<p>Autism spectrum disease (ASD) is a neurological disease that affects social interaction, communication and behaviour, thus early diagnosis is essential for treatment. Existing detection methods, such as 2D face scans and eye tracking data, are ineffective in capturing difficult facial features and minor asymmetries. Hence, this study presents a novel Convolutional Shape Appearance Model with Randomized Trees-based XGBoost to increase the accuracy and reliability of ASD detection using 3D facial landmark localization. Existing methods fail to capture small facial asymmetries and detailed landmarks essential for accurate diagnosis from 3D facial images, which limits the accuracy of diagnosing ASD-related behaviours. To address this, a novel convolutional neural network (CNN) with an active shape appearance model (ASAM) is introduced, which combines the strengths of CNNs in automatically learning intricate facial features from 3D images and ASAM’s ability to capture variations in facial shapes and textures, thereby enhancing the feature extraction process and accurate localization of facial landmarks. Furthermore, existing detection methods struggle to integrate and analyse facial asymmetry and the complex interactions between multiple facial features simultaneously, leading to reduced accuracy in detecting ASD. Thus, a Euclidian graph with a Randomized Trees<b>-</b>based XGBoost approach is utilized, which captures spatial relationships between facial landmarks and models complex feature interactions, thereby enhancing the accuracy and reliability of ASD detection. The findings demonstrate that the suggested model offers excellent levels of accuracy, precision, recall, detection accuracy, sensitivity and specificity when compared to other existing models.</p>