Automatic identification of the occurrence of heel strike/initial contact (IC) and toe off/foot off (FO) gait events is a desirable initial step for deriving fast and accurate, clinically relevant results from 3D gait analysis of a patient with paediatric pathological gait. Research has used various techniques, such as coordinate-based algorithms, velocity-based algorithms, rule-based algorithms, machine learning and fuzzy logic-based techniques, for the determination of IC of the foot and FO events automatically in pathological gait. With deep learning, sequence-to-sequence long short-term memory network (LSTM) models have used kinematic features, like the 3D position of foot markers and their velocities, to automatically detect the gait events successfully. False positives are a concern when detecting these events and have been addressed earlier using peak detection methods. We used a sequence-to-sequence LSTM in our study with various combinations of kinematic features given as input to the LSTM network to examine the impact of the expert-derived features on the performance of the deep learning model and to check for the reduction of false positives obtained. The study successfully shows that changing input features to a model with an otherwise fixed configuration improves the deep learning model’s performance and results in a reduction in the number of false positives on the test set.