Evaluation of the prognostic value of Ioflupane I-123 SPECT for predicting functional decline in Parkinson's Disease using machine learning

Aim: The course of Parkinson's disease (PD) in patients is highly variable and currently difficult to predict. Predicting functional decline in early PD would help patients and clinicians with planning care and improve therapeutic trials. The purpose of this study is to evaluate the prognostic value of Ioflupane I-123 SPECT for predicting functional decline in PD patients using machine learning. Materials and Methods: Parkinson's Progression Markers Initiative (PPMI) database was queried and yielded 552 patients (mean age = 62.1 +/- 9.8, 340 men, 212 women) with early PD and abnormal I-123 Ioflupane SPECT. Movement Disorder Society-Unified Parkinson's Disease Rating Scale was available for N = 425 patients at 1 year following SPECT imaging (UPDRS-1yr). Modified Schwab & England Activities of Daily Living Scale was available for N = 499 patients at 1 year following SPECT imaging (MSE-1yr). An artificial neural network was constructed with Striatal Binding Ratios (SBRs), current UPDRS, current MSE, symptomatic treatment status, and age as inputs and UPDRS and MSE at 1 year as targets. 70% of patient data was used for training, 15% for validation and 15% for testing. Learning was performed using Bayesian Regularization using between 1 and 3 hidden layers and sigmoid or ReLU activation functions. Normalized Mean Squared Error (NMSE) and Regression R value between outputs and targets were compared using training data with SBRs and without SBRs as inputs. Results: Artificial neural network with SBRs as inputs, 2 hidden layers and ReLU activation function resulted in lowest NMSE (0.0793) and highest R (0.741) compared to network without SBRs inputs (NMSE = 0.0858, R = 0.678) for prediction of UPDRS-1yr. The UPDRS-1yr target testing set with N = 63/425 yielded R of 0.678 and two tailed p < 0.00001. Artificial neural network with SBRs as inputs did not demonstrate a lower NMSE (0.005) or higher R (0.600) compared to network without SBRs inputs (NMSE = 0.005, R = 0.603) for prediction of MSE-1yr. Conclusion: Ioflupane I-123 SPECT SBRs can improve the prognostic value of an artificial neural network for predicting the course of PD assessed with UPDRS at 1 year following imaging.