Parkinson's disease (PD) is a degenerative, progressive disorder of the central nervous system that mainly affects motor control. The aim of this study was to develop data-driven methods and test their clinimetric properties to detect and quantify PD motor states using motion sensor data from leg agility tests. Nineteen PD patients were recruited in a levodopa single dose challenge study. PD patients performed leg agility tasks while wearing motion sensors on their lower extremities. Clinical evaluation of video recordings was performed by three movement disorder specialists who used four items from the motor section of the unified PD rating scale (UPDRS), the treatment response scale (TRS) and a dyskinesia score. Using the sensor data, spatiotemporal features were calculated and relevant features were selected by feature selection. Machine learning methods like support vector machines (SVM), decision trees, and linear regression, using ten-fold cross validation were trained to predict motor states of the patients. SVM showed the best convergence validity with correlation coefficients of 0.81 to TRS, 0.83 to UPDRS #31 (body bradykinesia and hypokinesia), 0.78 to SUMUPDRS (the sum of the UPDRS items: #26-leg agility, #27-arising from chair, and #29-gait), and 0.67 to dyskinesia. Additionally, the SVM-based scores had similar test-retest reliability in relation to clinical ratings. The SVM-based scores were less responsive to treatment effects than the clinical scores, particularly with regards to dyskinesia. In conclusion, the results from this study indicate that using motion sensors during leg agility tests may lead to valid and reliable objective measures of PD motor symptoms.