AMA Demands Federal Liability Shield as Doctor AI Use Doubles to 66%

医生AI使用率翻倍至66%，AMA要求联邦提供免责保护 除非统一的联邦标准能够将法律责任从一线医生身上转移，并强制推行严格的真实世界验证，否则医生将抵制在临床中采用AI。 AMA领导层Margaret Lozovatsky博士与David Aizuss博士探讨了在临床AI部署中由医生主导与共同设计的迫切需求，告诫切勿重蹈早期电子病历带来沉重行政负担的覆辙。他们指出了关键的结构性障碍，包括尚未解决的医疗法律责任、算法偏见以及各类预测模型对可解释性的必然要求。在强调医生应用率和热情迅速上升的同时，他们展示了环境临床记录听写助手、病历总结和眼科图像诊断等高影响力用例。他们全程强调，AI必须作为“增强智能”发挥作用，以减轻认知倦怠并维护医患之间的人文纽带。 美国医学会（AMA）强调医疗交付的核心在于人际信任、同理心与临床判断，AI 应被定义为辅助人类能力的“增强智能”，切忌重蹈电子病历（EHR）沦为沉重行政负担的覆辙。 (0:04) 黑盒算法的不透明性与法律责任归属的模糊，是医生采纳 AI 的重大障碍；缺乏本地工作流及人群验证的模型极具风险，例如曾出现将成人死亡率预测模型错误套用于儿科患者的严重隐患。 (3:37) AMA 调研显示临床应用正快速普及，医生在临床场景中积极使用 AI 的比例从 2023 年的 33% 翻倍至 2024 年的 66%，医生的角色正向协助患者解读数据的分析伙伴转变。 (9:20) AMA 呼吁制定统一协调的联邦监管框架，避免 50 个州碎片化的监管政策扼杀医疗创新，并强调一线医生必须全流程参与协同设计，以保障训练数据的多元与公平性。 (12:38)

Leaders from the American Medical Association argue that artificial intelligence must function as physician-guided augmented intelligence rather than automated checkbox medicine, asserting that sustainable clinical adoption depends on transparent validation, equitable liability models, and protecting the human physician-patient relationship. Clinical Primacy and Avoiding Past EHR Pitfalls (0:04) Dr. Margaret Lozovatsky and Dr. David Aizuss contend that healthcare delivery fundamentally depends on human relationships, clinical judgment, and empathy, rejecting industry visions where care is reduced to automated checkbox selections. Dr. Aizuss warns that deploying AI without frontline physician guidance risks repeating the missteps of electronic health record rollouts, which degraded into exhausting administrative burdens for practitioners. To prevent technological disruption from alienating clinicians, the AMA defines AI as "augmented intelligence" intended to enhance rather than replace physician decision-making. Patient sentiment reinforces this priority: survey data cited by the speakers indicates that while patients consult digital tools, they consistently demand that a human physician remain at the helm of their clinical care. Algorithmic Accountability, Validation, and Legal Exposure (3:37) A critical barrier to clinical AI adoption is unresolved medical-legal liability. Dr. Aizuss highlights the legal ambiguity that arises when an algorithm recommends an inappropriate or harmful treatment: current frameworks fail to delineate whether liability falls on the bedside physician, the deploying healthcare system, or the third-party software developers. Consequently, physicians cannot accept opaque black-box models without full visibility into training data, equity audits, and clinical validation. Dr. Lozovatsky emphasizes the real-world hazards of deploying unvalidated predictive tools, pointing to early sepsis alert failures and a personal clinical case involving a stable nine-month-old infant admitted for routine respiratory syncytial virus. The hospital's AI system generated an erroneous 89% mortality risk score because the model had never been tested or validated on pediatric patients under age 18. She argues that preventing such failures requires rigorous local workflow testing, cross-disciplinary collaboration, and physician pilot feedback prior to production release. Evolving Professional Identity and Adoption Metrics (9:20) Addressing the profession's trajectory, Dr. Aizuss argues that while AI will not replace clinicians, physicians who use AI will inevitably replace those who do not. With patients regularly arriving armed with self-directed research from "Dr. Google," the clinician's role has shifted from acting as an exclusive repository of medical facts to serving as an analytical interpreter who guides patients through complex data. AMA technology surveys document a significant cultural shift toward digital integration. Active physician AI usage across clinical settings doubled from 33% in 2023 to 66% in 2024, while the share of physicians refusing to adopt AI dropped by half. Additionally, 68% of surveyed physicians recognize tangible patient care advantages (a 5 percentage point increase), and 36% describe themselves as more excited than concerned about AI's clinical future. Federal Standards and AMA Governance Initiatives (12:38) Evaluating the federal government's "America's AI Action Plan," Dr. Aizuss views the framework as a constructive baseline, stressing the necessity of a coordinated national standard to prevent an unworkable patchwork of 50 distinct state regulations. He argues that federal policy must mandate physician co-design, establish liability protections for clinicians using third-party tools, and eliminate algorithmic bias by requiring representative training data. To institutionalize physician leadership across the technology lifecycle, the AMA convenes the AMA AI Collaborative with over 40 specialty societies to author clinical model cards and consensus standards. The association also incubates startups through Health2047, participates in coalitions like the Health AI Partnership and VALID AI, delivers continuing medical education modules on interpreting algorithmic outputs, and distributes an AI Governance Toolkit for health systems. Frontline Clinical Opportunities and Specialty Friction (23:42) At the bedside, ambient listening tools and automated chart summarization are providing urgent relief for clinical burnout, enabling physicians to complete documentation on time. Beyond clerical tasks, image-analysis algorithms—such as optical coherence tomography tracking for diabetic retinopathy and macular degeneration—streamline diagnostic monitoring, while conversational AI tools support pre-operative counseling to reinforce instructions for anxious surgical patients. However, Dr. Aizuss points to sharp friction between commercial software models and high-volume specialty practices. While ambient documentation adapts well to unstructured primary care encounters, complex specialty workflows remain poorly supported. He notes that vendors have deployed unrefined specialty tools despite clinician objections, compounding technical misalignments with rigid pricing structures—such as software caps of 200 patients per month—that high-volume surgical practices seeing 60 or more patients daily exhaust in days. While physician adoption of clinical AI has accelerated rapidly, the transition to safe implementation remains contingent on clarifying medical liability, aligning commercial software with complex specialty workflows, and establishing uniform federal standards that prevent unvalidated algorithms from entering frontline care.

美国医学会（AMA）领导层在 Ai4 会议上主张，医疗人工智能必须定位为“增强智能”（Augmented Intelligence）而非取代医生的黑盒系统，其核心使命是在减轻行政与认知负担的同时捍卫以人为本的医患关系。 吸取电子病历教训与避免黑盒责任困境 (0:04) AMA 理事会主席 David Aizuss 医生与数字健康创新副总裁 Margaret Lozovatsky 医生指出，AI 的推广绝不能重蹈电子病历（EHR）因缺乏一线临床医生深度参与而沦为繁重行政枷锁的覆辙。医疗不能退化为单纯勾选复选框并直接开药的机械流程，临床决策必须依赖人际信任与医生直觉。调查表明，即便患者参考 AI 意见，仍强烈要求人类医生主导治疗。与此同时，AI 的可解释性与医疗法律责任归属成为阻碍推广的制度瓶颈：当算法缺陷导致误诊或患者伤害时，法律尚未厘清责任应由主治医生、部署医院、算法开发商还是强制推行系统的医疗机构承担。如果无法打破“黑盒”模型并建立透明的模型训练、验证与安全审计标准，潜在的法律风险将严重挫伤临床医生的采纳意愿。 脱离实际情境的预测失败与落地风险 (5:10) Lozovatsky 医生以早期脓毒症预警算法和个人临床经历为例，强调模型若缺乏特定临床情境的严格验证将带来极大危害。脓毒症模型的失效根源在于训练采用的回顾性数据集未反映真实本地工作流和患者特征。更极端的案例发生在一名前往儿科病房住院的九个月大呼吸道合胞病毒（RSV）患儿身上：该患儿仅需半升辅助氧气，病情稳定且次日即可出院，病历系统却依据 AI 死亡率指数给出了惊人的“89% 死亡风险”。事后调查发现，该模型从未在 18 岁以下儿童群体中验证，却被草率推向儿科临床生产环境，此类指标若流入患者门户或误导医生，将引发严重恐慌与误判。相比之下，唯有数据科学家、临床信息学专家与一线医生展开跨学科合作，并在小规模受控试点中根据反馈动态校准模型，才能实现技术的安全落地。 医生角色重塑与临床采纳意愿翻倍 (9:20) 面对患者广泛查阅“谷歌医生”（Dr. Google）并带着先入为主的错误诊断就医的现状，Aizuss 医生认为，医生作为临床知识绝对垄断者的传统职业形象已不复存在。现代医生的定位必须转变为分析伙伴与信息解释者，引导患者理解复杂数据以建立信任。他强调，AI 不会淘汰医生，但善用 AI 的医生必然会取代抗拒技术的医生。AMA 的技术采纳调查数据印证了这一观念转折：68% 的受访医生认可 AI 在直接患者护理中的切实力度，年增长达 5 个百分点；36% 的医生对 AI 的潜力感到兴奋多于担忧；在临床场景中实际使用 AI 的医生比例在一年内翻倍，从 2023 年的 33% 跃升至 2024 年的 66%；明确拒绝使用任何 AI 的医生比例更直接减半。临床细微洞察和同理心无法从静态数据中提炼，但合规合伦理的 AI 工具正日益被视为不可或缺的诊疗辅助。 环境聆听与病历归纳：缓解行政过载的双重支柱 (15:52) AMA 将 AI 称为增强智能，核心目的在于辅助人类判断而非将医生边缘化。当前最高价值的应用聚焦于两大方向：行政减负与病历归纳。在行政减负方面，针对临床文书记录、预先授权（Prior Authorization）及电子收件箱的环境聆听（Ambient Listening）工具展现了巨大的情绪价值。许多首席医疗信息官（CMIO）反馈，自动文书记录让医生无需在深夜“加班写病历”，甚至激动到在医院走廊相拥而泣。第二个应用前沿是智能病历摘要，面对医疗数字化带来的海量、碎片化数据，自动提取纵向病史并生成实时临床洞察，能有效减轻医生的信息整合压力，从而将注意力重新聚焦在患者本身与临床判断上。 国家治理框架、专科实际断层与未来契机 (12:38) 在政策层面，Aizuss 医生评估了联邦政府新出台的《美国 AI 行动计划》，认为由联邦主导建立统一协调的标准框架至关重要，必须避免 50 个州各自为政的碎片化监管扼杀创新，同时要确保底层训练数据具有代表性以消除算法偏见，并建立公正的责任分担机制。在产业对接上，当前市场产品暴露出全科与专科工作流的严重脱节。环境聆听在非结构化的全科记录中表现优异，但在日接诊量超过 60 人的高负荷专科（如眼科手术门诊）中，缺乏针对性调优且附带脱离实际的用量定价上限（例如每月 200 名患者），导致仓促上线的工具屡遭抱怨。除文书外，影像辅助诊断（如英国开发的视网膜 OCT 扫描自动对比算法追踪黄斑变性和糖尿病视网膜病变）、术前缓解焦虑的交互式宣教工具以及实时双向多语言翻译，构成了未来真正造福一线医疗的广阔机遇。 AMA 正通过联合 40 多个专科医学会、运营 Health2047 孵化器、制定模型卡（Model Cards）及发布《AI 治理工具包》等举措将医生话语权置于技术生命周期核心。然而，如何促使软件供应商深入理解专科真实工作流、跨越商业定价与临床负荷的脱节，并在国家层面建立透明的算法责任与偏见监管，仍是医疗 AI 全面普及必须攻克的制度与技术挑战。

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 28 日发布，记录了美国医学会（AMA）数字健康创新副总裁 Margaret Lozovatsky 博士与 AMA 理事会主席 David Aizuss 博士展开的炉边对谈。根据页面简介，对谈围绕人工智能对临床决策、医患关系、医学教育及专业认同的影响展开，探讨医生如何作为负责任的引领者，确保技术赋能而非替代医学的艺术与科学。 ## TL;DR 两位 AMA 领袖在对话中指出，医疗 AI 应当被定义为“增强智能”（Augmented Intelligence），其核心在于协助而非取代医生的专业临床判断。他们吸取以往电子病历（EHR）推行中的教训，呼吁在 AI 研发全生命周期中深度引入一线医生参与，并解决责任归属、数据偏见与碎片化监管等关键问题。随着全美医生对 AI 的接纳度快速上升，AI 在环境拾音病历生成、海量数据与影像辅助诊断、术前宣教及多语言翻译等场景展现出巨大潜力，有望大幅减轻行政负担并让医生重回有温度的医患沟通。 ## Atmosphere & style 整场对话气氛融洽、务实且充满专业洞见，两位身处一线与管理前沿的资深医师展现了极具默契的互动。讨论既不乏对技术脱离临床实际、模型错误外推等现实风险的犀利剖析与幽默自嘲，又洋溢着对技术赋能医疗本质的积极期待。听众能真切感受到医学界在拥抱智能化转型时的审慎与担当，情感基调从对过往技术阵痛的警醒逐步走向对重塑医患关怀的坚定信心。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=4s) 技术不能取代临床本质，医疗照护的核心基石永远是不可替代的人性关怀、医患信任以及专业临床判断。 - [01:52](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=112s) 吸取电子病历（EHR）因缺乏一线医生深度参与而沦为繁重行政负担的教训，医疗 AI 必须在安全可靠、无缝融入工作流的前提下推进。 - [01:52](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=112s) AI 算法因医疗差错引发纠纷时的法律责任认定存在灰色地带，亟需提升算法训练数据与临床效用的透明度与可解释性。 - [05:10](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=310s) 医疗算法必须在本地具体临床场景中充分适配调优，若缺乏未成年人验证直接推向儿科会导致灾难性的死亡风险误报。 - [05:10](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=310s) 技术落地的“最后一公里”需要跨学科团队紧密协作，通过挑选骨干医生在真实临床场景进行小范围试点与闭环迭代。 - [09:39](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=579s) AI 不会取代医生，但善用 AI 的医生将取代拒绝 AI 的医生；医学界对 AI 的临床使用率与正面接纳度正呈现翻倍式增长。 - [11:21](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=681s) 在“谷歌医生”普及的背景下，现代医生不再仅是知识垄断者，而是善用数字工具分析庞大数据并向患者协作释疑的引导者。 - [12:42](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=762s) 支持建立联邦统一的医疗 AI 监管框架以避免各州碎片化，并要求在技术研发与全生命周期中始终保持临床医生的深度主导与合理责任分担。 - [15:52](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=952s) AMA 将 AI 明确定义为“增强智能”（Augmented Intelligence），旨在通过环境拾音监听等技术减轻行政泥潭并智能汇总病历数据。 - [19:25](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1165s) AMA 联合 40 多个专科医学会成立 AI 协作体制定专科“模型卡”，并通过产业孵化与贯穿全周期的医学教育构建行业治理话语权。 - [22:41](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1361s) 发布《AI 治理工具包》指导医疗机构系统化搭建基础设施，并倡导医学专业组织与科技产业界开展紧密跨界合作。 - [24:03](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1443s) 环境感知听力技术在全科表现优异，但在眼科等高度依赖结构化病历模板的专科仍面临成熟度不足与商业化计费错位的双重挑战。 - [25:47](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1547s) AI 在眼科序列 OCT 影像判读、交互式术前知情宣教以及跨语种实时双向口译等场景展现出巨大的赋能潜力与实用价值。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=4s) 开场介绍与医疗核心价值：技术与人性的平衡 对谈在热烈的开场氛围中展开。美国医学会（AMA）数字健康创新副总裁 Margaret Lozovatsky 博士首先感谢现场观众在临近傍晚、距离欢乐时光（Happy Hour）仅一步之遥的时刻依然坚守会场。她开门见山地指出，当今人工智能的发展速度空前迅猛，从减轻行政审批负担、推动前沿疗法研发，到疾病诊断辅助以及改善患者沟通互动，AI 带来的机遇几乎是无限的。 然而，Lozovatsky 博士着重强调，技术本身并不直接等同于临床医疗的本质。真正维系医疗照护能力的核心，永远是不可替代的人性关怀、医患之间的信任关系，以及沉淀在实践中的专业临床判断。作为一名至今仍在一线查房的儿科住院医生，Lozovatsky 博士与 AMA 理事会主席、资深眼科临床执业医师 David Aizuss 博士展开对话，共同探讨面对滚滚而来的 AI 浪潮，临床医生应当如何保持清醒与审慎，确保前沿技术能够真正服务于有温度的医疗。 ```mermaid flowchart TD accTitle: 医疗人工智能的核心原则架构 subgraph CoreValues["医疗实践核心"] Humanity["人性关怀与医患信任"] Judgment["医生专业临床判断"] end subgraph AIOpportunities["AI 带来的机遇"] Admin["减轻行政与文书负担"] Diag["辅助疾病诊断与前沿疗法"] Engagement["改善患者沟通互动"] end subgraph EssentialPillars["AMA AI 发展指导原则"] P1["安全可靠、无缝嵌入临床工作流"] P2["以患者为中心，确保透明度与可解释性"] end AIOpportunities --> EssentialPillars CoreValues --> EssentialPillars ``` ### [01:52](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=112s) 电子病历的前车之鉴与医疗责任的法律盲区 当被问及“关于医疗 AI，究竟是什么让你夜不能寐”时，Aizuss 博士幽默而犀利地回应，最让他彻夜难眠的，是前一位戴着牛仔帽的演讲嘉宾声称的“未来患者只需在屏幕上勾选几个复选框就能开药拿药”。他直言，这绝不是医学的执业方式，即便在高度发达的 AI 时代，医疗也永远不该退化为流水线式的打勾发药。患者调研同样印证了这一点——患者渴望的是由真实医生主导诊疗，而非向冷冰冰的机器托付生命。 Aizuss 博士将眼下的 AI 热潮与多年前电子病历（EHR）的推行进行了深刻对比。EHR 曾经也是一项初衷极好的技术，但由于开发过程中严重缺乏临床一线医生的深度参与，最终演变成压在医生肩头的一座繁琐行政大山。为了避免重蹈覆辙，AMA 确立了两大核心指导原则：其一，AI 工具必须值得信赖、高度可靠，且能无缝融入日常临床工作流；其二，必须始终将人性置于医疗的核心，医生必须主导建立安全护栏（Guardrails）与监管框架。 更为紧迫且常被行业忽视的隐患是法律责任（Liability）的归属。Aizuss 博士提出了一系列发人深省的现实拷问：当 AI 工具根据患者个人健康数据给出诊疗建议，而该建议最终被证实错误并导致患者受到伤害时，法律与道德责任究竟该由谁承担？ - 听从建议并开具处方的临床医生？ - 负责部署该算法模型的机构或医院？ - 设计与编写算法的技术工程师与研发团队？ - 还是强制要求医护人员使用该 AI 系统的医疗卫生体系？ Aizuss 博士指出，如果缺乏明确的透明度与可解释性（Explainability）——包括算法是如何训练的、使用了哪些验证数据集、其准确度、公平性与临床效用如何评估——这种责任认定的灰色地带将成为阻碍 AI 在医疗领域广泛落地的巨大绊脚石。 ### [05:10](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=310s) 临床实战案例：算法失灵、儿科死亡风险误报与跨学科协作经验 紧接 Aizuss 博士的见解，Lozovatsky 博士结合自己以往担任首席医疗信息官（CMIO）的实战经历，剖析了医疗算法在现实部署中的成败得失。早在 ChatGPT 轰动全球数十年前，医疗界就已开始探索脓毒症（Sepsis）预警算法。业界常常将过去的挫折简单归咎于“算法有缺陷、给出了错误建议”，但 Lozovatsky 博士指出，深层原因往往在于这些算法是用陈旧或不精准的数据集训练的，并且从未在具体的本地临床环境中进行充分的原位测试与适配调优。 随后，Lozovatsky 博士分享了一个令全场动容且后怕的亲身经历：在某年冬季的儿科病房，她接诊了一名因呼吸道合胞病毒（RSV）感染入院的 9 个月大婴儿。当时患儿仅需每分钟半升的低流量吸氧，各项生命体征平稳，预计次日即可停止吸氧顺利出院——这在冬季儿科病房属于最典型的常规轻症病例。然而，当她打开系统病历时，屏幕上的死亡指数预测模型赫然显示该婴儿的“死亡风险高达 89%”。 震惊之余，她立即联合技术团队展开排查，调查结果令人不寒而栗：该预测算法在设计和训练时完全基于成年人数据，根本从未在 18 岁以下的未成年人群中进行过验证，却被错误地直接推向了儿科临床生产环境。假使医生盲目信任算法而做出激进干预，或者该数据未经审核直接推送到患儿家属的移动端门户，将会引发难以想象的恐慌与灾难性后果。 与此形成鲜明对比的是 Lozovatsky 博士主导的一套成功推行模式。她强调，技术走向床旁的“最后一公里”必须依赖机制保障： - **跨学科紧密协同**：建立认知计算团队、数据分析团队、临床医生以及临床信息学专家（Clinician Informaticists）深度绑定的协作机制。 - **小范围试点验证**：率先挑选懂技术、具备信息学思维的骨干医生组成先锋测试小组，在真实场景下试用并高频反馈。 - **闭环迭代与全员推广**：基于前线反馈和运行数据反复修正模型，确信其安全有效后，再推广至全院医疗系统。这种前置铺垫让全院医生感到 AI 是真正帮得上忙的临床助手，而非徒增烦扰的系统累赘。 ### [09:39](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=579s) 医生的角色演进与态度剧变：拥抱技术而非被替代 谈到医生角色的未来演变，Aizuss 博士给出了掷地有声的研判：AI 绝不会取代医生，但未来的分水岭在于——“使用 AI 的医生，将会取代那些拒绝使用 AI 的医生”。那些故步自封、固守传统工作流的专业人员，终将被时代的浪潮所淘汰，这一规律在任何行业都同样适用。 实现 AI 广泛应用的前提，是同时赢得医生与患者的双向信任。Aizuss 博士援引了 AMA 针对全美医生技术采纳情况的最新权威调研数据，指出一向被视作技术保守群体的临床医师们，态度正在发生剧烈的积极转变： - **临床价值认同提升**：68% 的受访医生表示看到了 AI 在改善患者照护方面的显著优势，该比例在短短两三年内上升了 5 个百分点。 - **正面期待压倒担忧**：36% 的医生明确表示，相比于潜在风险，他们对 AI 带来的机遇“感到更加兴奋而非担忧”，这对于习惯规避风险的医生群体而言是一个极其罕见的信号。 - **临床使用率成倍暴增**：在各类多元医疗场景中，实际使用某种形式 AI 的医生比例从 2023 年的大约 33% 翻倍飙升至 2024 年的 66%。 - **排斥心理急剧消退**：明确表态“绝不使用任何形式 AI”的医生比例，在短短一年之内直接腰斩。 这一系列详实的数据表明，医学界对待 AI 的态度已经走过了最初的怀疑与抵触阶段。AI 不仅不再是洪水猛兽，而且正在深刻重塑新一代临床医生的专业认同、执业模式以及他们对未来医疗职责的自我期许。 ### [11:21](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=681s) 诊室中的“谷歌医生”与现代医生角色的演变 随着互联网信息的普及，医生作为“唯一临床知识来源”的传统身份正面临巨大冲击。大卫·艾祖斯（David Aizuss）医生指出，在如今“谷歌医生”（Dr. Google）普及的时代，几乎很少有患者在走进诊室前不提前搜索并预判自己的诊断结果。尽管多数情况下患者自行推导出的诊断并不准确，但医生的沟通方式已经发生转变——不再是简单否定，而是以协作的姿态发问：“您是想听谷歌医生的，还是想听我的？不如我们一起来看这些数据，好让我向您讲清楚我的依据，建立您的信心。”现代医生的核心能力不再局限于死记硬背医学知识，而是善于利用能够分析庞大数据集、辅助诊断并预测疗效的全新数字化工具。 与此同时，纯数据本身无法完全替代实际医疗场景中的复杂现实。理解临床现场环境、患者个体化诉求以及数据难以捕捉的细微关怀，依然是医生不可替代的价值所在。只有当 AI 工具的开发与部署能够紧密契合临床实用性、恪守医学伦理并真正以患者为中心时，医学界才能建立起对其应用的信心，广大临床医生也才会真正敞开怀抱接纳这些技术。 --- ### [12:42](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=762s) 联邦与州监管：AMA 对白宫人工智能行动计划的审视 谈及监管政策环境，对话转向了特朗普政府近期发布的《美国人工智能行动计划》（America's AI Action Plan）。艾祖斯医生坦言，尽管美国医学会（AMA）过去对该政府（特别是卫生与公众服务部 HHS 的领导层）提出过诸多批评，但他们认为这份旨在确立美国 AI 领导地位的路线图是一个“良好的第一步”。在关乎患者安全的极高风险领域，建立合理的监管框架以规避潜在风险并最大化发挥技术潜力至关重要。该计划在建立公众与专业信任、伦理监督以及加速推动安全性、性能与互操作性国家标准方面做出了积极布局。 在监管架构的设计上，艾祖斯强调了避免碎片化监管的紧迫性。虽然他个人并不支持大包大揽的法案，但他赞同此前特德·克鲁兹（Ted Cruz）参议员提出设立“国家统一监管框架”的诉求。如果全美 50 个州各自为政、自行出台 AI 监管细则，将严重阻碍技术的演进与落地；因此，必须由联邦与各州协同建立统一协调的监管体系。此外，AMA 强调了以下关键维度： - **全生命周期的医生参与**：临床工作流程的真实复杂性只有一线医生最清楚，因此在 AI 研发的“第一天起”以及全生命周期的每个阶段都必须有强有力的医生代表参与。 - **全政府协同与数据去偏见**：确保训练数据集消除偏见是建立临床信任的基石。尽管现任政府不愿过多提及“公平、多元与包容”（EDI），但数据的代表性必须成为伦理 AI 战略的核心基石。 - **合理的责任划分机制**：必须明确技术供应商与使用端之间的法律责任划分，不能让医生在部署新技术后独自承担不可预见的系统缺陷责任，从而消除医生的后顾之忧。 ```mermaid flowchart TD accTitle: 医疗 AI 治理与临床落地核心支撑框架 accDescription: 展示从联邦协同监管、全生命周期医生参与到伦理责任划分的医疗 AI 治理闭环 subgraph 顶层监管与标准 ["顶层监管与标准"] A["联邦统一监管框架<br/>(避免50州碎片化)"] B["国家统一标准<br/>(安全/性能/互操作性)"] end subgraph 临床生命周期介入 ["临床生命周期介入"] C["研发源头介入"] --> D["临床流程适配"] D --> E["部署与持续优化"] F["一线医生全流程主导"] -.-> C F -.-> D F -.-> E end subgraph 信任与保障机制 ["信任与保障机制"] G["数据集无偏见与多元代表性"] H["合理的责任划分机制 (免除医生孤立担责)"] end 顶层监管与标准 --> 临床生命周期介入 临床生命周期介入 --> 信任与保障机制 ``` --- ### [15:52](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=952s) “增强智能”赋能临床：减轻行政重负与病历汇总突破 玛格丽特·洛佐瓦茨基（Margaret Lozovatsky）从 AMA 首席医疗信息官（CMIO）的视角阐述了 AI 的定位。她强调 AMA 始终将 AI 定义为**“增强智能”（Augmented Intelligence）**，而非人工智能（Artificial Intelligence），这意味着技术应扮演“临床副驾驶”（Co-pilot）的角色来增强而非取代医生的临床决策。在临床实践中，她认为最具变革潜力的两大方向分别是减轻行政负担与数据智能汇总。 在减轻行政负担方面，临床医生正深陷文书记录、电子收件箱（In-basket）回复和保险预先授权（Prior Auth）等繁冗行政流程的泥潭。其中，环境拾音监听技术（Ambient Listening）的落地带来了立竿见影的效果。许多 CMIO 同行反馈，把该技术引入医院后，医生们激动地在走廊里拥抱他们甚至喜极而泣，因为他们终于能按时回家与家人共进晚餐，而不必在下班后把家庭时间全部耗费在整理病历上。尽管基于 AI 的收件箱自动回复功能目前更受前台行政及护理团队欢迎，但也已大幅缩短了患者获得答复的等待时间。 在数据汇总与洞察方面，随着电子病历的普及，临床端产生了海量的数据碎片，医生每天需要在大脑中强行记忆和追踪患者数年来的病程趋势。AI 正好可以发挥计算机最擅长的数据计算与归纳能力，将庞杂的医疗记录精炼提炼后呈现给医生，从而释放医生的临床判断力，专注于改善患者的最终预后。 --- ### [19:25](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1165s) AMA 全方位战略：标准制定、产业孵化与医学教育 AMA 正在通过多维度的布局全方位提升医生在 AI 发展中的领导地位。早在 2018 年，AMA 就率先推出了第一版《人工智能原则》（AMA Principles），并在后续持续保持动态更新，确保所有技术演进均以医生的真实诉求为中心，贯穿从早期初创企业对接、临床集成、落地应用到最终优化的完整链条。 为构建统一的行业话语权，AMA 成立了由 40 多个专科医学会组成的 **AMA AI 协作体（AMA AI Collaborative）**，致力于发布代表全体医生的统一声明，并针对不同专科的独特需求制定专属的“模型卡”（Model Cards）。此外，AMA 代表大会设立了专门的 AI 专家工作组，持续审视组织架构与支持赋能工具。在生态与教育层面，AMA 的核心布局包括： - **产业孵化（Health2047）**：AMA 旗下的创新孵化器 Health2047 坚持让医生深度参与技术底层开发，已成功孵化出多家医疗科技企业。 - **跨机构伦理联盟**：主导成立“健康 AI 伙伴关系”（Health AI Partnership），并深度参与 VALID AI 协作体、HIMSS、AMIA（美国医学信息学学会）以及 AMDIS（应用医学信息学学主任协会）等多方合作。 - **全周期医学教育培训**：面向医学生（本科前教育）、住院医规培（毕业后医学教育）以及执业医生（继续医学教育 CME）开发了系统化培训课程，帮助医生掌握解读 AI 输出结果的核心概念，并提供配套的 CME 学分。 ### [22:41](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1361s) AMA 的 AI 治理工具包与跨界合作 Margaret Lozovatsky 总结了美国医学会（AMA）在推动医疗机构 AI 落地方面开展的各项工作。AMA 组织医疗系统共同参与为期一年的协作项目，携手探索并在各机构内部搭建 AI 基础设施，这一模式获得了极为积极的反响。与此同时，AMA 在今年正式发布了《AI 治理工具包》（AI Governance Toolkit），为各类医疗机构和诊所提供了一套循序渐进的指导方针，指导管理者与临床医生如何系统性思考基础设施建设，从而确保 AI 工具的顺利实施。 ```mermaid flowchart TD accTitle: AMA 推进医疗 AI 实施与治理路径 subgraph Governance["治理与生态合作"] A["AMA AI 治理工具包<br/>AI Governance Toolkit"] --> B["医疗机构与诊所基础设施建设"] C["跨界生态合作<br/>技术专家 / 产业界 / 临床医生"] --> D["审慎推动技术走向临床床旁"] end subgraph Impact["临床赋能目标"] B --> E["减轻行政负担"] D --> E E --> F["重塑医患信任与面对面交流"] end ``` Lozovatsky 强调，在开发这些工具以及规划支持医生的培训项目时，AMA 深刻认识到跨领域合作的必要性。医疗专业组织必须主动走进技术与产业会议，与科技专家和产业界同行面对面深入对话，通过多方协作，审慎而负责任地将前沿技术真正带到患者的病床旁。随后，她向 AMA 理事会主席 David Aizuss 医生提出了本场对谈的收尾问题：立足于临床前沿展望未来，医疗 AI 最令人兴奋的潜力和前景究竟是什么？ - **AI 治理工具包**：提供步骤清晰的落地指南，帮助医疗机构与基层诊所评估基础设施需求与管理规范。 - **产业协同理念**：强调医学界不能闭门造车，必须与技术人员和产业研发端深度结盟。 --- ### [24:03](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1443s) 赋能医患关系：环境感知听力技术与专科落地挑战 作为一名日常接诊量极大的临床专科医生，David Aizuss 医生指出，AI 最令人振奋的核心价值在于它能够挽救并改善医患关系。当前繁琐的行政事务极大地挤占了医生的精力，如果 AI 能够接管病历撰写、预约排班、分诊等重复性行政流程，医生就能将宝贵的时间重新投入到与患者面对面的有温度的沟通中。然而，他也客观指出了当前技术在专科落地时面临的成熟度落差。 以大热的“环境感知听力”（Ambient Listening）自动生成病历技术为例，Aizuss 医生分享了他身边的真实案例：他的妹夫是一家大型医疗系统旗下的初级保健医生（PCP），对于病历格式相对非结构化的全科医生而言，这项技术堪称福音（"a boon"）。但对于眼科等依赖高度结构化病历模板的专科医生，现有产品还远远没有达到成熟水平。他所使用的电子病历（EHR）厂商曾在发布前两周向他演示其环境感知听力模块，他当时便直言“这还没准备好推向市场”（"This is not ready for prime time."）。遗憾的是，软件公司依然强行上线，导致用户群组中投诉爆棚。此外，该产品按“最多 200 名患者”设定的阶梯收费模式也脱离了高通量专科诊所的实际，诊所一天的就诊人数就能耗尽一整个月的额度。尽管痛点犹存，自诩为“技术极客”（"technology nerd"）的 Aizuss 医生表示，只要产品成熟且机制合理，他非常乐意成为早期采用者。 - **行政自动化诉求**：涵盖病历记录、智能预约、患者分诊，核心目标是夺回面对面的临床接诊时间。 - **全科与专科的代差**：全科医生的非结构化记录更容易适配环境听力生成，而专科的高度结构化记录仍需更精细的模型适配。 - **商业化与定价错位**：部分厂商按固定病患数量定价的收费模式，无法适应日均门诊量数十乃至上百名患者的高负荷诊所。 --- ### [25:47](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1547s) 影像决策辅助、术前沟通革新与多语言翻译前景 谈及临床决策支持，Aizuss 医生重点展望了医学影像领域的突破。与放射科、病理科和皮肤科一样，眼科也是图像密集型学科。例如英国开发的 AI 架构已能对连续的光学相干断层扫描（OCT，用于检测视网膜各层结构）进行序列化比对分析，精准辅助诊断黄斑变性与糖尿病视网膜病变。一旦这些算法下沉到私立私人诊所，将彻底把医生从逐张翻阅海量影像的繁重机械劳动（"grunt work"）中解放出来，由 AI 自动解析并即时提示病变演变，直接辅助临床决策。 ```mermaid flowchart LR accTitle: AI 赋能眼科临床与医患沟通场景 subgraph Imaging["智能影像决策"] I1["序列 OCT 影像"] --> I2["AI 自动比对解析"] I2 --> I3["黄斑变性 / 糖网病变即时预警"] end subgraph Communication["沟通与宣教支持"] C1["术前知情同意"] --> C2["AI 多模态重复宣教"] C3["多语种患者对话"] --> C4["AI 实时同声翻译"] end ``` 另一个充满潜力的应用方向是改善医患沟通与术前宣教。Aizuss 医生坦言，患者在得知自己需要手术的瞬间，往往会陷入焦虑、紧张与恐惧之中，医生在诊室里讲解的专业术语大部分都会被患者抛在脑后。哪怕他配备了现成的文字手册并反复叮嘱患者回家阅读，效果依然有限，以至于诊所不得不全职聘请三名手术宣教咨询员专门向患者反复解释手术细节。如果 AI 能够以更通俗易懂、具有交互性且不知疲倦的方式为患者重复讲解复杂的医疗信息，将极大地提升宣教质量。同时，针对多文化、多语种的患者群体，能够进行实时语音互译的 AI 翻译工具也将消除语言壁垒。 - **影像智能判读**：英国开发的序列 OCT 影像分析算法可自动检测视网膜层次变化，高效辅助黄斑变性与糖尿病视网膜病变诊断。 - **破解术前宣教困境**：利用 AI 化解患者面对手术通知时的焦虑与信息过载，分担专职宣教人员重复沟通的重负。 - **实时跨语言交流**：多语种环境下，AI 实时双向口译能够显著提升多元文化患者的就医体验与知情权保障。 ## Notable quotes > [01:52](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=112s) 未来患者只需在屏幕上勾选几个复选框就能开药拿药 — David Aizuss > [05:10](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=310s) 死亡风险高达 89% — Margaret Lozovatsky > [09:39](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=579s) 使用 AI 的医生，将会取代那些拒绝使用 AI 的医生 — David Aizuss > [09:39](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=579s) 感到更加兴奋而非担忧 — David Aizuss > [11:21](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=681s) 您是想听谷歌医生的，还是想听我的？不如我们一起来看这些数据，好让我向您讲清楚我的依据，建立您的信心。 — David Aizuss > [12:42](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=762s) 良好的第一步 — David Aizuss > [12:42](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=762s) 第一天起 — David Aizuss > [24:03](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1443s) This is not ready for prime time. — David Aizuss > [24:03](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1443s) 技术极客 — David Aizuss > [25:47](https://www.youtube.com/watch?v=G1FlEv9RoPQ&t=1547s) 机械劳动 — David Aizuss ## Mentioned Margaret Lozovatsky, David Aizuss, 美国医学会 (AMA), 电子病历 (EHR), ChatGPT, 谷歌医生 (Dr. Google), 《美国人工智能行动计划》(America's AI Action Plan), 美国卫生与公众服务部 (HHS), 特德·克鲁兹 (Ted Cruz), 环境感知听力技术 (Ambient Listening), AMA AI 协作体 (AMA AI Collaborative), Health2047, 健康 AI 伙伴关系 (Health AI Partnership), VALID AI, HIMSS, AMIA (美国医学信息学学会), AMDIS (应用医学信息学主任协会), 《AI 治理工具包》(AI Governance Toolkit), 光学相干断层扫描 (OCT) ## Video description Margaret Lozovatsky, VP of Digital Health Innovations, American Medical Association David Aizuss, Chair, AMA Board of Trustees Abstracts As AI transforms clinical practice, what does it mean to be a physician today—and tomorrow? In this fireside chat, we’ll explore how physicians can lead, shape, and safeguard the integration of AI in health care. Join us for a candid conversation about the opportunities and challenges AI presents to clinical decision-making, physician-patient relationships, medical education, and professional identity. The speakers will reflect on the physician’s role in being proactive stewards of responsible AI—ensuring that technology enhances, rather than replaces, the art and science of medicine. [Source](https://www.youtube.com/watch?v=G1FlEv9RoPQ)

[1:52] David Aizuss: 未来患者只需在屏幕上勾选几个复选框就能开药拿药 [5:10] Margaret Lozovatsky: 死亡风险高达 89% [9:39] David Aizuss: 使用 AI 的医生，将会取代那些拒绝使用 AI 的医生 [9:39] David Aizuss: 感到更加兴奋而非担忧 [11:21] David Aizuss: 您是想听谷歌医生的，还是想听我的？不如我们一起来看这些数据，好让我向您讲清楚我的依据，建立您的信心。 [12:42] David Aizuss: 良好的第一步 [12:42] David Aizuss: 第一天起 [24:03] David Aizuss: This is not ready for prime time. [24:03] David Aizuss: 技术极客 [25:47] David Aizuss: 机械劳动

[0:04] Dr. David Aizuss: What keeps me up at night is the guy in the cowboy hat who just said that you're going to click a bunch of check boxes and get your drugs. That's not how medicine is practiced. [3:37] Dr. David Aizuss: black boxes [5:10] Dr. Margaret Lozovatsky: faulty [5:10] Dr. Margaret Lozovatsky: 89% risk of mortality [9:20] Dr. David Aizuss: Dr. Google [12:38] Dr. Margaret Lozovatsky: America's AI Action Plan [12:38] Dr. David Aizuss: one big beautiful bill [15:52] Dr. Margaret Lozovatsky: augmented intelligence [23:42] Dr. David Aizuss: not ready for prime time

[0:04] Good afternoon and thank you for [0:06] sticking with us at this late hour in [0:08] the day. I know we stand between you and [0:10] the happy hours. So, thank you. Um, [0:13] we're going to today talk about the role [0:16] of the physician when it comes to [0:19] clinical technology. I think everyone in [0:22] this room already knows that AI is [0:24] moving at an unprecedented pace and we [0:27] know that the opportunities are infinite [0:30] from administrative burdens to [0:33] therapeutics, diagnostics to patient [0:35] engagement. [0:37] And we also know that technology doesn't [0:40] necessarily define the clinical care. [0:43] The humanity, the relationships, [0:47] the clinical judgment that is imperative [0:49] to our ability to deliver care is what [0:52] we're going to talk about today. And [0:55] we're going to ask a lot of questions [0:57] about what happens to the role of the [1:00] physician. How do we utilize this [1:02] technology thoughtfully in our clinical [1:04] care? [1:06] My name is Dr. Margaret Losski. I am a [1:10] general pediatrician. I still practice [1:12] as a pediatric hospitalist and I am the [1:14] VP of digital health at the American [1:16] Medical Association. And with me, I'm [1:19] thrilled to be joined by Dr. David Asus [1:23] who is the chair of our board of [1:26] trustees at the American Medical [1:28] Association and a practicing [1:30] opthalmologist who has lots of great [1:33] insights to share with you guys today. [1:35] And so as a first question, I'm going to [1:38] ask Dr. Jus wearing his physician leader [1:43] hat to share. We know there's a lot of [1:46] opportunities with AI, but what keeps [1:48] you up at night when you think about AI [1:50] and healthcare. [1:52] >> So what keeps me up at night is the guy [1:54] in the cowboy hat who just said that [1:56] you're going to click a bunch of check [1:57] boxes and get your drugs. That's not how [2:00] medicine is practiced. And even in the [2:02] the era of AI, that's not how medicine [2:04] will be practiced. Um, what we do have, [2:08] as you all know, is technology that's [2:09] changing incredibly rapidly. And we have [2:13] a cautionary tale in the implementation [2:15] of the electronic health record where it [2:18] was a well-meaning technology that was [2:20] largely developed without physician [2:22] involvement and has turned into another [2:25] burden for physicians to manage uh in [2:28] the clinical environment. So our AMI AMA [2:31] policy around AI and all of health [2:33] technology is guided by two principles [2:36] which is critical to the future that [2:37] we're stepping into. One is that it has [2:39] to be trustworthy, reliable and can be [2:42] seamlessly integrated into patient care [2:45] and that humanity is always at the [2:47] center. You're not going to be sitting [2:49] in front of a computer clicking some [2:51] check boxes and getting drugs. Um [2:55] and when we survey patients, patients [2:57] don't want that either. they still want [2:58] physicians at the center of their care [3:01] even if AI is providing uh some role in [3:04] giving them advice. [3:07] So our view is that physician leadership [3:09] has to maintain a critical role in [3:11] establishing the guardrails and as well [3:14] as the regulatory oversight framework [3:16] and that's the role of the AMA where we [3:18] elevate the needs of our patients their [3:21] trust and their welfare over whatever [3:23] technology is at hand. Um we have [3:27] principles on AI development, deployment [3:30] and use that highlights the importance [3:33] of both liability as well as uh [3:35] transparency [3:37] and um for physicians in particular [3:41] addressing the liability issues [3:43] surrounding AMA is an surrounding um AI [3:47] is an important barrier to widespread [3:49] adoption and we don't hear liability [3:51] being discussed very often. Um there's [3:54] liability that can occur from uh [3:56] inaccurate AI models or from their [3:59] misuse. Uh and this is already becoming [4:01] an area of great interest in the medical [4:03] legal field. For example, if an AI tool [4:07] recommends a certain course of treatment [4:09] for an individual patient based on their [4:11] personal health and [snorts] it turns [4:13] out that it's an inappropriate course of [4:15] treatment due to faulty AI, who bears [4:18] that responsibility? Is it the [4:21] physician? Is it the entity that [4:23] deployed the algorithm? Is it the [4:26] individual or team who developed the AI [4:28] algorithm? Is it the health system that [4:31] has required the use of that AI tool? So [4:35] it's very critical that physicians [4:38] understand that AIdriven technologies [4:41] um have uh an algorithm behind them and [4:45] we need to have information about how [4:47] those tools are being developed, how the [4:51] algorithms are trained, what data sets [4:53] are used to validate them, how we can [4:56] assess their quality, their performance, [4:58] their equity, their utility, you know, [5:01] so transparency and explanability become [5:03] very critical in the development of AI [5:06] in healthcare. [5:10] [snorts] [5:10] >> Thank you. And I really appreciated the [5:13] examples that Dr. Azus gave and I want [5:16] to share a few more examples about [5:20] AI algorithms that have been implemented [5:22] in my experience wearing my CMIO hat um [5:26] that worked well and some that did not [5:29] work as well. And I think when we all [5:32] remember the early days of utilization [5:35] of AI tools before anybody's heard of [5:37] Chad GPT decades ago when we started [5:39] putting these tools into practice the [5:42] one that often comes to mind is the [5:44] sepsis algorithms [snorts] and I'll [5:47] share with you a little bit about my [5:49] experience with it. It's probably [5:52] somewhat different than what we often [5:54] hear. We often hear that the algorithm [5:56] was faulty and gave bad suggestions and [6:00] that was true [6:03] but we don't talk about the reasons and [6:06] a lot of the reason was not the [6:07] algorithm itself it was the fact that [6:11] these algorithms and I think everyone in [6:13] this room understands that are built on [6:16] the existing data they were not [6:18] necessarily using the most accurate data [6:20] sets and they weren't tested in the [6:23] right environments we We're not testing [6:25] our algorithms in the local environment [6:28] and the criticality of understanding the [6:31] fact that these are technology tools [6:34] that need to be then thoughtfully [6:36] brought to the clinical care that will [6:39] potentially impact the patient outcomes [6:42] is where the rubber meets the road. And [6:45] it's really important for us as we think [6:47] about the future of these tools to [6:50] understand that we need processes in [6:53] place to test these models appropriately [6:57] to ensure that we have integrated them [6:59] thoughtfully so that the guidance [7:01] provided to clinicians is the [7:03] appropriate guidance. And so that was [7:05] some of the examples of things that [7:07] didn't go well. Um I'll give you another [7:09] example. This actually happened when I [7:10] was taking care of a patient. Um, I was [7:13] seeing a nine-month old baby that was [7:15] admitted for RSV that was in about a [7:18] half liter of oxygen. Healthy baby, this [7:21] is very standard in the winter. That's [7:23] half of our pediatric hospital was going [7:26] to be weaned off the oxygen the next day [7:28] or so and going home. I opened up the [7:30] chart and our mortality index said 89% [7:34] risk of mortality [snorts] on a [7:37] perfectly healthy baby that was clearly [7:39] going to go home the next day. And of [7:41] course that raised a lot of questions [7:44] from me to our development teams. And [7:47] after investigating what happened, we [7:49] learned that this algorithm was never [7:51] meant to be put in production on [7:53] pediatric patients because it was never [7:55] tested in patients under the age of 18. [7:58] And so I want to share that because a [8:01] lot of bad things could have happened, [8:02] right? I could have made clinical [8:04] decisions based on that that could have [8:07] been released in their patient portal to [8:09] their parents that would have had quite [8:11] a reaction to that. Um, we need to [8:14] understand that these decisions aren't [8:18] trivial when we bring these tools to the [8:21] bedside and the real patients that are [8:22] impacted. Conversely, [gasps] [8:26] we rolled out a lot of predictive uh [8:28] algorithms in my prior life when I was [8:32] in a CMIO role where we had a great [8:34] partnership with our cognitive computing [8:36] team, our analytics teams and our cli [8:39] clinicians, our clinic clinician [8:40] informaticist teams where we tested the [8:43] algorithms with the clinicians. We [8:45] rolled them out with a few select [8:47] technology [8:49] um knowledgeable clinicians that were [8:51] able to then get provide feedback. We [8:54] were able to gather data on how these [8:57] algorithms were functioning before we [8:59] rolled it out to all of the clinicians [9:01] in the system. And that worked [9:03] incredibly well and we got feedback [9:05] across the board from our clinicians [9:07] that they were surprised how helpful [9:09] these tools were because we had done a [9:12] lot of the leg work ahead of time. And I [9:14] think that's part of the messaging here [9:15] is we need to plan and think about how [9:18] to implement these tools. [9:20] So I know we talked a lot about the AI [9:24] tools. I want to shift gears a little [9:25] bit and ask Dr. Aus about his vision for [9:31] how the role of the physician is going [9:34] to evolve with the integration of AI in [9:37] healthcare. [9:39] So, you heard me say earlier that you're [9:40] not going to see AI replacing [9:42] physicians. Um, but it may uh you may [9:46] find that one day there'll be doctors [9:48] who use AI and doctors who don't. And [9:50] the ones who don't may be the ones who [9:52] go by the wayside. Um, and I think you [9:55] probably make a similar statement [9:56] professionals in most industries. I [9:58] think AI is here to stay and we're going [10:01] to be utilizing it in a variety of ways. [10:03] Um, one thing we know that for AI to [10:06] work in health care, it has to have both [10:08] the trust of physicians as well as the [10:11] trust of patients. And um, as physicians [10:15] begin to see its application in [10:17] practice, we will begin to trust it. The [10:20] AMA's latest survey on technology [10:22] adoption shows that more physicians are [10:25] taking the plunge on AI than before and [10:28] that a majority of physicians are [10:29] recognizing its benefits. 68% say that [10:32] they see some advantages in patient [10:34] care. That's up 5% just from a couple [10:37] years ago. But what's more interesting [10:39] is that um 36% of physicians are [10:42] reported being more excited than [10:44] concerned about AI. And given that [10:46] physicians tend to be a relatively [10:47] conservative lot in adopting things, [10:50] that says a lot. most importantly is [10:52] that the usage of AI in a variety of [10:55] scenarios has doubled from 66% or to [10:58] 666% in 2024 from like 33% in 2023. And [11:03] the percentage of physicians say they're [11:05] not going to use some form of AI has [11:07] dropped in half from a previous survey [11:09] just a year ago. So I think that says a [11:11] lot about physician attitudes towards [11:14] AI. [11:15] Um I think it's also starting to reshape [11:19] how we think about our role and our [11:21] identity. Um it is very challenging for [11:24] a physician to be the sole source of [11:26] clinical knowledge. Especially in the [11:28] era of Dr. Google when I sit in my exam [11:30] room it is the unusual patient now who [11:32] does not come in prepared with answers [11:34] that they've already researched to what [11:36] they are presumed diagnosises. Now [11:40] much of that time their presumed [11:41] diagnosis is incorrect. And you know, I [11:44] asked them, "Do you want to listen to [11:45] Dr. Google or do you want to listen to [11:47] me?" But let's look at this together so [11:49] I can inspire your confidence in what [11:51] I'm going to tell you. And I think [11:52] that's what you're going to see [11:53] physicians doing. They're going to [11:55] recognize that we have new tools that [11:57] can analyze data sets, support [11:58] diagnosis, help with our anticipated [12:01] outcomes. I think that's what the modern [12:02] physician will be doing. Um I also think [12:06] that uh physicians who understand the [12:11] clinical environment and patient needs [12:14] and nuances of care that is not captured [12:17] by data alone are going to be critical [12:20] for application of these things and as [12:24] AI tools are developed and deployed that [12:26] in ways that are clinically relevant and [12:27] ethically sound and patient centered [12:29] we're going to have confidence in its [12:31] adoption and uh physicians will embrace [12:34] this. [12:38] >> I love that optimistic view. That's [12:40] fantastic. [12:42] Well, let's talk a little bit about the [12:44] regulatory space. The Trump [12:46] administration just released the [12:48] America's AI action plan. That was about [12:51] two weeks ago with the roadmap of the US [12:54] becoming a leader in AI. Can you share a [12:56] little bit about the AMA's view of the [12:59] plan, some opportunities and some wins? [13:02] Well, I will say that we feel that the [13:05] plan that the Trump administration [13:07] released is a good first step [snorts] [13:10] and we are not typically uncritical of [13:12] what's coming out of the Trump [13:13] administration, particularly with HHS [13:15] and the leadership there. But with [13:17] stakes as high as they are, it is very [13:20] important that we have the appropriate [13:22] regulatory structure in place that [13:24] mitigates risk to patients and that will [13:27] position ourselves to get the most out [13:29] of this technology. In particular, uh [13:32] the Trump plan paid attention to [13:34] building public and professional trust [13:36] for AI technology, which is critically [13:38] important and ethical oversight and [13:42] accelerating national standards for [13:44] safety, performance, and [13:45] interoperability. I am personally not [13:48] someone who was a fan of the one big [13:50] beautiful bill. But one aspect of that [13:54] bill that didn't make it into the bill [13:56] was Ted Cruz's desire to have a national [13:58] regulatory framework. Um we all [14:01] recognize that if every state is [14:04] regulating the implementation of AI [14:06] individually, it is going to hamper the [14:09] advancement of this technology. So we [14:12] are going to need some sort of [14:14] well-coordinated federal federal [14:15] regulatory approach in concert with the [14:17] states so that we don't have 50 [14:20] different regulatory paradigms. Um [14:23] what's particularly important is having [14:25] strong physician representation at every [14:28] stage of the AI life cycle to ensure [14:30] that it reaches its full potential. [14:33] Physicians are really the only ones who [14:36] are fully versed in what the clinical [14:38] workflow is. And so as we develop these [14:42] AI technologies [14:44] um it's important that the physicians [14:47] are in from ground one um every stage of [14:51] the life cycle. Uh a transparent whole [14:55] of government approach is also [14:57] necessary. Ensuring that the data is [14:59] free from bias will ensure trust. Um, [15:03] this administration does not like to [15:05] talk about equity, diversity, inclusion, [15:08] but that is critically important in any [15:10] data set that forms the basis of a [15:13] clinical application. Um, and so there [15:16] was some lip service paid to that in in [15:20] the uh Trump action plan, but it is very [15:23] important that that that is a [15:26] significant cornerstone of an ethical AI [15:29] strategy. And lastly, it's very [15:31] important that whatever framework is [15:33] determined, as I mentioned earlier, is [15:35] that there's appropriate aortionment of [15:37] liability to ensure that physicians are [15:39] comfortable adopting this knowing that [15:42] they're not going to be left hanging out [15:43] there if there's a problem with the [15:45] technology as it's implemented. [15:50] >> Very good points. [15:52] >> So, I have a question for you. [15:56] Um, given your role as the CMIO of the [16:00] AMA, what do you see as the most [16:02] promising ways that uh, AI can augment [16:05] and not replace clinical decisionm and [16:08] the physician patient relationship? [16:11] >> Well, thank you for asking that question [16:13] and I want to start by sharing that at [16:15] the AMA we talk about an alternative [16:17] conceptualization of AI and so we define [16:21] it as augmented intelligence. Hence the [16:23] way the question was phrased because we [16:26] believe that the technology will be like [16:30] a co-pilot. It will augment the clinical [16:32] decisionm of our physicians. [16:36] And as these technologies are getting [16:38] more and more integrated, I would say [16:40] that there are two main areas where I [16:44] see the most potential. The first one is [16:47] decreasing administrative burden. [16:49] [clears throat] [16:49] And that is what we hear across the [16:51] board from our clinicians. There's so [16:53] much administrative burden. You heard [16:55] that in the last conversation. Um and [16:58] that's just one of the areas that people [17:00] are struggling with. Documentation, [17:03] the inbasket responses, the prior off, [17:07] those are all areas where people are [17:09] drowning in work that can be automated [17:12] for them. And we're seeing that already [17:15] in our clinical environment with ambient [17:18] listening. There are several vendors out [17:20] there and they're loved across the [17:22] board. Um, many of my CMIO colleagues [17:24] will say that they've never had people [17:26] hug them in the hallways as often as [17:28] they do now because they're the ones [17:30] that brought ambient to them. People are [17:32] crying saying that they get to see their [17:33] families for the first time over dinner [17:36] and they're not processing all of the [17:39] patient information during their family [17:42] time. And so that cognitive burden [17:44] decrease that these tools have the [17:46] potential to achieve is really [17:49] important. The automated inbasket [17:51] responses is another tool that has been [17:53] widely adopted in many organizations. [17:56] We're not necessarily finding that [17:57] physicians are the ones that are liking [17:59] this as much as the staff, but we're [18:01] definitely seeing huge decrease in time [18:04] in responses to patients. [18:06] And the second area where I think we [18:09] have a ton of potential is some of the [18:12] summarization technology. I would say [18:14] that's still early in its adoption curve [18:17] compared to ambient for example. And yet [18:20] many institutions are using it. And what [18:23] I find really exciting is that what when [18:26] we digitize records as Dr. Asus [18:29] mentioned, [18:31] we created a lot of extra steps for [18:33] clinicians [18:35] and with that came an extreme amount of [18:39] data that we have to process on a daily [18:41] basis that we're all holding in our [18:44] brains trying to trend patient [18:47] information over time. These AI tools [18:50] have the ability to summarize it for us [18:54] and do what computers do best, which is [18:57] data computation, and present it to us [19:00] as clinicians in a way that will allow [19:02] us to use our clinical judgment to truly [19:06] improve outcomes. And I think that [19:08] that's probably the most exciting piece [19:09] here. I think we have the opportunity to [19:13] impact patient outcomes, which is really [19:17] exciting. And if we do this well as an [19:20] industry, we can have great successes. [19:25] >> So tell me, how is the AMA supporting [19:27] and elevating physician leadership in AI [19:31] and what opportunities do you believe [19:32] exist for engagement? [19:35] >> Well, we do a lot of work in this space. [19:38] So really appreciate the opportunity to [19:40] share some of that. Um, Dr. Zeus had [19:43] alluded to our AMA principles. Those are [19:46] meant to support the needs of physicians [19:49] and we were very early in coming out [19:51] with those. The first ones released were [19:53] in 2018. We're updating them regularly [19:56] and we continue to ensure that we center [20:02] our work around the voice of the [20:03] physician. And so we do want to make [20:06] sure that we're including the physician [20:08] in every [20:10] step of the life cycle of technology. So [20:13] from conversations with early startups [20:18] to integrating that technology into [20:21] clinical care to implementation, [20:24] adoption and ultimately optimization. We [20:27] do work in all of those spaces. We have [20:30] an AMA AI collaborative that convenes um [20:34] many of the specialty societies. So we [20:37] have over 40 specialty societies and [20:39] together we're working on a unified [20:41] physician voice of what the needs of the [20:44] physicians are. So there will be a [20:45] statement that that group will be [20:46] releasing and we're also working on [20:48] model cards and those model cards will [20:51] reflect the needs of the various [20:52] specialties. We have an AMA task force [20:56] that has been convened by our house of [20:58] delegates with experts in this field [21:00] that is also looking at the structure at [21:03] our organization as well as what are the [21:06] deliverables that we need to continue to [21:08] support physicians. We have what we call [21:11] health 2047 that is an incubator that [21:16] focuses on development of new technology [21:19] with the support of physicians at the [21:20] table and several companies have come [21:23] out of that. We also have a health AI [21:25] partnership that works with multiple [21:28] organizations across the country to look [21:29] at the ethical framework for AI and [21:33] there's been some great work that has [21:34] come out of that group. We participate [21:36] in many organizations across the [21:40] country. So we're part of the valid AI [21:42] collaborative. We participate in HIMS. [21:44] We have partnerships with AMIA and with [21:47] AMDS and many of the physician groups [21:49] that are looking at how to thoughtfully [21:51] integrate technology into clinical [21:53] medicine. Um we also focus on medical [21:56] education and we have modules that are [22:00] available to folks at all levels of [22:03] training. So from um pre-graduate to [22:06] graduate medical education as well as [22:08] continuing medical education for all of [22:11] us that are past our med school and [22:13] residency days that introduce the main [22:17] concepts that clinicians need to [22:19] understand to be able to interpret the [22:22] outputs of AI. And um I personally took [22:25] them you get CME with it and they are [22:28] fantastic. Um, we also have policy that [22:32] ensures that we continue to focus on [22:33] medical education in this space. And [22:36] finally, I want to share a couple of [22:38] resources. We have a learning [22:39] collaborative where we bring health [22:41] systems together to work on how they're [22:44] going to develop the infrastructure for [22:46] AI at their institution. That has been [22:49] really wellreceived because it's a [22:50] year-long process where they work [22:52] together on a project. And we also just [22:55] this year released our AI governance [22:58] toolkit and that is a step-by-step [23:01] guideline of how to think about that at [23:04] healthcare institutions and practices [23:06] and to really consider the steps that [23:09] you need to build an infrastructure to [23:11] be successful in implementing these [23:13] tools. And finally, what I'll share is [23:17] that as we think about the development [23:19] of all of these tools and we think about [23:21] our programming and how we support [23:23] physicians, we recognize that it's a [23:25] partnership. And it's important for us [23:27] to be in rooms like this where we talk [23:30] to folks in industry, we talk to [23:32] technologists, and we partner across the [23:35] industry to make sure that we are [23:37] thoughtfully bringing these tools to the [23:40] bedside. [23:42] And with [snorts] that, I would like to [23:45] ask Dr. Asus our final question and then [23:48] we'll have some time for you guys to ask [23:50] questions. Um, we talked a lot about AI [23:54] in healthcare and as we think about the [23:57] future and the potential of these tools. [24:00] What excites you the most? [24:03] >> So, I'm a practicing physician. I see 60 [24:05] patients a day at a minimum in my [24:07] office. And uh I think what's most [24:10] exciting about this is that it has a [24:12] potential to improve the patient [24:14] physician relationship. And how can it [24:16] do that? First and foremost is more time [24:19] for physician patient human interaction. [24:23] Um if AI can automate some of the [24:26] administrative stuff that I deal with [24:28] like documentation, [24:29] scheduling, triaging, allowing us to [24:32] have face-to-face care that would be [24:34] fantastic. Um, you know, I don't have [24:38] ambient listening in my practice. My [24:40] brother-in-law is a primary care [24:41] physician in a large medical group [24:43] that's owned by a large system, and they [24:45] recently initiated the ambient uh, [24:47] listening. And for your typical primary [24:49] care physician who has an unstructured [24:51] progress note, it's a boon. For a [24:54] physician like me who's a specialist who [24:56] has a very structured type of progress [24:58] note, it's not quite there yet. The [25:01] software product that I use in my office [25:03] has released an ambient listening [25:05] module. They demonstrated to me two [25:07] weeks before it release. And I said, [25:09] "This is not ready for prime time." And [25:11] it's like your typical EHR company. They [25:13] released it anyway. And now on the users [25:16] group, they're being overloaded with [25:17] complaints because it's just not ready. [25:20] Another problem with it was the pricing [25:22] structure. They priced it that it would [25:24] be a certain price for up to 200 [25:26] patients. Well, I see 200 patients in my [25:29] office a day easily. So, I will exceed [25:31] on day one what they all lot for one [25:33] month. Um, so, you know, those issues [25:36] still need to be resolved. But if they [25:38] can get resolved, I will be an early [25:40] adopter as I have been for all sorts of [25:42] technology. I'm a technology nerd. Um, [25:47] what else? Enhanced decision support. [25:49] So, in my practice, I look at lots of [25:52] different images and we know that [25:53] imaging is a strong point for AI. It's [25:56] got applications in radiology, [25:58] pathology, dermatology and it will as [26:00] well in opthalmology. There have been uh [26:03] AI uh structures developed in the United [26:06] Kingdom that will analyze serial OCTs [26:09] which are um tests that look at the [26:11] layers of your retina and help with [26:13] macular degeneration diagnoses and [26:15] diabetic retinopathy diagnoses. And when [26:17] that comes down to the individual [26:19] private physician office like mine, that [26:21] will free me up from a lot of kind of [26:23] grunt work where I'm scrolling through [26:25] images. I'll have something that [26:26] interprets them for me and can tell me [26:28] immediately what the changes are that we [26:30] need to deal with to help in the [26:31] clinical care of our patients. And then [26:33] of course there's communication tools [26:34] that we're all looking forward to to [26:36] help us communicate complex medical [26:38] information to patients in a more [26:40] patient friendly way. Um I spend a lot [26:43] of my time doing pre-surgical counseling [26:46] and every single patient that I schedule [26:48] for surgery I have this standard speech [26:50] that I run through and in spite of that [26:54] I have it in writing. so they can take [26:56] it home and read through it. And I say [26:58] to every single patient, you're not [26:59] going to be able to absorb what I'm [27:01] going to tell you now. So, I want you to [27:03] go home, read this, and feel free to [27:05] call back with your questions so we can [27:07] come to a decision together on what's [27:10] best for you. In spite of that, it's not [27:13] that effective. And I have three [27:15] individuals that I employ as full-time [27:18] surgery counselors just to review this [27:20] information over and over and over [27:22] again. And if you've ever had surgery, [27:24] from the moment your physician says to [27:26] you, you need such and such a procedure, [27:28] and they start telling you things, it's [27:30] probably flying over your head because [27:32] you're upset, you're frightened, you're [27:34] nervous, you're anxious. And so, [27:36] anything that we can have that will help [27:38] us repetitively provide this information [27:41] to patients will be valuable. And also [27:44] AI translation tools. We I work in a [27:47] multicultural practice with patients [27:49] from all over the world, and there's [27:50] lots of different languages. And if [27:52] there's a way that uh an AI can listen [27:54] to what I'm saying and then translate it [27:57] real time for my patients, that would be [27:59] fantastic. So I I'm very excited about [28:01] the future of this technology and I'm [28:04] looking forward to future [28:05] implementation. [28:07] >> Well, thank you. Thank you for sharing [28:10] your insights with us. I think it was [28:12] valuable to all of our colleagues in the [28:14] audience. And I'd like to open it up for [28:17] a couple of minutes of questions.