Prediction of impulse control disorders in Parkinson's disease: a longitudinal machine learning study

Background Impulse control disorders (ICD) in Parkinson's disease (PD) patients mainly occur as adverse effects of dopamine replacement therapy. Despite several known risk factors associated with ICD development, this cannot yet be accurately predicted at PD diagnosis. Objectives We aimed to investigate the predictability of incident ICD by baseline measures of demographic, clinical, dopamine transporter single photon emission computed tomography (DAT-SPECT), and genetic variables. Methods We used demographic and clinical data of medication-free PD patients from two longitudinal datasets; Parkinson's Progression Markers Initiative (PPMI) (n=311) and Amsterdam UMC (n=72). We extracted radiomic and latent features from DAT-SPECT. We used single nucleotic polymorphisms (SNPs) from PPMI's NeuroX and Exome sequencing data. Four machine learning classifiers were trained on combinations of the input feature sets, to predict incident ICD at any follow-up assessment. Classification performance was measured with 10x5-fold cross-validation. Results ICD prevalence at any follow-up was 0.32. The highest performance in predicting incident ICD (AUC=0.66) was achieved by the models trained on clinical features only. Anxiety severity and age of PD onset were identified as the most important features. Performance did not improve with adding features from DAT-SPECT or SNPs. We observed significantly higher performance (AUC=0.74) when classifying patients who developed ICD within four years from diagnosis compared with those tested negative for seven or more years. Conclusions Prediction accuracy for later ICD development, at the time of PD diagnosis, is limited; however, it increases for shorter time-to-event predictions. Neither DAT-SPECT nor genetic data improve the predictability obtained using demographic and clinical variables alone.