Algorithms for Empathy: Using Machine Learning to Categorize Common Empathetic Traits Across Professional and Peer-Based Conversations.

Abstract

Introduction In this article, we describe our creation of a machine-learning model that uses a combination of rule-based and natural language processing (NLP) algorithms. We show how this "Empathy Algorithm" was developedand how its results compare to three datasets of professional counseling and peer-led conversations. Methods These conversation datasets were rated by people with varying degrees of empathetic expertise (from counselors to student volunteers)and labeled as either low- or high-quality empathy. Our methodology involved running both these "low-empathy" and "high-empathy" conversations through our algorithm and then looking for a correlation between conversations labeled "high empathy" and an increased presence of six empathy skills flagged by our algorithm. Results We found positive correlations between four ofthe six skills that our algorithm measures (i.e., four empathizing skills showed up the same or more in each of the "high-empathy" conversations within the three datasets). This suggests that certain empathizing skills are not only consistently present in effective conversationsbut also quantifiable enough to be measured by today's machine-learning models. Conclusion While limitations of language, binary classifications, and non-verbal cues remain as opportunities for further development, using algorithms to objectively assess empathic skills represents an important step to improveclient outcomes and refine communication practices for today's healthcare professionals.