Towards an autonomous clinical decision support system

Clinicians' decision-making is of utmost importance during critical situations. Thus, integrating Clinical Decision Support Systems (CDSS) may assist the medical staff by enhancing the decision-making process, eventually improving patient outcomes. The potential of an autonomous CDSS, proficient in predicting and guiding medical treatment, is significant-especially in situations where every second counts. We proposed a methodology to design a CDSS based on observational data of clinical procedures. This approach employs graph-convolutional networks (GCN) to encapsulate medical knowledge from simulated clinical procedures with sequential data. Consequently, our model can extrapolate from these procedures, identifying novel structural and characteristic combinations. This innovative method harnesses information that might elude human observers. Moreover, our model generates action sequences that a human physician has not previously executed. Traditional techniques tend to fall short in adapting to changing trends, thus failing to anticipate human actions. Conversely, advanced models like GCN have demonstrated promising potential in tasks like human action prediction, including activity recognition. We assessed these performances using benchmark datasets, which yielded encouraging results. Additionally, we constructed a graph-based CDSS to deliver pertinent medical advice. We outline a methodology to monitor the procedure's current stage and predict the physician's subsequent action, facilitating time-saving measures like pre-emptive instrument preparation. Our novel CDSS methodology achieved an F1-score of 0.899 and 0.714 when performing one and two-step predictions, respectively. Furthermore, our simulations illustrate a considerable time-saving potential, with an average reduction of approximately 00:01:28 +/- 00:01:15 min in the preparation time for adrenaline dosage, a crucial component for successful resuscitation.