Historical Data as the Foundation of Clinical Foresight

Abstract

Clinical Decision Support systems based on historical patient data represent a model for investigation of a category of Clinical Decision Support (CDS) applications that have recently gained traction — consumption of historical data to support clinical decision making. An overview of the category is provided and addressed based on five research questions: What are the properties of historical data that provide insights? How are these insights used in Analytical Methods for Clinical Decision Support based on Historical Data? How is the architecture of the Clinical Decision Support System represented? How is clinical validation and evaluation achieved? How is the presentation and user interaction with predictions controlled? Despite being a nascent area, numerous published works are available on the topic, and clinical validation efforts are in various stages of maturity.

Historical data have become increasingly more available, are generally regarded as safe, and, unlike predictive models constructed on Future Data, do not raise questions regarding biases introduced in the construction of supervised learning models. However, while safety and bias concerns are reduced, historical data are prone to their own logical shortcomings. Therefore, users must exercise caution when interpreting the insights derived from Historical Data Analytics. Properly defined and presented, such insights do not impose a decision making burden on clinicians but instead ease the decision making process. The existence of a predictive model that maps the Clinical Data relevant to a Clinical Decision to the Clinical Decision itself is often regarded as a requirement for CDS to be genuinely useful for everyday clinical practice. Such a statement suffers from circular reasoning. In summary, because no method of construction of predictive models can claim to be free from bias, the pannational nature of probability suggests that, provided sufficient Causally Uncorrelated Data points are available notwithstanding time, space and biological differences, the possibility of identifying and using a Supervised Learning Model that accurately predicts the Decision of a Clinician can not be excluded.