Performance of Multimodal Artificial Intelligence Chatbots Evaluated on Clinical Oncology Cases
Bibliographic record
Abstract
Importance: Multimodal artificial intelligence (AI) chatbots can process complex medical image and text-based information that may improve their accuracy as a clinical diagnostic and management tool compared with unimodal, text-only AI chatbots. However, the difference in medical accuracy of multimodal and text-only chatbots in addressing questions about clinical oncology cases remains to be tested. Objective: To evaluate the utility of prompt engineering (zero-shot chain-of-thought) and compare the competency of multimodal and unimodal AI chatbots to generate medically accurate responses to questions about clinical oncology cases. Design, Setting, and Participants: This cross-sectional study benchmarked the medical accuracy of multiple-choice and free-text responses generated by AI chatbots in response to 79 questions about clinical oncology cases with images. Exposures: A unique set of 79 clinical oncology cases from JAMA Network Learning accessed on April 2, 2024, was posed to 10 AI chatbots. Main Outcomes and Measures: The primary outcome was medical accuracy evaluated by the number of correct responses by each AI chatbot. Multiple-choice responses were marked as correct based on the ground-truth, correct answer. Free-text responses were rated by a team of oncology specialists in duplicate and marked as correct based on consensus or resolved by a review of a third oncology specialist. Results: This study evaluated 10 chatbots, including 3 multimodal and 7 unimodal chatbots. On the multiple-choice evaluation, the top-performing chatbot was chatbot 10 (57 of 79 [72.15%]), followed by the multimodal chatbot 2 (56 of 79 [70.89%]) and chatbot 5 (54 of 79 [68.35%]). On the free-text evaluation, the top-performing chatbots were chatbot 5, chatbot 7, and the multimodal chatbot 2 (30 of 79 [37.97%]), followed by chatbot 10 (29 of 79 [36.71%]) and chatbot 8 and the multimodal chatbot 3 (25 of 79 [31.65%]). The accuracy of multimodal chatbots decreased when tested on cases with multiple images compared with questions with single images. Nine out of 10 chatbots, including all 3 multimodal chatbots, demonstrated decreased accuracy of their free-text responses compared with multiple-choice responses to questions about cancer cases. Conclusions and Relevance: In this cross-sectional study of chatbot accuracy tested on clinical oncology cases, multimodal chatbots were not consistently more accurate than unimodal chatbots. These results suggest that further research is required to optimize multimodal chatbots to make more use of information from images to improve oncology-specific medical accuracy and reliability.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.019 | 0.135 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.004 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".