Fig. 1: Description of medical application of SSRL in the clinical settings and the flow of data. | npj Digital Medicine

Fig. 1: Description of medical application of SSRL in the clinical settings and the flow of data.

From: A scoping review of self-supervised representation learning for clinical decision making using EHR categorical data

Fig. 1

EHR data follow a cyclical process, beginning at health centers, where they are either used to train internal SSRL models (blue box) or directly supplied to external SSRL models (purple box). These models convert the data into efficient representations, which are then adapted based on the specific downstream tasks. The results from these downstream tasks are sent back to the health centers, facilitating the delivery of effective medicine and medical knowledge discovery. Blue and orange arrows represent unsupervised and supervised learning tasks, respectively. For efficient representations, the snowflake and cluster icons stand for obtaining efficient representations with respectively frozen (only inference) or trainable (with high computational resources such as high-performance computing) SSRL models. The gear icon signifies the training of downstream models using moderate resources, such as multiple GPUs. The potential use of externally developed SSRL models is highlighted in purple.

Back to article page