Automating hybrid collective intelligence in open-ended medical diagnostics.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Linguistics
• Medical Informatics

Background:

• Collective intelligence enhances decision-making in various fields but is underutilized in complex, open-ended medical diagnostics due to challenges in integrating diverse inputs.
• Existing applications of collective intelligence often focus on simpler tasks, limiting its potential in broad medical diagnosis scenarios.

Purpose of the Study:

• To develop and evaluate a fully automated approach for leveraging collective intelligence in general medical diagnostics.
• To overcome the challenge of integrating unstandardized diagnostic inputs from multiple clinicians using semantic knowledge graphs, natural language processing, and medical ontologies.

Main Methods:

• Developed an automated system integrating semantic knowledge graphs, natural language processing (NLP), and the SNOMED CT medical ontology.
• Applied the method to 1,333 medical cases from The Human Diagnosis Project, where each case was reviewed by ten diagnosticians.
• Compared the diagnostic accuracy of individual diagnosticians against collective diagnoses from groups of varying sizes.

Main Results:

• Individual diagnosticians achieved an average accuracy of 46%.
• Pooling decisions from ten diagnosticians using the automated approach increased accuracy to 76%.
• Accuracy improvements were observed across different medical specialties, chief complaints, and levels of diagnostician experience.

Conclusions:

• The developed automated approach significantly enhances diagnostic accuracy in general medical settings by effectively harnessing collective intelligence.
• This method demonstrates the potential to reduce medical diagnostic errors and improve patient safety by aggregating insights from a global community of medical professionals.
• The integration of knowledge graphs and NLP is crucial for overcoming data heterogeneity in open-ended medical diagnostic tasks.