This was a prospective, longitudinal cohort study, reported in accordance with the 2024 TRIPOD + AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis—Artificial Intelligence) guidelines for the development and validation of multivariable prediction models [22].

This study employed convenience sampling to recruit older adults who underwent hip surgery under general anesthesia at a tertiary hospital in China between March 2024 and March 2025. Eligible patients were followed prospectively for 6 months postoperatively, with follow-up assessments conducted from March 2024 to September 2025.

Inclusion criteria were age 60 years or older; scheduled for elective hip surgery, including total hip arthroplasty, hemiarthroplasty, or internal fixation; and sufficient communication ability to complete study assessments. Exclusion criteria were severe psychiatric disorders that prevented adequate cooperation; substantial hearing or visual impairments that precluded completion of questionnaires; palliative status (ie, life expectancy <6  mo); or current enrollment in other studies likely to affect frailty outcomes. Participants were withdrawn if they were lost to follow-up due to relocation or hospital transfer, voluntarily withdrew consent, or died of causes unrelated to the study protocol.

Baseline data and preoperative frailty assessments were collected through face-to-face interviews conducted by trained researchers. Follow-up assessments at 1, 3, and 6 months postoperatively were conducted via structured telephone interviews.

As there is currently no universally accepted sample size calculation method for trajectory-based multiclass machine-learning prediction models, the sample size was pragmatically estimated using a conventional epidemiological formula to ensure adequate baseline cohort representativeness. The required sample size was calculated as follows [23]: n=(Zδ)2ϕ^(1−ϕ^), where Z is the critical value at a 95% CI, ϕ∧ is the estimated incidence of frailty, and δ is the acceptable margin of error. Based on a pilot study indicating a 6-month postoperative frailty incidence of 14% and setting d at 5%, the calculated sample size was 185. After accounting for a 10% attrition rate, the final target sample size was set at no fewer than 206 cases to ensure the epidemiological representativeness of the baseline cohort.

Ethical approval for the study was obtained from the Ethics Committee of Sir Run Run Shaw Hospital, Zhejiang University School of Medicine (number 2024-2687-01). Informed consent was obtained from all participants before the study commenced.

Frailty was assessed preoperatively and at 1, 3, and 6 months postoperatively using a modified Frailty Index (FI) based on the cumulative deficit model developed by Rockwood and Mitnitski [24]. Tailored to the clinical characteristics of older adults undergoing hip surgery, we developed a 37-item population-specific FI covering 6 domains: chronic diseases, mental health, activities of daily living (ADLs), cognitive function, and symptoms/signs. Although several predictive variables share conceptual domains with components of the FI, none of the specific indicators used to construct the FI were directly included as predictors in the model. FI scores ranged from 0 to 1, with higher scores indicating greater frailty. A score <0.25 was classified as nonfrail and a score ≥0.25 was classified as frail [25]. The full list of FI items is provided in Multimedia Appendix 1.

Based on a review of the literature and expert consultation, potential risk factors were explored within the HEM, encompassing individual characteristics, behavioral and psychological factors, interpersonal networks, and the living environment. Key individual characteristics included demographic and socioeconomic factors (eg, gender, marital status, education level, and income), physiological indicators (high-sensitivity C-reactive protein [hsCRP]), health behaviors (smoking and drinking), and clinical history (number of comorbidities and surgery type). Functional status was assessed using the Barthel Index for ADLs, with scores indicating varying levels of independence [26]. Nutritional status was evaluated via the Mini Nutritional Assessment–Short Form [27], with scores categorizing participants as malnourished, at risk, or well-nourished.

Behavioral and psychological factors, including depressive symptoms, were measured using the 5-item Geriatric Depression Scale, while sleep quality was assessed using the Pittsburgh Sleep Quality Index. For interpersonal networks, family function was assessed using the Family APGAR (Adaptability, Partnership, Growth, Affection, Resolve) Index [28], and social support was measured using the Social Support Rating Scale. Finally, environmental factors were evaluated using a self-developed Living Environment Questionnaire (LEQ), which assesses home adaptability, perceived environmental burden, community resource accessibility, and residential satisfaction. The LEQ demonstrated strong content validity and was tailored to the needs of older adults. Detailed descriptions of these variables and the respective scales are provided in Multimedia Appendix 1.

A complete-case analysis was conducted. Missing data primarily arose from loss to follow-up during the longitudinal assessments, and participants with incomplete follow-up data were excluded from the analysis. To assess the missing data mechanism, the Little test for missing completely at random (MCAR) was performed. In addition, baseline characteristics between participants included in the final analysis and those who were lost to follow-up were compared to evaluate the potential for selection bias. Categorical variables are expressed as frequencies and percentages. Differences between the training and test sets were assessed using the chi-square test. Continuous variables with a normal distribution are presented as mean (SD), while nonnormally distributed variables are shown as median (IQR). Depending on the data distribution, the t test or the Mann-Whitney U test was applied.

Group-based trajectory modeling was utilized to identify frailty trajectories in older patients with hip fractures. The optimal model was selected based on clinical relevance, fit indices, and classification quality, with a preference for simpler models with fewer groups. Model fit was evaluated using the following criteria [29]: (1) Bayesian Information Criterion (BIC): lower values indicate a better fit; (2) average posterior probability (AvePP): AvePP greater than 0.7 for all groups, with the smallest group proportion exceeding 5%; (3) odds of correct classification (OCC): OCC was greater than 5; (4) relative entropy (E_k): E_k was higher than 0.8.

The data were randomly split into training and test sets in a 60:40 ratio using stratified sampling to preserve the original distribution of the 3 frailty trajectory classes across the dataset. To prevent overfitting, LASSO (least absolute shrinkage and selection operator) regression was applied to the training set for high-dimensional feature selection. The L1 regularization term enabled simultaneous variable selection and coefficient shrinkage. Model performance was evaluated using 10-fold cross-validation, with the optimal penalty parameter (λ) selected according to the 1 SE rule to ensure model parsimony and generalizability. To evaluate potential overlap and multicollinearity among predictors, variance inflation factors (VIFs) were calculated for variables retained after LASSO selection. A VIF <5 was considered to indicate no significant multicollinearity [30].

The selected features were used to construct the XGBoost model. Because no trajectory class represented an extremely small proportion of the sample, no additional class imbalance handling strategies (eg, oversampling or SMOTE [Synthetic Minority Oversampling Technique]) were applied during model training. To improve model robustness and reduce overfitting, regularization and subsampling strategies were incorporated into the XGBoost model. During training, a 3-fold inner cross-validation was employed to estimate performance. For the test set, the area under the receiver operating characteristic curve, accuracy, F₁-score, sensitivity, and specificity were computed. The calibration curve was used to compare predicted probabilities with actual outcomes across frailty trajectories. Decision curve analysis was performed to assess net clinical benefit. Finally, SHAP values were used to interpret the XGBoost model, ranking feature importance based on the mean absolute SHAP values. The XGBoost model was deployed on a cloud platform [31] using the “Shiny” package, creating an online prediction calculator for frailty trajectories following hip surgery.

Statistical analysis and graphical representations were performed using IBM SPSS 26.0, R 4.5.1 (R Core Team), and Python 3.9 (Python Software Foundation), with the “traj,” “glmnet,” “xgboost,” and “shap” packages. All statistical tests were 2-sided, and statistical significance was set at α=.05.

A total of 300 patients were screened for the study, of whom 78 were excluded due to severe hearing impairment, significant visual decline, communication barriers, or other factors. Four patients with incomplete data were also excluded, leaving 218 patients enrolled. Multimedia Appendix 2 presents the flowchart of the participants included in this study. During the follow-up period, 9 participants were lost to follow-up at 6 months, yielding an attrition rate of 4.1%. Given that frailty trajectories were identified using group-based trajectory modeling based on 3 repeated follow-up assessments, participants with incomplete longitudinal data were excluded to ensure the stability and interpretability of trajectory classification. To evaluate the missing data mechanism, we conducted a missing data analysis (Multimedia Appendix 2). The Little MCAR test indicated that the missing data were consistent with the assumption of MCAR (χ²₄=3.750, P=.44). In addition, no significant differences were observed in baseline characteristics between participants retained in the analysis and those lost to follow-up (Multimedia Appendix 2), suggesting a low risk of selection bias.

Frailty levels among older patients undergoing hip surgery exhibited a dynamic trend, with an initial increase followed by a gradual decline. The prevalence of frailty was 70.81% (148/209) before surgery, peaked at 84.21% (176/209) at 1 month postoperatively, and subsequently decreased to 63.16% (132/209) at 3 months and 45.45% (95/209) at 6 months.

A 3-class trajectory model demonstrated the best fit based on BIC, AvePP, and E_k values (Multimedia Appendix 3). It showed the lowest BIC change (59.36). All classes had AvePP values greater than 0.7 and OCC values exceeding 5. The E_k was 0.905, indicating high classification precision. Therefore, the 3-trajectory model was selected to characterize postoperative frailty trajectories among older adults undergoing hip surgery. Based on the estimated intercepts and slopes, trajectory 0 was labeled as low-fluctuation frailty, comprising 55 out of 209 individuals (26%), with a lower baseline frailty level, a modest increase at 1 month, and a subsequent decline. Trajectory 1 was labeled as high-improvement frailty, including 81 out of 209 (39%) individuals who had a higher baseline frailty level and experienced continuous functional recovery. Trajectory 2 was labeled as high-deterioration frailty, consisting of 73 out of 209 (35%) individuals who started with a high frailty level and demonstrated progressive worsening over time. Detailed parameter estimates and visual representations are provided in Multimedia Appendix 3. Given the distinct clinical patterns observed among the 3 classes, a multiclass modeling approach was employed to support the subsequent predictive analysis.

A total of 209 participants were randomly assigned to a training set (n=126) and a testing set (n=83) in a 60:40 ratio. The training set was used for model development, while the testing set was reserved for independent performance evaluation. Baseline demographic and clinical characteristics were compared between the 2 sets, with statistically significant differences observed only in marital status and type of medical insurance (Table 1).

Twenty-five candidate predictors were entered into the LASSO regression for feature selection. Using the 1 SE criterion (λ=0.0342), 12 variables were retained, including age, gender, living arrangement, albumin, hsCRP, drinking status, teeth number, comorbidities number, ADLs, nutritional status, social support level, and living environment. The feature selection process is illustrated in Figure 1, with additional details provided in Multimedia Appendix 4. Multicollinearity diagnostics showed that the VIF values of the retained predictors ranged from 1.07 to 1.67, with all values below 5, indicating no evidence of significant multicollinearity. Detailed results are presented in Multimedia Appendix 4. These variables were subsequently incorporated into the XGBoost model to predict postoperative frailty trajectories and to explore the relative contribution of each predictor to the classification outcomes.

To optimize XGBoost performance, key hyperparameters were tuned using grid search and cross-validation to improve model generalizability and accuracy. The model achieved excellent discriminative performance, with microaverage area under the receiver operating characteristic curve of 0.98 (95% CI 0.96‐0.99) in the training set (Figure 2A) and 0.93 (95% CI 0.90‐0.96) in the testing set (Figure 2B). Additional metrics, including accuracy, F₁-score, sensitivity, and specificity, are presented in Table 2. Calibration curves (Figure 2C) closely approximated the 45° line, indicating good agreement between predicted and observed probabilities, with a weighted Brier score of 0.085. Decision curve analysis demonstrated consistently greater net clinical benefit across a range of threshold probabilities compared to the “treat-all” and “treat-none” strategies, supporting the model’s potential for clinical applications (Figure 2D).

The SHAP algorithm was used to visualize and interpret the contribution of each predictor to the frailty trajectory model. As shown in Figure 3A, the bar plot displays the mean SHAP values of the 12 selected predictors, reflecting their average impact on the model output. ADLs had the greatest predictive importance, followed by the number of comorbidities and the number of teeth. Figure 3B presents the SHAP value distributions for each feature, with color gradients from blue (low values) to red (high values) representing the magnitude of each predictor and its influence on individual predictions.

To visualize case-specific model explanations, Figure 3C and D presents SHAP waterfall plots for illustrative patients from the low-fluctuation and high-deterioration frailty trajectories, respectively. For these samples, the predicted model outputs were −0.555 and −0.043, compared with baseline values of −0.226 and 0.037. For the low-fluctuation trajectory, favorable living environments and higher teeth counts showed strong negative contributions to the model output, offsetting the positive contributions of age and hsCRP. In contrast, the high-deterioration trajectory was primarily characterized by positive contributions from comorbidities and ADL limitations, whereas nutritional status, social support, dental health, and age contributed negatively to the model output.

A web-based frailty trajectory prediction calculator was developed and is freely available online [32]. The user interface of the application is shown in Figure 4. This tool allows users to input key patient characteristics and obtain estimated probabilities for each frailty trajectory class, which may support frailty risk assessment and individualized management planning.

This study identified 3 distinct postoperative frailty trajectories among older adults following hip surgery: low fluctuation, high improvement, and high deterioration. These findings highlight that frailty after hip surgery is a dynamic process with substantial interindividual variability rather than a static condition. By moving beyond traditional binary outcomes [17,18], this trajectory-based approach provides a more nuanced understanding of recovery patterns that are often obscured in aggregate analyses. Furthermore, by integrating multidimensional predictors within the HEM, we developed a prediction model that demonstrated good discrimination and calibration performance. Collectively, these findings provide a basis for early risk stratification and the development of targeted, individualized intervention strategies in postoperative care.

Although the overall prevalence of frailty followed a rise-then-fall trajectory over the 6-month postoperative period, consistent with prior studies such as Huang et al [33], this general trend concealed the diversity of individual recovery paths. The early peak in frailty may be attributed to systemic inflammatory responses, catabolic stress, and functional limitations induced by postoperative pain, while the subsequent decline likely reflects recovery mechanisms, including tissue healing, pain relief, and rehabilitation engagement. Trajectory modeling in this study revealed 3 distinct frailty patterns: low fluctuation, high improvement, and high deterioration. These heterogeneous trajectories suggest that despite experiencing the same surgical stressor, older patients vary substantially in their recovery due to differences in baseline physiological reserve, comorbid conditions, social support, and environmental context.

A closer examination of these subgroups further highlights their clinical relevance. The low fluctuation group, despite having a relatively low baseline frailty level, experienced a notable increase at 1 month postoperatively. This transient deterioration may indicate reduced physiological resilience. Prior studies suggest that abrupt fluctuations in frailty may signal underlying instability and predict adverse events such as falls or rehospitalization more accurately than static frailty status [34]. For this group, interventions should aim to maintain functional stability by implementing prehabilitation, early mobilization, multimodal pain control, and prevention of iatrogenic complications [12]. In contrast, the high-improvement group demonstrated a steady recovery trajectory, likely reflecting reversible functional impairments driven primarily by hip pathology. These patients may benefit from standard rehabilitation care without intensive resource input, allowing health care systems to prioritize those at greater risk. The high-deterioration group represents a clinically vulnerable subgroup with persistent decline, suggesting unresolved physiological or psychosocial deficits. Comprehensive geriatric assessments and tailored multidisciplinary interventions are warranted for this population, ideally guided by frameworks such as the World Health Organization (WHO)’s ICOPE (Integrated Care for Older People), which emphasizes multidomain support targeting mobility, cognition, nutrition, sensory function, and social engagement [35,36].

Based on the identified frailty trajectories, we developed a predictive model to support early risk identification and targeted intervention. Guided by the HEM, the model incorporated variables spanning individual characteristics, interpersonal networks, and living environment. Although some key predictors (eg, ADLs, comorbidity burden, and nutritional status) are conceptually related to domains of the FI, the multicollinearity analysis did not indicate significant statistical overlap among the predictors. This suggests that, despite conceptual relatedness, these variables do not represent redundant information from a statistical perspective. Nevertheless, these factors may partly reflect baseline vulnerability. Therefore, the model should not be interpreted as predicting frailty trajectory class membership entirely independent of baseline frailty severity. Instead, it is more appropriately viewed as an early risk stratification tool that integrates multidimensional information. Consistent with this interpretation, the SHAP analysis identified ADLs, number of comorbidities, number of teeth, age, and living environment as the top contributors to the model predictions.

In line with previous research [17,37], this study confirmed that impaired ADLs and a greater number of comorbidities were the strongest predictors of frailty deterioration. These variables represent fundamental aspects of functional dependence and chronic disease burden. Declining ADLs reflect a depletion of multisystem physiological reserves and may initiate a downward cycle in which the reduced function limits physical activity, weakens muscle strength, and accelerates frailty progression [37]. Similarly, multimorbidity contributes through persistent low-grade inflammation, polypharmacy, and cumulative organ damage, ultimately undermining mobility and physiological stability [38]. For individuals with impaired ADLs and multiple comorbidities, the early implementation of interventions such as individualized rehabilitation, optimized management of chronic conditions, and comprehensive medication review may be essential to interrupt this adverse trajectory [39,40]. In addition, age and teeth number were identified as important predictors. While age reflects irreversible biological decline, tooth loss serves as a modifiable marker of oral frailty. Reduced dentition in older adults impairs chewing function and nutritional intake, which in turn promotes sarcopenia and accelerates frailty [41]. These findings highlight the importance of incorporating oral health screening into perioperative management, along with timely interventions such as denture repair, chewing function training, and personalized nutrition support [14].

Among the identified predictors, living environment was the only environmental factor ranked among the top 5, highlighting its important role in frailty progression. Although the LEQ was a self-developed instrument, its core domains are conceptually aligned with the WHO age-friendly environment framework [42]. Previous studies have shown that environmental factors such as housing safety, community resources, social participation, and perceived environmental support are closely associated with frailty and its progression in older adults [43-45]. Notably, our findings demonstrated that living environment independently predicted frailty trajectories beyond individual-level characteristics. This finding is consistent with recent evidence showing that more positive perceptions of community environments are associated with lower frailty risk [46], as well as the findings of Mak et al [47], who reported sustained effects of environmental factors on frailty progression. For older adults undergoing hip surgery, environmental barriers after discharge may limit rehabilitation adherence and increase the risks of falls and readmission, thereby accelerating frailty progression. These findings emphasize the need to extend clinical care beyond hospital settings by incorporating environmental risk assessments into discharge planning and transitional care. Practical strategies may include predischarge home safety evaluations, caregiver education, and community linkage interventions to foster a more supportive recovery environment and reduce frailty-related risks [48].

In addition to the core predictors, the model identified several supplementary variables, including living arrangement, gender, nutritional status, drinking behavior, social support, albumin, and hsCRP. Although these variables contributed less individually, they enriched the understanding of frailty trajectories from physiological, behavioral, and demographic perspectives. These results indicate the importance of combining model-based risk identification with comprehensive biopsychosocial assessments to detect individualized frailty risk profiles and guide targeted interventions. Building on this, implementing a Fracture Liaison Service (FLS) may offer an effective integrative care pathway [49]. Through its structured components of assessment, education, intervention, and follow-up, Fracture Liaison Service enables continuous, coordinated management for older adults with hip fractures, supporting the transition from early risk prediction to precise clinical action.

This study has several limitations. First, it was a prospective single-center study with a relatively small sample size, focusing exclusively on older patients undergoing elective hip surgery under general anesthesia, which may limit the generalizability of the findings to broader populations such as individuals receiving emergency surgery, regional anesthesia, or in other age groups. In addition, the relatively limited sample size for multiclass machine-learning modeling may increase the risk of overfitting and unstable feature selection despite the use of LASSO-based dimensionality reduction and internal cross-validation. Therefore, both the current prediction model and the associated web-based calculator should be considered exploratory and proof-of-concept tools rather than clinically deployable systems. Prospective multicenter external validation, calibration assessment, and health-economic evaluation are warranted before routine clinical integration.

Second, a complete case analysis was used to handle missing data. Although the Little MCAR test suggested a completely-at-random mechanism and no baseline differences were observed between retained and excluded participants, departures from this assumption cannot be ruled out. If missingness were related to the severity of frailty or the presence of comorbidities, this could introduce bias and potentially affect trajectory classification and model performance. Therefore, the findings should be interpreted with caution. Future studies may consider more robust approaches for handling missing data, where appropriate, to further assess the robustness of the results.

Third, although the model incorporated multilevel predictors guided by the socioecological framework, some important but difficult-to-measure confounders, such as genetic predisposition or deeper physiological reserves, may have been missed. Future studies could leverage technologies such as wearable devices and multiomics approaches to improve measurement precision and expand variable inclusion. Finally, the 6-month follow-up period may capture only early or acute frailty changes and fail to reflect long-term trajectories. Longer follow-up and analytical methods such as competing risk models may help differentiate between temporary functional decline and persistent frailty, thus identifying critical intervention windows.

This study developed and internally validated a proof-of-concept machine-learning model to predict frailty trajectories within 6 months after hip surgery in older adults. Grounded in the HEM, the model integrated multidimensional predictors and showed good performance in discrimination, calibration, and clinical utility. Key contributing factors included functional status, comorbidity burden, oral health, age, and living environment. The model was translated into a web-based tool to support early identification and personalized frailty management; it currently serves as a foundational step. Comprehensive external validation across diverse clinical settings is strictly required before this tool can be deployed in routine clinical practice. These findings provide methodological and practical support for shifting perioperative care toward a proactive and individualized approach.