Baseline [¹⁸F]FP-CIT PET-based deep learning prediction of levodopa-induced dyskinesia in Parkinson's disease

We aimed to develop a convolutional neural network (CNN) model with multi-task learning to predict the onset of levodopa-induced dyskinesia (LID) in patients with Parkinson's disease (PD) using baseline [¹⁸F]FP-CIT PET images. In this retrospective, single-center study, 402 patients were classified based on whether they developed LID within 5 years after starting levodopa (within 5 years: n = 134; beyond 5 years or none: n = 268). The proposed CNN model achieved a mean AUROC ± SD of 0.666 ± 0.036. Model-derived probabilities were also incorporated into a Cox regression model, yielding a mean concordance index (C-index ± SD) of 0.643 ± 0.046, significantly outperforming the model based on specific/nonspecific binding ratios of striatal subregions (C-index = 0.392 ± 0.036) in four of five test configurations. These results suggest that model-extracted features from [¹⁸F]FP-CIT PET carry prognostic value for LID, although further performance improvements are needed for clinical application.