Development and temporal validation of a machine learning risk prediction model for anemia in older adults: A NHANES study of an inflammatory–nutritional–renal profile

Development and temporal validation of a machine learning risk prediction model for anemia in older adults: A NHANES study of an inflammatory–nutritional–renal profile | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and temporal validation of a machine learning risk prediction model for anemia in older adults: A NHANES study of an inflammatory–nutritional–renal profile Zhiyuan Niu, Lei Ren, Naibo Hu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9486902/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Anemia is common in older adults and is often associated with chronic inflammation, poor nutritional reserve, and renal dysfunction. However, risk prediction models for anemia in community-dwelling older adults based on nationally representative data remain limited. This study aimed to develop and temporally validate a machine learning model for predicting anemia risk in older adults using NHANES data, and to assess its interpretability and potential clinical applicability. Methods We used data from the National Health and Nutrition Examination Survey (NHANES). Participants aged ≥ 60 years from the 2015–2016 and 2017–2018 cycles were combined as the development cohort, and those from the 2021–2023 cycle served as the temporal validation cohort. Least absolute shrinkage and selection operator regression was used for feature selection. Five models were developed and compared: CatBoost, XGBoost, random forest, support vector machine, and logistic regression. Model performance was evaluated in terms of discrimination, calibration, and clinical utility, and SHapley Additive exPlanations were used to interpret the best-performing nonlinear model. Results A total of 6,311 older adults were included, of whom 3,665 were in the development cohort and 2,646 were in the temporal validation cohort. Eight predictors were retained: age, race/ethnicity, poverty-income ratio, estimated glomerular filtration rate, blood urea nitrogen, serum albumin, diabetes status, and high-sensitivity C-reactive protein. CatBoost showed the best overall performance, with area under the curve values of 0.834 in the development cohort and 0.773 in the temporal validation cohort. Logistic regression achieved a comparable area under the curve of 0.770 in the temporal validation cohort and showed good calibration (slope 0.965; intercept − 0.069). SHAP analysis identified race/ethnicity, albumin, age, estimated glomerular filtration rate, and blood urea nitrogen as major contributors to model prediction, with albumin emerging as the most important protective correlate. Conclusions Anemia risk in older adults may be understood as a composite susceptibility pattern shaped by inflammatory, nutritional, renal, socioeconomic, and age-related factors. CatBoost provided the best predictive performance, whereas logistic regression and its nomogram offered greater interpretability and potential for clinical translation. A prediction tool based on routinely available indicators may support early risk identification and stratification of anemia in older adults. older adults anemia NHANES temporal validation inflammatory–nutritional–renal profile clinical interpretability Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Background Anemia is a common yet long-underrecognized clinical problem among older adults. Systematic reviews and meta-analyses indicate that the prevalence of anemia is high in older adults and is even higher among hospitalized individuals and those in long-term care settings [1,2]. Global burden of disease data further suggest that anemia remains an important cause of disability worldwide [3]. Moreover, anemia in later life is closely associated with functional decline, reduced muscle strength, increased risk of hospitalization, and poor prognosis [4,5]. Therefore, anemia in older adults is not merely a laboratory abnormality, but an important geriatric health issue with public health and clinical management implications. Anemia in older adults is rarely driven by a single mechanism and more often reflects the combined effects of chronic low-grade inflammation, diminished nutritional reserve, and declining renal function [6–8]. Inflammation may promote anemia through hepcidin-mediated iron restriction and suppression of erythropoiesis [6,9]. Low albumin not only indicates inadequate nutritional reserve but may also reflect a higher chronic disease burden and impaired overall health status [8,10]. Meanwhile, renal aging and chronic kidney disease, with their associated reduction in erythropoietin production, constitute another important pathological basis of anemia in older adults [7,11]. Accordingly, identifying older adults at high risk of anemia from the integrated perspective of an inflammatory–nutritional–renal profile is biologically plausible and potentially clinically valuable. Despite the growing burden of anemia in older adults, current approaches to risk identification still rely largely on single indicators or traditional statistical analyses and therefore may fail to capture the complex nonlinear relationships among multidimensional clinical features. In recent years, machine learning has shown considerable promise in clinical prediction modeling through its capacity for high-dimensional data integration, complex pattern recognition, and individualized risk assessment [12,13]. Several studies have explored machine learning models for anemia prediction in outpatient or hospitalized older populations [14,15]; however, studies based on nationally representative community-dwelling older adults that simultaneously integrate inflammatory, nutritional, and renal indicators and incorporate temporal validation remain limited. Therefore, using NHANES data, we developed and temporally validated a risk prediction model for anemia in older adults, identified key predictors, and assessed its clinical interpretability and potential clinical applicability. Methods Study design and data source This study was based on publicly available data from the National Health and Nutrition Examination Survey (NHANES). NHANES uses a complex multistage, stratified probability sampling design and systematically collects demographic information, anthropometric measurements, laboratory results, and health-related data, thereby providing good national representativeness [16]. According to the prespecified study design, data from the 2015–2016, 2017–2018, and 2021–2023 cycles were included. The 2015–2016 and 2017–2018 cycles were combined as the development cohort for feature selection, model training, and internal optimization, whereas the 2021–2023 cycle served as the temporal validation cohort to evaluate model generalizability across different time periods. The manuscript was prepared in accordance with the TRIPOD + AI reporting framework [17]. The NHANES study protocol was approved by the Ethics Review Board of the National Center for Health Statistics (NCHS), and written informed consent was obtained from all participants [18]. Participant selection and outcome definition Participant selection and outcome definition Participants aged ≥ 60 years in NHANES with available hemoglobin measurements were included. Exclusion criteria were: (1) age < 60 years; and (2) missing hemoglobin data. The detailed sample selection process is shown in Fig. 1 . Anemia was defined using sex-specific hemoglobin thresholds: hemoglobin < 13.0 g/dL in men and < 12.0 g/dL in women [19]. Candidate predictors and data preprocessing Based on the clinical pathophysiology of anemia in older adults and data availability in NHANES, candidate variables initially included demographic features, inflammatory markers, nutrition-related indicators, renal function markers, and comorbidity status. To balance sample retention, model robustness, and clinical interpretability, a prespecified tiered missing-data strategy was used. For covariates with a missing rate < 5%, median imputation was used for continuous variables and mode imputation for categorical variables. For variables with 5%–20% missingness, multiple imputation by chained equations (MICE) was prioritized, with sensitivity analyses based on simple imputation when necessary. Variables with > 20% missingness were generally excluded from the main model unless they had clear clinical relevance and strong theoretical importance, in which case they were considered only in supplementary or sensitivity analyses [20,21]. Given the complex sampling design of NHANES, descriptive statistics and between-group comparisons incorporated sampling weights, strata, and cluster variables to obtain nationally representative estimates. Model development and temporal validation focused on individual-level risk stratification and discrimination; therefore, model training and testing were performed after unified data preprocessing without additional survey weighting [16]. Feature selection and model development In the development cohort, 13 candidate variables were prespecified for least absolute shrinkage and selection operator (LASSO) regression based on the clinical relevance of anemia in older adults and the availability of NHANES data. These variables included age, sex, race/ethnicity, poverty-income ratio, body mass index (BMI), estimated glomerular filtration rate (eGFR), blood urea nitrogen (BUN), serum albumin, diabetes status, hypertension status, smoking status, neutrophil-to-lymphocyte ratio (NLR), and high-sensitivity C-reactive protein (hs-CRP). Together, these variables represented demographic, socioeconomic, inflammatory–nutritional–renal, comorbidity, and lifestyle domains. Ten-fold cross-validation was then used to determine the optimal penalty parameter λ and to identify the predictors most strongly associated with anemia risk in older adults [22]. Under the optimal penalty parameter, LASSO retained 8 variables for subsequent model construction and effect estimation. Five prediction algorithms were then developed and compared, including CatBoost, XGBoost, random forest (RF), support vector machine (SVM), and logistic regression (LR) [23–26]. All models underwent hyperparameter tuning using 5-fold cross-validation within the development cohort and were subsequently evaluated in the temporal validation cohort for generalizability. Model evaluation, interpretation, and clinical translation Model performance was evaluated from three aspects: discrimination, calibration, and net clinical benefit. Discrimination was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC); calibration was evaluated using calibration curves; and clinical utility was assessed using decision curve analysis (DCA) [27,28]. To emphasize clinical interpretability, we also reported adjusted effect estimates from multivariable logistic regression and constructed a nomogram based on the logistic model as a potential tool for clinical translation. To improve the interpretability of the best-performing nonlinear model, we further applied the SHapley Additive exPlanations (SHAP) framework to quantify the contribution of each variable to anemia risk prediction and generated both SHAP summary and dependence plots [29]. Thus, model presentation in this study simultaneously emphasized predictive performance, temporal validation stability, and clinical interpretability. Results Study population and flow The participant selection process is shown in Figure 1. In the development cohort (2015–2018), 19,225 participants were initially assessed, including 9,971 from the 2015–2016 cycle and 9,254 from the 2017–2018 cycle. After excluding 15,174 individuals aged <60 years and 386 with missing hemoglobin data, 3,665 participants were ultimately included for model development, including 1,725 from 2015–2016 and 1,940 from 2017–2018. In the temporal validation cohort (2021–2023), 11,933 participants were initially assessed. After excluding 8,429 aged <60 years and 858 with missing hemoglobin data, 2,646 participants were included. The final analytic sample comprised 6,311 older adults. Figure 1. Flowchart of participant selection across NHANES cycles. Baseline characteristics Baseline characteristics of the development and temporal validation cohorts are shown in Table 1. Weighted analyses indicated that the two cohorts were generally comparable with respect to the main demographic characteristics. The mean age was 69.67 ± 6.75 years in the development cohort and 69.74 ± 6.58 years in the temporal validation cohort, with no significant difference (P=0.839, SMD=0.010). Similarly, no significant between-cohort differences were observed for sex distribution (P=0.465, SMD=0.017) or poverty-income ratio (3.12 ± 1.59 vs. 3.11 ± 1.57, P=0.943, SMD=0.006). For laboratory and clinical indicators, BMI (P=0.305, SMD=0.044), eGFR (P=0.564, SMD=0.025), hs-CRP (P=0.085, SMD=0.045), and neutrophil-to-lymphocyte ratio (NLR; P=0.159, SMD=0.046) did not differ significantly between cohorts. By contrast, the temporal validation cohort had slightly lower albumin levels than the development cohort (4.02 ± 0.32 vs. 4.14 ± 0.33 g/dL, P<0.001, SMD=0.360), and BUN also showed a mild difference (16.00 [13.00–20.00] vs. 17.00 [14.00–21.00] mg/dL, P=0.019, SMD=0.083). In addition, the SMD for race/ethnicity distribution was 0.106, suggesting mild structural heterogeneity, although the overall distribution did not differ significantly (P=0.413). The prevalence of anemia was not significantly different between the two cohorts (9.8% vs. 11.3%, P=0.165, SMD=0.052). Overall, the cohorts remained largely comparable across most baseline features while retaining limited temporal heterogeneity. Table 1. Weighted baseline characteristics of the development and temporal validation cohorts. Values are presented as mean ± SD, median [IQR], or n (weighted %). P values were calculated using survey-weighted t tests, survey-weighted rank tests, or Rao–Scott chi-square tests, as appropriate. SMD > 0.10 was considered indicative of meaningful imbalance. For multi-category variables, SMDs were summarized as the maximum absolute level-specific standardized mean difference. Abbreviations: BMI, body mass index; hs-CRP, high-sensitivity C-reactive protein; NLR, neutrophil-to-lymphocyte ratio; eGFR, estimated glomerular filtration rate; IQR, interquartile range; SMD, standardized mean difference. Variable selection and multivariable association In the development cohort, LASSO regression included 13 prespecified candidate variables, namely age, sex, race/ethnicity, poverty-income ratio, BMI, eGFR, BUN, albumin, diabetes, hypertension, smoking status, NLR, and hs-CRP. The coefficient profiles and cross-validation results indicated that 8 variables with non-zero coefficients were ultimately retained (Figure 2A–B): age, race/ethnicity, poverty-income ratio, eGFR, BUN, serum albumin, diabetes status, and hs-CRP. Sex, BMI, hypertension, smoking status, and NLR were not retained in the final model. These 8 retained variables were subsequently entered into a multivariable logistic regression model to estimate their adjusted effects within the same model (Figure 2C). Overall, Figure 2 illustrates the sequential analytical process of candidate-variable shrinkage and selection, final variable retention, and effect estimation, suggesting that the risk structure of anemia in older adults is primarily organized around an inflammatory–nutritional–renal profile, with albumin showing the strongest protective association. Figure 2. LASSO-based feature selection and multivariable logistic regression results. (A) Coefficient profiles of the 13 candidate predictors in the LASSO model. (B) Ten-fold cross-validation for selection of the optimal penalty parameter λ. (C) Adjusted odds ratios and 95% confidence intervals for the 8 variables retained by LASSO and entered into the multivariable logistic regression model. Comparison of model performance The predictive performance of the five machine learning algorithms is shown in Table 2 and Figure 3. In ROC analyses, XGBoost and random forest demonstrated high discrimination in the development cohort, with AUCs of 0.883 and 1.000, respectively. However, the AUC of random forest declined to 0.745 in the temporal validation cohort, suggesting substantial overfitting. By contrast, CatBoost achieved AUCs of 0.834 in the development cohort and 0.773 in the temporal validation cohort, showing the best performance in the validation set and slightly outperforming logistic regression (AUC=0.770) and XGBoost (AUC=0.758), thus indicating better generalization stability. In addition, CatBoost yielded a negative predictive value of 0.937 in the temporal validation cohort, suggesting good performance in identifying low-risk individuals. Considering discrimination, temporal validation stability, and model interpretability together, CatBoost was identified as the best-performing nonlinear model in this study and was therefore selected for subsequent SHAP interpretation. Logistic regression, in contrast, was considered more suitable as the foundation for a clinically translatable tool because it provides more direct clinical interpretability. Figure 3. Receiver operating characteristic curves in the development and temporal validation cohorts. Table 2. Predictive performance of candidate models in the development and temporal validation cohorts. SHAP-based interpretation of the best-performing model Given that CatBoost showed the best and relatively stable overall performance in the temporal validation cohort, SHAP analysis was further performed to interpret its prediction process (Figure 4A–B). The SHAP summary plot showed that race/ethnicity, albumin, age, eGFR, and BUN were the main contributors to model output; among the laboratory indicators, albumin was the most important. The SHAP dependence plot further indicated that lower albumin levels were associated with a markedly greater positive contribution to predicted anemia risk. This finding further supports the view that anemia risk in older adults is not determined by any single marker, but more closely resembles a susceptibility pattern driven by an integrated inflammatory–nutritional–renal profile. Figure 4. SHAP-based interpretation of the best-performing CatBoost model. Nomogram construction and clinical utility Decision curve analysis further showed that, across a relatively wide range of threshold probabilities, risk stratification based on this model provided greater net benefit than strategies of treating all or treating none (Figure 5). To improve clinical usability, we further constructed a nomogram based on the core variables (Figure 6A). Calibration curves showed good agreement between predicted and observed risk in the temporal validation cohort (Figure 6B), with a calibration slope of 0.965 and an intercept of −0.069, indicating good consistency between predicted probabilities and observed outcomes. Because logistic regression preserves a more direct path of parameter interpretation, this model offers advantages in terms of clinical interpretability and potential clinical translation. These findings suggest potential clinical value and support the use of the model for early risk identification in older adults. Discussion Using nationally representative NHANES data from older adults, we developed and temporally validated a risk prediction model for anemia that integrates inflammatory, nutritional, renal, and socioeconomic information. The main findings can be summarized in three points. First, among the 13 candidate variables prespecified for LASSO, 8 core features were ultimately retained, yielding a risk structure centered on an inflammatory–nutritional–renal profile and further shaped by age and socioeconomic vulnerability; among them, albumin showed the strongest protective association in multivariable logistic regression. Second, CatBoost maintained good discrimination in both the development and temporal validation cohorts, suggesting relatively stable generalization across time-separated samples of older adults. Third, SHAP analysis further showed that race/ethnicity, albumin, and related variables contributed substantially to model output, whereas logistic regression achieved an AUC close to that of CatBoost in the validation cohort while demonstrating better calibration and stronger clinical interpretability. Overall, these findings are consistent with prior work viewing anemia in older adults as a geriatric syndrome and emphasizing its close links with frailty and adverse outcomes [29–31]. From a pathophysiological perspective, our findings further support the view that anemia in older adults is not driven by a single mechanism, but more likely arises from the combined effects of inflammation, diminished nutritional reserve, declining renal function, and age-related physiological vulnerability. In this context, the final model does not capture one specific etiologic signal; rather, it more closely reflects a multidimensional anemia-prone phenotype [29,30,32]. Among the retained variables, albumin likely represents more than nutritional status alone. As a negative acute-phase reactant, lower albumin may also indicate increased chronic inflammatory burden, frailty, and reduced overall physiological reserve, which may explain its prominent contribution in both multivariable logistic regression and SHAP analysis [33]. In addition, reviews on nutrition-related anemia indicate that deficiencies in iron, folate, and vitamin B12 remain important contributors to anemia in older adults, while low albumin can serve as a clinical surrogate for nutritional risk and insufficient overall health reserve [34]. The inclusion of hs-CRP in the final model further supports the potential role of low-grade systemic inflammation in anemia among older adults, with plausible mechanisms involving inflammation-related iron restriction, hepcidin-mediated dysregulation of iron homeostasis, and suppression of erythropoiesis [35,36]. The renal function indicators identified in this study are likewise biologically plausible. Reduced eGFR and elevated BUN not only indicate renal impairment, but may also reflect insufficient erythropoietin production, suppressed erythroid activity, and shortened red blood cell survival, all of which are important mechanisms underlying anemia in older adults [37]. In addition, the retention of poverty-income ratio and race/ethnicity suggests that anemia risk in older adults is influenced not only by biological vulnerability but also by broader social vulnerability. These variables may, to some extent, reflect differences in nutritional access, chronic disease burden, healthcare utilization, and the cumulative impact of long-standing structural disadvantage. Previous population-based research has also suggested that education level, marital status, residential environment, and other socioeconomic factors are associated with the distribution of anemia among older adults [38]. Taken together, our model integrates both biological and social dimensions of susceptibility to anemia, which may partly explain why it retained relatively stable predictive performance in time-separated cohorts [32]. At the modeling level, the present findings have important translational implications. Recent studies have shown growing potential for machine learning in geriatric medicine and anemia screening [12–15]. However, for clinical research, the best-performing model is not necessarily the model most suitable for implementation. In the present study, CatBoost showed a slight advantage in discrimination, and we therefore further applied SHAP analysis after model comparison to reveal its key driving features. Logistic regression, however, performed very similarly in the later cycle and offered greater transparency, parameter interpretability, and ease of tool construction. Based on these findings, we present CatBoost as the main high-performance prediction model while also constructing a nomogram based on logistic regression. This parallel strategy of a performance-oriented model together with an interpretable model is better aligned with real-world clinical demands for accuracy, temporal validation stability, and clinical interpretability. The strengths of this study include the use of nationally representative NHANES data from older adults, the inclusion of routinely available clinical indicators, and the use of an independent subsequent survey cycle as a temporal validation cohort, all of which enhance the robustness and practical relevance of the findings to some extent. Nevertheless, several limitations should be acknowledged. First, NHANES is a cross-sectional survey, and the present study identifies a pattern of risk associated with coexisting anemia rather than strict causal relationships. Second, the validation data were drawn from different cycles within the same national survey system; thus, while this represents independent temporal validation, it cannot substitute for genuine external population-based validation. Third, some indicators closely related to etiologic subtyping of anemia, such as ferritin, transferrin saturation, vitamin B12, and folate, were not consistently available across cycles, limiting the model’s ability to further characterize anemia subtypes. Finally, machine learning models applied to complex survey data still require cautious interpretation with regard to modeling and generalizability. Future work should therefore validate model performance in external cohorts, prospective studies, and datasets containing more complete etiologic indicators, and explore whether the model improves clinical outcomes and decision-making. Conclusions Using NHANES data from older adults, we established and temporally validated a risk prediction model for anemia integrating an inflammatory–nutritional–renal profile. The findings suggest that anemia risk in older adults is not driven by a single factor, but is better understood as a composite susceptibility phenotype shaped jointly by inflammation, insufficient nutritional reserve, declining renal function, and socioeconomic vulnerability. Among the predictors, albumin emerged as the most important protective correlate. CatBoost demonstrated relatively stable predictive performance, whereas logistic regression and its nomogram offered stronger clinical interpretability. A prediction tool based on routinely available indicators may support the early identification and risk stratification of anemia in older adults, although further external validation and prospective research are still needed. Abbreviations AUC: Area under the curve BMI: Body mass index BUN: Blood urea nitrogen DCA: Decision curve analysis eGFR: Estimated glomerular filtration rate hs-CRP: High-sensitivity C-reactive protein IQR: Interquartile range LASSO: Least absolute shrinkage and selection operator LR: Logistic regression MICE: Multiple imputation by chained equations NCHS: National Center for Health Statistics NHANES: National Health and Nutrition Examination Survey NLR: Neutrophil-to-lymphocyte ratio RF: Random forest ROC: Receiver operating characteristic SHAP: SHapley Additive exPlanations SMD: Standardized mean difference SVM: Support vector machine TRIPOD+AI: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence XGBoost: Extreme Gradient Boosting Declarations Ethics approval and consent to participate The present study was based on publicly available, de-identified data from the National Health and Nutrition Examination Survey (NHANES). The NHANES study protocol was approved by the Ethics Review Board of the National Center for Health Statistics (NCHS), and written informed consent was obtained from all participants by NHANES. Because this study involved secondary analysis of publicly available de-identified data, additional ethical approval was not required. All methods were carried out in accordance with relevant guidelines and regulations. Clinical trial registration Not applicable. This study was a secondary analysis of publicly available, de-identified NHANES data and did not involve assignment of participants to interventions or registration as a clinical trial. Consent for publication Not applicable. This manuscript does not contain any individual person’s identifiable data in any form. Availability of data and materials The data that support the findings of this study are publicly available from the National Health and Nutrition Examination Survey (NHANES). Analytic code is available from the corresponding author on reasonable request. Competing interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Authors’ contributions Zhiyuan Niu conceived and designed the study. Lei Ren performed data curation, formal analysis, and visualization. Zhiyuan Niu drafted the manuscript. Naibo Hu and Zhiyuan Niu critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript. Acknowledgements Not applicable. Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this manuscript, the authors used AI-assisted language tools to support translation and language polishing. After using these tools, the authors carefully reviewed, revised, and edited the manuscript as needed and take full responsibility for the content of the final submitted version. References 1. Mohammadi A, Kazeminia M, Chogan A, Jalali A. Prevalence of anemia in older adults: A systematic and meta-analysis study. Int J Afr Nurs Sci. 2024;20:100739. doi:10.1016/j.ijans.2024.100739. 2. Guralnik JM, Eisenstaedt RS, Ferrucci L, Klein HG, Woodman RC. Prevalence of anemia in persons 65 years and older in the United States: evidence for a high rate of unexplained anemia. Blood. 2004;104(8):2263–2268. doi:10.1182/blood-2004-05-1812. 3. GBD 2021 Anaemia Collaborators. Prevalence, years lived with disability, and trends in anaemia burden by severity and cause, 1990–2021: findings from the Global Burden of Disease Study 2021. Lancet Haematol. 2023;10(9):e713-e734. doi:10.1016/S2352-3026(23)00160-6. 4. Culleton BF, Manns BJ, Zhang J, et al. Impact of anemia on hospitalization and mortality in older adults. Blood. 2006;107(10):3841–3846. doi:10.1182/blood-2005-10-4308. 5. Penninx BWJH, Pahor M, Cesari M, et al. Anemia is associated with disability and decreased physical performance and muscle strength in the elderly. J Am Geriatr Soc. 2004;52(5):719–724. doi:10.1111/j.1532-5415.2004.52208.x. 6. Wacka E, Nicikowski J, Jarmuzek P, Zembron-Lacny A. Anemia and Its Connections to Inflammation in Older Adults: A Review. J Clin Med. 2024;13(7):2049. doi:10.3390/jcm13072049. 7. Santos S, Lousa I, Carvalho M, Sameiro-Faria M, Santos-Silva A, Belo L. Anemia in Elderly Patients: Contribution of Renal Aging and Chronic Kidney Disease. Geriatrics (Basel). 2025;10(2):43. doi:10.3390/geriatrics10020043. 8. Stauder R, Valent P, Theurl I. Anemia at older age: etiologies, clinical implications, and management. Blood. 2018;131(5):505–514. doi:10.1182/blood-2017-07-746446. 9. Ganz T. Anemia of inflammation. N Engl J Med. 2019;381(12):1148–1157. doi:10.1056/NEJMra1804281. 10. Riviati N, Legiran, Indrajaya T, Saleh I, Ali Z, Irfannuddin, et al. Serum Albumin as Prognostic Marker for Older Adults in Hospital and Community Settings. Gerontol Geriatr Med. 2024;10:23337214241249914. doi:10.1177/23337214241249914. 11. Koury MJ, Haase VH. Anaemia in kidney disease: harnessing hypoxia responses for therapy. Nat Rev Nephrol. 2015;11(7):394–410. doi:10.1038/nrneph.2015.82. 12. Stahl D. New horizons in prediction modelling using machine learning in older people’s healthcare research. Age Ageing. 2024;53(9):afae201. doi:10.1093/ageing/afae201. 13. Olawade DB, Adeniji YJ, Olatunbosun FA, Egbon E, David-Olawade AC. Artificial intelligence for anemia screening, diagnosis, and management. Curr Res Transl Med. 2025;74(1):103560. doi:10.1016/j.retram.2025.103560. 14. Göl M, Aktürk C, Talan T, Vural MS, Türkbeyler İH. Predicting malnutrition-based anemia in geriatric patients using machine learning methods. J Eval Clin Pract. 2025;31(2):e14142. doi:10.1111/jep.14142. 15. Bi S, Meng G, Wei N, Zhou Q, Wang C. Influential factors and predictive model of anemia in hospitalized elderly patients. Sci Rep. 2025;15(1):35324. doi:10.1038/s41598-025-19397-6. 16. Centers for Disease Control and Prevention. NHANES Tutorials: Sample Design Module. Available from: https://wwwn.cdc.gov/nchs/nhanes/tutorials/sampledesign.aspx. Accessed April 11, 2026. 17. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378. 18. Centers for Disease Control and Prevention. Ethics Review Board Approval. Available from: https://www.cdc.gov/nchs/nhanes/about/erb.html. Accessed April 11, 2026. 19. World Health Organization. Guideline on haemoglobin cutoffs to define anaemia in individuals and populations. Geneva: World Health Organization; 2024. 20. Sterne JAC, White IR, Carlin JB, et al. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. BMJ. 2009;338:b2393. doi:10.1136/bmj.b2393. 21. van Buuren S, Groothuis-Oudshoorn K. mice: Multivariate imputation by chained equations in R. J Stat Softw. 2011;45(3):1–67. doi:10.18637/jss.v045.i03. 22. Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Series B. 1996;58(1):267–288. 23. Prokhorenkova L, Gusev G, Vorobev A, et al. CatBoost: unbiased boosting with categorical features. Adv Neural Inf Process Syst. 2018;31:6638–6648. 24. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016:785–794. doi:10.1145/2939672.2939785. 25. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32. doi:10.1023/A:1010933404324. 26. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273–297. doi:10.1007/BF00994018. 27. Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010;21(1):128–138. doi:10.1097/EDE.0b013e3181c30fb2. 28. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565–574. doi:10.1177/0272989X06295361. 29. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30. 30. Katsumi A, Abe A, Tamura S, Matsushita T. Anemia in older adults as a geriatric syndrome: A review. Geriatr Gerontol Int. 2021;21(7):549–554. doi:10.1111/ggi.14183. 31. Roy CN. Anemia in frailty. Clin Geriatr Med. 2011;27(1):67–78. doi:10.1016/j.cger.2010.08.005. 32. Palmer K, Vetrano DL, Marengoni A, Tummolo AM, Villani ER, Acampora N, et al. The Relationship between Anaemia and Frailty: A Systematic Review and Meta-Analysis of Observational Studies. J Nutr Health Aging. 2018;22(8):965–974. doi:10.1007/s12603-018-1049-x. 33. Gado K, Khodier M, Virag A, Domjan G, Dornyei G. Anemia of geriatric patients. Physiol Int. 2022;109(2):119–134. doi:10.1556/2060.2022.00218. 34. Cabrerizo S, Cuadras D, Gomez-Busto F, Artaza-Artabe I, Marin-Ciancas F, Malafarina V. Serum albumin and health in older people: Review and meta analysis. Maturitas. 2015;81(1):17–27. doi:10.1016/j.maturitas.2015.02.009. 35. Bianchi VE. Role of nutrition on anemia in elderly. Clin Nutr ESPEN. 2016;11:e1-e11. doi:10.1016/j.clnesp.2015.09.003. 36. Ferrucci L, Semba RD, Guralnik JM, Ershler WB, Bandinelli S, Patel KV, et al. Proinflammatory state, hepcidin, and anemia in older persons. Blood. 2010;115(18):3810–3816. doi:10.1182/blood-2009-02-201087. 37. Ferrucci L, Guralnik JM, Woodman RC, Bandinelli S, Lauretani F, Corsi AM, et al. Proinflammatory state and circulating erythropoietin in persons with and without anemia. Am J Med. 2005;118(11):1288.e11-1288.e19. doi:10.1016/j.amjmed.2005.06.039. 38. Portoles J, Martin L, Broseta JJ, Cases A. Anemia in Chronic Kidney Disease: From Pathophysiology and Current Treatments, to Future Agents. Front Med (Lausanne). 2021;8:642296. doi:10.3389/fmed.2021.642296. 39. Styszynski A, Mossakowska M, Chudek J, Puzianowska-Kuznicka M, Klich-Raczka A, Neumann-Podczaska A, et al. Prevalence of anemia in relation to socio-economic factors in elderly Polish population: the results of PolSenior study. J Physiol Pharmacol. 2018;69(1):75–81. doi:10.26402/jpp.2018.1.08. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9486902","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":629435711,"identity":"e9f2428d-bf4b-49f1-92bc-148dc6f1466c","order_by":0,"name":"Zhiyuan Niu","email":"","orcid":"","institution":"Tianjin Haihe Hospital","correspondingAuthor":false,"prefix":"","firstName":"Zhiyuan","middleName":"","lastName":"Niu","suffix":""},{"id":629435712,"identity":"f9852f30-cc96-48ea-9aa2-9e4b61d85fcf","order_by":1,"name":"Lei Ren","email":"","orcid":"","institution":"Tianjin Armed Police Corps Hospital","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Ren","suffix":""},{"id":629435713,"identity":"10aaed95-d911-4e7d-8e18-29f708e29583","order_by":2,"name":"Naibo Hu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIiWNgGAWjYFACxgaGDzz/eNjYG4AcAwvitDDOkDkgw8dzAKRFgjh7mHlsDtjISSSA2ERokZ/d3CY5I+cOD5vk86sbfhRIMPC3dyfgd9acg20SH84842GTzim72QN0mMSZsxvwO0oisU1yZg8zSEvaDR6gFgOJXPxa2IBapHn/AbVInkm7+YcYLTwgLTw8h3nYJNiP3SbKFgmJxGbLGTxpPGw8OWy3ZQwkeAj6RX5G+sMbH3hs7OXbjz+7+eaPjRx/ey9+LchuNACTxCoHAfYHpKgeBaNgFIyCEQQA1jJBcguIZvUAAAAASUVORK5CYII=","orcid":"","institution":"Tianjin Haihe Hospital","correspondingAuthor":true,"prefix":"","firstName":"Naibo","middleName":"","lastName":"Hu","suffix":""}],"badges":[],"createdAt":"2026-04-21 16:09:34","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9486902/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9486902/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108072928,"identity":"827ded2d-8ea3-468f-9661-4d896ea5cde6","added_by":"auto","created_at":"2026-04-29 06:16:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":305335,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of participant selection across NHANES cycles.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/aaef1cfcc835b236831e9030.png"},{"id":108072929,"identity":"026bf493-c08e-4ff3-a8f8-0a6663cddceb","added_by":"auto","created_at":"2026-04-29 06:16:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":641309,"visible":true,"origin":"","legend":"\u003cp\u003eLASSO-based feature selection and multivariable logistic regression results. (A) Coefficient profiles of the 13 candidate predictors in the LASSO model. (B) Ten-fold cross-validation for selection of the optimal penalty parameter λ. (C) Adjusted odds ratios and 95% confidence intervals for the 8 variables retained by LASSO and entered into the multivariable logistic regression model.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/f4949029ae23fae7f8359ef0.png"},{"id":108072931,"identity":"9d43c7a0-45eb-4dcf-9c1c-af40cdbd7d0b","added_by":"auto","created_at":"2026-04-29 06:16:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":156569,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic curves in the development and temporal validation cohorts.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/0fbea7362bd5f05c66a74059.png"},{"id":108072930,"identity":"b8f8af8a-77a2-452a-bff3-348ff6c88966","added_by":"auto","created_at":"2026-04-29 06:16:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":514652,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP-based interpretation of the best-performing CatBoost model.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/b5a5367d2a4939e978ff463a.png"},{"id":108072932,"identity":"7475ae94-751c-440e-9b33-8fe5d6eba109","added_by":"auto","created_at":"2026-04-29 06:16:49","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":82834,"visible":true,"origin":"","legend":"\u003cp\u003eDecision curve analysis for CatBoost and logistic-regression models.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/08e5552320cef7d147b27ba2.png"},{"id":108072933,"identity":"cff18edb-6d2f-43e8-8ebd-81569f0a5a3e","added_by":"auto","created_at":"2026-04-29 06:16:49","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":337677,"visible":true,"origin":"","legend":"\u003cp\u003eNomogram and calibration performance of the logistic-regression model.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/08c31545fc35b381db5e5b78.png"},{"id":108390799,"identity":"b9ad0597-98c5-4b20-800e-a639f1ad305e","added_by":"auto","created_at":"2026-05-04 06:57:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2342871,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9486902/v1/6feaea4f-c752-4eec-a01f-aa2fc94426df.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and temporal validation of a machine learning risk prediction model for anemia in older adults: A NHANES study of an inflammatory–nutritional–renal profile","fulltext":[{"header":"Background","content":"\u003cp\u003eAnemia is a common yet long-underrecognized clinical problem among older adults. Systematic reviews and meta-analyses indicate that the prevalence of anemia is high in older adults and is even higher among hospitalized individuals and those in long-term care settings [1,2]. Global burden of disease data further suggest that anemia remains an important cause of disability worldwide [3]. Moreover, anemia in later life is closely associated with functional decline, reduced muscle strength, increased risk of hospitalization, and poor prognosis [4,5]. Therefore, anemia in older adults is not merely a laboratory abnormality, but an important geriatric health issue with public health and clinical management implications.\u003c/p\u003e \u003cp\u003eAnemia in older adults is rarely driven by a single mechanism and more often reflects the combined effects of chronic low-grade inflammation, diminished nutritional reserve, and declining renal function [6\u0026ndash;8]. Inflammation may promote anemia through hepcidin-mediated iron restriction and suppression of erythropoiesis [6,9]. Low albumin not only indicates inadequate nutritional reserve but may also reflect a higher chronic disease burden and impaired overall health status [8,10]. Meanwhile, renal aging and chronic kidney disease, with their associated reduction in erythropoietin production, constitute another important pathological basis of anemia in older adults [7,11]. Accordingly, identifying older adults at high risk of anemia from the integrated perspective of an inflammatory\u0026ndash;nutritional\u0026ndash;renal profile is biologically plausible and potentially clinically valuable.\u003c/p\u003e \u003cp\u003eDespite the growing burden of anemia in older adults, current approaches to risk identification still rely largely on single indicators or traditional statistical analyses and therefore may fail to capture the complex nonlinear relationships among multidimensional clinical features. In recent years, machine learning has shown considerable promise in clinical prediction modeling through its capacity for high-dimensional data integration, complex pattern recognition, and individualized risk assessment [12,13]. Several studies have explored machine learning models for anemia prediction in outpatient or hospitalized older populations [14,15]; however, studies based on nationally representative community-dwelling older adults that simultaneously integrate inflammatory, nutritional, and renal indicators and incorporate temporal validation remain limited. Therefore, using NHANES data, we developed and temporally validated a risk prediction model for anemia in older adults, identified key predictors, and assessed its clinical interpretability and potential clinical applicability.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and data source\u003c/h2\u003e \u003cp\u003eThis study was based on publicly available data from the National Health and Nutrition Examination Survey (NHANES). NHANES uses a complex multistage, stratified probability sampling design and systematically collects demographic information, anthropometric measurements, laboratory results, and health-related data, thereby providing good national representativeness [16]. According to the prespecified study design, data from the 2015\u0026ndash;2016, 2017\u0026ndash;2018, and 2021\u0026ndash;2023 cycles were included. The 2015\u0026ndash;2016 and 2017\u0026ndash;2018 cycles were combined as the development cohort for feature selection, model training, and internal optimization, whereas the 2021\u0026ndash;2023 cycle served as the temporal validation cohort to evaluate model generalizability across different time periods.\u003c/p\u003e \u003cp\u003eThe manuscript was prepared in accordance with the TRIPOD\u0026thinsp;+\u0026thinsp;AI reporting framework [17]. The NHANES study protocol was approved by the Ethics Review Board of the National Center for Health Statistics (NCHS), and written informed consent was obtained from all participants [18].\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eParticipant selection and outcome definition\u003c/h3\u003e\n\u003cdiv class=\"Heading\"\u003eParticipant selection and outcome definition\u003c/div\u003e \u003cp\u003eParticipants aged\u0026thinsp;\u0026ge;\u0026thinsp;60 years in NHANES with available hemoglobin measurements were included. Exclusion criteria were: (1) age\u0026thinsp;\u0026lt;\u0026thinsp;60 years; and (2) missing hemoglobin data. The detailed sample selection process is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eAnemia was defined using sex-specific hemoglobin thresholds: hemoglobin\u0026thinsp;\u0026lt;\u0026thinsp;13.0 g/dL in men and \u0026lt;\u0026thinsp;12.0 g/dL in women [19].\u003c/p\u003e\n\u003ch3\u003eCandidate predictors and data preprocessing\u003c/h3\u003e\n\u003cp\u003eBased on the clinical pathophysiology of anemia in older adults and data availability in NHANES, candidate variables initially included demographic features, inflammatory markers, nutrition-related indicators, renal function markers, and comorbidity status. To balance sample retention, model robustness, and clinical interpretability, a prespecified tiered missing-data strategy was used. For covariates with a missing rate\u0026thinsp;\u0026lt;\u0026thinsp;5%, median imputation was used for continuous variables and mode imputation for categorical variables. For variables with 5%\u0026ndash;20% missingness, multiple imputation by chained equations (MICE) was prioritized, with sensitivity analyses based on simple imputation when necessary. Variables with \u0026gt;\u0026thinsp;20% missingness were generally excluded from the main model unless they had clear clinical relevance and strong theoretical importance, in which case they were considered only in supplementary or sensitivity analyses [20,21].\u003c/p\u003e \u003cp\u003eGiven the complex sampling design of NHANES, descriptive statistics and between-group comparisons incorporated sampling weights, strata, and cluster variables to obtain nationally representative estimates. Model development and temporal validation focused on individual-level risk stratification and discrimination; therefore, model training and testing were performed after unified data preprocessing without additional survey weighting [16].\u003c/p\u003e\n\u003ch3\u003eFeature selection and model development\u003c/h3\u003e\n\u003cp\u003eIn the development cohort, 13 candidate variables were prespecified for least absolute shrinkage and selection operator (LASSO) regression based on the clinical relevance of anemia in older adults and the availability of NHANES data. These variables included age, sex, race/ethnicity, poverty-income ratio, body mass index (BMI), estimated glomerular filtration rate (eGFR), blood urea nitrogen (BUN), serum albumin, diabetes status, hypertension status, smoking status, neutrophil-to-lymphocyte ratio (NLR), and high-sensitivity C-reactive protein (hs-CRP). Together, these variables represented demographic, socioeconomic, inflammatory\u0026ndash;nutritional\u0026ndash;renal, comorbidity, and lifestyle domains. Ten-fold cross-validation was then used to determine the optimal penalty parameter λ and to identify the predictors most strongly associated with anemia risk in older adults [22]. Under the optimal penalty parameter, LASSO retained 8 variables for subsequent model construction and effect estimation. Five prediction algorithms were then developed and compared, including CatBoost, XGBoost, random forest (RF), support vector machine (SVM), and logistic regression (LR) [23\u0026ndash;26]. All models underwent hyperparameter tuning using 5-fold cross-validation within the development cohort and were subsequently evaluated in the temporal validation cohort for generalizability.\u003c/p\u003e\n\u003ch3\u003eModel evaluation, interpretation, and clinical translation\u003c/h3\u003e\n\u003cp\u003eModel performance was evaluated from three aspects: discrimination, calibration, and net clinical benefit. Discrimination was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC); calibration was evaluated using calibration curves; and clinical utility was assessed using decision curve analysis (DCA) [27,28]. To emphasize clinical interpretability, we also reported adjusted effect estimates from multivariable logistic regression and constructed a nomogram based on the logistic model as a potential tool for clinical translation.\u003c/p\u003e \u003cp\u003eTo improve the interpretability of the best-performing nonlinear model, we further applied the SHapley Additive exPlanations (SHAP) framework to quantify the contribution of each variable to anemia risk prediction and generated both SHAP summary and dependence plots [29]. Thus, model presentation in this study simultaneously emphasized predictive performance, temporal validation stability, and clinical interpretability.\u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003eStudy population and flow\u003c/h2\u003e\n\u003cp\u003eThe participant selection process is shown in Figure 1. In the development cohort (2015\u0026ndash;2018), 19,225 participants were initially assessed, including 9,971 from the 2015\u0026ndash;2016 cycle and 9,254 from the 2017\u0026ndash;2018 cycle. After excluding 15,174 individuals aged \u0026lt;60 years and 386 with missing hemoglobin data, 3,665 participants were ultimately included for model development, including 1,725 from 2015\u0026ndash;2016 and 1,940 from 2017\u0026ndash;2018. In the temporal validation cohort (2021\u0026ndash;2023), 11,933 participants were initially assessed. After excluding 8,429 aged \u0026lt;60 years and 858 with missing hemoglobin data, 2,646 participants were included. The final analytic sample comprised 6,311 older adults.\u003c/p\u003e\n\u003cp\u003eFigure 1. Flowchart of participant selection across NHANES cycles.\u003c/p\u003e\n\u003ch2\u003eBaseline characteristics\u003c/h2\u003e\n\u003cp\u003eBaseline characteristics of the development and temporal validation cohorts are shown in Table 1. Weighted analyses indicated that the two cohorts were generally comparable with respect to the main demographic characteristics. The mean age was 69.67 \u0026plusmn; 6.75 years in the development cohort and 69.74 \u0026plusmn; 6.58 years in the temporal validation cohort, with no significant difference (P=0.839, SMD=0.010). Similarly, no significant between-cohort differences were observed for sex distribution (P=0.465, SMD=0.017) or poverty-income ratio (3.12 \u0026plusmn; 1.59 vs. 3.11 \u0026plusmn; 1.57, P=0.943, SMD=0.006).\u003c/p\u003e\n\u003cp\u003eFor laboratory and clinical indicators, BMI (P=0.305, SMD=0.044), eGFR (P=0.564, SMD=0.025), hs-CRP (P=0.085, SMD=0.045), and neutrophil-to-lymphocyte ratio (NLR; P=0.159, SMD=0.046) did not differ significantly between cohorts. By contrast, the temporal validation cohort had slightly lower albumin levels than the development cohort (4.02 \u0026plusmn; 0.32 vs. 4.14 \u0026plusmn; 0.33 g/dL, P\u0026lt;0.001, SMD=0.360), and BUN also showed a mild difference (16.00 [13.00\u0026ndash;20.00] vs. 17.00 [14.00\u0026ndash;21.00] mg/dL, P=0.019, SMD=0.083). In addition, the SMD for race/ethnicity distribution was 0.106, suggesting mild structural heterogeneity, although the overall distribution did not differ significantly (P=0.413). The prevalence of anemia was not significantly different between the two cohorts (9.8% vs. 11.3%, P=0.165, SMD=0.052). Overall, the cohorts remained largely comparable across most baseline features while retaining limited temporal heterogeneity.\u003c/p\u003e\n\u003cp\u003eTable 1. Weighted baseline characteristics of the development and temporal validation cohorts.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/69519_bce2c0439cd956a6/69519_custom_files/img1777403583.png\"\u003e\u003c/p\u003e\n\u003cp\u003eValues are presented as mean \u0026plusmn; SD, median [IQR], or n (weighted %). P values were calculated using survey-weighted t tests, survey-weighted rank tests, or Rao\u0026ndash;Scott chi-square tests, as appropriate.\u003c/p\u003e\n\u003cp\u003eSMD \u0026gt; 0.10 was considered indicative of meaningful imbalance. For multi-category variables, SMDs were summarized as the maximum absolute level-specific standardized mean difference.\u003c/p\u003e\n\u003cp\u003eAbbreviations: BMI, body mass index; hs-CRP, high-sensitivity C-reactive protein; NLR, neutrophil-to-lymphocyte ratio; eGFR, estimated glomerular filtration rate; IQR, interquartile range; SMD, standardized mean difference.\u003c/p\u003e\n\u003ch2\u003eVariable selection and multivariable association\u003c/h2\u003e\n\u003cp\u003eIn the development cohort, LASSO regression included 13 prespecified candidate variables, namely age, sex, race/ethnicity, poverty-income ratio, BMI, eGFR, BUN, albumin, diabetes, hypertension, smoking status, NLR, and hs-CRP. The coefficient profiles and cross-validation results indicated that 8 variables with non-zero coefficients were ultimately retained (Figure 2A\u0026ndash;B): age, race/ethnicity, poverty-income ratio, eGFR, BUN, serum albumin, diabetes status, and hs-CRP. Sex, BMI, hypertension, smoking status, and NLR were not retained in the final model. These 8 retained variables were subsequently entered into a multivariable logistic regression model to estimate their adjusted effects within the same model (Figure 2C). Overall, Figure 2 illustrates the sequential analytical process of candidate-variable shrinkage and selection, final variable retention, and effect estimation, suggesting that the risk structure of anemia in older adults is primarily organized around an inflammatory\u0026ndash;nutritional\u0026ndash;renal profile, with albumin showing the strongest protective association.\u003c/p\u003e\n\u003cp\u003eFigure 2. LASSO-based feature selection and multivariable logistic regression results. (A) Coefficient profiles of the 13 candidate predictors in the LASSO model. (B) Ten-fold cross-validation for selection of the optimal penalty parameter \u0026lambda;. (C) Adjusted odds ratios and 95% confidence intervals for the 8 variables retained by LASSO and entered into the multivariable logistic regression model.\u003c/p\u003e\n\u003ch2\u003eComparison of model performance\u003c/h2\u003e\n\u003cp\u003eThe predictive performance of the five machine learning algorithms is shown in Table\u0026nbsp;2\u0026nbsp;and Figure 3. In ROC analyses, XGBoost and random forest demonstrated high discrimination in the development cohort, with AUCs of 0.883 and 1.000, respectively. However, the AUC of random forest declined to 0.745 in the temporal validation cohort, suggesting substantial overfitting. By contrast, CatBoost achieved AUCs of 0.834 in the development cohort and 0.773 in the temporal validation cohort, showing the best performance in the validation set and slightly outperforming logistic regression (AUC=0.770) and XGBoost (AUC=0.758), thus indicating better generalization stability.\u003c/p\u003e\n\u003cp\u003eIn addition, CatBoost yielded a negative predictive value of 0.937 in the temporal validation cohort, suggesting good performance in identifying low-risk individuals. Considering discrimination, temporal validation stability, and model interpretability together, CatBoost was identified as the best-performing nonlinear model in this study and was therefore selected for subsequent SHAP interpretation. Logistic regression, in contrast, was considered more suitable as the foundation for a clinically translatable tool because it provides more direct clinical interpretability.\u003c/p\u003e\n\u003cp\u003eFigure 3. Receiver operating characteristic curves in the development and temporal validation cohorts.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;2. Predictive performance of candidate models in the development and temporal validation cohorts.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/69519_bce2c0439cd956a6/69519_custom_files/img1777403614.png\"\u003e\u003c/p\u003e\n\u003ch2\u003eSHAP-based interpretation of the best-performing model\u003c/h2\u003e\n\u003cp\u003eGiven that CatBoost showed the best and relatively stable overall performance in the temporal validation cohort, SHAP analysis was further performed to interpret its prediction process (Figure 4A\u0026ndash;B). The SHAP summary plot showed that race/ethnicity, albumin, age, eGFR, and BUN were the main contributors to model output; among the laboratory indicators, albumin was the most important. The SHAP dependence plot further indicated that lower albumin levels were associated with a markedly greater positive contribution to predicted anemia risk. This finding further supports the view that anemia risk in older adults is not determined by any single marker, but more closely resembles a susceptibility pattern driven by an integrated inflammatory\u0026ndash;nutritional\u0026ndash;renal profile.\u003c/p\u003e\n\u003cp\u003eFigure 4. SHAP-based interpretation of the best-performing CatBoost model.\u003c/p\u003e\n\u003ch2\u003eNomogram construction and clinical utility\u003c/h2\u003e\n\u003cp\u003eDecision curve analysis further showed that, across a relatively wide range of threshold probabilities, risk stratification based on this model provided greater net benefit than strategies of treating all or treating none (Figure 5).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo improve clinical usability, we further constructed a nomogram based on the core variables (Figure 6A). Calibration curves showed good agreement between predicted and observed risk in the temporal validation cohort (Figure 6B), with a calibration slope of 0.965 and an intercept of \u0026minus;0.069, indicating good consistency between predicted probabilities and observed outcomes. Because logistic regression preserves a more direct path of parameter interpretation, this model offers advantages in terms of clinical interpretability and potential clinical translation. These findings suggest potential clinical value and support the use of the model for early risk identification in older adults.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eUsing nationally representative NHANES data from older adults, we developed and temporally validated a risk prediction model for anemia that integrates inflammatory, nutritional, renal, and socioeconomic information. The main findings can be summarized in three points. First, among the 13 candidate variables prespecified for LASSO, 8 core features were ultimately retained, yielding a risk structure centered on an inflammatory–nutritional–renal profile and further shaped by age and socioeconomic vulnerability; among them, albumin showed the strongest protective association in multivariable logistic regression. Second, CatBoost maintained good discrimination in both the development and temporal validation cohorts, suggesting relatively stable generalization across time-separated samples of older adults. Third, SHAP analysis further showed that race/ethnicity, albumin, and related variables contributed substantially to model output, whereas logistic regression achieved an AUC close to that of CatBoost in the validation cohort while demonstrating better calibration and stronger clinical interpretability. Overall, these findings are consistent with prior work viewing anemia in older adults as a geriatric syndrome and emphasizing its close links with frailty and adverse outcomes [29–31].\u003c/p\u003e\n\u003cp\u003eFrom a pathophysiological perspective, our findings further support the view that anemia in older adults is not driven by a single mechanism, but more likely arises from the combined effects of inflammation, diminished nutritional reserve, declining renal function, and age-related physiological vulnerability. In this context, the final model does not capture one specific etiologic signal; rather, it more closely reflects a multidimensional anemia-prone phenotype [29,30,32]. Among the retained variables, albumin likely represents more than nutritional status alone. As a negative acute-phase reactant, lower albumin may also indicate increased chronic inflammatory burden, frailty, and reduced overall physiological reserve, which may explain its prominent contribution in both multivariable logistic regression and SHAP analysis [33]. In addition, reviews on nutrition-related anemia indicate that deficiencies in iron, folate, and vitamin B12 remain important contributors to anemia in older adults, while low albumin can serve as a clinical surrogate for nutritional risk and insufficient overall health reserve [34]. The inclusion of hs-CRP in the final model further supports the potential role of low-grade systemic inflammation in anemia among older adults, with plausible mechanisms involving inflammation-related iron restriction, hepcidin-mediated dysregulation of iron homeostasis, and suppression of erythropoiesis [35,36].\u003c/p\u003e\n\u003cp\u003eThe renal function indicators identified in this study are likewise biologically plausible. Reduced eGFR and elevated BUN not only indicate renal impairment, but may also reflect insufficient erythropoietin production, suppressed erythroid activity, and shortened red blood cell survival, all of which are important mechanisms underlying anemia in older adults [37]. In addition, the retention of poverty-income ratio and race/ethnicity suggests that anemia risk in older adults is influenced not only by biological vulnerability but also by broader social vulnerability. These variables may, to some extent, reflect differences in nutritional access, chronic disease burden, healthcare utilization, and the cumulative impact of long-standing structural disadvantage. Previous population-based research has also suggested that education level, marital status, residential environment, and other socioeconomic factors are associated with the distribution of anemia among older adults [38]. Taken together, our model integrates both biological and social dimensions of susceptibility to anemia, which may partly explain why it retained relatively stable predictive performance in time-separated cohorts [32].\u003c/p\u003e\n\u003cp\u003eAt the modeling level, the present findings have important translational implications. Recent studies have shown growing potential for machine learning in geriatric medicine and anemia screening [12–15]. However, for clinical research, the best-performing model is not necessarily the model most suitable for implementation. In the present study, CatBoost showed a slight advantage in discrimination, and we therefore further applied SHAP analysis after model comparison to reveal its key driving features. Logistic regression, however, performed very similarly in the later cycle and offered greater transparency, parameter interpretability, and ease of tool construction. Based on these findings, we present CatBoost as the main high-performance prediction model while also constructing a nomogram based on logistic regression. This parallel strategy of a performance-oriented model together with an interpretable model is better aligned with real-world clinical demands for accuracy, temporal validation stability, and clinical interpretability.\u003c/p\u003e\n\u003cp\u003eThe strengths of this study include the use of nationally representative NHANES data from older adults, the inclusion of routinely available clinical indicators, and the use of an independent subsequent survey cycle as a temporal validation cohort, all of which enhance the robustness and practical relevance of the findings to some extent. Nevertheless, several limitations should be acknowledged. First, NHANES is a cross-sectional survey, and the present study identifies a pattern of risk associated with coexisting anemia rather than strict causal relationships. Second, the validation data were drawn from different cycles within the same national survey system; thus, while this represents independent temporal validation, it cannot substitute for genuine external population-based validation. Third, some indicators closely related to etiologic subtyping of anemia, such as ferritin, transferrin saturation, vitamin B12, and folate, were not consistently available across cycles, limiting the model’s ability to further characterize anemia subtypes. Finally, machine learning models applied to complex survey data still require cautious interpretation with regard to modeling and generalizability. Future work should therefore validate model performance in external cohorts, prospective studies, and datasets containing more complete etiologic indicators, and explore whether the model improves clinical outcomes and decision-making.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eUsing NHANES data from older adults, we established and temporally validated a risk prediction model for anemia integrating an inflammatory\u0026ndash;nutritional\u0026ndash;renal profile. The findings suggest that anemia risk in older adults is not driven by a single factor, but is better understood as a composite susceptibility phenotype shaped jointly by inflammation, insufficient nutritional reserve, declining renal function, and socioeconomic vulnerability. Among the predictors, albumin emerged as the most important protective correlate. CatBoost demonstrated relatively stable predictive performance, whereas logistic regression and its nomogram offered stronger clinical interpretability. A prediction tool based on routinely available indicators may support the early identification and risk stratification of anemia in older adults, although further external validation and prospective research are still needed.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAUC: Area under the curve\u003c/p\u003e\n\u003cp\u003eBMI: Body mass index\u003c/p\u003e\n\u003cp\u003eBUN: Blood urea nitrogen\u003c/p\u003e\n\u003cp\u003eDCA: Decision curve analysis\u003c/p\u003e\n\u003cp\u003eeGFR: Estimated glomerular filtration rate\u003c/p\u003e\n\u003cp\u003ehs-CRP: High-sensitivity C-reactive protein\u003c/p\u003e\n\u003cp\u003eIQR: Interquartile range\u003c/p\u003e\n\u003cp\u003eLASSO: Least absolute shrinkage and selection operator\u003c/p\u003e\n\u003cp\u003eLR: Logistic regression\u003c/p\u003e\n\u003cp\u003eMICE: Multiple imputation by chained equations\u003c/p\u003e\n\u003cp\u003eNCHS: National Center for Health Statistics\u003c/p\u003e\n\u003cp\u003eNHANES: National Health and Nutrition Examination Survey\u003c/p\u003e\n\u003cp\u003eNLR: Neutrophil-to-lymphocyte ratio\u003c/p\u003e\n\u003cp\u003eRF: Random forest\u003c/p\u003e\n\u003cp\u003eROC: Receiver operating characteristic\u003c/p\u003e\n\u003cp\u003eSHAP: SHapley Additive exPlanations\u003c/p\u003e\n\u003cp\u003eSMD: Standardized mean difference\u003c/p\u003e\n\u003cp\u003eSVM: Support vector machine\u003c/p\u003e\n\u003cp\u003eTRIPOD+AI: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence\u003c/p\u003e\n\u003cp\u003eXGBoost: Extreme Gradient Boosting\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThe present study was based on publicly available, de-identified data from the National Health and Nutrition Examination Survey (NHANES). The NHANES study protocol was approved by the Ethics Review Board of the National Center for Health Statistics (NCHS), and written informed consent was obtained from all participants by NHANES. Because this study involved secondary analysis of publicly available de-identified data, additional ethical approval was not required. All methods were carried out in accordance with relevant guidelines and regulations.\u003c/p\u003e\n\u003cp\u003eClinical trial registration\u003c/p\u003e\n\u003cp\u003eNot applicable. This study was a secondary analysis of publicly available, de-identified NHANES data and did not involve assignment of participants to interventions or registration as a clinical trial.\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable. This manuscript does not contain any individual person\u0026rsquo;s identifiable data in any form.\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are publicly available from the National Health and Nutrition Examination Survey (NHANES). Analytic code is available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026rsquo; contributions\u003c/p\u003e\n\u003cp\u003eZhiyuan Niu\u0026nbsp;conceived and designed the study.\u0026nbsp;Lei Ren\u0026nbsp;performed data curation, formal analysis, and visualization.\u0026nbsp;Zhiyuan Niu\u0026nbsp;drafted the manuscript.\u0026nbsp;Naibo Hu\u0026nbsp;and\u0026nbsp;Zhiyuan Niu\u0026nbsp;critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003eDeclaration of generative AI and AI-assisted technologies in the manuscript preparation process\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this manuscript, the authors used AI-assisted language tools to support translation and language polishing. After using these tools, the authors carefully reviewed, revised, and edited the manuscript as needed and take full responsibility for the content of the final submitted version.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e1. Mohammadi A, Kazeminia M, Chogan A, Jalali A. Prevalence of anemia in older adults: A systematic and meta-analysis study. Int J Afr Nurs Sci. 2024;20:100739. doi:10.1016/j.ijans.2024.100739.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e2. Guralnik JM, Eisenstaedt RS, Ferrucci L, Klein HG, Woodman RC. Prevalence of anemia in persons 65 years and older in the United States: evidence for a high rate of unexplained anemia. Blood. 2004;104(8):2263\u0026ndash;2268. doi:10.1182/blood-2004-05-1812.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e3. GBD 2021 Anaemia Collaborators. Prevalence, years lived with disability, and trends in anaemia burden by severity and cause, 1990\u0026ndash;2021: findings from the Global Burden of Disease Study 2021. Lancet Haematol. 2023;10(9):e713-e734. doi:10.1016/S2352-3026(23)00160-6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e4. Culleton BF, Manns BJ, Zhang J, et al. Impact of anemia on hospitalization and mortality in older adults. Blood. 2006;107(10):3841\u0026ndash;3846. doi:10.1182/blood-2005-10-4308.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e5. Penninx BWJH, Pahor M, Cesari M, et al. Anemia is associated with disability and decreased physical performance and muscle strength in the elderly. J Am Geriatr Soc. 2004;52(5):719\u0026ndash;724. doi:10.1111/j.1532-5415.2004.52208.x.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e6. Wacka E, Nicikowski J, Jarmuzek P, Zembron-Lacny A. Anemia and Its Connections to Inflammation in Older Adults: A Review. J Clin Med. 2024;13(7):2049. doi:10.3390/jcm13072049.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e7. Santos S, Lousa I, Carvalho M, Sameiro-Faria M, Santos-Silva A, Belo L. Anemia in Elderly Patients: Contribution of Renal Aging and Chronic Kidney Disease. Geriatrics (Basel). 2025;10(2):43. doi:10.3390/geriatrics10020043.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e8. Stauder R, Valent P, Theurl I. Anemia at older age: etiologies, clinical implications, and management. Blood. 2018;131(5):505\u0026ndash;514. doi:10.1182/blood-2017-07-746446.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e9. Ganz T. Anemia of inflammation. N Engl J Med. 2019;381(12):1148\u0026ndash;1157. doi:10.1056/NEJMra1804281.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e10. Riviati N, Legiran, Indrajaya T, Saleh I, Ali Z, Irfannuddin, et al. Serum Albumin as Prognostic Marker for Older Adults in Hospital and Community Settings. Gerontol Geriatr Med. 2024;10:23337214241249914. doi:10.1177/23337214241249914.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e11. Koury MJ, Haase VH. Anaemia in kidney disease: harnessing hypoxia responses for therapy. Nat Rev Nephrol. 2015;11(7):394\u0026ndash;410. doi:10.1038/nrneph.2015.82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e12. Stahl D. New horizons in prediction modelling using machine learning in older people\u0026rsquo;s healthcare research. Age Ageing. 2024;53(9):afae201. doi:10.1093/ageing/afae201.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e13. Olawade DB, Adeniji YJ, Olatunbosun FA, Egbon E, David-Olawade AC. Artificial intelligence for anemia screening, diagnosis, and management. Curr Res Transl Med. 2025;74(1):103560. doi:10.1016/j.retram.2025.103560.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e14. G\u0026ouml;l M, Akt\u0026uuml;rk C, Talan T, Vural MS, T\u0026uuml;rkbeyler İH. Predicting malnutrition-based anemia in geriatric patients using machine learning methods. J Eval Clin Pract. 2025;31(2):e14142. doi:10.1111/jep.14142.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e15. Bi S, Meng G, Wei N, Zhou Q, Wang C. Influential factors and predictive model of anemia in hospitalized elderly patients. Sci Rep. 2025;15(1):35324. doi:10.1038/s41598-025-19397-6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e16. Centers for Disease Control and Prevention. NHANES Tutorials: Sample Design Module. Available from: https://wwwn.cdc.gov/nchs/nhanes/tutorials/sampledesign.aspx. Accessed April 11, 2026.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e17. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e18. Centers for Disease Control and Prevention. Ethics Review Board Approval. Available from: https://www.cdc.gov/nchs/nhanes/about/erb.html. Accessed April 11, 2026.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e19. World Health Organization. Guideline on haemoglobin cutoffs to define anaemia in individuals and populations. Geneva: World Health Organization; 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e20. Sterne JAC, White IR, Carlin JB, et al. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. BMJ. 2009;338:b2393. doi:10.1136/bmj.b2393.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e21. van Buuren S, Groothuis-Oudshoorn K. mice: Multivariate imputation by chained equations in R. J Stat Softw. 2011;45(3):1\u0026ndash;67. doi:10.18637/jss.v045.i03.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e22. Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Series B. 1996;58(1):267\u0026ndash;288.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e23. Prokhorenkova L, Gusev G, Vorobev A, et al. CatBoost: unbiased boosting with categorical features. Adv Neural Inf Process Syst. 2018;31:6638\u0026ndash;6648.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e24. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016:785\u0026ndash;794. doi:10.1145/2939672.2939785.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e25. Breiman L. Random forests. Mach Learn. 2001;45(1):5\u0026ndash;32. doi:10.1023/A:1010933404324.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e26. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273\u0026ndash;297. doi:10.1007/BF00994018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e27. Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010;21(1):128\u0026ndash;138. doi:10.1097/EDE.0b013e3181c30fb2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e28. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565\u0026ndash;574. doi:10.1177/0272989X06295361.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e29. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e30. Katsumi A, Abe A, Tamura S, Matsushita T. Anemia in older adults as a geriatric syndrome: A review. Geriatr Gerontol Int. 2021;21(7):549\u0026ndash;554. doi:10.1111/ggi.14183.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e31. Roy CN. Anemia in frailty. Clin Geriatr Med. 2011;27(1):67\u0026ndash;78. doi:10.1016/j.cger.2010.08.005.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e32. Palmer K, Vetrano DL, Marengoni A, Tummolo AM, Villani ER, Acampora N, et al. The Relationship between Anaemia and Frailty: A Systematic Review and Meta-Analysis of Observational Studies. J Nutr Health Aging. 2018;22(8):965\u0026ndash;974. doi:10.1007/s12603-018-1049-x.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e33. Gado K, Khodier M, Virag A, Domjan G, Dornyei G. Anemia of geriatric patients. Physiol Int. 2022;109(2):119\u0026ndash;134. doi:10.1556/2060.2022.00218.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e34. Cabrerizo S, Cuadras D, Gomez-Busto F, Artaza-Artabe I, Marin-Ciancas F, Malafarina V. Serum albumin and health in older people: Review and meta analysis. Maturitas. 2015;81(1):17\u0026ndash;27. doi:10.1016/j.maturitas.2015.02.009.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e35. Bianchi VE. Role of nutrition on anemia in elderly. Clin Nutr ESPEN. 2016;11:e1-e11. doi:10.1016/j.clnesp.2015.09.003.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e36. Ferrucci L, Semba RD, Guralnik JM, Ershler WB, Bandinelli S, Patel KV, et al. Proinflammatory state, hepcidin, and anemia in older persons. Blood. 2010;115(18):3810\u0026ndash;3816. doi:10.1182/blood-2009-02-201087.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e37. Ferrucci L, Guralnik JM, Woodman RC, Bandinelli S, Lauretani F, Corsi AM, et al. Proinflammatory state and circulating erythropoietin in persons with and without anemia. Am J Med. 2005;118(11):1288.e11-1288.e19. doi:10.1016/j.amjmed.2005.06.039.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e38. Portoles J, Martin L, Broseta JJ, Cases A. Anemia in Chronic Kidney Disease: From Pathophysiology and Current Treatments, to Future Agents. Front Med (Lausanne). 2021;8:642296. doi:10.3389/fmed.2021.642296.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e39. Styszynski A, Mossakowska M, Chudek J, Puzianowska-Kuznicka M, Klich-Raczka A, Neumann-Podczaska A, et al. Prevalence of anemia in relation to socio-economic factors in elderly Polish population: the results of PolSenior study. J Physiol Pharmacol. 2018;69(1):75\u0026ndash;81. doi:10.26402/jpp.2018.1.08.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"older adults, anemia, NHANES, temporal validation, inflammatory–nutritional–renal profile, clinical interpretability","lastPublishedDoi":"10.21203/rs.3.rs-9486902/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9486902/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eAnemia is common in older adults and is often associated with chronic inflammation, poor nutritional reserve, and renal dysfunction. However, risk prediction models for anemia in community-dwelling older adults based on nationally representative data remain limited. This study aimed to develop and temporally validate a machine learning model for predicting anemia risk in older adults using NHANES data, and to assess its interpretability and potential clinical applicability.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe used data from the National Health and Nutrition Examination Survey (NHANES). Participants aged\u0026thinsp;\u0026ge;\u0026thinsp;60 years from the 2015\u0026ndash;2016 and 2017\u0026ndash;2018 cycles were combined as the development cohort, and those from the 2021\u0026ndash;2023 cycle served as the temporal validation cohort. Least absolute shrinkage and selection operator regression was used for feature selection. Five models were developed and compared: CatBoost, XGBoost, random forest, support vector machine, and logistic regression. Model performance was evaluated in terms of discrimination, calibration, and clinical utility, and SHapley Additive exPlanations were used to interpret the best-performing nonlinear model.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eA total of 6,311 older adults were included, of whom 3,665 were in the development cohort and 2,646 were in the temporal validation cohort. Eight predictors were retained: age, race/ethnicity, poverty-income ratio, estimated glomerular filtration rate, blood urea nitrogen, serum albumin, diabetes status, and high-sensitivity C-reactive protein. CatBoost showed the best overall performance, with area under the curve values of 0.834 in the development cohort and 0.773 in the temporal validation cohort. Logistic regression achieved a comparable area under the curve of 0.770 in the temporal validation cohort and showed good calibration (slope 0.965; intercept\u0026thinsp;\u0026minus;\u0026thinsp;0.069). SHAP analysis identified race/ethnicity, albumin, age, estimated glomerular filtration rate, and blood urea nitrogen as major contributors to model prediction, with albumin emerging as the most important protective correlate.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eAnemia risk in older adults may be understood as a composite susceptibility pattern shaped by inflammatory, nutritional, renal, socioeconomic, and age-related factors. CatBoost provided the best predictive performance, whereas logistic regression and its nomogram offered greater interpretability and potential for clinical translation. A prediction tool based on routinely available indicators may support early risk identification and stratification of anemia in older adults.\u003c/p\u003e","manuscriptTitle":"Development and temporal validation of a machine learning risk prediction model for anemia in older adults: A NHANES study of an inflammatory–nutritional–renal profile","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-29 06:16:44","doi":"10.21203/rs.3.rs-9486902/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e2473b74-6f1d-4033-a960-3043b1cd003f","owner":[],"postedDate":"April 29th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-04T06:46:28+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-02T13:41:58+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-02T13:40:59+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T06:56:50+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-29 06:16:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9486902","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9486902","identity":"rs-9486902","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}