This research is conducted for developing an automated method to recognize variations in the Newborn Life Support (NLS) procedure. Compliance with the NLS standard guideline is essential to prevent any adverse consequences for the newborn. Video recordings of resuscitation are frequently used in research to identify types of variations and understand how to minimize unwanted ones. Despite their benefits, it takes a significant amount of time and human resources to manually evaluate the procedure from videos. Therefore, an automated method could help. In this study, a variation recognition based on an action recognition technique is built. In the first step, automatic object segmentation is performed on every NLS action image. In the second stage, a number of features involving the proportion of medical objects availability and their movement, as well as association among actions are extracted and fed into machine learning models. The results show that the strategy of considering actions' associations and preliminary prediction of actions succeeded in improving the model performance. However, the whole recognition system still works fairly, and it is only for the wet towel removal step in the procedure, yet it has been useful to inform the adherence of the recorded procedure to the NLS guideline. This study is an initial work that will advance toward the integration of automated variation recognition with reliability modeling work on the NLS procedure.