflammation in age-related physical decline. Although neither biomarker was selected in our models, our findings on GGT and albumin are consistent with emerging geriatric evidence. In a Korean population-based cohort, Kang et al [26] reported that elevated GGT levels were 26% higher in frail male participants than in nonfrail male participants, with higher GGT quartiles showing a 2.08-fold increased odds of frailty. Elevated GGT has also been associated with poorer muscle function—including reduced jump power and diminished handgrip strength—independent of body-composition factors [40], and with increased risk of hip fracture in postmenopausal female participants, irrespective of alcohol consumption or chronic liver disease [41]. Because both diminished handgrip strength and elevated risk of hip fracture are strongly associated with frailty in older adults, these observations support the potential relevance of GGT in frailty risk estimation [3,11]. A recent systematic review further emphasized that hypoalbuminemia is associated with a significant risk of frailty and adverse health consequences in hospital and community contexts [24]. Together, these findings strengthen the biological plausibility of our results, suggesting that GGT and albumin reflect convergent pathways of oxidative stress, malnutrition, and chronic inflammation that are central to geriatric syndromes [42].

From a model-performance perspective, the diagnostic criteria used to compare candidate logistic models suggested acceptable exploratory fit [20,33,43]. More importantly, from a clinical standpoint, the models showed that higher frailty probability was better captured by combinations of laboratory variables than by most isolated biomarkers alone. This supports the idea that frailty-related vulnerability reflects interacting biological domains rather than a single abnormal test result and helps explain why composite terms contribute more consistently than most individual variables. Notably, the strongest correlations were seen between CRF functions and both FTS-12 values and the associated probabilities, regardless of sex. However, substantial dispersion in observed FTS-12 values persisted for each value of the composite risk functions, highlighting the complex interplay of factors across the 7 dimensions underlying frailty assessment.

Placed in the context of existing approaches for laboratory-based frailty screening, the discriminatory performance reported in the Results section falls within the moderate-to-good range typically described for tools that rely exclusively on routine biochemical and hematological variables [11,21]. AUC values close to 0.80 for both sexes are consistent with what may reasonably be expected when functional, cognitive, and self-reported components are not incorporated, and they reinforce the view that the present models should be regarded as a complementary first layer of risk stratification rather than as a stand-alone diagnostic instrument. Beyond their overall discriminatory performance, the biological composition of the dominant composite functions is also informative: the prominence of the albumin-age composite in female participants is consistent with the well-established association of hypoalbuminemia and anemia with adverse outcomes in older adults [24,25], while the relevance of the hemoglobin-age composite in male participants is compatible, with erythropoietic decline acting as a sex-sensitive marker of physiological reserve. The recurrent involvement of GGT and alkaline phosphatase further suggests that subclinical hepatobiliary and oxidative pathways contribute to the laboratory signature of frailty [26], in line with the broader role of inflammatory, nutritional, and oxidative mechanisms previously implicated in the syndrome [3,39,42].

Variable selection was conducted using stepwise multiple regression guided by the AIC and Mallows C_p statistic. Although stepwise procedures have recognized limitations, this approach was used for exploratory purposes within a proof-of-concept framework, and the results should therefore be interpreted with appropriate caution.

This study has several limitations that should temper the interpretation of its findings. First, its cross-sectional design prevents causal inference and may be susceptible to reverse causation; for example, elevated GGT could reflect unmeasured liver disease or alcohol intake rather than etiological pathways to frailty [26,44]. Second, all biochemical measurements were obtained at a single baseline visit, so temporal fluctuations and dynamic changes in frailty status were not evaluated. Third, the cohort was drawn from a single European population (Toledo, Spain), which may limit external validity in ethnically diverse or resource-limited settings. Finally, model development and internal validation were performed using the same dataset, which may increase the risk of overfitting despite the application of bootstrapping procedures. Future studies should therefore prioritize external or independent validation to assess model performance on unseen data and the generalizability of the proposed models.

Furthermore, rather than serving as a general clinical cutoff for the FTS-12 scale, the binary logistic analysis based on the upper quartile should be understood as an internal analytical technique to identify the subgroup with the highest frailty burden within this cohort. We recognize that dichotomization may reduce information and affect apparent model discrimination. The current proof-of-concept study did not include formal calibration analyses (calibration plots, calibration slope/intercept, and Brier score), threshold-specific sensitivity and specificity, decision-curve analysis, or net-benefit evaluation, although discrimination was investigated using ROC curve analysis and multicollinearity was confirmed to be nonproblematic using VIFs.

As noted in the Methods section, calibration in the probabilistic sense did not apply to the multiple regression model, for which performance was instead evaluated through goodness-of-fit measures and residual diagnostics. For the logistic models, formal calibration was not assessed. The present models should therefore be interpreted as exploratory, discrimination-based tools to determine whether routinely available laboratory data contain meaningful frailty-related signals, rather than as deployment-ready clinical prediction instruments. Further development toward clinical implementation will require calibration-focused evaluation, external validation, and implementation-oriented testing. Although bootstrap procedures were used to estimate CIs for selected model parameters, this does not constitute a full internal validation; optimism-corrected evaluation of overall model performance was beyond the scope of this secondary exploratory analysis.

More stringent development and validation techniques—penalized regression (eg, least absolute shrinkage and selection operator), repeated cross-validation, formal calibration evaluation, clinical utility analyses, and benchmarking against parsimonious and machine-learning alternatives—should be incorporated into model refinement in later phases. These procedures were outside the scope of the current work because the study was not preregistered and was intended as a secondary exploratory proof-of-concept analysis based on an existing cohort dataset. Complete-case analysis was used due to low missingness, and sex-stratified modeling was given precedence over formal interaction testing. Variables such as alcohol use, liver disease, medication exposure, inflammatory markers, nutritional variables, and comorbidity burden were not available in sufficiently complete form for inclusion, so part of the observed signal for biomarkers such as GGT and albumin may reflect unmeasured clinical contexts rather than direct frailty-specific pathways. The present revision does not yet provide TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis)-compliant reporting, optimism-corrected full model validation, or direct benchmarking against simpler baseline models (eg, age-only or age-plus-albumin), all of which should be incorporated in future validation-oriented refinement.

Our proof-of-concept analysis supports the viability of this modeling approach, and the next logical step is model refinement. A prospective, multicenter cohort with repeated biomarker assessments and adjudicated clinical outcomes would provide a more appropriate framework to confirm temporality and improve generalizability. External validation in Latin American populations, integration of additional biological data layers, and comparison with machine-learning approaches may further improve predictive performance. Because the proposed models rely on routinely available laboratory parameters, they are particularly suitable for implementation in primary care and resource-limited settings, where they could help prioritize older adults for further geriatric assessment when specialized frailty evaluations are not immediately available. Embedding the frailty-probability equations within electronic health record dashboards or mobile screening tools could enable real-time, low-cost risk stratification in primary care and remote geriatric care. By facilitating earlier identification of risk patterns from routinely available laboratory data, this approach may contribute to more individualized and sex-sensitive frailty management strategies, consistent with the broader goal of precision geriatric care.

Recent discussions in clinical AI emphasize that real-world implementation requires explainability, fairness, human oversight, and careful workflow integration, so computational models should be developed to augment clinicians through earlier detection, smarter triage, and more targeted prevention rather than replace clinical judgment [45]. Likewise, work in adjacent computational health fields reinforces the need for interpretable pipelines, rigorous validation, and multicohort generalizability before routine implementation [46]. Taken together, these considerations position the present work not only as a statistical proof of concept but also as an early translational step toward clinically embedded, digitally supported frailty screening in aging care.

From routinely available biochemical and hematological data, frailty, operationalized using the FTS-12 score, was modeled with moderate accuracy; sex-specific models demonstrated consistent predictive performance in both male and female participants. The final models included albumin, age, GGT, and a few composite risk functions that are biologically plausible correlates of frailty-related risk. For transparency and future validation, the final equations (albumin in g/dL and GGT in U/L) are given in Multimedia Appendix 1. Currently, these models should not be viewed as nomograms or screening tools with clinically useful biomarker thresholds, but rather as instruments for continuous frailty-risk estimation.

The models offer clinically interpretable signals that could support future workflow-integrated frailty risk stratification, despite the modest explained variance. This strategy may support digitally enabled screening, earlier frailty-oriented assessment, and prioritization for comprehensive geriatric assessment in primary care, geriatric care, and remote care settings following appropriate external and prospective validation. Rather than establishing definitive clinical decision rules, this work was designed to evaluate feasibility, computational stability, and interpretability as a secondary exploratory proof-of-concept analysis based on an existing cohort. Before regular clinical use, additional validation, calibration, and implementation-focused research will be necessary.