PhenoDP: leveraging deep learning for phenotype-based case reporting, disease ranking, and symptom recommendation

Area of Science:

• Computational biology
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Current diagnostic tools face challenges in rare disease prioritization due to incomplete data and complex presentations.
• Existing methods lack patient-centered insights and symptom recommendation capabilities for differential diagnosis.

Purpose of the Study:

• To develop an advanced deep learning toolkit, PhenoDP, for enhancing Mendelian disease diagnosis.
• To improve disease prioritization, clinical summarization, and symptom recommendation in rare disease diagnostics.

Main Methods:

• PhenoDP utilizes a three-module deep learning approach: Summarizer (distilled LLM for HPO term summarization), Ranker (combining similarity measures for disease prioritization), and Recommender (contrastive learning for HPO term suggestion).
• The Summarizer generates patient-centered clinical summaries from Human Phenotype Ontology (HPO) terms.
• The Ranker integrates information content, phi-based, and semantic similarity measures for disease ranking.

Main Results:

• PhenoDP's Summarizer provides more coherent and patient-centered summaries compared to FlanT5.
• The Ranker achieves state-of-the-art diagnostic performance, outperforming existing methods on simulated and real-world data.
• The Recommender demonstrates superior performance over GPT-4o and PhenoTips in enhancing diagnostic accuracy through suggested HPO terms.

Conclusions:

• PhenoDP significantly enhances Mendelian disease diagnosis using deep learning for summarization, ranking, and symptom recommendation.
• The toolkit offers a valuable clinical solution with potential to accelerate diagnosis and improve patient outcomes.
• PhenoDP is an open-source tool available for broader clinical adoption and research.