Quantitative mobility analysis using wearable sensors has potential to identify, characterize and manage patients with movement disorders, including Parkinson’s disease (PD). Nonetheless, such sensors are not yet part of routine clinical examinations, in large part because it is still unclear which mobility tasks and which sensor-derived features per task should be analyzed to optimize/maximize the yield of this type of mobility analysis. To address this gap of knowledge, data from 262 participants with PD and 50 controls performing a series of motor tasks with a single wearable sensor on the lower back were analyzed using ensembles of heterogeneous machine learning models incorporating a wide range of classifiers and trained on a large set of features calculated from triaxial accelerometer and triaxial gyroscope signals. Our data show that sensor data analyzed with an ensemble of models effectively differentiate between participants with PD and controls. Furthermore, feature importance analysis revealed that a small number of more complex mobility tasks contribute the most informative features for accurate predictions, suggesting potential simplifications in wearable sensor-based mobility testing without sacrificing predictive performance.