Accuracy of Artificial Intelligence in Predicting the Treatment Effects of Headgear and/or Functional Appliance on the Maxillo-Mandibular Growth in Preadolescent Patients With Skeletal Class II Malocclusion.

Abstract

Objective: To evaluate the accuracy of artificial intelligence (AI) in predicting the effects of headgear (HG) and/or functional appliance (FA) on the maxillo-mandibular growth in preadolescent patients with skeletal Class II (C-II) malocclusion.

Materials and methods: The study included 206 Japanese preadolescent C-II patients treated with HG and/or FA, with serial lateral cephalograms taken at ages of 8 (T0) and 10 (T1). A single orthodontist with 7 years of experience identified 28 hard-tissue cephalometric landmarks. A Treatment Prediction Graph Convolutional Neural Network (TP-GCNN), integrating a high-resolution network and a graph neural network, was trained and validated using the landmarks' x- and y-coordinates. Data was split into training, validation and testing sets (ratio of 8:1:1; n = 164, n = 21 and n = 21). Model performance was assessed using the values of prediction error (PE, excellent, ≤ 0.5 mm; very good, 0.5-1.0 mm; good, 1.0-1.5 mm; acceptable, 1.5-2.0 mm; unsatisfactory, > 2.0 mm) and the degree of accurate prediction percentage (APP; very high, ≥ 90%; high, 70%-90%; medium, 50%-70%; low, < 50%).

Results: The mean PE value was 1.45 mm. In terms of PE, all landmarks showed the accuracy above the 'acceptable' category. In terms of APP, 'High' APP was observed at Hinge axis, Pterygoid point, A-point, PNS, ANS, R1, R3 and Articulare. However, 'Low' APP was noted for Pm, Pogonion, B-point and Menton. The remaining landmarks demonstrated 'Medium' APP.

Conclusion: This study demonstrates the potential of AI to reliably predict the effects of HG and/or FA treatment in preadolescent patients with Class II malocclusion.