Abstract
Introduction
Femoral neck fractures are a heterogeneous entity with wide variability in patient profiles and outcomes. Single-variable classification may not capture the multidimensional interactions that influence prognosis. This study aimed to identify clinically meaningful patient subgroups using hierarchical cluster analysis and to compare mortality, postoperative mobility, and complication rates between these subgroups.
Materials and methods
This retrospective cohort study included 672 patients treated for intracapsular femoral neck fractures classified as AO Foundation/Orthopaedic Trauma Association (AO/OTA) 31-B between 2009 and 2023, including 586 who underwent hemiarthroplasty (HA) and 86 who underwent total hip arthroplasty (THA). Unsupervised hierarchical clustering was performed using the Gower distance matrix and Ward’s D2 linkage method based on 20 preoperative and intraoperative clinical and surgical variables, excluding outcome measures. The optimal number of clusters was determined by silhouette analysis. Mortality predictors were assessed using Cox proportional hazards regression, and predictors of good postoperative mobility were evaluated by binary logistic regression.
Results
Two distinct clinical phenotypes were identified: a high-risk cluster (n = 607; mean age 78.3 years, American Society of Anesthesiologists [ASA] physical status III 86.2%, HA 96.4%, dementia 14.8%) and a low-risk cluster (n = 65; mean age 63.4 years, ASA II 81.5%, THA 98.5%, dementia 0%). Cox regression identified age (p < 0.001), male sex (p < 0.001), dementia (p = 0.029), ASA score (p = 0.010), and red-cell distribution width coefficient of variation (RDW-CV; p = 0.011) as independent mortality predictors, whereas procedure type was not significant (hazard ratio [HR] = 0.659, 95% confidence interval [CI]: 0.427–1.015, p = 0.058). In contrast, procedure type was independently associated with good postoperative mobility (odds ratio [OR] = 3.944, 95% CI: 1.516–10.260, p = 0.005).
Conclusions
Hierarchical cluster analysis identified two clinically distinct patient phenotypes with significantly different outcomes. Procedure type did not independently predict mortality, whereas THA was independently associated with good postoperative mobility, suggesting that the apparent survival advantage of THA may largely reflect indication-based selection bias. These findings support individualised, multidimensional risk assessment rather than single-variable decision-making. Because the identified phenotypes incorporate intraoperative variables, they characterise prognosis rather than guiding preoperative arthroplasty selection.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/s00402-026-06487-x.
Keywords: Mortality, Comorbidity, Patient Phenotyping, Arthroplasty, Treatment Outcome
Introduction
Intracapsular femoral neck fractures (FNF), corresponding to AO Foundation/Orthopaedic Trauma Association (AO/OTA) 31-B injuries, remain among the most common and consequential injuries in the elderly population, carrying substantial morbidity and mortality [1]. This distinction is clinically important because extracapsular proximal femoral fractures are generally managed with fixation or intramedullary nailing, whereas displaced intracapsular femoral neck fractures in older adults are commonly treated with arthroplasty. Although typically caused by low-energy falls, their occurrence is facilitated by underlying risk factors such as osteoporosis, frailty, and polypharmacy. With population ageing, the prevalence of FNF has increased considerably, and fracture risk is estimated to double with each decade of life beyond the age of 50 [2]. Surgical management generally involves either internal fixation or arthroplasty, with the latter including hemiarthroplasty (HA) and total hip arthroplasty (THA). HA is associated with shorter operative time, lower blood loss, and lower dislocation risk, whereas THA may provide better functional outcomes and quality of life in selected patients [3–5]. The HEALTH trial found no significant difference in secondary procedure rates between THA and HA, although THA provided a modest functional advantage [6]. Subsequent meta-analyses have generally confirmed that THA may offer better medium-term function, while HA remains advantageous in terms of operative burden and hospital stay [7, 8].
A central challenge in managing FNF is the marked heterogeneity of the patient population. Age, comorbidity burden, cognitive status, functional reserve, and operative risk vary widely, yet conventional studies often stratify patients according to a single variable, such as age, American Society of Anesthesiologists (ASA) grade, or comorbidity index. This approach may overlook the interaction among multiple prognostic factors that collectively influence both treatment selection and outcome. Earlier studies have used cluster analysis to identify clinically distinct hip-fracture subgroups. Eastwood et al. [9] described eight patient subgroups with divergent six-month outcomes, while Penrod et al. [10] identified seven homogeneous clusters based on prefracture characteristics. More recently, supervised machine-learning methods have been applied to predict mortality and length of stay after femoral neck fracture [11, 12]. In contrast, unsupervised methods can uncover latent patient phenotypes without predefined outcome labels; however, their use in arthroplasty-treated FNF patients and their relationship with surgical outcomes remain relatively limited [13].
The present study therefore applied hierarchical cluster analysis to a cohort of patients with FNF treated with arthroplasty. The objectives were: first, to identify clinically meaningful patient subgroups based on demographic, clinical, and surgical characteristics; and second, to compare these subgroups with respect to mortality, major complications, and postoperative mobility. We hypothesised that hierarchical clustering would reveal distinct patient phenotypes associated with different clinical outcomes.
Materials and methods
This retrospective single-center cohort study was conducted at a tertiary referral center and included all patients who underwent arthroplasty for intracapsular femoral neck fractures classified as AO/OTA 31-B between January 2009 and December 2023. The study protocol was approved by the local institutional ethics committee (approval date:3 February 2026; decision no: 3272) and was conducted in accordance with the Declaration of Helsinki. This study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines.
Patients were included if they had a confirmed diagnosis of intracapsular femoral neck fracture corresponding to AO/OTA 31-B, had been treated surgically with either HA or THA, and had clinical and radiological follow-up data available in the hospital records. Extracapsular proximal femoral fractures and femoral head fractures were not included. Exclusion criteria comprised pathological fractures secondary to tumour or metastasis (n = 35), concomitant ipsilateral or contralateral lower-extremity fractures (n = 75), treatment with internal fixation rather than arthroplasty (n = 178), incomplete medical records precluding data extraction (n = 55), and a history of prior hip surgery on the affected side (n = 56). Of the 1,071 patients initially assessed, 399 were excluded, leaving a final study cohort of 672 patients (586 HA, 86 THA). The patient selection process is summarised in Fig. 1. The cohort had a mean age of 76.85 ± 10.50 years (range 38–98); 425 patients (63.2%) were female and 247 (36.8%) were male.
Fig. 1.
Patient selection flowchart
Clinical data were retrieved retrospectively from the hospital information management system and individual patient charts. Recorded variables were organised into the following domains: demographics (age, sex, body mass index); fracture characteristics (laterality, Garden classification, Pauwels classification); comorbidity indices (Charlson Comorbidity Index [CCI], age-adjusted Charlson Comorbidity Index [aCCI], ASA physical status, presence of dementia); surgical parameters (procedure type, anesthesia mode, surgical approach, cementation status, femoral head diameter, time from admission to surgery, operative duration); preoperative laboratory values (hemoglobin, Systemic Immune-Inflammation Index [SII], red-cell distribution width coefficient of variation [RDW-CV]); and outcome measures (mortality, survival time, major complications, postoperative mobility score, limb-length discrepancy).
The age-adjusted Charlson Comorbidity Index (aCCI) was categorized into three groups: low (aCCI ≤ 3), moderate (aCCI 4–5), and high (aCCI ≥ 6) [14, 15].
Primary and secondary outcomes
The primary outcome was all-cause mortality and survival time after surgery. Mortality status was obtained from hospital records and verified against the Turkish National Death Notification System and the national civil registry (MERNIS). Vital status was ascertained for all patients through 3 February 2026, with no loss to follow-up for the mortality endpoint. Survival time was calculated from the date of surgery to death or administrative censoring at the date of vital-status ascertainment. Mortality and survival were analysed using Kaplan–Meier survival analysis, the log-rank test, and Cox proportional hazards regression.
The secondary outcomes were postoperative mobility and major complications. Postoperative mobility was graded on a five-point clinical scale and dichotomised as good mobility, defined as independent ambulation or walking with a single aid, versus poor mobility, defined as requiring two aids or being unable to walk. Postoperative mobility was analysed using between-group comparisons and binary logistic regression among 30-day survivors. Patients who died within the first 30 postoperative days (n = 30) were excluded from the postoperative mobility analysis, yielding an evaluable population of 642 patients. The 30-day cutoff was used to identify early postoperative deaths, because these patients did not survive long enough for a clinically meaningful assessment of postoperative ambulation after the immediate perioperative period and early rehabilitation.
Major complications were defined as deep infection, dislocation, postoperative periprosthetic fracture, revision surgery, and intraoperative fracture. Postoperative periprosthetic fractures referred to fractures occurring after the index operation, whereas intraoperative fractures were recorded separately as fractures occurring during the surgical procedure. Because of the limited number of complication events, major complications were reported descriptively.
Sample size and power considerations
Because this was a retrospective complete-cohort study, no formal a priori sample-size calculation was performed before data collection. The sample size was determined by the number of consecutive eligible patients who underwent arthroplasty for intracapsular femoral neck fractures during the study period. To assess the adequacy of the available cohort for the planned multivariable analyses, we considered the number of outcome events relative to the number of predictors in the final models. For mortality, 437 deaths were available for the Cox regression model, which included six predictors, providing more than 70 events per predictor. For postoperative mobility, 642 patients were evaluable, with 308 good-mobility outcomes and four predictors in the final logistic regression model, providing more than 75 events per predictor. For a two-sided α of 0.05 and 80% power, this cohort was sufficient to detect a hazard ratio of approximately 1.5 for all-cause mortality given the observed event distribution, and provided more than 99% power for the postoperative mobility comparison; we report this design-based precision rather than post-hoc observed power, which is statistically uninformative. In contrast, because only 53 major-complication events occurred, complication outcomes were reported descriptively and no multivariable complication model was constructed.
Statistical analysis
Continuous variables were reported as mean ± standard deviation and categorical variables as frequency and percentage. Normality was assessed using the Kolmogorov–Smirnov test. Because continuous variables were non-normally distributed, between-group comparisons were performed using the Mann–Whitney U test, and categorical variables were compared using the chi-square test.
The median potential follow-up for the whole cohort was estimated using the reverse Kaplan–Meier method. Effect estimates from regression models were reported as hazard ratios (HRs) or odds ratios (ORs) with 95% confidence intervals (CIs), as appropriate. Because cementation differed between procedure types and patient clusters and may influence postoperative morbidity and mortality, an additional cementation-adjusted Cox sensitivity model was performed. This model included the variables from the primary Cox model with the addition of cementation status.
Propensity-score sensitivity analysis
To address indication-based selection bias in the comparison between THA and HA, an inverse-probability-of-treatment weighting analysis was performed as a sensitivity analysis. The propensity score, defined as the probability of receiving THA, was estimated using logistic regression. Only baseline and preoperative variables were entered into the propensity model: age, sex, body mass index, Charlson Comorbidity Index, ASA physical status, fracture side, Garden classification, Pauwels classification, time from admission to surgery, preoperative hemoglobin, SII, and RDW-CV. Dementia was not included in the propensity model because no patient in the THA group had dementia, which would have caused complete separation. Average-treatment-effect-in-the-treated weights were applied, with THA patients assigned a weight of 1 and HA patients weighted according to the estimated propensity score; weights were truncated at the 1st and 99th percentiles. Weighted Cox and logistic regression models were then used to assess mortality and good postoperative mobility, respectively. A doubly robust model additionally adjusted for age and the Charlson Comorbidity Index. Covariate balance before and after weighting, was assessed using absolute standardized mean differences (SMDs), with values below 0.10 regarded as adequate balance. Because the propensity-score distributions of the two procedures showed limited overlap (mean propensity score 0.61 vs. 0.06), ATT weighting substantially reduced but did not fully eliminate baseline imbalance, and the effective sample size of the weighted hemiarthroplasty group fell to approximately 43; residual imbalance persisted for several covariates, most notably comorbidity burden. For this reason, doubly robust models additionally adjusting for age and the Charlson Comorbidity Index were used as the primary propensity-based estimates. Balance diagnostics (SMDs before and after weighting) are provided in Online Resource 1.
Complete data were available for all 20 clustering variables and for mortality and survival outcomes; therefore, no imputation was performed. Postoperative mobility was not evaluable in 30 patients who died within 30 days, because early postoperative death precluded meaningful assessment of ambulation status; these patients were therefore excluded only from the postoperative mobility analysis, leaving 642 evaluable patients.
A two-sided p value < 0.05 was considered statistically significant. Descriptive and regression analyses were performed using IBM SPSS Statistics version 29 (IBM Corp., Armonk, NY, USA), and propensity-score analyses were performed in R using the survey package.
Hierarchical cluster analysis
Hierarchical clustering is an unsupervised method that groups patients according to overall similarity across multiple variables, without using outcome labels. An agglomerative hierarchical clustering procedure was used to identify latent patient subgroups. Outcome variables and postoperative complication-related variables were excluded from the clustering input to avoid outcome leakage. The final clustering model included 20 preoperative and intraoperative clinical and surgical variables.
Because the dataset included both continuous and categorical variables, the Gower distance metric was used. Four linkage methods were compared using the agglomerative coefficient (AC), and Ward’s D2 method showed the highest value (AC = 0.9741) and was selected. The optimal number of clusters was assessed using the gap statistic and silhouette analysis. Because the gap statistic increased monotonically without a clear elbow, silhouette analysis was used as the primary criterion. The highest mean silhouette width was observed for k = 2 (0.303), and this solution was selected because it provided the best internal separation among the tested solutions and yielded clinically interpretable phenotypes.
As a robustness check, the primary solution was compared with a model that additionally included the four postoperative variables; the solutions were concordant (adjusted Rand index = 0.59), and the cluster structure and outcome associations were preserved. After cluster assignment, patient profiles and outcomes were compared between clusters using the Mann–Whitney U test and chi-square test, as appropriate. Clustering was performed in R using the cluster, factoextra, dendextend, and ape packages.
Results
Of the 672 patients included in the study, 425 (63.2%) were female and 247 (36.8%) were male. The mean age was 76.85 ± 10.50 years. HA was performed in 586 patients (87.2%) and THA in 86 patients (12.8%). Most patients were classified as ASA III (79.5%), and the mean aCCI was 4.70 ± 1.50. Garden type IV and Pauwels type III fractures accounted for 68.0% and 81.0% of cases, respectively (Table 1). The median potential follow-up for the whole cohort, estimated using the reverse Kaplan–Meier method, was 2,150 days, and the maximum potential follow-up was 5,474 days.
Table 1.
Comparison of Demographic, Clinical, and Surgical Characteristics by Procedure Type
| Variable | Total (n = 672) | HA (n = 586) | THA (n = 86) | p |
|---|---|---|---|---|
| Demographics | ||||
| Age (years) | 76.85 ± 10.50 | 78.54 ± 9.37 | 65.33 ± 10.62 | < 0.001 |
| Female sex, n (%) | 425 (63.2) | 377 (64.3) | 48 (55.8) | 0.158 |
| BMI (kg/m²) | 26.61 ± 4.83 | 26.53 ± 4.86 | 27.22 ± 4.60 | 0.083 |
| Comorbidity | ||||
| CCI | 4.40 ± 1.77 | 4.76 ± 1.50 | 1.91 ± 1.50 | < 0.001 |
| aCCI | 4.70 ± 1.50 | 4.97 ± 1.33 | 2.85 ± 1.29 | < 0.001 |
| aCCI category | — | — | — | < 0.001 |
| Low (≤ 3), n (%) | 107 (15.9) | 50 (8.5) | 57 (66.3) | — |
| Moderate (4–5), n (%) | 388 (57.7) | 361 (61.6) | 27 (31.4) | — |
| High (≥ 6), n (%) | 177 (26.3) | 175 (29.9) | 2 (2.3) | — |
| ASA grade | — | — | — | < 0.001 |
| ASA II, n (%) | 118 (17.6) | 58 (9.9) | 60 (69.8) | — |
| ASA III, n (%) | 534 (79.5) | 509 (86.9) | 25 (29.1) | — |
| ASA IV, n (%) | 20 (3.0) | 19 (3.2) | 1 (1.2) | — |
| Dementia, n (%) | 90 (13.4) | 90 (15.4) | 0 (0.0) | < 0.001 |
| Fracture Characteristics | ||||
| Left side, n (%) | 411 (61.2) | 368 (62.8) | 43 (50.0) | 0.031 |
| Garden classification | — | — | — | 0.984 |
| Garden I, n (%) | 8 (1.2) | 7 (1.2) | 1 (1.2) | — |
| Garden II, n (%) | 52 (7.7) | 46 (7.8) | 6 (7.0) | — |
| Garden III, n (%) | 155 (23.1) | 136 (23.2) | 19 (22.1) | — |
| Garden IV, n (%) | 457 (68.0) | 397 (67.7) | 60 (69.8) | — |
| Pauwels classification | — | — | — | 0.179 |
| Pauwels I, n (%) | 28 (4.2) | 26 (4.4) | 2 (2.3) | — |
| Pauwels II, n (%) | 100 (14.9) | 82 (14.0) | 18 (20.9) | — |
| Pauwels III, n (%) | 544 (81.0) | 478 (81.6) | 66 (76.7) | — |
| Surgical Parameters | ||||
| Spinal anesthesia, n (%) | 556 (82.7) | 514 (87.7) | 42 (48.8) | < 0.001 |
| Cemented, n (%) | 350 (52.1) | 331 (56.5) | 19 (22.1) | < 0.001 |
| Head diameter (mm) | 47.78 ± 4.15 | 47.64 ± 3.98 | 48.77 ± 5.23 | 0.001 |
| Time to surgery (days) | 6.27 ± 11.98 | 6.29 ± 12.52 | 6.14 ± 7.34 | 0.495 |
| Operative time (min) | 131.70 ± 46.17 | 129.85 ± 44.75 | 144.27 ± 53.50 | 0.061 |
| LLD (mm) | 3.05 ± 8.40 | 3.17 ± 8.50 | 2.21 ± 7.75 | 0.350 |
| Laboratory Values | ||||
| SII | 1399 ± 1067 | 1433 ± 1051 | 1169 ± 1154 | 0.001 |
| Preop hemoglobin (g/dL) | 11.76 ± 1.61 | 11.71 ± 1.60 | 12.08 ± 1.63 | 0.062 |
| RDW-CV (%) | 14.67 ± 2.25 | 14.64 ± 2.28 | 14.82 ± 2.06 | 0.247 |
| Outcomes | ||||
| Mortality, n (%) | 437 (65.0) | 409 (69.8) | 28 (32.6) | < 0.001 |
| Survival (days) | 1174 ± 1082 | 1110 ± 1044 | 1608 ± 1236 | < 0.001 |
| Major complication, n (%) | 53 (7.9) | 41 (7.0) | 12 (14.0) | 0.043 |
| Loosening, n (%) | 30 (4.5) | 22 (3.8) | 8 (9.3) | 0.041 |
| Acetabular erosion, n (%) | 38 (5.7) | 32 (5.5) | 6 (7.0) | 0.750 |
| HO, n (%) | 106 (15.8) | 95 (16.2) | 11 (12.8) | 0.513 |
Compared with the HA group, patients in the THA group were younger and had lower CCI, aCCI, and SII values. Significant between-group differences were also observed in ASA grade, dementia, anesthesia type, surgical approach, cementation, mobility status, and complication rate, whereas sex, Garden classification, and Pauwels classification did not differ significantly (Table 1).
Continuous variables are presented as mean ± SD (Mann–Whitney U test). Categorical variables are presented as n (%) (chi-square test). Bold p values indicate statistical significance (p < 0.05). HA: hemiarthroplasty, THA: total hip arthroplasty, BMI: body mass index, CCI: Charlson Comorbidity Index, aCCI: age-adjusted CCI, SII: Systemic Immune-Inflammation Index, RDW-CV: red cell distribution width coefficient of variation, LLD: limb length discrepancy, HO: heterotopic ossification. For categorical variables with more than two levels (aCCI category, ASA grade, Garden classification, and Pauwels classification), the p value refers to the overall group comparison (chi-square test) on the classification heading row; dashes (—) indicate sublevels covered by that single test.
Hierarchical clustering using the Gower distance matrix and Ward’s D2 method identified two patient subgroups. The agglomerative coefficient was 0.9741, and the highest mean silhouette width was observed for k = 2 (0.303) (Online Resource 2). Cluster 1 included 607 patients (90.3%) and was defined as the high-risk group, whereas Cluster 2 included 65 patients (9.7%) and was defined as the low-risk group (Fig. 2).
Fig. 2.
Two patient phenotypes identified by hierarchical clustering. a Dendrogram of the hierarchical cluster analysis of 672 patients using Gower distance and Ward’s D2 linkage based on 20 preoperative and intraoperative variables. b Defining clinical features compared between clusters. c Clinical outcomes by cluster. d Median survival by cluster. Cluster 1 represents the high-risk phenotype, and Cluster 2 represents the low-risk phenotype. *p < 0.05, **p < 0.01, ***p < 0.001; ns, not significant
Cluster 1 was characterized by older age, higher comorbidity burden, ASA III predominance, dementia, and HA dominance. In contrast, Cluster 2 consisted of younger patients with lower comorbidity burden, ASA II predominance, absence of dementia, and THA dominance. No significant differences were observed between clusters for sex, RDW-CV, time to surgery, operative time, Garden classification, or Pauwels classification (Table 2).
Table 2.
Cluster Profiles: Comparison of Demographic, Clinical, and Surgical Characteristics Between Patient Subgroups Identified by Hierarchical Clustering Analysis
| Variable | Cluster 1 (n = 607) | Cluster 2 (n = 65) | p |
|---|---|---|---|
| Continuous Variables | |||
| Age (years) | 78.29 ± 9.54 | 63.38 ± 9.56 | < 0.001 |
| BMI (kg/m²) | 26.51 ± 4.84 | 27.56 ± 4.59 | 0.037 |
| CCI | 4.70 ± 1.54 | 1.60 ± 1.34 | < 0.001 |
| aCCI | 4.91 ± 1.37 | 2.72 ± 1.26 | < 0.001 |
| SII | 1,418 ± 1,050 | 1,223 ± 1214 | 0.011 |
| Preop hemoglobin (g/dL) | 11.70 ± 1.59 | 12.28 ± 1.68 | 0.008 |
| RDW-CV (%) | 14.66 ± 2.27 | 14.70 ± 2.01 | 0.576 |
| Head diameter (mm) | 47.73 ± 4.06 | 48.28 ± 5.12 | 0.037 |
| Time to surgery (days) | 6.28 ± 12.35 | 6.15 ± 7.78 | 0.500 |
| Operative time (min) | 130.9 ± 46.0 | 139.4 ± 47.4 | 0.267 |
| Categorical Variables | |||
| Female sex, n (%) | 389 (64.1) | 36 (55.4) | 0.212 |
| ASA III, n (%) | 523 (86.2) | 11 (16.9) | < 0.001 |
| Procedure type: HA, n (%) | 585 (96.4) | 1 (1.5) | < 0.001 |
| Spinal anesthesia, n (%) | 523 (86.2) | 33 (50.8) | < 0.001 |
| Cemented, n (%) | 346 (57.0) | 4 (6.2) | < 0.001 |
| Dementia, n (%) | 90 (14.8) | 0 (0.0) | 0.002 |
| Garden classification | — | — | 0.708 |
| Garden I, n (%) | 7 (1.2) | 1 (1.5) | — |
| Garden II, n (%) | 46 (7.6) | 6 (9.2) | — |
| Garden III, n (%) | 137 (22.6) | 18 (27.7) | — |
| Garden IV, n (%) | 417 (68.7) | 40 (61.5) | — |
| Pauwels classification | — | — | 0.402 |
| Pauwels I, n (%) | 27 (4.4) | 1 (1.5) | — |
| Pauwels II, n (%) | 88 (14.5) | 12 (18.5) | — |
| Pauwels III, n (%) | 492 (81.1) | 52 (80.0) | — |
Continuous variables are presented as mean ± SD (Mann–Whitney U test). Categorical variables are presented as n (%) (chi-square test). Bold p values indicate statistical significance (p < 0.05). Cluster 1 represents the high-risk group characterized by older age, higher comorbidity burden, ASA III predominance, and HA. Cluster 2 represents the low-risk group characterized by younger age, lower comorbidity, ASA II predominance, and THA. HA: hemiarthroplasty, THA: total hip arthroplasty, BMI: body mass index, CCI: Charlson Comorbidity Index, aCCI: age-adjusted CCI, SII: Systemic Immune-Inflammation Index, RDW-CV: red cell distribution width coefficient of variation. For Garden and Pauwels classifications, the p value refers to the overall group comparison (chi-square test) on the classification heading row; dashes (—) indicate sublevels covered by that single test.
Clinical outcomes differed significantly between clusters. Mortality was higher in Cluster 1 than in Cluster 2 (69.2% vs. 26.2%, p < 0.001), median survival was shorter (675 vs. 1,054 days, p < 0.001), and good mobility was less frequent (42.7% vs. 95.3%, p < 0.001). Major complication rates did not differ significantly between clusters (7.6% vs. 10.8%, p = 0.506) (Table 3).
Table 3.
Comparison of Clinical Outcomes Between Patient Clusters
| Outcome | Cluster 1 (n = 607) | Cluster 2 (n = 65) | p |
|---|---|---|---|
| Mortality, n (%) | 420 (69.2) | 17 (26.2) | < 0.001 |
| Median survival (days) | 675 | 1,054 | < 0.001 |
| Mean survival (days) | 1,037.1 ± 991.0 | 1,478.4 ± 1,156.5 | < 0.001 |
| Good mobility, n (%)* | 247 (42.7) | 61 (95.3) | < 0.001 |
| Major complication, n (%) | 46 (7.6) | 7 (10.8) | 0.506 |
Survival times were compared using the Mann–Whitney U test. *Good mobility was defined as independent ambulation or ambulation with a single walking aid. Mobility analysis was performed after excluding patients who died within 30 days of surgery. Complication subtypes are reported descriptively due to small sample sizes. The difference in major complication rates between clusters was not statistically significant (p = 0.506); given the small number of events, no causal interpretation of this numerical difference is made. Bold p values indicate statistical significance (p < 0.05)
The overall mortality rate was 65.0% (n = 437). Median survival was longer in the THA group than in the HA group (1,062 vs. 661 days; log-rank p < 0.001). Survival also differed significantly across aCCI categories (log-rank p < 0.001) (Fig. 3).
Fig. 3.
Kaplan-Meier survival analysis. a Survival by procedure type (log-rank p < 0.0001). Cumulative survival was higher in the THA group than in the HA group. b Survival by age-adjusted Charlson Comorbidity Index category (log-rank p < 0.001). Patients with low comorbidity (aCCI ≤ 3) demonstrated markedly superior survival. Number at risk tables are presented below each curve
In the multivariable Cox regression model, age, male sex, dementia, ASA score, and RDW-CV were independent predictors of mortality. Procedure type was not an independent predictor of mortality after adjustment (p = 0.058) (Table 4).
Table 4.
Multivariate Cox Proportional Hazard Regression Analysis for Mortality
| Variable | B | SE | Wald | p | HR | 95% CI |
|---|---|---|---|---|---|---|
| Age | 0.043 | 0.006 | 51.36 | < 0.001 | 1.044 | 1.032–1.056 |
| Male sex | 0.354 | 0.099 | 12.79 | < 0.001 | 1.424 | 1.171–1.731 |
| Dementia | 0.315 | 0.144 | 4.79 | 0.029 | 1.370 | 1.033–1.815 |
| ASA score | 0.356 | 0.139 | 6.56 | 0.010 | 1.427 | 1.087–1.874 |
| RDW-CV | 0.051 | 0.020 | 6.50 | 0.011 | 1.052 | 1.012–1.093 |
| Procedure type (THA) | –0.417 | 0.220 | 3.59 | 0.058 | 0.659 | 0.427–1.015 |
Bold p values indicate statistical significance (p < 0.05). Procedure type was not an independent predictor of mortality after adjusting for age, sex, dementia, ASA score, and RDW-CV
HR hazard ratio, CI confidence interval, SE standard error, RDW-CV red cell distribution width coefficient of variation, THA total hip arthroplasty
Because cementation differed substantially between groups, a cementation-adjusted Cox sensitivity model was additionally performed. In this model, procedure type remained non-significant for mortality (HR = 0.690, 95% CI: 0.446–1.069, p = 0.097), and cementation itself was not independently associated with mortality (HR = 1.144, 95% CI: 0.939–1.394, p = 0.182) (Online Resource 3).
Postoperative mobility analysis was performed in 642 patients with a recorded postoperative mobility assessment (after excluding early deaths). Good mobility was more frequent after THA than after HA (92.9% vs. 41.1%, p < 0.001). Binary logistic regression identified procedure type, aCCI, dementia, and age as independent predictors of good postoperative mobility (Table 5). Regression diagnostics are reported in Online Resource 4.
Table 5.
Binary Logistic Regression Analysis for Good Postoperative Mobility
| Variable | B | SE | Wald | p | OR | 95% CI |
|---|---|---|---|---|---|---|
| Procedure type (THA) | 1.372 | 0.488 | 7.91 | 0.005 | 3.944 | 1.516–10.260 |
| aCCI | –0.611 | 0.093 | 43.15 | < 0.001 | 0.543 | 0.452–0.652 |
| Dementia | –0.904 | 0.330 | 7.50 | 0.006 | 0.405 | 0.212–0.773 |
| Age | –0.034 | 0.011 | 9.56 | 0.002 | 0.967 | 0.946–0.988 |
OR odds ratio, CI confidence interval, SE standard error, aCCI age-adjusted Charlson Comorbidity Index, THA total hip arthroplasty. Model: p < 0.001, Nagelkerke R² = 0.364. Good mobility was defined as independent ambulation or ambulation with a single walking aid. Analysis was performed on 642 patients after excluding those who died within 30 days of surgery. Bold p values indicate statistical significance (p < 0.05)
Major complications occurred in 53 patients (7.9%), including infection in 22 patients (3.3%), dislocation in 13 (1.9%), postoperative periprosthetic fracture in 9 (1.3%), revision surgery in 7 (1.0%), and intraoperative fracture in 2 (0.3%).
Weighting improved covariate balance for several baseline variables, particularly age and ASA physical status; however, residual imbalance remained for comorbidity measures, including CCI/aCCI. Therefore, the doubly robust sensitivity model additionally adjusted for age and CCI, and the propensity-score findings were interpreted cautiously. Standardized mean differences before and after weighting are presented in Online Resource 1; weighting reduced but did not fully remove covariate imbalance, and the doubly robust estimates are therefore emphasized. In the propensity-score–weighted sensitivity analysis, the crude survival advantage of THA was progressively attenuated and was no longer significant after doubly robust adjustment (crude HR 0.310; inverse-probability-weighted HR 0.516, p = 0.015; doubly robust HR 0.766, p = 0.319). In contrast, the association between THA and good mobility remained significant across all models (crude OR 18.86; inverse-probability-weighted OR 5.79; doubly robust OR 4.65, p = 0.003).
Receiver operating characteristic (ROC) analysis showed limited discriminatory performance of individual biomarkers for mortality. The highest area under the curve (AUC) values were observed for aCCI (AUC = 0.598) and RDW-CV (AUC = 0.572), whereas SII was not statistically significant (AUC = 0.535) (Table 6).
Table 6.
Receiver Operating Characteristic (ROC) Analysis for Mortality Prediction
| Biomarker | AUC | SE | 95% CI | p |
|---|---|---|---|---|
| aCCI | 0.598 | 0.023 | 0.554–0.643 | < 0.001 |
| RDW-CV (%) | 0.572 | 0.023 | 0.527–0.617 | 0.002 |
| SII | 0.535 | 0.023 | 0.490–0.580 | 0.135 |
| Preop hemoglobin (g/dL) | 0.425 | 0.023 | 0.380–0.470 | 0.001 |
AUC area under the curve, CI confidence interval, SE standard error, aCCI age-adjusted Charlson Comorbidity Index, RDW-CV red cell distribution width coefficient of variation, SII Systemic Immune-Inflammation Index
SII was not a significant predictor of mortality. Preoperative hemoglobin showed an inverse association (AUC < 0.5), indicating that lower hemoglobin levels were associated with higher mortality. Bold p values indicate statistical significance (p < 0.05).
Discussion
In this study, hierarchical clustering identified two clinically meaningful phenotypes among 672 patients who underwent arthroplasty for intracapsular femoral neck fractures. The high-risk phenotype was characterized by advanced age, greater comorbidity burden, ASA III predominance, dementia, cementation, and HA dominance, whereas the low-risk phenotype consisted of younger patients with lower comorbidity burden, ASA II predominance, preserved cognitive status, low cementation rate, and THA dominance. These phenotypes differed markedly in mortality, survival, and postoperative mobility, supporting the concept that patients with femoral neck fractures represent a heterogeneous population rather than a uniform clinical group [9, 10, 13]. Importantly, because the clusters were derived from both preoperative and intraoperative variables, they should be interpreted as prognostic phenotypes reflecting real-world treatment patterns rather than as a purely preoperative tool for selecting the arthroplasty procedure.
Regarding the primary objective, mortality and survival differed markedly between the derived phenotypes, with substantially higher mortality and shorter survival in the high-risk cluster. Although crude survival was longer after THA than after HA, procedure type was not an independent predictor of mortality after adjustment, and the apparent survival advantage of THA was attenuated in the propensity-score–weighted sensitivity analysis. This suggests that survival differences were driven primarily by baseline patient selection rather than by a direct mortality effect of the procedure, consistent with known selection patterns in which younger, less comorbid, and cognitively preserved patients are more likely to undergo THA [16, 17]. The additional cementation-adjusted Cox sensitivity model further supported this interpretation, as procedure type remained non-significant for mortality and cementation itself was not independently associated with mortality. These findings are consistent with randomized trials, meta-analyses, and population-based studies reporting no consistent mortality advantage of THA over HA in displaced femoral neck fractures [4, 6, 17–19]. Instead, mortality in this cohort was more strongly related to age, male sex, dementia, ASA score, and RDW-CV, which is in line with previous studies identifying frailty, comorbidity, cognitive impairment, and haematologic or inflammatory markers as important mortality predictors after hip fracture [20–22]. Although SII and other isolated biomarkers have previously been linked to prognosis after hip fracture, their limited discriminatory performance in the present study supports the view that multidimensional clinical models are more informative than single-biomarker approaches [23–28]. The overall mortality rate of 65.0% should also be interpreted in the context of cumulative all-cause mortality over a long observation period rather than as a fixed one-year mortality estimate [29].
Regarding secondary outcomes, THA was independently associated with better postoperative mobility, whereas major complication rates did not differ significantly between the two phenotypes. This supports the interpretation that the main potential advantage of THA in selected patients relates to postoperative mobility rather than survival. These findings are consistent with previous trials and meta-analyses suggesting that THA may provide better postoperative mobility or functional outcomes in appropriately selected patients [3, 4, 6, 30–33], although it may also be associated with a greater risk of hip-specific complications such as dislocation [6, 17, 31, 32]. In the present cohort, the limited number of major complication events precluded multivariable complication modelling; therefore, complication findings should be interpreted descriptively. From a clinical perspective, these results suggest that perioperative risk assessment and patient counselling should be based on the multidimensional patient profile rather than on procedure type alone. In high-risk patients, minimizing perioperative burden and setting realistic mobility goals may be the priority, whereas in lower-risk patients with greater physiological reserve, the potential mobility benefit of THA may be more relevant.
The main limitations of this study include its retrospective single-center design, non-randomized treatment allocation, limited number of complications, and the use of a clinical mobility scale rather than a standardized hip-specific functional score. Because the study was conducted at a single tertiary referral center in one country, the findings may not be fully generalizable to other healthcare systems, institutions with different arthroplasty practices, or populations with different frailty and rehabilitation profiles. Although propensity-score weighting was used to mitigate treatment-selection bias, residual confounding remained; dementia could not be incorporated into the propensity model because no THA patient had dementia, and comorbidity indices were not fully balanced after weighting reflecting limited propensity-score overlap between procedures. Moreover, postoperative mobility was evaluated only in 30-day survivors; because patients who died early could not undergo meaningful mobility assessment, this analysis is conditioned on survival and is therefore susceptible to survivor bias. The cohort also spanned a 15-year period from 2009 to 2023, during which implant designs, cementation practices, surgical technique, anesthesia protocols, perioperative care, rehabilitation pathways, and procedure-selection criteria may have changed. Therefore, temporal trends may have influenced both treatment allocation and outcomes. Although we performed an additional cementation-adjusted Cox sensitivity analysis, detailed intraoperative data on bone cement implantation syndrome were not available; therefore, the potential effect of cement-related physiological events could not be directly evaluated. For these reasons, the findings should be interpreted as associations rather than causal effects. Nevertheless, the large sample size, the use of a clustering method appropriate for mixed data types, and the strong association between the identified phenotypes and clinically relevant outcomes represent important strengths.
Conclusion
Hierarchical clustering identified two distinct clinical phenotypes among patients undergoing arthroplasty for femoral neck fracture, with significant differences in mortality, survival, and postoperative mobility. Procedure type was not an independent predictor of mortality, suggesting that the crude survival difference between THA and HA largely reflects selection bias, whereas THA was independently associated with better postoperative mobility, although residual confounding cannot be fully excluded. These findings support individualised, multidimensional risk assessment rather than single-variable decision-making; because the phenotypes incorporate intraoperative variables, they are prognostic rather than a preoperative selection tool. Future multicenter prospective studies are warranted to validate these findings.
Supplementary Information
Below is the link to the electronic supplementary material.
Author contributions
E.I. and O.T.E. conceived the study. E.I., M.A.B., and O.T.E. designed the methodology. Y.A., B.B., and H.U. curated the data. Y.A., B.B., H.U., and M.A.B. conducted the investigation. I.D. and O.T.E. performed the surgical procedures. M.A.B. and Y.A. carried out the radiological evaluation. E.I. and M.A.B. performed the formal analysis. E.I. wrote the main manuscript text. Y.A., B.B., M.A.B., H.U., I.D., and O.T.E. reviewed and edited the manuscript. O.T.E. and I.D. supervised the study. E.I. administered the project. All authors reviewed and approved the final manuscript.
Funding
No funding was received for this study.
Data availability
The datasets generated and/or analyzed during the current study are not publicly available because they contain potentially identifiable patient information and are subject to institutional restrictions. De-identified data may be made available from the corresponding author on reasonable request and with permission of the relevant institutional authority.
Declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Ethical approval
The study protocol was approved by the local institutional ethics committee (approval date: 3 February 2026; decision no: 3272) and was conducted in accordance with the Declaration of Helsinki.
Informed Consent
The requirement for informed consent was waived owing to the retrospective design of the study.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Luo S, Qin W, Yu L et al (2023) Total hip arthroplasty versus hemiarthroplasty in the treatment of active elderly patients over 75 years with displaced femoral neck fractures: a retrospective study. BMC Musculoskelet Disord 24. 10.1186/s12891-023-06860-6 [DOI] [PMC free article] [PubMed]
- 2.Kates SL, Kates OS, Mendelson DA (2007) Advances in the medical management of osteoporosis. Injury 38:17–23. 10.1016/j.injury.2007.08.007 [DOI] [PubMed] [Google Scholar]
- 3.Tang X, Wang D, Liu Y et al (2020) The comparison between total hip arthroplasty and hemiarthroplasty in patients with femoral neck fractures: a systematic review and meta-analysis based on 25 randomized controlled trials. J Orthop Surg Res 15. 10.1186/s13018-020-02122-6 [DOI] [PMC free article] [PubMed]
- 4.Ekhtiari S, Gormley J, Axelrod DE et al (2020) Total Hip Arthroplasty Versus Hemiarthroplasty for Displaced Femoral Neck Fracture: A Systematic Review and Meta-Analysis of Randomized Controlled Trials. J Bone Jt Surg 102:1638–1645. 10.2106/JBJS.20.00226 [DOI] [PubMed] [Google Scholar]
- 5.Lewis DP, Wæver D, Thorninger R, Donnelly WJ (2019) Hemiarthroplasty vs Total Hip Arthroplasty for the Management of Displaced Neck of Femur Fractures: A Systematic Review and Meta-Analysis. J Arthroplasty 34:1837–1843e2. 10.1016/j.arth.2019.03.070 [DOI] [PubMed] [Google Scholar]
- 6.Bhandari M, Einhorn TA, Guyatt G et al (2019) Total Hip Arthroplasty or Hemiarthroplasty for Hip Fracture. N Engl J Med 381:2199–2208. 10.1056/NEJMoa1906190 [DOI] [PubMed] [Google Scholar]
- 7.Peng W, Bi N, Zheng J, Xi N (2020) Does total hip arthroplasty provide better outcomes than hemiarthroplasty for the femoral neck fracture? A systematic review and meta-analysis. Chin J Traumatol – 23 356–362. 10.1016/j.cjtee.2020.09.005 [DOI] [PMC free article] [PubMed]
- 8.Wang Z, Bhattacharyya T (2017) Outcomes of Hemiarthroplasty and Total Hip Arthroplasty for Femoral Neck Fracture: A Medicare Cohort Study. J Orthop Trauma 31:260–263. 10.1097/BOT.0000000000000814 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Eastwood EA, Magaziner J, Wang J et al (2002) Patients with hip fracture: Subgroups and their outcomes. J Am Geriatr Soc 50:1240–1249. 10.1046/j.1532-5415.2002.50311.x [DOI] [PubMed] [Google Scholar]
- 10.Penrod JD, Litke A, Hawkes WG et al (2007) Heterogeneity in hip fracture patients: Age, functional status, and comorbidity. J Am Geriatr Soc 55:407–413. 10.1111/j.1532-5415.2007.01078.x [DOI] [PubMed] [Google Scholar]
- 11.Zhong H, Wang B, Wang D et al (2021) The application of machine learning algorithms in predicting the length of stay following femoral neck fracture. Int J Med Inf 155. 10.1016/j.ijmedinf.2021.104572 [DOI] [PubMed]
- 12.Li Y, Chen M, Lv H et al (2021) A novel machine-learning algorithm for predicting mortality risk after hip fracture surgery. Injury 52:1487–1493. 10.1016/j.injury.2020.12.008 [DOI] [PubMed] [Google Scholar]
- 13.Kruse C, Eiken P, Vestergaard P (2017) Clinical fracture risk evaluated by hierarchical agglomerative clustering. Osteoporos Int 28:819–832. 10.1007/s00198-016-3828-8 [DOI] [PubMed] [Google Scholar]
- 14.Charlson ME, Pompei P, Ales KL, MacKenzie CR (1987) A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. J Chronic Dis 40:373–383. 10.1016/0021-9681(87)90171-8 [DOI] [PubMed] [Google Scholar]
- 15.Quan H, Li B, Couris CM et al (2011) Updating and validating the charlson comorbidity index and score for risk adjustment in hospital discharge abstracts using data from 6 countries. Am J Epidemiol 173:676–682. 10.1093/aje/kwq433 [DOI] [PubMed] [Google Scholar]
- 16.Tohidi M, Mann SM, Groome PA (2023) Total hip arthroplasty versus hemiarthroplasty for treatment of femoral neck fractures: a population-based analysis of practice variation in Ontario, Canada. Bone Joint J 105–B:180–189. 10.1302/0301-620X.105B2.BJJ-2022-0878.R1 [DOI] [PubMed] [Google Scholar]
- 17.Edelstein AI, Dillingham TR, McGinley EL, Pezzin LE (2023) Hemiarthroplasty Versus Total Hip Arthroplasty for Femoral Neck Fracture in Elderly Patients: Twelve-Month Risk of Revision and Dislocation in an Instrumental Variable Analysis of Medicare Data. J Bone Jt Surg 105:1695–1702. 10.2106/JBJS.23.00247 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Chammout G, Kelly-Pettersson P, Hedbeck CJ et al (2019) HOPE-Trial: Hemiarthroplasty Compared with Total Hip Arthroplasty for Displaced Femoral Neck Fractures in Octogenarians A Randomized Controlled Trial. JBJS Open Access 4:E0059. 10.2106/JBJS.OA.18.00059 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Mukka S, Sjöholm P, Chammout G et al (2019) External Validity of the HOPE-Trial Hemiarthroplasty Compared with Total Hip Arthroplasty for Displaced Femoral Neck Fractures in Octogenarians. JBJS Open Access 4:E0061. 10.2106/JBJS.OA.18.00061 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Fakler JKM, Pieroh P, Höch A et al (2024) Predictors of long-term mortality in older patients with hip fractures managed by hemiarthroplasty: a 10-year study based on a population registry in Saxony, Germany. Patient Saf Surg. 10.1186/s13037-024-00398-9. 18: [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Nguyen BTT, Tran DNA, Nguyen TT et al (2024) The Association between Red Blood Cell Distribution Width and Mortality Risk after Hip Fracture: A Meta-Analysis. Med (Kaunas) 60:485. 10.3390/medicina60030485 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Wang N-J, Zhang Y-M, Zhang B-F (2023) The Association Between Red Cell Distribution Width (RDW) and All-Cause Mortality in Elderly Patients with Hip Fractures: A Retrospective Cohort Study. Int J Gen Med 16:3555–3566. 10.2147/ijgm.s417079 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Liu ZJ, Li GH, Wang JX et al (2024) Prognostic value of the systemic immune-inflammation index in critically ill elderly patients with hip fracture: evidence from MIMIC (2008–2019). 10.3389/fmed.2024.1408371. Front Med 11: [DOI] [PMC free article] [PubMed]
- 24.Wang ZC, Jiang W, Chen X et al (2021) Systemic immune-inflammation index independently predicts poor survival of older adults with hip fracture: a prospective cohort study. BMC Geriatr 21:155. 10.1186/s12877-021-02102-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Su YH, Huang YW, Kuo YJ et al (2025) Prognostic value of three inflammatory markers for mortality after hip fracture in the geriatric population. Sci Rep 15. 10.1038/s41598-025-29665-0 [DOI] [PMC free article] [PubMed]
- 26.Choi YH, Wong PY, Fong MK et al (2025) A potential novel predictor of 1-year mortality in elderly hip fracture patients-albumin haemoglobin index. Osteoporos Int 36:1557–1564. 10.1007/s00198-025-07566-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Fisher A, Fisher L, Srikusalanukul W (2024) Prediction of Osteoporotic Hip Fracture Outcome: Comparative Accuracy of 27 Immune–Inflammatory–Metabolic Markers and Related Conceptual Issues. J Clin Med 13. 10.3390/jcm13133969 [DOI] [PMC free article] [PubMed]
- 28.Fisher A, Fisher L, Srikusalanukul W, Smith PN (2018) Usefulness of simple biomarkers at admission as independent indicators and predictors of in-hospital mortality in older hip fracture patients. Injury 49:829–840. 10.1016/j.injury.2018.03.005 [DOI] [PubMed] [Google Scholar]
- 29.Asrian G, Suri A, Rajapakse C (2024) Machine learning-based mortality prediction in hip fracture patients using biomarkers. J Orthop Res 42:395–403. 10.1002/jor.25675 [DOI] [PubMed] [Google Scholar]
- 30.Migliorini F, Trivellas A, Driessen A et al (2020) Hemiarthroplasty versus total arthroplasty for displaced femoral neck fractures in the elderly: meta-analysis of randomized clinical trials. Arch Orthop Trauma Surg 140:1695–1704. 10.1007/s00402-020-03409-3 [DOI] [PubMed] [Google Scholar]
- 31.Hansson S, Bülow E, Garland A et al (2020) More hip complications after total hip arthroplasty than after hemiarthroplasty as hip fracture treatment: analysis of 5,815 matched pairs in the Swedish Hip Arthroplasty Register. Acta Orthop 91:133–138. 10.1080/17453674.2019.1690339 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Burgers PTPW, Van Geene AR, Van Den Bekerom MPJ et al (2012) Total hip arthroplasty versus hemiarthroplasty for displaced femoral neck fractures in the healthy elderly: A meta-Analysis and systematic review of randomized trials. Int Orthop 36:1549–1560. 10.1007/s00264-012-1569-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Frihagen F, Comeau-Gauthier M, Axelrod D et al (2022) Who, if anyone, may benefit from a total hip arthroplasty after a displaced femoral neck fracture? A post hoc subgroup analysis of the HEALTH trial. Bone Jt Open 3:611–617. 10.1302/2633-1462.38.BJO-2022-0074.R1 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets generated and/or analyzed during the current study are not publicly available because they contain potentially identifiable patient information and are subject to institutional restrictions. De-identified data may be made available from the corresponding author on reasonable request and with permission of the relevant institutional authority.