Aharonyan Artur, Anwar Syed, Choo HyeRan
Division of Plastic and Reconstructive Surgery, Neonatal and Pediatric Craniofacial Airway Orthodontics, Department of Surgery, Stanford University School of Medicine, 770 Welch Road, Palo Alto, CA, 94394, USA.
Sheikh Zayed Institute, Children's National Hospital, 111 Michigan Ave NW, Washington, DC, 20010, USA.
Int J Comput Assist Radiol Surg. 2025 Sep 9. doi: 10.1007/s11548-025-03515-w.
Alveolar molding plate treatment (AMPT) plays a critical role in preparing neonates with cleft lip and palate (CLP) for the first reconstruction surgery (cleft lip repair). However, determining the number of adjustments to AMPT in near-normalizing cleft deformity prior to surgery is a challenging task, often affecting the treatment duration. This study explores the use of machine learning in predicting treatment duration based on three-dimensional (3D) assessments of the pre-treatment maxillary cleft deformity as part of individualized treatment planning.
Digital 3D models of maxillary arches were collected from 35 infants with unilateral CLP. Key anatomical landmarks were labeled on the models, and the distances between these landmarks were calculated and fed into the model as features. A multi-layer perceptron (MLP) neural network was trained on this data and applied to predict the treatment duration. The model's performance was evaluated using regression metrics such as mean absolute error (MAE), Pearson's correlation, and coefficient of determination (R-squared: R), to assess predictive accuracy.
Performance metrics of our model revealed a correlation of 0.96, R of 0.91, and a mean absolute error of 3.03 days. The most significant features influencing the predictions were landmarks around the alveolar gap and distances delineating the overall alveolar gap width.
The results suggest that our model can reliably predict the treatment duration required for AMPT in neonates with unilateral CLP with a potential to contribute to developing a fully personalized yet efficient AI-based treatment pipeline.
牙槽成型板治疗（AMPT）在为唇腭裂（CLP）新生儿进行首次重建手术（唇裂修复）做准备方面起着关键作用。然而，在手术前确定将腭裂畸形调整至接近正常所需的AMPT调整次数是一项具有挑战性的任务，这通常会影响治疗持续时间。本研究探索使用机器学习，基于治疗前上颌腭裂畸形的三维（3D）评估来预测治疗持续时间，作为个体化治疗计划的一部分。
从35例单侧唇腭裂婴儿中收集上颌牙弓的数字3D模型。在模型上标记关键解剖标志点，并计算这些标志点之间的距离，将其作为特征输入模型。基于这些数据训练多层感知器（MLP）神经网络，并将其应用于预测治疗持续时间。使用平均绝对误差（MAE）、皮尔逊相关性和决定系数（R平方：R）等回归指标评估模型性能，以评估预测准确性。
我们模型的性能指标显示相关性为0.96，R为0.91，平均绝对误差为3.03天。影响预测的最显著特征是牙槽间隙周围的标志点以及描绘牙槽间隙总宽度的距离。
结果表明，我们的模型能够可靠地预测单侧唇腭裂新生儿AMPT所需的治疗持续时间，有可能为开发一个完全个性化且高效的基于人工智能的治疗流程做出贡献。
