Knowledge Discovery for Patient Survival in a Clinical Discharge Dataset Using Causal Graph Ontological Framework

Abstract

Knowledge mining from clinical datasets is a critical task in healthcare as well as other fields. While the existing methods, such as randomized controlled trials (RCT) and other automatic machine extraction, have been helpful, they have become increasingly insufficient to keep pace with time, and robust models are required for clinical decisions. In this paper, we present a new method to address this challenge by using the Causal graph ontological model. Our study used a semi-structured textual clinical discharge dataset from the Statewide Planning and Research Cooperative System (SPARCS) to design and validate the patient survival rate assumptions from the dataset. We extracted the clinical information and organized it according to medically relevant fields for decision-making (Diseases, confounders, treatment, and the survival rate). The initial assumptions model was validated using the conditional independent test (CIT) criteria. The outputs of the LocalTest validation showed that the conceptual assumptions of the causal graph hold since the Pearson correlation coefficient ranges between -1 and 1, the p-value was (>0.05), and the confidence intervals of 95% and 25% were satisfied. Furthermore, we used Shapley values to perform sensitivity analysis on the features. Our analysis showed that two variables, such as gender and diseases, contributed little to the survival rate prediction. Our study concludes that the combination of causal graph ontological framework and sensitivity analysis to discover knowledge from the clinical text could help improve the quality of clinical decisions in the text, remove bias in the assumption in medical applications, and serve as a premise for modelling causal data for natural Language machine learning predictions.