1 INTRODUCTION

Cleft lip and/or palate (CL/P) is the most common congenital craniofacial anomaly caused by the fusion failure of facial tissues. The incidence rate of CL/P ranges from 0.57 to 1.57 per 1000 live births; in the Japanese population, the rate is 1.56 per 1000 births.1 The frequency of clefts varies widely by race, and Asians tend to have higher rates of CL/P.2 CL/P has a multifactorial and polygenic aetiology, and various environmental influences (eg maternal use of retinoids or other drugs, maternal temperature, maternal smoking, and diet) and genetic links (eg IRF 6 in CP cases) have been reported.3

Patients with CL/P generally undergo a multidisciplinary and comprehensive clinical protocol that includes lip and palate repair, orthodontic treatments and speech therapy.4 It has been reported that patients who did not undergo surgical repair for various reasons tend to show normal maxillary developmental potential.5 In contrast, it is common for patients with a repaired CL/P to exhibit a maxillary growth deficiency, likely due to the surgical repair of the hard palate.6-8 This suggests that the scar tissue from hard palatal surgery results in characteristic growth patterns in patients with CL/P, including maxillary retrusion.9 Although several clinical approaches, involving various surgical protocols and orthodontic treatments, have been applied to improve this growth inhibition, severe cases may still require orthognathic surgery (OGS) to correct skeletal discrepancies.10

Several studies have attempted to identify prognostic factors for OGS in patients with CL/P. The most common predictors of OGS in patients with CL/P are skeletal features obtained by lateral cephalograms.11 In a study involving 61 patients with unilateral CLP (UCLP), maxillo-mandibular discrepancy and small maxillary length at 11 years of age were found to predict the need for OGS at 18 years of age, with 86.9% accuracy.12 Another study of 131 patients with cleft lip and alveolus (CLA), unilateral CLP (UCLP), and bilateral CLP (BCLP) also suggested that skeletal Class III tendency at 9 years of age was a predictor of OGS at 17 years of age.13 The resultant scar from palatal repair surgery has also been regarded as a primary cause of maxillary growth inhibition in children with CL/P.14, 15 In a study that evaluated the dental arch relationship of patients with UCLP at 11 years of age according to their hard palatal repair techniques, 47 patients were treated with a method that exposes the bone surface, while 61 patients received vomerplasty to tightly close the soft tissues.16 The group with bone exposure showed a significantly higher EUROCRAN Index score (3.04 in the bone exposure group versus 2.63 in the vomerplasty group), indicating an unfavourable dental arch relationship. A systemic review of the influence of the surgery method for palatal closure on the maxillary growth in patients with UCLP found that delayed hard palatal repair resulted in a more positive prognosis in comparison with early hard palatal repair.17 In a six-centre international study of patients with CL/P, palatal repair surgery was also suggested to be a factor influencing surgical outcomes.18

The inherited conditions of patients with CL/P have also been found to be related to their growth and treatment needs. A study comparing the cleft types of 131 male patients with CLA, UCLP and BCLP revealed that CLP had a significantly higher prevalence of OGS than CLA, and BCLP had a significantly higher prevalence of OGS than UCLP.13 In that study, the prognosis at 17 years of age varied according to cleft type, and the prevalence of OGS in patients with CLA, UCLP and BCLP was 8.5%, 21.4% and 30.0%, respectively.

Another prognostic factor for impaired craniofacial growth in patients with CL/P is missing permanent teeth, the prevalence of which is significantly higher in patients with cleft than in non-cleft populations.19 A study evaluating the lateral cephalograms of 73 UCLP patients, taken from 7 to 11 years of age, reported that the total number of congenitally missing permanent teeth in the maxillary dental arch influenced unique craniofacial characteristics.20 The anterior facial height, length of the maxillary dental arch and overjet were significantly smaller as the number of missing teeth increased. Various studies have shown that certain genetic factors, such as MSX1, contribute to both missing teeth and CL/P.21-24

Furthermore, the severity of lip separation at birth, determined by the absence of Simonart's band, has also been reported to be weakly correlated with the clockwise rotation of maxillary growth, based on lateral cephalograms obtained at 17 years of age in 90 patients with UCLP from three different centres.25 In another study, the prevalence of missing lateral incisor at 7 years of age in patients with and without Simonart's band was 40.5% and 45.9%, respectively.26

As described above, a number of factors have been found to affect the prognosis or maxillary growth of patients with CL/P. However, most of these factors were not considered simultaneously, and the influence of various confounders remains unclear. As individuals with CL/P undergo multidisciplinary treatments from birth to adulthood,27, 28 determining prognostic factors and predicting the need for OGS in children with CL/P are needed, not only for reducing the risk of deterioration but also for developing effective early-stage treatment plans to reduce the clinical burden. Recently, artificial intelligence (AI) systems using machine learning have become available for prognostic prediction based on the comprehensive examination of multiple variables.29 Machine learning is a subfield of AI that depends on algorithms to predict prognosis based on the dataset.

Thus, the purpose of this study was to develop comprehensive mathematical models to determine the prognostic factors and predict the need for OGS in mixed dentition patients with CL/P by AI systems using machine learning. The factors that were analysed included congenital conditions (cleft type, the severity of lip separation at birth, the number of missing teeth, sex), clinical history (surgery methods of palatal closure and surgeons, duration of maxillary protraction and expansion), and lateral cephalogram data.

2 METHODS 2.1 Samples

In this retrospective cohort study, patients who visited the orthodontic department of a university dental hospital between 2002 and 2014 were consecutively recruited. The inclusion criteria were as follows: (a) Japanese patients with bilateral and unilateral CL/P with/without syndromes; and (b) patients for whom at least three lateral cephalograms were obtained at 7 (T1), 10 (T2) and 15 (T3) years of age. The exclusion criteria were as follows: (a) submucous cleft palate and cleft of soft palate; and (b) missing records of palatal repair surgery. The institutional treatment protocol used for patients with CL/P during the study period was as follows: (a) lip repair surgery was performed at between 3 and 6 months of age; (b) palatal closure surgery was performed at around 12 months of age using the palatal push-back method or between 12 and 18 months of age using Furlow's method (the main method for palatal surgery before 2000 was the one-step push-back method; after 2000 it was the two-step Furlow's method); (c) if needed, the first-phase orthodontic treatment, including maxillary protraction and expansion, was initiated at approximately 7 years of age; (d) alveolar bone graft was performed at approximately 9 years of age, and treatment plans were re-evaluated at 10 years of age or after incisor eruption was completed; and (e) the second-phase orthodontic treatment was initiated at or after 15 years of age.

The following factors were examined as predictive variables: the type of cleft, the severity of lip separation at birth, the number of congenitally missing teeth, sex, surgery method and surgeon for palatal closure, cephalometric variables at T1 and T2 (Table S1), growth between T1 and T2, and the total duration of orthodontic treatment (maxillary protraction and expansion from T1 to T3). The severity of lip separation at birth was determined by the absence of Simonart's band (a band of soft tissue that joins the cleft in the lip) before the lip repair surgery. Patients with CP were considered to have the soft tissue continuity of the lip. The number of congenitally missing teeth was evaluated by the orthopantomograms from T1 to T3. This included congenitally missing permanent teeth in both maxilla and mandible. The third molars, deciduous teeth and extracted permanent teeth were excluded. The cephalometric landmarks of lateral cephalograms taken at T1, T2 and T3 were digitized with a customized version of the MATLAB software program (MathWorks) by a single operator (J.L). Thus, a total of 22 and 30 predictors at T1 and T2, respectively, were enrolled in the prediction analysis.

As outcome variables, the clinical decision regarding the need for OGS, skeletal discrepancy (ANB ≥ 0 or <0°) and overjet (≥0 or <0 mm) at T3 was examined. Patients were classified into the OGS and non-OGS groups according to whether they were diagnosed OGS at T3 by a consensus of more than two orthodontic specialists through a diagnostic conference. Skeletal and dental discrepancies at T3 were also considered as principal outcomes to determine the need for OGS, and these were thus enrolled in the prediction outcomes. An AI system with machine learning was used to enter combinations of the predictive variables observed at T1 and T2 into multivariable logistic regression models and predict the outcomes (RStudio 1.10, Boston, MA, USA). The association between risks and outcomes was evaluated using the odds ratio (OR). Predictive accuracy was evaluated using the area under the receiver operating characteristic curve (AUC). A post hoc statistical power analysis was performed.

This study was approved by the research ethics committee of the Graduate School of Dentistry, Osaka University (H29-E11-1).

2.2 Statistical analysis

A multivariate analysis is a set of techniques used for information extraction, regression or classification from correlated data sets. Among them, two methods could be applied to the present data: a logistic regression analysis and a discriminant analysis. Both are appropriate for evaluating relationships between multiple dependent variables (possible risk factors) and one independent variable (treatment outcome), which are represented as 0 or 1. A logistic regression analysis is particularly efficient when evaluating the prevalence of each risk factor using odds ratios. Thus, we selected a logistic regression analysis for the present study. Furthermore, a receiver operating characteristic curve, which is a graphical plot that illustrates the diagnostic ability of a binary classifier system (as its discrimination threshold is varied), was applied with the regression analysis in the present study.

3 RESULTS 3.1 Samples

A total of 7285 patients who visited the orthodontic department of a university dental hospital between 2002 and 2014 were recruited. Among them, 1502 had CL/P. Among these patients, lateral cephalograms were obtained at T1, T2 and T3 for 130 patients. Four patients were excluded as follows: 3 patients had incomplete clefts (2 with submucous cleft palate, 1 with cleft of the soft palate), and the record of palatal repair surgery was missing for 1 patient. Finally, in this study, we investigated 126 patients with CL/P. Two patients who had syndromes were included.

3.2 Patient characteristics

This study included 126 patients with bilateral and unilateral CL/P (70 males, 56 females; 19 with BCLP, 5 with BCLA, 71 with UCLP, 20 with UCLA and 11 with CP). All study participants had at least three lateral cephalograms at 7 (T1; mean age, 7.03; standard deviation, 0.68), 10 (T2; mean age, 9.97; standard deviation, 0.82) and 15 (T3; mean age, 14.94; standard deviation, 0.68) years of age. Simonart's band (or continuity of the lip) was observed in 80 patients and was absent in 46. For palatal closure surgery, 55 patients received the push-back method, 30 received Furlow's method, 18 received another method, and 25 were patients with CLA did not receive palatal closure surgery. Regarding the surgeons, 57 patients were treated by operator A, 22 were treated by operator B, 10 were treated by operator C, and 14 were surgeons other than operator A, B or C. Regarding orthodontic treatments, 62 patients received maxillary protraction for a mean duration of 977.15 days, and 95 patients received maxillary expansion for a mean duration of 361.48 days. A total of 80 patients required OGS, 74 had an ANB angle of <0° at T3, and 62 had overjet of < 0 mm at T3.

3.3 Development of prediction models

A total of six prediction models was developed using AI systems. The mean AUC was 0.93 (0.91 to 0.99), indicating high predictive accuracy (AUC > 0.9; Table 1; Figure S1). The observed statistical power of the post hoc analysis was 1.00 for the models.

TABLE 1. The sensitivity, specificity and power of the prediction models using variables at 7 (T1) and 10 (T2) years of age for predicting outcomes at 15 y of age (T3) Outcome at T3 Variables at AUC 95% CI Sensitivity Specificity R 2 Power Need for orthognathic surgery T1 0.91 0.85-0.96 0.78 0.87 .42 1.00 T2 0.91 0.85-0.96 0.78 0.87 .47 1.00 Skeletal discrepancy (ANB angle < 0°) T1 0.95 0.91-0.99 0.96 0.88 .55 1.00 T2 0.99 0.97-1.00 0.96 0.92 .65 1.00 Overjet < 0 mm T1 0.91 0.85-0.96 0.78 0.87 .46 1.00 T2 0.95 0.91-0.99 0.89 0.90 .55 1.00 3.4 Prognostic factors

The prognostic factors that were determined are shown in Tables 2, 3 and 4. In short, the predictive factors for OGS, ANB angle < 0° at T3 and overjet < 0 mm at T3 were as follows, as ranked by order of OR (highest to lowest): the number of clefts in the lip and alveolus, the use of the palatal push-back technique, male sex, smaller protrusion and length of the maxilla, smaller length of the anterior cranial base, greater mandibular protrusion, greater increase in lower facial height, smaller clockwise rotation of the palatal and mandibular plane between T1 and T2, smaller increase in palatal length, the number of missing teeth, and steeper palatal and mandibular plane at T1 and T2 (Tables 2, 3, 4; Figure 1, Figure S2). In particular, the number of clefts in the lip and alveolus showed relatively high ORs for OGS (OR: 6.35 and 13.15 at T1 and T2, respectively), as did anterior crossbite at T3 (OR: 15.63 and 89.29 at T1 and T2, respectively); these findings indicate that patients with BCL/P had a higher probability for these outcomes than others. Achieving palatal closure with the push-back method, rather than Furlow's method, was also found to be a predictive factor for anterior crossbite at T3, with high ORs (2.45 and 9.04 at T1 and T2, respectively). On the other hand, cleft in the palate, the severity of lip separation at birth, the surgeon performing palatal closure and the total duration of orthodontic treatment were not found to be prognostic predictors.

TABLE 2. Prognostic factors at 7 (T1) and 10 (T2) years of age for orthognathic surgery at 15 years of age (T3) Prognostic factors (orthognathic surgery at T3) T1 T2 OR P-value 95% CI OR P-value 95% CI Upper bound Lower bound Upper bound Lower bound Existence of cleft in palate 6.758 .120 0.606 75.319 14.667 .064 0.860 250.156 Number of clefts in the lip and alveolus 6.692 .006 * 1.711 26.167 13.644 .002 * 2.627 70.850 Male 0.643 .502 0.177 2.338 0.467 .336 0.099 2.200 Push-back method 1.124 .867 0.286 4.422 1.623 .563 0.314 8.381 Other than push-back/Furlow's method 0.473 .481 0.059 3.794 0.326 .374 0.028 3.859 Existence of Simonart's band 0.991 .217 0.976 1.006 0.989 .248 0.971 1.008 Number of missing teeth 1.712 .044 * 1.015 2.885 1.660 .096 0.914 3.016 Duration of maxillary protraction 0.999 .127 0.998 1.000 0.999 .166 0.998 1.000 Duration of maxillary expansion 1.001 .346 0.999 1.003 1.001 .499 * 0.998 1.003 Operator A 1.745 .551 0.280 10.892 0.747 .779 0.098 5.685 Operator B 6.008 .160 0.493 73.256 3.782 .397 0.175 81.878 Operator C 6.541 .227 0.310 138.061 2.200 .664 0.063 76.690 Other than operator A, B, C 1.494 .755 0.120 18.533 0.943 .967 0.060 14.839 S-N 1.166 .327 0.857 1.587 0.718 .130 0.468 1.103 SNA 0.761 .115 0.542 1.069 0.916 .692 0.594 1.414 A-Ptm 0.958 .788 0.702 1.308 1.823 .019 * 1.106 3.005 SNB 1.739 .008 * 1.158 2.612 1.193 .356 0.820 1.737 Ar-Me 0.836 .165 0.649 1.076 0.862 .350 0.631 1.177 SN-PP 1.065 .493 0.890 1.273 0.995 .963 0.799 1.239 SNMP 1.025 .779 0.864 1.216 1.050 .638 0.857 1.286 Me-PP 1.197 .261 0.875 1.637 1.126 .560 0.755 1.678 ∆S-N – – – – 1.289 .606 0.492 3.376 ∆SNA – – – – 1.152 .576 0.702 1.889 ∆A-Ptm – – – – 1.078 .727 0.708 1.641 ∆SNB – – – – 0.723 .315 0.384 1.361 ∆Ar-Me – – – – 0.895 .642 0.560 1.429 ∆SN-PP – – – – 0.781 .114 0.575 1.061 ∆SNMP – – – – 0.805 .344 0.515 1.261 ∆Me-PP – – – – 1.172 .573 0.675 2.035 Abbreviations: ∆, difference between T1 and T2;APtm, maxillary length; Ar-Me, mandibular length; Me-PP, lower facial height; SN, cranial base length; SNA, maxillary prominence; SNB, mandibular prominence; SN-MP, mandibular plane inclination to cranial base; SN-PP, palatal plane inclination to cranial base. TABLE 3. Prognostic factors at 7 (T2) and 10 (T2) years of age for skeletal discrepancy (ANB angle < 0°) at 15 years of age (T3) Prognostic factors (ANB angle < 0° at T3) T1 T2 OR P-value 95% CI OR P-value 95% CI Upper bound Lower bound Upper bound Lower bound Existence of cleft in palate 2.068 .692 0.057 75.397 1.22E−02 .401 4.15E−07 3.57E+02 Number of clefts in the lip and alveolus 3.085 .153 0.657 14.480 1.98E+01 .188 2.33E−01 1.68E+03 Male 2.236 .357 0.403 12.398 5.64E+02 .109 2.42E−01 1.31E+06 Push-back method 4.358 .156 0.571 33.288 1.60E+05 .099 1.05E−01 2.45E+11 Other than push-back/Furlow's method 15.706 .051 0.987 249.968 2.33E+02 .219