Google DeepMind 训练模型模拟精神科问诊，鉴别诊断率翻倍至 68.3%
DeepMind 用纯模拟训练 psychiatric 问诊模型，32 名医生当病人测试，鉴别诊断率从 33% 涨到 68%，但漏判风险仍有一成多
Google DeepMind 用完全模拟的方式训练模型进行精神科问诊访谈，并让 32 名精神科医生扮演患者进行测试。在与相同提示词的基础模型对比中，新模型在 60 组匹配病例中盲评胜出 76.7%，恰当的鉴别诊断比例从 33.3% 提升到 68.3%。在对抗性场景中，急性风险漏判率从 30.5% 降至 12.9%，但仍有约八分之一的漏判。该工作由 @SRSchmidgall 团队完成，展示了模拟交互对临床问诊能力提升的作用。
Google DeepMind trained a model to run psychiatric intake interviews entirely in simulation, then tested it with 32 psychiatrists playing patients.
Against its identically prompted base model, it won 76.7% of 60 matched cases, rated blind by a separate psychiatrist panel. Appropriate differential diagnoses roughly doubled, from 33.3% to 68.3%.
The safety numbers cut both ways: under-triage of acute risk in adversarial scenarios fell from 30.5% to 12.9%. A big improvement, and still roughly one in eight.
Interesting work from @SRSchmidgall and the team on how simulated interactions can improve clinical interviewing.
