Enhancing the Medical Diagnosis System and Treatment by Counterfactual Diagnostic Algorithm

Abstract

Clinical diagnosis and decision-making stand to be radically altered by machine learning. By identifying the underlying medical conditions, doctors may better explain their patients' symptoms via medical diagnosis. There is a substantial correlation between a patient's symptoms and medical history, yet current diagnostic algorithms only detect illnesses that are associated with each other. Suboptimal or even harmful diagnoses may arise from this failure to differentiate correlation from cause, as this article demonstrates. To get around this, they create novel algorithms for counterfactual diagnostics and reframe diagnosis as a work of counterfactual inference. In this work, researchers demonstrate that this method greatly enhances the reliability and security of the subsequent diagnosis while getting closer to the way doctors think about health problems. Using a battery of clinical scenarios, they evaluate the counterfactual method against 45 medical professionals, the gold standard Bayesian diagnosis system, and other algorithms. Compared to the counterfactual approach, which reaches expert clinical accuracy, the Bayesian algorithm obtains correctness, ranking in the top 26% of clinicians in the study unit. Changing this querying strategy alone yields this improvement, eliminating the need for any further model enhancements. Based on these results, the use of machine learning in medical diagnosis is incomplete without counterfactual thinking.