Prediction Algorithm for Surgical Intervention in Neonatal Brachial Plexus Palsy

Prediction Algorithm for Surgical Intervention in Neonatal Brachial Plexus Palsy Abstract BACKGROUND Neonatal brachial plexus palsy (NBPP) results in reduced function of the affected arm with profound ramifications on quality of life. Advances in surgical technique have shown improvements in outcomes for appropriately selected patients. Patient selection, however, remains difficult. OBJECTIVE To develop a decision algorithm that could be applied at the individual patient level, early in life, to reliably predict persistent NBPP that would benefit from surgery. METHODS Retrospective review of NBPP patients was undertaken. Maternal and neonatal factors were entered into the C5.0 statistical package in R (The R Foundation). A 60/40 model was employed, whereby 60% of randomized data were used to train the decision tree, while the remaining 40% were used to test the decision tree. The outcome of interest for the decision tree was a severe lesion meeting requirements for surgical candidacy. RESULTS A decision tree prediction algorithm was generated from the entered variables. Variables utilized in the final decision tree included presence of Horner's syndrome, presence of a pseudomeningocele, Narakas grade, clavicle fracture at birth, birth weight >9 lbs, and induction or augmentation of labor. Sensitivity of the decision tree was 0.71, specificity 0.96, positive predictive value 0.94, negative predictive value 0.79, and F1 score 0.81. CONCLUSION We developed a decision tree prediction algorithm that can be applied shortly after birth to determine surgical candidacy of patients with NBPP, the first of its kind utilizing only maternal and neonatal factors. This conservative decision tree can be used to offer early surgical intervention for appropriate candidates. Brachial plexus, Decision tree analysis, Neonatal brachial plexus palsy, Nerve injury ABBREVIATIONS ABBREVIATIONS CT computed tomographic MR magnetic resonance NBPP neonatal brachial plexus palsy Neonatal brachial plexus palsy (NBPP) occurs in approximately 0.5 to 3 per 1000 live births.1-4 The injury occurs before, during, or after labor and parturition as a result of a stretching of the nerves of the brachial plexus. A number of risk factors for the incidence of NBPP have been identified. However, the most important question for practitioners treating patients with NBPP, including physiatrists and nerve surgeons, is whether or not the neurological injury and resultant deficits will be persistent. Answering this question helps the treating practitioner determine whether or not surgical intervention will be potentially helpful in aiding arm function of the patient. Until recently, studies had examined risk factors for incidence, but incidence and persistence are really separate questions. We, along with others, have identified several risk factors for persistent NBPP including cephalic presentation, induction or augmentation of labor, birth weight >9 lbs, and the presence of Horner's syndrome.5-8 We also identified cesarean delivery and Narakas grade I/II injury as reducing the likelihood of NBPP persistence.5 Except in the case of flail arm, the decision to operate is a difficult one, with no consensus guidelines. The decision is made more difficult by the fact that 2 competing interests are at play. On the one hand, practitioners try to allow enough time to prove that the injury will be persistent rather than resolve spontaneously, but on the other hand, earlier intervention may be associated with improved outcomes.9 Thus, ideally there would be a decision algorithm that, rather than relying on serial physical examinations, would incorporate maternal and neonatal factors and could be applied early in life to reliably predict those patients who are likely to have persistent NBPP and would benefit from surgery. While the identification of risk factors for persistence is useful data, it is population-based data that cannot be applied to the individual patient presenting for surgical evaluation. Our objective was to develop a decision algorithm that could be applied at the individual patient level, early in life, in order to reliably predict persistence of NBPP that would benefit from surgery. Validation of this algorithm has the potential to allow earlier intervention with associated improved outcomes, while avoiding unnecessary surgery in those patients unlikely to have a persistent injury severe enough that surgery would be beneficial. FIGURE 1. View largeDownload slide University of Michigan NBPP Treatment Pathway used to determine management of patients with NBPP. FIGURE 1. View largeDownload slide University of Michigan NBPP Treatment Pathway used to determine management of patients with NBPP. METHODS Study Design Data were obtained from the Interdisciplinary Brachial Plexus Program data repository, and included patients evaluated for NBPP between July 2005 and June 2015. The data set analyzed was the same data set utilized in our previous study.5 This retrospective cohort study was approved by the Institutional Review Board. Due to the retrospective nature of the study, consent was not sought for inclusion, and the need for consent was waived by the Institutional Review Board. Outcome of Interest The outcome of interest was persistent NBPP for which surgery was recommended according to the University of Michigan NBPP Treatment Pathway (Figure 1). Of note, this differs significantly from our previous analysis where a strict definition of persistence was used.5 For this study, patients were reclassified as either meeting the criteria to have surgery recommended or not meeting the criteria. Serial physical examination data were used to determine surgical candidacy according to the NBPP Treatment Pathway. Whether surgical intervention actually took place did not factor into the classification. Patients without sufficient physical examination data to implement the treatment pathway and determine surgical candidacy were excluded from the study. Statistical Analysis R statistical software was used to generate a decision tree utilizing the C5.0 algorithm.10,11 A 60/40 model was employed, whereby the data were randomized and 60% were used to generate the decision tree (training data) and the remaining 40% were used to test and validate the decision tree that was generated (test data). The sensitivity [true positives/(true positives + false negatives)], specificity [true negatives/(true negatives + false positives)], positive predictive value [true positives/(true positives + false positives)], and negative predictive value [true negatives/(true negatives + false negatives)] of the decision tree were calculated. The F1 score was also calculated [2 * ((positive predictive value * sensitivity)/(positive predictive value + sensitivity))]. Input variables included presence of Horner's syndrome, induction/augmentation of labor, use of a birth aid (vacuum or forceps), Apgar score <7 at 5 min, birth weight >9 lbs, Narakas grade, cephalic vs breech presentation, cesarean vs vaginal delivery, shoulder dystocia, clavicle fracture, maternal age at birth, and presence of a pseudomeningocele on high-resolution magnetic resonance (MR) myelography or computed tomographic (CT) myelography. Table 1 shows the number of patients that had each specific data point available for analysis. For Narakas grading, the neurological examination was assessed at first evaluation in the brachial plexus clinic and then applied to the originally described Narakas scale. This deviates from the originally described Narakas grading in that the scale was not necessarily applied at 2 to 3 wk postdelivery as described by Narakas and recommended by Birch.12,13 Univariate logistic regression analysis was performed analyzing the ability of the outcome of the decision tree prediction algorithm to predict surgical candidacy according to the University of Michigan NBPP Treatment Pathway. P < .05 was considered significant. TABLE 1. Number of Patients Included in the Study for Which Each Data Point was Available for Analysis   Data point available  Variable  (Total n = 333)  Horner's syndrome  333 (100.0%)  Induction/augmentation of labor  298 (89.5%)  Use of a birth aid (vacuum/forceps)  328 (98.5%)  Apgar score at 5 min  182 (54.7%)  Birth weight  333 (100.0%)  Narakas grade  314 (94.3%)  Delivery presentation (cephalic/breech)  330 (99.1%)  Route of delivery (cesarean section/vaginal)  333 (100.0%)  Shoulder dystocia  308 (92.5%)  Clavicle fracture  331 (99.4%)  Maternal age at birth  333 (100.0%)  Imaging for pseudomeningocele  68 (20.4%)    Data point available  Variable  (Total n = 333)  Horner's syndrome  333 (100.0%)  Induction/augmentation of labor  298 (89.5%)  Use of a birth aid (vacuum/forceps)  328 (98.5%)  Apgar score at 5 min  182 (54.7%)  Birth weight  333 (100.0%)  Narakas grade  314 (94.3%)  Delivery presentation (cephalic/breech)  330 (99.1%)  Route of delivery (cesarean section/vaginal)  333 (100.0%)  Shoulder dystocia  308 (92.5%)  Clavicle fracture  331 (99.4%)  Maternal age at birth  333 (100.0%)  Imaging for pseudomeningocele  68 (20.4%)  View Large C5.0 The C5.0 algorithm works on the principal of maximizing normalized information gain to select attributes that optimally divide the total population. To create the first node, the software takes the first attribute and uses it to split the total population into target outcome-based groups—in this case, NBPP patients for whom surgery was recommended vs those for whom surgery was not recommended due to sufficient spontaneous recovery—and calculates the normalized information gain ratio for the given attribute. It uses an iterative process to calculate the normalized information gain ratio for each possible attribute and selects the attribute with the highest normalized information gain. It then takes the remaining population and repeats this process to create subsequent nodes, selecting the attribute with the highest normalized information gain for each node in the final decision tree. The algorithm uses a process of boosting and winnowing in order to optimize and improve the accuracy of the decision tree. RESULTS A total of 382 patients were evaluated in the Interdisciplinary Brachial Plexus Program clinic during the study period. Sufficient serial physical examination data to evaluate patient surgical candidacy according to the University of Michigan NBPP treatment pathway were available for 333 patients; the remaining 49 patients were excluded from the study. The 49 excluded patients presented after 6 mo of age and were not able to be classified based on the treatment pathway. Overall, 145 (43.5%) patients were classified as surgical candidates, while 188 (56.5%) patients were classified as nonsurgical candidates. Variables were entered into the C5.0 statistical package in R for 60% of the randomized data (training data). Variables included in the final decision tree (Figure 2) included presence of Horner's syndrome, presence of a pseudomeningocele, Narakas grade, clavicle fracture at birth, birth weight >9 lbs, and induction or augmentation of labor. The following are examples of how to utilize the decision tree using 2 test patients. Test patient 1 does not have Horner's syndrome, has not had imaging so pseudomeningocele status is unknown, has a Narakas grade III injury, did not have a clavicle fracture at birth, had a birth weight of 7 lbs, and underwent pharmacologic induction of labor. Test patient 2 does not have Horner's syndrome, has no pseudomeningocele on imaging, has a Narakas grade III injury, had a clavicle fracture at birth, had a birth weight of 10 lbs, and had no induction or augmentation of labor. For test patient 1, we follow the decision tree and with no Horner's syndrome present, we look at pseudomeningocele status; with pseudomeningocele status unknown, we look at Narakas grade; with a Narakas grade III injury, we look at whether a clavicle fracture occurred; no clavicle fracture occurred, so we look at birth weight; the birth weight was 7 lbs, so we look at whether induction or augmentation of labor occurred; since pharmacologic induction occurred, the decision tree predicts surgical candidacy. For test patient 2, there is no Horner's syndrome present, so we look at pseudomeningocele status; there is no pseudomeningocele on imaging, so we look at Narakas grade; with a Narakas grade III injury, we look at whether a clavicle fracture occurred; the patient had a clavicle fracture at birth, so the decision tree predicts surgical candidacy. FIGURE 2. View largeDownload slide Decision tree prediction algorithm generated using the C5.0 software package in R. This decision tree can be applied to patients with NBPP to determine surgical candidacy according to the University of Michigan NBPP Treatment Pathway. Y = Yes, N = No, U = Unknown. FIGURE 2. View largeDownload slide Decision tree prediction algorithm generated using the C5.0 software package in R. This decision tree can be applied to patients with NBPP to determine surgical candidacy according to the University of Michigan NBPP Treatment Pathway. Y = Yes, N = No, U = Unknown. The training data were then sorted by the decision tree that was generated. Figure 3 shows the results of putting the training data through the decision tree and the accuracy of each node. The remaining 40% of the data (test data) were used to test the decision tree (Table 2). The sensitivity of the decision tree was 0.71, specificity 0.96, positive predictive value 0.94, negative predictive value 0.79, and F1 score 0.81. A Receiver Operating Characteristic curve was created for the test data (Figure 4). The area under the curve was 0.78. FIGURE 3. View largeDownload slide Randomized training data were used to generate the decision tree. When the training data were then put back through the decision tree, this figure shows the accuracy of each node (A-G) for prediction of surgical candidacy. “Correct” refers to a correct prediction, while “incorrect” refers to an incorrect prediction. FIGURE 3. View largeDownload slide Randomized training data were used to generate the decision tree. When the training data were then put back through the decision tree, this figure shows the accuracy of each node (A-G) for prediction of surgical candidacy. “Correct” refers to a correct prediction, while “incorrect” refers to an incorrect prediction. FIGURE 4. View largeDownload slide Receiver Operating Characteristic curve associated with the NBPP Decision Tree Prediction Algorithm based on the test data. AUC = area under the curve. FIGURE 4. View largeDownload slide Receiver Operating Characteristic curve associated with the NBPP Decision Tree Prediction Algorithm based on the test data. AUC = area under the curve. TABLE 2. Comparison of Predicted Results of Surgical Candidacy Based on the Decision Tree Generated in This Study vs the Actual Surgical Candidacy of Patients Included in the Test Data     Predicted      Surgery  No surgery  Actual  Surgery  45  18    No surgery  3  68      Predicted      Surgery  No surgery  Actual  Surgery  45  18    No surgery  3  68  View Large Univariate logistic regression analysis was performed, analyzing the patients in the test data set. Patients for whom the decision tree prediction algorithm predicted surgery were 56.7 (95% confidence interval 18.1-253.0; P <.001) times more likely to actually be surgical candidates according to the University of Michigan NBPP Treatment Pathway compared to those patients for whom the decision tree prediction algorithm predicted no surgery. DISCUSSION With advances in surgical techniques, the door has been opened to improved outcomes via surgical intervention for patients with NBPP. This has made the decision-making process significantly more difficult for clinicians evaluating and treating these patients. Whereas until the 1990s NBPP was considered a nonsurgical condition, now the clinician is faced with the dilemma of determining which patients are likely to benefit from surgical intervention and which are likely to resolve spontaneously to the point that the natural recovery process outpaces any possible surgical intervention. These injuries are not infrequent, occurring in up to 3 per 1000 live births.1-4 Some existing data suggest that the incidence is decreasing in the United States, but it continues to occur at a not infrequent rate.4,14 The objective of this study was to develop a decision tree predictive algorithm for early determination of surgical candidacy in patients with NBPP. We generated a prediction tree that, rather than relying on serial physical examination alone, uses combined maternal and neonatal factors to allow application shortly after birth. Currently, there are no consensus guidelines available to determine the appropriateness of surgical intervention. Most clinicians use a combination of imaging, electrodiagnostics, and serial physical examination findings when the patient is between 3 and 6 mo of age to determine whether surgery is indicated. One attempt at earlier dichotomization was successful at predicting severe lesions at 1 mo of age, based on a combination of clinical examination findings and electrodiagnostics. This combination of studies was unsuccessful at earlier time points, however.15 Data have shown that nerve reconstruction improves motor outcome in comparison to the natural history when biceps function has not recovered by 4 mo of age.9 What is not clear is the optimal timing or method of nerve reconstruction.16-19 This is especially true given that the indications for primary nerve surgery may need to expand as we further understand the natural history of NBPP, such as when biceps recovery is adequate but shoulder function recovery is poor.20 NBPP also has significant effects outside the realm of motor function. NBPP has effects on limb preference/use, language development, and emotional/behavioral outcomes, to name a few.21-23 What remains to be seen is whether early intervention has any effect on these functional domains. In order to potentially move surgical intervention to an earlier time point, a strategy must be developed that can dichotomize patients into surgical candidates that will not spontaneously recover and patients likely to spontaneously recover who should be managed nonoperatively, and this strategy must be able to be employed early in life. We believe we have accomplished this by developing our current decision tree prediction algorithm. The decision tree generated here utilizes only factors that can be obtained shortly after birth and is the first to combine a set of maternal and neonatal factors into a decision algorithm. To fully employ the decision tree, the necessary components include the birth record, a physical examination after birth, and a high-resolution MR myelogram or CT myelogram. The final variables included in the decision tree were presence of Horner's syndrome on physical examination, presence of a pseudomeningocele on imaging, Narakas grade of injury, presence of a clavicle fracture at birth, birth weight, and utilization of induction or augmentation of labor. The only included factor that may alter the practice pattern of most clinicians would be routinely obtaining a high-resolution MR myelogram or CT myelogram for all NBPP patients being evaluated. This may argue for obtaining imaging and implementing this decision tree at 3 mo of age for those patients who have not already spontaneously recovered. The decision to operate is a question not of incidence of NBPP but rather of persistence of NBPP. We have previously identified risk factors for persistent NBPP.5 While these population-based data are useful for determining which patients are likely to have persistent NBPP in a general sense, these data are not applicable at the individual patient level. This is the advantage of the decision tree prediction algorithm developed here. This decision tree can be applied at the individual patient level and predicts surgical candidacy for the patient being evaluated. For the first time, a clinician faced with deciding whether a patient will go on to have persistent NBPP and benefit from surgery can now make a data-driven decision by applying this decision tree. The ideal prediction algorithm would have both a high positive predictive value and negative predictive value. However, a conservative prediction algorithm would have a high positive predictive value and the negative predictive value would be less important. The decision tree generated here fits this description with a high positive predictive value (94%) and a moderate negative predictive value (79%). What this means is that when the decision tree predicts surgical candidacy, it is highly accurate and unnecessary surgery would be performed only a small percentage of the time. Having a moderate negative predictive value means only that when the decision tree predicts no surgery, the patient must be followed for 6 mo as is currently being done, since they may still later fall into the surgical category a significant percentage of the time. Thus, having a moderate negative predictive value only means that early nerve surgery cannot be performed for those patients but no unnecessary surgery is performed, and these patients would continue to get the current typical evaluation for surgery. The decision tree developed here could certainly be improved upon, particularly the negative predictive value, but represents a conservative decision tree, which is ideal as a preliminary prediction method to be subsequently improved upon. An aggressive prediction algorithm that leads to myriad unnecessary surgeries is not an appropriate baseline for later improvement. Limitations Several shortcomings of the current decision tree present opportunities for future improvement. First, the decision tree was generated without complete records for every patient; complete records would likely improve the decision tree. One example where lack of complete records is likely evident is the Narakas grade branch in the decision tree. Not all patients with Narakas grade I/II injuries who do not have pseudomeningoceles on imaging will spontaneously recover, but that is what the decision tree would suggest. Narakas grade only quantifies the extent of injury but does not in any way indicate severity of the injury to involved segments of the brachial plexus. For example, a mild, neurapraxic injury to C5-7 would be a Narakas grade II injury likely to recover, but avulsion of C5 and C6 would be a Narakas grade I injury with no chance of spontaneous recovery. In addition to this decision tree branch incorrectly predicting lack of surgical appropriateness at a reasonable frequency (Figure 3), it is also likely reflective of incomplete records. Narakas grade was unknown for a number of patients, and having the complete set of data would likely help reflect reality, which is that Narakas grade I/II patients do not all spontaneously recover. Despite this, Narakas grade I/II patients are more likely to spontaneously recover than Narakas grade III/IV. Thus, this reflects the conservative nature of the decision tree, since it predicts no surgery for the population of patients that is, in fact, more likely to recover. The decision tree also does not incorporate any physical examination findings that occur in a delayed fashion, such as physical examination at 1 or 3 mo of age. Given that ultraearly surgical intervention has not been shown to improve outcomes to this point, it may be that incorporation of physical examination findings, such as simple motor grading, motor grading systems such as the Active Movement Scale, or specific tests such as the cookie test at 3 mo, may improve the positive and negative predictive value of the decision tree while still allowing for early intervention and potential for the associated improved outcomes.24,25 It may also be possible to combine this decision tree with the decision tree proposed by Malessy and colleagues15 that utilized physical examination and electrodiagnostic findings at 1 mo. Future versions of the decision tree will likely be improved by having complete records and by incorporating early physical examination findings and/or electrodiagnostic testing. The final limitation is the use of the University of Michigan NBPP Treatment Pathway as the basis of determining surgical candidacy. This treatment pathway has not been validated against the natural history of NBPP. As a result, the decision tree algorithm created here predicts surgical candidacy according to the treatment pathway, but due to the lack of validation it cannot be concluded that the decision tree algorithm predicts those patients for whom surgery will improve upon the natural history. The 2 main decision points for surgery in the University of Michigan NBPP Treatment Pathway are the presence of a flail limb or a lack of biceps function at 6 mo. While the specific pathway has not been validated against the natural history, a lack of biceps function at 6 mo has been shown to correlate with abnormal upper extremity motor function. Patients with a flail limb and those not recovering biceps function by 6 mo of age have been shown to have better outcomes with surgical management than nonsurgical management.9,26 On this basis, we believe, but cannot firmly conclude, that the decision tree produced here will predict patients for whom surgery will outpace the natural history. To more directly address this, however, either the University of Michigan NBPP Treatment Pathway could be validated against the natural history in the future or future iterations of the decision tree could use a validated predictor such as the Toronto Movement Scale. CONCLUSION We developed a decision tree prediction algorithm that can be applied shortly after birth to determine surgical candidacy of patients with NBPP, the first of its kind utilizing only maternal and neonatal factors. It is a conservative prediction algorithm with high positive predictive value and moderate negative predictive value. This decision tree can be utilized to offer early surgical intervention for appropriate candidates to determine if such intervention improves long-term outcomes. While the conservative nature of the decision tree is advantageous for the first iteration, future iterations of the decision tree will likely have improved positive and negative predictive value with incorporation of complete records and early physical examination findings and/or electrodiagnostic studies. Disclosure The authors have no personal, financial, or institutional interest in any of the drugs, materials, or devices described in this article. REFERENCES 1. Bager B. Perinatally acquired brachial plexus palsy–a persisting challenge. Acta Paediatr . 1997; 86( 11): 1214- 1219. Google Scholar CrossRef Search ADS PubMed  2. Dawodu A, Sankaran-Kutty M, Rajan TV. Risk factors and prognosis for brachial plexus injury and clavicular fracture in neonates: a prospective analysis from the United Arab Emirates. Ann Trop Paediatr . 1997; 17( 3): 195- 200. Google Scholar CrossRef Search ADS PubMed  3. Evans-Jones G, Kay SP, Weindling AM et al.   Congenital brachial palsy: incidence, causes, and outcome in the United Kingdom and Republic of Ireland. Arch Dis Child Fetal Neonatal Ed.  2003; 88( 3): F185- F189. Google Scholar CrossRef Search ADS PubMed  4. Foad SL, Mehlman CT, Ying J. The epidemiology of neonatal brachial plexus palsy in the United States. J Bone Joint Surg Am.  2008; 90( 6): 1258- 1264. Google Scholar CrossRef Search ADS PubMed  5. Wilson TJ, Chang KW, Chauhan SP, Yang LJ. Peripartum and neonatal factors associated with the persistence of neonatal brachial plexus palsy at 1 year: a review of 382 cases. J Neurosurg Pediatr.  2016; 17( 5): 618- 624. Google Scholar CrossRef Search ADS PubMed  6. Mollberg M, Lagerkvist AL, Johansson U, Bager B, Johansson A, Hagberg H. Comparison in obstetric management on infants with transient and persistent obstetric brachial plexus palsy. J Child Neurol . 2008; 23( 12): 1424- 1432. Google Scholar CrossRef Search ADS PubMed  7. Al-Qattan MM, Clarke HM, Curtis CG. The prognostic value of concurrent Horner's syndrome in total obstetric brachial plexus injury. J Hand Surg . 2000; 25( 2): 166- 167. Google Scholar CrossRef Search ADS   8. Gherman RB, Ouzounian JG, Satin AJ, Goodwin TM, Phelan JP. A comparison of shoulder dystocia-associated transient and permanent brachial plexus palsies. Obstet Gynecol . 2003; 102( 3): 544- 548. Google Scholar PubMed  9. Waters PM. Comparison of the natural history, the outcome of microsurgical repair, and the outcome of operative reconstruction in brachial plexus birth palsy. J Bone Joint Surg Am.  1999; 81( 5): 649- 659. Google Scholar CrossRef Search ADS PubMed  10. Babic SH, Kokol P, Podgorelec V, Zorman M, Sprogar M, Stiglic MM. The art of building decision trees. J Med Syst . 2000; 24( 1): 43- 52. Google Scholar CrossRef Search ADS PubMed  11. Scheer JK, Osorio JA, Smith JS et al.   Development of validated computer based pre-operative predictive model for proximal junction failure (PJF) or clinically significant PJK with 86% accuracy based on 510 ASD patients with 2-year follow-up. Spine . 2016; 41( 22): E1328- E1335. Google Scholar CrossRef Search ADS PubMed  12. Narakas AO. Obstetrical brachial plexus injuries. In: Lamb DW, ed. The Paralysed Hand. The Hand and Upper Limb . New York, NY: Churchill Livingstone; 1987: 116- 135. 13. Birch R. Obstetric brachial plexus palsy. J Hand Surg . 2002; 27( 1): 3- 8. Google Scholar CrossRef Search ADS   14. Chauhan SP, Blackwell SB, Ananth CV. Neonatal brachial plexus palsy: incidence, prevalence, and temporal trends. Semin Perinatol . 2014; 38( 4): 210- 218. Google Scholar CrossRef Search ADS PubMed  15. Malessy MJ, Pondaag W, Yang LJ, Hofstede-Buitenhuis SM, le Cessie S, van Dijk JG. Severe obstetric brachial plexus palsies can be identified at one month of age. PloS One . 2011; 6( 10): e26193. Google Scholar CrossRef Search ADS PubMed  16. Ali ZS, Bakar D, Li YR et al.   Utility of delayed surgical repair of neonatal brachial plexus palsy. J Neurosurg Pediatr.  2014; 13( 4): 462- 470. Google Scholar CrossRef Search ADS PubMed  17. Little KJ, Zlotolow DA, Soldado F, Cornwall R, Kozin SH. Early functional recovery of elbow flexion and supination following median and/or ulnar nerve fascicle transfer in upper neonatal brachial plexus palsy. J Bone Joint Surg Am.  2014; 96( 3): 215- 221. Google Scholar CrossRef Search ADS PubMed  18. Smith NC, Rowan P, Benson LJ, Ezaki M, Carter PR. Neonatal brachial plexus palsy. Outcome of absent biceps function at three months of age. J Bone Joint Surg Am . 2004; 86-A( 10): 2163- 2170. Google Scholar CrossRef Search ADS PubMed  19. Seruya M, Shen SH, Fuzzard S, Coombs CJ, McCombe DB, Johnstone BR. Spinal accessory nerve transfer outperforms cervical root grafting for suprascapular nerve reconstruction in neonatal brachial plexus palsy. Plast Reconstr Surg . 2015; 135( 5): 1431- 1438. Google Scholar CrossRef Search ADS PubMed  20. Bade SA, Lin JC, Curtis CG, Clarke HM. Extending the indications for primary nerve surgery in obstetrical brachial plexus palsy. BioMed Res Int . 2014; 2014: 627067. Google Scholar CrossRef Search ADS PubMed  21. Chang KW, Yang LJ, Driver L, Nelson VS. High prevalence of early language delay exists among toddlers with neonatal brachial plexus palsy. Pediatr Neurol . 2014; 51( 3): 384- 389. Google Scholar CrossRef Search ADS PubMed  22. Yang LJ, Anand P, Birch R. Limb preference in children with obstetric brachial plexus palsy. Pediatr Neurol . 2005; 33( 1): 46- 49. Google Scholar CrossRef Search ADS PubMed  23. Bellew M, Kay SP, Webb F, Ward A. Developmental and behavioural outcome in obstetric brachial plexus palsy. J Hand Surg . 2000; 25( 1): 49- 51. Google Scholar CrossRef Search ADS   24. Curtis C, Stephens D, Clarke HM, Andrews D. The active movement scale: an evaluative tool for infants with obstetrical brachial plexus palsy. J Hand Surg . 2002; 27( 3): 470- 478. Google Scholar CrossRef Search ADS   25. Borschel GH, Clarke HM. Obstetrical brachial plexus palsy. Plast Reconstr Surg . 2009; 124( 1 suppl): 144e- 155e. Google Scholar CrossRef Search ADS PubMed  26. Waters PM. Update on management of pediatric brachial plexus palsy. J Pediatr Orthop B.  2005; 14( 4): 233- 244. Google Scholar CrossRef Search ADS PubMed  COMMENT The authors have created a thoughtful decision tree to use maternal and neonatal factors to attempt to predict need for surgery in neonatal brachial plexus injury. They have validated this retrospectively using 333 patients, 145 of whom (43.5%) were indicated for surgery, and 188 (56.5%) who were deemed non-surgical candidates. The decision tree is shown in Figure 2 and uses the following data points: presence of Horner's syndrome, presence of pseudomeningocele, Narakas grade, presence of clavicle fracture at birth, birth weight >9 lbs, and induction vs augmentation of labor. The positive predictive value for surgery is 0.94, with a negative predictive value of 0.79, which indicates a conservative algorithm. The authors correctly propose the algorithm as a guideline, which may assist in the decision to take these children to surgery. The authors point out in the discussion that unlike clinical exam findings such as the Toronto Active Movement Scale, or "cookie test" which are associated with physical exam findings at 3-6 months, the current algorithm is designed to provide predictive value based on information present just following birth. The authors correctly state that, to date, operating very early (less than 3 months) has, to this point, not been shown to improve outcomes, and they go on to indicate that the algorithm could likely be improved by incorporation of these physical exam findings over time, which would likely close the gap and improve the negative predictive value of 79%, which is the single biggest criticism of version #1 of the algorithm. This work represents a solid contribution and another tool in the armamentarium used by neonatal brachial plexus surgeons in their decision tree as they work with infants and their families to come to the right decision regarding the role and timing of surgery, especially at very early timepoints. One minor implication from the algorithm is that patients with a Narakas grade III or IV injury with a clavicle fracture would necessarily be indicated for surgery, which appears not to have been reported previously. Chandan Gopal Reddy Rochester, Minnesota Copyright © 2017 by the Congress of Neurological Surgeons http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Neurosurgery Oxford University Press

Prediction Algorithm for Surgical Intervention in Neonatal Brachial Plexus Palsy

Loading next page...
 
/lp/ou_press/prediction-algorithm-for-surgical-intervention-in-neonatal-brachial-P9xusl7bL6
Publisher
Oxford University Press
Copyright
Copyright © 2017 by the Congress of Neurological Surgeons
ISSN
0148-396X
eISSN
1524-4040
D.O.I.
10.1093/neuros/nyx190
Publisher site
See Article on Publisher Site

Abstract

Abstract BACKGROUND Neonatal brachial plexus palsy (NBPP) results in reduced function of the affected arm with profound ramifications on quality of life. Advances in surgical technique have shown improvements in outcomes for appropriately selected patients. Patient selection, however, remains difficult. OBJECTIVE To develop a decision algorithm that could be applied at the individual patient level, early in life, to reliably predict persistent NBPP that would benefit from surgery. METHODS Retrospective review of NBPP patients was undertaken. Maternal and neonatal factors were entered into the C5.0 statistical package in R (The R Foundation). A 60/40 model was employed, whereby 60% of randomized data were used to train the decision tree, while the remaining 40% were used to test the decision tree. The outcome of interest for the decision tree was a severe lesion meeting requirements for surgical candidacy. RESULTS A decision tree prediction algorithm was generated from the entered variables. Variables utilized in the final decision tree included presence of Horner's syndrome, presence of a pseudomeningocele, Narakas grade, clavicle fracture at birth, birth weight >9 lbs, and induction or augmentation of labor. Sensitivity of the decision tree was 0.71, specificity 0.96, positive predictive value 0.94, negative predictive value 0.79, and F1 score 0.81. CONCLUSION We developed a decision tree prediction algorithm that can be applied shortly after birth to determine surgical candidacy of patients with NBPP, the first of its kind utilizing only maternal and neonatal factors. This conservative decision tree can be used to offer early surgical intervention for appropriate candidates. Brachial plexus, Decision tree analysis, Neonatal brachial plexus palsy, Nerve injury ABBREVIATIONS ABBREVIATIONS CT computed tomographic MR magnetic resonance NBPP neonatal brachial plexus palsy Neonatal brachial plexus palsy (NBPP) occurs in approximately 0.5 to 3 per 1000 live births.1-4 The injury occurs before, during, or after labor and parturition as a result of a stretching of the nerves of the brachial plexus. A number of risk factors for the incidence of NBPP have been identified. However, the most important question for practitioners treating patients with NBPP, including physiatrists and nerve surgeons, is whether or not the neurological injury and resultant deficits will be persistent. Answering this question helps the treating practitioner determine whether or not surgical intervention will be potentially helpful in aiding arm function of the patient. Until recently, studies had examined risk factors for incidence, but incidence and persistence are really separate questions. We, along with others, have identified several risk factors for persistent NBPP including cephalic presentation, induction or augmentation of labor, birth weight >9 lbs, and the presence of Horner's syndrome.5-8 We also identified cesarean delivery and Narakas grade I/II injury as reducing the likelihood of NBPP persistence.5 Except in the case of flail arm, the decision to operate is a difficult one, with no consensus guidelines. The decision is made more difficult by the fact that 2 competing interests are at play. On the one hand, practitioners try to allow enough time to prove that the injury will be persistent rather than resolve spontaneously, but on the other hand, earlier intervention may be associated with improved outcomes.9 Thus, ideally there would be a decision algorithm that, rather than relying on serial physical examinations, would incorporate maternal and neonatal factors and could be applied early in life to reliably predict those patients who are likely to have persistent NBPP and would benefit from surgery. While the identification of risk factors for persistence is useful data, it is population-based data that cannot be applied to the individual patient presenting for surgical evaluation. Our objective was to develop a decision algorithm that could be applied at the individual patient level, early in life, in order to reliably predict persistence of NBPP that would benefit from surgery. Validation of this algorithm has the potential to allow earlier intervention with associated improved outcomes, while avoiding unnecessary surgery in those patients unlikely to have a persistent injury severe enough that surgery would be beneficial. FIGURE 1. View largeDownload slide University of Michigan NBPP Treatment Pathway used to determine management of patients with NBPP. FIGURE 1. View largeDownload slide University of Michigan NBPP Treatment Pathway used to determine management of patients with NBPP. METHODS Study Design Data were obtained from the Interdisciplinary Brachial Plexus Program data repository, and included patients evaluated for NBPP between July 2005 and June 2015. The data set analyzed was the same data set utilized in our previous study.5 This retrospective cohort study was approved by the Institutional Review Board. Due to the retrospective nature of the study, consent was not sought for inclusion, and the need for consent was waived by the Institutional Review Board. Outcome of Interest The outcome of interest was persistent NBPP for which surgery was recommended according to the University of Michigan NBPP Treatment Pathway (Figure 1). Of note, this differs significantly from our previous analysis where a strict definition of persistence was used.5 For this study, patients were reclassified as either meeting the criteria to have surgery recommended or not meeting the criteria. Serial physical examination data were used to determine surgical candidacy according to the NBPP Treatment Pathway. Whether surgical intervention actually took place did not factor into the classification. Patients without sufficient physical examination data to implement the treatment pathway and determine surgical candidacy were excluded from the study. Statistical Analysis R statistical software was used to generate a decision tree utilizing the C5.0 algorithm.10,11 A 60/40 model was employed, whereby the data were randomized and 60% were used to generate the decision tree (training data) and the remaining 40% were used to test and validate the decision tree that was generated (test data). The sensitivity [true positives/(true positives + false negatives)], specificity [true negatives/(true negatives + false positives)], positive predictive value [true positives/(true positives + false positives)], and negative predictive value [true negatives/(true negatives + false negatives)] of the decision tree were calculated. The F1 score was also calculated [2 * ((positive predictive value * sensitivity)/(positive predictive value + sensitivity))]. Input variables included presence of Horner's syndrome, induction/augmentation of labor, use of a birth aid (vacuum or forceps), Apgar score <7 at 5 min, birth weight >9 lbs, Narakas grade, cephalic vs breech presentation, cesarean vs vaginal delivery, shoulder dystocia, clavicle fracture, maternal age at birth, and presence of a pseudomeningocele on high-resolution magnetic resonance (MR) myelography or computed tomographic (CT) myelography. Table 1 shows the number of patients that had each specific data point available for analysis. For Narakas grading, the neurological examination was assessed at first evaluation in the brachial plexus clinic and then applied to the originally described Narakas scale. This deviates from the originally described Narakas grading in that the scale was not necessarily applied at 2 to 3 wk postdelivery as described by Narakas and recommended by Birch.12,13 Univariate logistic regression analysis was performed analyzing the ability of the outcome of the decision tree prediction algorithm to predict surgical candidacy according to the University of Michigan NBPP Treatment Pathway. P < .05 was considered significant. TABLE 1. Number of Patients Included in the Study for Which Each Data Point was Available for Analysis   Data point available  Variable  (Total n = 333)  Horner's syndrome  333 (100.0%)  Induction/augmentation of labor  298 (89.5%)  Use of a birth aid (vacuum/forceps)  328 (98.5%)  Apgar score at 5 min  182 (54.7%)  Birth weight  333 (100.0%)  Narakas grade  314 (94.3%)  Delivery presentation (cephalic/breech)  330 (99.1%)  Route of delivery (cesarean section/vaginal)  333 (100.0%)  Shoulder dystocia  308 (92.5%)  Clavicle fracture  331 (99.4%)  Maternal age at birth  333 (100.0%)  Imaging for pseudomeningocele  68 (20.4%)    Data point available  Variable  (Total n = 333)  Horner's syndrome  333 (100.0%)  Induction/augmentation of labor  298 (89.5%)  Use of a birth aid (vacuum/forceps)  328 (98.5%)  Apgar score at 5 min  182 (54.7%)  Birth weight  333 (100.0%)  Narakas grade  314 (94.3%)  Delivery presentation (cephalic/breech)  330 (99.1%)  Route of delivery (cesarean section/vaginal)  333 (100.0%)  Shoulder dystocia  308 (92.5%)  Clavicle fracture  331 (99.4%)  Maternal age at birth  333 (100.0%)  Imaging for pseudomeningocele  68 (20.4%)  View Large C5.0 The C5.0 algorithm works on the principal of maximizing normalized information gain to select attributes that optimally divide the total population. To create the first node, the software takes the first attribute and uses it to split the total population into target outcome-based groups—in this case, NBPP patients for whom surgery was recommended vs those for whom surgery was not recommended due to sufficient spontaneous recovery—and calculates the normalized information gain ratio for the given attribute. It uses an iterative process to calculate the normalized information gain ratio for each possible attribute and selects the attribute with the highest normalized information gain. It then takes the remaining population and repeats this process to create subsequent nodes, selecting the attribute with the highest normalized information gain for each node in the final decision tree. The algorithm uses a process of boosting and winnowing in order to optimize and improve the accuracy of the decision tree. RESULTS A total of 382 patients were evaluated in the Interdisciplinary Brachial Plexus Program clinic during the study period. Sufficient serial physical examination data to evaluate patient surgical candidacy according to the University of Michigan NBPP treatment pathway were available for 333 patients; the remaining 49 patients were excluded from the study. The 49 excluded patients presented after 6 mo of age and were not able to be classified based on the treatment pathway. Overall, 145 (43.5%) patients were classified as surgical candidates, while 188 (56.5%) patients were classified as nonsurgical candidates. Variables were entered into the C5.0 statistical package in R for 60% of the randomized data (training data). Variables included in the final decision tree (Figure 2) included presence of Horner's syndrome, presence of a pseudomeningocele, Narakas grade, clavicle fracture at birth, birth weight >9 lbs, and induction or augmentation of labor. The following are examples of how to utilize the decision tree using 2 test patients. Test patient 1 does not have Horner's syndrome, has not had imaging so pseudomeningocele status is unknown, has a Narakas grade III injury, did not have a clavicle fracture at birth, had a birth weight of 7 lbs, and underwent pharmacologic induction of labor. Test patient 2 does not have Horner's syndrome, has no pseudomeningocele on imaging, has a Narakas grade III injury, had a clavicle fracture at birth, had a birth weight of 10 lbs, and had no induction or augmentation of labor. For test patient 1, we follow the decision tree and with no Horner's syndrome present, we look at pseudomeningocele status; with pseudomeningocele status unknown, we look at Narakas grade; with a Narakas grade III injury, we look at whether a clavicle fracture occurred; no clavicle fracture occurred, so we look at birth weight; the birth weight was 7 lbs, so we look at whether induction or augmentation of labor occurred; since pharmacologic induction occurred, the decision tree predicts surgical candidacy. For test patient 2, there is no Horner's syndrome present, so we look at pseudomeningocele status; there is no pseudomeningocele on imaging, so we look at Narakas grade; with a Narakas grade III injury, we look at whether a clavicle fracture occurred; the patient had a clavicle fracture at birth, so the decision tree predicts surgical candidacy. FIGURE 2. View largeDownload slide Decision tree prediction algorithm generated using the C5.0 software package in R. This decision tree can be applied to patients with NBPP to determine surgical candidacy according to the University of Michigan NBPP Treatment Pathway. Y = Yes, N = No, U = Unknown. FIGURE 2. View largeDownload slide Decision tree prediction algorithm generated using the C5.0 software package in R. This decision tree can be applied to patients with NBPP to determine surgical candidacy according to the University of Michigan NBPP Treatment Pathway. Y = Yes, N = No, U = Unknown. The training data were then sorted by the decision tree that was generated. Figure 3 shows the results of putting the training data through the decision tree and the accuracy of each node. The remaining 40% of the data (test data) were used to test the decision tree (Table 2). The sensitivity of the decision tree was 0.71, specificity 0.96, positive predictive value 0.94, negative predictive value 0.79, and F1 score 0.81. A Receiver Operating Characteristic curve was created for the test data (Figure 4). The area under the curve was 0.78. FIGURE 3. View largeDownload slide Randomized training data were used to generate the decision tree. When the training data were then put back through the decision tree, this figure shows the accuracy of each node (A-G) for prediction of surgical candidacy. “Correct” refers to a correct prediction, while “incorrect” refers to an incorrect prediction. FIGURE 3. View largeDownload slide Randomized training data were used to generate the decision tree. When the training data were then put back through the decision tree, this figure shows the accuracy of each node (A-G) for prediction of surgical candidacy. “Correct” refers to a correct prediction, while “incorrect” refers to an incorrect prediction. FIGURE 4. View largeDownload slide Receiver Operating Characteristic curve associated with the NBPP Decision Tree Prediction Algorithm based on the test data. AUC = area under the curve. FIGURE 4. View largeDownload slide Receiver Operating Characteristic curve associated with the NBPP Decision Tree Prediction Algorithm based on the test data. AUC = area under the curve. TABLE 2. Comparison of Predicted Results of Surgical Candidacy Based on the Decision Tree Generated in This Study vs the Actual Surgical Candidacy of Patients Included in the Test Data     Predicted      Surgery  No surgery  Actual  Surgery  45  18    No surgery  3  68      Predicted      Surgery  No surgery  Actual  Surgery  45  18    No surgery  3  68  View Large Univariate logistic regression analysis was performed, analyzing the patients in the test data set. Patients for whom the decision tree prediction algorithm predicted surgery were 56.7 (95% confidence interval 18.1-253.0; P <.001) times more likely to actually be surgical candidates according to the University of Michigan NBPP Treatment Pathway compared to those patients for whom the decision tree prediction algorithm predicted no surgery. DISCUSSION With advances in surgical techniques, the door has been opened to improved outcomes via surgical intervention for patients with NBPP. This has made the decision-making process significantly more difficult for clinicians evaluating and treating these patients. Whereas until the 1990s NBPP was considered a nonsurgical condition, now the clinician is faced with the dilemma of determining which patients are likely to benefit from surgical intervention and which are likely to resolve spontaneously to the point that the natural recovery process outpaces any possible surgical intervention. These injuries are not infrequent, occurring in up to 3 per 1000 live births.1-4 Some existing data suggest that the incidence is decreasing in the United States, but it continues to occur at a not infrequent rate.4,14 The objective of this study was to develop a decision tree predictive algorithm for early determination of surgical candidacy in patients with NBPP. We generated a prediction tree that, rather than relying on serial physical examination alone, uses combined maternal and neonatal factors to allow application shortly after birth. Currently, there are no consensus guidelines available to determine the appropriateness of surgical intervention. Most clinicians use a combination of imaging, electrodiagnostics, and serial physical examination findings when the patient is between 3 and 6 mo of age to determine whether surgery is indicated. One attempt at earlier dichotomization was successful at predicting severe lesions at 1 mo of age, based on a combination of clinical examination findings and electrodiagnostics. This combination of studies was unsuccessful at earlier time points, however.15 Data have shown that nerve reconstruction improves motor outcome in comparison to the natural history when biceps function has not recovered by 4 mo of age.9 What is not clear is the optimal timing or method of nerve reconstruction.16-19 This is especially true given that the indications for primary nerve surgery may need to expand as we further understand the natural history of NBPP, such as when biceps recovery is adequate but shoulder function recovery is poor.20 NBPP also has significant effects outside the realm of motor function. NBPP has effects on limb preference/use, language development, and emotional/behavioral outcomes, to name a few.21-23 What remains to be seen is whether early intervention has any effect on these functional domains. In order to potentially move surgical intervention to an earlier time point, a strategy must be developed that can dichotomize patients into surgical candidates that will not spontaneously recover and patients likely to spontaneously recover who should be managed nonoperatively, and this strategy must be able to be employed early in life. We believe we have accomplished this by developing our current decision tree prediction algorithm. The decision tree generated here utilizes only factors that can be obtained shortly after birth and is the first to combine a set of maternal and neonatal factors into a decision algorithm. To fully employ the decision tree, the necessary components include the birth record, a physical examination after birth, and a high-resolution MR myelogram or CT myelogram. The final variables included in the decision tree were presence of Horner's syndrome on physical examination, presence of a pseudomeningocele on imaging, Narakas grade of injury, presence of a clavicle fracture at birth, birth weight, and utilization of induction or augmentation of labor. The only included factor that may alter the practice pattern of most clinicians would be routinely obtaining a high-resolution MR myelogram or CT myelogram for all NBPP patients being evaluated. This may argue for obtaining imaging and implementing this decision tree at 3 mo of age for those patients who have not already spontaneously recovered. The decision to operate is a question not of incidence of NBPP but rather of persistence of NBPP. We have previously identified risk factors for persistent NBPP.5 While these population-based data are useful for determining which patients are likely to have persistent NBPP in a general sense, these data are not applicable at the individual patient level. This is the advantage of the decision tree prediction algorithm developed here. This decision tree can be applied at the individual patient level and predicts surgical candidacy for the patient being evaluated. For the first time, a clinician faced with deciding whether a patient will go on to have persistent NBPP and benefit from surgery can now make a data-driven decision by applying this decision tree. The ideal prediction algorithm would have both a high positive predictive value and negative predictive value. However, a conservative prediction algorithm would have a high positive predictive value and the negative predictive value would be less important. The decision tree generated here fits this description with a high positive predictive value (94%) and a moderate negative predictive value (79%). What this means is that when the decision tree predicts surgical candidacy, it is highly accurate and unnecessary surgery would be performed only a small percentage of the time. Having a moderate negative predictive value means only that when the decision tree predicts no surgery, the patient must be followed for 6 mo as is currently being done, since they may still later fall into the surgical category a significant percentage of the time. Thus, having a moderate negative predictive value only means that early nerve surgery cannot be performed for those patients but no unnecessary surgery is performed, and these patients would continue to get the current typical evaluation for surgery. The decision tree developed here could certainly be improved upon, particularly the negative predictive value, but represents a conservative decision tree, which is ideal as a preliminary prediction method to be subsequently improved upon. An aggressive prediction algorithm that leads to myriad unnecessary surgeries is not an appropriate baseline for later improvement. Limitations Several shortcomings of the current decision tree present opportunities for future improvement. First, the decision tree was generated without complete records for every patient; complete records would likely improve the decision tree. One example where lack of complete records is likely evident is the Narakas grade branch in the decision tree. Not all patients with Narakas grade I/II injuries who do not have pseudomeningoceles on imaging will spontaneously recover, but that is what the decision tree would suggest. Narakas grade only quantifies the extent of injury but does not in any way indicate severity of the injury to involved segments of the brachial plexus. For example, a mild, neurapraxic injury to C5-7 would be a Narakas grade II injury likely to recover, but avulsion of C5 and C6 would be a Narakas grade I injury with no chance of spontaneous recovery. In addition to this decision tree branch incorrectly predicting lack of surgical appropriateness at a reasonable frequency (Figure 3), it is also likely reflective of incomplete records. Narakas grade was unknown for a number of patients, and having the complete set of data would likely help reflect reality, which is that Narakas grade I/II patients do not all spontaneously recover. Despite this, Narakas grade I/II patients are more likely to spontaneously recover than Narakas grade III/IV. Thus, this reflects the conservative nature of the decision tree, since it predicts no surgery for the population of patients that is, in fact, more likely to recover. The decision tree also does not incorporate any physical examination findings that occur in a delayed fashion, such as physical examination at 1 or 3 mo of age. Given that ultraearly surgical intervention has not been shown to improve outcomes to this point, it may be that incorporation of physical examination findings, such as simple motor grading, motor grading systems such as the Active Movement Scale, or specific tests such as the cookie test at 3 mo, may improve the positive and negative predictive value of the decision tree while still allowing for early intervention and potential for the associated improved outcomes.24,25 It may also be possible to combine this decision tree with the decision tree proposed by Malessy and colleagues15 that utilized physical examination and electrodiagnostic findings at 1 mo. Future versions of the decision tree will likely be improved by having complete records and by incorporating early physical examination findings and/or electrodiagnostic testing. The final limitation is the use of the University of Michigan NBPP Treatment Pathway as the basis of determining surgical candidacy. This treatment pathway has not been validated against the natural history of NBPP. As a result, the decision tree algorithm created here predicts surgical candidacy according to the treatment pathway, but due to the lack of validation it cannot be concluded that the decision tree algorithm predicts those patients for whom surgery will improve upon the natural history. The 2 main decision points for surgery in the University of Michigan NBPP Treatment Pathway are the presence of a flail limb or a lack of biceps function at 6 mo. While the specific pathway has not been validated against the natural history, a lack of biceps function at 6 mo has been shown to correlate with abnormal upper extremity motor function. Patients with a flail limb and those not recovering biceps function by 6 mo of age have been shown to have better outcomes with surgical management than nonsurgical management.9,26 On this basis, we believe, but cannot firmly conclude, that the decision tree produced here will predict patients for whom surgery will outpace the natural history. To more directly address this, however, either the University of Michigan NBPP Treatment Pathway could be validated against the natural history in the future or future iterations of the decision tree could use a validated predictor such as the Toronto Movement Scale. CONCLUSION We developed a decision tree prediction algorithm that can be applied shortly after birth to determine surgical candidacy of patients with NBPP, the first of its kind utilizing only maternal and neonatal factors. It is a conservative prediction algorithm with high positive predictive value and moderate negative predictive value. This decision tree can be utilized to offer early surgical intervention for appropriate candidates to determine if such intervention improves long-term outcomes. While the conservative nature of the decision tree is advantageous for the first iteration, future iterations of the decision tree will likely have improved positive and negative predictive value with incorporation of complete records and early physical examination findings and/or electrodiagnostic studies. Disclosure The authors have no personal, financial, or institutional interest in any of the drugs, materials, or devices described in this article. REFERENCES 1. Bager B. Perinatally acquired brachial plexus palsy–a persisting challenge. Acta Paediatr . 1997; 86( 11): 1214- 1219. Google Scholar CrossRef Search ADS PubMed  2. Dawodu A, Sankaran-Kutty M, Rajan TV. Risk factors and prognosis for brachial plexus injury and clavicular fracture in neonates: a prospective analysis from the United Arab Emirates. Ann Trop Paediatr . 1997; 17( 3): 195- 200. Google Scholar CrossRef Search ADS PubMed  3. Evans-Jones G, Kay SP, Weindling AM et al.   Congenital brachial palsy: incidence, causes, and outcome in the United Kingdom and Republic of Ireland. Arch Dis Child Fetal Neonatal Ed.  2003; 88( 3): F185- F189. Google Scholar CrossRef Search ADS PubMed  4. Foad SL, Mehlman CT, Ying J. The epidemiology of neonatal brachial plexus palsy in the United States. J Bone Joint Surg Am.  2008; 90( 6): 1258- 1264. Google Scholar CrossRef Search ADS PubMed  5. Wilson TJ, Chang KW, Chauhan SP, Yang LJ. Peripartum and neonatal factors associated with the persistence of neonatal brachial plexus palsy at 1 year: a review of 382 cases. J Neurosurg Pediatr.  2016; 17( 5): 618- 624. Google Scholar CrossRef Search ADS PubMed  6. Mollberg M, Lagerkvist AL, Johansson U, Bager B, Johansson A, Hagberg H. Comparison in obstetric management on infants with transient and persistent obstetric brachial plexus palsy. J Child Neurol . 2008; 23( 12): 1424- 1432. Google Scholar CrossRef Search ADS PubMed  7. Al-Qattan MM, Clarke HM, Curtis CG. The prognostic value of concurrent Horner's syndrome in total obstetric brachial plexus injury. J Hand Surg . 2000; 25( 2): 166- 167. Google Scholar CrossRef Search ADS   8. Gherman RB, Ouzounian JG, Satin AJ, Goodwin TM, Phelan JP. A comparison of shoulder dystocia-associated transient and permanent brachial plexus palsies. Obstet Gynecol . 2003; 102( 3): 544- 548. Google Scholar PubMed  9. Waters PM. Comparison of the natural history, the outcome of microsurgical repair, and the outcome of operative reconstruction in brachial plexus birth palsy. J Bone Joint Surg Am.  1999; 81( 5): 649- 659. Google Scholar CrossRef Search ADS PubMed  10. Babic SH, Kokol P, Podgorelec V, Zorman M, Sprogar M, Stiglic MM. The art of building decision trees. J Med Syst . 2000; 24( 1): 43- 52. Google Scholar CrossRef Search ADS PubMed  11. Scheer JK, Osorio JA, Smith JS et al.   Development of validated computer based pre-operative predictive model for proximal junction failure (PJF) or clinically significant PJK with 86% accuracy based on 510 ASD patients with 2-year follow-up. Spine . 2016; 41( 22): E1328- E1335. Google Scholar CrossRef Search ADS PubMed  12. Narakas AO. Obstetrical brachial plexus injuries. In: Lamb DW, ed. The Paralysed Hand. The Hand and Upper Limb . New York, NY: Churchill Livingstone; 1987: 116- 135. 13. Birch R. Obstetric brachial plexus palsy. J Hand Surg . 2002; 27( 1): 3- 8. Google Scholar CrossRef Search ADS   14. Chauhan SP, Blackwell SB, Ananth CV. Neonatal brachial plexus palsy: incidence, prevalence, and temporal trends. Semin Perinatol . 2014; 38( 4): 210- 218. Google Scholar CrossRef Search ADS PubMed  15. Malessy MJ, Pondaag W, Yang LJ, Hofstede-Buitenhuis SM, le Cessie S, van Dijk JG. Severe obstetric brachial plexus palsies can be identified at one month of age. PloS One . 2011; 6( 10): e26193. Google Scholar CrossRef Search ADS PubMed  16. Ali ZS, Bakar D, Li YR et al.   Utility of delayed surgical repair of neonatal brachial plexus palsy. J Neurosurg Pediatr.  2014; 13( 4): 462- 470. Google Scholar CrossRef Search ADS PubMed  17. Little KJ, Zlotolow DA, Soldado F, Cornwall R, Kozin SH. Early functional recovery of elbow flexion and supination following median and/or ulnar nerve fascicle transfer in upper neonatal brachial plexus palsy. J Bone Joint Surg Am.  2014; 96( 3): 215- 221. Google Scholar CrossRef Search ADS PubMed  18. Smith NC, Rowan P, Benson LJ, Ezaki M, Carter PR. Neonatal brachial plexus palsy. Outcome of absent biceps function at three months of age. J Bone Joint Surg Am . 2004; 86-A( 10): 2163- 2170. Google Scholar CrossRef Search ADS PubMed  19. Seruya M, Shen SH, Fuzzard S, Coombs CJ, McCombe DB, Johnstone BR. Spinal accessory nerve transfer outperforms cervical root grafting for suprascapular nerve reconstruction in neonatal brachial plexus palsy. Plast Reconstr Surg . 2015; 135( 5): 1431- 1438. Google Scholar CrossRef Search ADS PubMed  20. Bade SA, Lin JC, Curtis CG, Clarke HM. Extending the indications for primary nerve surgery in obstetrical brachial plexus palsy. BioMed Res Int . 2014; 2014: 627067. Google Scholar CrossRef Search ADS PubMed  21. Chang KW, Yang LJ, Driver L, Nelson VS. High prevalence of early language delay exists among toddlers with neonatal brachial plexus palsy. Pediatr Neurol . 2014; 51( 3): 384- 389. Google Scholar CrossRef Search ADS PubMed  22. Yang LJ, Anand P, Birch R. Limb preference in children with obstetric brachial plexus palsy. Pediatr Neurol . 2005; 33( 1): 46- 49. Google Scholar CrossRef Search ADS PubMed  23. Bellew M, Kay SP, Webb F, Ward A. Developmental and behavioural outcome in obstetric brachial plexus palsy. J Hand Surg . 2000; 25( 1): 49- 51. Google Scholar CrossRef Search ADS   24. Curtis C, Stephens D, Clarke HM, Andrews D. The active movement scale: an evaluative tool for infants with obstetrical brachial plexus palsy. J Hand Surg . 2002; 27( 3): 470- 478. Google Scholar CrossRef Search ADS   25. Borschel GH, Clarke HM. Obstetrical brachial plexus palsy. Plast Reconstr Surg . 2009; 124( 1 suppl): 144e- 155e. Google Scholar CrossRef Search ADS PubMed  26. Waters PM. Update on management of pediatric brachial plexus palsy. J Pediatr Orthop B.  2005; 14( 4): 233- 244. Google Scholar CrossRef Search ADS PubMed  COMMENT The authors have created a thoughtful decision tree to use maternal and neonatal factors to attempt to predict need for surgery in neonatal brachial plexus injury. They have validated this retrospectively using 333 patients, 145 of whom (43.5%) were indicated for surgery, and 188 (56.5%) who were deemed non-surgical candidates. The decision tree is shown in Figure 2 and uses the following data points: presence of Horner's syndrome, presence of pseudomeningocele, Narakas grade, presence of clavicle fracture at birth, birth weight >9 lbs, and induction vs augmentation of labor. The positive predictive value for surgery is 0.94, with a negative predictive value of 0.79, which indicates a conservative algorithm. The authors correctly propose the algorithm as a guideline, which may assist in the decision to take these children to surgery. The authors point out in the discussion that unlike clinical exam findings such as the Toronto Active Movement Scale, or "cookie test" which are associated with physical exam findings at 3-6 months, the current algorithm is designed to provide predictive value based on information present just following birth. The authors correctly state that, to date, operating very early (less than 3 months) has, to this point, not been shown to improve outcomes, and they go on to indicate that the algorithm could likely be improved by incorporation of these physical exam findings over time, which would likely close the gap and improve the negative predictive value of 79%, which is the single biggest criticism of version #1 of the algorithm. This work represents a solid contribution and another tool in the armamentarium used by neonatal brachial plexus surgeons in their decision tree as they work with infants and their families to come to the right decision regarding the role and timing of surgery, especially at very early timepoints. One minor implication from the algorithm is that patients with a Narakas grade III or IV injury with a clavicle fracture would necessarily be indicated for surgery, which appears not to have been reported previously. Chandan Gopal Reddy Rochester, Minnesota Copyright © 2017 by the Congress of Neurological Surgeons

Journal

NeurosurgeryOxford University Press

Published: Mar 1, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off