Why your doctor is not an algorithm: Exploring logical principles of different clinical inference methods using liver transplantation as a model.

Abstract

The development of machine learning (ML) tools in many different medical settings is largely increasing. However, the use of the resulting algorithms in daily medical practice is still an unsolved challenge. We propose an epistemological approach (i.e., based on logical principles) to the application of computational tools in clinical practice. We rely on the classification of scientific inference into deductive, inductive, and abductive comparing the characteristics of ML tools with those derived from evidence-based medicine [EBM] and experience-based medicine, as paradigms of well-known methods for generation of knowledge. While we illustrate our arguments using liver transplantation as an example, this approach can be applied to other aspects of the specialty. Regarding EBM, it generates general knowledge that clinicians apply deductively, but the certainty of its conclusions is not guaranteed. In contrast, automatic algorithms primarily rely on inductive reasoning. Their design enables the integration of vast datasets and mitigates the emotional biases inherent in human induction. However, its poor capacity for abductive inference (a logical mechanism inherent to human clinical experience) constrains its performance in clinical settings characterized by uncertainty, where data are heterogeneous, results are highly influenced by context, or where prognostic factors can change rapidly.