Objective To determine the validity of ground reaction forces (GRF) during running estimated from 3D lower body landmark coordinates obtained via a markerless system using the Long Short-Term Memory (LSTM) neural network model. Methods 59 recreational runners were recruited. The video and GRF during running were collected by the motion capture system and force plates (FP). The 3D coordinates of 11 lower body landmarks, obtained via the markerless system, were used as inputs in LSTM model to estimate 3D GRF. The estimation performance was evaluated using correlation coefficients r, root mean square error (RMSE) and normalized root mean square error (nRMSE) by comparing LSTM model estimation and FP measurement. Statistical Parametric Mapping was used to analyze differences in GRF curves between the LSTM model and FP, while paired t-tests assessed differences in GRF characteristics. Results A strong correlation (r>0.85,P<0.001) and lower error (RMSE<0.3 Body Weight,nRMSE＜15%) was found between the LSTM model estimation and FP measurements. No significant difference area was found in GRF curves between LSTM model estimation and FP measurements. About the GRF characteristics, there was no significant difference between LSTM model estimation and FP measurements(P＞0.372). Conclusion With the 3D coordinates of lower body landmarks based on markerless system as inputs,the 3D GRF curves could be accurately estimated by LSTM model. The LSTM model developed in this study can be used to monitor running injury risks in outdoor environments.