The aim of this study was to develop a simple clinical algorithm for prediction of donor and recipient death using ‘yes'or ‘no’ questions through the process of recursive partitioning for patients undergoing laser therapy for twin to twin transfusion syndrome (TTTS). The intent was to identify a subset of patients with very high specificity to whom clinical decisions would be simplified.
Secondary analysis of data retrospectively collected from laser procedures was performed for TTTS at NAFTNet centers from 2002 to 2009. Preoperative factors associated with donor and recipient death were identified by recursive partitioning regression analysis. Classification And Regression Trees (CARTs) were developed to refine specificity for prediction of death.
There were 466 TTTS patients from eight centers. CARTs were obtained for prediction of donor death. Improved specificity was achieved through recursive partitioning as demonstrated in receiver operator characteristic curves for prediction of death of the donor. There was less than optimal predictive ability for prediction of death in the recipient, as demonstrated by lack of generation of CARTs.
Twin to twin transfusion syndrome (TTTS) is a disease of monochorionic pregnancies produced by a circulatory imbalance of blood and other vasoactive substances from one fetus to another through placental vascular connections. Informed decisions for management may be optimized by a reasonable understanding of the natural evolution of the disease and the expected outcomes for various therapeutic interventions. We recently reported the development of models for prediction of outcomes based on preoperative ultrasound and Doppler factors in a large cohort of patients undergoing laser therapy for TTTS and demonstrated moderate predictive ability for individual fetuses when including 6–10 preoperative factors. The current report is a secondary analysis of data from the same cohort. The objective of this current study was to determine whether the statistical process of recursive partitioning could refine the preoperative prediction of fetal and neonatal death of individual twin fetuses for pregnancies treated by endoscopic laser ablation of placental vessels in TTTS.
The North American Fetal Therapy Network (NAFTNet) is a collaborative initiative that currently comprises 22 fetal therapy centers in USA and Canada. TTTS pregnancies are cared for at NAFTNet centers, and clinical data and outcomes are routinely recorded prospectively as part of clinical care. This study is a secondary retrospective analysis of TTTS pregnancies that underwent endoscopic laser ablation of placental vessels at eight NAFTNet centers from January 2002 to June 2009. Weill Cornell College of Medicine was the data coordinating center. Local institutional review board approval was obtained at each participating center. For each center, consecutive cases were included. Clinical, ultrasound, and Doppler data were used to identify predictive factors for fetal and neonatal death (Table 1). Details of diagnostic criteria, the therapeutic laser technique, and sample size calculations are in the previous report.
Table 1. Ultrasound and Doppler preoperative factors assessed in the donor and recipient
Intermittent or persistent.
Amniotic fluid volume by deepest vertical pocket measurement (cm)
Estimated fetal weight EFW (grams)
Percent difference in EFW between donor and recipient (EFW percent difference = 100 X (Recipient EFW – Donor EFW)/Recipient EFW)
Presence of fetal bladder
Absent end diastolic velocity in the umbilical artery*
Reverse end diastolic velocity in the umbilical artery*
Hydrops fetalis (Presence of two or more body cavity effusions)
Mitral or Tricuspid cardiac valvular regurgitation (More than trace)
Global cardiac dysfunction (Abnormal Tei index, ventricular dyskinesia, abnormal ejection fraction for either ventricle, abnormal cardiac “score” or other measure of cardiac dysfunction)
Endovaginal ultrasound cervical length
Termination of pregnancy and spontaneous fetal deaths were recorded after all procedures. Cases of selective feticide or termination of the entire pregnancy after laser therapy were included in the denominator of fetal deaths because it was felt that these cases represented untoward events after laser treatment. Neonatal death was defined as the death of a liveborn neonate within 30 days of birth. De-identified data were submitted from each center using a centralized web-based site using a ‘point and click’ data entry tool using minimal free text entry. Data were analyzed by one author (DWS). DWS are the initials of the first author. This is common practice to list the initials in this way.
The analysis focused on factors independently associated with fetal and neonatal death for each fetus (donor and recipient). The process used was logistic regression analysis with recursive partitioning.[3, 4] The products obtained were Classification And Regression Trees (CARTs) and were developed using simple ‘yes’ or ‘no’ questions derived from the preoperative variables. CARTs will not be generated if the predictive ability does not allow a ‘yes’ answer (greater than 50% chance). Receiver operator characteristic (ROC) curves were plotted from the predictive specificity and sensitivity for logistic regression equations for the outcome variables. All graphs and statistical analyses were generated using the ‘R’ software package (http://www.r-project.org). Statistical significance was defined as p < 0.05.
There were a total of 466 cases of laser ablation of placental vessels for TTTS. There were laser procedures performed at nine participating centers as follows: Evergreen Hospital Seattle, 168; Children's Hospital of Philadelphia, 106; Texas Children's Fetal Center, 50; University of Toronto, 44; Brown University, 42; Ohio State University, 30; University of North Carolina, 25; and Yale University, 1. All pregnancies were confirmed to be monochorionic by ultrasound criteria or pathologic examination of the placenta after birth. Complete termination of pregnancy occurred after laser therapy in three cases – two for worsening TTTS and one for worsening condition of the recipient. Selective feticide occurred in seven cases as follows: four recipients (two for persistent TTTS, one for worsening fetal condition, and one for fetal bradycardia) and three donors (two for worsening fetal condition and one for intracranial hemorrhage seen on ultrasound). Thirty cases were lost to neonatal follow-up (6.4%). Overall neonatal survival was 596/872 (68.3%). Donor fetal and neonatal survival were 355/466 (76.2%) and 280/436 (64.2%), respectively. Recipient fetal and neonatal survival were 389/466 (83.5%) and 316/436 (72.5%), respectively.
Logistic regression identified models that were a combination of preoperative factors that best predicted fetal and neonatal death for each individual fetus (donor or recipient). These predictive characteristics and ROC curves are seen in our previous publication. These preoperative factors (Table 1) were then used in a separate logistic regression process using recursive partitioning to develop CARTs and ROC curves. Baseline data showing the type and frequency of abnormalities discovered during the preoperative assessment are seen in Table 2.
Table 2. Baseline preoperative variables
Donor (n = 466)
Recipient (n = 466)
GA = gestational age, EFW = estimated fetal weight, AF = amniotic fluid.
EDV = end diastolic velocity, UV = umbilical vein, DV = ductus venosus.
Percent difference in EFW between donor and recipient = EFW percent difference = 100 X (Recipient EFW – Donor EFW)/Recipient EFW.
GA at diagnosis (in weeks) Median (range)(N)
20 (12 – 26) (465)
Percent EFW difference at procedure Median (range) (N)
24.6 (-36.9 – 64.9) (416)
Donor larger than recipient by EFW
EFW at procedure (g) Mean +/- SD (N)
289 +/- 137 (418)
387 +/- 180 (417)
Maximum vertical pocket AF (cm) Mean +/- SD (N)
0.70 +/- 0.88 (464)
11.2 +/- 3.1 (465)
Absent umbilical arterial EDV
Reversed umbilical arterial EDV
Absent a wave in the DV
Reversed a wave in the DV
Tricuspid or mitral regurgitation
Abnormal global cardiac function
CARTs were obtained for prediction of death of the donor and not for recipient. The CART for donor fetal death is seen in Figure 1. The CART for donor ‘any’ death is seen in Figure 2. ‘Any’ death was defined as either fetal death or neonatal death. The ROC curve for donor fetal death is seen in Figure 3. Compared with the predictive model for donor fetal death previously generated, this curve sacrifices sensitivity in order to improve specificity. This can be seen by the upper right part of the ROC curve dropping down and the lower left part of the curve shifting toward the left (closer to the line that marks 100% specificity). The analysis for fetal donor death (corresponding with the CART seen in Figure 1) showed overall sensitivity of 32.5%, specificity 96.4%, positive predictive value 65%, negative predictive value 87.3%, and overall accuracy 85.4%. The analysis for any donor death (corresponding with the CART seen in Figure 2) showed overall sensitivity of 7.6%, specificity 99.4%, positive predictive value 75%, negative predictive value 82.9%, and overall accuracy 82.8%.
According to the CARTs generated, the factors most highly and significantly associated with donor fetal death included percent difference in estimated fetal weight (EFW) between the twins and global cardiac dysfunction of the donor. There was less ability to predict recipient death as demonstrated by the lack of CARTs.
We found that the specificity of prediction of death for the donor was refined or improved with the use of the recursive partitioning process, but we were not able to identify a group of patients where the prediction of subsequent death was nearly perfect.
In cases of TTTS, the option of selective feticide has been used by a number of centers when, by ultrasound and Doppler criteria, the physicians felt that an individual fetus was moribund or close to death.[5-7] Using these same ultrasound and Doppler criteria, we were unable to identify a subset of patients in whom the specificity for prediction of death was close to 100%.
Of the numerous preoperative factors assessed, the percent difference in EFW between the twins and global cardiac dysfunction in the donor were the strongest individual factors independently associated with donor death. One of these factors, the percent difference in EFW between the twins, is similar to a predictive parameter in a previous report. Zikulnig et al. retrospectively studied 121 TTTS cases after laser therapy and found that the intertwin discrepancy in abdominal circumference (AC) was significantly predictive of survival or demise. The difference in AC or EFW between donor and recipient needs to be prospectively evaluated.
The finding that intertwin EFW difference is a strong independent factor for donor fetal death is biologically plausible. As we previously reported, the larger the discordance between the fetuses, the more likely the donor is to die: a given rate of twin-twin transfusion is relatively more harmful to a small donor than to a large, resistant recipient.[2, 9] We believe this increased propensity for donor death is due to a much smaller volume of placenta for the donor compared to the recipient in many cases of TTTS.[10, 11]
The finding of global cardiac dysfunction of the donor as a predictor of death is also consistent with what is known about the pathophysiological process of TTTS. Although not commonly seen in the donor—and seen much more frequently in the recipient—global cardiac dysfunction probably represents a very late finding in a donor fetus attempting to compensate for chronic blood loss and hypovolemia incompatible with sustained vital organ perfusion.
Classical (Quintero) staging is widely used to predict prognosis. Indeed, classical stage is the prospectively evaluated predictor of outcome in TTTS that has shown the most consistent results.[12, 13] It can be argued that stage should be evaluated in the current analysis. However, this study design was chosen to identify the individual “best” predictive factors. Statistically, it is not appropriate to look also at the staging score. If stage were the best way to predict death, the stepwise algorithm should have selected the components of the staging score in the best predictive model, and this did not occur.
Endovaginal ultrasound cervical length did not appear to be significantly associated with donor or recipient neonatal death. Due to the retrospective nature of the study, we cannot eliminate the possibility that many cases with a very short cervix were excluded from laser therapy by individual centers, thus affecting the ability of cervical length to predict neonatal death from severe prematurity. Neonatal death is difficult to predict due to events that may occur after the preoperative assessment or after the delivery, including premature rupture of membranes, preterm birth, and numerous neonatal complications. Although we expected to discover that we could not reliably predict neonatal death, we believe these negative results are important and thus they are included.
The inability to reliably predict recipient death is of some concern. Some centers offer selective feticide in the most severe cases of recipient global cardiac dysfunction when the physicians suspect that fetal demise is impending. Based on our data, we believe this may not represent a decision based on sound predictive ability. We would suggest a reconsideration of this practice, although we would also suggest larger cohort studies to address this issue.
The sensitivity in the recursive partitioning models is poor. Most fetal losses will not be predicted by this type of analysis, but the certainty of fetal demise is high for the subsets of patients identified in this study. This may be important for the surgeon, particularly in technically challenging cases where surgical complications after laser therapy are thought to be increased (i.e., complete anterior placenta with no window for procedural access). In these cases, if prediction of fetal/neonatal demise is very high, selective feticide may be chosen as a therapy with a potentially lower rate of surgical complications.
The process of recursive partitioning requires a large number of patients to have significant power. Although this was the largest reported series of treated TTTS patients at the time of our original publication, it is a modest number for the recursive partitioning process to allow prediction, and this is a limitation of this study. The retrospective nature of the study is another. Possible selection bias due to the availability of selective feticide at some centers is another limitation.[14, 15] Prospective validation is necessary for the predictive factors identified in this secondary analysis and those found in our previous report.
The enhancement of patient autonomy in the setting of invasive prenatal diagnosis and therapy has been demonstrated to be of benefit. Although the predictive ability achieved in this study is not as close to 100% specificity as we hoped, it is still a stride forward. Use of these techniques provides information for patient counseling that will enhance autonomy, providing patients with a greater understanding of the possible outcomes, and a greater ability to choose from the available therapeutic options.
In summary, the process of recursive partitioning improves the specificity and refines the prediction of donor fetal demise in TTTS treated with laser therapy, and has the potential to improve therapeutic choices and refine counseling regarding outcomes. A prospective application of these techniques to a large cohort of patients is necessary.
WHAT'S ALREADY KNOWN ABOUT THIS TOPIC?
Classical TTTS Stage is the most consistent combination of factors that can predict outcome in TTTS.
WHAT DOES THIS STUDY ADD?
This article adds models that can predict death after laser therapy for TTTS using a large cohort that is unique in the fetal therapy literature.