Neurosymbolic, Multiagent AI Speeds Oncology Clinical Trial Matching by Fourfold

Objective:

To evaluate the effectiveness of a neurosymbolic, multi-agent AI system in improving oncology clinical trial matching, highlighting its significance in enhancing patient access to trials.

Key Findings:

• The AI system achieved an F1 score of 0.82, significantly outperforming GPT-4 zero-shot (0.47) and chain-of-thought (0.67) baselines, with a sensitivity of 0.8375 and specificity of 0.8359.
• Median screening time decreased from 120 minutes to approximately 30 minutes with the AI system.
• The system processed 157,000 pages and produced 17,912 confirmed matches within a median time-to-recommendation of fewer than 7 days.
• No demographic subgroup showed an F1 gap greater than 10 percentage points, with the largest gap being about 7 points.

Interpretation:

The study demonstrates that a neurosymbolic, multi-agent AI system can effectively enhance the speed and accuracy of clinical trial matching in oncology, with minimal performance gaps across demographic subgroups.

Limitations:

• The study did not declare any funding sources, which may raise questions about potential biases.
• The performance evaluation was based on a specific cohort and may not generalize to all patient populations.

Conclusion:

The findings indicate that AI can operationalize clinical trial matching at the necessary speed and complexity, potentially increasing patient access to trials and paving the way for future research in AI applications in oncology.