人工智能心理健康问题
大家好，
我想征求大家对一个“自学习”人工智能工具在心理健康领域应用的看法。
面部表情与抑郁症相关。基于小型数据集（来自少量参与者的4分钟面部视频）的机器学习模型在预测患者抑郁评分方面的表现优于医生的评分。其他与智能手机兼容的生物标志物，如瞳孔测量、眼动、音频和运动学，也与抑郁症相关。目前有几家公司销售使用这些平台的人工智能工具，但我不知道有任何大规模的多模态模型。
精神病学对主观数据的依赖削弱了其可靠性和信任度。尽管人工智能评分是基于主观数据训练的，但它们是客观的，能够解决这一问题。
如果一个非营利组织：
(a) 推出一个免费的智能手机应用程序，实时生成基于面部表情的抑郁评分。
(b) 定期要求用户填写PHQ8（8题抑郁调查问卷），例如每第四次使用时，并存储PHQ和视频数据。
(c) 利用这些数据扩展模型训练数据库；后续使用应能找到更稳健的模型。
(d) 扩展到其他数据模式，并考虑年龄、性别、种族、文化等因素。
(e) 从一个主动应用（“1分钟内获取您的评分”）演变为一个被动应用（“在用户同意下保持运行，以便随时间跟踪心理健康”）。
(f) 实现盈利：
a. 保护用户数据，保持应用对数据捐赠者免费。
b. 将算法授权给医疗专业人士。
c. 在初始慈善启动期后，将收益再投资于资金维护和开发。
面部表情模型的概念验证在学术界已有，但需要为智能手机重新开发。编码、建模和心理健康系统设计仍面临挑战。非营利结构可能对鼓励数据捐赠至关重要。
未解的问题包括：(a) 模型的保真度如何随数据规模变化，(b) 纵向系统的性能（随时间跟踪视频片段），(c) 可靠监测所需的最低视频时长，(d) 环境对视频捕捉的影响（例如，是否需要情感刺激的提示等）。
批评者可能会称这是一种噱头，因为用户已经在自我评估抑郁（而自我评估被用作训练的真实依据），系统也可能被伪造。然而，客观的纵向数据可能对个人有益，足够的数据可能使伪造检测成为可能。高使用率也可能通过提供客观数据来重新定义病症和评估治疗，从而改变心理健康研究。
虽然存在监管方面的担忧（例如，避免“诊断，确保知情同意”），但似乎对于早期版本的管理是可行的。
期待听到大家的想法。
查看原文
All,<p>Soliciting reactions on a “self-learning” AI tool for mental health.<p>Facial expressions correlate with depression. ML models of small datasets (4 mins. of facial video from small #s of particpants) predict patient depression scores better than doctor’s scores do. Other smart-phone-compatible biomarkers like pupillometry, eye saccades, audio, and kinematics, also correlate with depression. Several companies sell AI tools using such platforms, but I’m unaware of any massively multi-modal models.<p>Psychiatry’s reliance on subjective data undermines reliability and trust. AI scores, though trained on subjective data, are objective and could address.<p>What if a non-profit:
(a) Launched a free smart-phone app generating real-time depression score based on facial-expressions.
(b) Asked users to complete a PHQ8 (8 question depression survey) periodically (e.g. every fourth use) and store both PHQ and video data.
(c) Used this data to expand model training database; subsequent use should find a more robust model.
(d) Expanded to include other data modes and account for fators like age, sex, race, culture, etc.
(e) Evolve from an active app (“get your score in 1 minute”), to a passive one app (“leave running,, with user consent, to chart mental health over time”)
(f) Monetized:
a. Protecting user data, keeping the app free for data donors.
b. Licensed algorithms to health care professionals.
c. Reinvested proceeds of fund maintenance and development after an initial philanthropical startup period<p>Proof of concept for facial expression models exist in academia, but need redevelopment for smartphones. Coding, modeling, and mental health system design challenges remain. A non-profit structure may be critical to encourage data donations.<p>Unanswered questions include: (a) how model fidelity scales with data, (b) performance of longitudinal systems (tracking video snippets over time), (c) minimum video duration for reliable monitoring, (d) environmental impacts on video capture (e.g., need for emotionally provocative prompts, etc.)<p>Critics may call this a gimmick as, a users already self-assess depression (and self-assessment is used as the ground-truth for training) and the system could be spoofable.
However, objective, longitudinal could benefit individuals, and sufficient data might enable spoof-detection. High usage could also transform mental health research by providing objective data to redefine conditions and evaluate treatments.<p>Regulatory concerns (e.g., avoiding “diagnosis, ensuring informed consent) exist but seem management for an early version.<p>Curious to hear your thoughts.