Journal of Pediatric Surgery xxx (2014) xxx–xxx
Contents lists available at ScienceDirect
Journal of Pediatric Surgery journal homepage: www.elsevier.com/locate/jpedsurg
A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF)☆,☆☆,★,★★,☆☆☆,☆☆☆☆ Benjamin Turner a, Roshni Dasgupta b, Mary Elizabeth Brindle a,⁎ a b
Department of Surgery, University of Calgary, AB Canada Department of Surgery University of Cincinnati, Cincinnati OH USA
a r t i c l e
i n f o
Article history: Received 24 August 2014 Accepted 5 September 2014 Available online xxxx Key words: Tracheo-esophageal fistula Esophageal atresia Clinical prediction Population-based research
a b s t r a c t Background/Purpose: Existing prediction models for tracheo-esophageal fistula (TEF) and esophageal atresia (EA) are derived from small single-institution populations treated over long periods. A prediction rule developed in a contemporary, multicenter cohort is important for counseling, tailoring therapy, and benchmarking outcomes. Methods: Data were obtained from the 2003, 2006, and 2009 editions of the HCUP Kids’ Inpatient Database. Subjects included patients with admission age b three days and ICD-9 diagnostic classification of EA or TEF or procedural coding for TEF repair. An internally validated prediction rule for survival to discharge was developed using a stepwise logistic regression selection algorithm. Predictors included were sex, birth weight, gestational age, cardiac anomalies (major and minor), and chromosomal, other gastrointestinal, central nervous system, and renal anomalies. The model was evaluated for discrimination and calibration and compared with that of Spitz. Results: An integer-based prediction model was created, identifying patients at high, intermediate, and low risk of death with very good discrimination (c = 0.723) and calibration. It is particularly effective at identifying the small population at highest risk of death. The model can be summarized as follows with patients first assigned a score for associated abnormalities: chromosomal abnormality = 6 points, major cardiac anomaly = 3 points, renal anomaly = 2 points, and weight less than 1500 g = 9 points. Point score cut-offs were 0–6 points low risk, 7–14 intermediate risk, and 15–20 high risk. Conclusions: This model compares well with existing prediction models and more effectively discriminates the highest risk patients who may require tailored therapy. The Spitz model is also validated. © 2014 Elsevier Inc. All rights reserved.
Abbreviations: ASD, atrial septal defect; AUC, area under the curve; CHD, congenital heart disease; EA, esophageal atresia; GI, gastrointestinal; HCUP, Healthcare Cost and Utilization Project; ICD-9, International Classification of Diseases Version 9; KID, Kid Inpatient Database. ☆ Funding Source: Personal Funds and Alberta Children’s Hospital Foundation Professorship. ☆☆ Financial Disclosures: No authors have financial relationships relevant to this article to disclose. ★ Conflicts of Interest: No authors have conflicts of interest to disclose. ★★ What’s Known on the Subject: Prediction rules have been developed to determine those infants at TEF at highest risk of death. Existing models, developed in single centers over extended periods of time, have not been found to be valid in a contemporary population. ☆☆☆ What this Study Adds: This study validates the Spitz rule. It also presents a validated clinical prediction rule derived from a logistic regression selection algorithm developed in a population-based cohort. This model identifies a population of infants with TEF at high risk of mortality. ☆☆☆☆ Contributor’s Statements: Ben Turner: Dr. Turner aided in study interpretation. He crafted the initial manuscript and aided in its revision. He approved the final manuscript as submitted.Mary Brindle: Dr. Brindle conceptualized and designed the study, performed the analyses, and edited the manuscript. She approved the final manuscript as submitted.Roshni Dasgupta: Dr. Dasgupta provided insight into study interpretation and application. She critically reviewed and revised the manuscript and approved the final manuscript as submitted. ⁎ Corresponding author at: 2888 Shaganappi Trail NW, Calgary, AB T3B6A8 Canada. Tel.: +1 403 955 2848. E-mail address: [email protected] (M.E. Brindle).
The pathological pattern comprising esophageal atresia and tracheoesophageal fistula (TEF/EA) is a relatively common congenital abnormality, with a prevalence estimated at 1 in 3000–4500 live births [1]. Despite a dramatic increase in the survival rate since Leven and Ladd described the first successful repairs in 1939 [2,3], it remains a cause of significant mortality and morbidity, with a death rate estimated at 12% [4]. Several authors have proposed classification systems for survival prediction. Such models are important for advising parents, tracking improvements in prognosis, objectively comparing outcomes between centers and tailoring therapy. For example, lower risk infants might be selected for a minimally invasive approach, while higher risk infants might be chosen for a delayed or staged repair [5]. 1. Background The first widely accepted prognostic model was proposed by Waterston in 1962 [6]. Patients were divided into three groups according to birth weight, congenital anomalies and presence and severity of pneumonia. Survival ranged from 95% in group A (over 5.5 lb and well) to 6% in group C (birth weight under 4 lb, severe pneumonia or severe congenital anomaly). Poenaru proposed the simpler ‘Montreal’ classification in 1993, based only on ventilator dependence [7]. In recent
http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013 0022-3468/© 2014 Elsevier Inc. All rights reserved.
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
2
B. Turner et al. / Journal of Pediatric Surgery xxx (2014) xxx–xxx
years, the continued relevance of existing classification systems has been called into question. For example, both Spitz [4] and Konkin [8] have found that survival in all three Waterston classes is now so high (95%–100% in groups A and B) as to render the model nearly useless. In 1994, in an attempt to respond to modern survival trends, Spitz proposed a new three-category model based on birth weight and cardiac malformations [4]. This has since been tested at a number of centers, with variable results [8–12]. These prediction models above share a number of limitations. First, they are based on small populations treated at single centers. Treatment algorithms vary considerably between centers, making generalization problematic. Second, both the series used to derive the existing models and those confirming their validity extend over periods exceeding a decade. The Spitz model is based on a data set now going back thirty-one years and may not reflect the changes in contemporary management and survival. This is illustrated by the fact that the Spitz classification had been a better predictor of mortality for the years 1980–1992 than 1993–2005 [13]. For a disease state whose management continues to undergo such rapid change, a truly contemporary predictive model is needed based on a large, generalizable population. This was the goal of our study. 2. Methods The study was approved by the University of Calgary Research Ethics Board. The study population was obtained from the 2003, 2006 and 2009 editions of the Healthcare Cost and Utilization project (HCUP) KID inpatient database, the largest all-payer database for pediatric inpatients in the United States. It is derived from 44 individual state inpatient databases including pediatric discharges and deaths. An 80% stratified sample is attained from all complicated births and other pediatric discharges. As well, a 10% stratified sample of uncomplicated births is included. The discharges are weighted to provide a nationally representative population. Additional information on weighting and sampling is provided by HCUP (http://www.hcup-us.ahrq.gov/db/nation/kid/ kidrelatedreports.jsp). Subjects included for analysis and model-building were those of age less than 3 days at admission and an ICD-9 diagnostic classification of 750.3 (esophageal atresia or tracheoesophageal fistula) as well as those that had an ICD-9 procedural coding for TEF repair 31.73 (closure of fistula of trachea). The following potential predictors were included as variables within the model: Birth-weight, determined by ICD-9 coding or direct entry of birth-weight, stratified into two categories, either N/= 1500 g or b 1500 g, sex, gestational age, minor congenital heart disease which included patent ductus arteriosus (PDA) or atrial septal defect (ASD), major congenital heart disease (cardiac malformations other than PDA or ASD), congenital central nervous system (CNS) malformations, renal malformations, other gastro-intestinal malformations and chromosomal abnormalities. These were identified as follows: Chromosomal anomalies were defined by patient ICD-9 codes beginning with 758
(including all those between 758.0 and 758.9) while CNS anomalies were defined as those with ICD-9 codes beginning with 740, 741 and 742. Renal anomalies were defined as those with ICD-9 codes beginning with 753; GI malformations were defined as those with ICD-9 codes beginning with 751. Major congenital heart defects were identified by the ICD-9 codes starting with 745, 746 and 747 with the exception of the codes for PDA and ostium secundum type ASD which were classified as minor heart defects. ICD-9 coding does not allow the classification of VSDs by severity, so all VSDs were designated as major CHD. The prediction model was generated using the primary outcome of death prior to discharge. Two populations were created by random division of the total study population. The crude distributions of baseline predictors for mortality were examined through univariate analysis of the derivation population. A conservative model was created using a stepwise logistic regression on the derivation population with a p value of 0.01 for inclusion and exclusion. Various combinations of birth-weight divisions were examined for their contribution to the model in the derivation set. Birthweight was looked at as a binary factor (greater or less than 1500 g) as well as an ordinal variable (less than 1000 g, 1000–1500 g and greater than 1500). Once the derivation model was refined, an integer-based model was created based on the OR for mortality for each of the separate variables within the model (rounding to the nearest whole integer). This model was then run on the derivation set and the discrimination for mortality was examined through comparison of mortality rates at differing total point scores. A further modification of the model was created to allow simplification into a high, intermediate and low risk model based on a total points score. The integer-based model and the simplified three-level model were refined for discrimination, calibration and goodness of fit and validated on the second population. As well, a validation of the Spitz classification model was performed. The performance of the model generated from the KID inpatient data was then compared to the Spitz classification. Statistical analyses were performed using SAS 9.3 (Cary, North Carolina).
3. Results There were 2,821,117 patients less than 3 days of age. Overall mortality of these neonates was 0.99%. 2432 patients had a tracheoesophageal fistula. The derivation set included 1212 patients. With weighting, this reflects 1868.3 patients. There were minimal missing data apparent with no missing outcome variables and 7 cases of missing subject sex which account for less than 1% of the dataset. Significant crude relationships were demonstrated between baseline predictors and mortality which are reflected in the results of the univariate analyses (Table 1). Within the stepwise selection regression model, CNS, renal, and chromosomal abnormalities, weight b1500 g and major congenital heart abnormality were identified as significant predictors of death (Table 2). Male sex and minor CHD were not predictors.
Table 1 Baseline characteristics of infants with esophageal atresia and tracheoesophageal fistula and univariate analysis of relation between baseline characteristics and death.
Male Sex (%) Birth weight b1500 g Associated CHD (%) Major CHD Minor CHD CNS abnormalities Renal abnormalities Other gastrointestinal abnormalities Chromosomal abnormalities
All Infants with TEF (n = 2832)
Number Missing
Survivors (n = 2579, 91.1%)
Non-survivors (n = 253, 8.9%)
P value
1485 (52.6%)
7
1355 (52.6%)
130 (52.4%)
0.95
296 (10.4%)
0
195 (7.6%)
101 (39.8%)
b0.0001
755 (27.0%) 874 (30.9%) 129 (4.6%) 324 (11.4%) 372 (13.2%) 207 (7.3%)
0 0 0 0 0 0
613 (23.8%) 785 (30.4%) 102 (4.0%) 271 (10.5%) 319 (12.4%) 138 (5.3%)
142 (56.3%) 90 (35.5%) 27 (10.8%) 53 (20.9%) 53 (21.0%) 70 (27.6%)
b0.0001 0.10 b0.0001 b0.0001 0.0001 b0.0001
TEF tracheoesophageal fistula (includes esophageal atresia), CHD congenital heart disease, CNS central nervous system.
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
B. Turner et al. / Journal of Pediatric Surgery xxx (2014) xxx–xxx Table 2 OR for mortality of predictors in the derivation model (stepwise selection with p = 0.01 for inclusion and exclusion). Risk Factor
Adjusted OR (95%CI)
Weight Less than 1500 g Chromosomal Abnormality Major cardiac anomaly Renal anomaly
9.05 (6.21,13.20) 5.80 (3.71,9.06) 2.68 (1.90,3.79) 1.89 (1.20,2.97)
OR odds ratio, CI confidence interval.
The stepwise logistic regression model yielded excellent discrimination, with an AUC of 0.818. The model created in the derivation set was run in the validation set of 1220 patients (reflecting 1904 patients after weighting). The model still had excellent performance with an AUC of 0.816 and was well calibrated by the Hosmer and Lemeshow Goodness of Fit Statistic (Table 3). The ORs from the derivation set were used to create an integer-based risk score for mortality. ORs were rounded to the nearest integer to give points for the presence of each risk factor, as follows: chromosomal abnormality = 6 points, major cardiac anomaly = 3 points, renal anomaly = 2 points, weight less than 1500 g = 9 points. Using only the single variable risk score, there remained excellent discrimination with an area under the curve of 0.826 in the derivation set and 0.811 in the validation set. We next divided the risk score into low (0–6 points), medium (7–14) and high (15–20) risk categories based on natural divisions of risks observed in the derivation group. The three-category predictive model maintained good discrimination, with an AUC of 0.723 (Validation population). We also assessed the Spitz model’s performance in our validation population. The Spitz model displayed good discrimination, with an AUC of 0.758 (Table 3). The specific ability to discriminate the intermediate and high risk patients was examined both for our model and for the Spitz model (Table 3). Our model was able to identify a higher proportion of patients within clinically relevant higher risk categories. 4. Discussion We have derived and validated a clinical prediction model that effectively identifies those infants with TEF/EA at highest risk of mortality. Premature and small infants are at lower risk of death than in the past and many congenital cardiac diseases are no longer associated with dismal survival [14]. Early surgical repair of TEF is standard for infants judged to be at low risk of mortality and minimally invasive surgery is increasingly being applied in the surgical treatment of low risk
3
patients. Delayed and staged surgeries are reserved for high risk patients [5]. The performance of prior prediction models in detecting those infants with TEF at highest risk of mortality has declined in contemporary populations [13]; yet there still exists a small but defined population of infants who remain at high risk of death. Our model identifies this high risk population. Our clinical prediction model identifies 15% of the population of infants with TEF who have a risk of mortality greater than 25% (Table 4). Further study may demonstrate that this infant population could benefit from a delayed repair or may be more suitable to an open approach for TEF repair. There is already evidence that low birth weight infants benefit from a delayed repair [5], but further study is necessary to determine whether using this prediction rule to identify infants who would benefit from delayed repair results in improved overall survival. We know of no comparison between open and minimally invasive repairs in infants of differing risk groups. We believe that there is an intuitive argument in favor of open repair for the higher risk groups in our study, since this involves shorter operating times, does not require pneumothorax and minimizes the technical difficulty involved in suturing structures that may be smaller than in a low risk infant. This model also identifies a very small proportion of infants who have a very high estimated mortality rate for whom an appropriate multidisciplinary involvement and counseling could be warranted. The points-based predictive model also exhibited excellent discrimination; performing better than previous models in terms of discrimination and calibration. When the model is simplified into three risk strata, discrimination measures are similar to the Spitz model but, importantly, this model performs better at identifying the group of infants at highest risk of mortality. This study also represents the first validation of the Spitz model using contemporary national data. The Spitz model, despite its drawbacks, performs well when compared to a regression-generated clinical prediction rule. The performance of the regression-generated model, a small population at highest risk of mortality can be separated from the majority of patients who are at much lower risk of mortality. The Spitz model identifies a relatively large proportion of patients (35%) at moderate risk of mortality (greater than 16.5%) (Table 5). The current model identifies a smaller subset of the population (15%) at higher risk (greater than 25% mortality) with a very small proportion of extremely high risk patients with a mortality of 69% (Table 4). The inclusion of additional congenital anomalies and chromosomal anomalies in a prediction model is of considerable value. Previous studies have demonstrated that although the presence of cardiac disease in TEF was associated with a higher mortality, only 10% of those deaths
Table 3 Model characteristics in derivation and validation set for initial model, stratified risk model as well as Spitz Model. Derivation model Validation Model Performance Measures of Model Comparing AIC 993.146 Derivation and Validation Groups Goodness of Fit (Hosmer and Lemeshow) C statistic (area under curve) Difference in C statistic
0.818
Stratified Score⁎ Spitz Model⁎ (low, intermediate, high)
786.276 860.771 1.0426 (p = 0.7909) DF = 3 0.816 0.723 0.002 (0.2%)
859.791
0.758
AIC Akaike Information Criteria, DF degrees of freedom. ⁎ Both the stratified risk model and the Spitz model were tested in the validation group.
Table 4 Mortality in differing risk strata for TEF stratified risk model.
Low Risk (0–6) Intermediate Risk (7–14) High Risk (15–20)
Derivation Total Patients (% of total)
Derivation Deaths (% mortality of category)
Validation Total Patients (% of total)
Validation Deaths (% mortality of category)
1022 (84.3%) 166 (13.7%) 24 (2.0%)
50 (4.9%) 58 (34.9%) 16 (66.7%)
1032 (84.7%) 171 (14.0%) 16 (1.3%)
40 (3.9%) 39 (22.8%) 11 (68.8%)
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
4
B. Turner et al. / Journal of Pediatric Surgery xxx (2014) xxx–xxx
Table 5 Mortality in differing risk strata for Spitz model.
Low Risk Intermediate Risk High Risk
Patients (% of total)
Deaths (% mortality of category)
795 (65.2%) 383 (31.4%) 41 (3.4%)
20 (2.5%) 51 (13.3%) 19 (46.3%)
were related to the heart disease itself while the other deaths were a result of other anomalies, chromosomal abnormalities or respiratory issues [15]. Our findings support the notion that the presence of a significant cardiac lesion in combination with TEF may not predict a dire outcome in the absence of other risk factors as previously thought. Ensuring that pediatric surgical patients have the appropriate environment for care is an important quality initiative being championed by governing surgical associations. The model can not only be used to benchmark quality outcomes for individual institutions; but can identify those high risk patients that may require a more specialized neonatal care environment that would optimize their overall outcome. There are several limitations of the model. Detailed information is limited as the source of data is a large administrative discharge database. One example is the inability to differentiate between different types of TEF. Current ICD-9 diagnostic coding does not identify anatomic subtypes. The pooling of all types is likely to reflect the mortality risks of the most common phenotype, namely EA with distal TEF, but raises significant difficulty for the future use of the model on the less common phenotypes such as pure esophageal atresia or H-type tracheoesphageal fistula. Using ICD-9 coding to identify cases and comorbidities can limit accuracy as administrative databases may miss cases and misclassify diagnoses. The ICD-9 coding for TEF/EA has not been validated for the KID inpatient database. Finally, ICD-9 coding does not permit documentation of comorbidities in as much detail as might be desired; the most relevant example is that all VSDs had to be counted as major CHD. There is an important limitation to the practical application of the model, namely the late presentation of some chromosomal abnormalities. Some, such as the trisomies 13, 18 and 21, produce characteristic phenotypic features and may be diagnosed early on clinical grounds, and others may be picked up on prenatal karyotyping, but this is not universally the case. Fortunately, chromosomal abnormalities that are missed in the calculation of a risk score may correspond in many cases to abnormalities missed in the KID data capture, tending to cancel the error. In addition, we recognize that mortality is only one of many valid outcome measures. With improvements in survival, our ability to detect those infants at risk for significant morbidity is becoming increasingly important. Patients who survive admission for TEF are at high risk for early and late complications. Existing prognostic models for mortality do not appear to predict morbidity [16]. Since survival is now approximately 90%, it would seem that the time has come for more robust studies of morbidity prediction. Success in this realm might allow welldesigned surveillance regimens for early detection and intervention in complications, and might shed light on the morbidity of different surgical approaches within high and low risk groups. Significant hurdles in the way of such studies include the low prevalence of any individual complication and the lack of an existing large database enumerating morbidity accurately. For example, stricture and leak would not be discernible in the KID database. A future study in this setting might profitably examine the rate of Nissen fundoplication by ICD-9 coding, as a surrogate for reflux. Morbidity is likely influenced to a great degree by
phenotype, which is not included in any mortality prediction model to date, or in any large database of which we are aware. It would be highly desirable to establish such a database, or such a function of an existing database, before undertaking morbidity prediction. 5. Conclusions We have designed a clinical prediction model for death prior to discharge in patients with TEF, which exhibits good discrimination. The Spitz model has also been demonstrated to be both convenient and clinically useful in this validation study. The new clinical prediction model we have developed allows for comparison of outcomes between centers and selection for delayed or staged repairs. In the development of regionalization policies targeting high risk and complex patients, clinical prediction rules offer possible methods of patient selection for transfer to high-volume centers. Future research will investigate the ability of this model to discriminate those patients at highest risk of morbidity and to assess this model’s performance for different anatomic variations of TEF. Acknowledgments All authors contributed significantly to this manuscript. No authors have conflicts of interest or disclosures in relation to this research. We would like to acknowledge all the HCUP data partners that contribute to the KID inpatient database (www.hcup-us.ahrq.gov/ hcupdatapartners.jsp). References [1] Knottenbelt G, Skinner A, Seefelder C. Tracheo-oesophageal fistula (TOF) and oesophageal atresia (OA). Best Pract Res Clin Anaesthesiol 2010;24(3):387–401. [2] Ladd W. The surgical treatment of esophageal atresia and tracheoesophageal fistulas. N Engl J Med 1944;230:625–37. [3] Leven N. Congenital atresia of the esophagus with tracheo-esophageal fistula. J Thorac Cardiovasc Surg 1941;10:648–57. [4] Spitz L, Kiely E, Morecroft J, et al. Oesophageal atresia: at-risk groups for the 1990s. J Pediatr Surg 1994;29(6):723–5. [5] Petrosyan M, Estrada J, Hunter C, et al. Esophageal atresia/tracheoesophageal fistula in very low-birth-weight neonates: improved outcomes with staged repair. J Pediatr Surg 2009;44(12):2278–81. [6] Waterston D, Carter R, Aberdeen E. Oesophageal atresia: tracheo-oesophageal fistula. A study of survival in 218 infants. Lancet 1962;1:819–22. [7] Poenaru D, Laberge JM, Neilson IR, et al. A new prognostic classification for esophageal atresia. Surgery 1993;113(4):426–32. [8] Konkin DE, O'Hali WA, Webber EM, et al. Outcomes in esophageal atresia and tracheoesophageal fistula. J Pediatr Surg 2003;38(12):1726–9. [9] Yagyu M, Gitter H, Richter B, et al. Esophageal atresia in Bremen, Germany—evaluation of preoperative risk classification in esophageal atresia. J Pediatr Surg 2000; 35(4):584–7. [10] Lopez P, Keys C, Pierro A, et al. Oesophageal atresia: improved outcome in high-risk groups? J Pediatr Surg 2006;41(2):331–4. [11] Choudhury SR, Ashcraft KW, Sharp RJ, et al. Survival of patients with esophageal atresia: influence of birth weight, cardiac anomaly, and late respiratory complications. J Pediatr Surg 1999;34(1):70–3 [discussion 4]. [12] Driver C, Shankar K, Jones M, et al. Phenotypic presentation and outcome of esophageal atresia in the era of the Spitz classification. J Pediatr Surg 2001;36(9):1419–21. [13] Okamoto T, Takamizawa S, Arai H, et al. Esophageal atresia: prognostic classification revisited. Surgery 2009;145(6):675–81. [14] Oster ME, Lee KA, Honein MA, et al. Temporal trends in survival among infants with critical congenital heart defects. Pediatrics 2013;131(5):e1502–8. [15] Leonard H, Barrett AM, Scott JE, et al. The influence of congenital heart disease on survival of infants with oesophageal atresia. Arch Dis Child Fetal Neonatal Ed 2001;85(3):F204–6. [16] Alshehri A, Lo A, Baird R. An analysis of early nonmortality outcome prediction in esophageal atresia. J Pediatr Surg 2012;47(5):881–4.
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
Contents lists available at ScienceDirect
Journal of Pediatric Surgery journal homepage: www.elsevier.com/locate/jpedsurg
A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF)☆,☆☆,★,★★,☆☆☆,☆☆☆☆ Benjamin Turner a, Roshni Dasgupta b, Mary Elizabeth Brindle a,⁎ a b
Department of Surgery, University of Calgary, AB Canada Department of Surgery University of Cincinnati, Cincinnati OH USA
a r t i c l e
i n f o
Article history: Received 24 August 2014 Accepted 5 September 2014 Available online xxxx Key words: Tracheo-esophageal fistula Esophageal atresia Clinical prediction Population-based research
a b s t r a c t Background/Purpose: Existing prediction models for tracheo-esophageal fistula (TEF) and esophageal atresia (EA) are derived from small single-institution populations treated over long periods. A prediction rule developed in a contemporary, multicenter cohort is important for counseling, tailoring therapy, and benchmarking outcomes. Methods: Data were obtained from the 2003, 2006, and 2009 editions of the HCUP Kids’ Inpatient Database. Subjects included patients with admission age b three days and ICD-9 diagnostic classification of EA or TEF or procedural coding for TEF repair. An internally validated prediction rule for survival to discharge was developed using a stepwise logistic regression selection algorithm. Predictors included were sex, birth weight, gestational age, cardiac anomalies (major and minor), and chromosomal, other gastrointestinal, central nervous system, and renal anomalies. The model was evaluated for discrimination and calibration and compared with that of Spitz. Results: An integer-based prediction model was created, identifying patients at high, intermediate, and low risk of death with very good discrimination (c = 0.723) and calibration. It is particularly effective at identifying the small population at highest risk of death. The model can be summarized as follows with patients first assigned a score for associated abnormalities: chromosomal abnormality = 6 points, major cardiac anomaly = 3 points, renal anomaly = 2 points, and weight less than 1500 g = 9 points. Point score cut-offs were 0–6 points low risk, 7–14 intermediate risk, and 15–20 high risk. Conclusions: This model compares well with existing prediction models and more effectively discriminates the highest risk patients who may require tailored therapy. The Spitz model is also validated. © 2014 Elsevier Inc. All rights reserved.
Abbreviations: ASD, atrial septal defect; AUC, area under the curve; CHD, congenital heart disease; EA, esophageal atresia; GI, gastrointestinal; HCUP, Healthcare Cost and Utilization Project; ICD-9, International Classification of Diseases Version 9; KID, Kid Inpatient Database. ☆ Funding Source: Personal Funds and Alberta Children’s Hospital Foundation Professorship. ☆☆ Financial Disclosures: No authors have financial relationships relevant to this article to disclose. ★ Conflicts of Interest: No authors have conflicts of interest to disclose. ★★ What’s Known on the Subject: Prediction rules have been developed to determine those infants at TEF at highest risk of death. Existing models, developed in single centers over extended periods of time, have not been found to be valid in a contemporary population. ☆☆☆ What this Study Adds: This study validates the Spitz rule. It also presents a validated clinical prediction rule derived from a logistic regression selection algorithm developed in a population-based cohort. This model identifies a population of infants with TEF at high risk of mortality. ☆☆☆☆ Contributor’s Statements: Ben Turner: Dr. Turner aided in study interpretation. He crafted the initial manuscript and aided in its revision. He approved the final manuscript as submitted.Mary Brindle: Dr. Brindle conceptualized and designed the study, performed the analyses, and edited the manuscript. She approved the final manuscript as submitted.Roshni Dasgupta: Dr. Dasgupta provided insight into study interpretation and application. She critically reviewed and revised the manuscript and approved the final manuscript as submitted. ⁎ Corresponding author at: 2888 Shaganappi Trail NW, Calgary, AB T3B6A8 Canada. Tel.: +1 403 955 2848. E-mail address: [email protected] (M.E. Brindle).
The pathological pattern comprising esophageal atresia and tracheoesophageal fistula (TEF/EA) is a relatively common congenital abnormality, with a prevalence estimated at 1 in 3000–4500 live births [1]. Despite a dramatic increase in the survival rate since Leven and Ladd described the first successful repairs in 1939 [2,3], it remains a cause of significant mortality and morbidity, with a death rate estimated at 12% [4]. Several authors have proposed classification systems for survival prediction. Such models are important for advising parents, tracking improvements in prognosis, objectively comparing outcomes between centers and tailoring therapy. For example, lower risk infants might be selected for a minimally invasive approach, while higher risk infants might be chosen for a delayed or staged repair [5]. 1. Background The first widely accepted prognostic model was proposed by Waterston in 1962 [6]. Patients were divided into three groups according to birth weight, congenital anomalies and presence and severity of pneumonia. Survival ranged from 95% in group A (over 5.5 lb and well) to 6% in group C (birth weight under 4 lb, severe pneumonia or severe congenital anomaly). Poenaru proposed the simpler ‘Montreal’ classification in 1993, based only on ventilator dependence [7]. In recent
http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013 0022-3468/© 2014 Elsevier Inc. All rights reserved.
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
2
B. Turner et al. / Journal of Pediatric Surgery xxx (2014) xxx–xxx
years, the continued relevance of existing classification systems has been called into question. For example, both Spitz [4] and Konkin [8] have found that survival in all three Waterston classes is now so high (95%–100% in groups A and B) as to render the model nearly useless. In 1994, in an attempt to respond to modern survival trends, Spitz proposed a new three-category model based on birth weight and cardiac malformations [4]. This has since been tested at a number of centers, with variable results [8–12]. These prediction models above share a number of limitations. First, they are based on small populations treated at single centers. Treatment algorithms vary considerably between centers, making generalization problematic. Second, both the series used to derive the existing models and those confirming their validity extend over periods exceeding a decade. The Spitz model is based on a data set now going back thirty-one years and may not reflect the changes in contemporary management and survival. This is illustrated by the fact that the Spitz classification had been a better predictor of mortality for the years 1980–1992 than 1993–2005 [13]. For a disease state whose management continues to undergo such rapid change, a truly contemporary predictive model is needed based on a large, generalizable population. This was the goal of our study. 2. Methods The study was approved by the University of Calgary Research Ethics Board. The study population was obtained from the 2003, 2006 and 2009 editions of the Healthcare Cost and Utilization project (HCUP) KID inpatient database, the largest all-payer database for pediatric inpatients in the United States. It is derived from 44 individual state inpatient databases including pediatric discharges and deaths. An 80% stratified sample is attained from all complicated births and other pediatric discharges. As well, a 10% stratified sample of uncomplicated births is included. The discharges are weighted to provide a nationally representative population. Additional information on weighting and sampling is provided by HCUP (http://www.hcup-us.ahrq.gov/db/nation/kid/ kidrelatedreports.jsp). Subjects included for analysis and model-building were those of age less than 3 days at admission and an ICD-9 diagnostic classification of 750.3 (esophageal atresia or tracheoesophageal fistula) as well as those that had an ICD-9 procedural coding for TEF repair 31.73 (closure of fistula of trachea). The following potential predictors were included as variables within the model: Birth-weight, determined by ICD-9 coding or direct entry of birth-weight, stratified into two categories, either N/= 1500 g or b 1500 g, sex, gestational age, minor congenital heart disease which included patent ductus arteriosus (PDA) or atrial septal defect (ASD), major congenital heart disease (cardiac malformations other than PDA or ASD), congenital central nervous system (CNS) malformations, renal malformations, other gastro-intestinal malformations and chromosomal abnormalities. These were identified as follows: Chromosomal anomalies were defined by patient ICD-9 codes beginning with 758
(including all those between 758.0 and 758.9) while CNS anomalies were defined as those with ICD-9 codes beginning with 740, 741 and 742. Renal anomalies were defined as those with ICD-9 codes beginning with 753; GI malformations were defined as those with ICD-9 codes beginning with 751. Major congenital heart defects were identified by the ICD-9 codes starting with 745, 746 and 747 with the exception of the codes for PDA and ostium secundum type ASD which were classified as minor heart defects. ICD-9 coding does not allow the classification of VSDs by severity, so all VSDs were designated as major CHD. The prediction model was generated using the primary outcome of death prior to discharge. Two populations were created by random division of the total study population. The crude distributions of baseline predictors for mortality were examined through univariate analysis of the derivation population. A conservative model was created using a stepwise logistic regression on the derivation population with a p value of 0.01 for inclusion and exclusion. Various combinations of birth-weight divisions were examined for their contribution to the model in the derivation set. Birthweight was looked at as a binary factor (greater or less than 1500 g) as well as an ordinal variable (less than 1000 g, 1000–1500 g and greater than 1500). Once the derivation model was refined, an integer-based model was created based on the OR for mortality for each of the separate variables within the model (rounding to the nearest whole integer). This model was then run on the derivation set and the discrimination for mortality was examined through comparison of mortality rates at differing total point scores. A further modification of the model was created to allow simplification into a high, intermediate and low risk model based on a total points score. The integer-based model and the simplified three-level model were refined for discrimination, calibration and goodness of fit and validated on the second population. As well, a validation of the Spitz classification model was performed. The performance of the model generated from the KID inpatient data was then compared to the Spitz classification. Statistical analyses were performed using SAS 9.3 (Cary, North Carolina).
3. Results There were 2,821,117 patients less than 3 days of age. Overall mortality of these neonates was 0.99%. 2432 patients had a tracheoesophageal fistula. The derivation set included 1212 patients. With weighting, this reflects 1868.3 patients. There were minimal missing data apparent with no missing outcome variables and 7 cases of missing subject sex which account for less than 1% of the dataset. Significant crude relationships were demonstrated between baseline predictors and mortality which are reflected in the results of the univariate analyses (Table 1). Within the stepwise selection regression model, CNS, renal, and chromosomal abnormalities, weight b1500 g and major congenital heart abnormality were identified as significant predictors of death (Table 2). Male sex and minor CHD were not predictors.
Table 1 Baseline characteristics of infants with esophageal atresia and tracheoesophageal fistula and univariate analysis of relation between baseline characteristics and death.
Male Sex (%) Birth weight b1500 g Associated CHD (%) Major CHD Minor CHD CNS abnormalities Renal abnormalities Other gastrointestinal abnormalities Chromosomal abnormalities
All Infants with TEF (n = 2832)
Number Missing
Survivors (n = 2579, 91.1%)
Non-survivors (n = 253, 8.9%)
P value
1485 (52.6%)
7
1355 (52.6%)
130 (52.4%)
0.95
296 (10.4%)
0
195 (7.6%)
101 (39.8%)
b0.0001
755 (27.0%) 874 (30.9%) 129 (4.6%) 324 (11.4%) 372 (13.2%) 207 (7.3%)
0 0 0 0 0 0
613 (23.8%) 785 (30.4%) 102 (4.0%) 271 (10.5%) 319 (12.4%) 138 (5.3%)
142 (56.3%) 90 (35.5%) 27 (10.8%) 53 (20.9%) 53 (21.0%) 70 (27.6%)
b0.0001 0.10 b0.0001 b0.0001 0.0001 b0.0001
TEF tracheoesophageal fistula (includes esophageal atresia), CHD congenital heart disease, CNS central nervous system.
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
B. Turner et al. / Journal of Pediatric Surgery xxx (2014) xxx–xxx Table 2 OR for mortality of predictors in the derivation model (stepwise selection with p = 0.01 for inclusion and exclusion). Risk Factor
Adjusted OR (95%CI)
Weight Less than 1500 g Chromosomal Abnormality Major cardiac anomaly Renal anomaly
9.05 (6.21,13.20) 5.80 (3.71,9.06) 2.68 (1.90,3.79) 1.89 (1.20,2.97)
OR odds ratio, CI confidence interval.
The stepwise logistic regression model yielded excellent discrimination, with an AUC of 0.818. The model created in the derivation set was run in the validation set of 1220 patients (reflecting 1904 patients after weighting). The model still had excellent performance with an AUC of 0.816 and was well calibrated by the Hosmer and Lemeshow Goodness of Fit Statistic (Table 3). The ORs from the derivation set were used to create an integer-based risk score for mortality. ORs were rounded to the nearest integer to give points for the presence of each risk factor, as follows: chromosomal abnormality = 6 points, major cardiac anomaly = 3 points, renal anomaly = 2 points, weight less than 1500 g = 9 points. Using only the single variable risk score, there remained excellent discrimination with an area under the curve of 0.826 in the derivation set and 0.811 in the validation set. We next divided the risk score into low (0–6 points), medium (7–14) and high (15–20) risk categories based on natural divisions of risks observed in the derivation group. The three-category predictive model maintained good discrimination, with an AUC of 0.723 (Validation population). We also assessed the Spitz model’s performance in our validation population. The Spitz model displayed good discrimination, with an AUC of 0.758 (Table 3). The specific ability to discriminate the intermediate and high risk patients was examined both for our model and for the Spitz model (Table 3). Our model was able to identify a higher proportion of patients within clinically relevant higher risk categories. 4. Discussion We have derived and validated a clinical prediction model that effectively identifies those infants with TEF/EA at highest risk of mortality. Premature and small infants are at lower risk of death than in the past and many congenital cardiac diseases are no longer associated with dismal survival [14]. Early surgical repair of TEF is standard for infants judged to be at low risk of mortality and minimally invasive surgery is increasingly being applied in the surgical treatment of low risk
3
patients. Delayed and staged surgeries are reserved for high risk patients [5]. The performance of prior prediction models in detecting those infants with TEF at highest risk of mortality has declined in contemporary populations [13]; yet there still exists a small but defined population of infants who remain at high risk of death. Our model identifies this high risk population. Our clinical prediction model identifies 15% of the population of infants with TEF who have a risk of mortality greater than 25% (Table 4). Further study may demonstrate that this infant population could benefit from a delayed repair or may be more suitable to an open approach for TEF repair. There is already evidence that low birth weight infants benefit from a delayed repair [5], but further study is necessary to determine whether using this prediction rule to identify infants who would benefit from delayed repair results in improved overall survival. We know of no comparison between open and minimally invasive repairs in infants of differing risk groups. We believe that there is an intuitive argument in favor of open repair for the higher risk groups in our study, since this involves shorter operating times, does not require pneumothorax and minimizes the technical difficulty involved in suturing structures that may be smaller than in a low risk infant. This model also identifies a very small proportion of infants who have a very high estimated mortality rate for whom an appropriate multidisciplinary involvement and counseling could be warranted. The points-based predictive model also exhibited excellent discrimination; performing better than previous models in terms of discrimination and calibration. When the model is simplified into three risk strata, discrimination measures are similar to the Spitz model but, importantly, this model performs better at identifying the group of infants at highest risk of mortality. This study also represents the first validation of the Spitz model using contemporary national data. The Spitz model, despite its drawbacks, performs well when compared to a regression-generated clinical prediction rule. The performance of the regression-generated model, a small population at highest risk of mortality can be separated from the majority of patients who are at much lower risk of mortality. The Spitz model identifies a relatively large proportion of patients (35%) at moderate risk of mortality (greater than 16.5%) (Table 5). The current model identifies a smaller subset of the population (15%) at higher risk (greater than 25% mortality) with a very small proportion of extremely high risk patients with a mortality of 69% (Table 4). The inclusion of additional congenital anomalies and chromosomal anomalies in a prediction model is of considerable value. Previous studies have demonstrated that although the presence of cardiac disease in TEF was associated with a higher mortality, only 10% of those deaths
Table 3 Model characteristics in derivation and validation set for initial model, stratified risk model as well as Spitz Model. Derivation model Validation Model Performance Measures of Model Comparing AIC 993.146 Derivation and Validation Groups Goodness of Fit (Hosmer and Lemeshow) C statistic (area under curve) Difference in C statistic
0.818
Stratified Score⁎ Spitz Model⁎ (low, intermediate, high)
786.276 860.771 1.0426 (p = 0.7909) DF = 3 0.816 0.723 0.002 (0.2%)
859.791
0.758
AIC Akaike Information Criteria, DF degrees of freedom. ⁎ Both the stratified risk model and the Spitz model were tested in the validation group.
Table 4 Mortality in differing risk strata for TEF stratified risk model.
Low Risk (0–6) Intermediate Risk (7–14) High Risk (15–20)
Derivation Total Patients (% of total)
Derivation Deaths (% mortality of category)
Validation Total Patients (% of total)
Validation Deaths (% mortality of category)
1022 (84.3%) 166 (13.7%) 24 (2.0%)
50 (4.9%) 58 (34.9%) 16 (66.7%)
1032 (84.7%) 171 (14.0%) 16 (1.3%)
40 (3.9%) 39 (22.8%) 11 (68.8%)
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
4
B. Turner et al. / Journal of Pediatric Surgery xxx (2014) xxx–xxx
Table 5 Mortality in differing risk strata for Spitz model.
Low Risk Intermediate Risk High Risk
Patients (% of total)
Deaths (% mortality of category)
795 (65.2%) 383 (31.4%) 41 (3.4%)
20 (2.5%) 51 (13.3%) 19 (46.3%)
were related to the heart disease itself while the other deaths were a result of other anomalies, chromosomal abnormalities or respiratory issues [15]. Our findings support the notion that the presence of a significant cardiac lesion in combination with TEF may not predict a dire outcome in the absence of other risk factors as previously thought. Ensuring that pediatric surgical patients have the appropriate environment for care is an important quality initiative being championed by governing surgical associations. The model can not only be used to benchmark quality outcomes for individual institutions; but can identify those high risk patients that may require a more specialized neonatal care environment that would optimize their overall outcome. There are several limitations of the model. Detailed information is limited as the source of data is a large administrative discharge database. One example is the inability to differentiate between different types of TEF. Current ICD-9 diagnostic coding does not identify anatomic subtypes. The pooling of all types is likely to reflect the mortality risks of the most common phenotype, namely EA with distal TEF, but raises significant difficulty for the future use of the model on the less common phenotypes such as pure esophageal atresia or H-type tracheoesphageal fistula. Using ICD-9 coding to identify cases and comorbidities can limit accuracy as administrative databases may miss cases and misclassify diagnoses. The ICD-9 coding for TEF/EA has not been validated for the KID inpatient database. Finally, ICD-9 coding does not permit documentation of comorbidities in as much detail as might be desired; the most relevant example is that all VSDs had to be counted as major CHD. There is an important limitation to the practical application of the model, namely the late presentation of some chromosomal abnormalities. Some, such as the trisomies 13, 18 and 21, produce characteristic phenotypic features and may be diagnosed early on clinical grounds, and others may be picked up on prenatal karyotyping, but this is not universally the case. Fortunately, chromosomal abnormalities that are missed in the calculation of a risk score may correspond in many cases to abnormalities missed in the KID data capture, tending to cancel the error. In addition, we recognize that mortality is only one of many valid outcome measures. With improvements in survival, our ability to detect those infants at risk for significant morbidity is becoming increasingly important. Patients who survive admission for TEF are at high risk for early and late complications. Existing prognostic models for mortality do not appear to predict morbidity [16]. Since survival is now approximately 90%, it would seem that the time has come for more robust studies of morbidity prediction. Success in this realm might allow welldesigned surveillance regimens for early detection and intervention in complications, and might shed light on the morbidity of different surgical approaches within high and low risk groups. Significant hurdles in the way of such studies include the low prevalence of any individual complication and the lack of an existing large database enumerating morbidity accurately. For example, stricture and leak would not be discernible in the KID database. A future study in this setting might profitably examine the rate of Nissen fundoplication by ICD-9 coding, as a surrogate for reflux. Morbidity is likely influenced to a great degree by
phenotype, which is not included in any mortality prediction model to date, or in any large database of which we are aware. It would be highly desirable to establish such a database, or such a function of an existing database, before undertaking morbidity prediction. 5. Conclusions We have designed a clinical prediction model for death prior to discharge in patients with TEF, which exhibits good discrimination. The Spitz model has also been demonstrated to be both convenient and clinically useful in this validation study. The new clinical prediction model we have developed allows for comparison of outcomes between centers and selection for delayed or staged repairs. In the development of regionalization policies targeting high risk and complex patients, clinical prediction rules offer possible methods of patient selection for transfer to high-volume centers. Future research will investigate the ability of this model to discriminate those patients at highest risk of morbidity and to assess this model’s performance for different anatomic variations of TEF. Acknowledgments All authors contributed significantly to this manuscript. No authors have conflicts of interest or disclosures in relation to this research. We would like to acknowledge all the HCUP data partners that contribute to the KID inpatient database (www.hcup-us.ahrq.gov/ hcupdatapartners.jsp). References [1] Knottenbelt G, Skinner A, Seefelder C. Tracheo-oesophageal fistula (TOF) and oesophageal atresia (OA). Best Pract Res Clin Anaesthesiol 2010;24(3):387–401. [2] Ladd W. The surgical treatment of esophageal atresia and tracheoesophageal fistulas. N Engl J Med 1944;230:625–37. [3] Leven N. Congenital atresia of the esophagus with tracheo-esophageal fistula. J Thorac Cardiovasc Surg 1941;10:648–57. [4] Spitz L, Kiely E, Morecroft J, et al. Oesophageal atresia: at-risk groups for the 1990s. J Pediatr Surg 1994;29(6):723–5. [5] Petrosyan M, Estrada J, Hunter C, et al. Esophageal atresia/tracheoesophageal fistula in very low-birth-weight neonates: improved outcomes with staged repair. J Pediatr Surg 2009;44(12):2278–81. [6] Waterston D, Carter R, Aberdeen E. Oesophageal atresia: tracheo-oesophageal fistula. A study of survival in 218 infants. Lancet 1962;1:819–22. [7] Poenaru D, Laberge JM, Neilson IR, et al. A new prognostic classification for esophageal atresia. Surgery 1993;113(4):426–32. [8] Konkin DE, O'Hali WA, Webber EM, et al. Outcomes in esophageal atresia and tracheoesophageal fistula. J Pediatr Surg 2003;38(12):1726–9. [9] Yagyu M, Gitter H, Richter B, et al. Esophageal atresia in Bremen, Germany—evaluation of preoperative risk classification in esophageal atresia. J Pediatr Surg 2000; 35(4):584–7. [10] Lopez P, Keys C, Pierro A, et al. Oesophageal atresia: improved outcome in high-risk groups? J Pediatr Surg 2006;41(2):331–4. [11] Choudhury SR, Ashcraft KW, Sharp RJ, et al. Survival of patients with esophageal atresia: influence of birth weight, cardiac anomaly, and late respiratory complications. J Pediatr Surg 1999;34(1):70–3 [discussion 4]. [12] Driver C, Shankar K, Jones M, et al. Phenotypic presentation and outcome of esophageal atresia in the era of the Spitz classification. J Pediatr Surg 2001;36(9):1419–21. [13] Okamoto T, Takamizawa S, Arai H, et al. Esophageal atresia: prognostic classification revisited. Surgery 2009;145(6):675–81. [14] Oster ME, Lee KA, Honein MA, et al. Temporal trends in survival among infants with critical congenital heart defects. Pediatrics 2013;131(5):e1502–8. [15] Leonard H, Barrett AM, Scott JE, et al. The influence of congenital heart disease on survival of infants with oesophageal atresia. Arch Dis Child Fetal Neonatal Ed 2001;85(3):F204–6. [16] Alshehri A, Lo A, Baird R. An analysis of early nonmortality outcome prediction in esophageal atresia. J Pediatr Surg 2012;47(5):881–4.
Please cite this article as: Turner B, et al, A contemporary prediction rule for esophageal atresia (EA) and tracheo-esophageal fistula (TEF), J Pediatr Surg (2014), http://dx.doi.org/10.1016/j.jpedsurg.2014.09.013
