Propulsion power is an important metric in wheelchair racing. For a flat surface, it can be estimated from the sum of friction power and kinetic power. Usually, to determine friction power, the rolling and air drag coefficient first need to be determined with coast-down tests or other time-consuming methods. The aim of this paper was to investigate whether friction power could be estimated from Inertial Measurement Unit (IMU) data during wheelchair propulsion without the need for previously determining these coefficients. Two approaches were investigated using the kinematic data of a wheelchair athlete measured by a wheel, frame and trunk IMU. Firstly, an approach was used that considers the recovery phase of the propulsion cycle to be a coast-down period. Secondly, a machine-learning approach (Random Forest Regressor) was implemented. Coast-down tests were used to calculate a reference power with which the results from the two approaches could be compared. Results indicate that the machine-learning approach is more promising than the recovery phase analysis. However, whether the current machine-learning model can predict friction power for unseen subjects and surfaces should still be determined with inter-subject validation. Otherwise, it is recommended that a machine-learning model is trained for multiple subjects and a variation in conditions affecting friction force (surface, tyre pressure, wind, slope) to achieve a more robust model.