Digital Health Journalist, Founder & Editor-in-Chief of aboutDigitalHealth.com, Content Designer/Writer, Keynote Speaker, Moderator, Author
#GenAI won't be soon applied in #medicine due to #bias, #hallucination and #privacy concerns? According to an analysis of 120 articles published by March 2024 on PubMed, there are ways to mitigate these challenges. ๐ด ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ ๐บ๐ผ๐ฑ๐ฒ๐น๐ ๐ฎ๐ฟ๐ฒ ๐ฏ๐ถ๐ฎ๐๐ฒ๐ฑ โ Solution: multidimensional fairness evaluations; cross-disciplinary collaboration with ethicists, social scientists, and domain experts; reinforcement learning from expert feedback ๐ด ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ ๐ฐ๐ฎ๐ป ๐ฐ๐ผ๐บ๐ฝ๐ฟ๐ผ๐บ๐ถ๐๐ฒ ๐ฑ๐ฎ๐๐ฎ ๐ฝ๐ฟ๐ถ๐๐ฎ๐ฐ๐ โ Solution: localized architecture, federated learning methods, privacy-aware model training and deployment. ๐ด ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ ๐บ๐ถ๐๐๐ป๐ฑ๐ฒ๐ฟ๐๐๐ฎ๐ป๐ฑ๐ ๐ฝ๐ฟ๐ผ๐บ๐ฝ๐๐ โ Solution: understanding some heuristics for crafting effective prompts and algorithms to reduce adversarial attacks and ensure the responses align with human values ๐ด ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ ๐ต๐ฎ๐น๐น๐๐ฐ๐ถ๐ป๐ฎ๐๐ฒ๐ โ Solution: external review by experts, modifying model parameters, adjusting "temperature parameter" in GPT-4, adversarial testing or out-of-distribution evaluation ๐ด ๐ ๐ผ๐๐ ๐ด๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ ๐ฑ๐ฒ๐๐ฒ๐น๐ผ๐ฝ๐บ๐ฒ๐ป๐ ๐ถ๐ ๐ณ๐ผ๐ฐ๐๐๐ฒ๐ฑ ๐ผ๐ป ๐น๐ฎ๐ป๐ด๐๐ฎ๐ด๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น๐ โ Solution: generative AI with nontext data, multimodal foundations models ๐ด ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ ๐๐๐๐๐ฒ๐บ๐ ๐ฎ๐ฟ๐ฒ ๐ฑ๐๐ป๐ฎ๐บ๐ถ๐ฐ โ Solution: increased attention to HIPAA and GDPR compliance, greater regulation; using private health information to train models and, instead, using synthetic data sets Link to the full paper: https://lnkd.in/eyRpU9zT
