The goal of this study was to develop an algorithm that automatically quantifies motor states (off, on, dyskinesia) in Parkinson's disease (PD), based on accelerometry during a hand pronation-supination test. Clinician's ratings using the Treatment Response Scale (TRS), ranging from -3 (very Off) to 0 (On) to +3 (very dyskinetic), were used as target. For that purpose, 19 participants with advanced PD and 22 healthy persons were recruited in a single center open label clinical trial in Uppsala, Sweden. The trial consisted of single levodopa dose experiments for the people with PD (PwP), where participants were asked to perform standardized wrist rotation tests, using each hand, before and at prespecified time points after the dose. The participants used wrist sensors containing a three-dimensional accelerometer and gyroscope. Features to quantify the level, variation, and asymmetry of the sensor signals, three-level discrete wavelet transform features, and approximate entropy measures were extracted from the sensors data. At the time of the tests, the PwP were video recorded. Three movement disorder specialists rated the participants' state on the TRS. A Treatment Response Index from Sensors (TRIS) was constructed to quantify the motor states based on the wrist rotation tests. Different machine learning algorithms were evaluated to map the features derived from the sensor data to the ratings provided by the three specialists. Results from cross validation, both in tenfold and a leave-one-individual out setting, showed good predictive power of a support vector machine model and high correlation to the TRS. Values at the end tails of the TRS were under and over predicted due to the lack of observations at those values but the model managed to accurately capture the dose-effect profiles of the patients. In addition, the TRIS had good test-retest reliability on the baseline levels of the PD participants (Intraclass correlation coefficient of 0.83) and reasonable sensitivity to levodopa treatment (0.33 for the TRIS). For a series of test occasions, the proposed algorithms provided dose-effect time profiles for participants with PD, which could be useful during therapy individualization of people suffering from advanced PD.