Upper limb robotic rehabilitation devices can collect quantitative data about the user's movements. Identifying relationships between robotic sensor data and manual clinical assessment scores would enable more precise tracking of the time course of recovery after injury and reduce the need for time-consuming manual assessments by skilled personnel. This study used measurements from robotic rehabilitation sessions to predict clinical scores in a traumatic cervical spinal cord injury (SCI) population. A retrospective analysis was conducted on data collected from subjects using the Armeo®Spring (Hocoma, AG) in three rehabilitation centres. 14 predictive variables were explored, relating to range of motion, movement smoothness, and grip ability. Regression models using up to 4 predictors were developed to describe the following clinical scores: the GRASSP (consisting of four sub-scores), the ARAT, and the SCIM. The resulting adjusted R^2 value was highest for the GRASSP Quantitative Prehension component (0.78), and lowest for the GRASSP Sensibility component (0.54). In contrast to comparable studies in stroke survivors, movement smoothness was least beneficial for predicting clinical scores in SCI. Prediction of upper-limb clinical scores in SCI is feasible using measurements from a robotic rehabilitation device, without the need for dedicated assessment procedures.