AI-Assisted Hypothesis Generation to Address Challenges in Cardiotoxicity Research: Simulation Study Using ChatGPT With GPT-4o

Background: Cardiotoxicity is a major concern in heart disease research because it can lead to severe cardiac damage, including heart failure and arrhythmias.

Objective: This study aimed to explore the ability of ChatGPT with GPT-4o to generate innovative research hypotheses to address 5 major challenges in cardiotoxicity research: the complexity of mechanisms, variability among patients, the lack of detection sensitivity, the lack of reliable biomarkers, and the limitations of animal models.

Methods: ChatGPT with GPT-4o was used to generate multiple hypotheses for each of the 5 challenges. These hypotheses were then independently evaluated by 3 experts for novelty and feasibility. ChatGPT with GPT-4o subsequently selected the most promising hypothesis from each category and provided detailed experimental plans, including background, rationale, experimental design, expected outcomes, potential pitfalls, and alternative approaches.

Results: ChatGPT with GPT-4o generated 96 hypotheses, of which 13 (14%) were rated as highly novel and 62 (65%) as moderately novel. The average group score of 3.85 indicated a strong level of innovation in these hypotheses. Literature searching identified at least 1 relevant publication for 28 (29%) of the 96 hypotheses. The selected hypotheses included using single-cell RNA sequencing to understand cellular heterogeneity, integrating artificial intelligence with genetic profiles for personalized cardiotoxicity risk prediction, applying machine learning to electrocardiogram data for enhanced detection sensitivity, using multi-omics approaches for biomarker discovery, and developing 3D bioprinted heart tissues to overcome the limitations of animal models. Our group's evaluation of the 30 dimensions of the experimental plans for the 5 hypotheses selected by ChatGPT with GPT-4o revealed consistent strengths in the background, rationale, and alternative approaches, with most of the hypotheses (20/30, 67%) receiving scores of ≥4 in these areas. While the hypotheses were generally well received, the experimental designs were often deemed overly ambitious, highlighting the need for more practical considerations.

Conclusions: Our study demonstrates that ChatGPT with GPT-4o can generate innovative and potentially impactful hypotheses for overcoming critical challenges in cardiotoxicity research. These findings suggest that artificial intelligence-assisted hypothesis generation could play a crucial role in advancing the field of cardiotoxicity, leading to more accurate predictions, earlier detection, and better patient outcomes.