MedHELM: Holistic Evaluation of Large Language Models for Medical Tasks

While large language models (LLMs) achieve near-perfect scores on medical licensing exams, these evaluations inadequately reflect the complexity and diversity of real-world clinical practice. We introduce MedHELM, an extensible evaluation framework for assessing LLM performance for medical tasks with three key contributions. First, a clinician-validated taxonomy spanning 5 categories, 22 subcategories, and 121 tasks developed with 29 clinicians. Second, a comprehensive benchmark suite comprising 35 benchmarks (17 existing, 18 newly formulated) providing complete coverage of all categories and subcategories in the taxonomy. Third, a systematic comparison of LLMs with improved evaluation methods (using an LLM-jury) and a cost-performance analysis. Evaluation of 9 frontier LLMs, using the 35 benchmarks, revealed significant performance variation. Advanced reasoning models (DeepSeek R1: 66% win-rate; o3-mini: 64% win-rate) demonstrated superior performance, though Claude 3.5 Sonnet achieved comparable results at 40% lower estimated computational cost. On a normalized accuracy scale (0-1), most models performed strongly in Clinical Note Generation (0.73-0.85) and Patient Communication & Education (0.78-0.83), moderately in Medical Research Assistance (0.65-0.75), and generally lower in Clinical Decision Support (0.56-0.72) and Administration & Workflow (0.53-0.63). Our LLM-jury evaluation method achieved good agreement with clinician ratings (ICC = 0.47), surpassing both average clinician-clinician agreement (ICC = 0.43) and automated baselines including ROUGE-L (0.36) and BERTScore-F1 (0.44). Claude 3.5 Sonnet achieved comparable performance to top models at lower estimated cost. These findings highlight the importance of real-world, task-specific evaluation for medical use of LLMs and provides an open source framework to enable this.

@article{bedi2025_2505.23802,
  title={ MedHELM: Holistic Evaluation of Large Language Models for Medical Tasks },
  author={ Suhana Bedi and Hejie Cui and Miguel Fuentes and Alyssa Unell and Michael Wornow and Juan M. Banda and Nikesh Kotecha and Timothy Keyes and Yifan Mai and Mert Oez and Hao Qiu and Shrey Jain and Leonardo Schettini and Mehr Kashyap and Jason Alan Fries and Akshay Swaminathan and Philip Chung and Fateme Nateghi and Asad Aali and Ashwin Nayak and Shivam Vedak and Sneha S. Jain and Birju Patel and Oluseyi Fayanju and Shreya Shah and Ethan Goh and Dong-han Yao and Brian Soetikno and Eduardo Reis and Sergios Gatidis and Vasu Divi and Robson Capasso and Rachna Saralkar and Chia-Chun Chiang and Jenelle Jindal and Tho Pham and Faraz Ghoddusi and Steven Lin and Albert S. Chiou and Christy Hong and Mohana Roy and Michael F. Gensheimer and Hinesh Patel and Kevin Schulman and Dev Dash and Danton Char and Lance Downing and Francois Grolleau and Kameron Black and Bethel Mieso and Aydin Zahedivash and Wen-wai Yim and Harshita Sharma and Tony Lee and Hannah Kirsch and Jennifer Lee and Nerissa Ambers and Carlene Lugtu and Aditya Sharma and Bilal Mawji and Alex Alekseyev and Vicky Zhou and Vikas Kakkar and Jarrod Helzer and Anurang Revri and Yair Bannett and Roxana Daneshjou and Jonathan Chen and Emily Alsentzer and Keith Morse and Nirmal Ravi and Nima Aghaeepour and Vanessa Kennedy and Akshay Chaudhari and Thomas Wang and Sanmi Koyejo and Matthew P. Lungren and Eric Horvitz and Percy Liang and Mike Pfeffer and Nigam H. Shah },
  journal={arXiv preprint arXiv:2505.23802},
  year={ 2025 }
}