Mayo Clinic Deploys 128 AI Tools as Tech Capital Spend Hits $450 Million

Mayo Clinic部署128项AI工具，技术资本支出达4.5亿美元 该医疗系统通过将患者疗效置于短期投资回报之上，并将分级风险治理与无缝工作流整合相结合，在临床业务中拓展了128项活跃AI工具的应用规模。 Micky Tripathi详细阐述了Mayo Clinic将人工智能从理论研究转化为实际临床工作流的策略。他概述了从纸质EHR记录数字化，到在Mayo Clinic庞大的临床业务中部署预测性、生成式及智能体AI系统的技术演变。重点举措涵盖自动化传真处理、数字病理学、语音声学和预测性影像，所有这些均依托于分级风险治理模型的支持；该模型将患者疗效置于单纯的ROI之上，同时将100多项活跃的AI解决方案直接融入临床医生的工作流程中。 医疗技术正从基础数字化与预测模型加速演进为生成式 AI 与目标导向的 Agentic AI 工作流；同时，患者正利用法定病历访问权和消费级 AI 模型在就诊前自主分析检查结果。 (5:01) Mayo Clinic 领导层认为，医疗 AI 最大的风险并非模型幻觉，而是错失拯救生命技术所带来的机会成本；其落地战略核心在于通过消除行政负担“降低下限”，并通过尖端诊疗突破“拔高上限”。 (13:09) 在过去一年中，Mayo Clinic 投入 4.5 亿美元用于整体技术，其中 1 亿至 1.25 亿美元专用于 AI 与 Agent 举措；目前储备的 450 至 460 个 AI 方案中有 128 个已投入临床，且 70%–80% 由一线医护人员自主开发。 (13:09) 在落地治理上，Mayo Clinic 采用风险分级评估机制，强调严防“把算法误当产品”的误区；评估基准应参考现行医疗标准而非脱离实际的完美指标，确保节省的时间切实转化为患者获益。 (34:50)

Mayo Clinic is directing substantial capital and governance resources toward translating artificial intelligence from theoretical research into daily clinical practice. According to Chief AI Implementation Officer Micky Tripathi, the health system views the opportunity cost of failing to deploy life-saving tools as far greater than the risk of implementation errors, prioritizing direct patient impact over immediate financial returns. Evolution from Data Digitization to Agentic AI (5:01) Healthcare technology has evolved from basic digitization toward autonomous systems. Following the American Recovery and Reinvestment Act's $35 billion public investment alongside equal private capital, hospital electronic health record adoption expanded from roughly 10% in the early 2000s to between 95% and 98%. However, this initial wave largely left clinical information siloed in electronic storage. The sector is now moving along an evolutionary trajectory from predictive AI—which classifies multi-modal structured and unstructured data—to generative AI that synthesizes narrative notes, images, and audio, and ultimately to agentic AI that pursues clinical goals autonomously. In parallel, consumer access to models from providers like OpenAI, Anthropic, and Microsoft enables patients to analyze their own diagnostic records, while automated platforms like Doctronic in Utah handle tasks such as prescription renewals without human intervention. Enterprise Scale and Clinical Deployment Pipeline (13:09) Mayo Clinic's AI strategy is driven by its operational mandate to manage serious and complex illnesses, with half of its 1.3 million annual patients traveling from outside regions. The health system conducts 5 million outpatient visits and receives 5 million incoming faxes each year, requiring rapid data synthesis to identify critical clinical details. To modernize this infrastructure, Mayo invested $450 million in technology over the past year, targeting $100 million to $125 million specifically at AI and agentic software to normalize and vectorize disparate records. Rather than relying solely on external software vendors, between 70% and 80% of Mayo's deployed solutions are designed internally by frontline clinicians. The active pipeline contains 450 to 460 solutions, with 128 integrated into daily clinical care and more than 300 under evaluation. Routine administrative utilities, such as ambient clinical documentation, have already reached an organizational adoption rate of approximately 80%. Targeted Diagnostic and Workflow Implementations (13:09) Mayo Clinic deploys specialized AI tools across pathology, radiology, and administrative processing to shorten diagnostic timelines. The internal 'Record Time' application categorizes and summarizes millions of external faxes and network documents. For diagnostic workflows, an image encoder developed with Microsoft automates tube and line detection in portable chest X-rays, reducing radiologist review times by 30 to 45 seconds per image across 125 daily scans. Advanced models are also being applied to early disease detection and surgical pathology. Mayo's 'Atlas' digital pathology model uses 17 million digitized slides to perform virtual staining, reducing intraoperative frozen section evaluations from 20–25 minutes down to 5 minutes. Additionally, predictive models reviewing retrospective imaging have identified visually occult pancreatic cancer 18 to 36 months before human diagnosis, a development that could potentially raise 5-year survival rates from 3% to 39%. Risk-Tiered Governance and Operational Pitfalls (33:05) To prevent internal innovations from stalling, Mayo Clinic established an accelerated, risk-tiered governance process. Low-risk applications in non-critical settings bypass exhaustive front-end evaluations in favor of post-deployment monitoring and feedback loops. Conversely, high-risk, complex algorithms undergo rigorous evidentiary assessments before reaching patients. Tripathi emphasizes that an accurate algorithm does not equal a functional clinical product without a usable interface and workflow integration. Furthermore, implementation teams are instructed to benchmark tools against existing standards of care—such as an established 73% diagnostic baseline—rather than holding deployment hostage to arbitrary perfection targets like 92%. Because Tripathi reports directly to the CEO, the institution bypasses conventional line-of-sight return on investment, requiring instead that administrative time savings translate directly into tangible patient benefits. Mayo Clinic's AI deployment model demonstrates that bridging the gap between algorithmic research and front-line practice requires strong executive sponsorship, risk-tiered governance, and direct clinician ownership. A remaining operational test will be whether the institution can consistently convert time saved across hundreds of disparate AI tools into measurable, broad-scale improvements in patient survival and clinical care.

妙佑医疗国际（Mayo Clinic）首席人工智能实施官 Micky Tripathi 指出，人工智能正推动现代医学经历史上最深刻的变革，其实施核心已从理论研究转向以前线临床应用和患者结局为导向的全面落地。依托每年逾亿美元的专项投入与内部临床驱动模式，妙佑医疗正通过重构数据底座与分级治理，系统化推进AI工具在严重与复杂疾病诊疗中的实用转化。 医疗数据范式跃升与患者自主权演进 (5:01) 回顾医疗信息技术历程，美国在2000年代初期的电子病历（EHR）普及率仅约10%，绝大多数关键医疗数据被封闭在纸质档案中。随后《美国复苏与再投资法案》通过350亿美元公共资金及同等规模的民间资本投入，推动医院与非卧床医疗机构的认证EHR采纳率分别跃升至95%至98%和80%至85%。然而，单纯的数字化并未自发形成跨机构的深度洞察，医疗数据长期受困于新的数字孤岛。 Tripathi认为，当前医疗正沿着由预测性AI（分类复杂多模态数据）向生成式AI（整合非结构化叙述、音频、影像与病程记录），再向智能体AI（Agentic AI，自主执行既定目标任务）的技术曲线演进。与此同时，患者借助OpenAI、Anthropic及微软等基础模型，能够自主分析检查结果与影像，以更充分的准备参与就诊决策；而在犹他州等地，类似Doctronic的自主化智能体平台已在监管许可下实现处方续方自动化，标志着医疗服务正逐步具备全天候自动化交付的能力。 资本配置策略与双轨实施管线 (13:09) 妙佑医疗每年接诊约130万名独立患者并完成500万次门诊，其中50%的患者系长途跋涉寻求复杂重症治疗，伴随而来的还有每天数以百万计的外部病历与传真数据。首席执行官 Gianrico Farrugia 提出，医疗AI面临的最大风险并非系统幻觉或技术误差，而是未能拥抱救命技术所招致的“机会成本”。为此，妙佑医疗在过去一年总计投入4.5亿美元技术资金，其中1亿至1.25亿美元专用于AI与智能体软件开发，核心任务是将碎片化数据归一化、标记化与向量化以构建就绪的AI底座。 面对科技巨头数以百亿美元计的研发规模，妙佑医疗采取了内外结合的部署策略，一方面与头部技术机构战略合作，另一方面聚焦内部一线人员的自主发明。在其实施管线中，约70%至80%的方案由临床人员自主提出并牵头设计。整体战略分为“降低下限”（通过普及环境临床记录与收件箱自动化工具削减行政负荷，当前采纳率已达约80%）与“拉高上限”（打造尖端医学突破）。截至目前，其管线内储备了450至460个AI解决方案，其中128个已在临床日常实践中正式应用，另有300余个处于测试评估阶段。 多模态精准医疗的临床纵深实践 (1:00) 为了从零散的点状应用迈向构建患者“数字孪生”的精准医学，妙佑医疗在多项临床场景中部署了具体工具。自主研发的“Record Time”系统专门用于对每年数百万份外部传真和每月数千万条网络数据进行自动归类与提炼；护理团队主导的虚拟助手在交接班时捕捉隐性与显性知识以消除信息差；用于基因咨询的交互式虚拟化身则化解了长达9个月的人工咨询排队瓶颈。 在核心诊疗与早期筛查领域，多模态工具展现出显著临床效益：妙佑与微软合作开发的移动胸片图像编码方案，使放射科医生在日均125张影像的工作中每张节省30至45秒的管线识别时间；基于1700万张数字病理切片建立的“Atlas”大模型实现了虚拟H&E染色，将术中冷冻切片评估时间从20至25分钟大幅压缩至5分钟；针对隐匿性胰腺癌的早期识别模型可在临床确诊前18至36个月发现影像异常，有望将5年生存率从3%提升至39%；此外，心电图算法派生出主动脉瓣狭窄等9项生理指标，声纹特征分析支持心血管与认知衰退筛查，智能病房的环境传感器亦逐步替代了传统的跌倒报警床垫。 分级治理机制与避免临床落地的认知误区 (33:05) 为防止基层借助低门槛开发工具产生的创新沦为“科研玩具”，妙佑医疗建立了敏捷治理与风险分级评估框架。对于低复杂度、低风险的应用，免除繁复的前置实证审查，依托部署后的持续监测与反馈回路加速临床上线；而对于算法复杂、具有潜在非预期后果的高风险系统，则施加严格的前置验证机制，确保患者安全底线。 在具体推进中，Tripathi直接向首席执行官汇报，确保管理层以“改善患者结局”而非“短期可见的投资回报率（ROI）”作为最高考量标准。妙佑医疗总结出三大关键落地误区：第一，切忌将算法等同于产品，缺乏直观用户界面与无缝临床工作流集成的算法无法进入实用；第二，确立临床评估基准时切忌脱离实际现状盲目设定完美指标，若现有医疗标准准确率仅为73%，一款能带来切实改善的工具便不应因未达92%的理想阈值而被扣留；第三，效率节约（如节约医生时间）必须实质性转化为患者能够直接感知的照护获益，方具备立项价值。 妙佑医疗的经验表明，医疗AI从研究迈向实效的核心壁垒不在于算法精度本身，而在于工作流工程、基准对照与组织治理的紧密嵌合。然而，该模式高度依赖顶级医疗机构巨额的非营利资本缓冲与自研临床人才密度，对于缺乏同等资金与技术沉淀的中小型医疗体系而言，如何低成本复制此类深度集成的智能体管线仍是有待解答的行业挑战。

## Background 本内容由 Ai4 于 2026 年 8 月 20 日在 YouTube 平台发布，记录了妙佑医疗国际（Mayo Clinic）首席人工智能实施官米奇·特里帕蒂（Micky Tripathi）的专题演讲。根据页面简介，演讲探讨了医疗机构如何在确保安全性、医患信任与人文关怀的前提下，加速临床 AI 从科研实验室向日常诊疗实践的转化落地，最终改善患者预后。 ## TL;DR 特里帕蒂回顾了医疗 IT 二十年的数字化演进与孤岛困境，指出医疗 AI 正经历从预测型、生成式向智能体 AI 的范式跃迁。妙佑医疗国际秉持百年来“联合力量”与“以患者利益为中心”的信条，设立 4.5 亿美元研发基金并孵化了 450 余项自研管线，推行“托底提效”与“突破上限”并重的双轮战略。团队展示了病历极速检索、Atlas 数字病理大模型、胰腺癌超早筛、声纹生物标志物以及环境感知病房等多项前沿落地实践，并正向全人“数字孪生”迈进。最后，演讲分享了将治理定位从被动合规重塑为战略赋能的经验，强调算法必须转化为深度融入工作流的落地产品，切实让患者获益。 ## Atmosphere & style 整场演讲节奏紧凑且充满激情，特里帕蒂以“系好安全带”的直率开场，展现出技术变革浪潮中的紧迫感与雄心壮志。叙事风格兼具顶层战略的宏大视野与临床一线的扎实细节，穿插着母亲住院时利用大模型沟通与反思传统防跌报警等温情而生动的个人亲历。现场氛围务实而发人深省，既有对技术突破的由衷赞叹，也有对“算法不等于产品”的严厉敲打，让听众深刻感受到一家百年顶尖医疗机构在拥抱前沿科技时的敬畏与笃定。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=-JtNchTupWA&t=4s) 妙佑医疗国际（Mayo Clinic）将当前视作医学史上最具变革性的时刻，其 AI 实施的核心使命是将前沿研发成果从实验室“科学实验”切实转化为一线临床生产力。 - [01:34](https://www.youtube.com/watch?v=-JtNchTupWA&t=94s) 过去二十年全美通过数十亿美元投资实现了电子健康档案（EHR）的普及，但医疗数据依然深陷于多模态数字孤岛，难以自动化提炼深层临床见解。 - [04:51](https://www.youtube.com/watch?v=-JtNchTupWA&t=291s) 医疗 AI 正在经历三级能力跃迁：从分类整理数据的预测型 AI、提炼见解的生成式 AI，演进到具备目标驱动与自主执行能力的智能体 AI（Agentic AI）。 - [05:43](https://www.youtube.com/watch?v=-JtNchTupWA&t=343s) 大模型普及重塑了医患关系，医疗机构应当积极拥抱患者赋权，指导患者科学使用 AI，并探索 7×24 小时可信的自动化医疗前沿服务。 - [09:59](https://www.youtube.com/watch?v=-JtNchTupWA&t=599s) 妙佑传承百年的“联合力量”（Union of Forces）哲学在数字化时代焕发新生，正演进为多学科临床经验、海量多模态数据与先进 AI 算法的深度融合协同。 - [11:15](https://www.youtube.com/watch?v=-JtNchTupWA&t=675s) 面对疑难重症带来的海量病历处理挑战，医疗机构最大的风险并非技术偶发缺陷，而是因迟疑不决不敢采用 AI 进而剥夺患者康复机会的“机会成本风险”。 - [13:46](https://www.youtube.com/watch?v=-JtNchTupWA&t=826s) 妙佑专项投入过亿美元布局 AI，在储备的 450 余个 AI 管线方案中，有 75% 至 80% 由一线医护人员基于真实临床痛点自研孵化。 - [16:09](https://www.youtube.com/watch?v=-JtNchTupWA&t=969s) 临床 AI 落地遵循“托底提效”（消除行政冗余）与“突破上限”（超越人类感知极限推动医学科学）的双轮驱动战略。 - [20:00](https://www.youtube.com/watch?v=-JtNchTupWA&t=1200s) 基于 1700 万张玻璃切片数字化的 Atlas 病理基础大模型结合虚拟染色技术，将手术中切缘冰冻切片的等待耗时从 20-25 分钟压缩至约 5 分钟。 - [21:48](https://www.youtube.com/watch?v=-JtNchTupWA&t=1308s) 通过对经典临床数据（如历史影像、常规心电图）的算法回溯挖掘，可在症状显现前 18 至 36 个月实现胰腺癌等凶险疾病的早期筛查与预警。 - [25:29](https://www.youtube.com/watch?v=-JtNchTupWA&t=1529s) 通过探索声纹生物标志物以及投资 90 亿美元构建环境被动感知病房，医疗系统正在实现非接触式无感健康监测，兼顾临床体征捕捉与患者尊严。 - [29:09](https://www.youtube.com/watch?v=-JtNchTupWA&t=1749s) 打破各专科独立的“单点解决方案”信息烟囱，将跨模态推理结果综合提炼为全局临床洞见，是一体化医疗 AI 发展的核心瓶颈与方向。 - [29:44](https://www.youtube.com/watch?v=-JtNchTupWA&t=1784s) 妙佑正在构建全人维度的系统级器官模型与“数字孪生”（Digital Twin），旨在结合相似病例画像进行多情景预后推演，实现真正的精准医疗。 - [32:11](https://www.youtube.com/watch?v=-JtNchTupWA&t=1931s) AI 治理体系需从被动“合规约束”转向“战略赋能者”，通过分级风险治理（Risk Tiering）确保前端探索自由与后端临床转化的高效严谨。 - [34:52](https://www.youtube.com/watch?v=-JtNchTupWA&t=2092s) 医疗 AI 落地必须规避三大误区：明确算法不等于临床产品、避免脱离现有标准疗法空谈指标，且所有效率提升必须最终切实转化为患者获益。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=-JtNchTupWA&t=4s) 医疗史上最具变革性的时刻与 AI 落地实施的使命 妙佑医疗国际（Mayo Clinic）首席人工智能实施官（Chief AI Implementation Officer）米奇·特里帕蒂（Micky Tripathi）在开场中难掩兴奋之情。他直言自己面临的最大挑战是如何将满满的激情浓缩在短短 20 分钟内，并呼吁观众“系好安全带”，因为接下来将涵盖极其密集的信息。在妙佑医疗国际看来，当前正处于现代医学史上最具变革性的关键时刻；正如其首席执行官吉安里科·法鲁贾（Gianrico Farrugia）所言，这甚至是“整个医学史上最具变革性的时刻”。面对如此巨大的历史机遇，妙佑医疗国际正全力以赴展开战略布局。 特里帕蒂强调，业内已经充斥了太多关于抽象理论框架的演讲，但他更关注实际落地的具体解决方案。作为首席人工智能实施官，他的核心职责在于“实施”本身——如何打破实验室的局限，避免让 AI 仅仅停留在“科学实验”阶段，而是切实交付到一线临床医生手中，以最终改善患者护理。整体分享将围绕真实落地案例与赋能型治理体系展开，展现技术转化为临床生产力的全貌。 - **核心使命**：将前沿 AI 研发成果转化为一线临床医生的实用工具，打破“科学实验”与“临床落地”之间的鸿沟。 - **管理层愿景**：妙佑医疗国际 CEO 吉安里科·法鲁贾明确将 AI 浪潮定性为医学发展史上最重大的变革契机。 --- ### [01:34](https://www.youtube.com/watch?v=-JtNchTupWA&t=94s) 从纸质病历到数字孤岛：医疗 IT 二十年的演进与困境 特里帕蒂回顾了自己投身医疗信息技术（Health IT）二十余年的从业历程。早在波士顿咨询集团（BCG）担任项目经理时，他便参与了印第安纳波利斯的一项医疗信息化项目，试图将各医院现有的电子健康档案（EHR）系统连接起来，搭建能够从临床数据中持续学习的信息共享平台。然而，20 年前全美电子健康档案的普及率仅有 10% 左右。整个医疗系统深陷于纸质病历的泥潭之中，数据被牢牢锁在物理文件里，管理者和医生甚至连基础的医疗测量与系统性评估都无法做到。 转折点发生在 2000 年代初的十年后，美国通过《美国复苏与再投资法案》（ARRA / HITECH），投入了 350 亿美元的公共资金以及相当规模的民间资本，推动全美医疗体系告别纸质时代。如今，全美已有 95% 至 98% 的医院以及 80% 至 85% 的门诊机构全面应用了认证电子健康档案系统。然而，数字化并没有自动带来预期的洞察力跃升。数据虽然被数字化，但依然分散在各种非结构化文本、结构化字段、医学影像和音频等多模态孤岛中，仅仅满足了最初的“病历记录”功能，难以支撑跨维度的证据生成和深层临床洞察。 ```mermaid flowchart TD accTitle: 医疗信息化二十年演进路径 subgraph S1["第一阶段：纸质时代（约20年前）"] A["纸质病历文件"] --> B["普及率仅约 10%"] B --> C["数据被困纸端 / 无法系统化评估"] end subgraph S2["第二阶段：数字化跨越（ARRA法案驱动）"] D["$350亿公共与私人投资"] --> E["医院普及率 95-98% / 门诊 80-85%"] E --> F["完成数据要素数字化"] end subgraph S3["当前挑战：数字孤岛"] F --> G["多模态碎片化数据（文本/影像/音频）"] G --> H["困于基础记录 / 缺乏深层临床洞察"] end ``` - **数字化普及成就**： - 全美医院认证 EHR 普及率达 **95% - 98%**。 - 全美门诊医疗服务机构 EHR 普及率达 **80% - 85%**。 - 公共与私人领域累计投资数百亿美元完成底层基础设施改造。 - **遗留核心瓶颈**：海量数据处于“数字孤岛”状态，难以跨模态提炼知识，如同在数字干草堆里找一根针。 --- ### [04:51](https://www.youtube.com/watch?v=-JtNchTupWA&t=291s) 从预测、生成到智能体：AI 技术范式的跃迁 为了破除数据沉睡在孤岛中的僵局，医疗领域正在经历从单纯数据记录向高阶 AI 赋能的演进。特里帕蒂将这一演进路径清晰划分为三个层级：首先是**预测型 AI（Predictive AI）**，负责对海量数据进行分类、结构化梳理与初步解析；其次是**生成式 AI（Generative AI）**，基于多模态数据提炼临床知识并生成深度见解；最终跃升至**智能体 AI（Agentic AI）**。 智能体 AI 的出现被视作革命性的转折点。它不再需要工程师预先编写并细化每一步操作路径，而是能够根据给定的最终临床或运营目标，自主拆解任务、调用工具并执行复杂行动。对于拥有庞大历史沉淀与海量临床数据的妙佑医疗国际而言，这种技术演进为真正挖掘多模态数据价值、实现全流程自动化提供了前所未有的强大引擎。 ```mermaid flowchart LR accTitle: 医疗 AI 三级能力跃迁 A["预测型 AI<br/>(Predictive AI)"] -->|"分类、解析与整理数据"| B["生成式 AI<br/>(Generative AI)"] B -->|"跨模态见解提取与知识生成"| C["智能体 AI<br/>(Agentic AI)"] C -->|"目标驱动、自主规划与行动闭环"| D["临床与运营全流程自动化"] ``` --- ### [05:43](https://www.youtube.com/watch?v=-JtNchTupWA&t=343s) 患者赋权与自主医疗前沿：重新定义医患关系 在分析宏观技术趋势时，特里帕蒂强调必须“滑向冰球未来的位置”。AI 技术正以肉眼可见的速度每天迭代进化，尽管业内普遍关注幻觉（Hallucination）等潜在风险并保持审慎，但其演进速度与潜力毋庸置疑。更深远的变革发生在患者端：OpenAI、Anthropic、微软等底层大模型工具直接面向消费者开放，借助联邦法律赋予患者调取自身医疗记录的权利，患者可以在踏入医院前将 X 光片、化验单上传至手机上的大模型，提前获取病因推测与可能的多条治疗路径。 特里帕蒂分享了切身体会：其母亲近期住院时，他每天都会把母亲的各项检查结果导入大模型进行深入分析，以便在医生查房时展开更高质量、更有针对性的沟通。妙佑医疗国际坚决主张**不应阻碍或打击患者使用 AI 的积极性**，而应主动拥抱这种赋权趋势，指导患者在哪些场景下可以信任 AI、在哪些环节需要保持警惕，并与医生形成新型合作模式。 与此同时，医疗自动化也突破了传统监管认知。例如在犹他州，一款名为 Doctronic 的智能体应用依托州政府政策支持，在无需人类介入的前提下实现了处方自动续方（Prescription Renewals），直接切入传统监管边界地带。面对技术与实践走在监管前面的现实，妙佑医疗国际正积极探索将哪些医疗环节转化为受信任的 7×24 小时全天候自动化服务，以妙佑的权威临床背书为患者提供兼具高便捷性与高安全性的医疗体验。 - **个人亲历故事**：特里帕蒂在母亲住院期间每日使用大模型分析化验与影像数据，以便与主治医生开展高质量对话。 - **患者赋权立场**：鼓励患者利用大模型理解自身病情，医护人员的角色转变为提供场景化指导与辨真伪的专业伙伴。 - **处方自动化先例**：犹他州已允许 Doctronic 智能体在无人工参与的情况下执行处方自动续方操作，展现出智能体自动化医疗的前沿探索。 --- ### [09:07](https://www.youtube.com/watch?v=-JtNchTupWA&t=547s) 传承“联合力量”哲学：妙佑医疗国际的创立初心与 AI 新融合 特里帕蒂回顾了妙佑医疗国际百年历史底蕴。妙佑医疗国际由妙佑老医生及其两位儿子（即著名的妙佑兄弟）共同创立，其百年来恪守的核心信条是“患者的最佳利益是妙佑医疗国际唯一需要考虑的利益”（The best interest of the patient is the only interest to be considered）。在 1910 年代，妙佑兄弟提出了“联合力量”（Union of Forces）的理念。 在当时的时代背景下，“联合力量”的内涵是将全科医学与各个专科医生聚集在同一个屋檐下开展多学科协作诊疗（Multispecialty Practice）。而面对当今 AI 与数字化浪潮，妙佑医疗国际正赋予这句百年箴言崭新的时代意义——将先进的数据技术、AI 算法与多学科专家的临床经验深度融合，形成新时代的“力量联合”，以持续兑现以患者为中心的根本承诺。 ### [09:59](https://www.youtube.com/watch?v=-JtNchTupWA&t=599s) 从气动传输管到智能体：妙佑“力量联合”的百年演进 在妙佑医疗国际（Mayo Clinic）的历史传统中，“力量联合”（Union of Forces）始终是其立院基石。这一理念最早体现在制度设计上：所有医生实行固定薪酬制，彻底剥离个人创收激励，使医生的唯一出发点与机构及患者的最佳利益保持一致。在临床实践上，力量联合则具象化为“整合式病历系统”。在明尼苏达州罗彻斯特院区，至今仍保留着交织穿梭于各个大楼之间的气动传输管道系统。早在二十世纪初，亨利·普拉默医生（Dr. Henry Plummer）便开创了这一物理整合病历的先河——心内科、肿瘤科和放射科医生在各自诊室完成手写记录后塞入气动管道，统一汇总至中央病历室，如今妙佑基于 Epic 的电子病历系统依然被尊称为“普拉默病历”（Plummer Chart）。 随着技术的演进，妙佑亚利桑那院区首席执行官里克·格雷医生（Dr. Rick Gray）对“力量联合”赋予了全新的时代定义：下一阶段的协同核心在于全面引入人工智能与智能体（Agentic）解决方案。对于妙佑而言，这项技术的战略意义尤为深远，因为妙佑在医疗生态中专注于疑难重症诊疗（Serious and Complex Care），来到这里的患者通常已经历漫长而曲折的就医旅程，携带了来自全美各地的海量历史诊疗资料。 ### [11:15](https://www.youtube.com/watch?v=-JtNchTupWA&t=675s) 疑难重症的“大海捞针”与 AI 最大的机会成本风险 疑难重症患者往往辗转多家医疗机构无果后才来到妙佑，全院超过 50% 的患者来自其他地区。面对每年 130 万独立患者、500 万次门诊量以及每天多达 500 万份传真病历的庞大体量，医生面临着极其严峻的“大海捞针”挑战——必须以极高效率快速解析这些外部病历，捕捉关键临床线索，为绝望中的病患提供前沿诊疗方案并打破医学天花板。 在拥抱技术创新的过程中，妙佑医疗国际总裁兼首席执行官吉安里科·法鲁吉亚（Dr. Gianrico Farrugia）提出了一个振聋发聩的核心命题：面对人工智能，医疗界首要担心的风险不应是“系统表现是否不及预期”，而是“因犹豫不决而未能全力拥抱 AI，进而无法将医疗提升到全新高度”。 - 医疗机构固然需要防范 AI 滥用带来的安全风险，但更必须警惕因不敢采用 AI 而产生的“机会成本风险”——即剥夺患者本可通过前沿技术获得更好康复结果的机会。 - 妙佑坚守自身的独特使命，致力于通过深入挖掘海量医疗信息中的高阶推理洞见，持续推高顶尖医学科学的诊疗上限。 ### [13:46](https://www.youtube.com/watch?v=-JtNchTupWA&t=826s) 4.5 亿美元科技投入与 75% 自研管线的创新生态 为了将前沿技术切实落地为临床生产力，妙佑在过去一年中投入了约 4.5 亿美元的技术研发资金，其中 1 亿至 1.25 亿美元直接专项用于人工智能与智能体解决方案。面对长期存在的数据孤岛问题，工程团队正系统性地推进底层数据的标准化、分词化（Tokenization）与向量化（Vectorization），使全院临床数据达到原生适配 AI 和智能体的标准。作为非营利性医疗机构，妙佑明确自身的定位并非与科技巨头竞争算力，而是与最顶尖的技术企业建立深度战略合作，共同探索前沿技术的成熟度，并在第一时间将其引入临床环境。 这种机制孕育出了极其充沛的自研创新生态。目前妙佑的 AI 研发管线中储备了 450 至 460 个解决方案，其中已有 128 个全面嵌入日常临床工作流程，另有 300 多个正处于研发各阶段。尤为独特的是，这些方案中有 75% 至 80% 均来自妙佑内部一线医护人员的自研孵化，形成了由临床痛点驱动创新、再由机构赋能推广至全院患者的闭环。 ```mermaid flowchart TD accTitle: 妙佑医疗国际临床 AI 体系架构 subgraph Pipeline["AI 研发管线与落地规模"] P1["450+ 储备管线项目"] --> P2["300+ 处于各研发阶段"] P1 --> P3["128 项已在临床常态化运行"] P4["75%-80% 内部医护一线自研"] -.-> P1 end subgraph Strategy["临床 AI 双轮驱动战略"] T1["托底提效 (Lowering the Floor)<br/>消除行政冗余 / 提升流转效率"] T2["突破上限 (Raising the Ceiling)<br/>超越人类感知极限 / 推动医学科学"] end Pipeline --> Strategy ``` ### [16:09](https://www.youtube.com/watch?v=-JtNchTupWA&t=969s) 托底提效与突破上限：临床 AI 的两大维度与落地实践 妙佑将临床 AI 解决方案划分为两大核心维度：一是“托底提效”（Lowering the Floor），旨在消除人为行政低效并优化就医流程；二是“突破上限”（Raising the Ceiling），通过 AI 赋能医护人员完成仅凭人类感官无法实现的医学突破。在托底提效层面，环境智能临床记录（Ambient Documentation）在妙佑的采纳率已高达约 80%，近乎成为基础设施；而在突破上限与流程重塑方面，妙佑正密集落地多项代表性工具： - **病历极速检索系统（Record Time）**：针对每年 500 万份外部传真及全国互联网络每月调用的数千万份病历，该系统能自动化完成海量异构文本的扫描、分类、编目与智能摘要生成，使接诊医生在患者开启为期三天的妙佑初诊行程前，便能全面掌握其过往病史全貌。 - **护士虚拟助手（Nurse Virtual Assistant）**：针对临床交接班这一信息传递黑洞，由一线护理团队主导开发的虚拟助手利用 AI 帮助护士提炼、沉淀并无缝交接显性与隐性护理知识，极大减轻了手工交接的认知负担。 - **遗传咨询虚拟化身（Avatar）**：针对全美遗传咨询专业人员极度短缺的现状，妙佑推出了面向患者的交互式数字分身。患者无需排队 9 个月等待真人首诊，可在周日夜晚等个人空闲时间直接与数字分身完成初诊评估，大幅压缩就医等待周期并精准分流后续诊疗。 - **胸部 X 光图像编码器（Image Encoder）**：与微软深度合作开发，专门用于应对床旁移动 X 光摄片中因患者身上交错缠绕的各类导管、监测传感器线路等视觉干扰，帮助放射科医生在极其复杂的影像背景中快速精准阅片。 ### [20:00](https://www.youtube.com/watch?v=-JtNchTupWA&t=1200s) 放射影像与数字病理的自动化突破：从管线识别到虚拟染色 在临床一线，放射科医生每天面临着巨大的工作负荷。以胸部 X 光片（Chest X-ray）为例，医生每天必须逐张审阅，费时费力地去辨认每一根管道或导管——哪条是插管，哪条是静脉通路。妙佑医疗国际（Mayo Clinic）研发了一套自动化解决方案，能够自动标定图像中的各种置管与管线，让医生能够直接聚焦于底层解剖结构的病变分析。这项改进看似细微，却能为每张 X 光片节省 30 至 45 秒。妙佑医疗的放射科医生平均每天要阅片 125 张，累积起来的效率提升极为可观。该技术目前正进一步拓展至更多影像模态，并用于自动生成结构化报告。 与此同时，妙佑医疗在数字病理（Digital Pathology）领域推进了一项宏大工程。团队将 1700 万张实体玻璃切片全部扫描转化为高分辨率数字病理图像，并基于这些海量数据训练并发表了名为 Atlas 的病理基础大模型。该模型能够直接对病理图像进行深度推理，正在手术室环境中掀起一场效率革命。在传统的肿瘤切除术中，外科医生为了确认切缘阴性（Clean margins），必须在手术中途切取组织样本进行术中冰冻切片（Frozen section）。病理医生需要将样本带至专门房间，经历冷冻、实体切片、化学染色等繁复工序后才能在显微镜下阅片，这一过程通常需要耗费 20 到 25 分钟。 通过 Atlas 模型结合虚拟苏木精-伊红染色（Virtual H&E staining），整个实体染色过程被算法数字染色所取代。原本近半小时的漫长等待被直接压缩至短短 5 分钟左右。在外科手术中，这 20 分钟的缩短具有非凡的资源配置意义：主刀医生、麻醉师、手术室护士以及昂贵的手术室资源都无需无谓空转，而承受麻醉与手术创伤的患者更是直接受益者。 - **胸部 X 光读片减负**：自动识别并勾勒体内置管和管线，每张片子为放射科医生节省 30 至 45 秒，单个医生日均处理 125 张。 - **1700 万张玻璃切片数字化**：构建病理基础大模型 Atlas，实现高精度病理特征推理。 - **术中冰冻切片时间压缩**：利用虚拟 H&E 数字染色技术，将传统的 20-25 分钟等待时间压缩至约 5 分钟。 ```mermaid flowchart TD accTitle: 术中病理切缘分析流程对比 accDescr: 对比传统术中冰冻切片与基于虚拟H&E的AI病理工作流耗时差异 subgraph Traditional["传统术中冰冻切片流程 (20-25 分钟)"] A["切取病理组织"] --> B["送往病理室并物理冷冻"] B --> C["机械切片与化学染色 (H&E)"] C --> D["病理医生镜下阅片判读切缘"] end subgraph AI_Pathology["Atlas 虚拟数字化流程 (约 5 分钟)"] E["切取样本数字化扫描"] --> F["Atlas 基础模型虚拟 H&E 染色"] F --> G["快速切缘推理与即时反馈"] end ``` --- ### [21:48](https://www.youtube.com/watch?v=-JtNchTupWA&t=1308s) 经典模态的算法复用：胰腺癌早筛、心电图与基因多组学 将已有的临床数据与已知的随访终点进行“回溯匹配”，是推动医学前沿突破的核心研发范式。以早期胰腺癌为例，传统临床诊断极其困难，发现时往往已是晚期。妙佑医疗调取了确诊胰腺癌患者以往所有的历史影像切片，将当年放射科医生认为“未见异常”的历史底片输入 AI。经过特定的图像去噪（Denoising）与视觉隐匿性胰腺癌（Visually occult pancreatic cancer）特征提取算法，模型成功识别出了人眼根本无法捕捉的早期微弱癌变信号。这项发表的研究成果曾获 NBC News 深度报道，目前正在开展更广泛人群的临床试验。 胰腺癌与脑胶质瘤素来被称为“所有人的噩梦”，常规 5 年生存率极低（约 3% 左右）。而通过 AI 早期筛查工具，有望在患者症状显现前的 18 至 36 个月提前预警。若能在极早期介入，胰腺癌患者的 5 年生存率可实现几何级跨越，从惨淡的 3% 提升至 39% 左右，并在约 50% 的高危受试人群中展现出了提前筛查的潜力。这种超越人类肉眼感官极限的能力，正是医疗 AI 最具颠覆性的价值所在。 这种将经典数据重构发掘的思路，被生动地形容为“洗发水配方式的循环复用”（Lather, rinse, repeat），并在多个医学领域全线铺开： 1. **常规心电图（ECG）挖掘**：一张简易的心电图被 AI 拓展出 9 个不同的量化分析仪表盘，不仅能检测主动脉瓣狭窄等传统肉眼不易察觉的器质性心脏病，甚至能计算患者的“生理年龄”（Physiological age），以对比其与生物学实际年龄的健康差距。 2. **脑部影像与痴呆亚型**：通过头颅影像算法实现各类阿尔茨海默病及痴呆亚型的超早期识别。尽管目前尚无根治药物，但早诊断能让神经内科团队为患者尽早量身定制生活方式干预和照护规划，极大提升生存质量。 3. **基因组学与药物反应预测**：将患者基因分型与临床用药反应做关联学习。例如甲氨蝶呤（Methotrexate）是类风湿关节炎的临床金标准用药，但约 70% 的患者其实对其反应不佳，且伴有严重的毒副作用。AI 旨在提前预测患者是否属于那 30% 的有效响应人群，避免患者遭受无效治疗与药物损伤。 --- ### [25:29](https://www.youtube.com/watch?v=-JtNchTupWA&t=1529s) 突破既有维度：声纹生物标志物与 90 亿美元环境感知病房 除了在既有影像和生理信号上不断迭代，妙佑医疗正在探索前所未有的全新感知模态，其中最具代表性的是声带声纹（Voice Biomarkers）。人体解剖结构中，有一根迷走神经的分支紧贴心脏上方走行，随后回折穿过声带旁。当患者出现主动脉瓣狭窄、心室肥大或心瓣膜关闭不全等心脏病变时，肿大的心血管组织会压迫该神经，进而引发微弱的声带振动改变。这种变化人耳根本无法辨识，但 AI 却能通过精密的声学分析捕捉到心血管病变信号。如今在妙佑医疗的心内科，声纹采样已与验血、组织活检一样成为常规项目——患者只需在手机 App 上朗读一段文本，系统便能建立基线并持续动态监测。神经科团队也开始利用语音语调、语速停顿及细微的语义结构改变，来敏锐侦测认知功能减退与潜在的精神疾病发作。 在更宏观的物理空间层面上，妙佑医疗正斥资 90 亿美元进行医院基础设施的全面升级。无论是明尼苏达州罗切斯特（Rochester）总部还是佛罗里达州杰克逊维尔（Jacksonville）院区，天际线上如今林立着密密麻麻的建筑吊车。所有新建病房均全面配备了“环境被动感知”（Ambient detection）系统，彻底颠覆了传统的住院安全管理模式。 发言人动情地分享了自己母亲住院的切身经历：母亲曾因跌倒风险被置于传统防跌监测垫上，只要她稍微起身，床边警报器就会发出刺耳的尖叫，“把她吓得魂飞魄散”。母亲曾无奈地抱怨自己“仿佛置身于监狱，只要稍动一下就警铃大作、医护蜂拥而至”。而新一代环境智能感知能够以非接触、无感的方式实时看护患者。更重要的是，这套系统就像自动驾驶汽车在行驶中不断学习人类驾驶习惯一样，能够持续记录患者的步态（Gait）、起卧动力学与姿态变化，提炼出前所未有的临床体征数据，真正践行了“让每一位走进医院的患者，都使医疗系统变得更聪明”的愿景。 - **声纹心血管筛查**：利用喉返神经与心脏的空间解剖关系，通过算法识别心瓣膜与心室病变导致的细微发音改变。 - **90 亿美元重塑院区**：在罗切斯特和佛罗里达广泛部署非侵入式环境智能感知病房。 - **动态步态与跌倒预警**：替代机械式防跌噪音警报，以无感视频和空间遥测技术持续积累运动步态数据，大幅改善患者住院尊严与心理体验。 --- ### [29:09](https://www.youtube.com/watch?v=-JtNchTupWA&t=1749s) 从孤立单点工具到全模态综合决策的跨越 在回顾了上述所有令人惊叹的技术突破后，发言人话锋一转，指出了当前医疗人工智能生态中的核心短板：目前所展示的一切成就，本质上都依然是孤立的“单点解决方案”（Point solutions）。 无论是针对心电图的特定心血管算法、针对 X 光或 CT 的早期胰腺癌筛查模块、针对手术切缘的病理分析工具，还是针对语音信号的认知与心脏评估模型，各自都在特定领域表现卓越，向临床医生分别提供着割裂的患者局部洞察：“心内科工具评估了心脏指标，神经科模型推断了认知状态，放射科模型给出了肺部读数。”然而，真实的患者是一个不可分割的有机整体。妙佑医疗接下来的核心战略目标，正是要打破这些各自为战的信息烟囱，将跨多模态的单点推理结果综合提炼（Synthesize），打造能够统筹全局临床信息、赋能全流程决策的真正一体化医疗 AI 体系。 ### [29:44](https://www.youtube.com/watch?v=-JtNchTupWA&t=1784s) 从单点解决方案迈向“数字孪生”与全人建模 梅奥医学中心（Mayo Clinic）并不满足于仅仅关注患者身体的某个割裂局部，而是致力于将患者视为一个完整的生命整体。演讲者指出，虽然团队目前开发了许多针对具体环节的单点方案，但真正的价值在于双向延展：向后追溯，可以更深层次地理解底层生物学机制，把握科学本质；向前推进，则能够构建器官系统模型——例如循环系统模型、胰腺系统模型以及内分泌系统模型。这些系统级模型让临床医生能够以更加复杂而全面的视角，推演不同类型诊断下的多维情景。 最终的宏伟愿景是构建患者的“数字孪生”（Digital Twin）。通过全方位汲取患者数据，系统可以根据个体独特的生物特征创建个性化的数字孪生体。医生借此能够与患者共同探讨各种潜在的诊断可能性，更关键的是评估不同场景下的预后情况并优化治疗方案。借助庞大数据库中与当前患者高度相似的既往病例特征与诊疗经历，系统能推演并呈现出三到四种具体的可能情景。尽管这一构想目前尚未投入实际临床生产环境，仍处于模块构建阶段，但团队已经着手设计临床操作看板，并探索智能体架构（Agentic Framework）与底层 AI 框架的接入方式，力求为“精准医疗”（Precision Medicine）注入真正的精准度。 演讲者强调，如果不能将这些模型无缝嵌入到临床工作流和优秀的用户体验中，让医生能够整合信息并据此采取行动，那么所有技术都无异于一场孤立的“科学实验”。随着更广泛的数据漏斗持续接入，核心在于不仅要提供针对患者个体的诊断信息，还要结合类似患者群体的真实画像以及数字孪生预测，形成切实可行的临床指导。 ```mermaid flowchart TD accTitle: 全人多维建模与数字孪生架构 subgraph S1["数据输入与生物机制"] A["多源患者数据漏斗"] --> B["底层生物机制解析"] end subgraph S2["系统级模型与数字孪生"] B --> C["器官系统模型<br/>(循环/胰腺/内分泌)"] C --> D["个性化数字孪生体"] end subgraph S3["临床推演与应用落地"] D --> E["相似病例对照与多情景推演"] E --> F["集成工作流看板与智能体架构"] F --> G["临床决策与精准治疗优化"] end ``` ### [32:11](https://www.youtube.com/watch?v=-JtNchTupWA&t=1931s) 治理变革：从合规约束转变为战略赋能 在推进先进医疗 AI 的过程中，治理体系是不可或缺的基石。演讲者回顾道，在约14个月前加入梅奥医学中心时，健康信息技术（Health IT）领域的治理普遍被当作一项纯粹的“合规流程”（Compliance Exercise）。然而，如果认定 AI 带来的是一个颠覆性的历史时刻，就绝不能用合规的被动思维来束缚它，而必须将其视为核心战略驱动力。因此，梅奥将治理定位彻底转变为“战略赋能者”（Strategic Enabler）。 梅奥做出了一个明确的战略抉择：在前端探索与发现阶段必须保持“天马行空、不加缰绳”（Unbridled）。无论是普通员工还是临床医生，都可以依据自身的专业判断，利用机构提供的先进工具和便捷开发能力（例如氛围编程 Vibe Coding）去探索他们认为有价值的解决方案。然而，一旦方案进入转化阶段，就必须匹配一套快速高效的治理通道，确保它们迅速跨越门槛，从实验室中的“科学实验”转变为能够长期融入临床工作流、稳定带来医疗成效的实际产品，以此保障患者的切身利益。 为了实现这一目标，梅奥建立了一套分级风险治理机制（Risk Tiering）： - **低复杂度与低风险方案**：若工具的 AI 功能相对有限，且仅应用于风险极低的临床场景，便直接定级为低复杂度方案。此类方案无需经历冗长严苛的循证审批周期，而是快速放行，并重点依赖部署后的运行监控（Post-deployment Monitoring）机制进行持续反馈。 - **高复杂度与高风险方案**：将有限的精力集中投入到真正复杂、潜在非预期后果（Unintended Consequences）较大的方案中，在前端进行极其严谨的技术与伦理把控。 - **分级策略矩阵**：从低复杂度逐步过渡到高复杂度，制度约束与评估标准层层递进，既保证了创新活力，又筑牢了安全防线。 ### [34:52](https://www.youtube.com/watch?v=-JtNchTupWA&t=2092s) 核心落地经验与避坑指南 在演讲的最后阶段，演讲者总结了梅奥医学中心在 AI 落地探索中的核心经验与常见陷阱。首先，改善患者的健康结局是唯一“不可妥协”（Non-negotiable）的核心底线。梅奥拥有得天独厚的资源储备，这使得团队目前不需要过度纠结于短期的投资回报率（Line of Sight ROI）；只要确信某项技术能切实改善患者结局，机构就会全力推进其实际应用。 其次，在组织机制上，方案必须由产品发起人（Product Proponents）负全责，而不能推给泛化的机构层面，梅奥内部约 80% 的项目都由内部产品负责人直接担责。更为关键的是决策层级：作为 AI 实施的最高推动者，该岗位直接向首席执行官（CEO）汇报。任何 AI 解决方案在进入临床实践前，都必须经过该部门的审批。直接向最高管理层汇报，确保了所有决策都具备全院顶层视角，能够彻底跳出传统合规思维，将患者利益与梅奥的长期战略优先级放在首位进行综合权衡。 在实际执行过程中，团队也频繁遭遇几个极具代表性的误区： - **算法不等于产品**：临床或研发人员常兴奋地汇报某个算法性能表现极其卓越，但当被问到“它有用户交互界面吗？考虑过融入临床工作流吗？临床医生当前能用起来吗？”时，答案往往是否定的。只要缺乏这些要素，它就不是一个合格的产品，管理层也绝不会批准其上线。 - **脱离当前标准疗法空谈指标**：很多研发人员常执念于主观设定的高门槛，例如声称“模型准确率必须达到 92% 才能上线”。然而追问其依据时，往往只是因为“听起来是个好数字”；实际上，当前临床的现有标准疗法基准可能只有 73%。只要新技术能在此基础上带来有意义的显著提升，就应当果断推进，而不是被臆想的指标束缚。 - **未将效率提升转化为患者获益**：许多方案在推介时习惯宣称“能为某类人员节省大量时间”。面对此类陈述，梅奥的严格要求是：必须明确把节省的时间转化为患者最终能切身体会到的实质利益。单纯说“能帮放射科医生省时间”固然很好，但如果无法证明这对患者护理带来了具体改变，就无法确立上线的正当理由。所有技术努力的终点，必须落在患者可感知的体验与疗效改善之上。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=-JtNchTupWA&t=4s) 整个医学史上最具变革性的时刻 — 吉安里科·法鲁贾 (Gianrico Farrugia) > [09:07](https://www.youtube.com/watch?v=-JtNchTupWA&t=547s) 患者的最佳利益是妙佑医疗国际唯一需要考虑的利益（The best interest of the patient is the only interest to be considered） — 妙佑兄弟 (Mayo Brothers) > [11:15](https://www.youtube.com/watch?v=-JtNchTupWA&t=675s) 面对人工智能，医疗界首要担心的风险不应是‘系统表现是否不及预期’，而是‘因犹豫不决而未能全力拥抱 AI，进而无法将医疗提升到全新高度’。 — 吉安里科·法鲁贾 (Gianrico Farrugia) > [16:09](https://www.youtube.com/watch?v=-JtNchTupWA&t=969s) 托底提效（Lowering the Floor）旨在消除人为行政低效，而突破上限（Raising the Ceiling）则是赋能医护人员完成仅凭人类感官无法实现的医学突破。 — 米奇·特里帕蒂 (Micky Tripathi) > [21:48](https://www.youtube.com/watch?v=-JtNchTupWA&t=1308s) 将经典数据重构发掘的思路，就像‘洗发水配方式的循环复用’（Lather, rinse, repeat）。 — 米奇·特里帕蒂 (Micky Tripathi) > [25:29](https://www.youtube.com/watch?v=-JtNchTupWA&t=1529s) 让每一位走进医院的患者，都使医疗系统变得更聪明。 — 米奇·特里帕蒂 (Micky Tripathi) > [29:09](https://www.youtube.com/watch?v=-JtNchTupWA&t=1749s) 目前所展示的一切成就，本质上都依然是孤立的单点解决方案。 — 米奇·特里帕蒂 (Micky Tripathi) > [32:11](https://www.youtube.com/watch?v=-JtNchTupWA&t=1931s) 在前端探索与发现阶段必须保持‘天马行空、不加缰绳’（Unbridled），但一旦进入转化阶段，就必须匹配快速高效的治理通道。 — 米奇·特里帕蒂 (Micky Tripathi) > [34:52](https://www.youtube.com/watch?v=-JtNchTupWA&t=2092s) 改善患者的健康结局是唯一‘不可妥协’（Non-negotiable）的核心底线。 — 米奇·特里帕蒂 (Micky Tripathi) > [34:52](https://www.youtube.com/watch?v=-JtNchTupWA&t=2092s) 算法不等于产品。只要缺乏用户交互界面、没有融入临床工作流，管理层就绝不会批准其上线。 — 米奇·特里帕蒂 (Micky Tripathi) ## Mentioned 米奇·特里帕蒂 (Micky Tripathi), 吉安里科·法鲁贾 (Gianrico Farrugia), 里克·格雷 (Rick Gray), 亨利·普拉默 (Henry Plummer), 妙佑兄弟 (Mayo Brothers), 妙佑医疗国际 / 梅奥医学中心 (Mayo Clinic), 波士顿咨询集团 (BCG), 美国复苏与再投资法案 (ARRA / HITECH), Epic Systems, 普拉默病历 (Plummer Chart), OpenAI, Anthropic, 微软 (Microsoft), Doctronic, Atlas 病理基础大模型, NBC News, Record Time (病历极速检索系统), Nurse Virtual Assistant (护士虚拟助手), Avatar (遗传咨询虚拟化身), Image Encoder (胸部 X 光图像编码器) ## Video description Micky Tripathi - Chief Artificial Intelligence Implementation Officer, Mayo Clinic AI holds enormous promise to improve patient care, but only if it moves beyond research and into everyday clinical practice. Today, the biggest challenge is not developing AI, but ensuring it is safely implemented, trusted by clinicians, and meaningfully integrated into care delivery. This session will explore how healthcare organizations can accelerate the adoption of clinical AI while maintaining a focus on safety, trust, and the human connection to deliver better outcomes for patients. [Source](https://www.youtube.com/watch?v=-JtNchTupWA)

[0:04] 吉安里科·法鲁贾 (Gianrico Farrugia): 整个医学史上最具变革性的时刻 [9:07] 妙佑兄弟 (Mayo Brothers): 患者的最佳利益是妙佑医疗国际唯一需要考虑的利益（The best interest of the patient is the only interest to be considered） [11:15] 吉安里科·法鲁贾 (Gianrico Farrugia): 面对人工智能，医疗界首要担心的风险不应是‘系统表现是否不及预期’，而是‘因犹豫不决而未能全力拥抱 AI，进而无法将医疗提升到全新高度’。 [16:09] 米奇·特里帕蒂 (Micky Tripathi): 托底提效（Lowering the Floor）旨在消除人为行政低效，而突破上限（Raising the Ceiling）则是赋能医护人员完成仅凭人类感官无法实现的医学突破。 [21:48] 米奇·特里帕蒂 (Micky Tripathi): 将经典数据重构发掘的思路，就像‘洗发水配方式的循环复用’（Lather, rinse, repeat）。 [25:29] 米奇·特里帕蒂 (Micky Tripathi): 让每一位走进医院的患者，都使医疗系统变得更聪明。 [29:09] 米奇·特里帕蒂 (Micky Tripathi): 目前所展示的一切成就，本质上都依然是孤立的单点解决方案。 [32:11] 米奇·特里帕蒂 (Micky Tripathi): 在前端探索与发现阶段必须保持‘天马行空、不加缰绳’（Unbridled），但一旦进入转化阶段，就必须匹配快速高效的治理通道。 [34:52] 米奇·特里帕蒂 (Micky Tripathi): 改善患者的健康结局是唯一‘不可妥协’（Non-negotiable）的核心底线。 [34:52] 米奇·特里帕蒂 (Micky Tripathi): 算法不等于产品。只要缺乏用户交互界面、没有融入临床工作流，管理层就绝不会批准其上线。

[9:07] the best interest of the patient is the only interest [13:09] Micky Tripathi: lowering the floor [13:09] Micky Tripathi: raising the ceiling

[0:04] Hi everyone. I'm Mickey Chopathy. I'm the chief AI  implementation officer at Mayo Clinic and really [0:10] excited to be here. And my biggest challenge is  going to be um how to contain all of my excitement [0:15] in a short 20 minutes. So, um, buckle your seat  belts cuz we're going to cover a lot of ground, [0:20] but wanted to give you a perspective on  how we're thinking about AI at Mayo Clinic, [0:25] which is to say that we believe that this is  one of the most transformative, if not the most [0:30] transformative moment in modern medicine. If you  ask our CEO, Jean Rico Fuja, he would say that [0:35] this is the most transformative uh moment in the  history of medicine. Um, so we're taking it very [0:40] seriously and we believe that this is a huge  opportunity um for uh for healthcare. Um and [0:47] uh and uh so I'll give a little bit of perspective  on that. Then I'll dive into giving you snapshots [0:52] of the different solutions that we're putting into  place. Um because I think that you know we can [0:55] you probably seen many many many presentations on  frameworks and this framework and that framework. [1:01] I like get the excitement from the actual  solutions. Um and I'll give you um some snapshots [1:06] of the different solutions and you know how we're  thinking about them. And then finally, how do we [1:09] think about the governance part of it? Because  we see that as an enabler and that's my job. Um, [1:13] as the name chief AI implementation officer  suggests, my focus is on the implementation. [1:19] How do we get these things that would otherwise  just be science experiments into the hands of [1:23] our front-facing clinicians so that we improve  patient care? Um, because at the end of the day, [1:27] that's what it's all about. So, let me dive in.  First thing, um, I've been in health IT for about [1:34] 20 years. When I started, I was a manager at the  Boston Consulting Group and started at a project [1:40] um in Indianapolis to connect up the existing  health um uh the existing electronic health [1:46] record systems there across the hospitals um to be  able to start to enable the sharing of information [1:51] and to create like a learning platform for how can  we actually learn from this clinical data aside [1:57] from just its immediate use for the provision of  clinical care. uh back then 20 years ago there was [2:03] something like 10% adoption of electronic health  records and the biggest challenge we faced in the [2:07] um in health care was that we didn't even have an  ability to even learn what was going on across the [2:13] healthcare delivery system right most people were  living this world and I apologize to those of you [2:18] who are getting PTSD from just looking at this  slide um and all of the issues that were related [2:24] to um data that is trapped in these paper files  and the inability for us to be able to just get [2:29] a better perspective on care itself and on care  delivery what's going on with the patients but [2:35] how do we actually elevate to understand what's  going on in the health care system um how do we [2:39] even understand how we have the ability  to measure which we literally didn't have [2:44] um a short uh this is in the early 2000s a short  10 years later we um uh had uh uh the um the um uh [2:52] the American uh recovery and re reinvestment act  that put $35 billion in public dollars and I'm or [3:00] at least an equal amount in private dollars to a  large-scale investment to transform the healthcare [3:05] delivery system in the US from a paperbased system  to one that is now essentially digitally native is [3:11] now 95% 98% of hospitals across the US are using  certified electronic health record products and [3:18] 80 85% of ambulatory providers across the country.  So we have that digital foundation now. And what [3:24] that did is it transformed all of that paper data  into these digital data elements. So it digitized [3:31] that system. But one of the things that we found  was that that still didn't get us to the kinds of [3:36] learning capabilities, the evidence generation  that we wanted to be able to get to that next [3:40] level to say what are the inferences and the  insights that we can derive from that all of [3:45] that digital information. Yes, we made it digital  which was hugely important. It was a really really [3:50] big first step. Um but now how do we take that  data that was in all these different formats and [3:55] still is in all these formats, all these different  um unstructured data elements, structured data [4:00] elements, narratives, images, um audio, more  and more modalities of information. How do you [4:07] synthesize all of that to be able to get that next  level of insight that you want to be able to get? [4:12] um as people started to recognize and certainly  at Mayo Clinic as we recognize that most of that [4:18] information was trapped in these digital silos  um and what we were finding was we still couldn't [4:24] get to that next level of insight that we wanted  to be able to get to. Um that we still couldn't [4:28] use that data beyond the original documentation  intent. Right? You have to document the care, you [4:33] document that, but we weren't able to say to how  do we actually find the needle in the haststack? [4:38] um how do we um get those kinds of inferences  to be able to get the genuine insights that we [4:42] want to be able to get. Um so we see this next  generation of technologies in the you know sort [4:51] of in that trajectory of saying we started to  digitally enable all that data. We made the data [4:55] digital. We start to get um different analytic  capabilities but now we have these amazing AI [5:01] um tools that allow us to first use predictive AI  to be able to categorize that data to be able to [5:06] start to make sense of it. Generative AI that  starts to derive insights from the knowledge [5:11] gained from that data. And then finally, Agentic  AI to say how do we actually take action from the [5:16] inferences and the insights that we get from  that to be able to automate more and more of [5:20] these capabilities to point these tools at a goal  and not have to define every step of how to get [5:26] there. Just say here's the goal and allow these  agentic systems to be able to do that. Right? So [5:31] really really powerful moment for us as we start  at Mayo start to say we have all of this data. How [5:36] do we make use of it now to be able to get to that  next level? some of the basic underlying trends [5:43] that we also want to be able to say and recognize  within Mayo to in the spirit of you know let's [5:47] go to where the puck is heading is first off AI  is getting better every single day every single [5:53] day as we sit here it is getting better right  and we can certainly point to many areas where [5:58] hallucinatory behavior other kinds of behaviors  which um you know which uh uh give us pause and [6:04] caution as we think about the application of these  um technologies in healthcare but we also need to [6:09] recognize that They're getting better and better  and better every single day and they're incredibly [6:12] powerful and we need to be able to harness that  and marshall that as much as we possibly can. [6:18] We also recognize at Mayo that the ability for  patients to be able to be newly empowered with [6:25] these technologies is going to fundamentally alter  the way we think about the patient experience, the [6:31] patient journey, our relationship to patients and  patients relationship to the health care system [6:36] and to Mayo Clinic. Um this is just a depiction  of the different foundation model capabilities [6:42] that are now offered to consumers where chat where  open AI um anthropic Microsoft other techn other [6:50] companies are basically saying to patients here on  your phone is the AI that your doctor uses go and [6:57] get your records that the federal government has  given you a right to get upload that into these [7:02] tools and we will give you the insights before you  even show up at Mayo Clinic. will tell you what [7:07] does that X-ray say, what do those labs say, what  do those say together, and what might be different [7:11] treatment scenarios, what might be the different  care pathways for that, right? And unbelievably [7:15] powerful. There are all sorts of issues of whether  that's a good thing or a bad thing. It is a thing, [7:19] right? I do it every single day. My mother  just came out of a care experience in the [7:24] hospital and I was on these tools every single  day taking those records, putting them in saying, [7:29] "What does that mean?" So that I could have a more  informed discussion with her clinicians when they [7:33] came into the room. Right? Very, very powerful.  But we at Mayo are also talking to our clinicians [7:37] about let's not discourage patients from doing  this. Let's encourage them to do this but let's [7:42] give them the context for saying where are these  things good and where should you be cautious and [7:47] where should we partner with you on these uh on  these because um patients should absolutely be [7:51] doing these and getting information that we don't  even have right how do we how do we think about [7:56] that as well as we think about wearables and then  the last thing is more and more technologies this [8:01] is a depiction of a of a company called doctronic  which lives in this weird demilitarized zone of [8:07] the regulatory space where it does an it's It's an  agentic tool that automates prescription renewals. [8:13] Um, hasn't gone through the FDA for clearance or  approval. Is available through the state of Utah. [8:18] Literally the state government making it available  to patients in Utah that automates prescription [8:23] renewals as a first use case for the automation  of care without a human involved in that. Again, [8:28] we may, you know, think that's a good thing,  that's a bad thing. It is a thing, right? [8:32] It is absolutely happening. And we are looking  thinking hard at Mayo about how do we think about [8:37] what parts of our care delivery can be automated,  right? what are the things that we want to be able [8:41] to make available 24/7 so a patient can absolutely  get on to a trusted Mayo capability to be able to [8:47] say here here's a whole bunch of stuff that we can  automate for you and it's with the trust that you [8:52] should have in Mayo Clinic um that we have you  know sort of vetted that tool so you should have [8:56] trust in it but also providing the convenience to  you the availability to you and the power of these [9:01] technologies and Mayo Clinic expertise to be able  to get you um to where you want to be so how do [9:07] we think about this in particular at Mayo those  are all industry trends that we are intersecting [9:12] with and trying to you know sort of um uh you  know help to um accelerate and also um take the [9:17] best advantage of for our patients. Um the Mayo  brothers there really are Mayo brothers. It was [9:22] the Mayo father and then the Mayo brothers who  were the inspiration and the founders of the Mayo [9:26] Clinic. And one of the um you know the the the  famous quotes that they had famous within Mayo I [9:31] know it's not famous nationally but famous within  Mayo was about the best the best interest of the [9:36] patient is the only interest at Mayo Clinic. And  what they said was a union of forces is necessary [9:42] to be able to provide for that care. What did they  mean by a union of forces back then in 1910? What [9:48] they meant um was things like multispety practice,  right? Core to Mayo is that you should have all [9:55] the different specialties as well as primary care  under the same under the same roof because those [9:59] clinicians need to work together on behalf  of the patient. Another thing they meant was [10:04] that physicians should be salaried. they should  absolutely be salaried so that their incentive is [10:09] what is the best for Mayo Clinic which is what is  best for the enterprise. And then one of the other [10:14] things that they thought about in Union of Forces  is an integrated medical record. Um which back [10:18] then if you ever go to Rochester, Minnesota, there  are pneumatic tubes that interlace Rochester, [10:24] Minnesota because back then Dr. Plameumber who  is our our EHR now is called the plumber chart. [10:30] It's epic but we call it the plumber chart  after Dr. Plameumber. But he came up with [10:33] the idea of an integrated medical record. So the  cardiologist who's over in that building can do [10:38] whatever documentation they do, stuff it into  a pneumatic tube. It gets sent to the central [10:42] medical records department where the cardiologist  and the oncologist and the radiologists are all [10:47] sending their records. So you have the physical  integrated medical record back at the time. That [10:51] was the union of forces then. Um now the union of  forces oops I went the wrong way. Sorry. Now the [10:57] union of forces is um uh AI and agentic solutions  as we think about it. And this is the modification [11:04] from Dr. Dr. Rick Gray who is the CEO of our  Arizona campus um who now talks about the next [11:09] level of care in that union of forces being about  um the use of agents and AI in particular and for [11:15] Mayo these technologies are particularly powerful  and important because of um uh because one of the [11:23] things that we uh have to do and that are that's  acutely important to us in um in where we sit in [11:28] you know sort of the care u the care spectrum is  that we are involved in serious and complex care [11:33] which means that patients come to us at the end  of a long journey. Well, not at the end hopefully, [11:39] but after a long journey that they're in where  they've been to many providers and now they're [11:44] coming to Mayo for the specialty of care that  we can provide for the serious and complex care. [11:49] What does that mean? It means that they come  with a long set of records already from many [11:55] many many different places and we have to have  the ability to say how do I get those records [11:59] and how do I make sense of them in rapid cycle  faction to be able to understand what's going [12:04] on with this patient and what are those needles  in the haststack that we can identify to be able [12:09] to provide hope and the next level of care to um  to patients right so that's a particular problem [12:14] that needle in the haststack problem that's more  acute for us than it is for many many other places [12:18] 50% of Our patients come from other places to come  to Mayo. They come with a stack of records, faxes. [12:24] We get 5 million faxes a day, 1.3 million unique  patients um per year, 5 million outpatient visits [12:30] per year, right? So, we need to rapidly be able  to assess that information and then be able to [12:34] make best use of it. The other thing we need to do  is be able to say, how do we get inferences from [12:40] all of this information in ways that allow us to  deliver that top of science care? because patients [12:47] are coming to us seeking answers where they  haven't been able to get answers before, right? [12:52] Someone needs to be pushing the ceiling on that  and we see that as one of our very unique roles [12:57] is to keep pushing that ceiling. So, how do we get  the best insight um from all of that information [13:02] to be able to do that? So, one of the the other  thing that we keep in mind is Dr. Fuja's um uh you [13:09] know sort of um imperative that the risk number  one as we think about AI is not that um that the [13:18] systems may not perform the way we want. It's that  we actually don't fully embrace them to get to the [13:23] next level of care. The other way to think about  this is yes, we need to worry about the misuse [13:28] of these systems, but we also need to worry about  the misuse that comes from the opportunity cost, [13:35] the misuse um uh of the opportunity for patients  to be able to get a better outcome from the [13:41] uses of these technologies, right? So, we need to  take both of those things into account. So, we're [13:46] investing a ton. Um in the last year, we invested  something like $450 million in technology itself, [13:53] aboutundred 125 million of that 100 million 125  million of that was on AI things that you could [13:59] specifically point to to say those this is about  AI and Aentic solutions a lot of data in a lot of [14:04] different pockets that siloing that I was talking  about we're doing a ton of work now to say how do [14:09] we actually normalize that data tokenize that data  vectorize that data to not only make it to make it [14:16] AI and agentic ready essentially um because that's  the core to these AI systems I think as um as all [14:21] of you can appreciate the other thing that we're  doing is we are forging partnerships with many of [14:26] the leading in organizations involved on the tech  side because we as I just said you know as I just [14:32] said we will spend tens of millions hundreds  of millions of dollars but we are at the end [14:36] a nonprofit healthcare delivery system we're a  nonprofit we're a healthcare delivery system we're [14:41] not a big tech player we can't invest billions of  dollars in the technology um but we can partner [14:47] with the organizations who are investing billions  and billions of dollars in technology to say how [14:52] do we figure out and have them help us figure  out what's the bleeding edge of that and when [14:57] is that mature enough to be able to say how do we  prov um get that into the clinical environment as [15:03] fast as we possibly can. So we've got a number  of partnerships going on there. So what does [15:07] that mean about where we are today? We have 450  460 solutions in our pipeline today. 128 of them [15:16] are actually implemented in the practice. So we  are using these um uh day-to-day in the clinical [15:21] practice. I'll give you a very quick snapshot of  this um uh in a second and then 300 plus that are [15:27] you know somewhere in the pipeline. Another thing  that's unique about Mayo is that 80% 70 to 80% of [15:33] these solutions are actually internally developed  solutions. So we certainly have vended products [15:37] and you know we have Epic and number of other  you know um uh vendors who we work very closely [15:42] with. So we're getting more and more of those  solutions um through that pipeline but something [15:46] like 75% as I said are internally developed  solutions because we have very much a culture of [15:51] um in innovation happening internally empowering  our individual frontline clinicians and staff to [15:58] take an idea and turn that into an answer to  a problem and then we say how can we actually [16:04] um uh tap into that and make that something that's  available to all Mayo Clinic patients. So I like [16:09] to categorize these solutions into two categories.  One is what are the things that can just help us [16:15] increase efficiency just improving care delivery.  So in a way lowering the floor and then we want to [16:20] raise the ceiling to say there's the opportunity  space that's expanding for the improvement of [16:25] clinical care. So we have a variety of solutions  that are in this category of um of lowering the [16:31] floor um ambient documentation something like 80%  adoption across Mayo almost commodity solution [16:36] now the epic in inbox management and a variety  of solutions and I'll give you a sense of these [16:40] other solutions that really just help with human  created administrative inefficiency and how can we [16:46] get through um to get through that and use these  tools for that but the other more exciting part [16:52] and I would argue would argue or the part that  we don't give enough um play to in the industry [16:57] to help patients understand here's the potential  here's what's in it for you is the advancement [17:04] the advancing of medical science using these AI  technologies to do things that humans ourselves [17:10] with our own senses cannot do without these  technologies and I'll give you a sense of what [17:15] those are today um so I'm going to go through  some of these you know pretty quickly um but in [17:20] the improving care delivery we have an internally  developed solution called record time as I said [17:24] we get 5 million faxes a year. And that doesn't  count all the other records that we get through [17:28] the interoperability networks nationwide where  we get tens of millions of records per month from [17:33] these nationwide interoperability networks. But we  need to be able to make sense of them. We need to [17:37] be able to process them. We have a solution  called record time that takes all of that, [17:41] scans through it, categorizes and cataloges all of  that, and then gives us um summaries of that. So, [17:47] the clinicians when that patient shows up  for their 3-day journey at in Rochester or in [17:51] uh in um uh Arizona, that clinician has a really  good sense of what do those outside materials say [17:58] so that they can then figure out what should um  uh be going on with the patient. We have a nurse [18:02] virtual assistant. As I said, my mother was just  in the hospital. And one of the things that we [18:06] found in the hospital, she was there for 5 days,  not a de Mayo Clinic site. Um um but she was there [18:11] for 5 days. Is that many of you may know this is  the nurse shift change is kind of a black hole, [18:16] right? Don't ask the nursing staff for anything  during the shift change. Um it's not their fault, [18:21] but the passing of the explicit and tacet  knowledge from one shift to the next shift [18:26] is a really big deal and takes a long time and  it's very manual effort right now in most places. [18:30] We have a virtual assistant that was created by  the nursing staff themselves to say how do I how [18:35] do we use AI technologies so that they can take  that tacet knowledge that they have and seamlessly [18:40] pass it to the next um to the next next nursing  shift um uh you know in a way that uh minimizes [18:46] that burden. We have a a set of avatars that we're  testing. Genetic counseling being one of them. Um [18:52] patientf facing avatars that allow us to expand  the reach of our genetic counseling staff. Very [18:57] short supply of genetic counseling staff. So we're  able to offer to patients. You can see a person. [19:02] You can see it and have an in-person first visit  with a person if you want. That might take nine [19:07] months. or Sunday night when your only free time  that you have because you're working all week, [19:13] you can log on to this avatar and get that first  visit in that first um that first encounter with [19:19] the avatar that can then help us direct where your  next visit should be and dramatically shorten um [19:24] the time to be able to get you into Mayo Clinic.  We've worked with Microsoft on on an image encoder [19:30] technology to allow us to process um review chest  review of chest X-rays. One of the things that um [19:37] uh that I also experienced with my mother was she  had all these different lines and tubes on her. [19:42] She had the tube, she had all these sensors for,  you know, the um the vital the vitals monitoring. [19:47] They were doing blood pressure, SPO2, all of that.  And they would come in two or three times a day to [19:52] do a mobile chest X-ray. Every single time those  tubes were in a different place, right? They would [19:57] try to, you know, sort of work their way through  them. And then imagine, you know, the radiologist [20:00] now is taking that every day. And every day they  are to figure having to figure out all right wait [20:04] a minute which one of these is a line which is a  tube. We have the solution that now automates that [20:08] and says here are the lines in the tubes so that  they can start to look at the underlying anatomy [20:12] and that saves them 30 to 45 seconds per chest  X-ray. Our radiologists are seeing 125 chest [20:18] X-rays a day. Um so that's a pretty significant  thing. Now we're now expanding it to other types [20:23] of modalities as well as um being able to do  other things like generate um uh automated [20:28] um reports. Um we have a big effort in digital  pathology. We've taken 17 million glass slides [20:35] and turned those into digital images, dim digital  pathology slides essentially, and now are driving [20:42] inferences from those from a foundation model  called Atlas um that we've published on as well [20:47] that allows us to draw inferences on the pathology  images themselves and allows us to now do things [20:52] that we're starting to experiment with in surgery,  for example, to be able to dramatically shorten [20:57] the time for of frozen sections. For those  of you who are familiar with that, it's where [21:00] they basically in the middle of a surgery when  they're, you know, sort of extracting a tumor, [21:04] for example, being able to say, "All right, I  need to do the pathology to make sure I have [21:08] clean margins." That's a process that can take 20  to 25 minutes right now. Because it literally will [21:12] take slices of tissue, and then the um and then  the pathologist takes it into a different room, [21:17] freezes it, physically dyes it, and stains it, and  then looks and says, "Yes, that's a clean margin. [21:22] No, that's not a clean margin." using a virtual  H& which is what's depicted here. You can do that [21:27] digital staining and compress all of that time  to something like five minutes and you don't have [21:32] and think about that from a resource allocation  perspective. The staff, the nurses, the surgeons, [21:37] the operating room and most important the patient,  right? All of that um is just one simple example [21:43] of the kinds of things you can get from that. Let  me jump ahead. I know I've run out of time here um [21:48] to a couple of examples of the advancing medicine  um advancing medical science because I think it's [21:52] you know really really powerful. So we've looked  at um a number of cases and this will be a little [21:57] bit of lather rinse repeat for you as I go through  some of these modalities. In this case looking at [22:01] pancreatic um uh cancer patients. So we've taken  the um the slides and the images from patients [22:08] who we know got pancreatic cancer right we know  it. And then we look back and we say, "Well, [22:13] wait a minute. We have all the images from those  patients going all the way back." And we know when [22:18] we as humans diagnosed that that patient had that  that patient had pancreatic cancer. Let's run that [22:24] through the AI and say, "Well, when does the AI  say that that patient got pancreatic cancer?" [22:29] Right? And so it goes through a dino denoising um  and then it goes through what's called identifying [22:33] visually occult panc pancreatic cancer, which  basically just means um pancreatic cancer that was [22:38] not visual that was not visually apparent. and it  finds it right in the um in the in the uh in the [22:45] image. What we're now finding is that that is a  tool that allows the early detection of pancreatic [22:51] cancer in patients who have been identified as  being potentially high risk for cancer. And this [22:55] was a um a paper that was um uh you know that  was covered in NBC News. But in particular, [23:01] what that's what we're finding with that and  we're now putting that through a set of trials [23:05] with a broader set of patients is it can reduce  the um uh the uh uh increase the survival rate for [23:12] patients over 5 years from an order of magnitude  of 3% 39%. If you can detect that cancer 18 months [23:19] to 36 months in advance, which is what this tool  is now giving an early indication of being able [23:24] to do in something like 50% of patients, right?  Everyone's worst nightmare nightmare. um brain [23:30] cancer and pancreatic cancer, right? Um so an  amazing breakthrough with something that again [23:35] that that the AI can uniquely do that a patient  that we can't do with um with the uh with our [23:40] human senses. Now I'll just go through a series  of examples of where we do the same kind of thing [23:46] with ECG for example. We take a simple ECG and  then break that out now and into nine different [23:52] dashboards that the AI can read different things  from the ECG that a physician was not able to read [23:59] different types of um of risks. Um aortic stenosis  for example or now we even are starting to do [24:06] physiological age. Here's what your geog what  your biological age is, but your physiological [24:11] age based on this looks like it's hopefully it's  more than your uh uh than your biological or less [24:16] than your biological age. Um um looking at brain  scans, same thing, early detection of different [24:24] types of dementia in this case. Not that there's  a cure for it, but it allows the neurology team [24:29] to then work with patients on what might be the  different kinds of considerations from a quality [24:33] of life perspective, from a treatment plan for  perspective, um from a lifestyle perspective [24:38] as we think about the different types of early  indications of dementia that we weren't able um [24:42] see before and the different subtypes of dementia.  Same exact thing, lather, rinse, repeat. Genomic [24:49] information, right? We have genetic information.  We know the outcomes. we start to say what are [24:53] the connections between genetic a certain type  of genetic profile and what might be an early [24:59] indicator of disease or different treatment  options. One of the things we're looking at [25:03] is patients who may be more or less responsive to  methtoate for rheumatoid arthritis for example. [25:09] It's a gold standard of care for metho for  rheumatoid arthritis but something like 70% [25:13] of patients don't actually respond well to it  but we don't know which 30 or 70% of methtoate [25:19] actually has really bad side effects. So you want  to prevent people giving that drug to people if [25:23] you don't have to. Right? That's the kind of thing  that this can um enable. And then finally, this is [25:29] not in the category of lather, rinse, repeat.  This is actually in lather right now, which is [25:33] what about new modalities. So we're starting to  look at voice. It turns out these are vocal cords. [25:38] Um amazingly enough, you think about the subtlety  of what's going on with your vocal cord and the [25:44] human limitation of being able to make sense of  that. But an AI being able to say, "Wait a minute, [25:49] an AI can actually look at that subtlety and  start to detect things um that it turns out can [25:55] be indicators of cardiac disease, for example. So  it turns out that there's a nerve that goes right [26:01] over your heart and then it passes right next to  your vocal cords. And if you have certain types [26:05] of cardiac conditions, aortic stenosis,  enlargement of a ventricle, malperforming [26:10] uh valve for example, that can impinge on the  nerve and change make subtle changes in your [26:16] voice that the human ear can't detect but the AI  can detect. And we're finding tremendous success [26:20] now in looking in the early indications of the  ability to be able to identify cardiac disease to [26:25] the point that we're now for cardiology patients,  we actually take a voice sample of a patient when [26:30] they come. Just like we'll take blood, we'll take  a tissue sample, we'll say here, here's an app on [26:35] your phone. Please read into this so that now we  have a baseline and we can track your voice every [26:39] single time you come. Um, and our neurology team  is looking at different things that are language [26:45] dependent that it different indications of  cognitive changes that again the human might not [26:50] be able to detect a different change in a cadence.  For example, we all know word salad which is the [26:55] extreme version of that. But what about subtle  changes in the language that might be indicative [26:59] of a cognitive change or a mental illness um onset  for example that you're able to detect through the [27:05] AI? And then the last thing I'll just cover is uh  in terms of in in terms of you know lathering new [27:11] types of data. What about telemetry on the patient  as a whole? So we have $9 billion of of investment [27:18] going on now in new physical um capabilities. If  you come to Rochester, Minnesota, you will see the [27:24] skyline is construction cranes. And in Florida  as well, in Jacksonville, a ton of construction [27:29] cranes. And all of those rooms were now outfitting  to be able to do ambient detection. Um, and what [27:35] that does is, you know, not only is it just a  basic safety thing and a B patient experience [27:40] thing. My mother was in the hospital. I don't know  if all of you know that. Um, but she was in the [27:44] hospital. And one of the current standard of care  right now is they'll put a detector for a patient [27:49] who has a fall risk. um for example, they'll  put a detector under it so that they stand up, [27:54] it screams, right? And for my mother, that was  like it scared the be Jesus out of her every [27:59] single time. And also, she said, "I feel like I'm  in prison, right? If I just get out of the out of [28:04] my bed, all of a sudden this alarm goes off as if  I'm in prison and people come rushing in." Well, [28:10] what about a solution that allows a better way  of doing that? So you can have a person who is [28:14] able to detect that in a more passive kind of way  and be able to um uh respond to that. But also and [28:20] perhaps more important think about all of the data  that is gathered from that patient um uh from the [28:27] you know sort of the gate of the patient and all  of that. If you think about the data that's used [28:31] in self-driving cars every time I drive a car it  informs the you know the self-driving um companies [28:37] about the behavior of you know sort of the  environment, the behavior of me all of that. Think [28:42] about all of the data that you're able to get from  that about what are the changes in a patient's [28:47] gate that give you a better indication of fall  risk versus not non-fall risk for example and this [28:53] doesn't even include the audio data right so we  literally start to look at this as every patient [28:58] makes us smarter and that's a you know that's a  real picture actually that's a real video from [29:03] um from one of our Florida sites so um and I know  I'm going over time here but let me just uh cover [29:09] a couple more slides everything I showed you was  kind of point solutions though, right? I mean, as [29:15] phenomenal as I think they are and I think they're  really just amazing. Every one of them is a point [29:19] solution. Here I showed you a cardiology one  and here's, you know, a cancer one and here's an [29:24] imaging one over here and here's another one over  here. Right? Every one of them a point solution [29:29] that says, "Oh, okay. I can tell you something  about this patient from a heart perspective. I [29:32] can tell you something about this patient from  a cognitive perspective. Um, I can tell you [29:36] something about this patient from a chest X-ray  perspective." Every one of them point solutions. [29:40] What we want to be able to do is synthesize all of  that, right? Is to say, well, we're not interested [29:44] in the different pieces of the patient. We're  interested in the patient as a whole. So, how do [29:48] we do that? We do all those point solutions, but  then we start to say, all right, well, that can [29:53] help me go backward to get a better understanding  of the underlying biology. So, I actually [29:58] understand the science better and what's going on  there. But it also allows me to go forward to say, [30:03] how do I start to build models of organ systems? I  can build a model of the circulatory system. I can [30:08] build a model of the pancreatic system for example  and um the endocrine system for example that [30:14] then allows me to start to think about different  scenarios for different types of diagnoses that I [30:20] might be able to think about in a more complex um  kind of way and to ultimately start to think about [30:26] more of a digital twin. So, I'm be able to, you  know, sort of learn from all of that patient data, [30:31] but then be able to say, how do I create a  digital twin of you or you or you based on [30:37] your individual patient characteristics that then  allows me to start to say, how do I work with the [30:42] patient on thinking about what are different  diagnostic possibilities here? And importantly, [30:48] what are different potential prognosis or therapy  optimizations in different scenarios? literally [30:54] say based on all the data we have and thinking  about patients who might be in our database who [31:00] look a lot like you actually and what was their  experience and now I say oh here are four three [31:05] or four different scenarios for that right that's  the goal um and I'm not saying that we're there [31:11] but that's the goal is to try to say how do we get  to um you know to all of that and importantly all [31:16] of that's a science experiment until unless we can  get it into some kind of workflow integration and [31:23] user experience where the clinician is able  to integrate all of that information and be [31:28] able to take action on it. So we have more data  coming in a broader funnel of data. um we have [31:34] the ability to start to say all right what are  the different kinds of diagnostic information [31:37] that is more that is specifically important to you  as the patient and then finally how do I project [31:44] different scenarios both based on other patients  who may be like you in terms of you know what the [31:50] what they're what that profile looks like but also  looking at the different kinds of predictions that [31:55] we can make based on you know sort of a digital  twin uh you know kind of approach. So again, [32:00] none of this is in production because  we're still building all those pieces, [32:03] but we're starting to now say, well, what would  that dashboard look like? And how would you build [32:06] the kind of agentic framework and the kind of  AI framework to be able to plug those things in [32:11] bringing precision to precision medicine? Finally,  none of that works without governance. Um, one of [32:19] the things that we have found, and I've just sorry  I'm going over just two or three more slides. Um, [32:23] thank you for staying. um is that um when I came  uh to Mayo about 14 months ago, we had kind of [32:30] thought about governance was kind of a compliance  exercise in health IT, right? It was really about [32:34] compliance. And one of the challenges and one of  the things that we have found is as I described [32:38] going all the way to the beginning, we see this as  a transformative moment. And if you see something [32:43] as a transformative moment, the things that you  think are going to drive that, you can't think [32:47] about as compliance things, right? You have to  think about those things as strategic things. And [32:53] so we've kind of turned this to say wait a minute  governance is actually a strategic enabler of all [32:58] of this. And at Mayo we very much believe that the  discovery part needs to be essentially unbridled. [33:05] And again this is a strategic choice that we've  made which is to say that all of our staff all of [33:10] the clinicians are essentially doing discovery  according to what they think is important as [33:16] clinicians and using these tools that we provide  to them and try to make available to them. [33:21] um that lowers you know that that obviously are  make it much easier to create these tools. If you [33:26] think about vibe coding and you know the abilities  of the capabilities for them to be able to create [33:29] pretty powerful solutions but as those start to  go through the process we start to say wait a [33:35] minute we need to have an accelerated governance  process for getting them over the line. So they're [33:39] not just science experiments they actually are  solutions that are going to be integrated in [33:43] workflow and that are going to be able to deliver  outcomes on an ongoing basis so that we actually [33:50] have the assurance that patients are benefiting  from these at the end of the day. And that's my [33:55] job. One of the things that we've done is put in  a risk taring. So we actually have a better way of [34:00] saying what's a lowrisk solution. And if it's got  a pretty limited AI functionality, it's going to [34:05] be used in pretty lowrisk clinical environments.  Let's just call that a low complexity solution and [34:12] allow it to get through much faster. we don't have  to put it through all these, you know, evidentiary [34:17] um uh, you know, sort of, uh, evaluations. If at  the end of the day, it's relatively low risk and [34:22] we're going to put it in an environment where  we can actually do post-eployment monitoring [34:25] and put that into place and then allow that to be  the feedback and then spend our time on the ones [34:30] that are genuinely more more complex and have the  greater potential for unintended consequences that [34:36] we need to be able to manage up front. And that's  leads to a set of policies that are of increasing [34:42] you know sort of rigor as you get from low slow  low complexity solution to a high complexity [34:48] solution. So what are the lessons learned so far?  And this is my last slide I promise. Um there are [34:55] many of them and we're all on this journey. Um  for you know from our part um a couple of things [34:59] I'll just highlight and those are in those are  in yellow. For us better patient outcomes is the [35:04] non-negotiable at the end of the day. We do have  the the fortune of having resources where we don't [35:10] have to think about line of sight ROI right now.  We certainly will, right? Um, but a part of being [35:14] Mayo Clinic is that we're going to think about  the patient outcome first and say, "Let's get that [35:18] into p. If we believe it's going to improve an  outcome, let's get it into practice." So, that's [35:23] our non-negotiable outcome for every one of these.  Um, the other is that we need to have the product [35:28] proponents being the ones who are responsible  for this. The enterprise is not responsible [35:33] for making this real. The product owner and the  product proponent, 80% of them internal to Mayo [35:39] are the ones who are responsible for it. And then  executive level decision-making. I report directly [35:44] to the CEO. That's an important construct within  the Mayo enterprise, which is that the driver, [35:50] in this case me, of that um process for every  solution that has to go through me and my [35:56] department before it's implemented in the clinical  practice. I report directly to the CEO so that the [36:01] entire enterprise understands that we're speaking  at the top level for all the decisions we're [36:06] making. And it can only be an executive level  that takes this out of a compliance mindset and [36:12] turns it into something that says here's how we  weigh the pros and cons of each of these solutions [36:17] and take patient benefit and the Mayo um strategic  imperative um explicitly into mind as we're making [36:23] the decisions about which ones get implemented.  And then finally, there are certainly many [36:27] pitfalls that we're still encountering. One is we  got lots of great ideas, lots of great algorithms, [36:32] but they need to be a product. It's like great.  Hey, this algorithm is perform per performing [36:36] really well. It's like awesome. Is there a user  interface on any of that? Have you thought about [36:41] the workflow on any of that? Could any clinician  or any of our clinical team be able to use that [36:46] in its current form? Well, no. Well, then it's not  a product. Um, and we're not going to consider it. [36:51] Um, lack of comparison with the current standard  of care. Um, many many times we see, you know, [36:57] we're not going to use this or we're not going to  do this because it has to hit 92%. Where'd you get [37:00] 92%. um well, you know, we just thought that was  a really good number. It's like great, we think [37:06] our current standard of care is like 73%. So maybe  we ought to be thinking about it in that context [37:11] and say if we can make improvements, meaningful  improvements on the way we do things today, we [37:15] should be doing that. Um and the last thing I'll  just point out is often we get solutions that say [37:20] it's going to save this person time. It's going  to save this thing time. And one of the things [37:24] we come back is and say is please translate that  into how a patient is going to benefit at the end [37:29] of the day. I understand a radiologist is going to  save some time. That's great. That's cool. But you [37:34] need to translate that into how a patient's going  to benefit. Otherwise, we can't justify that to [37:39] ourselves, right? It's got to be something that we  can capture at the end of the day that a patient [37:43] can actually experience. So, I really apologize  for going over. I hope this gives you a good [37:47] sense of how we at Mayo think that this really is  a transformative moment and what we're doing to, [37:53] you know, sort of for in our part to help all  of us move forward. So, thank you very much.