Fig. 4.

A receiver operating characteristic (ROC) curve and confusion matrix for the convolutional neural network (CNN) model utilizing windowed time-lagged cross-correlation (WTLCC) in two experiments. The first experiment (‘WTLCC utility’) aimed to predict short-term outcome and was trained and validated on the Wroclaw University Hospital (WUH) database (P_training = 31 with 1,079 matrices, P_val = 16 with 573 matrices). The second experiment (‘WTLCC generalizability’) aimed to predict long-term outcome and was trained on the CENTER-TBI database and validated on WUH (P_training = 100 with 17,062 matrices, P_val = 47 with 6,220 matrices). Abbreviations: Fav. – Favourable outcome; Unfav.– Unfavourable outcome