Human evaluation of large language models in healthcare: gaps, challenges, and the need for standardization
Raghav Awasthi, Atharva Bhattad, S. Ramachandran, Shreya Mishra, Ashish K. Khanna, Jacek B. Cywiński, Kamal Maheshwari, Dwarikanath Mahapatra, Izabella DiRosa, Anabelle Cohen, Hajra Arshad, Aarit Atreja, Asma Alshukaili, Aryan Vohra, Nishant Singh, Francis Papay, Ashish Atreja, Rahul Kashyap, Piyush Mathur
Abstract
Publications related to experimentation with Large Language Models (LLMs) in healthcare are rapidly increasing. While human evaluation remains the gold standard for evaluating LLMs, there is still a lack of standardization in its implementation. In this review article, we systematically examine studies involving LLMs in healthcare that have conducted human evaluations. We analyze the metrics used, assess their variability across studies. We also propose a standardized framework along with an interactive open web application HumanELY, to facilitate human evaluation. We believe that use of HumanELY will provide an opportunity for consistent, comprehensive, reliable, reproducible, and measurable human evaluations of LLM in healthcare. HumanELY is publicly available at https://www.brainxai.com/humanely .