[Construction of a risk prediction model for moderate to severe orthodontic-induced inflammatory root resorption of maxillary incisors based on cone beam CT radiomics and clinical features].

Abstract

Objective: To develop a risk prediction model for moderate to severe orthodontic-induced inflammatory root resorption (OIIRR) of maxillary incisors based on cone beam CT (CBCT) radiomics features and clinical characteristics of the orthodontic patients. Methods: Clinical and CBCT data from 101 orthodontic patients treated by the same attending orthodontist in the Department of Orthodontics, Stomatology Hospital of Tianjin Medical University from January 2019 to January 2024 were retrospectively collected. The sample included 42 class Ⅰ patients, 52 class Ⅱ patients and 7 class Ⅲ patients [age: (19.7±6.3) years], and a total of 394 maxillary incisors were analyzed. Potential influencing factors for moderate to severe OIIRR (root volume resorption rate≥10%) were collected from the patients' CBCT and medical records, including initial age, gender, treatment duration, Angle's classification, extraction or not, type of orthodontic appliance (fixed or clear aligner), changes in root inclination, root movement distance and direction, pre-treatment cephalometric measurements, pre-treatment root-bone relationship, pre-treatment root length, and pre-treatment radiomics features of the teeth. Univariate analysis was initially performed to screen for factors influencing moderate to severe OIIRR. Subsequently, least absolute shrinkage and selection operator (LASSO) regression, best subset regression, and random forest were used for feature selection to construct the OIIRR risk prediction model. The discrimination, calibration, and net benefit of the three risk prediction models were evaluated, and the optimal model was displayed using a nomogram. Results: LASSO regression identified clinical features including initial age (LASSO coefficient 0.052), treatment duration (LASSO coefficient 0.024), pre-treatment root length (LASSO coefficient -0.023), and vertical root movement distance (LASSO coefficient -0.029). Initial age and treatment duration were positively correlated with the severity of OIIRR, while root length and vertical root movement distance were negatively correlated. A total of 14 radiomics features were identified, including 2 original image features and 12 wavelet features. Best subset regression identified vertical root movement distance as the clinical feature and 7 radiomics features, including 1 original image feature and 6 wavelet features. The random forest model identified 8 wavelet features as important predictors, and all of which were radiomics features. Model performance evaluation showed that the random forest model had the highest discrimination, calibration, and net benefit, making it the optimal model, with radiomics features being the most important predictors. Conclusions: Based on the data from this study, radiomics features were identified as the most important predictors by the optimal model for OIIRR risk prediction. Predicting the occurrence of moderate to severe OIIRR before orthodontic treatment held potential clinical application value.