Predicting Perceived Profile Attractiveness From Cephalometric Measurements Using Machine Learning Authors: Zeynab Pirayesh, Fatemeh Sohrabniya, Seyed AmirHossein Ourang, Hossein Mohammad-Rahimi, Saharnaz Esmaeili, Sarvin Khosravani, Saeed Reza Motamedian, Falk Schwendicke, Antonin Tichy Published: 2026 ABSTRACT ObjectivesOrthodontic treatment planning requires careful consideration of facial aesthetics. This study aimed to compare machine learning (ML) models to predict perceived attractiveness based on lateral cephalometric measurements and to identify parameters that significantly impact perceived attractiveness.MethodsLateral cephalometric radiographs and corresponding facial photographs of 400 patients were utilized in this study. Twenty-six raters, including 20 laypeople and 6 domain experts (art, maxillofacial surgery, and orthodontics), evaluated each photograph using a 1 to 5 Visual Analogue Scale (VAS), and cephalometric analyses using the Dolphin software served as the reference standard. Five ML models, linear regression, Support Vector Machine, XGBoost Regressor, Random Forest, and artificial neural network, were trained to predict attractiveness scores based on the reference standard, and their performance on a test set (n = 59) was evaluated using root mean square error (RMSE). Subsequently, feature importance analysis was conducted to identify cephalometric indices that had the largest impact on VAS scores.ResultsExperts consistently rated facial attractiveness higher than laypeople, especially in age groups >15 years. Furthermore, female raters rated female profiles significantly higher and male profiles significantly lower than male raters, with differences evident in >15 age groups. Among the 5 ML models, Random Forest demonstrated the best performance (RMSE = 0.50). The feature importance analysis revealed that soft tissue parameters, especially those related to facial convexity, maxillary and mandibular prognathism, and vertical proportions, were most predictive of attractiveness scores.ConclusionML models showed promising results in predicting facial attractiveness based on cephalometric measurements and highlighted the correlation of facial aesthetics with soft tissue parameters. The proposed ML-based approach can aid clinicians in the objective assessment of treatment outcomes. Introduction Facial attractiveness has long been a subject of interest in psychology, biology, and aesthetic medicine, with most studies focusing on frontal facial features such as symmetry, averageness, and proportions.1,2 However, the profile view, which plays a crucial role in facial aesthetics, remains comparatively underexplored.3 While research has highlighted the significance of profile characteristics, such as nasal and chin protrusion, lip positioning, and skeletal convexity,4 in determining attractiveness, the majority of aesthetic evaluations have prioritized frontal assessments. This gap underscores the need for further investigation into the profile view, as it provides unique insights into facial harmony and balance, making it essential in aesthetic procedures such as rhinoplasty, lip augmentation, and facial contouring.5,6 In orthodontics, profile view analysis plays a critical role in diagnosis and treatment planning by enabling the identification and correction of structural disharmonies to attain facial balance.7,8 By focusing on profile analysis, clinicians and researchers can gain a more comprehensive understanding of facial attractiveness and design personalized interventions that optimize both functional and aesthetic patient outcomes. When assessing facial attractiveness and influential anatomical features, combining objective measurements with subjective evaluations from laypeople is a well-established approach9 applied in various fields, such as plastic surgery and facial aesthetics.10 Similarly, in orthodontics, integrating cephalometric analysis with subjective assessments is a routine practice for evaluating facial profiles and planning treatments aimed at improving facial harmony.11 Studies have demonstrated that cephalometric indices, such as the nasolabial angle and chin prominence, significantly influence perceived attractiveness, with evaluations from both clinicians and laypeople providing valuable insights into the aesthetic impact of orthodontic interventions.7,8,12, 13, 14 This combined approach not only aids in diagnosing skeletal discrepancies but also facilitates treatment planning that aligns with both functional and aesthetic goals. Artificial intelligence (AI) and its subset, machine learning (ML), have transformed many areas of dentistry by providing robust tools to interpret complicated data and assist in clinical decision-making. ML algorithms that learn from examples without being explicitly programmed have the capability to reveal hidden patterns in large datasets, improve precision, and minimize human errors.15,16 In orthodontics, AI-driven technologies are increasingly used to automate time-consuming tasks such as cephalometric analysis, landmark detection, and treatment outcome predictions. Furthermore, they can reduce the need for human resources and improve accuracy and patient satisfaction simultaneously.17, 18, 19, 20, 21 Previous studies using AI for facial analysis focused primarily on frontal facial features.22,23 However, the integration of ML into profile analysis could provide significant advantages in orthodontic and aesthetic treatment planning. ML could not only predict attractiveness scores based on laypeople’s perceptions17 but also identify important features in cephalograms and photographs,24 offering clinicians valuable insights into which cephalometric indices have the greatest influence on perceived attractiveness. This data-driven approach allows for more precise treatment planning tailored to each patient’s unique profile. Given the challenges in scoring and judging attractiveness, the objectives of the present study were: 1) to develop ML models capable of predicting perceived facial attractiveness based on lateral cephalometric measurements, and 2) to identify the cephalometric indices that have the most significant impact on perceived attractiveness, aiding both orthodontic and aesthetic treatment planning. Methods Study design This study aimed to predict facial attractiveness based on cephalometric measurements using 5 ML models. Lateral cephalometric radiographs and corresponding profile photographs of 400 patients were collected from the archive of the School of Dentistry, Shahid Beheshti University of Medical Sciences (Tehran, Iran). Attractiveness ratings were obtained from 26 raters using a Visual Analogue Scale (VAS), with reliability tests ensuring consistency. All evaluators independently assessed the full set of facial profile photographs. Cephalometric landmark annotation was performed by 2 trained last year dental students using the Dolphin imaging software (Dolphin Imaging and Management Solutions, Chatsworth, CA, USA), with inter- and intra-rater reliability assessments. The annotated data were used for training ML models, with hyperparameter tuning and cross-validation to optimize performance. Model evaluation was conducted using Root Mean Square Error (RMSE), and feature importance analysis identified the cephalometric variables that were most predictive of attractiveness scores. This section provides an overview of the methodology, with each component discussed in detail in the following sections. Data and sampling This study protocol was reviewed and approved by the Research Ethics Committee of the School of Dentistry, Shahid Beheshti University of Medical Sciences [IR.SBMU.DRC.REC.1398.230]. Over 1000 images were reviewed, and a final sample of 400 patients (200 males and 200 females) was selected based on predefined criteria. Inclusion criteria were as follows: participants were included if they had no evident asymmetry between the two sides of the face. For participants older than 12 years, inclusion necessitated the presence of all permanent teeth except the third molars. Participants younger than 12 with missing teeth were excluded. To ensure data completeness, participants were only included if all required cephalometric landmarks and corresponding facial photographs were available and of sufficient quality. Cases with incomplete, poor-quality, or artifact-containing radiographs or photographs were excluded before analysis. Regarding exclusion criteria, patients with systemic diseases, trauma affecting growth, prior orthodontic or facial surgical history, or poor-quality radiographs or profile images were excluded. Additionally, any artifacts in prediction points, such as beards in male participants or tattoos in relevant facial areas, led to exclusion. Data collection continued until the predetermined sample size was reached; therefore, no missing or in-complete data remained in the final dataset. Consequently, no imputation or statistical handling of missing values was required. The dataset was structured to ensure an even distribution across 4 age groups (above 18 years, 15-18 years, 11-14 years, and 7-10 years), with 50 images per gender in each age group. Moreover, images in each group were selected as follows: 25 with Class I occlusion, 10 with half-Class II, 5 with full-Class II, 5 with half-Class III, and 5 with full-Class III. The lateral cephalometric radiographs were taken in the natural head position using 2 imaging systems, PaX-I (Vatech Co, Ltd, Hwaseong-si, Gyeonggi-do, South Korea) and Cranex (Soredex, a division of KaVo Kerr Group, Tuusula, Finland). The imaging setup maintained a 1.5 m distance between the X-ray tube and the patient, ensuring perpendicular alignment, with the film placed 30 cm from the patient’s head. Exposure parameters varied by age group, ranging from 80 to 90 kVp with an exposure time of 6-18 s. Facial profile photos were also captured in the natural head position from a distance of 1.5 m using a digital SLR camera (Nikon D750 Body only, Tokyo, Japan). To ensure clarity of facial features, subjects refrained from wearing eyeglasses, and no hair covered the face. The images were then processed using Adobe Photoshop CC 2020 (Adobe, San Jose, CA, USA) and edited to encompass the region from 1 cm above the glabella to the front of the ear and 1 cm below the menton, to prepare them for assessment. The photographs were desaturated to remove colour and then adjusted for skin tone darkness to match a standard reference image, aiming to minimize any potential evaluative bias. To soften the skin texture, the 'Layer Mask' feature and 'Gaussian Blur' filter in the image editing software were used. Remaining imperfections, such as scars or skin roughness, were further smoothed out utilizing the 'Paint Bucket' and 'Healing Brush' tools. In the final step of anonymization, the subjects' eyes and eyebrows were obscured with a black rectangle in the profile images. The boundaries of this rectangle were defined by the eyebrow as the upper edge, the lower eyelid as the lower edge, and the medial and lateral canthi of the eye as the lateral edges. Reference standard Facial profile photographs scoring A group of 26 examiners, comprising an equal number of males and females ranging from 20 to 40 years old, assessed the attractiveness of the facial profiles. To ensure diversity in perceptions of facial attractiveness, the panel included both expert and non-expert raters. Six participants were considered experts: 2 orthodontists, 2 oral and maxillofacial surgeons, and 2 visual arts specialists. The remaining 20 raters had no formal education or professional background in aesthetics. Initially, all participants completed a consent form. Then, facial photographs were evaluated using VAS, a tool commonly employed to measure subjective qualities. This scale features a horizontal line marked from 1 (signifying very unpleasant) to 5 (signifying extremely pleasant) at its endpoints. Facial images were displayed on a 15.6-inch monitor within a dimly lit environment to maintain consistency in visual perception. The presentation of photographs was randomized, and adjustments were made to ensure the size of the images was appropriate for the age group being represented. This approach was designed to eliminate any potential bias that might arise from the variation in image size relative to the actual size of the faces or the sequence in which the photographs were viewed. Each rater was allotted 20 s to evaluate an image and took a 15-minute break after every 50 assessments. The entire process was concluded over a span of 2 days. The final score for each photograph was calculated based on the average of the scores provided by the raters. To evaluate the reliability of the raters, ten of them were selected to reassess a random set of 25 images after a 2-week interval. The intra-class correlation coefficients (ICCs) varied between 0.73 and 0.98 across the 25 photographs. Paired-samples t-tests were conducted to compare the evaluations of laypeople and expert raters within each age group. Additionally, a two-way repeated measures ANOVA was employed to examine the effect of rater gender on scores across male and female profiles. To further assess the robustness of the model across demographic subgroups, independent-samples t-tests with Welch’s correction were performed separately for each age group to compare mean RMSE prediction errors between male and female subjects. Lateral cephalometric radiographs The Dolphin imaging software was utilized to delineate cephalometric landmarks on the provided data. The data were independently annotated by 2 trained senior dental students, who performed 80 predefined cephalometric measurements of both hard and soft tissues in each radiograph. Any discrepancies in landmark detection were resolved through consultation with an experienced orthodontist, and annotations were subsequently reviewed and verified by another experienced orthodontist. The analyses incorporated in this study included various established orthodontic assessments such as Legan and Burstone, Holdaway, Ricketts, Z angle, Epker, Steiner, McNamara, Downs, Tweed, Wits, Jarabak, Bjork, Moreno, Farkas, Powell N, Wisth PJ, and SJ Chakonas (please refer to Figure 1 and Table S1 for details). All cephalometric measurements were standardized using z-score normalization prior to model training. Further details regarding the linear and angular measurements for both hard and soft tissues, as well as the specific analyses they were included in, can be found in the publication by Esmaeil et al.25Fig. 1Lateral cephalogram landmarks annotated using the Dolphin software.Fig 1 dummy alt text Lateral cephalogram landmarks annotated using the Dolphin software. To determine the annotation reliability, the 2 students re-analysed a set of 25 lateral cephalograms after a 2-week interval. The mean ICC for inter-examiner and intra-examiner reliability was 0.904 and 0.905, respectively. The reliability of the cephalometric analysis was assessed in a prior work by the authors, and it was reported that the Dolphin imaging software is a reliable tool for most measurements.26 Data partitions To reduce sampling bias and improve model generalizability, a stratified sampling approach was adopted during test set selection. Stratified sampling is known to preserve the underlying distribution of target variables, particularly useful in datasets with subjective outcomes such as aesthetic scores, by ensuring proportional representation of different subgroups in training and evaluation sets.27 To define strata, the mean and standard deviation (SD) of the average VAS scores were calculated. Based on this distribution, the data were categorized into 3 groups:1.Rank 1: scores > mean +0.4 SD2.Rank 2: scores within mean ±0.4 SD3.Rank 3: scores < mean −0.4 SD 1.Rank 1: scores > mean +0.4 SD Rank 1: scores > mean +0.4 SD 2.Rank 2: scores within mean ±0.4 SD Rank 2: scores within mean ±0.4 SD 3.Rank 3: scores < mean −0.4 SD Rank 3: scores < mean −0.4 SD This resulted in 134 samples in Rank 1, 124 in Rank 2, and 142 in Rank 3. From each group, 15% of the samples were randomly selected to form a stratified test set (n = 59 total; Rank 1: 20, Rank 2: 18, Rank 3: 21). The remaining 341 samples were used in the model development process. For model evaluation, 5-fold cross-validation was applied using the full remaining dataset (n = 341), without stratification by rank. In each fold, 20% of the data (n ≈ 68) was used for validation, while the remaining 80% (n ≈ 273) served as the training set. This process was repeated 5 times to ensure comprehensive exposure of the model to all data during both training and validation. Model Initially, the following models were trained for performance comparison: linear regression, Support Vector Machine (SVM), XGBoost Regressor, Random Forest (RF), and Artificial Neural Network (ANN). The ANN consisted of 2 hidden layers, ReLU activations, a single linear output neuron, and the Adam optimizer. The number of training epochs was 200. Following an initial experiment, the RF model was identified as the most effective algorithm and selected for further analysis. After selecting the RF model, feature importance scores were then evaluated to assess each feature's contribution to model performance. To determine the optimal number of features for accurate prediction, a stepwise feature addition approach was implemented. Beginning with the single most important feature, features were incrementally added in order of decreasing importance. After each addition, the model was retrained, and the RMSE was recorded. This process continued until all features had been incorporated. Training All ML models were implemented using the Python programming language and the sci-kit learn library. Hyperparameter tuning was done using randomized search across the following parameter ranges:1)Node size or the maximum number of leaf nodes (set from 2 to 32);2)Number of trees or estimators in the forest (set from 3 to 500);3)Number of features used for the best split (set from 5 to 13);4)Using bootstrap samples in the trees building (True or False). 1)Node size or the maximum number of leaf nodes (set from 2 to 32); Node size or the maximum number of leaf nodes (set from 2 to 32); 2)Number of trees or estimators in the forest (set from 3 to 500); Number of trees or estimators in the forest (set from 3 to 500); 3)Number of features used for the best split (set from 5 to 13); Number of features used for the best split (set from 5 to 13); 4)Using bootstrap samples in the trees building (True or False). Using bootstrap samples in the trees building (True or False). A total of 1000 random combinations of these hyperparameters were evaluated on the validation sets. Given the training process with cross-validation, each sample was once employed in the validation set, providing a more robust evaluation of model performance. After determining the optimal hyperparameters, the final models were retrained on the entire training set (n = 341). Then, model performance was assessed on unseen data, ie, the mentioned holdout test set. The final tuned hyperparameters for all trained models, including Linear Regression, SVM, RF, XGBoost, and the ANN, are fully summarized in Supplementary Table 2. Evaluation The performance of the final models on the test set was compared using RMSE (Eq. 1).(1)RMSE=1∑i=1n(Ai−Pi)2n,where Ai represents the actual value, Pi represents the predicted value, and n is the total number of samples. Feature selection was performed to find the most relevant cephalometric measurements contributing to the predictions, ie features most strongly associated with the VAS scores. This was achieved using Gini impurity (Eq. 2) and Mean Decrease in Impurity (MDI), a standard approach for evaluating feature importance in tree-based models. The top 55 factors were used to evaluate the influence of the number of features on model performance. Specifically, the best-performing model was repeatedly trained using an incremental subset of the top-ranked features (top 2, top 3, top 4, and so on, up to top 55).(2)G(S)=∑k=1cPk(1−Pk) Each model was trained 5 times on various seeds, and the average RMSE was reported for each training. The performance was reported separately for 8 standard cephalometric analyses (Legan and Bustone, Holdaway, Epker, Steiner, McNamara, Downs, Bjork, Jarabak). Figure 2 illustrates the overall workflow of the proposed method.Fig. 2Flowchart summarizing the study workflow.Fig 2 dummy alt text Flowchart summarizing the study workflow. Results Facial attractiveness scores The distribution of the mean VAS score was approximately normal (Figure 3). Table 1 summarizes attractiveness scores for each age group of subjects, separated by gender. Overall, laypeople and expert raters gave comparable scores for younger subjects with no significant difference. Only a slight, non-significant trend was observed in the early adolescent group (11-14 years), but clear differences emerged in older age groups. Experts’ scores were significantly higher than laypersons’ scores for the 15 to 18 year-old adolescents (P < .001) and for adults (> 18 years; P < .001).Fig. 3Histogram of mean VAS scores. The distribution of the data was normal.Fig 3 dummy alt textTable 1Summary of attractiveness scores for male and female profile photographs across age groups presented as mean ± standard deviation (SD) and median (Q1-Q3).Table 1 dummy alt textAge group Mean±SDMedian (Q1; Q3)Gender7-102.52 ± 0.592.57 (2.13; 2.99)Females2.85 ± 0.582.85 (2.67; 3.39)Males11-142.79 ± 0.592.88 (2.38; 3.45)Females2.89 ± 0.572.97 (2.57; 3.46)Males15-182.70 ± 0.532.75 (2.43; 3.25)Females2.59 ± 0.482.66 (2.43; 3.00)Males>182.83 ± 0.642.84 (2.52;3.32)Females2.66 ± 0.572.74 (2.28; 3.25)Males Histogram of mean VAS scores. The distribution of the data was normal. Table 1Summary of attractiveness scores for male and female profile photographs across age groups presented as mean ± standard deviation (SD) and median (Q1-Q3).Table 1 dummy alt textAge group Mean±SDMedian (Q1; Q3)Gender7-102.52 ± 0.592.57 (2.13; 2.99)Females2.85 ± 0.582.85 (2.67; 3.39)Males11-142.79 ± 0.592.88 (2.38; 3.45)Females2.89 ± 0.572.97 (2.57; 3.46)Males15-182.70 ± 0.532.75 (2.43; 3.25)Females2.59 ± 0.482.66 (2.43; 3.00)Males>182.83 ± 0.642.84 (2.52;3.32)Females2.66 ± 0.572.74 (2.28; 3.25)Males Summary of attractiveness scores for male and female profile photographs across age groups presented as mean ± standard deviation (SD) and median (Q1-Q3). Female profiles No significant differences were observed between lay and expert raters in the age groups below 14 years (P = .16 and 0.52). However, in the 15 to 18 and >18 age groups, experts consistently assigned significantly higher attractiveness scores than laypeople (P < .001). Significant differences were observed between female and male raters in the 11 to 14 (P = .040), 15 to 18 (P = .017), and >18 (P = .012) age groups. Overall, female raters assigned higher attractiveness scores to female profiles compared to male raters (Table 2).Table 2VAS scores for female profiles across age groups stratified by rater type and gender.Table 2 dummy alt textAge groupOverall scoreLayExpertFemaleMale7-102.52 ± 0.592.52 ± 0.602.51 ± 0.642.53 ± 0.662.50 ± 0.5411-142.79 ± 0.592.77 ± 0.592.83 ± 0.682.82 ± 0.64‡2.75 ± 0.5715-182.70 ± 0.532.65 ± 0.542.86 ± 0.58*2.73 ± 0.56‡2.66 ± 0.52>182.83 ± 0.642.76 ± 0.643.05 ± 0.69*2.88 ± 0.73‡2.77 ± 0.57⁎Significant difference between expert and lay raters (P < .001 for >15).‡Significant difference between female and male raters (P = .40 for 11-14, P = .017 for 15-18 and P = .012 for >18).Presented as mean ± SD. Table 2VAS scores for female profiles across age groups stratified by rater type and gender.Table 2 dummy alt textAge groupOverall scoreLayExpertFemaleMale7-102.52 ± 0.592.52 ± 0.602.51 ± 0.642.53 ± 0.662.50 ± 0.5411-142.79 ± 0.592.77 ± 0.592.83 ± 0.682.82 ± 0.64‡2.75 ± 0.5715-182.70 ± 0.532.65 ± 0.542.86 ± 0.58*2.73 ± 0.56‡2.66 ± 0.52>182.83 ± 0.642.76 ± 0.643.05 ± 0.69*2.88 ± 0.73‡2.77 ± 0.57⁎Significant difference between expert and lay raters (P < .001 for >15).‡Significant difference between female and male raters (P = .40 for 11-14, P = .017 for 15-18 and P = .012 for >18).Presented as mean ± SD. VAS scores for female profiles across age groups stratified by rater type and gender. Significant difference between expert and lay raters (P < .001 for >15). Significant difference between female and male raters (P = .40 for 11-14, P = .017 for 15-18 and P = .012 for >18). Presented as mean ± SD. Male profiles The results of the paired t-test showed that expert raters assigned significantly higher attractiveness scores than laypeople in all groups (P < .001), except for the 11 to 14 age group (P = .191). Additionally, female raters assigned significantly lower scores to male profiles compared to male raters in both the 15 to 18 (P = .004) and >18 (P < .001) age groups (Table 3).Table 3VAS scores for male profiles across age groups stratified by rater type and gender.Table 3 dummy alt textAge groupOverall scoreLayExpertFemaleMale7-102.84 ± 0.582.86 ± 0.572.79 ± 0.71*2.87 ± 0.662.82 ± 0.5311-142.89 ± 0.572.88 ± 0.562.91 ± 0.692.92 ± 0.642.85 ± 0.5215-182.58 ± 0.482.54 ± 0.482.71 ± 0.57*2.52 ± 0.512.64 ± 0.49‡>182.66 ± 0.572.61 ± 0.552.83 ± 0.73*2.57 ± 0.642.75 ± 0.52‡⁎Significant difference between expert and lay raters (P < .001 for 7-10 and >15).‡Significant difference between female and male raters (P =0.004 for 15-18 and p<0.001 for >18).Presented as mean ± SD. Table 3VAS scores for male profiles across age groups stratified by rater type and gender.Table 3 dummy alt textAge groupOverall scoreLayExpertFemaleMale7-102.84 ± 0.582.86 ± 0.572.79 ± 0.71*2.87 ± 0.662.82 ± 0.5311-142.89 ± 0.572.88 ± 0.562.91 ± 0.692.92 ± 0.642.85 ± 0.5215-182.58 ± 0.482.54 ± 0.482.71 ± 0.57*2.52 ± 0.512.64 ± 0.49‡>182.66 ± 0.572.61 ± 0.552.83 ± 0.73*2.57 ± 0.642.75 ± 0.52‡⁎Significant difference between expert and lay raters (P < .001 for 7-10 and >15).‡Significant difference between female and male raters (P =0.004 for 15-18 and p<0.001 for >18).Presented as mean ± SD. VAS scores for male profiles across age groups stratified by rater type and gender. Significant difference between expert and lay raters (P < .001 for 7-10 and >15). Significant difference between female and male raters (P =0.004 for 15-18 and p<0.001 for >18). Presented as mean ± SD. Model performance On the test set, RF exhibited the lowest RMSE (0.50 ± 0.02), slightly outperforming SVM (0.51 ± 0.03) and other ML models (XGBoost Regressor 0.54 ± 0.03, linear regression 0.60 ± 0.03, ANN 0.64 ± 0.03). In addition, the comparison of mean absolute prediction errors between male and female subjects revealed no statistically significant differences (Welch’s t-test, P = .99), indicating that model performance was consistent across genders and did not vary systematically with age group. When important features were added one by one, as illustrated in Figure 4, RMSE did not significantly improve beyond the inclusion of the top 13 features. This suggests that the model can achieve strong predictive performance using only the top 13 features, thereby reducing complexity without sacrificing accuracy. So, RF performed well when the following 13 features were included in the model: G-Sn-Pg’ angle, G-Sn, G-Pg’, G-Sn/Sn-Me’, Sn-Gn’-C, Sn-Gn’/C-Gn’, Cm-Sn-Ls, Ls to Sn-Pg’, Li to Sn-Pg’, Si to Li-Pog’, Sn-Stms/Stms-Me’, Stms-Ui, Stms-Stmi which represented the facial convexity (contour) angle (°), the maxillary prognathism (mm), the mandibular prognathism (mm), the vertical height ratio, the lower facial throat angle (°), the lower vertical height–depth ratio, the nasolabial angle (°), the upper lip protrusion (mm), the lower lip protrusion (mm), the mentolabial sulcus depth (mm), the vertical lip chin ratio, the maxillary incisor exposure (mm), and the inter-labial gap (Figure 5).Fig. 4Root mean square error (RMSE) of the random forest model as a function of the number of features used. Features were added sequentially based on their importance ranking. After the inclusion of the top 13 features, further additions resulted in minimal improvement in RMSE, indicating diminishing returns beyond this point.Fig 4 dummy alt textFig. 5Feature importance. It was revealed that A-B to facial plane angle, facial convexity angle, distance from subnasale to vermilion, Wits appraisal, upper centralincisor protrusion, ANB angle, IMPA angle, Sn-Go-Gn angle and distance from subnasale to Soft Tissue Pogonion were the most important determinant factors in facial attractiveness and its score.Fig 5 dummy alt text Root mean square error (RMSE) of the random forest model as a function of the number of features used. Features were added sequentially based on their importance ranking. After the inclusion of the top 13 features, further additions resulted in minimal improvement in RMSE, indicating diminishing returns beyond this point. Feature importance. It was revealed that A-B to facial plane angle, facial convexity angle, distance from subnasale to vermilion, Wits appraisal, upper centralincisor protrusion, ANB angle, IMPA angle, Sn-Go-Gn angle and distance from subnasale to Soft Tissue Pogonion were the most important determinant factors in facial attractiveness and its score. Feature selection and model fitting performance To identify the most relevant cephalometric features, RMSE values were analysed across different sets of variables using the RF model (Table 4). The soft tissue analysis suggested by Legan and Burston, which had the lowest RMSE of the analyses listed, outperformed other hard tissue analyses in predicting subjective attractiveness. Figure 5 shows the most correlated factors ranked based on their relevance to the VAS scores.Table 4RMSE of RF models on various seeds and analyses.Table 4 dummy alt textAnalysisNumber of measurementsSeed = 15 (RMSE)Seed = 25 (RMSE)Seed = 35 (RMSE)Seed = 45 (RMSE)Seed = 55 (RMSE)Mean (RMSE) ± SDLegan and Bustone130.510.510.510.510.510.51 ± 0.0018Holdaway110.520.530.530.530.530.53 ± 0.0031Epker40.530.530.530.530.530.53 ± 0.0033Steiner100.530.540.520.520.520.53 ± 0.0090McNamara80.520.520.530.530.530.53 ± 0.0016Downs60.540.550.540.550.550.55 ± 0.0076Bjork50.560.560.560.550.560.56 ± 0.0039Jarabak80.550.550.550.550.560.55 ± 0.0021ALL810.500.490.500.510.510.50 ± 0.0080 Table 4RMSE of RF models on various seeds and analyses.Table 4 dummy alt textAnalysisNumber of measurementsSeed = 15 (RMSE)Seed = 25 (RMSE)Seed = 35 (RMSE)Seed = 45 (RMSE)Seed = 55 (RMSE)Mean (RMSE) ± SDLegan and Bustone130.510.510.510.510.510.51 ± 0.0018Holdaway110.520.530.530.530.530.53 ± 0.0031Epker40.530.530.530.530.530.53 ± 0.0033Steiner100.530.540.520.520.520.53 ± 0.0090McNamara80.520.520.530.530.530.53 ± 0.0016Downs60.540.550.540.550.550.55 ± 0.0076Bjork50.560.560.560.550.560.56 ± 0.0039Jarabak80.550.550.550.550.560.55 ± 0.0021ALL810.500.490.500.510.510.50 ± 0.0080 RMSE of RF models on various seeds and analyses. Discussion ML has become an integral tool in modern orthodontics, which offers significant advancements in diagnosis, treatment planning, and predictive analytics. Given that aesthetics are a crucial aspect of orthodontic care,28 with facial attractiveness influencing self-confidence and social interactions,29 there is increasing interest in using ML models to predict attractiveness based on cephalometric data, helping to achieve a balance between dental, skeletal, and soft tissue features for optimal facial proportions. The present study focused on traditional machine learning algorithms, linear regression, SVM, RF, XGBoost, and an ANN, because the dataset consisted of structured, tabular cephalometric variables rather than raw images. For this type of numerical data and moderate sample size (n = 400), classical models typically outperform deep learning architectures in both accuracy and reliability, while requiring fewer parameters and reducing overfitting risk.17,19,30 Moreover, these approaches enable feature importance analysis and allow clearer interpretation of the anatomical determinants of attractiveness, which is essential for clinical translation. Nevertheless, future studies could expand this framework by incorporating deep learning or multi-modal models that jointly analyse cephalometric data and facial photographs to capture more complex nonlinear relationships and improve generalizability across diverse populations.31 Among the tested models, RF demonstrated the highest predictive accuracy, achieving an RMSE of approximately 0.5 on the 5-point attractiveness scale. This error magnitude indicates that the model’s predictions differed from the raters’ mean scores by about half a point, which is comparable to the natural variability in human ratings (SD ≈ 0.5-0.6, Table 1, Table 2, Table 3). In other words, model’s deviation reflects the expected fluctuation in aesthetic judgments but not to an extent that would alter the overall ranking or perceived level of attractiveness. While this degree of variation may not allow for a strict cutoff between 'attractive' and 'unattractive' categories, it is sufficiently small to capture consistent trends across profiles and to provide clinically meaningful, reproducible estimates of perceived attractiveness. Therefore, the achieved accuracy can be regarded as acceptable for decision-support and comparative evaluations, complementing, rather than replacing, expert assessment. Feature importance analysis confirmed that soft tissue factors, particularly facial convexity, lip protrusion, and vertical proportions were the most influential predictors of perceived attractiveness.32 Although facial attractiveness is strongly influenced by soft-tissue features,13,32 the ability to approximate perceived attractiveness from cephalometric measurements offers practical value in everyday orthodontic planning. By providing an objective estimate of how soft- and hard-tissue relationships may be perceived aesthetically, the model can support clinicians in communicating treatment expectations, identifying profile imbalances early, and prioritizing interventions that most influence appearance. While such tools may enhance treatment planning by adding a reproducible, data-driven perspective to aesthetic assessments that are traditionally subjective, clinical applicability of the present model is limited because by the use of manual cephalometric measurements derived from the Dolphin software. Although they were proven reliable,26 it is a time-consuming task, hence it would be beneficial if the detection of cephalometric landmarks was automated, provided that the landmark detection model performs accurately as errors would directly affects the results of model training and attractiveness prediction. Beyond prediction, ML models also offer valuable insights into why certain faces are rated as more attractive by highlighting which features most influence model outputs. By analysing feature importance, we can better understand the specific cephalometric measurements that drive attractiveness perceptions. In this study, feature importance analysis revealed that soft tissue factors, particularly facial convexity, lip protrusion, and vertical proportions (Figure 4), played the most significant role in attractiveness perception. This aligns with established aesthetic principles emphasizing balanced soft tissue proportions over hard tissue structures.33,34 Parul et al,35 who analysed frontal and lateral photographs, also identified lip prominence, nasolabial angle, and mentolabial sulcus depth as key factors. However, their study relied on a small dataset (18 subjects), limiting its generalizability, as opposed to this study with 400 subjects. Previous studies focusing on profile analyses have demonstrated correlations between facial attractiveness and cephalometric indices, such as the inter-incisal angle,36 profile angle,37 nasolabial angle,28 facial convexity,37 and chin prominence.38,39 These features collectively influence the perception of a harmonious facial profile, which aligns with our findings showing that certain cephalometric variables significantly contribute to attractiveness as rated by laypeople. In a study by Yu et al,40 facial attractiveness was evaluated using an ML technique and superimposition algorithms. The authors chose orthodontists as raters, meaning that facial attractiveness was evaluated primarily based on professional ideals rather than general public perception. In contrast, our study incorporated ratings from both experts and lay people to ensure broader representativeness and to capture real-world variability and diversity in aesthetic judgment. Moreover, unlike Yu et al,40 we evaluated multiple ML algorithms to select the best-fitting predictive model. In another study,41 a harmony score after orthodontic treatment was predicted, along with analysing the cephalometric factors that influence facial harmony. While traditional studies often assume linear relationships between such indices and attractiveness, Wang et al suggested that these relationships might be non-linear.41 ML methods offer a valuable approach to capture non-linear relationships, providing a deeper understanding of the complex interplay between cephalometric measurements and aesthetic perceptions.30 In the study by Wang et al, the XGBoost model achieved a lower RMSE (0.34) than our RF model (0.50), despite using a smaller dataset of 180 cases.41 Given that their data was based on expert ratings after orthodontic treatment, this difference is likely due to the greater variability in layperson ratings and pre-treatment facial attractiveness used in our study. Both studies used feature importance analysis, but the most predictive features differed: Wang et al emphasized skeletal and dental factors (eg, ANB, L1/NB),41 while our findings highlighted soft tissue characteristics such as facial convexity and lip prominence, aligning more closely with established aesthetic principles.42 This could suggest that despite higher prediction variability, our approach may offer broader clinical relevance by targeting public perception of attractiveness in the diagnostic stage. One of the primary challenges in this study was the subjectivity of attractiveness scoring. Evaluating facial attractiveness is inherently subjective, with perceptions varying significantly among individuals. Research indicates that different people prefer distinguishable regions of the face-space, highlighting the essential subjectivity of the phenomenon.43 Additionally, studies have shown that self-perceived attractiveness often differs from external evaluations, suggesting that personal biases influence one's assessment of their own attractiveness.44 This variability introduces challenges in creating a standardized model for attractiveness, as the lack of consensus among raters can lead to inconsistencies in the data used for model training. Specifically, our results showed that expert raters consistently assigned higher attractiveness scores than laypeople in the 15 to 18 and >18 age groups, while no significant differences were observed in younger groups. This pattern aligns with prior research suggesting that laypeople and professionals may evaluate facial aesthetics differently, particularly as secondary sexual characteristics and structural facial features become more pronounced with age.45,46 In younger age groups, facial features are generally softer and less differentiated, which may contribute to greater agreement among raters regardless of expertise. Experts are trained to assess facial proportions, symmetry, and skeletal harmony, which may make them more attuned to features not as readily noticed by lay observers.47,48 In contrast, lay raters may prioritize overall appearance or emotional expression, leading to assessments that are less consistent with clinical standards.49,50 Gender-based differences were also observed in our data. Female raters consistently assigned lower attractiveness scores to male profiles in the 15 to 18 and >18 age groups compared to male raters. This suggests a more critical evaluation of male faces by female observers, potentially influenced by cultural or psychological factors.51 Our finding is in agreement with previous studies that reported similar gender-based discrepancies in aesthetic judgment,52 and it echoes broader trends in the literature, where female observers tend to demonstrate greater selectivity and sharper contrasts when evaluating male faces, an effect not as pronounced among male raters. Taken together, these findings underscore that both professional expertise and gender of raters can significantly influence aesthetic evaluations and should be carefully considered in both clinical assessments and ML model trained on subjective ratings. However, despite these differences in scoring patterns, Welch’s test showed no significant difference in RMSE between male and female subjects. This suggests that although the ratings may reflect observer-related preferences, these did not translate into sex-based differences in model prediction error. Another limitation arises from the image anonymization and standardization process. To protect patient privacy, we obscured the eye region in all images and converted them to grayscale. However, features such as skin tone and the eye area significantly influence attractiveness perceptions. Research indicates that skin tone impacts attractiveness judgments, and the eye area, including under-eye skin and crow's feet, affects perceived age and attractiveness.44 By standardizing images in this manner, we may have inadvertently excluded critical information pertinent to attractiveness assessments. Additionally, while using black and white photos may reduce the effect of skin tone, facial features and attractiveness perceptions are likely to vary between races. Differences in how profiles are perceived between races necessitate testing of the implemented models on external data from a different region.53 This would also help test the generalizability of the model, as it was trained on a dataset from a single academic institution representing an Iranian population. While this ensured methodological consistency in imaging and analysis, it also implies that the model may reflect region-specific patterns rather than universally applicable ones.54,55 Future investigations using multi-centre or multi-ethnic data are essential to confirm the model’s robustness and broaden its clinical relevance. While the RF model demonstrated the highest predictive accuracy in this study, its ensemble nature can obscure the rationale behind individual predictions. Unlike linear models, where coefficients directly indicate feature influence,56,57 RF aggregates numerous decision trees, making case-level interpretation less intuitive. Nonetheless, it offers partial transparency through global feature importance metrics and partial dependence plots, which help reveal the relative influence of cephalometric variables on attractiveness predictions.24,36 Other methods of explainable AI, such as SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations), allow visualization of feature contributions at both global and individual levels.58 Incorporating these tools in future studies could bridge the gap between predictive performance and clinical interpretability, enhancing confidence in AI-assisted orthodontic assessment and decision-making. To improve the predictive accuracy of attractiveness models, future studies could explore multi-modal approaches, incorporating 3D imaging, frontal facial analysis, and demographic data alongside cephalometric measurements. Additionally, integrating deep learning architectures with explainable AI techniques could help identify nonlinear relationships between skeletal structure, soft tissue features, and subjective attractiveness scores. By addressing these challenges, future research can contribute to the development of more robust and clinically useful AI models for facial aesthetics assessment. Conclusion This study demonstrated that ML models can successfully predict perceived facial profile attractiveness from cephalometric measurements. Consequently, the traditionally subjective appraisal of profile aesthetics can be translated into quantifiable metrics, effectively bridging the gap between aesthetic perception and objective analysis. Furthermore, the feature importance analysis revealed that soft tissue–related cephalometric parameters were the most influential predictors of attractiveness scores. In particular, measurements reflecting facial convexity and vertical facial proportions emerged as critical determinants of an attractive profile. In practice, the presented ML-based approach can serve as a standardized and data-driven tool to aid clinicians in assessing and planning treatment outcomes related to facial profile aesthetics. Author contribution Conceptualization: Saeed Reza Motamedian, Zeynab Pirayesh, Fatemeh Sohrabniya, Hossein Mohammad-Rahimi. Data acquisition & curation: Zeynab Pirayesh, Saharnaz Esmaeili, Sarvin Khosravani. Methodology & analysis: Zeynab Pirayesh, Fatemeh Sohrabniya, Seyed AmirHossein Ourang, Hossein Mohammad-Rahimi. Visualization & feature importance analysis: Zeynab Pirayesh, Fatemeh Sohrabniya. Writing – original draft: Zeynab Pirayesh, Fatemeh Sohrabniya, Seyed AmirHossein Ourang. Writing – review & editing: Zeynab Pirayesh, Fatemeh Sohrabniya, Seyed AmirHossein Ourang, Hossein Mohammad-Rahimi, Antonin Tichy, Saeed Reza Motamedian, Falk Schwendicke. Supervision & project administration: Saeed Reza Motamedian, Falk Schwendicke, Antonin Tichy. Accountability statement Each author confirms they have seen and approved the submitted version of the manuscript and accept responsibility for its integrity and accuracy. Conflict of interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Falk Schwendicke is cofounder of dentalxrai, a company focusing on AI-based analytics in dentistry. The current work was fully independent of this activity. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.