Epic Ambient Scribe Threatens AI Startups, Health Leaders Say

医疗行业领袖表示Epic环境记录功能威胁AI初创企业 医疗高管警告称，EHR原生AI工具与异步虚拟护理可能淘汰独立软件初创公司，并颠覆传统医院的收入模式。 医疗高管探讨了将AI整合至医疗提供机构的战略现实，涉及“自建还是采购”的权衡、EHR平台集成以及小型医疗系统所面临的财务壁垒。专家小组强调，临床应用在很大程度上取决于与EHR的无缝集成，并且需要超越简单的行政工作流程，迈向智能体自动化。他们还警告称，异步虚拟护理对传统实体医疗系统构成了生存威胁，同时强调了数据清洗、临床治理以及达到人类对等基准对于成功部署的重要性。 受制于严格的 FDA 监管、CPT 编码与医保报销壁垒，核心诊断和治疗类 AI 模型仍将由第三方主导开发；医疗系统自身则应聚焦于底层平台架构，通过智能体流水线将常规行政任务彻底去人工化，避免采购浅层 GUI 套壳应用所带来的系统集成与版本同步负担。 (18:08) 临床医生对 AI 工具的实际采纳高度依赖与电子病历（EHR）系统的无缝集成，任何外部跳转都会导致使用率暴跌；随着 Epic 推出原生环境录音（ambient scribe）功能，缺乏深度平台集成的单点垂直 AI 初创公司正面临严峻的生存挑战。 (21:40) 随着模型使用成本持续降低，小型诊所已能部署本地安全 LLM 或代理工作流来自动化处理转诊与术前检查文书；但面临劳动力短缺和利润挤压的单一医院系统若不加入区域医疗联盟共享技术，将面临掉队风险。 (27:27) 基于智能手机的会话式 AI 和异步虚拟护理正在颠覆传统实体医疗模式；若轻症患者完全被线上虚拟渠道分流，实体医院将只剩下病情最复杂的高危患者，从而引发严重的运营承载力与财务结构危机。 (32:06)

Care provider executives face complex operational trade-offs when implementing artificial intelligence, balancing platform architecture and workflow automation against strict electronic health record integration, variable data quality, and traditional fee-for-service reimbursement models. Platform Architecture and the Build-Versus-Buy Dilemma (18:08) Health systems can viably build in-house models for operational workflows and quality improvement, but panelist Adam Kaufman observes that few will develop internal diagnostic or therapeutic tools. These clinical applications require navigating steep FDA regulatory hurdles, complex reimbursement structures, and American Medical Association (AMA) CPT coding processes, leaving core clinical algorithms largely to specialized vendors. At the architectural level, Peter McCaffrey argues that health systems should avoid purchasing superficial applications that merely wrap graphical interfaces around large language models, which incur heavy integration and version-parity taxes. Instead, he advocates investing in platform-level architecture and agentic pipelines that bring tokens directly to enterprise data. McCaffrey asserts that healthcare's core objective is satisfying user intent rather than preserving legacy workflows, meaning AI should strive to remove human participation entirely from routine administrative tasks. EHR Integration and Resource Disparities Across Systems (21:40) Electronic Health Record (EHR) integration serves as the primary determinant of clinical adoption. Dr. Krishnaj Gourab warns that requiring clinicians to open external browser windows or third-party portals drastically reduces utilization, compelling systems to align their technical plans with major vendor roadmaps, such as Epic's three-year horizon. While Epic's native ambient scribe challenges standalone startups, Kaufman points out that expansive health system data lakes often reside outside the EHR on hyperscale cloud platforms like AWS or Microsoft Azure. Gourab also cautions that native EHR tools are not inherently superior, referencing historical deficiencies in Epic’s proprietary sepsis prediction model exposed by University of Michigan researchers. Resource constraints further divide provider types. Large academic centers maintain dedicated engineering teams—such as Gourab's 20- to 25-person technical group—while small practices can increasingly deploy local models or commercial tools under vendor Business Associate Agreements to automate administrative paperwork like surgical referrals. However, Kaufman warns that independent, single-hospital systems face severe operational risks from thin margins, labor costs, and Medicaid cuts unless they join regional technology consortia. Brick-and-Mortar Disruption and Reimbursement Bottlenecks (32:06) The panel outlines long-term structural risks to traditional healthcare facilities. McCaffrey projects that conversational AI and asynchronous mobile interfaces will increasingly become the initial entry point for patients, threatening major physical medical centers if traditional providers fail to integrate virtual interactions into clinical operations. Gourab cautions that if healthy patients migrate to virtual channels for low-acuity issues, brick-and-mortar hospitals will be left treating exclusively complex, high-acuity patients, which strains physical capacity and undermines hospital operating margins. Broader clinical shifts, including home-based infusions and GLP-1 medications, further reduce the need for physical hospital visits. Nevertheless, panel members emphasize that traditional delivery models remain anchored to legacy payment structures. Because provider revenues depend on clinicians spending 15-minute increments with patients to bill under fee-for-service rules, health systems will remain tied to physical time slots until payers like Medicare fundamentally alter reimbursement. Data Hygiene, Performance Parity, and Clinical Governance (37:13) Underlying data quality remains a persistent hurdle for healthcare AI. Gourab explains that at Innovation Harbor, technical teams spend three to six months strictly sanitizing and curating electronic medical record data before training machine learning models. Conversely, McCaffrey pushes back against demands for zero-error ground truth, noting that human-generated records have always exhibited significant inaccuracies. He argues that health systems should target parity with baseline human performance rather than unattainable perfection, using advanced reasoning and token management to manage imperfect data. To ensure sustained adoption and avoid clinician alienation, health systems are formalizing AI governance. Panelists point to strategies like those at Baylor Scott & White, where AI projects are linked directly to institutional key performance indicators such as reduced appointment wait times and no-show rates. Sustainable implementation also relies on dedicated clinical-technical leadership—such as Chief AI Officers or Chief Medical Information Officers—paired with governance pathways where clinician-initiated proposals must receive formal vetting from executive Chief Medical Officer councils. While artificial intelligence offers substantial opportunities to automate administrative overhead and shift care triage, health systems remain constrained by legacy IT architectures, variable electronic health record data, and fee-for-service reimbursement models that still tie revenue to physical clinician time.

在AI4医疗管理者战略研讨中，来自美国医学会、学术医疗系统与产业界的专家指出，医疗机构推进AI转型不仅面临自研与采购的权衡，更受到电子病历集成、按服务收费体制以及底层数据治理的现实制约；未来医疗组织若要在技术重塑中维持生存，必须建立平台级架构并推动以临床意图为核心的敏捷治理变革。 算法自研与外部采购的边界权衡 (18:08) Adam Kaufman指出，医疗机构可利用专有数据针对内部流程和质量改进开发本土模型；但面对诊断或治疗类核心应用，极少有医院会选择自研，因为这类模型面临美国FDA严苛的监管审批、复杂的医保报销途径以及申请美国医学会（AMA）CPT计费代码的高昂门槛，核心临床算法因此主要仍由第三方专业公司承接。 Peter McCaffrey则认为，大语言模型正在重塑传统软件的定义。医疗系统若盲目采购仅在底层大模型上套壳图形界面的浅层应用，将背负沉重的集成税与版本同步税。他主张将智能体管线直接部署至数据湖，并强调医疗转型的本质在于满足用户意图而非因循旧制，AI的真正价值在于把人类完全从常规行政事务中抽离，而不仅是修补既有工作流。 基础设施门槛与电子病历集成生态 (21:40) Christopher Khoury与Krishnaj Gourab指出，引入AI需要承担涵盖治理框架、硬件设施和人员培训的固定基础成本，但随着技术底座成熟，后续部署场景的边际成本将大幅下降。Gourab强调，临床医生对工具的接纳度几乎完全取决于其与电子病历（EHR）的嵌入深度，强迫医生跨系统或开启外部浏览器窗口会导致使用率锐减，因此内部技术选型必须与Epic等主要EHR厂商的三年路线图对齐。 针对Epic近期推出原生环境语音转录（ambient scribe）功能的举措，McCaffrey认为这将对缺乏深度集成的初创软件造成毁灭性打击。Kaufman则指出，许多医疗系统在EHR之外拥有庞大云端数据湖，更看重微软Azure或AWS等超大规模云底座。同时Gourab告诫，原生EHR模型并不等同于高质量，密歇根大学的学术研究就曾揭示Epic早期专有脓毒症预测模型存在严重的性能缺陷。 机构规模分化与资源限制下的落地路径 (27:27) 研讨认为，大型学术医学中心固然拥有20至25人的专业研发团队优势，但算力与模型成本的快速下降正为小型机构带来机会。基层诊所很快能够通过本地部署安全模型或与Anthropic等厂商签订业务伙伴协议（BAA），利用自主智能体分担初级保健中沉重的手术转诊、术前评估及预约排期等行政文书负担。 针对不同类型机构的抗风险前景，McCaffrey认为3至5人的小型技术团队借助现代开发工具亦可快速构建高效方案；Kaufman也看好独立门诊灵活引入成熟产品。然而Kaufman对独立单体医院表达了深切担忧，这类机构深陷微薄利润、人员紧缺及医疗补助（Medicaid）削减的多重挤压，若不加入区域技术联盟共享基础设施，将面临被技术浪潮进一步边缘化的风险。 实体诊疗模式重构与支付机制制约 (32:06) 传统实体医疗体系正受到移动端异步虚拟护理的深刻挑战。McCaffrey警告，患者愈发倾向于通过手机端对话式AI和语音交互寻求即时医疗答复，若传统医院无法将此类触点接入临床工作流，诸如德克萨斯医疗中心这类实体巨头将面临被架空的长期威胁。Gourab补充指出，若轻症患者大量流向线上虚拟通道，实体医院将被迫主要接收病情最重、利润最低的高危患者，进而严重破坏运营容量和住院收入模型。 此外，GLP-1药物与居家输液等治疗模式的兴起，进一步削减了患者到院就医的刚性需求。然而多位专家强调，当前的医疗收入依然牢牢绑定在按服务收费（Fee-for-Service）体制上，医院仍需依赖医生面对面问诊15分钟来获取账单补偿。在联邦医疗保险（Medicare）等主要支付方根本性调整报销标准之前，成熟医疗系统在打破实体诊疗依赖上面临现实制度掣肘。 医疗数据治理与临床变革管理机制 (37:13) 针对病历记录充斥混乱非结构化数据的普遍担忧，Gourab介绍了创新港（iHarbor）的应对方案，其跨学科团队在算法摄入前会投入3至6个月专门进行数据清洗与标定。McCaffrey则反对对AI实施双重标准，指出传统人工医疗记录原本就存在高变异率与历史错误；医疗系统应以对标人类表现为基准目标，通过优化推理链、Token预算控制和多数据源交叉校验来包容和化解数据缺陷。 在推行变革管理方面，专家借鉴贝勒斯科特与怀特健康等系统的经验，建议将AI项目紧密挂钩至减少候诊时间与降低失约率等全院关键绩效指标（KPI）。同时，机构应设立首席人工智能官（CAIO）或首席医疗信息官（CMIO）统筹路线图，建立由临床医生自下而上发起需求、再交由首席医疗官（CMO）委员会等高层严格审批的治理闭环，避免技术应用沦为强加于医护人员的外部行政负担。 总体来看，AI对医疗机构的重塑不仅关乎模型开发与工具选用，更是一场牵涉数据底座、实体运营与商业模式的系统性博弈。虽然技术门槛正在降低，但在按服务收费机制和严格审批监管的制约下，医疗管理者仍需在保障临床安全与推动机制创新之间寻求稳健平衡。

## Background 本内容由 Ai4 于 2025 年 8 月 29 日在 YouTube 平台上发布。视频记录了一场面向医疗机构高管的高端圆桌讨论，汇聚了来自马里兰大学医疗系统（UMMS）、德克萨斯大学医学分部（UTMB）、Private AI 以及美国医学会（AMA）的专家领袖，共同探讨在 AI 时代医疗机构如何推进战略转型、重塑护理交付模式、提高运营效率以及应对文化与组织变革挑战。 ## TL;DR 圆桌论坛围绕医疗机构在 2025 年推行 AI 战略的核心议题展开，深入剖析了如何将 AI 项目直接锚定高层战略目标与北极星指标，平衡行政提效与临床应用的 ROI 及风险。嘉宾们详细讨论了“自研 vs 外采”的权衡、EMR 平台的生态博弈与工作流摩擦，以及如何应对 Epic 等巨头推出原生 AI 带来的行业冲击。同时，专家们警示了异步虚拟医疗与医药模式转变给传统实体医院带来的生存危机，并就脏数据治理与一线临床医生的变革管理给出了实操解法。 ## Atmosphere & style 讨论充满高度的务实感与战略前瞻性，没有空洞的技术狂热，而是充斥着来自一线管理与临床实践的冷峻反思与敏锐洞察。主持人掌控自如，三位来自不同医疗体系与技术背景的嘉宾各抒己见、层层递进，既有对“实体医院边缘化”和“劣质 EMR 数据”的直白警示，也有对“AI RVU”等机制创新的大胆构想。整场对话既严肃专业又极富张力，宛如一场顶尖医疗智囊团关于行业生死存亡的战略闭门兵棋推演。 ## Key takeaways - [04:41](https://www.youtube.com/watch?v=c4lubsynndg&t=281s) 争取高层对齐的关键在于唤起对技术指数级进化的共识，医疗机构若脱离大模型的发展曲线，将面临复利式恶化的技术债务与负利润率。 - [04:41](https://www.youtube.com/watch?v=c4lubsynndg&t=281s) 获取 AI 预算的有效路径是将项目锚定在医院核心战略指标上，例如为实现安全目标而构建全天候运作的“虚拟住院医”。 - [09:33](https://www.youtube.com/watch?v=c4lubsynndg&t=573s) 行政管理类 AI（如医保拒付申诉、病床状态与利用度审查）拥有极度直接的近端 KPI 和快速 ROI，比临床核心场景更容易获得立项批准。 - [10:58](https://www.youtube.com/watch?v=c4lubsynndg&t=658s) “紧邻收入”且具伦理价值的用例（如筛查高阶手术候选人、追踪漏做检查）最易落地，并通过人机协同将算法风险化解在专家最终审核中。 - [12:48](https://www.youtube.com/watch?v=c4lubsynndg&t=768s) 克服一线医生的抵触需重塑激励机制（如构想“AI RVU”补贴）并将体验隐形化，将复杂算法转化为类似护理助理的分工，而非生硬改变习惯。 - [17:26](https://www.youtube.com/watch?v=c4lubsynndg&t=1046s) 自研与外采策略分工明确：机构应基于本地数据自研工作流与质控模型，而诊断与治疗级器械模型因需跨越 FDA 审批与 CPT 计费报销门槛，更适合外采。 - [17:26](https://www.youtube.com/watch?v=c4lubsynndg&t=1046s) 警惕过度采购仅为大模型套壳图形界面的 SaaS 软件，否则未来需承担高昂的系统集成税与版本同步税；AI 的终极价值在于响应临床意图而非妥协现有工序。 - [21:19](https://www.youtube.com/watch?v=c4lubsynndg&t=1279s) 电子病历（EMR）具备绝对的工作流重力，任何要求医生跳出主界面独立操作的 AI 工具都会遭遇采纳率断崖式下跌，必须与原生界面深度嵌合。 - [25:49](https://www.youtube.com/watch?v=c4lubsynndg&t=1549s) 虽然 Epic 等巨头推出原生环境转录会挤压专精厂商，但平台内置并不天然等同于卓越质量（如早期脓毒症模型的缺陷），且互操作性法规正逐步拆解数据护城河。 - [27:27](https://www.youtube.com/watch?v=c4lubsynndg&t=1647s) 中小机构无需重资产算法团队，利用开源模型与商业 API 智能体打通初级保健与手术转诊即可见效，创新的最大阻力往往是流程黑盒而非团队规模。 - [31:59](https://www.youtube.com/watch?v=c4lubsynndg&t=1919s) 异步虚拟医疗构成了实体医院的生存级风险，当就医首选入口转移至口袋端的大模型时，实体机构若不主动融入工作流将面临边缘化危机。 - [34:25](https://www.youtube.com/watch?v=c4lubsynndg&t=2065s) 医疗重心正在从“手术操作”向“药物治疗”（如 GLP-1）转移，轻症被线上分流导致实体医院留存极度重症患者，直接动摇了基于人工时长的收费运营模式。 - [37:06](https://www.youtube.com/watch?v=c4lubsynndg&t=2226s) 医疗界对 AI 幻觉存在双重标准，面对充满错漏的历史电子病历，工程策略应从追求绝对洁净的数据源转向利用推理逻辑与多步反思架构来弥补脏数据干扰。 - [39:21](https://www.youtube.com/watch?v=c4lubsynndg&t=2361s) 消除临床团队排异反应的关键在于让临床痛点驱动技术立项、设立具备统筹话语权的首席 AI 官（CAIO），并由首席医疗官（CMO）委员会把关治理准入。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=c4lubsynndg&t=4s) 开场介绍与圆桌嘉宾背景 会议主持、来自美国医学会（American Medical Association, AMA）战略洞察副总裁 Christopher Khoury 向与会者致意，并针对现场观众展开互动调研。现场不仅汇聚了大量一线临床医生与医疗系统从业者，还有众多来自医疗系统非临床及技术岗位的专业人士。Christopher Khoury 指出，AMA 作为全美最大的医师组织，长期关注各类前沿医疗变革与医师群体的核心需求。随后，三位在医疗健康与人工智能交叉领域深耕的资深专家依次进行了自我介绍： - **Adam Kaufman**：目前担任 Private AI 董事会成员。该机构专注于帮助各类机构（尤其是医疗健康组织）挖掘非结构化数据的价值，核心应用场景包括识别并脱敏受保护健康信息（PHI），生成脱敏数据或合成数据以供安全使用。在此之前，他长期为制药与医疗科技企业提供咨询，过去五年深入参与大型医疗系统建设，先后服务于梅奥诊所（Mayo Clinic，包括 Mayo Clinic Platform 的 AI 布局）以及德克萨斯州的 Baylor Scott & White Health。 - **Krishnaj Gourab 博士**：物理医学与康复科（PM&R）专家，主要负责脑卒中与脊髓损伤等患者的康复治疗；现任马里兰大学康复与骨科研究所（University of Maryland Rehab and Ortho Institute）首席医疗官。在 AI 领域，他兼任马里兰大学医疗系统（UMMS）创新部门“IH Harbor”的数字产品与 AI 医疗主管，并担任系统内 AI 治理委员会委员。他直言，当前 AI 在临床医学中的应用既令人兴奋又充满挑战与不确定性。 - **Peter McCaffrey 博士**：病理学家背景，现任德克萨斯大学医学分部（UTMB）首席 AI 与数字官，同时主持覆盖德克萨斯大学系统全部 8 家健康相关机构的“UT Real Health AI”倡议。其工作重点是在确保安全与合规治理的前提下，全方位评估 AI 在手术规划、患者照护、院内寻路导航、账单优化等全流程中创造价值的切入点与用例空间。 ### [03:34](https://www.youtube.com/watch?v=c4lubsynndg&t=214s) 2025年医疗系统面临的宏观挑战与AI转型背景 Christopher Khoury 为本次战略探讨梳理了宏观背景：步入 2025 年，美国医疗生态正经历前所未有的剧烈震荡与系统性重构。医疗机构当前面临着多重外在冲击与内在压力的叠加：一方面，政府医疗补助（Medicaid）削减幅度巨大，学术型医疗中心更遭遇了美国国立卫生研究院（NIH）科研经费的大幅缩水；另一方面，全行业承受着严峻的人才短缺、临床医生与护士群体的职业倦怠（Burnout），加之成人与儿童慢性病负担持续加重，以及日趋复杂的社会与政策博弈。 医疗系统被公认为当今社会最为复杂的社会、政治与经济复合体之一。在如此庞大且高度严密的有机系统中推行 AI 落地，绝非简单的技术引入，而是一场触及运营肌理的深层组织转型。要实现这一目标，首要前提便是从顶层设计出发，构建全院统一的战略愿景并获得最高管理层的真正认同。 ### [04:41](https://www.youtube.com/watch?v=c4lubsynndg&t=281s) 如何获取高层领导支持并构建AI战略愿景 针对如何赢得最高管理层对 AI 战略的支持，三位嘉宾从不同维度分享了各自的破局思路与方法论： Peter McCaffrey 博士强调，争取高层领导对齐的关键在于唤起对“当下历史节点特殊性”的共识。大语言模型（LLM）与生成式 AI 在过去近三年内展现出的能力扩展速度与通用性已彻底颠覆以往的技术迭代周期。如果技术底层的推理与抽象能力呈指数级增长，医疗机构一旦脱离这条增长曲线，其技术债务与负利润率也将以指数级急剧恶化。领导层必须意识到，低估这场变革所带来的“复利式恶化”（"compounding badness"）是任何医疗系统都无法承受的战略误判。 ```mermaid flowchart TD accTitle: 医疗系统 AI 战略落地与高层对齐框架 accDescr: 展示 AI 战略如何结合宏观增长曲线与医院顶层战略目标，并分类推进不同层级的 AI 工具。 subgraph StrategyAlign ["顶层对齐与战略驱动"] A["技术演进现实<br/>LLM 推理与通用能力指数级增长"] --> B["战略警示<br/>落后将导致技术债务与负利润率复利恶化"] C["医院核心战略目标<br/>例如：打造全州最安全的医疗系统"] --> D["AI 解决方案具象化<br/>构建全天候运转的虚拟住院医/实习生"] end subgraph ToolClassification ["AI 工具分层推进策略"] E["临床类 AI (Clinical AI)<br/>影像/病理/辅助诊断<br/>依循监管审批按需采纳"] F["行政类 AI (Administrative AI)<br/>环境声学听诊/排班/核保<br/>由直接运营效率驱动"] G["底座支撑 AI (Foundational AI)<br/>数据治理/PHI 脱敏/系统集成<br/>由 IT 与工程团队直接决策"] end StrategyAlign --> ToolClassification ``` Krishnaj Gourab 博士则指出，无论是搭建 AI 治理体系还是针对本土人群数据自研定制化模型，均需投入大量资金与时间。获取高层支持的有效路径是将 AI 项目直接锚定在医院的战略北极星指标上。例如，当首席执行官提出“将本院打造为全州最安全的医疗系统”时，沿用传统手段显然无法破局。UMMS 由此提出了构建“虚拟住院医”（Virtual Resident / Intern）的概念——利用 AI 全面解析结构化指标与临床非结构化病历，沉淀医院全套制度知识库与资源调度规则（如检查预约通道与流程），打造一个永不疲惫、无工时限制且 24 小时全天候守卫患者安全的智能中枢，从而使高层欣然批准对前沿技术的投资。 Adam Kaufman 补充了另一种务实分类视角：并非所有 AI 工具都需要上升到医院最高战略愿景的层面去审批。他将 AI 工具划分为三类： 1. **临床类 AI（Clinical AI）**：如病理切片或计算机视觉辅助诊断工具，其采纳很大程度上取决于监管审批与科室层面的实际临床效用； 2. **行政与流程类 AI（Administrative AI）**：如环境声学聆听（Ambient Listening）病历记录、护士智能排班等，其价值在于提升效率，无需绑定宏观愿景即可凭借明确的效能提升直接推行； 3. **底座赋能类 AI（Foundational AI）**：涉及底层数据架构与工程能力，通常由 IT 和研发团队自行决断。 与此同时，各机构会根据自身品牌定位选择差异化的倾斜方向——梅奥诊所聚焦攻克复杂疑难病例的前沿创新，Baylor Scott & White 聚焦患者体验与“以客户为中心”，而马里兰大学则聚焦患者安全。 ### [09:33](https://www.youtube.com/watch?v=c4lubsynndg&t=573s) 行政类AI与临床类AI的投资回报率（ROI）与风险权衡 在探讨行政类工具与临床类工具的落地差异时，Christopher Khoury 提出疑问：鉴于医疗行业向来保持审慎规避风险的传统，医疗系统是否更容易在行政工具上达成投资回报率（ROI）共识并推进落地，抑或已有先行者敢于在关乎患者照护的核心临床领域大步迈进？ Peter McCaffrey 博士回应指出，行政管理类工作流的投资回报率确实极易量化与论证。诸如利用 AI 优化医保拒付申诉（Denial Appeals）流程效率，或是 UTMB 大力投入的病床状态判定与利用度审查（Utilization Review / Bed Status Adjudication），这些场景拥有极度直接的近端 KPI 指标，能够直接折算为挽回的收入或削减的硬性支出，因此在立项审批时极具说服力。然而令人惊喜的是，临床领域的探索也展现出强烈的内生需求——例如通过全院级算法自动排查患者未完成的诊疗事项，精准识别并弥合各类“照护脱节”（Gaps of Care），确保每位患者的必要临床干预都能得到闭环执行。 ### [10:58](https://www.youtube.com/watch?v=c4lubsynndg&t=658s) 收入相邻场景与放射科AI的前沿应用 在医疗AI的落地场景中，紧邻收入环节（Revenue-adjacent）且具备明确伦理价值的用例往往最容易获得机构支持。彼得·麦卡弗里（Peter McCaffrey）指出，这类用例包括识别潜在的手术或器械治疗候选人（例如植入式心律转复除颤器 ICD、经导管主动脉瓣置换术 TAVR、外科主动脉瓣置换术 SAVR 以及左心耳封堵器 Watchman 的适用患者），或是追踪放射科检查中的异常发现与后续随访（例如排查漏做的 PET-CT）。这些任务不仅满足了医疗质量的伦理要求，还能直接带来业务增量，同时非常适合采用“人机协同”（Human-in-the-loop）模式，将算法决策风险完全化解在人类专家的最终裁决与审核之中。此外，针对中风（Stroke）和肺栓塞（PE）等急重症的影像AI拥有极为坚实的循证医学证据，在全院层面的推广阻力相对较小。 克里希纳吉·古拉布（Krishnaj Gourab）对此表示赞同，并补充了行政与临床两个维度的AI采纳现状。在行政端，拒付管理（Denials Management）与提前干预已在全美多家医院广泛应用，AI会在提交理赔单前全面核查合规性，纯行政属性使其落地极其迅速；在临床端，放射科医生是全行业AI采纳度最高的群体，已广泛用于脑卒中病灶定位、胸部 X 光片分类以及影像片读取的优先级分诊。随着自然语言处理（NLP）与大语言模型（LLM）的成熟与负责任应用，AI正在从过去“读取像素或体素（Pixels or Voxels）的数值”进化为“直接阅读与理解病历记录中的文本与临床笔记”，这为将AI普及到更广泛的非放射科临床科室打开了大门。 --- ### [12:48](https://www.youtube.com/watch?v=c4lubsynndg&t=768s) 临床落地与推广瓶颈：工作流重塑与“AI RVU”构想 将AI技术推向临床一线时，组织面临着工作流程重塑与医生行为改变的巨大阻力。针对主持人克里斯托弗·库里（Christopher Khoury）关于模型卡片（Model Cards）与各类框架是否真能驱动临床采纳的提问，麦卡弗里坦言落地过程极其艰难。以生成式收件箱（InBasket）自动回复、环境智能语音转录（Ambient AI）以及 Microsoft 365 Copilot 为例，最初设想这些工具能显著节省时间，但实际推行时却遇到了医生个体偏好的巨大差异。面对庞大而分散的用户群，尽管团队不断组织用户交流会并反复进行现场演示，但随着人员流动，培训效果依然会逐渐衰减。为此，他们正在探讨一种激进的激励机制——设立类似相对价值单位（RVU）的“AI RVU”：机构为医生的AI使用行为提供补贴或计分奖励，“如果全局来看价值不足那是管理层的责任，但参与使用是临床医生的职责”，以此通过机制设计驱动临床采纳。 针对一线医生在经历电子病历系统（EMR）推广后产生的技术创伤与对“AI幻觉”、被剥夺决策权的顾虑，古拉布强调了沟通与教育策略的重要性。在推进落地时，团队通常会建议医生“忘掉它是AI、算法或软件”，转而从业务本质提问：“如果你身边配有一名极聪明的实习医生或护理助理，你会希望他帮你做什么？”以此将复杂的算法映射为具体的分工。 亚当·考夫曼（Adam Kaufman）则指出，医疗行业本身的数字化成熟度并不高，AI的快速涌入本质上是传统的数字化转型与变革管理问题。在很多情况下，优秀的AI应该是隐形的——例如贝勒医疗（Baylor）在其患者端 App 中上线的导诊与临床决策分诊工具，背后虽有AI模型支持，但患者与护士只会觉得这是一个比传统电话呼叫中心更高效的“下一步最佳行动”（Next Best Action）路由工具，用户无需感知技术底层即可顺畅使用。 ```mermaid flowchart TD accTitle: 医疗AI临床采纳阻力与推进策略 subgraph Barriers["临床落地阻力"] B1["EMR系统带来的'技术创伤'"] B2["对AI幻觉与替代决策的恐惧"] B3["医生个体习惯与工作流冲突"] B4["培训成效随人员流动衰减"] end subgraph Strategies["推进与破局策略"] S1["心智重塑：视AI为'智能实习生'"] S2["机制激励：探索 'AI RVU' 补偿体系"] S3["体验隐形化：无缝嵌入分诊与流程"] S4["人机协同：人类专家保留最终审核权"] end Barriers --> Strategies ``` --- ### [17:26](https://www.youtube.com/watch?v=c4lubsynndg&t=1046s) 自研还是外采？医疗机构的AI模型策略与底层架构反思 在医疗器械与自研模型的路径选择上，全美已有约1000款获得 FDA 批准的AI/机器学习医疗设备（主要集中在影像领域）。对于医疗系统是否应该“自研模型”（Homegrown Models），考夫曼结合其在梅奥诊所平台（Mayo Clinic Platform）的经历进行了深度拆解。梅奥诊所平台致力于在安全合规的前提下，将梅奥及其他世界顶尖医疗系统的数据资产开放给开发者，并搭建模型部署的基础管道。 考夫曼认为，在**工作流优化**与**医疗质量改进**层面，医疗系统自建本地模型完全可行且极具优势，因为核心在于数据访问权以及结合基础模型进行本地微调；但在**疾病诊断与治疗器械**层面，普通医疗机构不宜盲目自研，因为这类产品不仅要经历漫长复杂的 FDA 审批，还要向美国医学会（AMA）申请全新的 CPT 计费代码并争取医保报销，更适合由第三方专业技术企业研发。 ```mermaid flowchart LR accTitle: 医疗AI自研与外采策略分工 subgraph LocalBuild["医疗机构自研 (Build)"] direction TB A1["质量改进与工作流优化"] A2["本地临床数据 + 基础模型微调"] A3["免除复杂器械审批，快速部署"] end subgraph ThirdParty["第三方供应商外采 (Buy)"] direction TB B1["诊断与治疗级器械模型"] B2["跨越 FDA 监管与审批门槛"] B3["争取 AMA CPT 代码与医保报销"] end ``` 麦卡弗里进一步从底层技术演进的角度反思了“自研与采购”的边界。前沿大模型的通用能力不断跃升，正在模糊传统“软件产品”的定义——许多所谓的商业产品本质上只是“给大模型套了一个图形用户界面（GUI）”。医疗机构若盲目采购大量针对特定工作流的独立 GUI 产品，未来将背负巨大的“系统集成税”（Integration Tax）和“版本同步税”（Version Parity Tax）。因此，更明智的架构策略是押注底层平台与 Agent 智能体管道，实现数据与算力的底层标准化。麦卡弗里最后挑战了医疗信息化领域“一切以工作流为中心”的传统信条： > “人们常说‘一切都取决于工作流’，但现实并非如此——没有人真正按‘工作流’去思考。人们思考的是‘意图’（Intent）以及他们希望达成的结果；工作流只是‘我想要的’与‘我不得不应付的现实’之间的妥协。AI真正的价值不是去优化这些琐碎的工作流，而是将人类从繁琐的操作中彻底解脱出来。” - **FDA 获批AI设备现状**：目前全美有大约 1000 款通过 FDA 认证的 AI/ML 医疗设备，其中绝大多数集中在医学影像与放射学领域。 - **软件产品形态的退化风险**：随着前沿大模型通用能力的渗透，许多 SaaS 工具退化为简单的“LLM 套壳 GUI”，过量采购会导致接口集成与版本维护成本剧增。 - **工作流本质的认知颠覆**：工作流是妥协的产物，医疗 AI 的终极方向不是单纯加速现有工序，而是基于临床“意图”消除不必要的人工参与环节。 ### [21:19](https://www.youtube.com/watch?v=c4lubsynndg&t=1279s) 采纳门槛与自研考量：EMR 工作流的绝对重力 在探讨医疗 AI 落地时，临床医生的工作流认知与工程师的产品思维往往存在脱节。许多方案在技术上看似合理，但在实际管理和临床路径中却难以融入。Adam Kaufman 强调，无论是医疗机构自身建设还是外部供应商销售，进入 AI 赛道都存在巨大的“准入门槛成本”——包括数据治理体系、底层基础设施、工具链配置以及人员技能培训。然而，这种投入具有极强的规模递延效应：组织在前期的基础建设走得越远，后续实施新用例的综合边际成本就越低。因此，战略重心不仅在于寻找一两个高价值的临床或管理单点用例，更在于搭建底层基础技术管线，使整个工作流能快速插拔新工具并实现自研规模化。 ```mermaid flowchart TD accTitle: Build vs Buy Decision Pipeline in Hospital AI subgraph Epic["Epic 核心生态"] E1["Epic 3年路线图评估"] --> E2{"是否在原生规划中？"} E2 -- 是 --> E3["优先采用原生/订阅模块"] E2 -- 否 --> E4["评估外部集成 / 自研"] end subgraph Workflow["临床采纳关键路径"] E4 --> W1{"是否需要切换窗口？"} W1 -- 需要跳出EMR --> W2["采纳率断崖式下跌 ❌"] W1 -- 无缝嵌入工作流 --> W3["临床实际采纳与推广 ✔️"] end ``` 针对自研与采购（Build vs Buy）的权衡，Krishnaj Gourab 补充了一个核心考量：医疗机构必须高度紧跟其电子病历（EMR）厂商（如 Epic）未来三年的产品路线图。他指出，临床实践中屡见不鲜的教训是，即使团队自研出功能极佳的 AI 独立系统，只要它独立于 EMR 之外、要求医生在看诊时切换到另一个窗口操作，临床采纳率就会遭遇断崖式下跌。因此，院内任何技术自研与采购的首要原则，都是必须无缝嵌入现有 EMR 的界面与工作流之中。 --- ### [23:13](https://www.youtube.com/watch?v=c4lubsynndg&t=1393s) Epic 自研环境感知转录震撼业界：平台巨头与专精厂商的博弈 Christopher Khoury 抛出了当时医疗 IT 界的一枚“重磅炸弹”：在即将于威斯康星州召开的 Epic 开发者大会前夕，Epic 正式宣布将在其原生平台内推出环境感知转录工具（Ambient Scribe），直接向行业内的第三方转录与语音 AI 厂商发起冲击。 ```mermaid flowchart LR accTitle: Healthcare AI Platform Architecture Options subgraph EMR_Native["方案 A：原生 EMR 闭环"] EpicApp["Epic 原生工作流"] --> EpicAmbient["Epic 内置环境感知 AI"] end subgraph Hyperscaler_Lake["方案 B：云端数据湖驱动"] DataSources["院内多源异构数据"] --> DataLake["云端数据湖 (Azure / AWS)"] DataLake --> CloudVM["私有安全 VM 运行大模型"] CloudVM -.-> EpicRecord["Epic 仅作为记录系统 (System of Record)"] end ``` 针对 Epic 的入局是否会彻底肃清市场，嘉宾们展开了深度剖析： * **对初创企业的毁灭性挤压**：Peter McCaffrey 认为这无疑是一次行业巨震。虽然 UTMB 作为 Epic 客户倾向于采用订阅模式尝试其新功能，但 Epic 很难在每个细分功能上都击败所有专精竞品。然而，这足以对腰部及初创企业关闭大门——如果一家初创公司既做不到 Epic 的原生集成，又达不到 Nuance DAX 或 Abridge 的规模，医院管理层就会倾向于认为“直接调用通用语音大模型（如 ChatGPT 语音转文字）或等待 Epic 原生功能即可”，小厂商的生存空间被严重挤压。 * **记录系统 vs. 业务计算引擎**：Adam Kaufman 指出，虽然环境感知语音非常依赖 EMR 原生界面，但许多大型医疗系统早在此轮生成式 AI 爆发前，就已经耗费巨资在公有云（如 Azure、AWS）上构建了独立的数据湖。大量非 EMR 数据沉淀在云端虚拟机内，医院更看重云厂商在数据安全、私有化部署环境上的承诺。未来的核心分歧在于：医疗机构究竟是将 Epic 仅仅当作底层“记录系统”（System of Record）并与外部大脑连接，还是将所有核心业务全面交给 Epic 原生驱动。 --- ### [25:49](https://www.youtube.com/watch?v=c4lubsynndg&t=1549s) 脓毒症模型的历史教训与微软 DAX 的正面对决 Krishnaj Gourab 与 Peter McCaffrey 对“Epic 必胜论”提出了审慎的质疑。Gourab 提醒大家回顾 Epic 当年推出的内置脓毒症预测模型（Sepsis Module）——以密歇根大学（University of Michigan）发表的权威临床验证为代表，多项独立研究均证实该原生模型存在严重的预测缺陷与误报漏报问题。这证明了平台巨头内置的 AI 并不天然等同于卓越的临床质量，专精型算法公司依然具备技术竞争空间。此外，医疗系统在技术更换上面临巨大的“黏性”（Stickiness），许多机构已经在 Nuance DAX 等方案上投入了大量资金与流程重塑，中途解绑并切换到新系统的转换成本极高。 Peter McCaffrey 进一步指出，Epic 正在遭遇两股强大外部势力的夹击： 1. **超大科技巨头的正面刚抢**：Nuance DAX 正在进化为深度融合在病历操作中的 Dragon Copilot，其背后站着市值近 4 万亿美元的科技巨头微软（Microsoft），其功能创新与算力支持直接与 Epic 形成正面交锋。 2. **互操作性法规对数据护城河的瓦解**：随着联邦监管政策的强力推进以及 FHIR（特别是 FHIR R5）标准的普及，法律上不再允许 EMR 厂商闭门锁国、垄断数据。在底层数据互操作性被政策强行打通、各大超大规模云服务商（Hyperscalers）长驱直入的背景下，Epic 夹在数据湖与临床应用之间的独占地位正在经受严峻考验。 --- ### [27:27](https://www.youtube.com/watch?v=c4lubsynndg&t=1647s) 资源悬殊下的破局：中小机构如何利用低成本 AI 与智能体突围 Christopher Khoury 提出质疑：头部学术医疗中心拥有顶尖资源和充沛预算，但对于广大中小型诊所、社区医院或资源匮乏的医疗机构而言，他们缺乏顶尖的 AI 算法专家与工程团队，是否注定在这场技术革命中被边缘化？ ```mermaid flowchart TD accTitle: Primary Care AI Agent Workflow for Referrals P1["初级保健医生 (PCP) 诊所"] --> Agent["本地/轻量化 AI 智能体"] Agent --> Process["自动扫描病历 & 整理前置检查"] Process --> Check{"术前材料是否齐全？"} Check -- 否 --> Complete["自动跟进补齐材料"] Check -- 是 --> Ready["无缝转诊至社区医院"] Ready --> Surgery["直接进入手术准备阶段 🚀"] ``` 嘉宾们针对不同层级医疗机构的生存策略进行了深度研讨： * **团队规模与技术门槛的解耦**：Krishnaj Gourab 坦言，UMMS 拥有由 20 到 25 名开发者、数据科学家和 AI 工程师组成的专业团队（iHARP/iHarbor），这种重资产配置确实难以被中小机构复制。但技术正在快速平民化——开源大模型与轻量化部署使得中小型诊所能够以极低成本在本地安全硬件内运行私有化模型，或通过签署商业伙伴协议（BAA）调用前沿厂商（如 Anthropic）的 API 搭建专属 AI 智能体（Agents）。 * **打通初级保健与手术转诊断点**：作为社区医院的首席医疗官，Gourab 分享了实际痛点：大量外科手术患者由初级保健医生（PCP）转诊而来，双方往往因材料缺失在繁重的文书往返中耗费大量时间。通过在 PCP 端部署轻量级 AI 智能体，自动处理转诊审批与术前检查缺漏项，患者在抵达医院时便已完成全部准备工作，大幅提升转诊效率。 * **阻碍创新的本质是流程黑盒**：Peter McCaffrey 指出，如今早已不是必须从头训练专属卷积神经网络（ConvNet）才能落地 AI 的时代，在 Claude Code 等开发者工具的辅助下，3 到 5 人的敏捷小团队完全可以实现惊人的交付效率。真正杀死项目的往往不是团队规模，而是组织内部的“未知阻碍”——不知道业务流程究竟归谁管、不知道核心数据究竟存在哪里、流程卡点无法厘清。 * **机构分化与单体社区医院的困境**：Adam Kaufman 对中小型门诊（PCP 或专科诊所）的转型保持乐观，认为其业务形态相对明确，主要通过“采购加配置”即可快速产生商业价值；但他对“单体综合社区医院”（Single-Hospital Integrated Systems）的前景深感担忧。这类单体医院在 AI 时代之前就已经深陷于劳动力短缺、成本攀升以及联邦医疗保险/医疗补助（Medicare/Medicaid）削减的重重危机之中，医院复杂的全业务流使得他们难以像轻资产门诊那样灵活突围，AI 的加速演进反而可能成为压垮这类机构的又一重负担。 ### [31:32](https://www.youtube.com/watch?v=c4lubsynndg&t=1892s) 区域医疗系统的破局之道：借助联盟力量下注 谈及中小型医疗机构是否会在这场技术变革中进一步落后于大型医疗集团，Adam Kaufman 指出，单打独斗确实难以为继，但区域性医疗系统可以通过加入各种医疗联盟来破局。例如梅奥诊所（Mayo Clinic）主导的联盟或其他区域合作网络，许多大型学术医学中心正在向基层及区域系统开放能力与支持。对于资源有限的机构而言，生存与发展的关键在于找准自身定位，融入现有的生态网络并借势而为，明确选定少数几个具有高回报潜力的切入点果断下注，然后顺应整个行业的技术浪潮共同前行。 ### [31:59](https://www.youtube.com/watch?v=c4lubsynndg&t=1919s) 未来交付模式与实体医疗的“生存危机” 主持人 Christopher Khoury 将讨论引向未来医疗交付模式的重塑。在过去几天的会议中，大家频繁听到关于“AI 原生”（AI-first）公司的讨论，甚至上一个场次还分享了一家完全基于 AI 构建和运行的钢铁切割企业，但医疗行业目前尚未出现类似的颠覆性实体。尽管亚马逊等科技巨头接连跨界涉足医疗，其核心运作依然高度依赖人工。美国医学会（AMA）始终将 AI 视为临床医生的“增强工具”，强调必须保持“人在回路”（Human-in-the-loop）乃至“人掌控全盘”。然而在大语言模型（LLM）爆发的当下，行业领导层对于传统砖石结构实体医院的未来走向正在展开更深层的探讨。 德克萨斯大学医学分校（UTMB）首席 AI 官 Peter McCaffrey 坦言，当他们与高管团队评估实体业务模式面临的“生存级风险”（existential risks）时，异步虚拟医疗（asynchronous virtual care）被公认为最大的威胁。他生动地举例道，休斯敦的德克萨斯医疗中心（Texas Medical Center）拥有全球规模最庞大的实体医疗“殿堂级”建筑群，然而其根基却正在被虚拟交互所撼动。无论是创业公司还是直接面向消费者的 ChatGPT，它们都在解决一个长期存在的痛点：传统医疗系统从来没有为患者提供过真正便捷易用的服务。 长久以来，医疗机构拥有足够的时间去自我革新，却始终步履沉重。如今患者正在主动寻找获取医疗服务的全新“入口”（on-ramp）。未来这个入口将越来越普遍地存在于每个人的口袋之中，化身为 Siri、生成式对话助手等虚拟工具。当 AI 逐渐包揽了医疗问诊的初始入口、诊疗过程乃至最终处方闭环时，传统的实体医疗机构甚至连建立临床医患关系的机会都被剥夺了。McCaffrey 警告说，如果传统医疗系统无法主动将自身工作流融入这种新模式——例如允许患者将与 ChatGPT 的问诊记录带入诊室，医生快速复核后直接出具处方以提升效率——那么实体机构必将在庞大的定向直接患者护理领域被彻底边缘化。 ```mermaid flowchart TD accTitle: 医疗入口的范式转移与实体机构的边缘化危机 subgraph S1["传统医疗路径"] A["患者身体不适"] --> B["传统挂号/漫长等待"] B --> C["实体医院就诊"] end subgraph S2["新兴虚拟路径（脱敏与旁路）"] D["口袋端交互<br>(Siri / ChatGPT)"] --> E["异步虚拟诊断/方案"] E --> F["线上完成处置/出具处方"] end subgraph S3["融合应对方案"] E -.-> G["结构化聊天记录导入门诊"] G --> H["临床医生快速复核确认"] H --> I["保持医患纽带与合规落地"] end ``` ### [34:25](https://www.youtube.com/watch?v=c4lubsynndg&t=2065s) 像亚马逊一样便捷：重症留滞与从“手术”到“药物”的范式转移 马里兰大学医疗系统（UMMS）的 Krishnaj Gourab 对此高度赞同，他强调就医体验必须进化到如同网上购物一样丝滑流畅。面对日常简单病症，患者自然会倾向于线上寻求即时医疗服务，这直接导致传统医院面临另一重生存危机：轻症患者被数字渠道分流，留在实体医院里的全都是“最危重的重症患者”（the sickest of the sick）。重症监护救治难度极高，且在现行医疗体制下的运营成本与收入风险极其严峻。如果实体医疗不能迅速完成自我演进，让患者像在亚马逊购物一样轻松地与医院系统互动并即时获得反馈，而是继续让患者苦等电话回拨，医疗机构的财务平衡与生存基础将被打破。 Adam Kaufman 进一步补充了宏观结构上的技术与产业趋势。他援引咨询委员会（Advisory Board）关于未来十年的五大趋势预测，其中极为核心的一项正是“医疗重心从手术操作向药物治疗的转移”（shift from procedures to therapeutics）。长久以来，实体医院的核心价值在于“将患者收治进来，对其实施某种侵入性操作”。然而，一旦诊疗模式转变为患者自行服药即可解决问题，或是由护士上门进行输液护理，实体建筑存在的根本意义就面临拷问。诸如 GLP-1 这类颠覆性新药的普及，正从底层深刻重塑患者对实体就医的需求。 Kaufman 犀利地指出，当前支撑整个传统医疗帝国运转的底层支柱，本质上依然是按人工作业时间付费的医保报销规则（Fee-for-Service）。目前大多数支付仍然要求某位具体执业医生提供特定时长的服务才能出具账单，这使得实体医疗在与单纯自费的数字医疗平台竞争中暂时占据优势。然而，一旦联邦医疗保险（Medicare）在未来几年内突然改变支付规则，不再强制为纯人工耗时买单，外部竞争对手将迅速对传统医院形成降维打击。 ### [37:06](https://www.youtube.com/watch?v=c4lubsynndg&t=2226s) 提问环节：电子病历中的“劣质数据”与 AI 的双重标准 在观众提问环节，一位参会者尖锐地指出了医疗数据的真实困境：大家都在关注 AI 模型的幻觉，但在实际落地中，喂给模型的核心数据源通常是电子病历系统（EMR/EHR），而众多一线医生私下反馈 EMR 中充斥着大量的虚假、冗余或错误数据。从机器学习角度看，“垃圾进，垃圾出”（Garbage in, garbage out），底层脏数据若无法解决，模型构建将面临致命危机。 Krishnaj Gourab 率先回应了数据工程的现实投入。他以 UMMS 旗下的创新孵化团队 IH Harbor 为例解释道，当团队开启一个新项目时，前 3 到 6 个月往往需要 4 至 5 名专职人员完全扑在数据清洗上。在 AI 时代，算法推理高度依赖输入特征的精准度，如果数据未经过彻底治理就直接灌入模型，后续输出的结果将完全无法被临床信任，因此医疗机构必须在数据预处理阶段耗费极大的精力和成本。 Peter McCaffrey 则从认知哲学与实际效用的角度给出了不同的反思。他认为，整个医疗界目前对 AI 存在严重的“双重标准”。一旦 AI 产生非零概率的幻觉，所有人都会将其视为不可逾越的鸿沟；但回顾整个人类医疗史，人类医生的临床决策表现始终存在着巨大的方差和不稳定性，既有电子病历在 AI 诞生前本就充满了错漏与主观“幻觉”。McCaffrey 强调： - “我们的目标应当是达到与人类平齐的标准（par），而不是追求绝对的完美。” - 医疗创新的首要任务是尽快将价值交付到迫切需要的患者手中，过分苛求零失误只会导致技术停滞。 - 面对混乱的现实世界数据，工程团队无法在一夜之间将历史真实数据净化得毫无瑕疵，因此关注点应该从“追求绝对真实的数据源”转向“利用推理能力、多步反思架构（reflection）、增加 token 算力消耗以及引入辅助数据源”来辅助推导，通过模型自身的逻辑校验来抵御底层脏数据的干扰。 ### [39:21](https://www.youtube.com/watch?v=c4lubsynndg&t=2361s) 落地实施与变革管理：如何让临床带头人不再觉得“被强加”？ 来自马里兰州蒙哥马利县 Adventist HealthCare 的一位项目管理与实施负责人提出了关于变革管理的实操难题：在院内推行 IT 和 AI 落地时，临床倡导者（Clinical Champion）或行政管理领袖的全程深度参与决定了项目的成败。如果在项目初期未能深度绑定一线医护，实施团队往往会遭遇严重的心理抵触，医生们会抱怨“这是你们 IT 部门强加给我们的工作”。提问者希望专家分享，如何从一开始就将临床专家拉到同一张桌子前共同构思用例，实现“携手合作”而非“单向强加”。 Adam Kaufman 结合他在贝勒医疗系统（Baylor）的经历指出，解决这一矛盾的根基在于自上而下对齐战略目标（KPI）。高效的医疗机构不会凭空向一线推销某个软件，而是将全院级核心指标（如门诊候诊时间过长、预约取消率过高）层层分解到科室与个人。当临床医生、行政管理层和 IT 团队基于同一个痛点达成共识时，讨论的语境就不再是“IT 逼迫医生使用新工具”，而是“我们共同在解决一个大家都深恶痛绝的问题，现有的这个解决方案是否合理”。 Peter McCaffrey 和 Krishnaj Gourab 则从组织架构与权力下放角度给出了补充： - **设立专门的领军职能**：医院必须在顶层设立具备跨界话语权的正式职位，无论是首席 AI 官（CAIO）、首席医疗信息官（CMIO）还是两者的结合体。如果机构缺乏一个能够统揽战略、为临床发声并明确界定资源倾斜方向的专业领袖，任何变革在复杂的医院政治中都会寸步难行。没有这些岗位的机构应当立即设立。 - **自下而上的用例来源**：业务用例必须真正萌生自一线临床医生的实际工作痛点，由医生提出设想，再经由技术治理委员会进行专业评估与工程转化，唯有让临床团队成为创意的原作者，才能彻底消除推行过程中的排异反应。 ### [41:48](https://www.youtube.com/watch?v=c4lubsynndg&t=2508s) 临床 AI 项目的治理决策机制与圆桌结语 在评估与推进具体的医疗 AI 项目时，严密且以临床为中心的治理流程是确保技术落地的关键。以针对医生的应用场景（physician-centric）为例，团队首先会将提案提交给整个医疗系统的首席医疗官委员会（CMO Council）。在该委员会上，团队会详细陈述项目的核心构想与潜在价值，征询委员会的专业意见：“这是一个绝佳的想法，你们认为我们是否应该进行试验和试点（pilot）？”如果 CMO 委员会给出肯定的答复，团队才会着手进行构建或进一步考察相关产品；如果未能获得认可，该项目就会直接被否决。这种严格的自上而下与临床专业把关机制，确保了技术资源始终聚焦在真正有临床价值的需求上。 随着讨论进入尾声，主持人对全体专家嘉宾表达了诚挚的感谢。主持人总结道，这是一场内容极其丰富、观点多元碰撞的深度对话，不仅自己从中受益匪浅，也希望现场的每一位观众都能满载而归。在全场热烈的掌声中，本次关于医疗机构高管 AI 战略与医疗变革的圆桌研讨会圆满结束。 - **临床主导的准入决策**：针对医生端的 AI 项目必须经过系统级 CMO 委员会的审查与批准，只有获得临床领导层明确的试点许可后才会立项开发或采购。 - **全场圆满落幕**：涵盖医疗系统高管、AI 负责人及行业协会专家的深度探讨为医疗领域的 AI 转型提供了全方位的实战视角与战略指引。 ## Notable quotes > [04:41](https://www.youtube.com/watch?v=c4lubsynndg&t=281s) 低估这场变革所带来的“复利式恶化”（"compounding badness"）是任何医疗系统都无法承受的战略误判。 — Peter McCaffrey > [12:48](https://www.youtube.com/watch?v=c4lubsynndg&t=768s) 如果你身边配有一名极聪明的实习医生或护理助理，你会希望他帮你做什么？ — Krishnaj Gourab > [17:26](https://www.youtube.com/watch?v=c4lubsynndg&t=1046s) 人们常说‘一切都取决于工作流’，但现实并非如此——没有人真正按‘工作流’去思考。人们思考的是‘意图’（Intent）以及他们希望达成的结果；工作流只是‘我想要的’与‘我不得不应付的现实’之间的妥协。AI真正的价值不是去优化这些琐碎的工作流，而是将人类从繁琐的操作中彻底解脱出来。 — Peter McCaffrey > [12:48](https://www.youtube.com/watch?v=c4lubsynndg&t=768s) 如果全局来看价值不足那是管理层的责任，但参与使用是临床医生的职责。 — Peter McCaffrey > [37:06](https://www.youtube.com/watch?v=c4lubsynndg&t=2226s) 我们的目标应当是达到与人类平齐的标准（par），而不是追求绝对的完美。 — Peter McCaffrey > [41:48](https://www.youtube.com/watch?v=c4lubsynndg&t=2508s) 这是一个绝佳的想法，你们认为我们是否应该进行试验和试点（pilot）？ — Krishnaj Gourab > [04:41](https://www.youtube.com/watch?v=c4lubsynndg&t=281s) 将本院打造为全州最安全的医疗系统 — Krishnaj Gourab > [39:21](https://www.youtube.com/watch?v=c4lubsynndg&t=2361s) 这是你们 IT 部门强加给我们的工作 ## Mentioned Christopher Khoury, Adam Kaufman, Krishnaj Gourab, Peter McCaffrey, American Medical Association (AMA), Private AI, Mayo Clinic, Mayo Clinic Platform, Baylor Scott & White Health, University of Maryland Rehab and Ortho Institute, University of Maryland Medical System (UMMS), IH Harbor (iHARP), University of Texas Medical Branch (UTMB), UT Real Health AI, National Institutes of Health (NIH), Epic, Microsoft, Nuance DAX, Abridge, Dragon Copilot, Amazon, Claude Code, Anthropic, ChatGPT, Adventist HealthCare, Texas Medical Center, University of Michigan, FHIR, Medicare, Medicaid, FDA ## Video description Krishnaj Gourab, Chief Medical Officer - University of Maryland Rehab and Ortho. System Medical Director, Post Acute Care - UMMS. Adjunct Faculty - Biomedical Informatics and Data Sciences, Johns Hopkins University., University of Maryland Medical System Peter McCaffrey, Chief AI Officer, The University of Texas Medical Branch Adam Kaufman, Board Member, Private AI Christopher Khoury, Vice President of Strategic Insights, American Medical Association Abstract In this session, healthcare executives will discuss how they are navigating the strategic transformation of their organizations in the age of AI. Panelists will explore the role of AI in shaping future care delivery models, improving operational efficiency, and driving innovation. Learn how executive leaders are aligning AI initiatives with their broader organizational goals, managing change, and fostering a culture that embraces AI-driven healthcare. This session will provide insights into the leadership strategies that are positioning care providers to succeed in an AI-driven world. [Source](https://www.youtube.com/watch?v=c4lubsynndg)

[4:41] Peter McCaffrey: 低估这场变革所带来的“复利式恶化”（"compounding badness"）是任何医疗系统都无法承受的战略误判。 [12:48] Krishnaj Gourab: 如果你身边配有一名极聪明的实习医生或护理助理，你会希望他帮你做什么？ [17:26] Peter McCaffrey: 人们常说‘一切都取决于工作流’，但现实并非如此——没有人真正按‘工作流’去思考。人们思考的是‘意图’（Intent）以及他们希望达成的结果；工作流只是‘我想要的’与‘我不得不应付的现实’之间的妥协。AI真正的价值不是去优化这些琐碎的工作流，而是将人类从繁琐的操作中彻底解脱出来。 [12:48] Peter McCaffrey: 如果全局来看价值不足那是管理层的责任，但参与使用是临床医生的职责。 [37:06] Peter McCaffrey: 我们的目标应当是达到与人类平齐的标准（par），而不是追求绝对的完美。 [41:48] Krishnaj Gourab: 这是一个绝佳的想法，你们认为我们是否应该进行试验和试点（pilot）？ [4:41] Krishnaj Gourab: 将本院打造为全州最安全的医疗系统 [39:21] 这是你们 IT 部门强加给我们的工作

[0:04] Good morning or good midday here. Uh my name is Christopher Corey. Um really [0:09] excited to have this uh panel and my esteemed uh panelists uh on stage with [0:14] us today and great to be at AI4. So again we're talking about AI strategy [0:19] for care provider executives. Before I introduce my panelists, I'm just curious [0:23] in the audience here, do we have any practicing clinicians or those in health [0:27] system settings here? So we have a good number. That's that's great to see. [0:32] Welcome. And um folks that are in health systems but in nonclinical roles and [0:37] other technical roles also a good number. This is great. Well, [0:40] it's the definitely the uh the right audience for this conversation here. Um [0:45] so uh again just a a second about me. I lead up a group called the strategic [0:51] insights uh department at the American Medical Association. and we're the [0:55] largest physician uh voice in this country uh and work on a number of [1:00] issues uh emerging issues in healthcare uh in addition to uh physician needs uh [1:06] of all sorts. Uh let me start uh to my left here, Adam, and ask you to [1:10] introduce a little bit about yourself. Yeah, [1:12] >> sure. Pleasure. Well, thank you for having me. Thank you to AI4. Uh I'm [1:17] currently a board member at Private AI. Uh we're a company that helps [1:22] organizations of all types but focus mostly in healthcare uh realize value [1:26] out of their unstructured data. So help identify particular uh entities inside [1:31] that data uh and then help organizations act. One of the most common use cases [1:35] would be uh you know identify PHI redact it and either give you back the redacted [1:41] version or put some kind of synthetic data. Uh but we find that that model [1:45] applies to a lot of other use cases as well. Uh and then I spent a lot of time [1:49] consulting and advising uh both pharma and medtech uh and have spent the last [1:53] five years with large health systems. Uh first working with Mayo Clinic and the [1:56] Mayo Clinic platform uh so familiar with their AI work there uh and now I do a [2:01] lot of work with Baylor Scott and White the system out of uh Texas. [2:05] >> Dr. Gab, thank you. Yeah, I'm glad to be here and uh I'm Krishn Gorab. I'm a [2:10] physical medicine rehab doctor by occupation mostly treat patients with [2:14] stroke, spinal cord injury, etc. I'm also a chief medical officer at uh [2:18] University of Maryland Rehab and North Institute, one of the hospitals of the [2:22] University of Maryland Medical System. My AI hat is that inside University of [2:27] Maryland Medical System, we have IH Harbor, that's our innovation arm and I [2:31] serve as the medical head for uh digital products and AI in in the innovation arm [2:37] and I also sit in the AI governance committee. So as as we all know it's [2:41] very exciting time of in AI and clinical medicine exciting and sometimes scary. [2:46] So I try to keep my hands on all of those spots. So looking forward to this [2:50] session. Dr. McAffrey. Hey [2:53] >> everyone nice to be here and talk to you all. So my name is Peter McAffrey. I'm a [2:58] pathologist by training. I'm coming here from University of Texas. Um I'm our [3:02] chief AI and digital officer at University of Texas Medical Branch. uh [3:06] and I also chair now our initiative called UT real health AI across all [3:10] eight of our UT system uh what we call health related institutions and really [3:15] my focus is where and how can AI drive value provide you know safe and governed [3:20] and so forth where can it drive value in our enterprise really in anything from [3:24] surgical planning patient care wayfinding billing optimization and so [3:28] forth so every day it's sort of like use case scoping and value prediction [3:34] >> great and I just want to paint a little bit of a picture as we get into this [3:38] discussion today around the context that all these our our panelists here have [3:43] been and are working in today. So we're this is 2025. We're in a really dramatic [3:49] transformation period uh broadly in this country, but certainly for what our [3:54] health systems look like. There's a lot of forces at play ranging from dramatic [3:59] cuts to Medicaid to those of you in academic settings, NIH funding cuts also [4:04] dramatic um you know different um workforce challenges around shortages [4:09] around clin uh clinician nurse burnout uh and other workforce challenges [4:15] chronic disease burden both pediatric and adult populations uh now among other [4:21] sort of political and social dynamics here. So health systems may be [4:26] potentially the most complex you know social political economic uh in [4:31] environment uh today and so with that context we're talking about more [4:37] transformation now we're talking about implementation of AI in a very complex [4:41] uh organic system so I want to start at the top here uh to get your thoughts on [4:47] uh you know at what in your experience how have you built a strategic vision [4:53] inside your organization that buy in at the top leadership level because we know [4:57] these types of transformations don't really work without that uh support and [5:01] that vision at the top. So I wonder if you can you can comment and I happy to [5:05] start with uh the folks. Yeah. Uh Peter, do you want to take that for starts? [5:10] Yeah, >> sure. Yeah, I I think one of the things [5:12] that has helped gain uh sort of alignment and buyin from leadership is [5:17] that everyone I think appreciates there's a uniqueness of this moment in [5:22] time that the type of transformation we're seeing is different what we've [5:26] seen before that its impacts are different. So we appeal to that. You [5:30] know, we say you look at the the march of capability and generality of AI, [5:34] especially LLMs over the past, you know, almost three years. And you could be [5:39] wrong, but you'd have to at least acknowledge it could be a continual [5:42] expansion of capability. And if that's the case, you know, you would make [5:46] really bad missteps to miscalculate the extent that will transform. So we appeal [5:50] to these kind of fundamental things, which is workflows have to be [5:54] simplified. We have to make data more easy to access with agentic tools, LMS, [5:59] integrations and so forth. And that if we believe the real common core [6:03] reasoning, abstraction of these tools grows at an exponential rate and the way [6:09] we kind of phrase them is that if we fell off of that growth curve, then our [6:13] technical debt and our negative margin grows at an exponential rate. And we [6:18] don't have the luxury to be like wrong on that. Uh it would be a compounding [6:22] badness to be wrong on that. Yeah, please. [6:26] >> So, I agree with what Peter said. I love the com comment about compounding [6:30] badness. That's absolutely true. Uh apart from this, I mean what we try, it [6:37] takes investment both for governance of AI, investment of time and money as well [6:42] as creating AI products for our own population, our own data set which are [6:47] unique to us. So, how do we get executive buy in? We try to tie whatever [6:52] we are doing to our uh strategic goals. So for example, I'll give you an [6:56] example. Our CEO said that a couple of months back that we want to be the [7:01] safest healthcare system in the state and the question becomes can we re go [7:06] there by doing what we have done in the past and the answer is no. So what can [7:11] we do next? So imagine now we call this a uh virtual resident or or let's say an [7:17] intern. AI now has the capability of digesting structured data as well as [7:23] clinical notes has the capability of having the institutional knowledge base [7:28] about what works in a patient which number to call to get an MRI which are [7:32] the MR slots and we can build this things that can be awake at all times [7:38] doesn't go to sleep does not have duty hours and so on so forth and hence we [7:43] make the case of investing in these kind of technologies is with the ultimate [7:49] goal that will serve the big strategic topline goals. So that's how we try to [7:54] get buy in and that we have been fairly successful in that strategy till now. [7:58] >> Yeah, >> I think it's it's a really important [8:01] question but if you'll allow me I think there's a set of AI tools that really [8:07] probably don't need to appeal to the health system vision at the top level. I [8:11] mean I think we've laid out really compelling and at least a framework I [8:15] take is you know there's there's sort of clinical AI tools and um a lot of those [8:20] are pending uh you know regulatory approval right so if a computer vision [8:25] tool can replace or augment what a pathologist is doing I don't think the [8:29] health system is thinking at a at a vision level they're saying okay that's [8:32] now approved and it does or does not make sense in our uh in our setting uh I [8:37] think probably the the use cases we're most familiar with are the [8:39] administrative AI tools tools, ambient listening. Again, I don't think we're [8:42] appealing to the vision of the health system to say we can make our teams more [8:46] efficient. We can do nurse scheduling there. There's a whole bunch of stuff. [8:50] And then the last is, you know, we can think about foundational AI. So, you [8:53] know, AI that is enabling other processes and often that's at the, you [8:58] know, the IT teams themselves or the engineering teams are making those [9:01] decisions. There are some set and you know, the University of Maryland [9:05] focusing on safety. I can say at my time at Mayo we would say we want to be the [9:09] leading health system from an innovation the most complicated cases Baylor will [9:13] say something like we want to be the most customer centric and so you will [9:17] see different health systems picking different kinds of AI tools that apply [9:21] to their vision and leaning in uh and so you get different focus but there's a [9:25] lot of work that you know is is just trying to make us all better more [9:29] efficient more effective that doesn't need to go to that vision. [9:33] Yeah, that's great. Then thinking about that that uh dichotomy [9:39] between administrative uh a whole swath of administrative tools and solutions [9:44] and clinical tools. I mean uh health systems tend to be relatively riskaverse [9:48] for for these uh new forayas in technology. Um are they is it sort of [9:55] easier to build the case around ROI and investment commitment around [9:58] administrative tools because of this kind of perceived sort of lower risk [10:02] approach or are as you said are there organizations that are really saying we [10:06] want to be you know clin you know clinical leaders that that uh are able [10:12] to move the needle on patient care in a different way. What's been your [10:16] experience? Go ahead. >> Yeah, good question. Uh well I think the [10:19] ROI is sometimes easier to describe on the administrative workflow. So things [10:24] like improving the efficiency of appealing a denial or we've invested a [10:28] lot in AI for like bed status adjudication. So like utilization review [10:32] these have clear very proximal KPIs that are really translated into revenue money [10:38] versus savings. That's an easy appeal to make. But on the clinical side to my [10:44] surprise we've seen other things really take off that people interested in as [10:46] well. So like at at our institution, we want to be able to say that anyone who [10:51] has an action item that should be done, they're not sitting there with it not [10:55] being done. These are things like finding these gaps of care. Who's the [10:58] and some of them are very revenue adjacent. Who's our ICD candidate? Our [11:01] taver saver watchman candidate and so forth. Okay. Right. Yeah. Or things [11:05] like, you know, who are the people with radiology actual findings and [11:08] follow-ups. You said a PET CT didn't happen. How did it happen? Can we comb [11:11] through that? It meets an ethical imperative, but it's also revenue [11:15] adjacent and it's very minimal to human in the loop. So, we can kind of absolve [11:19] the risks into human adjudication and review and that's taken off really well [11:23] as well. Uh I think those I mean from my experience people are pretty open to [11:29] leaning into those kinds of use cases and those themes even now. um and things [11:34] like you know uh imaging AI for stroke P and so forth have a very strong evidence [11:38] base and and also not a hard sell I think uh to the enterprise [11:45] >> um yeah I would agree and one of the things that uh has gotten the most [11:49] traction as you mentioned denials management many hospitals across the [11:53] country use AI for denials management and mitigation even before we submit the [11:58] claim is everything okay or not so that's a purely administrative task [12:02] has got a good uh uptake and radiology and on the other side clinical I the [12:08] radiologists are probably the most ahead they're the farthest in AI adoption in [12:14] terms of detecting the area of the stroke triaging of uh which film should [12:19] be read first and chest X-rays and all kind of use cases and they're all [12:23] clinicians they're using it and I think now what has changed is that with the [12:26] use with advances in NLP and LLM use of LLMs responsible use of L&Ms what [12:33] radiology was able to do by reading the numbers of the pixels or vox cells we [12:38] can actually read and interpret the nodes and hence is the opportunity of [12:43] taking this adoption to a wider variety of clinicians [12:48] thank you I want to note a couple things I heard there one was um the ethical [12:52] imperative Peter that you mentioned it's good to hear that in a setting like this [12:56] that that's part of the decision making um so I I appreciate that and we might [13:01] want to come back to the that theme um as as we go further here. Um [13:06] I want to think uh ask you all to think about clinical implementation sort of in [13:11] the real world environment in healthcare and that means adoption right that means [13:15] getting um you know uh clinicians staff uh administrative staff to to use new [13:22] tools that potentially break their workflows or require retraining and so [13:26] on. Um, but certainly on the clinical side, I'm curious to hear how do you get [13:32] that clinical workforce to change the way they're doing things. Um, and I'm [13:38] also wondering, especially at this conference, we hear a lot about [13:40] frameworks and elsewhere too around guidelines, frameworks, model cards, [13:46] right? Different sort of uh um tools in the toolbox. Do any of those help or do [13:53] you have a better approach there to engage the clinical workforce for [13:56] adoption? >> Yeah. [13:58] >> Yeah. It it's hard. I'll tell you like just as one classic example where [14:03] adoption didn't go the way we thought was certainly in generative and basket [14:06] responses. You know this first we thought it obviously saves time and then [14:09] you realize that it's a lot of variable physician preference. Some don't want to [14:13] use it and so forth and so forth. And it's been true of ambient. It's been [14:16] true of M365 co-pilot and so forth and so forth. It's a diffuse. It's a broad [14:21] user base. We try to avoid deploying to a broad user base if we can, but some of [14:25] these things are for a broad user base. Uh what we've really tried to do is, you [14:30] know, very attentive like user groups and meetups and how you're doing it and [14:33] demo and demo and show and demo and show and over and over and over again. And [14:37] this helps to some degree, but like with anything with your turnover, it decays [14:41] as well. There's a challenge there. What I think what we've talked about, we've [14:45] not done this, but I think was very interesting is we begin to monetize and [14:49] reward some level of AI utilization >> almost like an RVU, like an AI RVU, like [14:55] if you use it at least, we're going to support you for that. It's on us if it's [14:59] not valuable globally, but it's on you to participate and that we can actually [15:03] drive. And I think we'll probably get to do something like that in the next year [15:06] or so. >> Yeah. in a RVU that's an interesting [15:09] concept to think some more about any other comments there. Yeah. [15:14] >> Yeah, I think that's a very interesting concept. A RV. Yeah. The uh one thing [15:19] people don't like. So, one of the things that has happened with AI is that there [15:24] is some amount of especially from the clinicians probably we are all burned by [15:28] the implementation of the EMR is that uh there's some amount of push back. Okay. [15:34] AI hallucinates. AI will do decisions for me. I think education is a key and [15:40] often we when we go to implement something we try to say forget it's AI, [15:44] forget its algorithm or software. If you had a very smart intern with you or a [15:50] very smart nursing assistant, what would you want this person to do and then take [15:55] it back to our use case and we have had some successful implementation and some [15:59] implementations that have not been successful. But I think as we are going [16:03] forward with this, I think educating our workforce is becoming the key so as they [16:08] can utilize these things effectively. >> Yeah. Go ahead Adam. And I think just to [16:12] add um you know healthcare is not particularly [16:16] digital to begin with or I'd say we're in significant phase of still sort of [16:21] implementing digital as we know it in other industries and so AI is coming [16:26] faster probably into healthcare given the maturity of that and so we have a [16:30] behavior change around just your patient might be doing self-management things we [16:35] might be sending them education they're getting emails they're getting a lot of [16:38] stuff that is you know didn't start yesterday but it's relatively new uh and [16:42] so AI brings I think to the point uh that they were raising questions about [16:46] hallucination can I trust the model but a lot of this is just standard change [16:50] management that we're doing with the basic technologies that we're rolling [16:53] out in addition and AI often is invisible I mean at Baylor we've [16:58] implemented a very powerful you can go onto our app and find essentially a next [17:03] best action you know a triage tool that takes clinical decision support has some [17:08] AI modeling will take you to a nurse There's some AI behind it, but most [17:13] people aren't thinking about that. They're just thinking this is actually a [17:16] better way than calling the nurse call center every time to know where to go. [17:19] >> And so I think a lot of it is embedded in that kind of processes that we need [17:22] to educate our staff and our patients to to understand. [17:26] >> Yeah. Thank you. Um and just you one data point out there. So we're we're [17:30] talking about you know the clinical administrative AI. There's also the [17:33] world of AI and medical devices, right? Especially in uh some of your worlds. [17:37] There's something like,00 devices today cleared by the FDA that [17:43] claim to have some AI or machine uh learning component to them right now. [17:47] Most of them are imaging and radiological but I'm also interested to [17:51] hear that how organizations are developing their own um you know [17:57] homegrown models. Is that a viable path? Um uh what what's sort of dominating the [18:03] discussion right now in a health system? Is it about these medical device vendors [18:08] that are saying use our stuff or something else? Go ahead, Adam. Yeah, [18:12] >> I I can speak to a little to that because uh this is part of the work that [18:15] we did at the Mayo Clinic platform and I' I've had the good fortune to advise a [18:19] couple uh you know computer vision technology companies. [18:22] >> Can you can you say one second about the Mayo Clinic platform for the audience? [18:25] >> Yeah, sure. So Mayo Clinic platform is a division of the Mayo Clinic. It's at the [18:29] level of its uh you know its cancer center and other research centers really [18:32] dedicated to um you know making Mayo Clinic and a number of other you know [18:38] worldrenowned health systems data first available to people building solutions [18:42] in a safe and secure way and then building we'd call it the piping or the [18:46] infrastructure so that those solutions can be deployed and part of that is to [18:50] give those health systems that are using it the ability to build their own [18:52] models. Um, what I would say that, and this is not a Mayo statement, it's been [18:57] a while since I've been there, but I would say in the quality improvement in [19:01] the workflow, I think it's very very viable for health systems to build their [19:05] own models. Right? This is mostly about access to data. Can they get the [19:09] infrastructure to do it and and that local model maybe trained with another [19:13] foundational model is probably better at the actual level of therapeutics of [19:18] diagnostics. I don't think most health systems will do that because you have to [19:22] go through a complicated FDA process. You're going to fight a reimbursement [19:25] issue. You're going to come to the AMA and ask for a CPT code. And so I see [19:29] those applications being built by third parties. [19:33] >> Yeah, I I would echo a lot of that. I think that you know our our stance has [19:36] been on build versus buy is that one of the big consequences if the trend of [19:43] increasing generality of at least frontier models and LMS continues to [19:47] grow even like at the pace machine or or frankly doesn't even grow but just [19:51] penetrates further into the workflow is that it really blurs the significant [19:56] margin of like what is a product actually because a lot of times you know [20:00] there were products where well it's a guey it's a gooey and a database and an [20:03] application layer and now It's kind of like more like a gooey wrapping an LLM. [20:07] And at some point there's a level of abstraction where it's like just bring [20:12] the tokens to the data as cleanly and consistently as possible. We've started [20:17] working more with like platform type either products or even companies and [20:22] doubling and tripling down on like a certain system of you know an agentic [20:27] pipeline or review abstraction action type of systems and many many many many [20:31] workflows where they are similar because we're afraid that if we don't do that [20:35] and we invest in kind of the gooey product that seems to fit the workflow [20:40] now we will have this proliferation at that level and then we will be [20:44] encumbered again by the integration tax and the version parity tax and all these [20:49] things from those living at that at that level. There is this axiom we often say [20:54] which is you know it's all about workflow but I think the reality is it [20:58] isn't I have to say it isn't actually about workflow. No one thinks in [21:01] workflows. They they think in intent and what they want to happen they [21:06] communicate it work by workflow. It's a compromise of what I want and what I [21:09] have to deal with. And I think what we're going to see with AI is like some [21:13] of these don't need to be optimized. They need to be completely taken away [21:16] from human participation for everyone's betterment, especially on the [21:19] administrative side. And that really challenges where does a product fit if [21:24] no one even sees the workflow anymore. I >> I love that point about no one thinks of [21:28] it in workflows. I wonder if there's sort of a an engineer mindset clashing [21:32] with a clinician mindset or something happening there with some of these [21:35] solutions and just reconciling that. Yeah, great point. [21:40] Could I just make a comment that coming back to cost because I think what Peter [21:43] is saying is really important is as you think about your own health systems or [21:47] for those of you selling to health systems there is a cost to just be in [21:52] the AI game right whether we talk about governance whether we talk about the [21:55] infrastructure whether you talk about tool sets whether we talk about the [21:58] staff knowing it and so the further along the organization is in that the [22:03] cheaper and I don't mean that just money the cheaper in all senses it becomes to [22:06] implement next so there's really a return to doing this. Um, and I think [22:11] that's why a lot of people think about what are the easy places to go. Uh, but [22:14] I think to Peter's point, a lot of that is can I get a a use case that's [22:18] valuable to a clinician? Can I get a use case valable administration, but at the [22:21] same time, can I build foundational technology that make the whole whether [22:25] you call it the pipeline or the process or the workflow easier at all points [22:28] because then you can start plugging things in and and really get that rapid [22:31] scale and can start doing it for yourselves. [22:34] >> I'll just add one thing to this buy. We always struggle with the buy versus [22:38] build and uh is that we also look very closely to what our our EMR vendor is [22:43] epic. What is their u their u road map over the next three years what are they [22:50] promising to deliver how realistic is that because we have often seen that [22:54] when we build a product that is outside the EMR clinician has to go to another [22:58] window if such if they have to the adoption even if it's a great product [23:03] the adoption doesn't happen as well. So even if we build our products, we try to [23:07] deliver it within the EMR or we see if Epic has a better product or they are on [23:11] the road map. That's another consideration. [23:13] >> So that's a very timely point. I I'll come back to you Adam in a second. Um, [23:18] and not really in our original discussion here, but Epic kind of [23:21] dropped a bombshell in the past week. And there's I think the Epic user [23:24] meeting starts uh next week in Wisconsin, uh, which is the the offering [23:30] of an ambient scribe tool inside the native Epic platform as I understand it [23:35] and sort of, you know, sort of shaking up the foundations of the other players [23:39] in that industry. any any reactions there about whether that is a small [23:44] deal, a big deal or or whether there's still competition or room for [23:48] competition? I I think it's definitely a big deal. [23:52] Um, so we're Epic customers well and sort of our philosophy with this is, you [23:55] know, Epic's AI can be mostly consumed. I don't know about ambient what they [23:58] will do, but like as a subscription based model, so why not just try it all? [24:02] That's kind of our thing. >> We have no intention to sort of compete [24:05] with Epic obviously even in what we procure elsewhere. But the reality I [24:09] think that is challenging is are you gonna go into every one of these areas [24:13] and be better than the sum of all competitors in every area and it's [24:16] always probably not. What's the middle ground? I don't know. We'll certainly [24:20] try it. I think it would probably make it harder though for new incumbents in [24:25] this market without a doubt because if you're not Epic or if you're not DAX or [24:29] a bridge and you're small but you have a great feature set then it's like well if [24:34] you're small enough then can't I just use like chat GPT voice to text and so [24:38] it's one of these things where it I think it's going to shut the door on a [24:42] lot of people >> it was actually the comment I was going [24:45] to make and I think it fits in this so I think ambient listening is probably a [24:50] good case of Epic will win. To Peter's point, it's in the workflow of your EMR, [24:55] but a number of the health systems I work with have spent a lot independent [24:59] of AI to build a comprehensive data lake because they have a lot of information [25:03] that isn't in their EMR and they're running most of the new models against [25:06] the data lake originally and not Epic. And so, this is where I'm saying [25:11] understanding the infrastructure of the system where you are will have a lot to [25:14] do with what tooling you need. Uh, a comment I made when we were preparing is [25:18] um the health systems I've seen are locked into Epic. I've only worked with [25:22] Epic Systems, but they're actually more interested in what Azure is providing or [25:25] what AWS or where their cloud provider is because that's where their data is. [25:29] They get a guarantee of security. It can live inside of their virtual machine. So [25:33] I do think the big platform players will win. But whether it's Epic and the EMRs [25:37] or, you know, our more traditional tech vendors, I think depends on whether you [25:42] ultimately think of Epic as a system of record and you need to connect it to [25:45] others or you actually going to run most of your business out of it. [25:49] >> Just to add to that two comments. One is that Epic if you remember Epic came up [25:53] with the sepsis module too. >> Yeah. And of course we all know from the [25:58] publications of University of Michigan mostly and others too that there were [26:02] significant deficiencies in that model. So lot of the there have been models by [26:08] other players which are probably better or are competing now. So we will see [26:12] what what the offering from Epic is. It's not a given that if Epic is making [26:18] something as you said it will be better than everybody else. So that's part one. [26:23] >> Yeah. >> Second. [26:24] >> Yeah. Go ahead. you did >> Hello. [26:26] >> Okay. No, sorry. So, uh that's that's what we'll see and then we have also [26:32] invested in in DAX and it would it is typically difficult to uncouple a [26:37] technology to get another new technology in that that becomes challenging [26:41] >> stickiness of it all. Yeah. >> Yeah. [26:43] >> We also are DAX customer here and I think that the challenge the rock and a [26:47] hard place for Epic will be both that now you have DAX going to Dragon Copilot [26:51] which is just openly confrontational in the feature set. I think you know it's [26:54] in your EMR does things you would do in your EMR and you're talking about an [26:59] almost $4 trillion company powering that like that's significant and then just [27:03] the push for openness you look at what like fire is supporting fire R5 it just [27:09] liquidates the mode around data interoperability further too with like [27:13] federal leverage and so it that is a hard place to be in where you got all [27:17] the hyperscalers coming at the data you sit in between but legally you can't [27:23] really sit in between Yeah. And I I don't know how that will play out. [27:27] >> Yeah. Um I have so many more questions and not enough time. So I'm going to try [27:32] and prioritize in my head here quickly. One of the things I wanted to push back [27:35] on though, you all uh do work or have worked in worldclass institutions with a [27:40] lot of resources, right? And so and you know Adam your comment around the [27:45] infrastructure and what's the sort of scaffolding inside the organization in [27:49] the build versus buy uh debate um for the rest of us for you know medium to [27:56] large practices to less resource institutions [27:59] can they even sort of play in this space do they can they compete for the right [28:05] talent I mean even for the larger some of the institutions you're at do you [28:09] have the right talent in place in addition to the vision um and access to [28:14] that. >> Yeah, this could be for anybody [28:18] obviously. So let's hear from all of you. Yeah. [28:22] >> So we are lucky as I talked about I harbor we have a group of around 20 20 [28:26] 25 individuals, developers, uh very good data scientists, AI engineers working on [28:33] this. So that way we are lucky and that is not true for every healthcare [28:36] organization that cannot be replicated in medium or smalls size practices. But [28:40] what I think is the silver lining is that the cost of the technology is [28:45] decreasing so fast now that it's it's not it is not impossible that medium and [28:51] smalls size um practices can have their own in-house LLM secured within their [28:58] own box not going out somewhere with complete data security or probably do a [29:04] BA with Anthroporic or somebody else and create their own agents um that will [29:10] actually help them to take care of all of the road paperwork that the primary [29:14] care practitioners are doing and specifically about PCP. One of the [29:17] things that I'm the CMO of a hospital which is a community hospital which [29:21] means that lot of our patients for surgery they're referred to by PCP and [29:27] we then send them back this is pending that is pending it's a lot of paperwork [29:31] but that all we can probably deploy agents at the PCP's office that will [29:35] take care of all of this so when the patient comes to us they're ready for [29:38] our s they're ready for surgery so I think these kind of technology adop or [29:43] creating these kind of technology the the cost of that will decrease so That's [29:47] that's the silver lining. >> Yeah, I think that's maybe comforting to [29:52] smaller places that the barrier to entry to build things is definitely going [29:55] down. You you uh you know the idea of like you know even not so long ago a [30:01] team making a custom confinet for image analysis still is valid but it is [30:05] certainly not the only way you can deliver AI into practice. And I think [30:09] it's more builder friendly uh small teams be more efficient with like claw [30:13] code and so forth. It's more just about unencumbering them from the things that [30:18] kill every project which is I don't know who knows the workflow where the data [30:22] lives how I get from here and there. If those things can be clarified I think [30:26] even a team of three to five people can make tremendous progress if they can [30:29] iterate. But the challenge I've seen at our institution is it's always some [30:35] unknown that like hard stops iteration and then it doesn't matter if your [30:39] team's five people or 50 people nobody's going anywhere. [30:41] >> Good point. I'm pretty bullish on, you know, the the [30:46] ambulatory practices, whether they're small or big. I think whether it's the [30:50] PCPs or specialty practices, I think they'll find real value, you know, [30:55] mostly buying and configuring. Um I'm I'm a little less optimistic on the sort [31:01] of the smaller integrated health systems. Um, you know, I think if you've [31:05] got a one hospital system, even before AI, I mean, you were [31:10] struggling, right, with with cost, rising costs, labor shortage, cuts to [31:15] Medicaid. I I haven't followed, but I don't know what the the latest story on [31:18] Medicare today is, but you know, so you've got a whole host of things going [31:22] and your business is not simple, right? If you're PCP, I'm not saying a PCP is [31:25] simple, but compared to running a local community hospital, uh, it would seem [31:29] that way. And so I fear that this is just one more thing where they're going [31:32] to fall further behind what the big systems are doing. The one thing I would [31:36] say again whether it's with Mayo or other consortium there are consortium [31:40] that are solving this problem together and some of the larger academic centers [31:44] are providing support to these regional systems. So I think on their own it's [31:47] going to be very hard but they can figure out how to be part of that [31:50] network and consortium and leverage that and then they just have to know where [31:53] they're going to make those few bets and then let the the tide come with them. [31:59] Thanks, Adam. Um, I'm going to kind of get get to a final topic here in brief [32:06] and then and then be able to have at least a few minutes for questions, I [32:10] hope. Um, and I want to I want to ask the panelists to think a little bit [32:14] around future delivery models. In fact, for the past couple days, we've heard [32:18] around heard about companies designed to be AI first from the ground up. Even the [32:25] prior panel in this room or discussion talked about a aluminum and steel [32:29] cutting company that was built from the ground up is completely run by AI. We [32:34] don't have a health system today that is like that. We've got lots of forayas [32:38] from you know the likes of Amazon and and uh other big tech players that are [32:43] trying to offer healthcare offerings. Still though very human right there's [32:46] still people there. Um but as you think about future delivery models and what [32:52] some of the organizations you know so well the bricks and mortar organizations [32:55] can look like what and we at the I'll just say at the AMA we think a lot about [33:00] augmentation AI as augmentation of the clinician and uh with the humans still [33:05] not only in the loop but in charge right are you thinking differently now or [33:09] having different kinds of conversations especially in the LLM era on that and [33:14] Peter if you'd like to start. >> Absolutely. So the when we have our [33:18] discussions with leadership about what we think are existential risks to our [33:22] brick-andmortar business model, this is one that comes up and we we sit around [33:25] and we say, you know, Texas Medical Center in Houston has got, you know, the [33:28] biggest brickandmortar cathedral of healthcare real estate. And yet you look [33:31] at it and you're like, well, the existential threat to all of this is [33:34] asynchronous virtual care. Like the fact that whether it's a company or whether [33:38] it's just Chad GBT, they featured in their own launch that you know, we've [33:42] not made healthcare easy to use for pretty much anybody involved. We've had [33:45] lots of time to make it better. we've not and they're going to seek that [33:49] on-ramp to healthcare where they can get it. It will be increasingly virtual in [33:53] the pocket. It'll be Siri, it'll be chat, it'll be so forth. And as that [33:56] expands to become the on-ramp and the middle ground, maybe the off-ramp in [34:01] some cases, we were never even part of the the clinical relationship. And that [34:06] we've never had to deal with that before. And if we don't find a way to [34:09] dovetail into that meaning like you know take that chat conversation with GBT [34:14] bring it to the clinic let me quickly review it and write your script up make [34:17] your life more efficient somehow uh then we will just be marginalized out of a [34:21] massive sector of directive patient care. [34:25] >> Thank you. >> Yeah I I would agree it has to be our [34:27] healthcare has to be as easy as interacting with the web when we are [34:31] buying something. The other challenge to that is for simple things people will go [34:35] that's this route. I often use other things on the web to get simple health [34:40] care. The challenge to that is then hospitals or health system would be left [34:44] with caring for the sickest of the sick patient which is traditionally very [34:49] difficult to do and has challenges with revenue too. So [34:53] as you said it's existential crisis. If we are left only with the sickest of the [34:58] sick patient it will be difficult for us to manage. So we have to rapidly evolve [35:03] to make it so easy for all of our patients that they are they find it easy [35:06] enough as they're shopping in Amazon to interact with us to get an answer and [35:11] not wait for phone call back and things like that. So I think that's that that's [35:15] where we are headed. it this is not from any particular [35:21] health system point of view but I think um I think it was advisory board and I [35:25] think it was this year but it could have been the year before you know five [35:28] trends for the next decade and one of them was the shift from procedures to [35:31] therapeutics and I think I I would couple the digital transformation and AI [35:36] with this move you know if we think of our hospital systems it's it's mostly to [35:39] get someone in to do something to them right but if we don't need to see them [35:43] to treat and we don't need to see them to talk to them and now we don't need to [35:46] see them because they and take the medicine themselves or we can send a [35:50] home a nurse to do the infusion at home. What is the role of that physical [35:54] location to begin with? How do we think about that? And then how do you [35:57] reconstruct around that? And so I think AI will make this harder to get right, [36:01] but I don't think it's primarily an AI thing. I think it's a combination of [36:05] what care is like. Um the other thing I would say is we're still living in a [36:08] world where the vast majority of things is paid for for the amount of time that [36:12] was spent or actually a human being doing something. And so I think while [36:15] that exists, we'll still be in that business as health systems. I be [36:20] interesting to know if 3 years from now, you know, Medicare just decides to start [36:24] paying differently, if we could keep up or someone out competes us. We win [36:27] against Amazon and all these other state because it's a niche market and cash and [36:30] other things that can capture those, but most people still need 15 minutes of X- [36:35] provider to get the bill paid and you need a person at the end of that line. [36:40] >> Thanks. And I love that point around just a reminder for everyone that AI [36:44] huge transformation but we have all these other kind of secular trends [36:47] around therapeutics GLP1s right these things that are transforming how people [36:52] seek care and what they need in those settings in those bricks and mortar [36:55] settings. Um I'm going to sort of pause here now just so we can squeeze in I [37:01] think a couple questions. Uh does anyone have one? I see a co uh I'll in the back [37:06] there with the GL. That's you. Yeah. Yeah. I got you. [37:13] >> I don't mean to dis um mediate the problem of hallucinations and the AIs [37:19] themselves, but what about the data that's feeding the AI? And it seems to [37:23] me that one of the primary sources is the health the EMR. And I've talked to a [37:29] lot of doctors and they say there's a lot of crap in there. And it's like from [37:33] an a from a machine learning perspective, if you don't have good [37:36] data, you're in trouble to begin with. Anybody have any thoughts on how to [37:41] improve that? >> Not necessarily fix it. [37:43] >> Let me take a step. So we talked about the development team in IH Harbor, our [37:47] innovation arm. So I think when we are doing a project, our first three three [37:52] to six months for four or five people, they are just cleaning up the data [37:57] before it can be ingested in any kind of model because in AI now we have we trust [38:03] the model to tell us what's next. We we trust the model data to build the [38:08] algorithm. So that's a very important point and we spend a huge amount of time [38:12] just cleaning up the data so that our models are accurate. [38:17] >> Yeah. I think uh a a big way we can handle this is we have to get rid of [38:22] some of the double standarding we do for AI verse non AAI like just because AI [38:27] has non-zero hallucinations we always tend even in healthcare circles to talk [38:30] about how this is a huge problem even though it's your point healthcare has [38:34] been wildly variable in performance the entire time that's existed you know but [38:40] when it's AI it's got to be perfect and the reality is like we have to deliver [38:44] this value to patients as soon as we poss possibly can because they need it [38:48] and it won't be perfect. We're not perfect. We have to be clear that par is [38:52] the goal here, not perfection. The record is full of hallucinations and [38:56] inaccuracies prior to AI. And the thing that we can control with AI is not going [39:01] to be the, you know, perfection of the ground truth data. We can control the [39:05] token spend, the reasoning, the reflection, the bringing in of ancillary [39:09] data sources. reasoning our way through it should be the focus not necessarily [39:15] the point in time perfect accuracy because again we are far from that [39:21] let's squeeze in one did you have a question yeah front here yeah [39:29] >> so I come from a implementation background a project management [39:32] background um in our health care system I work for Adventist healthcare in in [39:36] Montgomery County Maryland and um one of the things that we find in implement [39:40] mentations is if you have a clinical champion beginning and throughout the [39:45] implementation process, it it goes off better. You talked about change [39:49] management, right? And uh what we see is that when you don't have that engagement [39:53] of your clinical champion or or your administrative leader uh champion uh [39:58] throughout the the process of that project and implementation, we get [40:03] feedback like you are doing this to us as clinicians. you and it are making us [40:09] do this. How do you now have the uh the clinical champions at the table with you [40:18] working with you to develop the use cases uh and then implement them uh so [40:23] that it's a matter of working with you and not doing something to you. [40:29] I the the only thing I can add because I think the clinicians can say it better [40:33] is um you know the systems I've seen do it and I I will point to Baylor where I [40:38] spent a lot of time have done a lot of work to align priorities. So you know we [40:43] have global KPIs the health system they roll down to and so the clinicians the [40:47] administrators it the innovate know that wait time is an issue canceled [40:52] appointment and so we are aligning around solving a problem that people are [40:56] connected to doesn't remove the fact that you still need a clinical champion [40:59] but you're not convincing them this is a problem we're solving you're saying this [41:02] is what we're doing doesn't make sense and so I think that starting from the [41:07] consensus on what is the the thing to solve helps make whatever solution you [41:11] get to easier Yeah, but just a second a clinical [41:14] champion is you need to have this kind of leader in these roles. And so I think [41:18] it's places should have people who serve in these positions. So whether it's a [41:22] chief a officer or a CMIO or a mixture of the two or whatever, someone who who [41:27] has to have a strategy, advocate for it, and specify like what we're leaning [41:32] into. Otherwise, it yeah, it's pretty much dead in the water. And I think if [41:35] people don't have those positions, they should create those positions because [41:38] they're they're really for that function. [41:41] One thing to add to that is that if there's an idea that comes some use [41:44] cases have to come from the clinicians and then we take it to the governance [41:48] council like for example we have the let's say it's a physiciancentric thing [41:51] we take it to the chief medical officer CMO council of the system say that this [41:56] is a great idea do you think we should try it violate it if the answer is yes [41:59] then we build it or then we look at that product otherwise we don't do it [42:06] >> well I'm going to take a moment to to thank our panelists I learned a lot this [42:10] was was a really uh rich conversation uh all your different perspectives um I [42:15] hope you all got a lot out of it also um uh thank you let's give a round of [42:20] applause to our panelists [Applause]