Figure 3.

Linear and recurrent neural network model architecture for computing vertical (vert) and anterior–posterior (A/P) ground reaction forces (GRF). The linear models predicted vertical and A/P GRFs from five different pressure regions (PR_1–5) divided by subject−specific body mass (BM). The coefficients (A_1–5, B_1–5) were computed using a least−square regression. The recurrent neural network predicted GRFs based on eight different inputs: PR_1–5, body mass, speed and slope. The bidirectional long short−term memory (LSTM) layers utilized information from the current sample (x_t) and from the future (x_t+1) and previous samples (x_t-1). For example, the GRF at 1% of the step is informed from the input at 0%, 1%, and 2% of the step. Both models were cross−validated using a leave−one subject out approach.