Posted 1/19/2024, 6:00:57 AM
AI Models Show Promise for Medicine But May Be Too Brittle for Reliable Clinical Use
- New study finds AI models predict patient outcomes well when trained and tested on same trial data, but fail to generalize across different studies
- Suggests machine learning models may be too "fragile" for reliable medical use, providing a "cautionary tale"
- Core issue is getting enough high-quality medical data to train robust AI models that perform consistently
- For psychiatry, expert predicts clinical use of these prediction models is at least 10 years away
- But better data practices could enable AI to generate next level of advancement in medicine long-term