Knowing the ground reaction force (GRF) helps to quantify biomechanical load during running. However, GRF measurement is restricted to the lab. Artificial neural networks (ANNs) are capable to model complex relations and thus could be used to predict GRF from inertial measurement units (IMUs). This prediction would be a step towards the quantification of biomechanical load in the field. PURPOSE: Show the possibilities of an ANN to estimate GRFs in vertical and posterior/anterior direction using three IMUs. METHODS: 6 experienced heel strike runners (3 F, 3 M; age 31.5y ± 11.7y, height 1.77 m ± 0.07 m, mass 60.7 kg ± 16.0 kg) ran 9 trials on a force-instrumented treadmill at combination of three velocities (10, 12 and 14 km/h) and three stride frequencies (preferred, -10% of preferred and + 10% of preferred). Subjects were instrumented with IMUs (240 Hz) mounted at both proximal tibias and pelvis. With a per subject approach, 40 strides were extracted per trial, 20% of this data was used as validation and 80% as training set. Two layer ANNs (250 and 100 neurons) were then trained with the gravity subtracted acceleration in the global frame as input to fit the vertical and posterior/anterior GRF. Performance of the models was analysed with the root mean squared error (RMSE) and the Pearson’s correlation coefficient (r). The mean of the absolute error of the vertical peak during stance was calculated as an absolute value and percentage. The absolute and percentual error of the posterior/anterior breaking and push-off force peaks were determined. RESULTS: The ANN modelled peak GRF forces with high accuracy (vertical: r > 0.99, mean absolute peak error < 2.4%; posterior/anterior: r > 0.96, mean absolute peak error < 11.0%; Table 1). CONCLUSION: Deployment of ANNs to predict 2D GRFs directly from gravity subtracted acceleration in the global frame is very promising with a subject specific approach.