AMA Chief Rejects Predictions AI Will Replace Radiologists

美国医学会负责人驳斥AI将取代放射科医生的预测 美国医学会执行副总裁 John Whyte 认为，人工智能应作为“增强智能”来强化医生的决策能力，不过按服务收费的支付模式与培训欠缺仍制约着其推广普及。 John Whyte 博士探讨了人工智能在医学领域日益广泛的应用，将其定位为必须支持而非取代医生决策的“增强智能”。他探讨了不同代际医生在接纳新技术上的差异、患者过度依赖消费级AI及未经证实的健康可穿戴设备的风险，以及临床医生面临的行政负担。针对AI将淘汰放射科等领域的论调，Whyte 予以反驳并强调，AI将使医疗人员能够充分发挥其最高级别的专业执业能力，同时他坚称医疗AI必须严格遵循“首先不伤害”的原则。 最新调查显示近 80% 的受访医生已在工作中使用 AI 工具，三年内普及率翻倍，目前主要集中在病历记录等行政任务；美国医学会将其定义为“增强智能”，强调必须在医生监督下辅助决策。 (0:04) AI 应用在医生群体中呈现代际差异，年轻医生更积极运用 LLM 进行鉴别诊断，这推动医学教育从死记硬背转向信息检索，但也对保留扎实的核心临床基本功提出挑战。 (2:52) 消费级健康可穿戴设备大多处于“健康指南”的监管灰色地带，部分指标较高的假阴性率容易给患者带来虚假安全感，从而导致延误必要的就医诊治。 (10:42) John Whyte 反驳了放射科医生将被取代的言论，强调 AI 的价值在于辅助医生处理复杂病灶并发挥最高专业水平；未来被淘汰的不是医生群体，而是不会使用 AI 的医生。 (19:49)

American Medical Association EVP and CEO Dr. John Whyte argues that artificial intelligence must function as physician-supervised 'augmented intelligence' to enhance clinical decision-making rather than replace clinicians. Realizing AI's diagnostic and operational benefits requires addressing pedagogical risks, regulatory ambiguities in consumer tech, and misaligned reimbursement incentives. Augmented Intelligence and Physician Retention (0:04) According to an AMA survey, physician adoption of AI doubled between 2023 and 2026, with nearly 80% using tools primarily focused on administrative documentation, ambient listening, and scheduling. While current implementations handle mundane tasks, Whyte maintains that the technology's most significant clinical value will stem from pattern recognition to support faster, personalized diagnostic decisions, such as flagging tumors in mammograms alongside radiologists. Whyte directly disputes predictions by tech figures like Nvidia CEO Jensen Huang, who expressed doubt about the future of radiology. Pointing out that forecasters have predicted the specialty's obsolescence for over a decade, Whyte highlights that radiologists remain indispensable. Rather than replacing medical specialists, AI is expected to automate routine film evaluations, freeing physicians across disciplines to practice at the top of their licenses and focus on ambiguous cases. Generational Adoption and Foundational Medical Training (2:52) Clinical adoption exhibits a distinct generational divide. Older physicians nearing retirement often feel overwhelmed by the rapid technological transition, whereas younger doctors natively consult tools such as OpenEvidence and UpToDate to formulate differential diagnoses and treatment plans. This rapid democratization of medical literature enables both trainees and patients to access structured clinical data quickly. However, early reliance on digital assistants raises serious pedagogical concerns comparable to the historic debate over calculators in mathematics education. Whyte warns that if medical students depend on ambient scribes and AI stethoscopes from their first day of training, they risk failing to master essential physical examination techniques and patient-interviewing skills. In response, the AMA has distributed grant funding to medical schools to evaluate curricula that balance automated efficiency with the retention of foundational clinical competencies. Tech-Driven Design, Diagnostic Hallucinations, and Wearables (8:41) Whyte cautions that healthcare innovation currently resembles 'the tail wagging the dog,' with tech developers pushing complex tools in search of problems rather than addressing authentic clinical bottlenecks. Without clinical context, large language models remain prone to misleading interpretations—illustrated by Alice Park's account of her 11-year-old son, who mistakenly concluded he was suffering a heart attack after querying ChatGPT about heat exhaustion without specifying his age or clothing. This gap between consumer perception and clinical reality is amplified by wearable wellness devices. Whyte highlights a 'regulatory gray zone' created by FDA wellness guidance, where manufacturers market trackers as non-medical devices while providing clinical-style metrics like heart rate variability and blood pressure indicators. Drawing on findings published in JAMA, Whyte notes that high false-negative rates in continuous wellness tracking can falsely reassure patients into skipping necessary medical care. He argues that reproducible, validated home blood pressure devices would provide far greater clinical utility than niche wellness metrics. Administrative Burden, Reimbursement Models, and Care Access (16:44) While AI drafting tools and ambient listeners can relieve documentation burdens—such as mandatory Medicare Advantage mental health screenings and escalating patient portal emails—systemic integration remains constrained by financial incentives. In traditional fee-for-service environments, providers lack financial incentives to deploy autonomous or agentic diagnostic tools without dedicated reimbursement mechanisms. Conversely, integrated managed care environments like Kaiser can leverage AI efficiencies directly to shift clinician focus toward complex diagnostics. Whyte also challenges assumptions that AI and telehealth innovations will organically expand healthcare access for vulnerable groups. Drawing an analogy to remote dermatology—which saw its highest uptake among insured New York City residents seeking convenience rather than rural populations—Whyte emphasizes that new technologies typically benefit well-insured, tech-literate populations first, leaving structural and economic barriers to care unresolved. Referencing AI researcher Geoffrey Hinton's call for protective AI systems and the medical pledge to 'first do no harm,' Whyte emphasizes that healthcare AI must be guided by safety and human oversight. The long-term integration of clinical AI remains contingent on whether the industry can realign payment models, validate consumer monitoring tools, and resolve structural access disparities without eroding foundational medical skills.

在Ai4大会上，美国医学会（AMA）执行副总裁兼首席执行官John Whyte博士指出，医疗人工智能应定位为受医生严格监管的“增强智能”而非替代工具。他强调，技术应用在提高行政与诊断效率的同时，医疗界必须警惕技术驱动导致的本末倒置、消费级设备的监管盲区以及医疗可及性的技术悖论，始终坚守“首先不伤害”的核心伦理。 增强智能与临床定位：驳斥放射科消亡论 (0:04) Whyte援引AMA调查数据表示，医生对AI工具的采用率在2023年至2026年间翻了一番，受访医生中已有近80%在实践中使用AI。然而，当前应用多集中于环境聆听记录、就诊后小结生成及预约等行政事务，临床潜能仍处于起步阶段。AMA主张将AI定义为“增强智能”（augmented intelligence），强调其应当在医生监督下协助识别细微病理模式（例如辅助放射科医生筛查乳腺钼靶检查中的肿瘤），以提高诊断精度和个性化方案制定能力，绝非单方面替代医生决断。 针对英伟达首席执行官Jensen Huang关于放射科职业前景暗淡的预测，Whyte明确反驳称，有关放射科即将消亡的论调已存在十至十五年，但放射科医生至今仍不可或缺且待遇优厚；未来的行业规则更可能是“使用AI的医生将替代不使用AI的医生”。他进一步指出，技术普及受制于医疗支付模型：在传统的按项目付费体系中，缺乏明确报销机制的高级工具难以推广；而在类似Kaiser的管理式医疗体系中，此类工具能有效释放人力，让专业医师专注于复杂的CT、MRI和不确定病例。 医学教育与技能传承：平衡效率与基本功 (2:52) 随着生成式语言模型进入临床，不同资历医生之间显现出明显的代际差异。Whyte指出，临近退休的年长医生对快速的技术转型容易产生畏难情绪，而年轻医师已能熟练利用OpenEvidence和UpToDate等工具辅助鉴别诊断与治疗规划。然而，这一现象引发了关于核心临床技能退化的担忧，尤其当学员从第一天起就依赖环境聆听转录或AI听诊器时，可能难以独立掌握病史采集与体格检查（H&P）或听辨心脏杂音等基础诊断技术。 针对这一类似于“数学教学中计算器争论”的教育困境，Whyte表示，现代医学教育的重点正从单纯的事实死记硬背，转向培养学生如何组织与驾驭海量信息流。为此，AMA正通过向医学院提供专项资助项目，研究如何在基础训练中安全、规范地整合AI工具，确保医学生在享受技术效率的同时守住核心技能底线。 技术驱动的本末倒置与模型幻觉风险 (8:41) Whyte警示称，当前医疗科技创新存在“本末倒置”（the tail wagging the dog）的严重倾向，科技企业频繁推出炫目的工具并四处寻找适用场景，而非依据临床实际痛点倒逼技术研发。大语言模型因缺乏现实情境感知能力，容易产生荒谬结论。Park在对话中分享其实际经历：其11岁的儿子在华氏90度高温下身穿连帽卫衣运动后出现恶心与焦虑，在向ChatGPT查询时因未输入年龄和着装背景，算法竟提示其可能遭遇了心脏病发作或哮喘发作。 除了消费端的误导，算法输出也给专业临床判断带来了心理干扰。Whyte指出，当大语言模型给出不同的诊断建议时，医生有时会对自己经过长期专业训练建立的临床直觉产生动摇和怀疑。他强调，技术工具不能替代面对面的个体化诊疗评估，必须由临床实际需求来主导技术开发的方向与节奏。 可穿戴设备监管盲区与消费级数据隐患 (10:42) 针对消费级穿戴设备的普及，Whyte在其发表于《美国医学会杂志》（JAMA）的文章中指出，FDA的健康指南催生了显著的“监管灰色地带”：厂商以非医疗级的“日常健康”设备进行包装（提供睡眠评分、心率变异性HRV、血氧及最大摄氧量等），但消费者却直接将这些指标视作具备临床指导价值的医学依据。 这种界限模糊构成了实质性的健康风险。Whyte援引JAMA近期关于高假阴性率的研究指出，采用颜色编码警示的非正式高血压监测工具若未能发出警报，极易让患者误以为身体无恙而推迟必要的线下就医。此外，持续追踪还会加剧患者的焦虑心理。AMA当前正与监管机构及行业合作确立基准标准，并指出开发具备高复现性、经过临床验证的居家日常血压监测工具，在防治心血管疾病方面的实际价值远大于监测HRV等小众指标。 行政减负潜力与医疗可及性的技术悖论 (16:44) 在工作流程优化层面，Whyte认同AI能帮助消除低价值的重复劳动。目前临床医生耗费了海量精力来满足各类行政考核，例如对所有联邦医疗保险优惠计划（Medicare Advantage）患者强制执行焦虑与抑郁筛查，以及处理数量激增的患者门户电子邮件。利用AI生成沟通草稿和整理文档，有助于缓解医生的行政负担，使医护人员更专注于自身执业范围内的核心工作。 然而，Whyte对AI与远程医疗自动改善弱势群体就医状况的论调表示怀疑。他以远程医疗的发展历史为例说明，曾被寄予厚望以惠及偏远农村的远程皮肤科服务，实际使用率最高的人群却是纽约市内拥有优良保险覆盖、仅为避免门诊排队的城市居民。这种“可及性悖论”表明，突破性技术往往优先惠及具备数字素养和经济基础的群体，单纯依赖技术手段无法从根本上消除无保险或缺乏资金等深层医疗结构壁垒。 对话最后呼应了AI先驱Geoffrey Hinton关于构建关怀型“母亲/父亲式AI”（mother/father AI）的设想，Whyte强调医学界“首先不伤害”（first do no harm）的核心准则必须成为指引医疗AI研发的伦理支柱。尽管AI展现出巨大的临床赋能潜力，但在支付机制未获理顺、监管标准尚存盲区以及医疗公平性未得保障的前提下，行业各方在追求创新的同时必须对潜在风险保持充分的警惕与审慎。

## Background 本内容由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台。视频记录了《时代周刊》（TIME）资深健康记者 Alice Park 与美国医学会（AMA）执行副总裁兼首席执行官 John Whyte 的炉边对话，围绕美国健康政策变革、生物医学预算调整、疫苗政策及医疗 AI 的深远影响展开探讨。 ## TL;DR 对话深入探讨了 AI 与可穿戴技术在医疗临床落地中的机遇与挑战。John Whyte 强调 AMA 将该技术定义为以医生监督为核心的“增强智能”，批评了科技界“尾巴摇狗”的工具导向思维，并指出模型幻觉与过早依赖技术对年轻医生临床基本功的潜在侵蚀。同时，他剖析了可穿戴健康设备的监管灰色地带、假阴性风险以及医疗可及性在现实支付体系下的局限，重申 AI 不会取代医生，但医疗创新必须始终坚守“首先，不造成伤害”的伦理底线。 ## Atmosphere & style 整场对话兼具专业权威与务实清醒的审慎感，既有行业高层的宏观政策视野，又不乏生动接地气的个人与临床轶事。两位对话者探讨节奏紧凑且默契，从容拆解科技狂热背后的监管盲区与伦理隐患，为听众营造出一种既拥抱技术潜力、又时刻保持临床警惕与人文关怀的理性思辨氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=4s) AMA 调查显示近 80% 的受访医生已采用某种 AI 工具，但当前应用多集中于行政与环境拾音等文书事务，临床潜力的挖掘仍处于初期阶段。 - [00:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=4s) AMA 将技术定性为“增强智能”（Augmented Intelligence）而非取代医生，强调核心价值在于跨维度识别病理模式，且必须在医师专业监督下运行。 - [02:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=172s) 医生群体在技术应用上呈现出明显的代际鸿沟：资深医生倾向维持现状，而年轻一代高度依赖 AI 进行鉴别诊断与方案推演。 - [02:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=172s) 临床决策工具（如 OpenEvidence）必须具备高保真度与清晰的证据来源追溯，以确保推荐结论有扎实的循证医学支撑。 - [04:40](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=280s) 医学教育面临“计算器效应”挑战：过早引入自动化工具可能导致医学生丧失病史采集、体检及基础听诊等核心临床感知技能。 - [08:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=532s) 大语言模型的幻觉与缺乏上下文推理易引发误诊，过度迷信算法甚至会导致患者或年轻医生动摇本应正确的临床判断。 - [08:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=532s) 医疗创新存在“尾巴摇狗”（喧宾夺主）倾向，科技企业往往拿着炫目工具在医疗中生搬硬套，而非以临床真实痛点为驱动逆向研发。 - [12:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=724s) 可穿戴设备游走于监管灰色地带：厂商宣称仅提供健康指标以规避医疗器械监管，但消费者的实际依赖易导致假阴性延误就医风险。 - [14:16](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=856s) 公众健康正面临“数据过载”陷阱，类似膳食补充剂概念的泛滥与对睡眠评分等数字的盲目依赖，易引发不必要的健康焦虑。 - [16:44](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1004s) AI 在行政减负方面价值显著，能有效分担价值医疗指标记录、抑郁焦虑筛查量表及激增的患者咨询邮件等归档压力。 - [16:44](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1004s) AI 提升医疗可及性的设想在现实中遇冷，以远程医疗为例，其往往优先惠及城市高净值人群，无法单靠技术解决医保与支付能力的底层鸿沟。 - [19:49](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1189s) 放射科等专科的未来并非被 AI 取代，而是“不用 AI 的医生会被使用 AI 的医生取代”，且技术落地的动力直接受医疗支付模式（按服务收费 vs 整合式医疗）制约。 - [22:13](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1333s) 医疗 AI 的应用必须坚守希波克拉底誓言中“首先，不造成伤害”（First, do no harm）的伦理底线，防止在技术狂热中忽视患者安全。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=4s) 医疗 AI 的现状与“增强智能”的真正价值 在午餐后的开场对话中，主持人 Alice Park 提到，作为全美规模最大的专业医师组织，美国医学会（AMA）在数月前发布的一项调查显示，其会员对 AI 工具的使用率在短短三年内实现了翻倍。然而，目前绝大多数应用仍集中在行政支持、环境语音监听（ambient listening）、病历代写（scribe）以及就诊后总结生成和日程排期等流程性事务上。Alice 向 John Whyte 提问：行业当下言必称生成式、智能体（agentic）与决策支持，医疗领域的 AI 究竟将走向何方？ Whyte 直言，当前医疗界对 AI 潜力的挖掘“连皮毛都还没有触及”。近八成的医生虽然已经在工作中使用某种形式的 AI 工具，但诸如撰写病历或生成患者出院小结等事务性功能，并不是 AI 真正的价值所在。AI 的核心价值在于跨维度识别复杂的病理模式，从而帮助临床医生更快、更精准地完成诊断，并制定高度个性化的治疗方案。他特别强调，这一切必须在“医师监督”（physician supervision）的框架下运行。AMA 内部更倾向于将 AI 称为“增强智能”（augmented intelligence），其本质是辅助医生做出更明智的决断，而不是取而代之。例如，利用 AI 工具辅助放射科医生识别乳腺钼靶检查中的肿瘤，就是这种增强协同的典型体现。 ```mermaid flowchart TD accTitle: AMA Augmented Intelligence Framework accDescr: Diagram illustrating AMA's concept of augmented intelligence where AI supports rather than replaces clinical decisions. subgraph Input["临床数据与感知"] A["患者数据 / 影像样本"] --> B["AI 模式识别与分析"] end subgraph Augmentation["增强智能协作"] B --> C["辅助决策支持"] C --> D["加速早期精准诊断"] C --> E["个性化诊疗方案建议"] end subgraph PhysicianSupervision["医师监督与决策"] D --> F["临床医生综合研判"] E --> F F --> G["最终医疗决策与执行"] end ``` - AMA 调查显示，医生群体对 AI 的采纳率在过去三年内实现翻倍，近 80% 的受访医生正在使用 AI 工具。 - 当前最普及的落地形态为行政与文书辅助工具，包括环境拾音记录、门诊总结生成及预约排程。 - AMA 将技术定位为“增强智能”（Augmented Intelligence），核心原则是以临床医生的专业监督为主体，推动早期精准诊断与个性化医疗。 --- ### [02:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=172s) 代际差异与临床技能保留的隐忧 随着 AI 向临床决策支持领域延伸，医生群体内部显现出鲜明的代际差异。Whyte 回顾自己 25 至 30 年的行医经历感慨道，部分距离退休仅剩五年的资深同僚普遍抱持“多一事不如少一事”的心态，只求平稳度过职业生涯晚期；而年轻一代医生在成长与受训过程中本就伴随着数字化浪潮，因而极为主动地拥抱新技术，积极运用 AI 来构建鉴别诊断（differential diagnosis）、拟定治疗策略并从海量医学文献中提炼关键信息。同时，这种技术的普及也正在推动医学信息的民主化，赋予患者更多知情权。 然而，这种技术依赖也引发了深层忧虑。年轻医生如果在受训阶段过度依赖自动化系统，是否会导致基础临床能力的退化或发育不全？例如，在从未独立经历过问诊（history taking）与体格检查（physical exam）训练的情况下，医生将难以建立起扎实的临床直觉。Whyte 表示，AMA 正在以审慎积极的态度应对这一转变，评估如 OpenEvidence 和 UpToDate 等前沿循证决策工具的应用。他强调，临床医生必须对底层数据的保真度（fidelity）拥有绝对信心，并且工具必须具备清晰的可追溯性，能够明确展示推荐结论的循证出处。 - 年长医生群体倾向于回避学习新技术，而年轻住院医及医学生则高度依赖 AI 进行鉴别诊断与方案推演。 - 临床信息民主化使得患者获得前所未有的知识获取渠道，但也对医生综合研判与信息把关的能力提出了更高要求。 - 临床决策工具（如 OpenEvidence）必须具备高保真度及透明的引用来源，以确保医生知晓建议背后的循证医学支撑。 --- ### [04:40](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=280s) 医学教育的重塑与“计算器效应”的权衡 Alice 以自身经历引入了医学教育面临的困境：她的父亲是一位老派外科医生，一生奉行严谨的“上手实践”经验，而当下医学生可能因随时能够查阅模型而丧失对底层知识的牢固掌握，这正如当年计算器的普及曾引发“学生不再懂得基础加减乘除”的恐慌一样。Whyte 完全赞同这一类比，并指出计算器的存在并未免除人们理解乘除法本质的必要性。创立于 1847 年的 AMA 最初便是为了规范医学教育、打击游医骗术（medical quackery）而诞生，如今面对 AI 冲击，AMA 同样承担着引导医学院校课程改革的重任。 目前，全美医学院院长和教学主管之间正在进行激烈交锋：究竟应当在多大程度上允许医学生接触和使用 AI？Whyte 指出，如果在求学伊始就任由环境监听代写工具接管一切，学生将永远无法掌握问诊的核心技巧，甚至连“该问哪些深入的问题”都不得要领，沦为机械对照 AI 生成清单勾选的照章操作者；同样，若从一开始就使用能自动识别杂音的 AI 听诊器，学生便再也无法培养出依靠双耳识别心音异常的实操本领。昔日医学训练奖励的是“死记硬背与背诵事实”，但在医学知识爆炸的今天，这种模式早已过时；AI 应当成为帮助学生梳理知识结构的脚手架，而非代替临床思考的大脑。为此，AMA 设立了专项拨款计划（grants program），资助医学院校深入探索平衡 AI 工具应用与核心临床技能锻造的教学方案。 - AMA 正在联合各大医学院院长及住院医师培训项目，设立专项资助基金探索 AI 时代的医学教育新范式。 - 核心分歧在于“介入时点”：过早引入自动化问诊与 AI 听诊器，会剥夺医学生掌握病史采集、体格检查及心脏杂音鉴别的基础感知机会。 - 医学考核的核心指标正由“事实记忆”转向“高阶临床信息整合与决策判断”。 --- ### [08:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=532s) 幻觉风险、网球营轶事与“尾巴摇狗”的行业现状 除能力退化外，大语言模型固有的大模型幻觉（hallucination）是另一个致命隐患。Whyte 分享了一则发生在他自己家里的啼笑皆非的真实故事：在酷暑的一周，他 11 岁的儿子参加网球夏令营，在接近华氏 90 度（约 32 摄氏度）的高温下坚持穿着连帽卫衣和长运动裤，随后出现恶心、焦虑和轻微胸痛等不适症状。回家后，孩子直接使用 ChatGPT 查阅症状，并煞有介事地告诉身为医生的父亲，自己可能得了“心脏病发作”（heart attack）或哮喘。Whyte 哭笑不得地反问：“你输入自己只有 11 岁了吗？你输入了你在 90 度高温天套着卫衣打球吗？”面对父亲“这只是单纯中暑过热”的专业判断，孩子甚至一度反驳、宁愿相信 ChatGPT 的诊断。Whyte 感慨道，普通用户甚至部分年轻医生往往误以为算法绝不会出错，将其输出奉为圭臬；在某些临床场景中，甚至有医生在模型给出不同答案后，开始自我怀疑、动摇原本准确的专业判断。 ```mermaid flowchart LR accTitle: Tech-Driven vs Problem-Driven Healthcare Innovation accDescr: Comparison between current technology-driven innovation and desired problem-driven clinical innovation. subgraph CurrentStatus["现状：喧宾夺主 (Tail Wagging the Dog)"] direction TB T1["科技界研发炫目工具"] --> T2["在医疗场景中强行寻找应用场景"] T2 --> T3["产生海量难以利用的非结构化数据"] end subgraph IdealModel["AMA 倡导：需求驱动"] direction TB P1["锚定临床痛点与患者真实诉求"] --> P2["逆向研发适配的技术解决方案"] P2 --> P3["产出具有可操作性 (Actionable) 的临床依据"] end ``` 针对此现象，Whyte 发出警告：当前医疗创新的主导权几乎被科技行业把持，呈现出“尾巴摇狗”（the tail wagging the dog）的严重倒挂局面——科技公司先造出花哨炫目的技术工具，再去医疗行业里生搬硬套寻找问题，而真正的创新应当始于医疗系统亟待解决的痛点。可穿戴设备的爆发式增长带来了类似难题，海量生理指标涌入但缺乏明确的临床解释路径。在被问及 AMA 是否在推动行业规范时，Whyte 证实，AMA 正在积极与科技巨头展开对话，牵头制定基准标准（benchmarks），力求实现数据保真度、透明度与“可操作性”（actionable）。只有当数据能被医生信任且能据此做出干预时，信息的民主化才具有真正价值。本周，他还在《美国医学会杂志》（JAMA）上发表了专论，针对 FDA 关于可穿戴设备的健康指导规范提出了深入的行业见解。 - 误导性的输入与缺乏上下文的模型推理极易导致误诊，甚至引发患者或初级医师对专业临床判断的无端动摇。 - Whyte 批评当前医疗技术创新存在“尾巴摇狗”（喧宾夺主）倾向，科技公司往往是“拿着锤子找钉子”，忽视了真实临床痛点的优先性。 - 可穿戴设备产生海量数据，若缺乏透明度、保真度与临床可操作性，不仅无法辅助治疗，反而会引发患者不必要的健康焦虑。 - AMA 正在推进与科技企业的对话以确立行业基准，Whyte 近期在《JAMA》撰文探讨 FDA 针对可穿戴健康设备的监管指引。 ### [12:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=724s) 可穿戴设备的“监管灰色地带”与心血管健康风险 可穿戴设备正在悄然滑入一个棘手的“监管灰色地带”。约翰·怀特（John Whyte）指着自己手上佩戴的设备指出，厂商普遍坚称自己的产品并非“医疗器械”，然而设备提供的健康仪表盘上却赫然列出了睡眠评分、血氧饱和度、生物特征数据、心率变异性（HRV）乃至最大摄氧量（VO2 max）。与十年前仅仅用来计算步数的简易计步器截然不同，如今的用户和患者正在根据这些复杂的数据直接指导健康行为并做出生活决策。问题在于，监管机构的现有标准是围绕制造商声明的“预期用途”（intended use）制定的，而公众在现实中的实际使用方式却早已脱节。 以高血压检测为例，改善血压控制是降低心血管风险最直接有效的手段。许多健康应用正相继推出高血压监测功能，但为了规避监管，它们巧妙地选用“指标提示”（indicator）或彩色代码等模糊措辞。怀特提到近期《美国医学会杂志》（JAMA）上刊登的一封信件，专门讨论了这类工具的“假阴性率”问题。普通消费者根本不会去深究统计学上的阳性预测值（PPV）或阴性预测值（NPV），一旦设备没有发出警报，患者就会误以为自己一切正常，从而放弃就医或测量真实血压，甚至不再关注健康饮食。这种潜在隐患往往导致患者在没有咨询医生的情况下做出错误判断。 为了保护患者并同时推动技术创新，怀特呼吁监管机构与行政部门探索新的监管框架。这并不意味着扼杀创新，而是建立基于证据的方法，包括强化上市后监测（post-market surveillance）和全面的风险评估，真正将患者安全置于技术落地的核心位置。 - 可穿戴设备呈现的数据维度：睡眠评分、血氧水平、HRV（心率变异性）、VO2 max（最大摄氧量）等。 - 监管落差：厂商合规口径采用“健康指标提示”而非“疾病诊断”，但消费者的实际行为将其视为临床诊断依据。 - 临床隐患：《JAMA》研究指出的假阴性风险导致患者产生虚假安全感，延误高血压的规范就诊与治疗。 ### [14:16](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=856s) 数据验证的必要性、膳食补充剂的教训与“数据过载”困境 针对设备厂商普遍宣称“我们不提供诊断”的说辞，怀特直言这正是监管逃避的典型写照，如同保健品与膳食补充剂行业的营销套路——到处宣称“改善记忆力”或“促进前列腺健康”，公众却难以追溯背后的真正科学证据。当下行业中甚至掀起了全民测蛋白质的狂热，但同样缺乏坚实的数据支撑。怀特强调，医疗行业需要的是可重复验证、可信赖的工具。相比于患者每年甚至每两年去一次诊所测得的单次血压读数，医生真正关心的恰恰是患者全天候的日常生活血压；如果行业能打造出精准可复现的日常监测工具，将对心血管疾病防治产生难以估量的正面影响。 然而，技术的普及也催生了另一种极端，即公众正被海量的健康数据所吞噬和绑架。怀特生动地分享了当天的亲身经历：他在前一天深夜抵达，清晨6点迎着室外高达 112 华氏度（约 44.4 摄氏度）的滚滚热浪出门散步。走到半途，他感到身体极度疲惫，脑海中冒出的第一个念头竟然是：“糟糕，我得赶紧查一下我的睡眠评分。”但他随即意识到这种思维的荒谬——自己明明已经真切感受到了疲劳，为何还需要一个冰冷的数字来告诉自己身体的感受？这种对数据的过度沉迷与焦虑，是当前数字健康领域必须正视的心理和行为陷阱。 ```mermaid flowchart TD accTitle: 可穿戴设备监管与数据应用框架对比 subgraph S1["现有监管与市场现状"] A["设备厂商声明<br/>'仅提供健康指标/非诊断'"] --> B["规避医疗器械监管<br/>类似膳食补充剂模式"] B --> C["消费者认知错位<br/>过度依赖或陷入数据焦虑"] end subgraph S2["理想的循证医疗闭环"] D["多方协作平台<br/>AMA + 监管层 + 科技厂商"] --> E["经过临床验证的标准化工具<br/>高可重复性与可靠性"] E --> F["改善真实临床结局<br/>捕捉日常全天候连续体征"] end ``` ### [16:44](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1004s) 医疗 AI 的落地现实：行政减负与可及性神话的破灭 话题转向医疗界知名人物对 AI 的看法时，主持人爱丽丝·帕克（Alice Park）引述了梅赫梅特·奥兹（Dr. Mehmet Oz）的观点：AI 有望提升临床医生的工作流效率，替医生剔除耗时低效的琐碎杂务，让医生有充裕精力投入到算法无法解决的复杂疑难病例中。怀特对此表示认同，但他补充指出，奥兹医生常谈论急诊室中结构化数据的自动整理，但现实中临床医生更沉重的负担来自浩繁的行政归档。例如在“基于价值的医疗”（value-based care）体系下，医生必须手动录入二三十项指标；依照联邦医疗保险优势计划（Medicare Advantage）的要求，无论长者因何种原因就诊，医生都必须对其进行焦虑与抑郁量表评估。加上患者电子邮件咨询量的爆炸式增长，AI 如果能分担这些文书与沟通重担，将极大解放独立执业的医生。 然而，对于业内普遍鼓吹的“AI 与技术能显著提升医疗可及性”的说辞，怀特却当场“皱起了眉头”。他直言理论与现实存在巨大断层。理论上，远程医疗应该打破地域鸿沟，为偏远乡村社区引入顶级专科资源；但在现实中，以远程皮肤科（Telederm）为例，其使用率最高的地区竟然是医疗资源极其密集的纽约市。究其原因，仅仅是因为纽约本地患者不愿意在诊室排队等待。这表明许多新工具并没有雪中送炭，反而进一步赋能了那些本就享有充分医疗资源和强烈健康意识的优势群体。对于缺乏医保、无法承担自费成本的弱势群体而言，技术再先进也无法跨越支付能力的现实门槛。 - 行政重负现状：单次诊疗需记录 20-30 项合规指标；Medicare Advantage 患者被强制要求筛查焦虑/抑郁评分；邮件沟通负荷激增。 - 可及性反差：远程医疗在实践中往往优先服务于城市高净值人群，而非预想中的乡村或医疗匮乏区。 - 核心制约因素：技术本身无法解决支付机制与医疗保险缺失带来的底层医疗不平等。 ### [19:49](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1189s) 放疗与影像科医生的未来辩论：支付模式决定技术落地 针对英伟达（NVIDIA）首席执行官黄仁勋（Jensen Huang）此前关于“放射科未来前景黯淡、将被技术深度取代”的言论，怀特予以了坚决驳斥。他指出，“放射科医生即将消亡”的唱衰论调在业内已经叫嚣了整整十到十五年，然而事实是放射科医生不仅依然存在，而且工作无可替代、收入依然可观。AI 的真正角色并不是替代放射科医生，而是作为辅助工具提升读片效率和病灶检出率。怀特重申了医疗界的经典共识：“AI 不会取代医生，但不用 AI 的医生会被使用 AI 的医生取代。” 怀特进一步指出，技术的落地命运在很大程度上取决于医疗系统的“支付模式”（payment model）。在传统的“按服务收费”（Fee-for-Service）模式下，如果智能影像读取工具缺乏明确的医保报销编码，医疗机构根本没有动力采购和使用；相反，在凯撒医疗（Kaiser）这类整合型管理式医疗（Managed Care）体系中，AI 协助筛查标准胸片所省下的时间，能让医生集中精力去完成更多的 CT、MRI 阅片或处理更复杂的影像诊断。随后他还提到了 OpenAI 的山姆·奥特曼（Sam Altman）的观点，奥特曼同样将医疗视作 AI 最具潜力的领域，其目标并非替换医护，而是赋能医生、护士、药剂师等全行业人员，让他们能够真正“在自身执业资质的最高水准上行医”（practice at the top of their license）。 ### [22:13](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1333s) 坚守“首先，不造成伤害”的医学伦理底线 对话尾声，爱丽丝·帕克提到了人工智能先驱杰弗里·辛顿（Geoffrey Hinton）在行业大会上的反思。辛顿提出应当警惕模型的自我保全倾向与幻觉缺陷，主张构建关爱、滋养人类的“母亲型 AI”（Mother AI）或“父亲型 AI”，确保 AI 是为了造福人类而非仅仅为了机器自身而演进。帕克指出，这种理念与源远流长的希波克拉底誓言（Hippocratic Oath）中“首先，不造成伤害”（First, do no harm）的原则不谋而合，医学领域恰恰是检验这套向善伦理的最佳阵地。 怀特对此深表赞同。他总结道，医学界由衷认可并拥抱这一哲学理念——首要原则永远是不造成伤害。AI 在重塑未来医疗格局方面孕育着无可比拟的巨大潜力，但在张开双臂迎接技术红利的同时，医疗行业与全社会必须对伴随而来的潜在危害保持高度清醒与警惕，决不能在乐观的技术狂热中忽视对患者安全的底线守护。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=4s) 连皮毛都还没有触及 — John Whyte > [00:04](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=4s) 增强智能 — John Whyte > [08:52](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=532s) 尾巴摇狗 — John Whyte > [14:16](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=856s) 我们不提供诊断 — John Whyte > [14:16](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=856s) 糟糕，我得赶紧查一下我的睡眠评分 — John Whyte > [19:49](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1189s) AI 不会取代医生，但不用 AI 的医生会被使用 AI 的医生取代 — John Whyte > [19:49](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1189s) 在自身执业资质的最高水准上行医 — John Whyte > [22:13](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1333s) 首先，不造成伤害 — Alice Park > [22:13](https://www.youtube.com/watch?v=GFiLHi5hFCk&t=1333s) 母亲型 AI — Alice Park ## Mentioned John Whyte, Alice Park, Mehmet Oz, Jensen Huang, Sam Altman, Geoffrey Hinton, American Medical Association (AMA), TIME, JAMA, FDA, OpenEvidence, UpToDate, ChatGPT, OpenAI, NVIDIA, Kaiser Permanente, Medicare Advantage, Hippocratic Oath ## Video description John Whyte - EVP & Chief Executive Officer, American Medical Association Alice Park - Senior Health Correspondent, TIME I will discuss w/Whyte the changing landscape of health-related policies in the US and the impact that has on the health of Americans, from cuts in biomedical research budgets to changes in vaccine policies. [Source](https://www.youtube.com/watch?v=GFiLHi5hFCk)

[0:04] John Whyte: 连皮毛都还没有触及 [0:04] John Whyte: 增强智能 [8:52] John Whyte: 尾巴摇狗 [14:16] John Whyte: 我们不提供诊断 [14:16] John Whyte: 糟糕，我得赶紧查一下我的睡眠评分 [19:49] John Whyte: AI 不会取代医生，但不用 AI 的医生会被使用 AI 的医生取代 [19:49] John Whyte: 在自身执业资质的最高水准上行医 [22:13] Alice Park: 首先，不造成伤害 [22:13] Alice Park: 母亲型 AI

[0:04] John Whyte: scratched the surface [0:04] John Whyte: augmented intelligence [2:52] John Whyte: medical quackery [8:41] John Whyte: the tail wagging the dog [10:42] John Whyte: regulatory gray zone [19:49] John Whyte: they're still here and they're still making a lot of money [19:49] John Whyte: physicians that don't use AI are going to be replaced by physicians that do [22:13] Alice Park: mother AI [22:13] Alice Park: father AI [22:13] John Whyte: first do no harm

[0:04] Thank you very much, Penny, and hello,  John. Hello, Alice. Hello, everyone here, [0:09] and welcome back from lunch. So, really  looking forward to this discussion. I think um [0:15] obviously at an AI conference, uh it was  great to see that AI uh in health was [0:21] one of the keynotes here. So, um really  heartened to see that we're starting to take it [0:27] very seriously, this um impact of technology on  health, and in particularly in medical care. So, [0:33] wanted to ask you, you head up one of the largest  professional medical organizations in the country, [0:40] and uh in a survey just a few months ago by the  American Medical Association, uh your members [0:46] reported a doubling in the use of AI from 2023.  So, that's just 3 years, a doubling in 3 years. Uh [0:53] we're going to go into some of the more the other  details of the survey, but at the moment, most [0:58] of that use was concentrated on administrative,  scribe-related, you know, summary-type duties. Um [1:08] and as you look at the landscape of where AI  is today, and what it could do in medicine, [1:16] um where do you see that moving? Do you see  that evolving, you know, away from more sort of [1:22] data collection-type duties to something that's  a little more, I guess the buzzword is agentic, [1:28] generative, you know, decision-support-type roles?  And Alice, what I would say is I don't think we've [1:37] even scratched the surface of the power of AI.  So, as you point out, and our survey demonstrate, [1:44] 80% almost of physicians are using some AI  tool, and it's mostly what we would call ambient [1:49] listening, uh a scribe that prepares the note for  the doctor, or prepares the note for you as you [1:55] leave the after-visit summary, it's helping with  scheduling. That's not the value of AI. The value [2:02] of AI is recognizing all these different patterns  that are helping in getting to a diagnosis sooner, [2:10] more accurately, and then actually helping deliver  a personalized treatment plan for you. But it's [2:18] all in the context of physician supervision,  and I think that's an important point. [2:23] We refer to it as AMA as augmented intelligence.  It's going to help me make a better decision. And [2:31] there's some good examples of that, you know, real  quickly. There's an AI tool that helps physicians [2:37] in diagnosing tumors on mammograms.  That's a good thing. It's not replacing [2:44] the physician. It's actually helping me do better.  That's the future of AI. From your membership, [2:52] are you seeing a generational difference in terms  of how uh doctors are using it in the clinic? You [3:00] know, as far as more decision support type um uses  versus note summaries and scribing. Absolutely. [3:09] And I will say there are and I've been a physician  for 25-30 years. I had some colleagues that are [3:15] probably 5 years from retirement, Alice, and  they're like, you know what? I'm just so worried [3:21] to get through the next 5 years. But the younger  physicians, and that's partly how they've trained, [3:27] are embracing it. And they're using it to help  with the diagnosis, with a differential diagnosis, [3:34] to help with the treatment plan. And that's  important because it goes to how do I synthesize [3:41] all this information that I have nowadays?  And what's also exciting about it is it's [3:48] democratizing information. So it's also allowing  patients to have access to this information. But, [3:55] you're right about that generational component  because younger people have grown up with it. [4:00] They're used to it. Now, the challenge with that  is are they losing certain skills that or didn't [4:06] develop certain skills because they never had to  take a history and physical or they never had a, [4:11] you know, present a diagnosis without the use of  tools. But, at AMA, we're really looking at it in [4:17] a positive way to say, how does it help me make  a better decision? Things like open evidence and [4:23] other tools where I can put in some information,  you know, and up-to-date where it can help me [4:29] um give you the best diagnosis, but I have to be  confident in the fidelity of that data and be able [4:35] to see where they got those recommendations.  And And many of them are actually doing that. [4:40] You brought up an interesting point,  which is medical education. And obviously, [4:44] we're seeing AI play a bigger role in that  in a lot of positive ways. But, one of the [4:50] one of the challenges, I think, is that experience  issue. My father's a surgeon and, you know, [4:56] he he loves He's an old-fashioned surgeon. Talks a  lot about the hands-on approach and the experience [5:03] and the importance of that. So, um at the AMA,  you know, since you are so involved in shaping, [5:10] you know, the next generation of clinicians, how  do you look at that and at the role of AI and [5:16] what it might do to education and knowledge  where you might have doctors in the future [5:24] not {quote} {unquote} knowing as  much because they can look it up, [5:29] you know, and of course, the best example is  like calculators and being able to, you know, [5:32] do math, right? Is there was a lot of concern  back when calculators came out that, you know, [5:38] kids would not know math and would not have  basic math skills. So, similar issue here, too, [5:44] right? With medicine. It's like, how concerned  and what is the AMA doing to address that issue [5:49] and and make sure that AI is a tool rather than a  replacement, right, for knowledge? Exactly right, [5:55] and I think the calculator is a good example.  We use calculators now, but we use it in a [6:00] different way. It's not trying to have kids not  understand how to do multiplication and division. [6:07] So, it's an important point that you bring up  in terms of how do we address these issues? And [6:14] the AMA's been around since 1847, a long time. And  it really started in terms of medical education, [6:20] trying to address medical quackery, which is  what we would talk about. And and some days [6:24] I think uh what old is new, uh but that's a  different now conversation. But the AMA has a [6:33] important role in terms of medical education.  So, we are trying to work with the deans in [6:38] the different medical schools to talk about how  should medical students use AI? Part of it is also [6:46] in residency training is point of care. When I'm  seeing you, how does it help me make a better [6:52] decision? But the interesting struggles that are  going on right now is how much AI do we allow [6:58] in medical school? Because if you start using  a scribe ambient listening from the beginning, [7:05] you'll never learn how to do a history and  physical. Or even know the right questions to ask, [7:11] right? Cuz AI is really good at generating  questions, so you could just go down on a list. [7:15] There is an AI stethoscope that will tell whether  you have murmurs or other elements. How will you [7:22] ever learn to diagnose it if you've never had to?  And it's different for someone like myself who's [7:29] been in training for past training for, you know,  20 25 years versus someone that's going on. So, [7:34] we actually have had a grants program to medical  schools to really think about how do we address [7:40] these issues? Because you're right, we want to  prepare the best physicians in in medical school [7:46] and have to have that training. And you have to  acknowledge the role of AI. But the big difference [7:51] that I would say is as I became a physician, you  were rewarded for being able to memorize and being [7:59] able to, you know, spew off facts. That's not how  it works now, and it shouldn't be because there's [8:06] just an explosion of medical technology  and medical information. AI can help you [8:13] structure that in a way that is going to help you  become a better doctor. And that's how we have [8:20] to help students think about it because the big  challenge and and I've seen this with students. [8:25] If they have never had to come up with  their own differential on their own, [8:29] they just rely upon an LLM to say, "Okay, here  are the five things that the patient might [8:35] the symptoms and get a list. going to learn.  Right. Right. And there's so many symptoms that [8:41] are very common in general, right? That  um you know, you know have to also know [8:46] how to ask to go deeper. And remember there's  hallucination as well, that it provides wrong [8:52] information. So even though we're talking  about physicians, I'll give the example, [8:56] uh my 11-year-old is in tennis camp for the  summer and um he insists on wearing a hoodie and [9:05] you know, and and uh long sweatpants and  it was very hot this week and he had some [9:12] he had to sit down because he had some nausea and  feeling a little anxious and he said chest pain. [9:19] So he comes home and he did use ChatGPT  and he comes to me and he says, you know, [9:26] it could be a heart attack or asthma. And  I say, "Did you put in that you're 11?" [9:34] He says, "No." I said, "You didn't have a heart  attack. You didn't have any shortness of breath. I [9:40] said you were overheated because you're wearing a  hoodie. You put in that you were wearing a hoodie [9:44] and it was like 90°. And he didn't agree with me.  He was like, "No, it could be heart attack." But [9:49] that's the challenge. It It matters what you put  in as well in terms of the information, you know, [9:55] that you're going to get out. So, people don't  think that these tools make mistakes. They're [10:01] taking it as this is correct. And I can tell  you that there are physicians who then start [10:07] to question their own judgment after they put  it in to an LLM because they think, "Oh, well, [10:13] you know, maybe that knows better." And it's a  It's a challenge and it's certainly something [10:17] that we have to work with industry to improve  because right now, Alice, it's the tech industry [10:24] that's running innovations in healthcare. It's  the tail wagging the dog. And what it should be is [10:31] let's start with the problems that we're trying  to solve. And then let's find solutions to do that [10:35] rather than finding some glitzy tool and then say,  "Okay, well, let's go find a problem for that." [10:42] I I think that's a good point that you make that  it's a tail wagging the dog at this point and it's [10:46] the industry that's leading because as we see from  the explosion of um popularity of wearables, too, [10:53] we now have people with armed with all kinds of  data and information, but we're not quite sure [10:59] what it means. Is the AMA playing any active  role in working with industry or, you know, [11:06] at least providing some guidance on how what to do  with all of that data and how to best interpret it [11:13] so that, you know, it helps people rather than  scares them? We're actually talking to industry [11:18] right now about can we set up some meetings  around developing benchmarks around what should be [11:26] the right standards. Look, we need to have  transparency of data. We need to have fidelity [11:32] that the information that you provide is useful  because here's what's happening. I talk about [11:37] we're democratizing information and making it  more available. But what do I care about as a [11:41] clinician? What do you care about as a patient?  That it's actionable. And if I don't know how [11:47] to utilize this data or I don't believe in it,  how am I going to act upon it? And that's a real [11:53] challenge. So, we want to work with industry. As  I mentioned, I wrote a perspective in JAMA this [11:58] week about that this the wellness guidance  from the FDA as it relates to wearables, [12:04] I call it a regulatory gray zone. Are they a  medical device or they're not a medical device? [12:09] They want to say they're not a medical device, but  with the wearables and I have one on, it's telling [12:15] me my sleep score, my oxygenation, my biometrics,  my HRV, my VO2 max. It's giving me a dashboard of [12:25] my health data. Patients are acting on the data.  They're not looking at it as a, you know, a step [12:33] meter that we had 10 years ago when it simply  told us the number of steps. That's the concern. [12:39] The regulatory standard goes around this intended  use that the manufacturer wanted, but how it's [12:44] actually being utilized is very different.  Let's take the example of hypertension. If [12:51] I wanted to do one thing to reduce cardiovascular  risk, it would improve blood pressure control. [12:57] And the apps are coming out with a hypertension  detection. They're using the term indicator. Yes, [13:04] indicator. Sometimes it's a color code. You know,  there's a lot of um debate about that, but there [13:09] is a letter in JAMA that recently appeared  that talked about the false negative rate. [13:15] Which means if I'm wearing a tool, I know  it's checking for high blood pressure, [13:20] and it doesn't give me an alert, maybe I  don't need to go to the doctor. Maybe I I need [13:25] to eat healthy. Maybe I don't need to check  my blood pressure cuz it would have told me. [13:31] Patients aren't looking at the, you know, the  positive predictive value, negative predictive [13:36] value. They're thinking if this is marketed as a  hypertension indicator, then, you know, I'll make [13:43] a decision based on what I hear. And the problem  is they're not even always coming to the doctor [13:47] to talk about it. So, we need to have more of an  evidence-based approach. Maybe it's post market, [13:56] maybe it's risk assessment, and we're you know,  very open to to talking to the administration [14:02] about addressing some of these issues in a  regulatory perhaps a new regulatory framework. We [14:07] want to promote innovation. We're not here to say  no. We're here to protect patients, and I don't [14:11] hear that enough, Alice. That we're saying, "How  do we protect patients with these tools?" Right? [14:16] I think that's an important thing that there's  a lot of stakeholders here. So, it looks like, [14:20] you know, professional organizations like the AMA,  regulatory agencies, even the federal government, [14:25] they're in the in the in the companies, right?  Need to They need There needs to be a a table [14:31] which all of you are sitting and discussing these  issues because if you talk to the companies, [14:37] they do say, you know, "We don't  diagnose." [laughter] But, as you said, [14:43] because that's the regulatory state. Exactly.  But, as you Wait, people use it, however, Exactly. [14:49] This is the supplement industry. What does it  improve memory, improve prostate? What does that [14:54] mean? I mean, everyone's measuring protein right  now, too. What's the data on on protein? Let's [15:01] have an evidence-based discussion about that that  relates to you. That's what we want to promote. [15:09] Better outcomes in a way that is on data that we  can rely upon. So, if you want to measure blood [15:16] pressure and talk about hypertension, then have  the validated tool. That Is that too much to ask [15:21] to to measure blood pressure that's reproducible  that I can rely upon because that's the value [15:27] of it. Because I Do I really care what your  blood pressure is the one time a year you come in [15:33] or the once every other year? No, I care what it  is on a daily basis. And if I could have a good [15:39] tool to measure that, that could have huge impact  on heart disease. But if it's used correctly, [15:46] right? And if it's interpreted correctly. Oh,  that's right. And and I wish we had more input [15:51] from industry and I know there's a lot of people  from industry here about having a good tool to do [15:56] that. That's going to have more impact than your  HRV, which does have value or or your oxygenation. [16:02] And there also has to be a recognition that  we don't want to be consumed by too much data. [16:08] And you know, we were talking earlier and I said I  got in last night and I didn't tell you this part. [16:14] I said I I woke up very early and walked outside  at 6:00 a.m. where it was like 112° already. [16:20] But as I was walking, I thought, oh, I'm  really tired and I thought to myself, oh, [16:24] I need to check my sleep score. Well,  I don't need to check my sleep score. [16:30] I can just be how I feel. And that's the challenge  that we're all so facing as well that people can [16:37] also be consumed by the data that we're collecting  as well. Okay, I want I now want to read to you [16:44] some statements from, you know, leaders in  healthcare and technology and get get your [16:49] perspective on them. Dr. Mehmet Oz, who we were  just talking about as well, recently said that he [16:55] sees AI in health providing more opportunities to  make the workflow for clinicians, for doctors, um, [17:04] better and and make it more efficient.  I.E., you know, perhaps as you were saying, [17:10] eliminate some of the tasks that, you know,  take up a lot of time but don't have a lot of [17:16] uh return, perhaps. And maybe even save that  time for more complicated cases that AI and, [17:23] you know, algorithms can't really solve. Do you  agree with that? I do, and I'm glad to hear that [17:28] Dr. Oz said that. I think we may have a little bit  of interpretation because I've heard him say that, [17:34] and he often talks about if you came to the  ER, how it could, you know, tabulate and [17:40] structure much of the data. And and that's  helpful, but still under physician input. [17:45] But what I'd also say, and I think Dr. Oz is open  to discussion, just the amount of administrative [17:49] burden that physicians have to document in terms  of value-based care. I got to document 20, 30 [17:57] things. That that takes up a lot of time. I have  to do a an anxiety and depression score on every [18:03] Medicare beneficiary that comes in under Medicare  Advantage, no matter what reason they came in for. [18:09] And I'm not always sure we're measuring the  right thing. So, if we can use AI to reduce [18:16] some of those administrative tasks, to help with  responding to emails, we've seen a huge increase [18:22] in the number of email exchanges that doctors  have to address, that allows us to see more time, [18:28] to spend more time with patients. And that  is something that CMS has talked about. [18:31] How do we empower independent physicians,  which is actually becoming a challenge. So, [18:35] I'm glad to hear he said that. Okay. And it's also  obviously important for increasing access to care, [18:40] too. I think a lot of people have looked at  technology and AI in particular as a tool to [18:46] to help um increase access so that more people  are able to connect with the health care system, [18:52] you know, either Yeah, I'm frowning  a little, Alice, cuz I don't know if [18:55] we really have realized that benefit. And and  the example that I would say is in telehealth. [19:02] In theory, it would be that I'm going to increase  access in rural communities to specialists. [19:09] Let's take telederm. You know where the  greatest use of Telederm is? In New York City. [19:16] Where all the They just have to wait too  long and they don't, you know, want to wait, [19:20] you know, any longer. New Yorkers don't  want to wait. I'm concerned sometimes [19:25] that some of these tools are just empowering  people that already have access and already are [19:30] interested in their health. So, I want to see  more data on that that it truly has improved [19:37] access because there's a cost component. Because  if you can't afford and you don't have insurance, [19:42] uh it it doesn't matter what these  tools do. Okay. Um Now I'm going to uh [19:49] read to you something Jensen Huang of Nvidia  recently said that one profession that he doesn't [19:55] see having a very rosy future in coming years in  medicine is radiology. Um he he definitely sees, [20:02] you know, AI and technology, I think. I don't  know if he would go so far as to replace, [20:07] but maybe. But at least um really take over  a lot of the duties that radiologists do. Um [20:14] Makes sense. I've I've heard that and I'm going  to push back, Alice, and I think you know this. [20:19] They've been saying the demise of  radiologists for 10 or 15 years. So, [20:24] this is not They're still here. And they're still  here and they're still making a lot of money. [20:29] Um and they serve an important role. I think  what it can do is perhaps free up some time [20:36] for radiologists to read more films, to read with  more accuracy. I'll tell you what's an important [20:44] point here is what's going to be the payment  model there. Because say I have an agentic tool, [20:50] you mentioned that. So, just for sake of argument,  and say it can read a chest x-ray better. [20:55] If you're in a fee-for-service system and there's  not a way to pay for it, no one's going to utilize [21:01] that. If you're in a managed care program like a  Kaiser and you have that service, that can free up [21:07] time for radiologists to do more CT scans, to do  more MRIs, to do, you know, Or to read the Or to [21:13] answer that AI finds. So, uh, people always love  to talk about the demise of various specialties. [21:21] I'm not too concerned about it, and I also  don't think that, uh, there's going to [21:25] physicians are going to be replaced by AI. I  think radiologists, physicians that don't use AI [21:32] are going to be replaced by physicians that do.  Because these tools can help me as a radiologist [21:38] detect more tumors, detect more abnormalities  in imaging, and actually improve patient care. [21:44] And also perhaps concentrate on the cases that AI  finds more challenging. That's right. You know, [21:49] can't provide a black and white answer for.  And I don't know if this is in your quotes, [21:52] but you probably know Sam Altman talked about  in terms of AI that healthcare is the space [21:59] that you need to be in. And he's not seeing it  from the point of replacing health professionals. [22:04] The opportunities. Yes, and to practice at the top  of your license. And that's not just for doctors, [22:08] that's for nurses, that's for pharmacists, it's  for all people in the health profession. Okay, [22:13] we're we're at time. So, one last thing at this  conference and even last year, Geoffrey Hinton, [22:18] um, talked about the need for really creating  AI that nurtures and helps us rather than, um, [22:28] you know, harms us and uh, creating making sure  that we develop AI that is for us rather than for [22:39] AI. You know, because any model that you create  is going to focus on self-preservation, right? [22:44] And and that's what we're seeing with some of the  hallucinations and some of the things that, um, [22:50] outcomes that we've seen with AI systems. So, it  seems that there's no better place in medicine, [22:55] right? With the Hippocratic oath, um, do no  harm, to really put this into play. Um, and to [23:01] find ways that AI algorithms can really benefit us  and nurture us. I think he calls it mother AI. Um, [23:08] Or father AI. Or father AI, whichever you want  to say. Hi. I'm so sorry, we're going to have to [23:14] end that. We have the next ones. So, just  final final thoughts on that. Sure. I I [23:19] love that approach. It is about first do no  harm. And I think AI has tremendous potential, [23:24] but we also have to recognize that there is  potential harm. We don't always acknowledge that. [23:29] All right. John, thank you very  much. Really appreciate it.