A phenotype-based AI pipeline outperforms human experts in differentially diagnosing rare diseases using EHRs

Xiaohao Mao; Yu Huang; Ye Jin; Lun Wang; Xuanzhong Chen; Honghong Liu; Xinglin Yang; Haopeng Xu; Xiaodong Luan; Ying Xiao; Siqin Feng; Jiahao Zhu; Xuegong Zhang; Rui Jiang; Shuyang Zhang; Ting Chen

doi:10.1530/ey.22.14.1

14.1

Prev Next

Section Contents Cite

ESPE Yearbook of Paediatric Endocrinology (2025) 22 14.1 | DOI: 10.1530/ey.22.14.1

ESPEYB25 14. The Year in Science and Medicine Artificial Intelligence (4 abstracts)

14.1. A phenotype-based AI pipeline outperforms human experts in differentially diagnosing rare diseases using EHRs

Xiaohao Mao , Yu Huang , Ye Jin , Lun Wang , Xuanzhong Chen , Honghong Liu , Xinglin Yang , Haopeng Xu , Xiaodong Luan , Ying Xiao , Siqin Feng , Jiahao Zhu , Xuegong Zhang , Rui Jiang , Shuyang Zhang & Ting Chen

NPJ Digit Med. 2025;8(1):68. doi: 10.1038/s41746-025-01452-1

Brief Summary: This study developed and evaluated PhenoBrain, an artificial intelligence (AI) pipeline for rare disease diagnosis using electronic health records (EHRs). It consists of two modules: 1) PBTagger, a deep learning-based Natural Language Processing (NLP) tool for extracting phenotypes from Chinese clinical texts, and 2) a differential diagnosis module using five novel machine learning models, including an Ensemble method. The system was tested on over 2,200 real and simulated cases from multiple international datasets and benchmarked against 50 physicians, ChatGPT, and GPT-4 in a controlled human-computer comparison.

This study addresses the global challenge of diagnosing rare diseases, which affect ~350 million people and are often misdiagnosed or diagnosed late. The authors introduce an AI system that automates phenotype extraction and disease ranking from EHRs. PhenoBrain outperformed 50 specialist physicians, ChatGPT, and GPT-4 in diagnostic accuracy across multiple datasets, achieving a top-3 recall of 0.61 and a top-10 recall of 0.81. When combined with human expertise, PhenoBrain further improved diagnostic performance, demonstrating its potential for integration into clinical workflows. It is also suggested to significantly reduce diagnostic delays and costs.

The study has several limitations, including the absence of multimodal data integration—such as genetic information, which is often essential for accurate rare disease diagnosis—and the current restriction to Chinese-language clinical texts. Furthermore, the diagnostic precision is constrained by incomplete or non-specific phenotype annotations available for many rare diseases.

Nevertheless, this study demonstrates that integrating AI predictions with physician expertise leads to superior diagnostic performance. Therefore, AI should be viewed not as a competitor, but as a collaborative partner in clinical decision-making.