ROLE OF ARTIFICIAL INTELLIGENCE IN PREDICTING OUTCOMES OF CLINICAL TRIALS IN ONCOLOGY

Abstract

The field of oncology continues to face high attrition rates in clinical trials, with nearly 90% of investigational agents failing before market approval. This places enormous financial and ethical burdens on research institutions and patients alike. The advent of Artificial Intelligence (AI), particularly large language models (LLMs) and specialized neural networks, presents a promising strategy to enhance predictive accuracy for trial outcomes.

Objective: This study evaluates and compares the predictive performance of multiple AI architectures—including GPT-3.5, GPT-4, GPT-4mini, GPT-4o, LLaMA3, and the HINT model—in forecasting success or failure of oncology clinical trials using real-world datasets.

Methods: Oncology trial data were extracted from ClinicalTrials.gov. Structured trial inputs—phase, sponsor type, duration, endpoints, intervention class—were standardized across models. GPT-family and LLaMA3 models were accessed via API and prompted to classify trial outcomes. HINT, a structured AI model, was used with molecular and ICD-10 inputs. Performance was evaluated using Balanced Accuracy, Matthews Correlation Coefficient (MCC), Recall, and Specificity.

Results: GPT-4o demonstrated the highest balanced accuracy (0.573) and robust recall (0.931), while HINT achieved the best specificity (0.541) and stable MCC (0.111). All models underperformed in predicting outcomes for complex oncology subtypes and longerduration trials.

Conclusion: AI models hold significant promise in improving predictive capabilities for oncology trial outcomes. GPT-4o and HINT exhibit complementary strengths, and ensemble approaches may optimize trial design, patient selection, and resource allocation.