Prediction of Antipsychotic Drug Doses for BPSD in Alzheimer's Disease Using Deep Learning Techniques

Background/Objectives: Antipsychotic dosing for behavioral and psychological symptoms of dementia (BPSD) in Alzheimer's disease remains empirical and variable. This study develops a deep learning model to predict individualized antipsychotic doses from structural MRI. Methods: A transfer learning approach with a cascaded ResNet (Cas-ResNet) was used. The model was first pre-trained on a large healthy aging dataset (CBMFM, n = 646) for brain age prediction, then fine-tuned on a BPSD dataset (SMHC, n = 86) to predict the defined daily dose (DDD) of antipsychotics. Model interpretability was performed using Grad CAM to identify predictive brain regions. Results: The proposed model achieved a mean absolute error of 0.19 and a Pearson correlation of 0.66 between predicted and actual doses, outperforming baseline 3DCNN, VGG, and DenseNet. Key contributing regions included the left inferior temporal gyrus, right parahippocampal gyrus, right putamen, left middle temporal gyrus, and left caudate. Conclusions: This proof-of-concept study demonstrates that deep learning can predict personalized antipsychotic doses from structural MRI, offering an objective tool to standardize BPSD pharmacotherapy and reduce empirical prescribing. The identified brain regions provide neurobiological insights into treatment response.