MODELING CYSTIC FIBROSIS PATIENT PROGNOSIS: NOMOGRAMS TO PREDICT LUNG TRANSPLANTATION AND SURVIVAL PRIOR TO HIGHLY EFFECTIVE MODULATOR THERAPY

Background. The duration of time a person with cystic fibrosis (pwCF) spends on the lung transplant waitlist is dependent on waitlist and post-transplant survival probabilities and can extend up to 2 years. Understanding the characteristics involved with lung transplant and survival prognoses may help guide decision making by the patient, the referring CF Center and the transplant team. Methods. This study seeks to identify clinical predictors of lung transplant and survival of individuals with CF using 29,847 subjects from 2003-2014 entered in the Cystic Fibrosis Foundation Patient Registry (CFFPR). Results. Predictors significant (p ≤ 0.05) in the final logistic regression model predicting probability of lung transplant/death were: FEV 1 (% predicted), BMI, age of diagnosis, age, number of pulmonary exacerbations, race, sex, CF-related diabetes (CFRD), corticosteroid use, infections with B. cepacia , P. aeruginosa, S. aureus , MRSA, pancreatic enzyme use, insurance status, and consecutive ibuprofen use for at least 4 years. The final Cox regression model predicting time to lung transplant identified these predictors as significant FEV 1 (% predicted), BMI, age of diagnosis, age, number of pulmonary exacerbations, race, sex, CF-related diabetes (CFRD), corticosteroid use, infections with B. cepacia , P. aeruginosa, S. aureus , MRSA, pancreatic enzyme use, and consecutive ibuprofen use for at least 4 years. The concordance indices were 0.89 and 0.92, respectively. Conclusions. The models are translated into nomograms to simplify investigation of how various characteristics relate to lung transplant and survival prognosis individuals with CF not receiving highly effective CFTR modulator therapy.


INTRODUCTION
The lung function complications due to cystic fibrosis (CF) lead to death in 63% of persons with CF, and as a result, persons with CF (pwCF) are often listed for a lung transplant [1].Since late 2019 CF transmembrane-conductance regulator (CFTR) modulation is now standard of care for up to 90% of pwCF in the U.S.These small molecules directly modulate the activity and trafficking of the defective CFTR protein [2].Of particular interest is "highly effective CFTR modulator therapy" (HEMT) which has resulted in exceptional improvement in mean lung function, body mass index, quality of life and rate of pulmonary exacerbations.HEMT commonly refers to use of ivacaftor in patients with CFTR gating mutations [4,5], and the triple combination elexacaftor, tezacaftor and ivacaftor (E/T/I) in pwCF who are either homozygous [2] or heterozygous for F508del mutation [3].
Prior to introduction of HEMT the waiting period on the lung transplant list could last 2 or more years and 23% of CF patients may die while on the list, understanding when to list them is a priority [6,7] The Lung Allocation Score, which is used by the Organ Procurement and Transplantation Network to assure equitable organ allocation, takes survival probability while on the list and after transplantation into account, but is not tailored specifically to CF [8].Following approval of E/T/I within the U.S. the 2020 CFFPR reported a notable reduction in the number of lung transplants performed [9].
Identifying the clinical characteristics involved in predicting when an individual will require a lung transplant may lead to a more personalized approach to transplant referral and listing.Additionally, these need to be communicated effectively from clinicians at CF Care Centers to pwCF and lung transplant centers [10].The CFF consensus guidelines recommends "routine clinician-led efforts to discuss disease trajectory and treatment options, including lung transplantation" and "communication between the CF and lung transplant care teams at least every 6 months and with major clinical changes" [10].
The purpose of this study is to develop models to predict probability of lung transplant or death and time to lung transplant or death of CF patients in the United States prior to HEMT, and translate these into nomograms.The nomograms are meant to personalize estimations of survival and potentially ease communication between the CF clinician and patient and the lung transplant care team about the characteristics of the CF patient affecting both their probability of lung transplant or death and their probability of lung-transplant-free survival.Nkam et al. (2017) developed a nomogram predicting a patient's 3-year risk of lung transplant or death given the patient's characteristics based on data from the French CF Registry [30].The current study aims to demonstrate the validity of this method using data pre-HEMT, and a step towards a current prediction of differing time periods of risk based on patient-specific characteristics to provide a more detailed understanding of a specific patient's prognosis.

Patients
This study was approved by the University of Wyoming Institutional Review Board.A subset of data from the CFF Patient Registry (CFFPR) was used in our study.The CFFPR collects data on CF patients who have been treated in CFF-accredited institutions and consented to be included in the CFFPR [41].The CFFPR is estimated to include approximately 81-84% of persons with CF in the United States [41].The data included 29,847 patients ages 6 to 40 years from from Jan. 1, 2003 to Dec. 31.2014.This was a retrospective study of deidentified data and the authors did not have access to the patient identification.These age limits were included in this study, because pulmonary function tests cannot be measured reliably until age 6 and estimates of survival after age 40 are subject to survivor bias and severe left truncation [42].The interval chosen pre-dates significant usage of HEMT in the US CF population.

Predictors and Outcome variables
Lung transplant or death, whichever occurred first, was the event of interest for both the logistic and Cox regression models, described in more detail below.For the Cox model, time to event was set up as the minimum of the age of lung transplant or death, with age in the most recent review year used for censored observations.The data cleaning and summary measures calculations were conducted with SAS 9.4.

Model Development
The data was analyzed using the software R 4.0.3.Multiple logistic regression modeling was used to predict lung transplant or death and predict the probability of lung transplant or death.Cox regression models were used to identify significant predictors of time to lung transplant or death and probability of time to lung transplant or death.Full logistic and Cox regression models were fit including all independent variables of interest.Significance of predictors (p ≤ 0.05), Akaike information criterion (AIC), and concordance index were used to select the best subset of predictors for the final models.The Holm method was used to correct p-values for multiple testing [43].The Cox proportional hazards assumption was checked using Schoenfeld residuals.Final models were used to generate nomograms of the probability of lung transplant or death and time to lung transplant-free survival.

Model Validation
Leave one out cross-validation was used to determine the estimated prediction error of the logistic regression model using R function cv.glm with the number of resamples equal to the sample size of 29,847.Since there is not an R function for leave one out cross-validation with Cox regression modeling, bootstrapped resampling with 1000 repetitions was used to validate the Cox regression model with R function validate.

Patient Characteristics
The subset of the CFFPR eligible for this study was 29,847 CF patients.Descriptive statistics of the study sample are shown on Table 1.Death or lung transplant was observed in 5206 (17.4%) patients with the remaining 24,641 patients (82.6%) censored in 2014 or if missing in 2014, the time of most current review prior to 2014.Significant differences between the death/transplant and the alive groups were observed in the following characteristics: age of diagnosis, BMI, height, weight, age, FEV1pp, FVCpp, and number of pulmonary exacerbations treated with IV antibiotics (NumPulmExacerbation).Significant associations were observed between vital status and the following categorical variables: race, sex, F508del genotype, CFRD, corticosteroid therapy, B. cepacia, P. aeruginosa, S. aureus, MRSA, pancreatic enzymes therapy, insurance status, and high dose ibuprofen use for at least 4 consecutive years.

Multiple Logistic Regression Model
The following characteristics were significant predictors of probability of lung transplant/death in the final logistic regression model (

Cox Multiple Regression Model
The final Cox regression model of time to lung transplant/death identified the following significant predictors as: FEV1pp, BMI, age of diagnosis, age, NumPulmExacerbation, race, sex, CFRD, corticosteroid therapy, B. cepacia, P. aeruginosa, S. aureus, MRSA, pancreatic enzyme use, and consecutive high dose ibuprofen use for at least 4 years (Table 3, Fig. 2).

Model Validation
The validation of the model was measured based concordance index measuring the proportion of the predictions of the models to predict lung transplant/death in patients that experienced lung transplant/death.The multiple logistic regression model had a concordance index of 0.89 and an estimated prediction error of 0.09 from leave one out cross-validation.The Cox multiple regression model performed similarly as well with a concordance index of 0.92 based on the bootstrap resampling validation.

DISCUSSION
This study developed models that were able to accurately predict probability of the lung transplant/death and time to lung transplant/death.The logistic and Cox regression models were internally validated with accuracy of prediction at 89% and 92%, respectively.The logistic regression model predicting probability of lung transplant/death identified FEV1pp, BMI, age of diagnosis, age, NumPulmExacerbation, race, sex, CFRD, corticosteroid therapy, B. cepacia, P. aeruginosa, S. aureus, MRSA, pancreatic enzyme usage, insurance status, and consecutive high dose ibuprofen use for at least 4 years.Similarly the following characteristics were identified by the Cox regression modeling time to lung transplant/death: FEV1pp, BMI, age of diagnosis, age, NumPulmExacerbation, race, sex, CFRD, corticosteroid therapy, B. cepacia, P. aeruginosa, S. aureus, MRSA, pancreatic enzyme use, and consecutive high dose ibuprofen use for at least 4 years.Significant differences between and associations with vital status were observed in all the characteristics identified by the logistic and Cox regression models.
The logistic and Cox regression models were translated into nomograms.These nomograms are intended to ease the communication between CF clinicians and people with CF, and the lung transplant care center.The characteristics included in the nomograms are measured at most encounter visits (see Methods section for more detail), thus, the CF clinician can routinely calculate the CF patient's probability of lung transplant and probability of lung transplantsurvival in 2 years and 5 years.This may enable the CF clinician and the individual with CF to stay on track with a timely lung transplant referral, to avoid complications and potential barriers to listing associated with late referrals [10].This is intended help to address morality prior to listing, which is reported as high as 50% in a recent study from France [44].The additional benefit of the modeling time to lung-transplant free survival with a Cox regression is the ability to predict probability of lung transplant/death in different time-points allowing investigators to choose time-points that are meaningful to a specific person with CF. 2017) also developed a nomogram to predict probability of lung transplant.Their model predicted probability of lung transplant in 3 years based on a smaller sample (n=2096) of the French CF Registry with only 3 years of data (2010-13) [30].Their model included the following categorical predictors: B. cepacia, hospitalization (yes/no), oral corticosteroids, longterm oxygen therapy, and non-invasive ventilation, and categorical versions of FEV1pp (≥ 60, , <30) and BMI (≥ 18.5, [16-18.5],<16).Although the current study's models are more complex in terms are the number of predictors, they can be more specific to each patient given the models use non-categorical versions of the quantitative predictors and can predict probability of lung transplant-free survival at different time-points.

Nkam et al. (
This study was limited by the variables available in the CFFPR.Characteristics suggested to exacerbate lung function and affect survival by other papers, such as infection with Stenotrophomonas maltophilia, 6 minute walking test, hypercarbia, and hypoxemia, were not available in the CFFPR [10,45,46]. Although nomograms allow for easy communication, there are limitations involved with applying them to predicting probability of and timing of lung transplant/death.Nomograms do not accommodate time-varying covariates, which could have been useful with conditions that change prognosis after they are present, e.g., B. cepacia.Given the necessity of nomograms to allow for clear, easy-to-implement-and-interpret, although other models suggested slope of FEV1 and interactions between pulmonary exacerbations and FEV1pp as predictors, adding these into the model would have made the nomogram more complex to interpret [25,38].Finally, the current nomograms were developed using data that generally predated widespread use of highly effective CFTR modulator therapy (HEMT) in the United States CF population.CFTR modulators improve outcomes for a majority of individuals with CF [2,3,5] and may deliver a sustained benefit over time [47].Clinical trials excluded patients with advanced lung disease, thus there is little available data to guide the degree of disease modification in the period included within this study.It is now apparent that most individuals receiving HEMT will experience slower progression towards transplant or death in this cohort, but current CFFsponsored guidelines for lung transplant referral suggest that transplant referral not be delayed based on use of CFTR modulators [2,3,5,47].Going forward, these findings can serve as a comparison for future cohorts who are on HEMT.
We have developed and internally validated nomograms to predict probability of lung transplant/death and probability of lung transplant-free survival in 2 year and 5 years.The nomograms are user-friendly and will facilitate further investigation into need for transplant and survival in people with CF with advanced lung disease.

Figure 1 .
Figure 1.Nomogram for probability of lung transplant or death.For each characteristic, find the number of points by drawing a vertical line to the Points scale.Total the points for all the characteristics and draw a vertical line from the Total Points scale to get the probability of lung transplant or death.

Figure 2 .
Figure 2. Nomogram for probability of lung transplant or death free survival for 2 years.

Table 1 .
Summary of demographic, annualized, and best yearly pulmonary characteristics by event.

Table 2 .
Logistic regression results for probability of lung transplant/death using complete dataset.CFRD categories are indicator variables for CFRD2= Impaired Glucose Tolerance, and CFRD3 = CFRD with or without fasting hyperglycemia with the reference category CFRD1 = Normal Glucose Metabolism.†The reference category for insurance status was federal. *

Table 3 .
Cox regression results for probability of lung transplant/death prevention for 5 years using complete dataset.*CFRD categories are indicator variables for CFRD2= Impaired Glucose Tolerance, and CFRD3 = CFRD with or without fasting hyperglycemia with the reference category CFRD1 = Normal Glucose Metabolism. *