We present an investigation into modeling of athetoid motion and prediction of user intent, for use in assistive computer interfaces during icon-clicking tasks. Data were recorded from three athetoid patients during unassisted icon-clicking trials with an isometric joystick. In order to facilitate development and testing of filter designs without the difficulty of repeated testing with human subjects, a quantitative model of the recorded patient data was developed using pseudoinverse methods. Using this model within the visuomotor control loop for the icon-clicking task, a prediction filter was then developed to reduce the target acquisition time. The filter is based on a novel "autoregressive stretching window" model which selects five data points evenly distributed across the input and output histories to predict the intended target, together with a second-order system that smoothes the movement of the cursor. On average, the filter demonstrated a reduction of target acquisition time by a factor of 2.7 in experiments with the patient models.