Introduction

Frailty, a multidimensional geriatric syndrome closely associated with aging, manifests as an excessive decline in reserve and function across multiple physiological systems, leading to diminished responsiveness to minor stressors [1]. It is highly prevalent among hospitalized patients and significantly increases the risk of adverse health outcomes, including hospitalization, falls, disability, and mortality [2,3]. It is noteworthy that heart failure (HF), similar to frailty, represents a complex clinical syndrome with a persistently increasing hospitalization prevalence, which has become a significant global public health concern [4–6].

However, previous research on frailty has primarily focused on stable chronic disease populations [7]. However, frailty-related risk factors in patients with coexisting acute and chronic conditions—a common yet understudied clinical scenario—remain largely unexplored. In HF with acute infection, the overlapping pathophysiology of fluid retention (HF) and fluid exudation (infection) may synergistically exacerbate fluid overload, manifesting as bilateral lower extremity edema and secondary muscle strength decline, thereby increasing the risk of prolonged bed rest and adverse outcomes [8]. These complications not only worsen prognosis but also impose substantial challenges to clinical management [9,10].

Moreover, existing frailty prediction models, whether based on conventional regression or machine learning approaches, have notable limitations. Logistic regression-based nomograms, while achieving acceptable discrimination (e.g., AUCs of 0.912 and 0.881 in diabetic patients [11]; AUC of 0.847 in elderly coronary heart disease patients [12]), often suffer from limited practical utility or small sample sizes. Machine learning-based models, such as an XGBoost model for chronic obstructive pulmonary disease (COPD) patients using China Health and Retirement Longitudinal Study (CHARLS) data (test AUC 0.942) [13], have demonstrated superior performance but frequently lack objective laboratory parameters. Moreover, most existing models overlook model interpretability—a critical barrier to clinical adoption, as clinicians require intuitive understanding of how individual predictors contribute to frailty risk at the patient level. SHAP (SHapley Additive exPlanations), a game theory-based approach, quantifies each feature’s marginal contribution to individual predictions, effectively bridging the gap between predictive accuracy and clinical transparency [14,15].

To solve these disadvantages, this study aimed to investigate frailty-associated risk factors in patients with concurrent HF and acute infections, and to develop a frailty prediction model utilizing multiple machine learning algorithms. Additionally, a clinically accessible online calculator was developed and deployed to accurately identify high-risk individuals and tailor appropriate interventions based on individual risk factors. Ultimately, this approach sought to reduce the incidence of frailty or even reverse its progression in affected patients.

Methods

Result

Sample characteristics

This study initially screened 11762 hospitalized patients, excluding 10264 cases that met the exclusion criteria. Ultimately, 1498 patients were included in the analysis, with 1203 (80.3%) diagnosed with frailty. The median age was 77 years, with 59.8% males and 40.2% females. The mean hospital stay was 10.5 days. Baseline characteristics of the full cohort are presented in S2 Table. There were no significant differences in variables between the training and testing sets (S3 Table). The schematic of the study workflow is detailed in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. The schematic of the study workflow.

Abbreviations: AUROC, Area Under The Receiver Operating Characteristic Curve; AUPRC, Area Under Precision-Recall Curve; DCA, Decision Curve Analysis.

https://doi.org/10.1371/journal.pone.0355848.g001

Variables selection

This study initially considered 81 clinically relevant variables. To identify variables potentially associated with frailty, we first conducted univariate analyses. For continuous variables, group differences were assessed using the independent samples t-test or Mann-Whitney U test, depending on distributional normality. For categorical variables, the chi-square test or Fisher’s exact test was applied as appropriate. Variables showing significant associations in univariate analyses (p < 0.05) were retained as candidate predictors for subsequent LASSO-based selection, resulting in 37 candidate predictors. (Table 1).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Demographics and potential risk factors of patients in the training set.

https://doi.org/10.1371/journal.pone.0355848.t001

To reduce dimensionality and select a parsimonious set of predictors, we employed LASSO logistic regression with 10-fold cross-validation. The optimal regularization parameter (λ) was chosen using the one-standard-error (1-SE) rule, which selects the largest λ within one standard error of the minimum cross-validated error—thereby favoring model simplicity while preserving predictive performance and mitigating overfitting. The selected λ was 0.0238 (S4 Table). At this penalty level, LASSO retained 11 variables with non-zero coefficients: age, serum albumin, eGFR, lymphocyte percentage, MCHC, capacity for action, LVEF, NYHA functional class, history of cerebral infarction, smoking status, and the use of thiazide diuretics. There was no substantial multicollinearity between the variables, with all VIF values<1.5 (S4 Table). These 11 variables constituted the final predictor set for all subsequent modeling. A flowchart illustrating the variable selection process is provided in S1 Fig.

Using the above 11 variables, we developed eight ML models: LR, RF, XGBoost, LightGBM, SVM, CatBoost, NB, and MLP. The hyperparameters of each model were optimized by grid search, and the specific parameters are shown in S5 Table.

Model performance

The predictive performance of eight machine learning models was evaluated on the testing set. The XGBoost model demonstrated superior overall performance, achieving an AUROC of 0.872 (95% CI: 0.835–0.909) and AUPRC of 0.969 (95% CI: 0.945–0.983) (Fig 2). Calibration performance was evaluated using calibration curves, which assess the agreement between predicted and observed probabilities, together with Brier scores, where lower values indicate better overall probability estimation (Fig 3). The calibration curves indicated that the XGBoost, MLP, and SVM models exhibited relatively better agreement with the ideal 45° reference line than the other models. Among the evaluated models, XGBoost achieved the lowest Brier score (0.106), followed by MLP (0.107) and SVM (0.110), suggesting relatively better overall probability estimation and satisfactory calibration performance than the other evaluated models. DCA revealed that XGBoost provided consistently higher standardized net benefit than other models across the clinically relevant threshold probability range of 60–90% (Fig 4). At the optimal threshold determined by the Youden index, XGBoost maintained strong performance metrics: accuracy (0.787), sensitivity (0.783), specificity (0.802), precision (0.948), recall (0.783), and F1-score (0.858) (Table 2). The training set performance metrics are shown in S6 Table. To further assess the robustness of the XGBoost model, we performed 5-fold stratified cross-validation (random_state = 42) on the full cohort (N = 1,498) using the same set of 11 predictors and hyperparameter configuration selected during the initial development phase. For each performance metric, we report the mean value across the five folds along with a 95% confidence interval estimated via non-parametric bootstrapping with 2,000 resamples (np.random.seed(42)) (S7 Table) The cross-validated performance remained highly consistent with the independent test set results: mean AUROC of 0.873 (vs. 0.872 on testing set), mean AUPRC of 0.967 (vs. 0.969), sensitivity of 0.751 (vs. 0.783), specificity of 0.881 (vs. 0.802), PPV of 0.963 (vs. 0.948), NPV of 0.474 (vs. 0.448), and F1-score of 0.842 (vs. 0.858). The close agreement between these results confirms the robustness of the model against data sampling variability. Based on these comprehensive evaluations, XGBoost was selected as the final prediction model.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. The performance metrics of the eight machine learning models on the testing set.

https://doi.org/10.1371/journal.pone.0355848.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. AUROC and AUPRC curves of the eight machine learning models.

(A), AUROC curves of the training set; (B), AUROC curves of the testing set; (C), AUPRC curves of the training set; (D), AUPRC curves of the testing set. Abbreviations: LR, Logistic Regression; RF, Random Forest; XGBoost, eXtreme Gradient Boosting; LightGBM, Light Gradient Boosting Machines; SVM, Support Vector Machine; CatBoost, Categorical Boosting; NB, Naive Bayes; MLP, Multilayer Perceptron; AUROC, Area Under The Receiver Operating Characteristic Curve; AUPRC, Area Under Precision-Recall Curve.

https://doi.org/10.1371/journal.pone.0355848.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Calibration curves (reliability curves) of eight machine learning models on the testing set.

The x-axis represents the mean predicted probability, and the y-axis represents the actual proportion of positive samples. The dashed line represents perfect calibration (where the predicted probability exactly matches the actual proportion). The values in the brackets are the Brier scores of each model, which are used to evaluate the reliability of the model’s probability prediction. Abbreviations: LR, Logistic Regression; RF, Random Forest; XGBoost, eXtreme Gradient Boosting; LightGBM, Light Gradient Boosting Machines; SVM, Support Vector Machine; CatBoost, Categorical Boosting; NB, Naive Bayes; MLP, Multilayer Perceptron.

https://doi.org/10.1371/journal.pone.0355848.g003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Decision curve analysis of eight machine learning models on the testing set.

The x-axis represents the threshold probability, and the y-axis denotes the net benefit. Curves for different models illustrate their net benefit across varying threshold probabilities. The “None” curve (dashed line) assumes no intervention, and the “All” curve (dotted line) assumes all cases are positive. Abbreviations: LR, Logistic Regression; RF, Random Forest; XGBoost, eXtreme Gradient Boosting; LightGBM, Light Gradient Boosting Machines; SVM, Support Vector Machine; CatBoost, Categorical Boosting; NB, Naive Bayes; MLP, Multilayer Perceptron.

https://doi.org/10.1371/journal.pone.0355848.g004

Model interpretation and application

Capacity for action, age, and eGFR were the three most significant factors influencing the prediction of frailty risk, according to an interpretability analysis of the optimal XGBoost model using the SHAP method. These were followed by serum albumin, MCHC, lymphocyte percentage, LVEF, NYHA functional class, history of cerebral infarction, smoking, and the use of thiazide diuretics (Fig 5A). Dependency plot revealed that while capacity for action, eGFR, serum albumin, MCHC, lymphocyte percentage, LVEF, and thiazide diuretics use were negatively associated with the risk of frailty, advanced age, NYHA functional class III-IV, history of cerebral infarction, and smoking were positively associated with the risk of frailty (Fig 5B). We developed an interactive prediction tool (https://frailty-risk-assessment.streamlit.app/) that calculates the probability of frailty risk in patients with HF co-infections in real time through a visual interface and dynamically displays the contributing weights of each clinical variable to provide an interpretable demonstration of model predictions (S2 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. SHAP summary plot for the eleven influential variables in the XGBoost model.

(A) The average absolute influence of each factor on the model output magnitude was presented in descending order of feature significance; (B) The graph depicted the dot estimate of the XGBoost model output, with each dot corresponding to a patient in the dataset.

https://doi.org/10.1371/journal.pone.0355848.g005

Discussion

This study developed a machine learning prediction model for the occurrence of frailty in HF patients with acute infections. The prevalence of frailty in our cohort was 80.3%, which is largely consistent with the findings reported by Vidán et al [30]. Notably, this high frailty prevalence leads to class imbalance in the dataset, a factor that may exert non-negligible impacts on model performance and result interpretation. Specifically, a skewed class distribution tends to induce model bias toward the majority class (frail patients), which may inflate the overall accuracy while masking the model’s potential deficiency in identifying the minority class (non-frail patients). Such bias could further lead to misleading judgments about the model’s actual discriminative ability, especially in distinguishing non-frail individuals who may require different clinical management strategies. To mitigate the potential influence of prevalence imbalance on model evaluation, we adopted a comprehensive assessment strategy incorporating discrimination, classification performance, calibration, and clinical utility rather than relying on a single performance metric. The study employed eight machine learning models, with the XGBoost model demonstrating optimal performance. Through XGBoost model analysis, significant predictors associated with frailty were identified, including thiazide diuretics use, serum albumin, eGFR, lymphocyte percentage, MCHC, capacity for action, age, LVEF, NYHA functional class, history of cerebral infarction, and smoking.

This study demonstrated the model’s innovation and practical value through the following aspects: first, the introduced XGBoost model exhibited exceptional predictive performance, with AUROC (0.872) and AUPRC (0.969) metrics significantly outperforming other comparative models. As one of the most widely utilized machine learning models in medical research, XGBoost became a vital tool in clinical decision support systems due to its computational efficiency and superior predictive accuracy, demonstrating significant clinical applicability [31]. Furthermore, the acceptable calibration performance of the XGBoost model supports the potential utility of its predicted probabilities for individualized risk assessment. However, these probability estimates should be further validated in external populations before being applied in routine clinical decision-making. Additionally, this study employed SHAP to enhance the interpretability of the XGBoost model. SHAP analysis provided both global and individual-level explanations by quantifying feature importance, illustrating the directionality of predictor effects, and visualizing patient-specific contributions through waterfall plots integrated into the online risk calculator. This provided reliable evidence-based support for developing personalized treatment strategies. Finally, we successfully developed an interactive web-based application to facilitate real-time clinical data input and visual risk prediction output. This tool significantly improved clinical workflow efficiency and allows clinicians to dynamically monitor disease progression trends, offering immediate decision-making support for precision medicine. In summary, this study extended the predictive model into an interactive visualization tool by integrating a full-process framework of “patient feature input, SHAP-based real-time computation, and visualized output.” This study not only provided a proof-of-concept demonstration of the potential application of machine learning models in clinical risk assessment but also provided a scalable technical approach for advancing intelligent clinical decision-making. Future development of intelligent clinical decision-support tools. However, this tool should be considered a proof-of-concept prototype rather than a clinically ready application, and further external validation is required before clinical implementation.

Compared with previous frailty prediction models for HF and elderly populations, the present study exhibits marked differences in study population, modeling strategy, and predictor composition. Previous models have been predominantly based on logistic regression or nomograms [22,32]. In contrast, our study is the first to focus on patients with HF and acute infections, systematically comparing multiple machine learning algorithms to select the optimal model and incorporating SHAP to enhance interpretability. The core predictors in our study mainly involve medication-, nutrition-, and inflammation-related indicators, which are more innovative than those in prior research. Overall, our model better aligns with the actual clinical needs of this specific patient population. Moreover, the predictors identified in this study implicate both core biological pathways of frailty and disease‑specific predictive factors. Specifically, albumin and lymphocyte percentage reflect nutritional–immune status, renal function and age represent cumulative physiological reserve, and capacity for action integrates multisystem aging effects. Through the interplay of systemic inflammation, nutritional depletion, and hemodynamic stress, these variables may enhance predictive performance in older patients with HF, consistent with previously proposed multidimensional frailty frameworks [33]. In addition, the catabolic syndrome posited by prior studies provides pathophysiological support for the predictive direction of metabolism-related factors observed in our analysis [34].

Currently, there is no universally established gold standard for the frailty prediction probability threshold that triggers intervention. The determination of such a threshold is inherently determined by the specific “risk-benefit ratio” of the intervention itself. Based on the DCA results, the model demonstrated potential clinical utility within a threshold probability range of 0.6 to 0.9 in the present cohort. However, these thresholds should not be interpreted as definitive intervention criteria, as clinical decisions require consideration of individual patient characteristics, intervention risks, and potential benefits. Therefore, practical intervention decisions should integrate multiple factors, including patient preference, overall health status, intervention costs, and associated risks, requiring clinicians to conduct a comprehensive evaluation and balance these considerations. The most feasible strategy is to incorporate the prediction model as an “intelligent early-warning tool” into the clinical workflow, with the final individualized decision made by the physician based on a holistic assessment of the patient’s information. If, upon clinical assessment, a patient is deemed to be at high risk and requires intervention initiation, the following integrated workflow can be implemented: First, the attending physician formulates a personalized intervention plan based on the patient’s specific characteristics. Subsequently, a geriatrician conducts a professional, comprehensive frailty intervention, encompassing nutritional support, exercise guidance, and other components. Finally, the patient’s frailty status is regularly assessed through follow-up to evaluate improvement, and the intervention strategy is dynamically adjusted based on the follow-up findings. The successful implementation of this workflow relies on corresponding resource support: Both the attending physician and the geriatrician must possess substantial clinical experience and appropriate qualifications, while the hospital needs to integrate resources across multiple departments to provide patients with systematic and continuous intervention support.

The findings of this study suggest thiazide diuretics are independently associated with a lower probability. This observation may be attributed to the following potential reasons. This observed association may be attributed to several mechanisms. First, thiazide diuretics primarily reduce cardiac preload and afterload by decreasing blood volume, thereby alleviating symptoms such as dyspnea and edema while improving exercise tolerance [35]. Second, a previous multicenter, randomized, double-blind clinical trial on acute decompensated HF demonstrated that adding oral thiazide diuretics to intravenous loop diuretics mitigated adverse effects like hypokalemia, consequently preventing muscle weakness. This suggests thiazide diuretics may offer more favorable potassium-sparing effects [36]. Furthermore, in infected patients, thiazide diuretics can effectively alleviate fluid exudation caused by primary lesions. Taking pulmonary infection as an example, increased sputum production often leads to persistent or refractory cough, with some patients showing poor response to conventional expectorant therapy. For HF patients during acute infections, thiazide diuretics not only exert their inherent diuretic effects but also significantly improve symptoms of cough and dyspnea associated with pulmonary infection by modulating infection-related fluid exudation mechanisms [37,38]. The explanation above is primarily inferred from the observation that thiazide diuretics can alleviate clinical symptoms in patients with HF and acute infections, which indirectly suggests their use is an independent predictor associated with lower frailty probability. However, this does not provide direct clinical evidence that such drugs can effectively prevent frailty. Therefore, the causal relationship between the two still requires further validation through additional research.

A comparative evaluation provides compelling evidence for the significance of thiazide diuretics. Clinically, loop diuretics, as first-line agents for HF treatment, exhibit significant clinical limitations despite their widespread use. The primary concern is their propensity to induce severe electrolyte imbalances, compounded by the increasingly prevalent phenomenon of diuretic resistance. Research demonstrates that combination therapy not only effectively addresses diuretic resistance but also significantly reduces adverse drug reactions [39]. However, current studies predominantly focus on combining loop diuretics with mineralocorticoid receptor antagonists [40], while the potential of thiazide diuretics remains underinvestigated. Compared to loop diuretics, thiazide diuretics exhibit a more moderate diuretic effect. However, they also carry the risk of causing electrolyte disturbances, which may consequently exacerbate patient frailty. Currently, the combined use of thiazide and loop diuretics has gained acceptance in clinical practice. This dual treatment strategy can synergistically enhance diuretic efficacy, aid in weight reduction, significantly improve edema symptoms, and may potentially reduce the risk of electrolyte imbalances [36]. However, the negative correlation observed in this study between thiazide diuretics and frailty risk is an exploratory finding. It neither supports modifications to current clinical prescribing practices nor carries direct therapeutic implications at this stage. Due to limitations inherent in its cross-sectional design, reliance on single-center data, and the absence of a control group, the current evidence cannot substantiate a preventive effect of the medication on frailty and merely suggests a potential link between the two. Furthermore, in this study, the use of thiazide diuretics was identified based on electronic prescription records, specifically including hydrochlorothiazide and indapamide. Inclusion criteria for medication required prescriptions issued within six months prior to admission, with continuous use documented for six months. However, this study did not record specific dosage ranges or assess patient medication adherence. Future studies should address these limitations through a more detailed investigation.

The findings revealed significantly lower serum albumin levels in the frail group compared to the non-frail group. The relationship between frailty and serum albumin can be explained through the following mechanisms: first, frail elderly individuals exhibit systemic functional decline, including compromised masticatory function. This impairment often leads to inadequate food breakdown during chewing, particularly affecting the digestion and absorption of protein-rich foods. Second, frailty adversely affects gastrointestinal function, further hindering nutrient assimilation. Third, during acute infections, frail patients demonstrate more pronounced protein catabolism compared to their non-frail counterparts [41]. Therefore, for patients with HF complicated by acute infections, nutritional management is critically important. Moreover, micronutrient supplementation (particularly of calcium, iron, zinc, and magnesium) may significantly aid in restoring serum albumin levels [42].

eGFR, a key indicator of renal function, demonstrated an inverse relationship with frailty in this study. A decline in eGFR leads to overactivation of the renin-angiotensin-aldosterone system (RAAS), exacerbating fluid retention. This fluid accumulation is particularly detrimental in HF patients, whose pre-existing fluid overload further restricts mobility and accelerates frailty progression [43]. While eGFR decline is irreversible, the judicious selection of nephroprotective pharmacological agents may help slow the progression of renal insufficiency, thereby delaying the onset of frailty. Furthermore, lymphocyte percentage serves as a key indicator of immune system function, with its decline reflecting impaired immunity and significantly elevating frailty risk [44]. During acute infections, lymphocyte percentage emerges as a clinically significant predictor of frailty among various infection markers. Notably, evidence suggests that enhanced physical activity may prevent immunosenescence and mitigate frailty progression [45].

MCHC, a key diagnostic parameter for anemia, reflects the hemoglobin concentration within red blood cells. A low MCHC often indicates potential iron deficiency in the body. Iron is an essential element for erythropoiesis, and its deficiency can impair hemoglobin synthesis, reducing the oxygen-carrying capacity of red blood cells. This leads to systemic tissue hypoxia, which may manifest as muscle weakness, fatigue, and decreased exercise tolerance [46,47]. Furthermore, chronic hypoxia can disrupt mitochondrial function, decreasing adenosine triphosphate (ATP) production and limiting energy supply for daily activities, thereby accelerating frailty progression [48]. Therefore, increasing dietary iron intake may help improve frailty status in these patients.

This study demonstrates that patients who remain ambulatory are less prone to frailty compared to those who are wheelchair-dependent or bedridden. Therefore, for patients with HF and acute infections, exercise-based rehabilitation should be prioritized to improve mobility. Although physical activity increases myocardial oxygen demand and may transiently worsen HF symptoms, judicious exercise promotes muscle recovery, enhances the capacity for action, and reduces frailty risk—provided the intensity remains within individualized thresholds [49]. This necessitates careful supervision by experienced clinicians to optimize the risk-benefit balance. In summary, personalized exercise prescriptions should be implemented for this population to achieve measurable benefits. In addition, our study found that advanced age significantly increases the likelihood of frailty development, which aligns with previous reports [50]. The study also highlights associations between decreased LVEF, elevated NYHA functional class, history of cerebral infarction, smoking, and elevated frailty risk.

Limitations

Although this study has yielded significant findings, several limitations should be acknowledged. First, this study cannot establish a causal relationship between the use of thiazide diuretics and the risk of frailty onset. The analysis is further limited by the lack of recorded dosage ranges and assessment of medication adherence. Furthermore, given the retrospective observational design of this study, the observed associations may be subject to multiple confounding factors, including confounding by indication, differences in disease severity, baseline functional status, renal function, physician treatment selection, and concomitant HF therapies. Additionally, due to factors such as acute infectious events and functional decline during hospitalization, the CFS assessment at discharge may be subject to reverse causation. Second, this study did not formally evaluate the inter-rater reliability of the CFS. ADL and IADL assessments primarily rely on self-reported data, which may introduce recall bias, subjective overestimation, or underestimation. Third, the data were derived exclusively from a subset of elderly patients at a specific tertiary class-A public hospital in Nanjing. As a single-center study, the generalizability of the results is inherently limited. Moreover, the predictive model has not been externally validated, and its generalizability across diverse populations requires further confirmation. Therefore, external validation using independent cohorts from multiple institutions is necessary before clinical implementation. Fourth, although missing values were handled using KNN imputation, the missing data mechanism was not formally assessed, and sensitivity analyses using alternative methods were not performed. Future studies should compare different imputation strategies to further validate model robustness. Finally, although class weighting was used to address the high frailty prevalence (80.3%), resampling methods such as SMOTE were not applied to models that lack native class-weighting support, including NB and MLP, which may affect cross-model performance comparisons. Additionally, the model’s generalizability to settings with lower frailty prevalence may be limited.

Conclusion

This study developed a frailty prediction model for patients with HF complicated by acute infections. The model incorporates 11 readily accessible predictors. This study demonstrated that the use of thiazide diuretics is an independent predictor associated with lower frailty probability. We developed an online calculator as a proof-of-concept tool to demonstrate the potential application of the predictive model and facilitate real-time risk estimation. This tool facilitates the early identification of high-risk individuals and supports the implementation of personalized interventions.

Supporting information

Acknowledgments

We gratefully acknowledge all research participants and individuals involved in data collection.