Date of Award

Spring 2021

Document Type


Degree Name

Master of Science in Education (MSEd)


Human Movement Sciences


Exercise Science

Committee Director

Hunter J. Bennett

Committee Member

Rumit Singh Kakar

Committee Member

Stacie Ringleb

Committee Member

Kevin Valenzuela


BACKGROUND: Knee joint contact (bone on bone) forces are commonly estimated using surrogate measures such as external knee adduction moments (with limited success) or musculoskeletal modeling (more successful). Despite its capabilities, modeling is not optimal for clinicians or persons with limited experience and knowledge. Therefore, the purpose of this study was to design a novel prediction method for knee joint contact forces that is equal or more accurate than modeling, yet simplistic in terms of required inputs. METHODS: This study included all six subjects’ (71.3±6.5kg, 1.7±0.1m) data from the opensource “Grand Challenge” datasets ( and two subjects from the "CAMS" datasets, consisting of motion capture and in-vivo instrumented knee prosthesis data (e.g. true knee joint contact forces). Inverse kinematics were used to derive three-dimensional hip, two-dimensional knee (sagittal & frontal), and one-dimensional ankle (sagittal) kinematics during the stance phase of normal walking for all subjects. Medial and lateral knee joint contact forces (normalized to body weight) and inverse kinematics were imported into MATLAB and normalized to 101 data points. A long-short term memory network (LSTM) was created to predict knee forces using combinations of the kinematics inputs. The Grand Challenge data were used for training, while the CAMS data were used for testing. Waveform accuracy was explained by the proportion of variance and root mean square error between network predictions and in-vivo knee joint contact forces data. RESULTS: The top five networks demonstrated excellent fit with the training data, achieving RMSE < 0.26BW for medial and lateral forces, R2 > 0.69 for medial forces, but only R2 > 0.15 for lateral forces. The overall best-selected network contained frontal hip and knee, and sagittal hip and ankle input variables and presented the finest visual waveform agreement with the in vivo data (R2=0.77, RMSE=0.27). CONCLUSIONS: The LSTM network designed in this study revealed knee joint forces could accurately be predicted by using only kinematic input variables. The network’s results outperformed most reports of root mean squared errors and correlation coefficients attained by musculoskeletal modeling and surrogate measures of KAMs.