Predictive Model for Congenital Muscular Torticollis: Analysis of 1021 Infants With Sonography

ORIGINAL ARTICLE Predictive Model for Congenital Muscular Torticollis: Analysis of 1021 Infants With Sonography Miao-Ming Chen, MD, Huan-Cheng Chang, MD, MBA, Chuan-Fa Hsieh, MS, Ming-Fang Yen, PhD, Tony Hsui-Hsi Chen, PhD 2199 ABSTRACT. Chen M-M, Chang H-C, Hsieh C-F, Yen M-F, Chen TH. Predictive model for congenital muscular torticollis: analysis of 1021 infants with sonography. Arch Phys Med Rehabil 2005;86:2199-203. Objective: To construct a predictive model to foretell congenital muscular torticollis (CMT) on the basis of clinical correlates. Design: Correlation study. Setting: Regional hospital. Participants: A consecutive series of 1021 newborn infants. Interventions: Not applicable. Main Outcome Measure: Participants underwent portable ultrasonography to diagnose CMT. Significant clinical correlates were identified to construct a predictive model using the logistic regression model. Results: Forty of 1021 infants were diagnosed with CMT using ultrasonography, yielding an overall incidence of 3.92%. Birth body length (odds ratio [OR] 1.38; 95% confidence interval [CI], 1.49 2.38), facial asymmetry (OR 21.75; 95% CI, 6.6 71.7), plagiocephaly (OR 22.3; 95% CI, 7.01 70.95), perineal trauma during delivery (OR 4.26; 95% CI, 1.25 14.52), and primiparity (OR 6.32; 95% CI, 2.34 17.04) were significant correlates. A predictive logistic regression model with the incorporation of these 4 correlates was developed. We used cross-validation with a receiver operating characteristic curve to validate the predictive model. Conclusions: Our study successfully developed a quantitative predictive model for estimating the risk of CMT on the basis of clinical correlates only. This model has good discriminative ability for classifying CMT and non-cmt by yielding acceptable values of false-negative and false-positive cases. Key Words: Projections and predictions; Rehabilitation; Torticollis; Ultrasonography. 2005 by the American Congress of Rehabilitation Medicine and the American Academy of Physical Medicine and Rehabilitation From the Department of Rehabilitation (M-M Chen), Division of Family Medicine (Chang), and Research Center (Hsieh), Li Shin Hospital, Taoyuan County, Taiwan; and Institute of Preventive Medicine, College of Public Health, National Taiwan University, Taipei, Taiwan (Yen, H-H Chen). No commercial party having a direct financial interest in the results of the research supporting this article has or will confer a benefit upon the author(s) or upon any organization with which the author(s) is/are associated. Reprint requests to Tony Hsui-Hsi Chen, PhD, Institute of Preventive Medicine, College of Public Health, National Taiwan University, Room 207 No 19, HsuChow Rd, Taipei, Taiwan, e-mail: stony@episerv.cph.ntu.edu.tw. 0003-9993/05/8611-9877$30.00/0 doi:10.1016/j.apmr.2005.05.010 CONGENITAL MUSCULAR torticollis (CMT) is a common disease in infants, second only in incidence to dislocation of the hip and clubfoot among congenital musculoskeletal disorders. 1 The incidence rate of CMT is approximately 0.4%, 2 ranging from 0.3% to 1.9%. 3-7 The major clinical feature is characterized by persistent fibrosis of the sternocleidomastoid muscle that may lead to head tilt in mild cases and specific findings of facial asymmetry in severe cases. This may include changes in the maxilla and mandible, characterized by flattening of the contralateral occipitoparietal region and the ipsilateral fronto-orbital region. In severe cases, the enlargement of a mass often occurs at 1 month of age, remains static for 2 to 3 months, dwindles degree by degree afterward, and eventually disappears for life. 2 The primary treatment of CMT is physical therapy (PT), including manual stretching by PT and an active positioning program. 2 Surgical release is recommended for refractory cases. 2,8,9 Although most patients have a favorable prognosis after treatment, a few are still subject to impairment of activity and severe fibrosis after therapy. This is mainly attributable to delay of treatment to an older age or the initial degree of muscle damage that produces restriction of movement and considerable fibrosis. The diagnosis of CMT usually relies on clinical expression and physical examination. A few cases were diagnosed by the supplemental use of radiography, computed tomography scan, and fine-needle biopsy. Ultrasonography has been shown to be useful for evaluating neck masses in children. 10-14 The advantage of using ultrasonography over traditional methods is also because of its high sensitivity and the facilitation of follow-up and evaluation after treatment. However, because the incidence rate of CMT is low, routine checkup for CMT with sonography may be costly and time-consuming. The alternative is to make use of information on clinical correlates in association with CMT with sonography by the use of a 2-stage method that identifies suspected cases at first stage with a questionnaire composed of a series of questions related to significant correlates. These suspicious cases are further referred for sonographic examination. Despite a number of studies addressing the elucidation of clinical determinants of CMT, the impact of clinical correlates on the diagnosis of CMT is poorly quantified. Furthermore, the findings are also heterogeneous. Most studies have found that intrauterine malposition or trauma in the labor process were highly associated with CMT. 15-17 One study found that perinatal compartment syndrome may lead to the occurrence of CMT. 18 Heredity may also account for CMT. 19 These findings together with wide ranges in the incidence of CMT in different countries suggest that factors accounting for CMT are heterogeneous and may vary from country to country. This suggests that the elucidation of relevant risk factors associated with CMT in different racial groups would be helpful. If the relation between reputed causes and the occurrence of CMT can be better defined, a predictive model based on significant clinical correlates or risk factors can be established. Such a model is useful for screening CMT in order to reduce the clinical burden on ultrasonographic examination. Therefore, the aim of this study was to estimate the magnitude of morbidity of CMT and to define better the association among relevant clinical correlates associated with CMT. The

2200 PREDICTIVE MODEL FOR CONGENITAL MUSCULAR TORTICOLLIS, Chen Table 2: Comparison of Categoric Variables Between Normative and Abnormal Sonography Findings Variable Normative Abnormal P Facial asymmetry 1.1 (11/981) 32.5 (13/40).05 Plagiocephaly 1.7 (17/981) 32.5 (13/40).05 Neck ROM limitation 0 (0/981) 27.5 (11/40).05 Abnormal palpation finding 0 (0/981) 10.0 (4/40).05 Abbreviation: ROM, range of motion. Fig 1. Clinical example of facial asymmetry. predictive model for foretelling CMT was developed on the basis of significant correlates identified in the current study. METHODS Study Participants A consecutive series of 1021 infants born within 24 to 72 hours at a regional hospital were enrolled in this study through the period from January to August in 2002. Because most newborn infants did not stay in the hospital longer than 72 hours, screening was performed during this period. Each underwent portable ultrasonography examination to assess the condition of soft tissue around bilateral sternocleidomastoid muscle and the functional activity of the head and neck. Ultrasonographic examinations were performed by 2 physicians who screened patients alternately. To assess interobserver variation, we randomly sampled 30 infants who received duplicate blinded examinations by both physicians. Diagnosis Abnormalities of the soft tissues found on sonographic images were classified into 4 categories: type 1, the presence of a small tumor; type 2, diffuse fibrosis without thickening or tightness of the sternomastoid muscle; type 3, diffuse fibrosis with abnormal muscle; and type 4, severe diffusive fibrosis. Table 1: Comparison of Continuous Variables Regarding Delivery History Between Normative and Abnormal Sonography Findings Variable Normative Abnormal P Body height (cm) 50.23 1.92 51.33 1.49.05 Body weight (gm) 3069.3 375.2 3246.7 268.3.05 Head circumference (cm) 33.23 9.36 33.99 1.05.14 Shoulder width (cm) 12.09 2.06 12.43 0.76.05 Cesarean section 32.9 (323/981) 32.5 (13/40).87 Primiparity 51.8 (508/981) 82.5 (33/40).05 Breech presentation 17.2 (169/981) 22.5 (9/40).24 Birth trauma 2.5 (25/981) 22.5 (9/40).05 Cord around neck 13.9 (136/981) 10.0 (4/40).76 Use of forceps 17.9 (118/658) 44.4 (12/27).05 NOTE. Values are mean standard deviation and percentage (n/n). Data Collection The measurement of neck motion includes rotation and lateral flexion. Rotation above 90 is defined as normal. For lateral flexion, normative range is defined by the head being flexed so that the ear touches the shoulder on the same side. In the abnormal subjects, the maximum degree of lateral flexion was recorded. Other information collected included sex, height, weight, order of birth, trauma of birth, instruments used in the delivery process, and facial asymmetry, an example of which is seen in figure 1. Statistical Analysis The comparison of continuous variables was made by the use of an independent t test. Categoric variables were assessed by the chi-square test. We used a multiple logistic regression model to ascertain relevant factors that were significant in univariate analysis. To predict the probability of being CMT, a weighted score (T) for predicting abnormal sonography findings was constructed as follows: T a 1 Birth body length a 2 Facial asymmetry a 3 Plagiocephaly a 4 Birth trauma a 5 Primiparity 100 where a 1 a 5 were estimated regression coefficients obtained from the multiple logistic regression model. Note that birth body length, facial asymmetry, plagiocephaly, birth trauma, and primiparity were significant predictors identified in the multiple logistic regression model (see below). The predicted probability (P) is calculated by P 1/[1 exp( T)]. The cross-validation was also performed to validate the predictive model mentioned above. Two thirds of the patients were randomly selected into the training sample and one third into the test sample. The training sample was used to fit the logistic regression model and estimate the regression coefficients, which were further applied to the test sample to obtain the predictive value. The predicted value generated from the test sample was compared with the observed value to estimate sensitivity (Se) and the false-positive rate (1 specificity). The 95% confidence interval (CI) of sensitivity (or false-positive rate) was calculated by the estimated value (Se) plus or minus 1.96 times the standard error (SE), which is the square root of (Se[1 Se])/n, where n is the number of CMT, following Baker et al. 20 To avoid random fluctuation, 100 iterations were performed with the same procedure mentioned above by randomly selecting a training and test data set at 1 time to fit the model and, in turn, to compute a receiver operating characteristic (ROC) curve in each step. To assess whether the model has good predictive validity, the ROC curve with the mean of sensitivity in the y axis and of (1 specificity) in the x axis from the 100 random test-sample sets was illustrated. The mean area under the ROC curves and its 2.5th and 97.5th percentile were also reported.

PREDICTIVE MODEL FOR CONGENITAL MUSCULAR TORTICOLLIS, Chen 2201 Table 3: Estimated Results of Multiple Logistic Regression Model Variable Regression Coefficients (a 1 a p ) Adjusted OR 95% CI Birth body length 0.633 1.88 1.49 2.38 Facial asymmetry 3.08 21.75 6.60 71.70 Plagiocephaly 3.10 22.30 7.01 70.95 Birth trauma 1.45 4.26 1.25 14.52 Primiparity 1.84 6.32 2.34 17.04 RESULTS Of the 1021 subjects, 40 infants were determined to be abnormal with sonography. The overall incidence rate diagnosed with use of ultrasonography was 3.92%. Boys had a higher rate than girls (57.5% vs 42.5%). However, no statistically significant difference was noted (P.51). The incidence rates were 0.98% for type 1 (n 1), 2.45% for type 2 (n 25), and 1.37% for type 3 (n 14). There was no type 4 case. The association between type and abnormal findings was statistically significant (P.01). Regarding location, the incidence rates were 2.4% for left involvement (n 25) and 1.5% for right involvement. The interexaminer concordance rate was 100% with a value of 1.00. Table 1 shows the comparisons of basic physiologic measurements between the abnormal and normative group. The abnormal group was statistically significantly longer (P.001), heavier (P.001), and with wider shoulder width (P.001) than the normative group. There was no substantial difference in head circumference between the 2 groups (P.14). Regarding birth history, no association between type of birth and abnormality was found (P.87). First birth order, trauma, and the use of instruments were highly associated with abnormal ultrasonographic findings. First birth order had a higher proportion of abnormality than subsequent birth orders (P.001). Facial asymmetry was statistically associated with abnormal findings (P.001). Limitation of neck motion was also highly associated with abnormality (P.001) (table 2). Association The univariate analysis of the logistic regression model shows that the higher risk associated with abnormal sonography finding was found for high weight (odds ratio [OR] 1.002; 95% CI, 1.001 1.002), long body length (OR 1.53; 95% CI, 1.29 1.82), the presence of facial asymmetry (OR 45.72; 95% CI, 19.02 109.95) or plagiocephaly (OR 29.40; 95% CI, 13.16 65.71), primiparity (OR 3.84; 95% CI, 1.76 8.40), the use of forceps (OR 3.25; 95% CI, 1.66 6.35), and occurrence of delivery trauma (OR 5.31; 95% CI, 1.92 14.67). Table 3 shows the results of multivariable logistic regression. After adjustment for all variables in each other, only Fig 2. The ROC curve for the predictive model for CMT. height, facial asymmetry or plagiocephaly, perineal trauma during delivery, and primiparity remained statistically significant. An increase of 1cm of body length conferred an extra 88% risk of abnormal findings. Pregnant women with perineal trauma had a 4-fold greater risk of abnormal findings. Primiparity was 6 times more likely to have abnormal finding than multiparity. Infants who had facial asymmetry or plagiocephaly had a 22-fold greater risk of abnormal findings than those with normal facial symmetry. Regression coefficients listed in table 3 give the weighted score (T) formula as follows: T score 63.3 (birth body length in centimeter) 308 (facial asymmetry; yes 0, no 1) 310 (plagiocephaly; yes 0, no 1) 145 (birth trauma; yes 0, no 1) 184.3 (primiparity; yes 0, no 1). Table 4 shows how the T score can be applied in calculating the predictive probability for 5 selected cases by using the formula (1/[1 exp( T)]). It can be seen that cases 1 through 3 had a higher likelihood of being abnormal than cases 4 and 5. If the cutoff point is set as 0.5, the predicted CMT is completely in agreement with the observed (see table 5, column 9). The optimal point (closest to the left-upper corner) selected from the ROC curve with the area equivalent to 91.40% (95% CI, 86.99% 95.81%) as shown in figure 2 was 85.37% (SE.056; 95% CI, 74.55% 96.19%) for sensitivity and 17.43% (SE.012; 95% CI, 15.06% 19.81%) for false-positive rate following the Baker method 20 mentioned in Methods. For cross-validation, the ROC curves from the 100 random test-sample sets are shown in figure 3. Because the mean area under the ROC curves (the line with bold) was 90.05% (95% credible interval, 79.37% 96.02%), the model has good ability to predict CMT. Case Height Plagiocephaly Table 4: Predictive Probability for Selected Cases Facial Asymmetry Birth Trauma Primiparity T Score Predictive Probability Sonography* 1 53 1 1 0 1 387.64.98 1 2 49 1 1 0 1 134.42.79 1 3 52 0 1 0 1 13.86.53 1 4 56 0 0 1 0 80.27.31 0 5 54 0 0 1 0 20.69.11 0 *1 abnormal; 0 normal.

2202 PREDICTIVE MODEL FOR CONGENITAL MUSCULAR TORTICOLLIS, Chen Fig 3. Cross-validation of ROC curves from the 100 random test sample sets and the 1 with the mean value of 100 random test samples. Legend: black line, ROC curve with the mean value of 100 random test samples; gray lines, ROC curves for 100 random test samples. DISCUSSION By assessing 1021 infants with sonography, the present study estimated the incidence of CMT and identified significant correlates associated with CMT by using a logistic regression model. The estimated results from a multiple logistic regression model were also used to develop a predictive model for predicting CMT using clinical correlates only. The incidence rate, estimated as 3.92% in the current study, was higher than figures reported in earlier studies, ranging from 0.3% to 1.9%. The principal reason is that our incidence was estimated on the basis of sonographically detected cases rather than by traditional clinical diagnostic methods as reported in previous studies. Using traditional diagnostic methods, the incidence of CMT subjects included in this study was estimated at 1.08%, which falls within the range of estimated figures from previous studies. In this sense, sonography may be more sensitive in detecting occult cases than traditional diagnostic methods. The participants in this study underwent ultrasonography screening during their hospital stay, which was within 24 to 72 hours after birth. However, we may have missed some muscle fibrosis cases caused by trauma during the labor process because wound healing in these cases may take more than 72 hours and may have been preceded by intramuscular hematoma. However, as pointed out earlier, our incidence rate using traditional clinical diagnostic methods falls within the range reported in previous studies. Missing cases of CMT due to trauma may be uncommon. From a clinical viewpoint, our predictive model can easily calculate weighted score by hand. We developed a range of scores and corresponding predictive probabilities at the 5% interval, as shown in table 5. Take a case for example in table 5: the predicted score was 204 and the predictive probability lies between 55% and 60% from table 5 instead of using complicated algebra as done in table 4. The predictive model is very useful to ascertain the optimal balance between false-negative cases and false-positive cases. Figure 4 shows the probability density function of CMT and non-cmt given the corresponding mean and standard deviation and normal distribution. To achieve 95% sensitivity, the cutoff point is 477.12, which leads to a 50.63% false-positive rate. To achieve 95% specificity, the cutoff point is 203.57, which results in a false-negative rate of 33.60%. By using a weighted score, figure 2 is also very helpful for taking into account false-negative cases and Table 5: Reference Table for Predicting Risk of CMT With Sonography Probability T Score.99 459.5.95 294.4.90 219.7.85 173.5.80 138.6.75 109.9.70 84.7.65 61.9.60 40.6.55 20.1.50 0.0.45 20.1.40 40.6.35 61.9.30 84.7.25 109.9.20 138.6.15 173.5.10 219.7.05 294.4.01 459.5

PREDICTIVE MODEL FOR CONGENITAL MUSCULAR TORTICOLLIS, Chen 2203 Fig 4. Probability density function of score for non-cmt and CMT patients under a normal distribution assumption. false-positive cases. If the weighted score is smaller than 500, it is less likely that false-negative cases will be identified with clinical assessment. Similarly, if the weighted score is larger than 0, there is a low possibility of being a false-positive. The suggested range of weighted score for the confirmation of CMT with ultrasonography is between 500 (sensitivity, 96.0%; specificity, 43.9%) and 0 (sensitivity, 31.6%; specificity, 99.8%). Only suspected cases with this range need to be referred for ultrasonography. Thus, the predictive model used in this study may be very useful, not only to enhance the effectiveness of early detection of CMT due to the reduction in false-negative cases, but also in reducing the costs related to false-positive cases. Significant correlates identified in the final multivariable model included body height and length, primiparity, birth trauma, facial asymmetry, and plagiocephaly. The first 3 correlates may be a reflection of complicated labor process, which have been deemed as significant clinical correlates in previous studies. 15-17 The latter 2 correlates were classical clinical expressions of CMT. Although our clinical correlates are commensurate with clinical experience, the use of a predictive model for early detection of CMT and further early intervention makes a significant contribution to reducing long-term complication of CMT, including orbital dystopia, malocclusion, severe facial disfigurement, and visual axis problems. Because PT is the major treatment for CMT, early detection of CMT by use of the predictive model in conjunction with the early intervention of PT may produce an 80% reduction in CMT morbidity and only a few cases will need further surgery. CONCLUSIONS Our study successfully developed a quantitative predictive model for estimating the risk of CMT on the basis of clinical correlates only. This model has good discriminative ability for classifying CMT and non-cmt by yielding acceptable values of false-negative and false-positive cases. References 1. Bredenkamp JK, Hoover LA, Berke GS, Shaw A. Congenital muscular torticollis. A spectrum of disease. Arch Otolaryngol Head Neck Surg 1990;116:212-6. 2. Porter SB, Blount BW. Pseudotumor of infancy and congenital muscular torticollis. Am Fam Physician 1995;52:1731-6. 3. Fabian K, Marshall M. Conservative and surgical treatment of congenital muscular torticollis: a literature review. Physiother Can 1984;36:146-51. 4. Staheli LT. Muscular torticollis: late results of operative treatment. Surgery 1971;69:469-73. 5. Cheng JC, Au AW. Infantile torticollis: a review of 624 cases. J Pediatr Orthop 1994;14:802-8. 6. Ling CM, Balanchandran N. A prospective study of sternomastoid tumour in a closed community. In: Proceedings of the 10th Singapore-Malaysia Congress of Medicine. Vol 10. Singapore: Academy of Medicine; 1975. p 233-6. 7. Suzuki S, Yamamuro T, Fujita A. The aetiological relationship between congenital torticollis and obstetrical paralysis. Int Orthop 1984:8:175-81. 8. Gonzales J, Ljung BM, Guerry T, Schoenrock LD. Congenital torticollis: evaluation by fine-needle aspiration biopsy. Laryngoscope 1989;99(6 Pt 1):651-4. 9. Cheng JC, Tang SP. Outcome of surgical treatment of congenital muscular torticollis. Clin Orthop 1999;May(362):190-200. 10. Cheng JC, Metreweli C, Chen TM, Tang SP. Correlation of ultrasonographic imaging of congenital muscular torticollis with clinical assessment in infants. Ultrasound Med Biol 2000;26:1237-41. 11. Sherman NH, Rosenberg HK, Heyman S, Templeton J. Ultrasound evaluation of neck masses in children. J Ultrasound Med 1985;4:127-34. 12. Friedman AP, Haller JO, Goodman JD, Nagar H. Sonographic evaluation of non-inflammatory neck masses children. Radiology 1983;147:693-7. 13. Kraus R, Han BK, Babcock DS, Oestreich AE. Sonography of neck masses in children. AJRH Am J Roentgenol 1986;146:609-13. 14. Glasier CM, Seibert JJ, Williamson SL, et al. High resolution ultrasound characterization of soft tissue masses in children. Pediatr Radiol 1987;17:233-7. 15. Binder H, Eng G, Gaiser JF, Koch B. Congenital muscular torticollis: results of conservative management with long-term follow-up in 85 cases. Arch Phys Med Rehabil 1987;68:222-5. 16. Canale ST, Griffin DW, Hubbard CN. Congenital muscular torticollis: a long-term follow-up. J Bone Joint Surg Am 1982; 64:810-6. 17. Cheng JC, Wong MW, Tang SP, Chen TM, Shum SL, Wong EM. Clinical determinants of the outcome of manual stretching in the treatment of congenital muscular torticollis in infants: a prospective study of eight hundred and twenty-one cases. J Bone Joint Surg Am 2001;83:679-87. 18. Davids JR, Wenger DR, Mubarak SJ. Congenital muscular torticollis: sequela of intrauterine or perinatal compartment syndrome. J Pediatr Orthop 1993;13:141-7. 19. Thompson F, McManus S, Colville J. Familial congenital muscular torticollis. Clin Orthop 1986;Jan(202):193-6. 20. Baker SG, Kramer BS, Srivastava S. Markers for early detection of cancer: statistical guidelines for nested case-control studies. BMC Med Res Methodol 2002;2:1-8.