Subthalamic nucleus deep brain stimulation (STN-DBS) is a successful treatment for patients with Parkinson’s disease but carries a risk for postoperative delirium. The study investigates radiomics as a predictive tool for identifying patients at risk and highlights the importance of developing a thalamic-hippocampal-amygdala network. Using magnetic resonance images (MRIs), machine learning and deep learning were applied for image segmentation of the amygdala, hippocampus, and thalamus regions, feature extraction and selection, and the development of predictive models. Eight different predictive models were trained, finding that regularized binary logistic regression and linear discriminant analysis using five and ten feature sets produced the highest precision. When increased to 20 radiomic features, the feed-forward neural network reached the highest predictive accuracy (99.28%), sensitivity (100.0%), specificity (98.57%), and area under the receiver operating characteristic curve (0.99) across all models. Additionally, the thalamic laterodorsal nucleus, reuniens nucleus, central medial nucleus, amygdalar basal and paralaminar nucleus, and hippocampal cornu ammonis were highlighted as key features to analyze in patients at risk of postoperative delirium due to their modifications in neural activity. While the study introduced promising predictive models for postoperative delirium risk in patients with Parkinson’s disease, there are limitations of the model having been developed across a limited data set, potential for feature selection bias, and the lack of external validation.