A systematic review and meta-analysis of diagnostic performance comparison between generative AI and physicians
Bibliographic record
Abstract
Abstract Background The rapid advancement of generative artificial intelligence (AI) has led to the wide dissemination of models with exceptional understanding and generation of human language. Their integration into healthcare has shown potential for improving medical diagnostics, yet a comprehensive diagnostic performance evaluation of generative AI models and the comparison of their diagnostic performance with that of physicians has not been extensively explored. Methods In this systematic review and meta-analysis, a comprehensive search of Medline, Scopus, Web of Science, Cochrane Central, and MedRxiv was conducted for studies published from June 2018 through December 2023, focusing on those that validate generative AI models for diagnostic tasks. The risk of bias was assessed using the Prediction Model Study Risk of Bias Assessment Tool. Meta-regression was performed to summarize the performance of the models and to compare the accuracy of the models with that of physicians. Results The search resulted in 54 studies being included in the meta-analysis. Nine generative AI models were evaluated across 17 medical specialties. The quality assessment indicated a high risk of bias in the majority of studies, primarily due to small sample sizes. The overall accuracy for generative AI models across 54 studies was 56.9% (95% confidence interval [CI]: 51.0–62.7%). The meta-analysis demonstrated that, on average, physicians exceeded the accuracy of the models (difference in accuracy: 14.4% [95% CI: 4.9–23.8%], p-value =0.004). However, both Prometheus (Bing) and GPT-4 showed slightly better performance compared to non-experts (-2.3% [95% CI: -27.0–22.4%], p-value = 0.848 and -0.32% [95% CI: -14.4–13.7%], p-value = 0.962), but slightly underperformed when compared to experts (10.9% [95% CI: -13.1–35.0%], p-value = 0.356 and 12.9% [95% CI: 0.15–25.7%], p-value = 0.048). The sub-analysis revealed significantly improved accuracy in the fields of Gynecology, Pediatrics, Orthopedic surgery, Plastic surgery, and Otolaryngology, while showing reduced accuracy for Neurology, Psychiatry, Rheumatology, and Endocrinology compared to that of General Medicine. No significant heterogeneity was observed based on the risk of bias. Conclusions Generative AI exhibits promising diagnostic capabilities, with accuracy varying significantly by model and medical specialty. Although they have not reached the reliability of expert physicians, the findings suggest that generative AI models have the potential to enhance healthcare delivery and medical education, provided they are integrated with caution and their limitations are well-understood. Key Points Question: What is the diagnostic accuracy of generative AI models and how does this accuracy compare to that of physicians? Findings: This meta-analysis found that generative AI models have a pooled accuracy of 56.9% (95% confidence interval: 51.0–62.7%). The accuracy of expert physicians exceeds that of AI in all specialties, however, some generative AI models are comparable to non-expert physicians. Meaning: The diagnostic performance of generative AI models suggests that they do not match the level of experienced physicians but that they may have potential applications in healthcare delivery and medical education.
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 | 0.015 | 0.077 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.010 | 0.022 |
| Bibliometrics | 0.006 | 0.007 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.003 | 0.002 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.005 | 0.000 |
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".