A novel risk score for predicting prolonged length of stay following pediatric kidney transplant

Kidney transplants (KT) are accepted as the kidney replacement therapy of choice for children with kidney failure. The surgery itself may be more difficult especially in small children, and often leads to significant hospital stays. There is little research on predicting prolonged length of stay (LOS) in children. We aim to examine the factors associated with prolonged LOS following pediatric KT to help clinicians make informed decisions, better counsel families, and potentially reduce preventable causes of prolonged stay. We retrospectively analyzed the United Network for Organ Sharing database for all KT recipients less than 18 years old between January 2014 and July 2022 (n = 3693). Donor and recipient factors were tested in univariate and multivariate logistic analysis using stepwise elimination of non-significant factors to create a final regression model predicting LOS longer than 14 days. Values were assigned to significant factors to create risk scores for each individual patient. In the final model, only primary diagnosis of focal segmental glomerulosclerosis, dialysis prior to KT, geographic region, and recipient weight prior to KT were significant predictors of LOS longer than 14 days. The C-statistic of the model is 0.7308. The C-statistic of the risk score is 0.7221. Knowledge of the risk factors affecting prolonged LOS following pediatric KT can help identify patients at risk of increased resource use and potential hospital-acquired complications. Using our index, we identified some of these specific risk factors and created a risk score that can stratify pediatric recipients into low, medium, or high risk groups. A higher resolution version of the Graphical abstract is available as Supplementary information A higher resolution version of the Graphical abstract is available as Supplementary information


Introduction
In children with kidney failure, kidney transplantation (KT) is accepted as the kidney replacement therapy of choice [1].In 2021, 819 kidney transplants were performed on recipients less than 18 years of age, with over one thousand children added to the transplant waitlist, according to Organ Procurement and Transplantation Network (OTPN) data as of November 23, 2022.Transplant surgery, especially in small children, often leads to significant hospital stays, but little research exists predicting which pediatric patients are at risk of prolonged length of stay (LOS).Knowing what factors are correlated with prolonged hospital stay following KT will help clinicians make more informed decisions, better counsel families, allow for better planning and management of patients, and potentially help reduce the preventable causes of prolonged stay.
Currently, most of the research regarding LOS following KT has focused on the adult population.Specifically, many multivariate analyses exist showing correlations between various donor and recipient factors and prolonged LOS [2][3][4][5][6].Furthermore, specific measures such as frailty [7,8], Surgical Apgar Score [9], and Short Physical Performance Battery [10] have been studied as potential objective surrogate markers to predict prolonged stay in the adult population.To our knowledge, few if any multivariate analyses looking at prolonged hospital stay following KT specifically in children have been done and none within the last 20 years.
Currently, there is no reliable index or measure used to accurately predict prolonged hospital stay in children following KT.Since LOS is a key determining factor of cost [11] and can predict poor graft and patient survival, especially when prolonged longer than 2 weeks [12], a risk factor score or equivalent for children is vital for optimizing outcomes in pediatric KT.
The aim of this project was to identify the factors associated with prolonged hospital stay following pediatric KT.With this information, we created a risk scoring system that can be used to predict whether an individual patient may be at a high risk of prolonged stay following the surgical procedure.Additionally, we seek to show that the kidney donor profile index (KDPI), a separate index used to aid in the allocation system for kidney transplants, is not an accurate predictor of prolonged stay.

Study population
We performed a retrospective analysis of the United Network for Organ Sharing (UNOS) database.Specifically, we used the kidney-pancreas dataset containing deidentified patient-level data collected by the OPTN.We examined data for all kidney transplant recipients less than 18 years of age between January 2014 and July 2022.Patients with a missing discharge date (n = 110) were excluded.Living donor transplants (n = 1970), recipients with a history of prior transplant (n = 437), recipients of combined or multi-visceral transplants (n = 304), recipients on multiple organ transplant waitlists (n = 3), and recipients who died less than 14 days post-transplant (n = 2) were excluded from the analysis to prevent the effect of these confounding variables.A total of 3693 recipients were followed from transplant date to hospital discharge and included in the analysis.Demographic information is contained in Table 1.

Statistical analysis
Data were analyzed and figures created using a standard statistical analysis package, Stata/BE® 17.0 (Stata Corp, College Station, TX 77845, USA).The primary effect we examined was length of hospital stay following KT, and logistic regression was used as time-to-event analysis.LOS was calculated as the difference between the date of transplant and date of hospital discharge.Fourteen days was chosen as the cut-off for prolonged LOS as this was approximately the 85 th percentile in our cohort, was shown to be associated with poor outcomes in previous studies [12], and allows for a simplified clinical cutoff.Hospital LOS greater than 14 days was the dependent variable.Donor and recipient factors were the independent variables.Independent variables were first tested in logistic regression univariate analysis, and those found to be significant were then assessed in a multivariate model.Contingency table analysis was used to test for significance in the multivariate model, and nonsignificant variables were eliminated to create the final logistic regression model.Significance was set at a value of 0.05.All p-values reported are two-sided.

Risk factors
Risk factors considered in the model are summarized in Table 2. Continuous variables were categorized with references chosen based on clinical relevance.Recipient risk factors were collected at the time of transplant unless otherwise noted.

Training and validation groups
Recipients were randomly assigned into one of two groups, either training or validation.Two-thirds of participants (n = 2462) were included in the training group, and this data set was used to create the risk score.The validation group (n = 1231) was then used as a reference to compare the accuracy of the model.

Risk score
Logistic regression analysis was used to determine the predictors of hospital stay longer than 14 days.Factors that were significant in the final multivariate 14-day logistic regression were then used to create the risk score.Starting at a score of 0, each increase or decrease in the odds ratio by 0.1 in each variable adds or subtracts 1 point from the risk score, respectively, with the final score being the sum of all points.We then stratified recipients into terciles based on their risk scores to create low, medium, and high risk of prolonged stay groups.Model accuracy was assessed using receiver operator characteristic curve (ROC) on both the training and validation groups.C-statistics were compared to validate the model.

Sensitivity analysis
We ran a post-hoc sensitivity analysis to identify if our results were overly dependent on the geographic region variable and if transplant center-specific factors were impacting our results.This was done by running our final risk score statistical model without the OPTN geographic region variable and comparing the results to our original model.

KDPI score concordance
We compared our generated risk score index to the kidney donor profile index (KDPI) score of each recipient to test if one, KDPI was an accurate predictor of prolonged hospital stay, and two, if our risk score could better predict this.We used scatter plots and linear regression to assess for the correlation between the risk score index and KDPI.KDPI was then used to assess for prolonged LOS using ROC, and the C-statistics were compared between the two indices.

Study population
We analyzed 3693 pediatric KT recipients in the USA between January 2014 and July 2022.Demographic characteristics of KT recipients and their donors are highlighted in Table 1.The median LOS was 8 (IQR 6) days, for a total of 38,917 days in the hospital in our cohort.

Data entry rate
Overall, variables were well populated, with a mean population of 99.84%.The lowest-populated variable was cold ischemia time, which was 98.21%.Full data entry rate information is included in Table 2.

Univariate and multivariate
Univariate logistic analysis was run on 35 recipient and donor factors to predict hospital stay longer than 14 days.Factors considered in the univariate analysis are included in Table 2. Factors found to be significant in the univariate were then included in a multivariate model predicting stay greater than 14 days.Factors insignificant on multivariate were eliminated, and a final multivariate risk score model was run from which the risk score was created.After the  3 and 4, respectively.All predictive models were fit on the training data set alone.

Risk score
The risk score model included variables found to be significant in predicting prolonged LOS (greater than 14 days) following stepwise elimination in the univariate and multivariate models.The C-statistic of the risk score model is 0.7308.Assigned points for risk score calculations can be found in Table 4.
The mean risk score in our cohort was 11.21 (11.15) and the median was 11.The ROC C-statistic when using the risk score is 0.7221.The validation data set was then used to test the accuracy of the score.When using the score on the training set alone the ROC was 0.7268, and on the validation set it was 0.7128 (p-value = 0.56).

Risk group stratification
Recipients were grouped into three equal groups, a low, medium, and high risk of prolonged hospital stay, based on their risk scores.Scores less than or equal to 5 defined the low-risk group, scores 6-17 the medium, and scores 18 and above the high.The mean risk score and LOS for each group is included in Table 5, along with the information for the 10 th and 90 th percentiles of risk score.Figure 1 shows the Kaplan-Meier curve for each risk group with the outcome of interest being hospital discharge.Figure 2 does the same for the 10 th and 90 th percentiles.

Sensitivity analysis
In our post hoc sensitivity analysis without the geographic region variable, the final risk score model shows no change in significance of the other variables.The primary diagnosis of FSGS, dialysis prior to transplant, and recipient weight are still significant predictors of prolonged LOS.The C-statistic of our risk score model without geographic region was 0.7186, which is similar to the C-statistic of the full model including geographic region.This shows that our model is not overly reliant on geographic region or being unduly influenced by center-specific differences as the model is still significant when excluding geographic region.

KDPI score
Using linear regression, we found that KDPI was not a good predictor of our risk score (RMSE = 0.09).We also found that KDPI was a poor predictor of prolonged length of stay, generating an ROC C-statistic of 0.5230, which was significantly different from the ROC of our risk score (p-value = 0.00).

Odds ratio calculation
We provided the risk score to be a simple to implement method of predicting prolonged hospital stay and aid in clinical decision-making.Here, we provide our full formula to calculate an odds ratio for hospital stay greater than 14 days using the risk score model.Each recipient starts with an odds ratio of

Discussion
Using univariate and multivariate logistic regression analysis, we generated a risk score that can predict prolonged hospital LOS in pediatric KT recipients.Our model and risk score implement four recipient factors: primary diagnosis of FSGS, dialysis prior to transplant, geographic region within the USA, and weight prior to transplant.Since our model does not include any donor factors, our score can be used to predict prolonged LOS prior to a donor being identified.Our risk score allowed us to stratify patients into low, medium, or high risk categories, which accurately predicted their relative risk of prolonged LOS compared to their peers.This is especially true for those at the highest and lowest risk, represented by the tenth and ninetieth percentile of our score.Finally, KDPI was demonstrated to be a poor predictor of prolonged LOS risk compared to our score.
There is prior evidence to suggest that children with a primary diagnosis of FSGS experience worse outcomes following KT, specifically, decreased graft survival [13,14].In Black children, diagnosis of FSGS has been shown to correlate with an increased LOS following KT [15].Our study validates and expands previous findings by showing that a primary diagnosis of FSGS correlates with increased LOS, regardless of a child's race.This could be because if FSGS recurs, heavy proteinuria usually happens within 24-48 hours of surgery, and subsequent treatment with apheresis and other IV medications likely contribute to prolonged LOS.
It is well known that earlier KT, especially prior to dialysis, can improve outcomes in children with kidney failure.Specifically, preemptive transplant has been shown to decrease risk of graft failure, graft loss, acute rejection, and death [16,17].Duration of dialysis appears to have an influence, with longer durations being associated with worse outcomes [16,18].This finding could be attributed to many different factors, but due to the observational nature of our work, we can only offer speculation.First, research has found increased rates of cardiovascular events associated with dialysis prior to transplantation, which could increase recovery times following surgery [19].There could also be a leadtime bias impacting results, as those who receive preemptive transplants will naturally be earlier in their disease course which could lessen recovery times and LOS [17].Finally, selection bias could play a role, as historically patients who were white and male were more likely to receive preemptive transplant [17], but of note our study did not find significant associations with LOS and race or gender.Ultimately, our study adds further evidence and expands upon prior work by showing that dialysis is not only associated with poor graft outcomes, but also prolonged LOS.This serves to highlight the importance of preemptive transplant when available to avoid the morbidity associated with dialysis.In our dataset, only 965 (26.13%) preemptive transplants occurred, compared to 2728 (73.87%) after the initiation of dialysis.Certain barriers to preemptive transplantation exist within the USA, primarily limited organ availability, but also late presentation when in need of kidney replacement therapy.Measures that increase the number of organs available for transplantation, including transitioning to an opt-out consent system of donation, has the potential to significantly increase the number of transplants per year and reduce waitlist morbidity and mortality [20].
The differences we found between geographic regions in the USA regarding risk of prolonged LOS could be attributed to many different factors.has shown previously that there do not appear to be differences in graft/patient survival following KT between different regions in the USA [21], but our study appears to show that LOS may vary, which is in line with similar geographical differences in LOS following liver transplantation [22].This could be attributed to differences in transplant volume per region, demographic factors in the region, or procedural differences across transplant centers.Specifically, center-level findings have been found in pediatric liver transplant, where it has been shown the LOS following transplantation varies significantly across different centers [23], and in pediatric KT, where it has been shown that listings at low-volume centers are associated with worse outcomes [24].In these other studies the regions/ centers were blinded, but further research should be done to identify if regions with increased LOS identified in our study correlate with increased LOS in the region in general or if our findings are unique to pediatric KT.Discharge policy differences across centers could also be contributing to our findings, as there is no universal definition in place for when a KT patient is cleared for discharge.Some centers have experimented and had success with expediated discharge policies, but this is not the norm and highlights the lack of universal criteria for discharge following KT [25].
Weight, specifically low weight, was one of the strongest predictors of prolonged LOS in our risk score.Also, higher weight was the only factor that was protective compared to the reference group in our model.Other studies have found that weight is a key determining factor in post-transplant outcomes, but many have focused on BMI instead of weight alone.This correlation with obesity-range BMI and prolonged stay has been found in both children [26] and adults [27][28][29][30].Few studies using weight alone have been performed, but one study found no difference between LOS in low and high weight children following transplant [31].Our study contradicts this finding possibly because the other study only stratified children into two groups, those below and those equal to or above 15 kg, whereas we stratified into more granular groups.Further, the large increase in risk as our weight groups decrease may reflect the increased technical difficulty for transplant surgery in smaller recipients [32].It should also be noted that our analysis did not identify age, BMI, growth, or weight-for-length as significant predictors of prolonged stay.
Our study confirms the findings of others that some potential risk factors are not associated with prolonged LOS following pediatric KT.Specifically, we found no significant effect on LOS related to the use of donation after circulatory death donors [33] or the use of pulsatile perfusion storage compared to static cold storage [34,35].We also found no association in the pediatric population with factors noted to be significant in the adult KT population, including African American race, an increase in panel reactive antibody (PRA) [4], diabetes, cold ischemia time, and donation after circulatory death [3].The only similarities between our study and adult studies were a difference in LOS related to preoperative weight [4,6] and dialysis [2,4,6].
No measure is currently available to predict prolonged LOS following pediatric KT.The closest measure would be the KDPI score, which we have shown in this study is a poor predictor of prolonged LOS, especially compared to the risk score we were able to develop.Our novel risk score could help identify pediatric candidates who are at increased risk of prolonged LOS, potentially targeting them for earlier interventions to reduce hospital costs via center-based quality improvement initiatives and improve outcomes.Beyond this, the risk score could be used to better inform families to plan for predicted LOS.
Our study is not without limitations.Namely, our sample size was limited to less than 4000 due to the relatively small number of pediatric kidney transplants that occur each year.In order to simplify our findings, living donors and those with history of prior transplant were excluded from analysis, but studies in adults have found that these factors are associated with different LOS [2,4].Future work should study these factors and their association with LOS in the pediatric population.Research should also be dedicated to understanding why there are such stark differences in the LOS between different geographic regions in the USA, how these may be decreased, and whether they are related to procedural, population, or other differences between specific high performing and low performing transplant centers.Finally, research into the implementation of enhanced recovery after surgery (ERAS) methods should be explored following pediatric KT, as various studies have shown its effectiveness in reducing LOS in the adult population [36][37][38][39][40].

Conclusion
Given that kidney transplantation is the modality of choice for long-term kidney replacement therapy in children with kidney failure, it is important for both physicians and families to be able to predict the LOS following KT.Using the UNOS database of all pediatric kidney transplants done between January 2014 and July 2022, we identified four factors associated with prolonged LOS: weight prior to transplant, primary diagnosis of FSGS, geographic region within the USA, and dialysis prior to transplant.We then calculated a risk scoring system that could be used to predict if individual patients may be at risk of LOS greater than 14 days (ROC = 0.7221).By using our risk score, physicians and transplant centers can better inform families' expectations.Since our score implements only recipient risk factors, it may be possible to identify those at risk of prolonged stay for potential interventions before a donor is ever identified.Lastly, our score is more accurate than KDPI at predicting prolonged LOS in the pediatric population.Future work should expand upon this study by examining geographic variability in LOS and by including more variables and living donor recipients for which data is available in the UNOS database.

Table 1
Demographic characteristics of the study cohort recipients and donors (standard deviation) AA, African American; COD, cause of death; CVA, cerebral vascular accident; OPTN, organ procurement and transplantation network stepwise elimination, only focal segmental glomerulosclerosis (FSGS) primary diagnosis, dialysis prior to transplant, geographic region within the USA, and weight prior to transplant were found to be significant predictors of prolonged hospital stay.Results from the first multivariate model and the final risk score model are included in Tables

Table 2
Univariate analysis of factors predicting LOS greater than 14 days following pediatric KT.All factors are recipient factors unless otherwise noted.All metrics unless otherwise noted are from time of transplant DF, degrees of freedom; BMI, body mass index; CMV, cytomegalovirus; COD, cause of death; FSGS, focal segmental glomerulosclerosis; EBV, Epstein-Barr virus; cPRA, calculated panel reactive antibody; HLA, human leukocyte antigen; OPTN, organ procurement and transplantation network [95% confidence interval].* = p-value < 0.05

Table 3
Initial multivariate analysis of factors predicting LOS greater than 14 day following pediatric KT.All factors are recipient factors unless otherwise noted.Recipient weight is weight on transplant candidate registration form.All other metrics are from time of transplant DF, degrees of freedom; CMV, cytomegalovirus; FSGS, focal segmental glomerulosclerosis; EBV, Epstein-Barr virus; OPTN, organ procurement and transplantation network [95% confidence interval].* = p-value < 0.05 FSGS, focal segmental glomerulosclerosis; OPTN, organ procurement and transplantation network [95% confidence interval].* = p-value < 0.05

Table 5
Mean risk scores and LOS for each risk score group.LOS is in days (standard deviation)