Reducing Burnout and Transforming the Clinician Experience with AI

借助AI缓解职业倦怠并重塑临床工作体验 来自Northwell Health、马里兰大学医疗系统（University of Maryland Medical System）及CI&T的医疗高管与医疗信息学领袖共同评估了企业级AI工具在积极缓解全院网络认知过载与文书事务摩擦方面的作用。专家组通过详细剖析下班后差错疲劳的风险，对传统的“人在回路”（human-in-the-loop）教条提出质疑，并提出了一个基于任务挫败感和错误波及范围的结构化优先级矩阵。展望未来，讲者们勾勒了迈向自主智能体工作流与多模态感知的演进路径，敦促医疗机构强制推行AI素养培训，并将采纳AI视为财务与人文层面紧迫的必然选择。 Northwell Health 构建了包含 25 个大语言模型且符合 PHI 规范的企业级“模型花园”，覆盖 11 万名员工，日均活跃用户约 2,500 人并已记录超 1,800 万次交互，在满足一线强烈需求的同时遏制非托管消费级工具的风险。 (2:27) 专家指出 AI 必须被视作数字化临床参考书而非绝对权威，法律与伦理责任始终由床旁临床医生承担；过度依赖 12 小时高强度倒班后身心俱疲的医生把关易产生严重的“差错疲劳”，使人机协同（human-in-the-loop）防线出现漏洞。 (5:30) 环境感知的病历转录（ambient scribing）每天可为临床医生节省 20 至 90 分钟，但只有在医疗系统保留医生自主权、而非利用节省的时间强加更高患者配额时，才能切实减轻职业倦怠。 (21:29) 高收益的 AI 落地场景应遵循优先级评估公式：针对耗时长、医护人员痛恨度高、数据就绪度好且出错时“破坏半径”（blast radius）极小的工作流程切入；医护人员并非抗拒变革，而是理性抵制增加点击、复杂性与专业风险的工具。 (25:46) 随着 AI 从预测、处方生成走向自主智能体（Agentic AI），以及病房视觉传感和可穿戴设备的普及，临床医生需从“人工智能”转向“增强智能”协作，但多源数据互操作性障碍与监管政策仍是决定完全自主化落地的关键边界。 (35:24)

Healthcare systems are deploying enterprise language models and ambient scribing tools to alleviate severe clinician burnout, but realizing sustainable relief requires rigorous use-case selection, collaborative governance, and resolving the risks of human oversight fatigue. Targeting High-Friction Administrative and Clinical Workflows (2:27) Health systems are targeting routine documentation and administrative chores rather than flashy diagnostic algorithms to address cognitive overload. At Northwell Health, an enterprise "model garden" hosting 25 secure large language models logged over 18 million interactions across 47,000 active staff users, serving both administrative drafting and de-identified clinical scenario analysis. At the University of Maryland Medical System (UMMS), ambient scribing deployed across outpatient clinics saved between 20 and 90 minutes per day per provider, while automated "care sequencing" models targeted inpatient case management—a field where registered nurses spend up to 90% of their working hours on manual documentation, phone calls, and faxing. Similarly, automating after-hours inbox messages and prescription triage saved roughly 900 hours per month across a single 20-person clinic network. The Cognitive Strain of Human Oversight (12:15) While health systems mandate a "human in the loop" as a primary safeguard against algorithmic errors, panelists warned that human oversight is vulnerable under real-world clinical conditions. Clinicians reviewing machine-generated documentation at the end of exhausting 12-hour shifts suffer from severe "error fatigue," heightening the risk that subtle AI hallucinations or transcription faults slip past validation. Furthermore, editing imperfect AI drafts can introduce a cognitive tax that requires more mental effort than drafting from scratch. Addressing legal liability, Dr. Eric Cruzen noted that because modern clinical training operates as an open-reference environment, legal and ethical responsibility remains strictly with the bedside clinician, who must treat AI outputs as digital reference texts rather than definitive medical authorities. Prioritizing Low-Risk Tasks and Cross-Functional Buy-In (25:46) Engineering and product teams evaluate potential healthcare AI use cases using a formula balancing four variables: time spent, clinician dissatisfaction ("level of hatred"), data readiness, and the "blast radius"—the potential clinical damage caused by an algorithmic error. High-friction, low-risk administrative tasks represent the ideal starting point to build workforce trust. The panel rejected the assumption that clinicians reflexively resist change, maintaining that clinicians instead resist added clicks, workflow complexity, and uncompensated liability. Successful adoption avoids top-down mandates by co-creating tools with frontline doctors and nurses from day one. Engaging legal, compliance, and clinical teams early—while enforcing ownership boundaries through RACI governance frameworks—turns potential institutional gatekeepers into active advocates. Transitioning to Agentic Autonomy and Technical Literacy (35:24) Clinical AI is evolving across three distinct phases: predictive risk alerts, prescriptive generative recommendations, and emerging agentic systems capable of autonomous task execution. While computer vision models already match or exceed human diagnostic accuracy in data-intensive specialties like radiology, regulatory protections currently maintain the clinician as the formal decision-maker. However, as in-room ambient computer vision and wearable sensors bridge the gap in physical patient evaluation, panelists questioned whether the insistence on human supervision will remain practical across all workflows. To prepare the workforce, medical schools like the University of Maryland are integrating mandatory AI literacy modules to teach incoming physicians how to spot hallucinations, challenge sycophancy, and use models as collaborative augmented intelligence. With clinician replacement costs reaching approximately $500,000 per departure, mitigating burnout through automation has become an economic imperative. However, health systems face an unresolved dilemma: determining how to maintain effective clinical accountability as autonomous agentic models increasingly outpace the cognitive reserves of exhausted human supervisors.

面对严峻的医护人员职业倦怠，医疗机构正积极引入生成式与环境感知AI来精减繁重的工作流；但多位临床及数字化转型主管强调，AI落地的成功关键不仅在于技术本身，更取决于防范“人在回路”中的审核疲劳、建立跨部门共创机制，以及在技术向自主智能体演进时坚守临床判断与伦理治理底线。 减轻认知负荷：环境听诊与流程自动化 (10:17) 马里兰大学医疗系统（UMMS）首席医疗官Krishnaj Gourab博士指出，过重的认知负荷是引发职业倦怠的主因。UMMS已在门诊向医生、执业护士（NP）和医生助理（PA）推广环境听诊工具，并部署了电子病历（EHR）24小时患者数据浓缩工具及智能收件箱分流系统。临床实践显示，环境听诊工具每日可为医护节省20至90分钟。此外，小型诊所网络通过自动化处理下班后的处方续方和患者邮件，每月累计节省了约900小时。麻省总医院（MGH）与耶鲁大学的多中心研究亦证实，引入相关工具后30至40天内，受试医护的身心耗竭评分出现显著下降。 Gourab博士特别强调，个案管理与利用度管理中的行政负担常被忽视，主要由注册护士担任的个案经理往往将高达90%的工时耗费在发传真、打电话与手工记录等重复事务上，导致高离职率。UMMS为此开发了用于“照护时序编排”的AI模型以接管出院协调琐事。针对护理团队“为何只减免医生负担”的诉求，Northwell Health等机构亦开始试点针对护理流程的语音自动录入系统，将第一年护士离职率作为评估AI成效的核心指标。 审核疲劳与“人在回路”的认知税陷阱 (12:15) 业内普遍将“人在回路”（Human-in-the-loop）视作保障医疗AI安全的不二法门，但研讨嘉宾直言该防线存在脆弱性。在经历长达12小时以上的高强度轮班后，医护人员的认知储备已近枯竭，此时要求他们审查机器生成的记录，极易诱发严重的“错误疲劳”，导致细微的AI幻觉或复制粘贴错误被忽略。在医疗失败代价极高的背景下，过度依赖疲惫的人工把关蕴含着深刻的临床与法律风险。 伦理学家Reid Blackman与Northwell Health首席数据信息官Eric Cruzen探讨了“自动化偏差”的危害。Cruzen认为管理AI如同面对“四刃剑”，即使系统整合了最新医学证据与医院规章，最终裁决权仍必须归属于床边医护。Cruzen指出，AI的定位应当是实时参考工具而非绝对权威，正如医生无法因教科书差错而免责一样，床边决策的法律与伦理责任始终由执业医师承担。与此同时，低质量的AI起草反而会迫使临床人员耗费巨额心智去校验修正，形成得不偿失的“认知税”。 规模化部署与“爆炸半径”评估模型 (25:46) 在企业级推广方面，Northwell Health打造了包含25种大型语言模型的安全隔离“模型花园”，向全院11万名员工开放。在上线六个月内，该平台吸引了约4.7万名独立活跃用户，年内交互量突破1800万次，且企业级许可模式使算力成本保持在可控范围内。其中三分之一的使用者为临床人员，其输入的提示词中逾半数直接用于临床场景总结与第二诊疗意见参考。 为了合理规划开发顺序，CI&T的高级副总裁Luiz Cieslak与小组提出了一套用例优先级评估体系，综合考量任务耗时、医护厌恶程度、数据就绪度以及“爆炸半径”（AI出错时引发的临床或运营损害）。优先选择高耗时、高厌恶感但爆炸半径极小的低风险繁琐事务，能以最快速度累积用户信任。Cruzen据此驳斥了“医生天然抗拒变革”的行业成见，明确指出医护人员并不排斥改变本身，他们真正抵制的是无实质临床获益的“额外点击、额外复杂度和额外风险”。 跨部门共创治理与智能体未来演进 (30:23) 研讨指出，医疗IT项目的成败“90%取决于人，10%取决于技术”。自上而下强行推行预设系统往往引发强烈抵触，成功的实施必须从第一天起就吸纳临床一线人员参与共创。同时，跨学科协作不可或缺：将法务与合规团队尽早引入设计阶段，可将其从被动把关的阻力转化为主动规避监管风险的推动者。为避免多方参与导致决策瘫痪，机构需采用RACI矩阵明确各方权责。 展望未来，临床AI正从单纯风险预警的“预测阶段”和提供建议的“处方性生成阶段”，向自主执行临床任务的“智能体（Agentic AI）阶段”迈进。伴随病房计算机视觉监测和可穿戴设备的普及，环境感知正逐步涉足体格检查这一传统人类领地。小组指出，在放射学等部分认知专科，计算机视觉的模式识别精度已超越人类医生，目前阻隔自动化全面替代的主要壁垒在于监管法规而非技术瓶颈。为此，马里兰大学等医学院已增设为期两个月的AI必修课程，着力培育医学生的AI素养与批判性审视能力，避免对模型产生盲目谄媚式信任。 医护职业倦怠不仅造成个人身心消耗，更会削弱医疗质量并推高离职率——替换单名离职医生的综合成本高达约50万美元。小组一致认为，应用AI解放生产力是一项紧迫的运营刚需，但医疗机构必须保障医生的时间自主权，避免将节省出的工时简单转化为强制增加的看诊配额；同时，在技术加速迈向自主智能体的进程中，医疗系统亟须化被动为主动，前瞻性构建起兼顾效率、伦理与法律责任的安全防护网。

## Background 该视频由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台。视频记录了一场关于医疗 AI 落地的圆桌对话，邀请了来自 Northwell Health、CI&T、马里兰大学医疗系统及 Virtue 咨询公司的专家，探讨如何利用环境感知听写、临床文书生成和流程自动化等 AI 工具减轻医护人员沉重的行政负担，从而缓解职业倦怠并重塑临床执业体验。 ## TL;DR 圆桌聚焦于医疗 AI 如何从疾病诊断的宏大叙事转向切实缓解医护人员的繁重事务与认知负荷。与会专家分享了包括私有模型花园、环境感知听写（Ambient Scribing）以及护理时序管理模型在内的落地实践，并深入剖析了自动化偏见、人在回路（HITL）的局限性与倦怠评估指标。讨论强调，AI 用例的筛选应聚焦高耗时、怨气重且爆炸半径小的场景，依托跨部门协同共创与 RACI 矩阵破除抗拒，最终借助多模态感知与智能体演进实现真正的人机协同，将医生从繁杂键盘操作中解放并还给病床旁。 ## Atmosphere & style 整场对话务实、严谨且充满幽默机锋，两位主持人分别从组织变革与伦理治理的视角展开审视，与一线医院管理者及技术专家形成了高质量的思维交锋。讨论直面 12 小时高压轮班下医护“差错疲劳”等残酷现实，打破了将人在回路视作绝对防线的迷思，充满了对一线临床体验的深刻同理心。叙事节奏明快且极具温度，在犀利剖析技术局限与制度痛点的同时，描绘了一幅由增强智能重塑医疗人文关怀的未来图景。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=4s) 医疗 AI 当前最紧迫的落地价值不在于宏大的诊断宣称，而在于切实削减沉重的行政与文书负担，从根本上缓解医护人员的职业倦怠。 - [02:27](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=147s) 诺斯韦尔健康（Northwell Health）构建了集成 25 个大语言模型的私有“模型花园”，在确保 PHI 隐私合规的前提下面向全院 11 万名员工开放，实现超过 1800 万次交互。 - [05:30](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=330s) 防范“自动化偏见”的关键在于强化临床教育；AI 与教科书地位类似，床旁的第一手观察与最终裁决权始终属于医生（Human in the Loop）。 - [09:42](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=582s) 马里兰大学医疗系统（UMMS）通过环境感知听写、EMR 摘要工具与护理时序模型，重点解决门诊录入及个案管理护士 90% 机械文书导致的极高流失率。 - [12:15](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=735s) 在经历极限轮班后，“人在回路”的疲劳医护人员往往成为安全链条中最脆弱的一环，极易引发“差错疲劳”问题。 - [13:54](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=834s) 职业倦怠的量化需结合病历完成时间、满意度量表与员工流失率等指标；麻省总医院与耶鲁的研究证实环境听写能显著降低医护情绪衰竭评分。 - [18:57](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1137s) AI 带来的时间结余核心价值在于降低认知负荷并归还医生的“自主权”，同时可在急诊滞留时间和住院天数等硬性运营指标上带来实质改善。 - [25:46](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1546s) 医疗 AI 用例筛选应遵循“三变量法则”，优先选择任务耗时长、医护厌恶程度高且出错爆炸半径小的速赢（Quick Wins）场景。 - [28:03](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1683s) 破解“医生抗拒变革”迷思的核心在于新技术能否让临床医生“在其执照的最高水平下执业”，剥离键盘打字等非必要复杂性。 - [30:23](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1823s) 信息化项目成功有赖于医护技法跨部门“协同共创”，立项之初拉入法务团队并依托 RACI 矩阵厘清权责，将阻碍者转变为推广者。 - [35:24](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2124s) 医疗 AI 正从预测分析、生成建议向自主协同（Agentic AI）演进，未来将结合病房多模态感知与可穿戴设备打破单次诊疗切片，提供全生命周期洞察。 - [41:08](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2468s) 医学教育正全面重塑，新一代医学生不仅需掌握 AI 底层机制以应对幻觉，更需培养批判性提问与交互能力，将 AI 视为增强智能的合作伙伴。 - [43:47](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2627s) 职业倦怠具有个体耗竭、组织敬业度滑坡至财务重创的三维连锁反应，重置单名医生的成本高达 50 万美元，改善医护体验直接关乎医院永续运营。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=4s) 圆桌主持与嘉宾背景介绍 圆桌在一阵轻松幽默的氛围中开场。联合主持人 Jana Brown 博士率先向观众致意，并调侃这场会议迎来了“双主持人”阵容。她说明了自己的背景：作为一名首席人力资源官（CHRO），她的博士学位专攻组织变革，因此她将主要从员工倦怠（burnout）与组织转型的视角引导讨论。另一位联合主持人是 AI 治理专家、Virtue 咨询公司首席执行官 Reid Blackman，他著有《Ethical Machines》与《The Ethical Nightmare Challenge》，拥有哲学伦理学博士背景，将从 AI 治理和伦理合规的维度展开审视。 随后，三位主讲嘉宾依次亮相并互相打趣： - **Eric Cruzen 博士**：急诊科医生，现任纽约大型医疗系统诺斯韦尔健康（Northwell Health）高级副总裁兼首席数据与信息官（CDIO）。他幽默地自嘲“看来我确实需要两位主持人好好约束一下”。 - **Krishnaj Gourab 博士**：马里兰大学医疗系统（University of Maryland Medical System, UMMS）急性后期护理（Post-Acute）副总裁兼该大学创新中心医学负责人。他介绍 UMMS 是马里兰州最大的住院医疗服务提供商，旗下拥有约 10 家医院和广泛的门诊诊所。 - **Luiz Cieslak**：CI&T 公司医疗健康与生命科学高级副总裁。他笑着调侃自己是台上唯一没有博士头衔的人，只能靠身高来“弥补”差距；CI&T 专注于 AI 赋能与数字化转型，致力于帮助医院和医疗系统将解决方案真正落地到生产环境中。 --- ### [02:27](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=147s) 诺斯韦尔健康（Northwell Health）的私有模型花园实践 Jana Brown 将议题引向核心痛点：当前医疗行业充斥着 AI 用于疾病诊断与药物研发的宏大叙事，但医疗机构究竟在采取哪些切实的 AI 举措来直接缓解临床医生的倦怠？ ```mermaid flowchart TD accTitle: Northwell Health's Model Garden Architecture subgraph Enterprise ["诺斯韦尔健康全员安全 AI 平台"] MG["内部私有模型花园 (Model Garden)<br/>集成 25 个大语言模型"] SEC["PHI / PII 数据安全与合规隔离"] EDU["全员 AI 素养与提词培训"] MG --- SEC MG --- EDU end subgraph Users ["110,000 名员工生态"] CLIN["临床医护人员 (约占 1/3)"] ADMIN["行政与运营人员 (约占 1/3)"] OTHER["其他职能部门"] end subgraph UseCases ["核心落地场景"] C1["临床病例咨询与诊疗决策辅助 (>50%)"] A1["In-basket 邮件处理 / 演示文稿制作 / 日常减负"] end Enterprise --> Users CLIN --> C1 ADMIN --> A1 ``` Eric Cruzen 详细分享了 Northwell 的战略路径。与很多医疗机构采用单一单点工具（Point Solutions）或直接接入外部公共平台（如 Gemini）不同，Northwell 在 ChatGPT 最初兴起时先全面封禁了公共模型访问，随后自主构建了一个包含 25 个模型的内部“模型花园”（Model Garden）。该平台完全自研闭环，严格符合患者隐私保护规范（PHI/PII compliant），向全系统 11 万名员工彻底开放，并配套了大规模的 AI 素养教育。 - **应用规模与活跃度**：全院 110,000 名员工中，已有约 47,000 人成为注册用户，每日独立活跃用户达 2,500 人；仅当年累计的模型交互调用量已突破 1,800 万次。 - **用户构成与使用场景**：活跃用户中约三分之一为一线临床医护人员，其使用量的 50% 以上集中在具体的临床场景——由于平台具备 PHI 合规保障，医生们直接输入真实患者病情与复杂病历特征，借助 LLM 快速获取极具临床参考价值的信息；另有三分之一为行政管理人员，用于从制作演示文稿（PowerPoint）到梳理日常文书的各种行政减负任务。 --- ### [05:30](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=330s) 自动化偏见、临床判断与问责边界 Reid Blackman 提出关键质疑：当医生使用 AI 获取诊断建议时，是否会产生“自动化偏见”（Automation Bias）？即因为输出结果来自计算机，医护人员便产生盲目顺从的本能，医疗系统该如何训练员工避免过度依从？ Eric Cruzen 直言这种情况比“双刃剑”更严峻，堪称“四刃剑”（Quadra-edged sword）。他指出，有供应商尝试将最新医学循证证据与医院自身的临床指南、操作流程进行深度结合，当两者发生冲突时由医生裁定。解决过度依赖的根本在于教育：“你成为一名临床医生是有原因的，你接受了多年专业教育。就像你手头有两本写着不同观点的教科书，最终必须由你站在病床前，亲手触碰患者、亲眼观察病情，运用专业判断做出最适合该患者的决定。” ```mermaid flowchart LR accTitle: Evolution of Medical Decision Making subgraph Traditional ["传统医学模式"] M1["死记硬背医学知识"] --> M2["纸质教科书静态参考"] --> M3["个体临床裁决"] end subgraph Modern ["现代 AI 增强模式"] A1["实时循证与 LLM 检索"] --> A2["警惕锚定偏见 / 自动化偏见"] --> A3["床旁医生终审 (Human in the Loop)"] end ``` 针对 Reid 追问的“法律问责困境”——如果医生采纳 AI 建议导致患者出问题，会被指责放弃专业判断；如果不采纳 AI 建议而患者恶化，又会被指责无视智能系统——Cruzen 风趣地回应道：“我不是律师，但你能起诉一本教科书吗？”他指出当今医学教育正迎来根本性变革：过去医学生必须死记硬背所有知识，而现在的医学执业考试基本上已是“开卷”模式，医学生被明确要求在真实诊疗中实时检索参考资料。Luiz Cieslak 补充强调，变革管理（Change Management）的核心不仅是把工具造出来，更是要持续训练临床人员保持批判性思维，利用 AI 克服传统诊断中的思维盲区（如锚定偏见与鉴别诊断盲区）。 - **偏见的演进**：医学院传统上重点防范“锚定偏见”等认知启发式偏见（Heuristics），如今技术带来的“技术偏见/自动化偏见”只是新增的一种需要被审视的变量。 - **临床实践本质**：任何数字工具与教科书地位类似，都属于参考源，床旁第一手观察与专业裁量权始终属于医生。 --- ### [09:42](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=582s) 马里兰大学医疗系统（UMMS）：环境感知听写与个案管理减负 Jana Brown 总结指出，数字化转型中最关键的一环，正是确立“何时由 AI 主导，何时由人类引领”的决策质量边界。随后，Krishnaj Gourab 介绍了马里兰大学医疗系统（UMMS）针对认知过载与临床倦怠落地的三大抓手： ```mermaid flowchart TD accTitle: UMMS Clinical and Administrative Burden Reduction subgraph ClinicalFrontline ["临床一线减负"] AMB["门诊环境感知听写<br/>(Ambient Scribing)"] --> F1["解放医生 / NP / PA 键盘录入"] EMR["电子病历文本摘要工具"] --> F2["提炼过去 24 小时病情与病程进展"] INBOX["In-basket 智能收件箱管理"] --> F3["分类、优先级排序与回复草拟"] end subgraph CareCoordination ["个案与利用度管理 (Care Sequencing)"] CM_BURDEN["护士个案经理面临 90% 繁琐文书<br/>(发传真 / 打电话 / 跨机构协调)"] AI_SEQ["AI 护理时序模型 (Care Sequencing Model)"] CM_BURDEN --> AI_SEQ AI_SEQ --> CM_RELIEF["降低岗位高空缺率与职业倦怠"] end ClinicalFrontline --- GOV["AI 治理底线: 始终保持人类在环 (Human in the Loop)"] CareCoordination --- GOV ``` 1. **门诊环境感知听写（Ambient Dictation / Scribing）**：在门诊全面推广环境录音与自动生成病历，覆盖执业医师、执业护士（NP）及医师助理（PA）等高级实践医疗人员（APP）。短期研究已明确证实该技术显著提升了医护满意度，降低了职业倦怠。 2. **电子病历（EMR）认知减负与收件箱管理**： - 部署病历文本摘要工具，自动归纳患者过去 24 小时内发生的所有病情变化与病程进展，免去医生逐页翻阅海量检验记录的精力消耗。 - 引入 In-basket（消息收件箱）管理工具，辅助医护人员对每天海量的内部及患者来信进行优先级排序，并在严格的安全护栏（Guardrails）下辅助生成答复草稿。 3. **个案管理与“护理时序”（Care Sequencing）AI 模型**： - Gourab 指出，当前利用度管理（Utilization Management）与个案管理（Case Management）领域的倦怠危机极度严重。个案经理多为资深护士，但其日常工作中 90% 沦为机械性的填表、发传真、打电话等文书劳作，导致全美该岗位出现极高的离职率与空缺率。 - UMMS 研发了专门的“护理时序”（Care Sequencing）AI 模型，自动化推演与规划患者出院及转诊流程中的繁琐事务，从根本上分担事务性重担。 4. **安全底线与治理机制**：作为系统内 AI 治理委员会成员，Gourab 强调防范自动化偏见与系统失效的终极防线在于：所有临床与运营决策流程中，必须始终贯彻“人类在环”（Human in the loop）原则。 ### [12:15](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=735s) “人机协同”的盲区：疲劳医护沦为最脆弱的薄弱环节 在探讨医疗流程中的技术干预时，现场先从电子病历（EMR）的现实痛点切入。台上向观众席发问有多少临床医生或接触过病历的人在场，台下纷纷响应。大家心知肚明，每天病历系统里充斥着大量的复制粘贴错误，而在每一个错误环节中，原本都设有一道“人在回路”（Human in the loop）的安全审查。多项实证研究已经证实，在临床现实中，“人在回路”往往是整条安全链条上“极其脆弱的一环”。 这一现象揭示了医疗行业与其他领域的根本差异。其他行业普遍将人机协同视作坚不可摧的安全防线，但在医疗行业，被拉入审核闭环的医护人员，往往刚经历长达 12 小时甚至更久的极限轮班。让他们在筋疲力尽的状态下去逐行复核、修改 AI 生成的内容，“差错疲劳”（error fatigue）问题尤为致命。加之医疗体系的试错成本和失败代价远高于其他任何行业，一旦审查失守，后果不堪设想。 由此，主持人 Reid Blackman 抛出一个关键问题：当机构试图通过 AI 解决医护人员的职业倦怠（burnout）时，究竟该如何科学评估其效果？是需要建立一套严格控制变量的复杂评估体系，去追踪倦怠成因与离职率，还是仅仅依靠医护人员每天身处一线的直观感受，来判定 AI 是否切实起到了正面作用？ --- ### [13:54](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=834s) 评估职业倦怠：从病历书写时间到多中心实证研究 谈及如何量化职业倦怠，Eric Cruzen 坦言这是一个复杂的系统性课题。以 Northwell Health 推广的环境智能病历听写（ambient scribing）为例，评估工作是在全院组织层面层层展开的。在操作层面，团队密切监测病历书写完成时间（note completion time）并收集一线临床医生的直接反馈；在组织与人力层面，则深入分析医护参与度评分、专项满意度调研、人员离职率以及技术的实际采纳与使用率。 令人意想不到的是，当系统率先向医生群体铺开时，管理团队收到的第一波强烈反馈竟然来自护士群体。护士们直接发问：“那我们呢？（What about us?）我们的职业倦怠同样真实而沉重。”这直接推动了针对护理流程的环境听写技术试点，覆盖护理流转单（flowsheets）、生命体征录入等繁重的高频事务。 ```mermaid flowchart TD accTitle: 职业倦怠量化评估与落地闭环 accDescr: 展示从临床微观指标到宏观实证研究的多层级评估流程 subgraph MicroLevel["微观操作层评估"] A["病历完成时间监测"] B["一线临床医护即时反馈"] end subgraph MacroLevel["组织人力层评估"] C["员工敬业度与调研评分"] D["人员流失率与采纳率"] E["护士首年流失率追踪"] end subgraph ClinicalEvidence["多中心学术验证"] F["麻省总医院与耶鲁研究"] G["情绪衰竭量表显著下降"] end MicroLevel --> MacroLevel MacroLevel --> ClinicalEvidence ``` Krishnaj Gourab 补充道，学术界已有权威数据支持这一趋势。麻省总医院（Mass General）与耶鲁大学（Yale）在多中心合作研究中，采用经过严格信效度验证的职业倦怠量表发现，在引入环境智能病历听写系统后，医护人员的精神衰竭与情绪衰竭评分（emotional exhaustion score）出现了实质性下降。尽管这些研究只对比了实施前后 30 到 40 天的数据，尚缺乏超长期跟踪，但已提供了扎实的早期证据。此外，在个案管理（case management）与利用度管理（utilization management）中，居高不下的“护士入职第一年流失率”正成为衡量此类技术投资回报的核心硬指标。 - **多中心实证**：麻省总医院与耶鲁大学团队的研究表明，部署环境听写 30 至 40 天即可观察到情绪衰竭评分显著改善。 - **护理痛点延伸**：Northwell Health 在医生端部署成功后，迅速拓展至护理流转单（flowsheets）及生命体征的自动化记录。 - **核心人力指标**：针对个案管理与高流失科室，护士第一年离职率成为衡量技术干预效果的关键风向标。 --- ### [16:25](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=985s) 超越医生的视野：护理实时归因与跨国医疗可及性突破 Luiz Cieslak 分享了两个展现技术纵深潜力的代表性案例。第一个项目聚焦于极度高压的临床护理环境。以往在抢救或复杂处理过程中，护士必须在分秒必争的高压状态下，手动记录每一分钟发生的事件、所涉人员与具体操作。团队正在开发一套语音自动化报告系统，利用 AI 实时监听环境对话、区分不同讲话者的身份，将对应医疗行为精准归因到具体人员。这种方式将护士从庞大的“认知负荷”（cognitive burden）中彻底解放出来，无需在全力救治的同时分心记录，事后只需快速审阅系统实时生成的记录并确认归档即可。 第二个项目则扎根于巴西的双轨制医疗环境（国家公共医疗与商业医保并存）。在当地，患者获得优质照护面临严重的地理与流程阻碍。团队正在利用 AI 优化医疗资源的智能分流与接入流程：根据患者症状精准推荐最近且适配的医生，免除跨城就医的奔波；同时智能判断医保属性，在线自动完成繁杂的预先授权（pre-authorization）。对于家中还有孩子需要照料的患者而言，这意味着离开家门就医的时间被压缩至极限。AI 不仅帮患者减负，更免去了医院门诊前台与后台大量的文书冗务，使医护能够全神贯注于诊疗本身。 Luiz 强调，这些案例生动诠释了行业理念的演进：医疗领域的 AI 绝非单纯的“人工智能”（Artificial Intelligence），而是真正意义上的“增强智能”（Augmented Intelligence）。在既有临床工作流程保持稳定的前提下，AI 正在利用现有资源不断拉高“可能性的艺术”（the art of the possible）。 --- ### [18:57](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1137s) 规模化落地的指标考量：海量调用、AI素养与认知负荷缓解 面对主持人关于 47,000 名用户规模与预期目标的追问，Eric Cruzen 直言这一数据已经远超团队最初的设想，但整个体系的胃口也随之变得更大。半年来系统共承载了高达 1,800 万次模型交互（18 million model interactions），出人意料的是，得益于灵活的企业级授权机制，如此巨大的调用量并未产生令人咋舌的高昂成本。Northwell Health 正在通过增加自主智能体（agents）、丰富任务类型、开放任务与智能体共享来全面放大草根创新。为防范潜在风险，管理层甚至正在研讨将“AI 素养”（AI literacy）列入全院每年必须强制完成的合规培训模块之中，以确保临床人员充分理解模型偏差，始终保持“双重甚至三重核对”的严谨警惕，防止对算法盲目信任。 关于时间节省的价值，现场展开了一场幽默却耐人寻味的探讨。Krishnaj Gourab 指出，环境智能 AI 每天为医生节省的绝对时间往往只有 15 到 20 分钟，部分场景可达 90 分钟，单纯看数字似乎不算惊人，更难以在财务 Excel 表格中直接折算成额外收益。但它带来的核心价值在于“认知负荷的骤降”——医生心理上明确知道繁重琐事已被托底。 ```text 现场关于 90 分钟结余时间的精彩辩论： 主持反问：“节省下来的 90 分钟应该用来做什么？是去多挣几个 RVU（相对价值单位），还是回家吃晚饭？” 嘉宾回应：“我肯定回家吃晚饭！能多陪陪家人才是真正消除倦怠的解药。” 深层共识：即便有医生选择在节省出来的时间里多看 3 个病人，只要决策权依然掌握在医生自己手中即可；“自主权（agency）的丧失才是职业倦怠更致命的元凶”。 ``` 除了主观维度的缓解，系统在硬性运营指标上的回馈同样显著。在成功采纳智能体模型的院区中，急诊留观滞留时间（ED boarding time）与平均住院日（length of stay）均出现了可测量的实际下降，证明 AI 助手在理顺诊疗节奏、加速床位流转方面已释放出实质效益。 - **调用规模与经济性**：半年内实现 1,800 万次交互，在合理的企业授权模式下运营成本极为可控。 - **治理前置**：拟将 AI 偏见防范与复核技能纳入全院年度强制性合规培训课程。 - **业务硬指标改善**：成功部署模型的分院在急诊滞留小时数（ED boarding hours）与住院时长（length of stay）上均录得显著下降。 ### [23:48](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1428s) 非工作时间文书痛点与 AI 减负的“认知悖论” 在小型诊所网络的实际落地场景中，由于未部署大型一体化电子健康档案系统（EHR），医护人员在下班后不得不耗费大量私人时间处理杂务。通过引入自动化与 AI 工具整合处方、电话回复和邮件沟通，一个仅 20 人的团队每月成功节省了约 900 小时的下班后文书工作时间。然而，这种提效也引发了更深层的战略思考：在所谓的“增强智能（Augmented Intelligence）”应用中，让 AI 起草内容有时反而会增加认知负担——逐句审阅、消化并修改一篇不合心意的草稿，往往比直接在一张白页上从头撰写更加耗费心神。 这种心理阻力在临床环境中尤为剧烈。临床医护人员本身就已经被繁复的新系统和新技术压得喘不过气。如果引入 AI 不能实现真正的认知减负（Cognitive Offloading），反而迫使医生耗费更多精力去逐项核实和纠错，技术应用就会适得其反。因此，如何在立项之初精准筛选出真正具备减负价值、而非自我消耗的合适用例，成为了医疗机构数字化转型的核心考量。 ### [25:46](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1546s) AI 用例评估的“三变量法则”与变革数学模型 针对如何科学筛选 AI 落地场景，工程与管理团队提出了一套结构化的优先级评估逻辑。在工程视角下，用例的选择主要取决于三个核心变量：**耗费的时间（Time Spent）**、**员工对该任务的厌恶程度（Level of Hatred）**，以及该任务出错时的**爆炸半径（Blast Radius）**与**数据就绪度（Data Readiness）**。最理想的切入点是那些耗时长、医护人员极度反感、数据基础完备且容错代价（爆炸半径）极低的环节。通过在这些场景率先拿下一批快速成效（Quick Wins），团队才能建立信任并推动后续推广。 ```mermaid flowchart TD accTitle: AI医疗用例筛选评估矩阵 subgraph Input ["评估维度 (Evaluation Dimensions)"] A["任务耗时 (Time Spent)<br/>高耗时优先"] B["厌恶程度 (Level of Hatred)<br/>痛点越深越好"] C["数据就绪度 (Data Readiness)<br/>具备即用数据"] D["爆炸半径 (Blast Radius)<br/>低临床风险"] end subgraph Decision ["决策引擎 (Decision Engine)"] E{"综合评分与优先级<br/>(RICE / 变革公式)"} end subgraph Output ["落地路径 (Implementation Path)"] F["第一优先级：认知减负速赢场景<br/>(低风险 + 高痛点 + 高耗时)"] G["后续扩展：核心复杂流程协同"] end A --> E B --> E C --> E D --> E E -->|筛选达标| F F -->|建立信任后推进| G ``` 从医疗机构的管理视角来看，这一逻辑与产品管理中的 RICE 模型以及变革管理方程不谋而合。推动组织变革需要权衡平台危机的紧迫性（Burning Platform）、变革的迫切需求以及阻力与投入成本。变革管理必须始终以“同理心（Empathy）”为主导，因为面对庞大的人员体量（例如正在向 11 万名员工推广大型 EMR 系统的庞大机构），任何系统切换都是极其艰难的过程，唯有流程得当、换位思考，变革才能真正走通。 - **用例筛选三要素**：优先锁定“耗时长”、“怨气重”、“爆炸半径小”的任务。 - **调查问卷的正确问法**：与其问一线员工“你们想要什么”，不如直接问“你们最讨厌做什么”。 - **同理心驱动的变革管理**：正视系统上线的痛苦周期，将技术推行转化为对员工痛点的实际缓解。 ### [28:03](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1683s) 破解“医生抗拒变革”的迷思与临床经验法则 在医疗行业中，“医生天生抗拒变革”常被当作推进受阻的借口，但这实际上是一个巨大的伪命题。在应对 COVID-19 疫情期间，面对层出不穷的新指南与严峻局势，临床医生展现出了极高的适应速度与调整意愿。医生本质上并不抗拒变革本身，他们抗拒的是**额外的系统点击（Extra Clicks）**、**额外的流程复杂性（Extra Complexity）**以及**潜在的临床风险（Extra Risk）**。 因此，临床 AI 的采纳遵循一条核心的经验法则（Rule of Thumb）：**新技术是否能够让临床医生“在其执照的最高水平下执业（Practice at the top of their license）”？** 如果工具能够将医生从繁杂的键盘操作中解放出来，让他们真正留在床旁，专注于与患者及家属沟通这些充满人性关怀的工作，这样的技术就会获得发自内心的拥护。此外，变革的时机与节奏（Cadence and Timing）也至关重要，必须确保组织在特定运营节点具备承载变革的精力和容量。 ### [30:23](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1823s) 协同共创、跨部门联盟与 RACI 权责机制 医疗信息化项目的成败，90% 取决于人，只有 10% 取决于技术。回顾历史项目，所有失败案例几乎都遵循同一种自上而下的强推模式——外部管理者或非临床部门单方面宣称“我懂你们的痛点，这是给你们的解决方案”，在缺乏同理心和现场调研的情况下强行下发。而最成功的项目，则是一开始就让一线护士、临床医生和后台运营人员坐到同一张桌子前进行“协同共创（Co-creation）”。由于医生更信任具有相同临床背景与思维方式的同行，让临床骨干成为发起者与代言人，能消除后续规模化推广中的巨大阻力。 ```mermaid flowchart LR accTitle: 医疗AI协同共创与推进机制 subgraph Alliance ["跨部门共创联盟 (Co-creation Coalition)"] P["临床医生 (Physicians)"] N["护理团队 (Nursing)"] L["法务合规 (Legal / Compliance)"] E["工程技术 (Engineering)"] end subgraph Governance ["权责治理 (Governance via RACI)"] R["明确最终决策者 (Accountable)"] S["划定执行与咨询职责 (R/C/I)"] end subgraph Outcome ["落地成效 (Transformation)"] O["从抵触阻挠转向主动推广<br/>(Promoters vs Blockers)"] end Alliance --> Governance Governance --> Outcome ``` 建立广泛的跨部门联盟同样适用于传统上被视为阻碍的法务部门（Legal）。业内常戏称“法务是创意走向死亡的地方”，但若在立项第一天就将法务团队纳入共创过程，提前围绕其合规与风险顾虑设计方案，法务人员不仅不会叫停项目，反而会因为方案能够降低合规风险而转变成为项目最积极的推广者（Promoters）。不过，在组建跨学科广泛联盟的同时，必须严格依托 **RACI 矩阵（Responsible, Accountable, Consulted, Informed）** 厘清权责，明确谁拥有最终裁决权并为项目的成败负全责，防止因声音过多而导致执行效率低下或推诿扯皮。 | 推进维度 | 失败模式（传统强推） | 成功模式（协同共创） | |:--- |:--- |:--- | | **参与时机** | 方案定型后作为测试或审批环节介入 | 从 Day 1 起共同参与需求定义与方案设计 | | **利益相关方** | 单一部门推动，忽视护理或法务诉求 | 建立涵盖医、护、技、法务的跨职能联盟 | | **法务角色** | 流程后期的“阻拦者/橡皮图章” | 前期合规设计者与方案主动“推广者” | | **权责划分** | 权责模糊，失败时相互推诿 | 依托 RACI 矩阵明确最终决策人与责任归属 | ### [35:24](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2124s) 具身与智能体演进：从“人在回路”到多模态主动感知的未来 在探讨未来 6 到 12 个月内最期待的 AI 应用场景及潜在阻碍时，嘉宾们先是用一句风趣的调侃打破了严肃气氛：“机器人马上就要统治世界了，这绝对会发生。”但玩笑过后，讨论迅速切入了核心哲学问题——“人在回路”（Human in the Loop, HITL）。当前业内总是将人在回路奉为不可逾越的信条，甚至让人觉得质疑这一观点都是某种禁忌。然而，随着智能体（Agentic AI）的迅猛发展，技术正无限接近脱离人工干预独立运作的临界点。医疗行业迫切需要摆脱被动追赶技术的局面，在强有力的护栏、治理与伦理框架下主动布局，引导无需人实时干预的自主智能体发挥正面效益。 回顾医疗 AI 的演进轨迹，几年前的技术重心主要停留在“预测分析”，即识别哪些患者处于高危状态；随后生成式 AI 与语言工具的出现，让系统能够进一步输出针对高危患者的“行动建议”；而眼下正在落地的智能体则开始实质性地“协同执行任务”。尽管目前临床医生的触诊、面对面交流及综合评估依然是不可替代的感知输入源，但随着病房视觉感知、非接触式摄像头与图像视频处理技术的飞跃，AI 正在接管更底层的物理感知能力。例如，团队目前正在研发一种在严格保护患者隐私的前提下，利用病房监测提前数小时预测患者躁动情绪的技术，这将彻底重塑重症监护与住院护理的工作流。 ```mermaid flowchart TD accTitle: 医疗AI技术演进与应用层级 subgraph S1["第一阶段：预测分析"] A["高危患者筛查与风险预测"] end subgraph S2["第二阶段：生成建议"] B["生成式AI提供干预与治疗方案"] end subgraph S3["第三阶段：自主协同"] C["Agentic AI 协助执行与流程闭环"] end subgraph S4["下一代前沿：多模态感知"] D["病房视觉与传感器捕捉体征"] --> E["提前预测患者情绪躁动"] end A --> B B --> C C -.-> D ``` --- ### [38:07](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2287s) 全生命周期数据整合与人机关系的辩论：颠覆还是赋能？ 除了住院环境内的监测，未来的医疗愿景更聚焦于跳出诊室物理边界的“全生命周期全景视图”。通过智能戒指、智能手表等多设备连续追踪，医生不仅能排除患者单次就诊时的“白大褂高血压综合征”，更能将生命体征与现实生活压力源精准挂钩——例如了解患者家中是否有生病的孩子或需要照料的老人，进而看清那些引发血压波动的具体生活时刻。实现这种“把患者当成完整的人而非时间切片”的愿景，最大瓶颈在于多源数据的互操作性（Interoperability）。医疗机构不仅需要打通来自可穿戴硬件、各类专科检查以及无论体量大小的各类 EHR 系统的数据孤岛，更关键的是避免盲目囤积数据，必须将海量数据提炼为可指导行动的临床洞察并实时反馈给一线医生。 在讨论中，主持人和嘉宾之间碰撞出了激烈的观点交锋：AI 究竟是在辅助医生，还是终将把医生降级为算法的附庸？经典的“AI 是否会抢走饭碗”再次被摆上台面。嘉宾给出了极为精辟的回答：“AI 不会淘汰你，但掌握 AI 的人会淘汰你。”虽然在当下将自己定位为“AI 的合作伙伴”可能听起来有些超前，但几年后回看录像必将印证这一趋势。在放射科等特定专科，多项研究已证明 AI 的模式识别与阅片能力已超越人类平均水平，目前唯一阻碍其全面接管的只有医疗监管法规。尽管未来五年内全自主机器人独立完成复杂外科手术并不现实，但在纯认知与影像主导的领域，监管护栏确实是维持人类主导权的最后边界。 - **连续监测与现实画像**：借助可穿戴设备打破传统诊疗切片，消除白大褂效应，洞察日常生活压力源对病情的动态影响。 - **互操作性数据阻碍**：打通异构 EHR 系统与物联网硬件，核心在于从“数据囤积”转向“临床决策级洞察产出”。 - **专科颠覆与法规边界**：放射学等影像依赖型学科中，算法表现已高度成熟，政策与合规是决定其落地边界的核心变量。 --- ### [41:08](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2468s) 白袍仪式上的嘱托：从医学教育重塑到提问的艺术 如果置身于医学院的白大褂授予仪式（White Coat Ceremony），面对即将步入临床实习、眼神充满憧憬的年轻医学生，资深专家们给出了极具时代印记的职业建议：首要任务就是“百分之百地去学习 AI”。医学生不仅需要掌握如何调用和解读 AI，甚至要理解其底层逻辑。只有了解底层技术机制，才能明白模型为何会出现“幻觉”（Hallucinations），或者为何会产生讨好用户的“谄媚效应”（Sycophancy）。为此，顶尖医学院已在课程体系中植入为期约两个月的专项 AI 培训模块，并计划将其贯穿住院医师规范化培训及非临床员工的全员数字素养提升计划中。 另一位嘉宾则强调了临床思维与交互范式的转变：AI 绝非单纯的“人工智能”，它更应该被理解为“增强智能”（Augmented Intelligence），其价值在于弥补人类的认知盲区。当前许多人仅仅把大模型当成“高配版 Google 搜索”，随手输入一段语焉不详的半成品提问，就妄想模型能完美解析并给出正确方案。新一代临床医生必须掌握提问与批判的艺术：学会提出更高质量的问题、批判性审视生成的结果，并在多轮反馈迭代中持续优化决策。正如当年医学院院长在白袍仪式上那句震撼人心的名言——“当你披上这件白大褂，你的人生将彻底改变”——在当今由算法重构医疗范式的全新纪元里，这句话依然成立，只是承载了截然不同的时代内涵。 --- ### [43:47](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2627s) 倦怠连锁反应的三维模型与收官总结 会议尾声，主持人从人力资源与组织运营的视角分享了评估医护人员职业倦怠（Burnout）的三维框架。该框架揭示了倦怠如何自下而上层层击穿医疗体系：最底层是**个体维度的侵蚀**（身心耗竭），紧接着蔓延至**组织层面的动荡**（员工敬业度丧失、工作投入度下降，进而危及医疗安全与患者护理质量），最终必然传导至**财务层面的重创**。由于离职率飙升，医疗系统重新招聘并培训一名新临床医生的直接重置成本高达约 50 万美元（$500,000），这对医院的整体营收与财务健康构成了巨大的结构性压力。 ```mermaid flowchart TD accTitle: 医护职业倦怠的三维连锁反应模型 subgraph L1["第一层：个体人本维度 (Human Side)"] A["身心严重耗竭 (Burnout) 与精力侵蚀"] end subgraph L2["第二层：组织协同维度 (Organization Side)"] B["工作敬业度丧失 (Loss of Engagement)"] C["潜在医疗风险与患者照护质量滑坡"] B --> C end subgraph L3["第三层：财务运营维度 (Financial Side)"] D["人员流失率激增 (High Turnover)"] E["单名临床医生重置成本约 50 万美元"] F["机构净收入与关键财务指标恶化"] D --> E E --> F end A --> B C --> D ``` 职业倦怠是真实且紧迫的危机，积极引入 AI 来削减非临床行政负荷，其本质不仅在于解救医护人员的身心健康，更在于保障医疗机构的永续运营。因为在医疗行业中，“医护人员的执业体验直接等同于患者的就医体验”。讨论在全场致谢中圆满收官，为 AI 时代下重塑医疗人文与运营效率提供了坚实的方法论支撑。 ## Notable quotes > [05:30](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=330s) 你成为一名临床医生是有原因的，你接受了多年专业教育。就像你手头有两本写着不同观点的教科书，最终必须由你站在病床前，亲手触碰患者、亲眼观察病情，运用专业判断做出最适合该患者的决定。 — Eric Cruzen > [05:30](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=330s) 我不是律师，但你能起诉一本教科书吗？ — Eric Cruzen > [13:54](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=834s) 那我们呢？（What about us?）我们的职业倦怠同样真实而沉重。 — 护士群体 > [18:57](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1137s) 节省下来的 90 分钟应该用来做什么？是去多挣几个 RVU，还是回家吃晚饭？ — Reid Blackman > [18:57](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1137s) 我肯定回家吃晚饭！能多陪陪家人才是真正消除倦怠的解药。 — 嘉宾 > [18:57](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1137s) 自主权（agency）的丧失才是职业倦怠更致命的元凶。 > [25:46](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1546s) 与其问一线员工‘你们想要什么’，不如直接问‘你们最讨厌做什么’。 > [28:03](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1683s) 在其执照的最高水平下执业（Practice at the top of their license） > [30:23](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=1823s) 法务是创意走向死亡的地方 > [38:07](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2287s) AI 不会淘汰你，但掌握 AI 的人会淘汰你。 — 嘉宾 > [41:08](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2468s) 当你披上这件白大褂，你的人生将彻底改变 — 医学院院长 > [43:47](https://www.youtube.com/watch?v=ZP3PKC7szSw&t=2627s) 医护人员的执业体验直接等同于患者的就医体验 — Jana Brown ## Mentioned Jana Brown, Reid Blackman, Eric Cruzen, Krishnaj Gourab, Luiz Cieslak, Northwell Health (诺斯韦尔健康), University of Maryland Medical System (UMMS / 马里兰大学医疗系统), CI&T, Virtue Consultants, Mass General (麻省总医院), Yale (耶鲁大学), Gemini, ChatGPT, Model Garden, Ambient Scribing (环境感知听写), EHR / EMR, Care Sequencing Model, RACI 矩阵, RICE 模型, Agentic AI ## Video description Eric Cruzen — SVP, Chief Data & Informatics Officer, Northwell Health Luiz Cieslak — SVP of Healthcare & Life Sciences, CI&T Krishnaj Gourab — Chief Medical Officer, University of Maryland Rehabilitation and Orthopaedic Institute Reid Blackman — CEO, Virtue Consultants One of the most immediate impacts of AI in healthcare isn’t flashy diagnostics—it’s in reducing the crushing administrative burden on providers. This panel will examine how AI is being used for ambient scribing, clinical documentation, revenue cycle automation, and virtual assistants, freeing up time for patient care. Panelists will share real-world examples of deployment, how to measure ROI in both time savings and provider satisfaction, and what cultural changes are needed for adoption. Expect a practical conversation on how AI can shift the clinician experience from burnout to balance, making care delivery more sustainable for the long run. [Source](https://www.youtube.com/watch?v=ZP3PKC7szSw)

[5:30] Eric Cruzen: 你成为一名临床医生是有原因的，你接受了多年专业教育。就像你手头有两本写着不同观点的教科书，最终必须由你站在病床前，亲手触碰患者、亲眼观察病情，运用专业判断做出最适合该患者的决定。 [5:30] Eric Cruzen: 我不是律师，但你能起诉一本教科书吗？ [13:54] 护士群体: 那我们呢？（What about us?）我们的职业倦怠同样真实而沉重。 [18:57] Reid Blackman: 节省下来的 90 分钟应该用来做什么？是去多挣几个 RVU，还是回家吃晚饭？ [18:57] 嘉宾: 我肯定回家吃晚饭！能多陪陪家人才是真正消除倦怠的解药。 [18:57] 自主权（agency）的丧失才是职业倦怠更致命的元凶。 [25:46] 与其问一线员工‘你们想要什么’，不如直接问‘你们最讨厌做什么’。 [28:03] 在其执照的最高水平下执业（Practice at the top of their license） [30:23] 法务是创意走向死亡的地方 [38:07] 嘉宾: AI 不会淘汰你，但掌握 AI 的人会淘汰你。 [41:08] 医学院院长: 当你披上这件白大褂，你的人生将彻底改变 [43:47] Jana Brown: 医护人员的执业体验直接等同于患者的就医体验

[5:30] Eric Cruzen: You are a clinician for a reason. You went to school for a reason... you're the one at the bedside with the hands and eyes on the patient. [13:59] What about us? Our burnout is real as well. [28:03] Eric Cruzen: I don't think clinicians are resistant to change per se; they're resistant to extra clicks, they're resistant to extra complexity, they're resistant to extra risk. [30:23] I know your pain, I know how to solve it, here it is, start using it. [34:01] 90% about people and 10% about technology [35:24] I think robots are going to take over the world. Yeah, for sure. That's going to happen. [35:24] Well, what if there wasn't a human in the loop? [35:24] Because this patient is at risk, these are the things you should do. [39:36] AI is not going to put you out of a job, but someone who knows how to work with AI is going to put you out of a job. [41:08] When you put this coat on, your life will never be the same.

[0:04] Hello everyone. Good afternoon. Hope you all are  having a great day. My name is Dr. Jana Brown and [0:12] um I am a co-odderator. You've been co-odderated.  That's right. We're going to have a really good [0:19] time with this. Uh so our topic reducing burnout  and transforming the clinical experience with AI. [0:27] A very very important topic. I'll share with  you the vantage point I will be speaking from. [0:33] I am a chief human resources officer. So when I  say doctor, it is on organizational change. So [0:42] thinking about the perspective of burnout that's  what I will bring uh to the conversation today and [0:49] I'll also somehow a moderator. There's a story  about how we both became moderators. It's not [0:55] that interesting at any rate. So I'm a moderator  as well. My name is Reed Blackman. Um, I'm the [1:00] author of a couple books, both on AI governance.  One called Ethical Machines, the other called The [1:04] Ethical Nightmare Challenge. So, my training, my  PhD is in philosophy. I was an ethicist. So, I'll [1:09] come at this from the sort of ethics perspective,  governance perspective. Hi, I'm Eric and [1:14] apparently I really need to be moderated, be duly  moderated. He's a lot to handle. My name is Eric [1:19] Cruzen. I'm an emergency physician and the chief  data and informatics officer for Northwell Health [1:25] um a large health care system in the New York City  area. Um Krishn Gorab, I'm the VP for postcute for [1:34] University of Maryland Medical System. I am also  the medical lead for our university's innovation [1:39] center. Um University of Maryland Medical System,  largest uh inpatient provider, state of Maryland. [1:46] We have around 10 hospitals, outpatient practices  and uh so yeah excited to be here. I realize that [1:54] I have to compensate the titles of all the other  doctors that I'm surrounded with the height. So [2:00] um I'm the only known doctor on the panel.  I just realized this. I didn't realize you [2:03] nobody's perfect. We always have opportunities  to continue. Uh I'm Louis Chislick. I run [2:09] healthcare life sciences for CINT. We are AI  enablement and digital transformation company. [2:14] So we work a lot with hospital and healthcare  systems implementing solutions that actually [2:19] get used in production with clients. Great. So  let's just jump right in. Uh the first question [2:27] just relates to what specific AI initiatives you  all have going on like what's the number one thing [2:34] you're working on right now to help to address  the issue of burnout? thinking about it from the [2:40] perspective of we hear in the healthc care arena  so much about um AI supporting everything with [2:50] regards to diagnosis with discovery uh but how  is it helping our physicians what are you doing [2:56] on that front sure uh well actually I'll talk a  little bit about uh an AI initiative that we're [3:03] doing that's u for more of just our physicians  although it certainly involves our physicians [3:08] uh So many health systems are rolling out  uh AI. Some are doing point solutions, [3:16] others are doing platforms like uh uh Gemini for  instance. But we decided actually to build a um [3:24] model garden of our own. So, we built a 25 model  um PHI PII compliant fully self-contained secure [3:36] uh model garden for all 110,000 of our employees  to use. Um, we did that with uh a large amount of [3:45] uh uh AI literacy education and the idea was just  to get it out there and to hope that folks would [3:53] um uh you know find some ease to their burnout in  in having tools in their arsenal. Now this fell [4:02] on the uh coattails of us blocking all the models  uh you know as uh when chat GPT first came out. Uh [4:10] and so we actually have have found that around  47,000 of our 110,000 employees uh are users. [4:19] Uh we have 2500 unique daily users uh and over 18  million model interactions so far just this year. [4:28] about a third of those users are uh clinicians  and they are using it more than 50% for [4:35] uh clinical use cases but um another third  are are administrative folks that are using [4:41] it for everything you can imagine from PowerPoint  creation uh on to just uh uh easing their days. [4:48] So when you say they're using it for clinical  purposes, what is that what does that look like? [4:52] Is that diagnostics? Is that what are what are the  actual things that they're using it for? Sure. So [4:57] there so we do use without naming some products  some products that are specifically created for [5:05] um uh asking specific u diagnostic questions  but we found that uh when we made it PHI and [5:13] PII compatible people are putting actual patient  scenarios in and and getting in in many cases very [5:19] meaningful um information back uh from them from  the LLMs uh in helping uh with patient care. Okay, [5:30] I'll ask one more one more question then um move  on. But is there a worry at all about things [5:35] like automation bias? So, you know, being overly  differential to the outputs of a computer because [5:40] it's a computer and so you know you're smiling  so I know you've got something to say about so I [5:44] guess the you know what how do you train doctors,  nurses, who'sever using the AIs not to be overly [5:50] differential? How do you sort of check that  I don't know sort of quasi natural impulse? [5:55] Oh, I have no worries at all. Um, no, it's a, you  know, it to say it's a double-edged sword is an [6:03] understatement. It's probably a quadraedged sword.  Um, you give people information and people will [6:10] use that information. So, you you you know, when  we were speaking with some of the vendors that [6:17] the affforementioned vendors that that produce  software specifically for this purpose, you know, [6:22] that exact question comes up. Uh in one case for  instance there there is a vendor that will uh take [6:30] evidence uh and will marry it with your particular  facilities uh policies protocols workflows and in [6:39] the event that there's a conflict uh you get  to decide how that conflict is resolved. So [6:45] education is really the answer. I mean you you are  a clinician for a reason. You went to school for [6:51] a reason. Um, just like if you had two textbooks  uh that said two different things, it's up to you [6:57] to use your judgment to decide what's right for  this patient in this scenario because you're the [7:02] one at the bedside with the hands and eye on the  patient. I was going to chime in and pick back on [7:08] your experience because it's there's a what I  see when we're talking to hospital systems is [7:13] the change management aspect. It's like you do  you build the tools, you give them the tools. [7:19] more often than not they're actually you build and  they actually will come in many cases contrary to [7:25] popular belief but the education the change  management giving that reminding people to be [7:30] critical of that outcome is a is a critical part  we were talking about earlier a little bit of the [7:36] the bias we're we're seeing some aspects also of  having helping educating people to use the AI to [7:45] be more open to actually help with some of that  blind the blind sides that they they realize like [7:51] differential diagnostics and all of that on like  very controlled kids. Well, if you think about it [7:56] in medical school, they teach us about heristics,  right? Anchoring bias. There's so many kinds of [8:01] bias. And now we just have one more bias to add  to our list. Uh you know, tech bias. Uh I'm sure [8:08] there probably is a name for it, but but is  is should you give deference to the machine? [8:16] um presuming that it's smarter than you are. Um  should you give deference to this textbook or [8:21] that textbook, are either one smarter than you? I  don't know. And maybe well, how do you think about [8:28] accountability in those kinds of contexts? I I can  imagine sort of doctors being in a tough spot. You [8:34] know, you defied the AI's recommendation and the  patients went south. Why did you not listen to the [8:38] AI? It's so smart. On the other hand, you listen  to the AI and now the patient's dead. Why didn't [8:43] you use your own professional judgment? So, how  are clinicians supposed to navigate that space? [8:49] I'm not a lawyer, but can you sue a textbook?  Not to my knowledge. I'm not a lawyer. I actually [8:55] don't know. Is there a lawyer? Uh, no. But I I  think it's this is a brave new world. And I feel [9:01] like I'm hijacking I'm going to hand you the mic  next. Um, I think this is a brave new world. Um, [9:06] and uh, probably that was a really good reference,  but but it it's it we don't know the answers to [9:13] these questions, but I can tell you that, you  know, the medical boards are essentially open [9:17] book now. Um, when I went to medical school,  we were told we had to route memorize things [9:22] because there wouldn't be any other way to know  in the moment. But that's just not the case [9:27] anymore. Medical students that that I train now,  they're expected to use references in real time. [9:35] So, it's a brave new world. 1984. Yeah. So,  that's a really good point and let's uh we'll [9:42] hold on that. I want to get your voices in and  then we're going to come back to that because [9:48] it is a breakthrough moment and decisions and  decision quality is so key and making decisions [9:57] on when AI or a digital tool will lead versus when  a human will lead. That's extremely critical in [10:07] transformation. So please share your story and  then I want to hear yours too. I know you piped [10:12] in around change management would also want to  hear your use case. So please share yours. The [10:17] things that we are implementing in our system and  similar to probably yours in terms of volume it's [10:23] the ambient dictation in the outpatient area for  for providers physicians nurse practitioners PS [10:30] advanced practice providers that has been a bigger  satisfier and there have been studies that have [10:35] shown some decrease in the burnout at least in the  short term. Um but shifting a little bit from that [10:41] we are also and I'm sure other hospital systems  are doing the same. We have to decrease the [10:47] uh cognitive burden because that's one of the  biggest factor for burnout. We have text tools [10:51] in our electronic medical record now that  will summarize the data uh of and summarize [10:56] the patients what has happened in the last 24  hours what has been the patients progress. There [11:01] is some tools for inbasket management. We get a  lot of incoming emails similar to other hospital [11:07] systems. we have uh tools to to to help with  the inbasket management, how to answer them, [11:12] how to prioritize. There are of course there are  guard rails around all of that. Um apart from that [11:18] a little little different from from from purely  provider workflows we are seeing a lot of need in [11:25] AI for utilization management for case management  u because these the lot of this is very ro stuff [11:34] and somehow most case managers are are trained  nurses and the work has become 90% paperwork [11:40] faxing and calling people and doing this and doing  that they have a high degree of burnout and and [11:45] and and vacancy rates are significant in many part  many parts of the country. So we are trying to use [11:51] AI to do lot of those planning. We call it care  sequencing. We have developed a model for that. AI [11:57] can help with that. Before I stop I'll just react  to two things that very good great points that [12:03] were mentioned over here. One is the automation  bias. Um I sit in the AI governance committee [12:08] for my system. The the fail safe for everything  mentioned is there is always a human in the loop. [12:15] That is that's a very very weak link. How many  clinicians are there in this room or anybody who [12:21] has looked in Yeah. Great. What happened to our  electronic medical records? How many copy paste [12:26] errors we see every day? And on all of them  there is a human in the loop. Right. So and [12:32] there are studies that show that human in the  loop is a very very very weak link. So that's [12:36] that's that's a word of caution in that. You  mentioned the the human in the loop. I think [12:42] one fascinating thing we keep talking about in  healthcare is human in the loop is the fail safe [12:47] across all the industries but very particular to  healthcare we have an industry where we're putting [12:54] him into the loop that is after 12 hours of shift  or at the end of the 12-hour shift m sometimes [13:00] even more trying to revisit and revise the content  that's been generated. So error fatigue is a big [13:06] much bigger deal in healthcare than it is in other  industries that I've seen and cost to failure. So [13:12] it cost is probably the highest as compared to any  other industry. How do you think about and maybe [13:19] maybe the answer is is obvious or you know it's  you want to reduce burnout. You introduce these AI [13:24] solutions. Is it the sort of thing for which you  think okay we need to measure what burnout looks [13:29] like or the sources of burnout or the cases in  which people have walked away because of burnout [13:34] and then control for the variables and see if the  AI is doing the right sort of thing to produce [13:38] burnout or is it look we live this stuff every  day. We can see that it's reducing burnout. We [13:43] don't need, you know, some complex metric regime.  We just we see it every day that it's doing, [13:48] you know, it's doing good. You'd like to tackle  that first and then I would go that's a difficult [13:54] question. So I let you go first. But how do  you measure burnout? So the extent to which [13:59] AI mitigates burnout. Yeah, I mean certainly, you  know, when we when we look when we for instance [14:08] did ambient uh scribing um you know we measured  it kind of at the workforce level. So we you [14:15] know we would measure no completion time and  we would measure uh you know we would actually [14:20] get um feedback from the providers. Uh um then at  the workforce level you look at things more like [14:30] provider engagement scores, you look at uh survey  results, you look at turnover rates, you look at [14:37] adoption utilization. So if this tool is helpful  and you know I in the case of ambient scribing I [14:48] certainly think it is and a great example is when  we rolled it out to our uh physicians the first [14:55] feedback we got was from the nurses. What about  us? You know that this this this burnout you know [15:03] our burnout is real as well. And so we're now  piloting actually a a ambient nurse an ambient [15:11] scribing technology uh for instance for flowheets  and for vital signs and for other nursing tasks. [15:17] So yeah, I would agree to all of that and that yes  there have been I think 2025 there were two large [15:24] studies one was from uh from Mass General and I  think another one was from Yale but multic-enter [15:31] study they showed with validated burnout measures  um for mental exhaustion um uh score emotional [15:40] exhaustion score they substantially decreased  but they were not the studies were done about [15:44] 30 or 40 days pre and post uh post uh ambient  scribe implementation. So not long-term studies [15:51] but there is some evidence with validated tools  about burnout. One of the other things uh I would [15:57] like to mention is that and we have not started  this project. We have created this technology. I [16:02] was mentioning about case management utilization  management. One of our outcome metrics would be uh [16:08] the turnover rate. Did the turnover rate decrease  on those those particular departments where we [16:13] are seeing significant turnover and and and you  mentioned nursing nursing first year turnover is [16:18] a huge problem for many many healthcare systems.  So that could be another great metric to measure [16:25] as before about the initiatives projects programs  I had two that are very near and dear because I [16:32] think really highlight the the power of AI. One of  them being associated to nurses. So we're working [16:38] with a client to build a solution for cleaning  automation of the reporting and generation the [16:45] how the food highly stressful falls onto the  nurse making sure that every single minute [16:50] and every single action has been documented. So  picking up AI, leveraging listening to what is it, [16:57] parsing out who said what to attribute the the  activities and the actions and taking that burden [17:02] off of the nurse so that instead of having to pay  attention to everything that's going on as well as [17:07] picking up names, happening events, and writing it  all down becomes then less of a cognitive burden [17:15] to the to them to the nurse uh team to then  revisit and revise what the AI is transcribing [17:20] in real time to to fill out the the form. cars. Uh  the other one that I that is very near and dear to [17:26] me is um come from Brazil. So access to to care is  challenging right and we're working with another [17:34] hospital system to use AI to facilitate access. So  do you really need to go cross out to get to that [17:43] particular clinic? Is there another doctor that I  can find that will treat you for those symptoms? [17:48] Do you have insurance? Do you not have insurance?  at this particular client is is in Brazil. So, [17:53] we have a national healthcare as well as  a private market. So, can I take you on a [17:59] private do you have a private insurance? Can I  get the pre-authorization done? Can I get all [18:03] of that done? So, when you're coming in because  you have kids at home that you need to baby leave [18:07] with somebody that to get care, we're taking you  out of your house for the minimal amount of time [18:13] as efficient as possible. reducing the time of  the nurse the front desk back office afterwards [18:20] to do all of the paperwork necessary for your  for your care to be focus focusing people on [18:26] their care. So to me those are two near and dear  projects like that helps with the burnout and the [18:32] change but highlights the what if the art of  the possible with AI think the workflow was [18:38] there has been there we haven't changed how do we  move from a like the artificial intelligence but [18:44] really an augmented intelligence how do we get  and elevate what the art of the possible is with [18:49] the resources that is need I love that and um  as I think about three very different use cases [18:57] right, of how AI is working. Um, I'd love to  hear about outcomes. What are you seeing? So, [19:04] you talked about 47,000 users now. Uh, did you  have a goal? Like, what was the goal? Was it 85%, [19:14] 50%, what was the goal? So, think about it. Uh,  for each of your use cases, what were the outcomes [19:21] and what are you seeing thus far? So, how did the  47,000 fit in your in what you were expecting? Uh, [19:29] more than we expected. Uh, but now, of course, we  want, you know, we want more. Of course, right. [19:35] Um, you know, interestingly, you would think  that, you know, 18 million model interactions [19:42] in half a year would cost a small fortune, but it  actually doesn't. um you know if you use some of [19:48] the um the you know the licensing models that it's  actually not that expensive versus some of the [19:54] other opportunities. So we're actually trying to  scale um and we are scaling by adding um uh agents [20:03] by adding tasks by being adding the ability to  share tasks uh publish agents. Uh so we're we're [20:13] actually trying to take that um kind of that  grassroots um interest and and really put it [20:21] out there. We're actually talking about I'm sure  most people are familiar with mandatory education. [20:26] Every year you have to take your compliance  education and everything. So, we're uh discussing [20:31] actually putting AI literacy uh into one of those  compliance modules that all employees have to take [20:38] because we're concerned that now that we've put  this tool in people's hands, will they know about [20:45] bias? Will they know to doublech checkck the AI's  work or triple check the AI's work? So, you're [20:52] right. You have to be careful what you what you  wish for. So those outcomes it sounds like are not [20:59] just utilization right but it's also ensuring that  you're still having the quality of care uh without [21:06] those biases etc. Right. And and we have done a  lot of um qualitative and quantitative research [21:13] around the usage as well. So we know what people  are doing. we're getting, you know, response back [21:19] about the, you know, what the experience is like  and whether people actually feel that the that the [21:24] um use cases are are useful. That's great. What  about the two of you? Um, similar kind of an [21:30] experience. Now, um, if I go back very quickly to  ambient AI, I'm love to get your perspective. The [21:36] the amount of time saving is not huge. It's  usually in minutes at the end of the day, [21:41] 90 minutes or 20 minutes, similar to what has  been there in published studies. But the I think [21:48] the greater benefit to that is the decrease in  the cognitive load, decrease in the burnout of [21:53] the providers feel that okay something has been  taken care of. So that's that is one of the we [21:58] would say an outcome metric although it cannot be  very nicely expressed in Excel sheet because 90 [22:04] minutes does not get you another patient encounter  which is which which goes into the Excel sheet. So [22:10] and should it sorry what should you do with that  90 minutes? Should you make more RV use or should [22:17] you go home and have dinner? I would go home and  have dinner, but yes, I would go home and have [22:23] dinner. You're actually spending good time with  your family because that's a lot of what happened, [22:26] right? Because that's what we're talking about  decreasing burnout. But I think it's an excellent [22:30] point is if you give a provider another 90  minutes, there are those who will say, well, [22:36] that's another, you know, that's another three  patient visits, right? Depends on how fast you [22:40] are. And that is okay. That is okay. I think it if  the agency is still with the provider saying that [22:46] I want to see another patient in the 19 minutes.  Yes, that is still okay because agency loss of [22:50] agency is probably a bigger factor of burnout.  So I just wanted to respond to that but apart [22:54] from that similar to you we have some process  metrics we are also seeing in terms of our general [22:59] workforce how many agents are being created we  create our agents in in in our enterprise system [23:05] and in terms of other outcome metrics they are not  very related to burnout but for the AI models that [23:12] we have implemented we have shown that there has  been a decrease in the ED boarding time how many [23:17] hours patients are boarding in the ED in decrease  in the length of stay for not all of our hospitals [23:23] some of our hospitals in which the adoption  has been appropriate. So we measure those [23:26] outcome measures too. Let's uh moving from the two  cases I mentioned because those are in flight so [23:33] we still don't we don't have the metrics of the  results. We're seeing very positive reassurance [23:38] that we're going the right track but we haven't  we don't have the measurement yet. Um but another [23:43] implementation we've done for primary care and  aggregating information from disparate systems. [23:48] They're not on one of the big EHRs. Um there  the one of the clients we have smaller clinics, [23:55] smaller network of clinics. We've seen um  like about 900 hours a month being saved on [24:01] a on a 20 people team of documentation being done  after hours aggregating prescription aggregating [24:10] uh phone calls uh replies emails and all of that  that were being done outside of the of the hours. [24:18] I think I have something of a maybe it's a  strategic question something like that. So [24:22] two two kinds of things that I have in the back  of my head or actually I suppose at the front one [24:27] is you we talk about augmented intelligence if I  ask AI to do something let's say draftto whatever [24:33] it is I find it way more taxing to edit to sort of  read the thing digest it edit it change it that's [24:40] not what I want that's not how I than if I just  took the blank page and just had at it myself. So [24:46] that's that's sort of one sort of comment. Second  thing, I'm not a clinician, but I have watched two [24:52] seasons of the pit. So I'm I'm a bit of an expert.  I'm almost an expert. Almost very realistic. And [24:58] one one thing that they portrayed which seemed  at least realistic and obviously you can push [25:02] back. It's it's TV is they seemed overburdened  by having to learn new technologies. And so look, [25:10] sometimes you're going to introduce AI and it's  not a brain drain and sometimes it is a brain [25:14] drain. That's that no that's I think that's that's  the sort of the setup. The more strategic question [25:19] is something like okay so how are you supposed to  go about figuring out what are the appropriate use [25:24] cases such that choose one that's it's it is a  genuine cognitive offloading in a in a good way [25:30] or actually no it's actually self- undermining  because now they're doing more work trying to [25:35] vet what the AI says. So are there questions that  you ask as you think about what's the appropriate [25:39] use case in that context? Are there criteria?  So how do you think about it? Can we use we try [25:46] to use it as a a mathematical formula, right?  I'm on the opposite side again. So I'm not the [25:51] physician. On the engineering side, we try to  measure it with three variables. Time spent and [25:58] level of hatred that people have for the task  of data readiness and the the blast radius. [26:05] So if I make a mistake here, what is the impact  that's going to cause? So you want to keep it low [26:10] on the on the blast radius, you want to have the  data and where you you gain that and where you [26:16] prioritize is where people spend the most time  and they hit the most. So that's that's where [26:21] we've been focusing on getting those first because  once you get the first once you get people to buy [26:26] in then the follow the next ones will follow.  So that sounds like a sort of you know a pretty [26:31] clear call to action. first thing or among the  first things you should do is survey the people [26:35] and ask not not what do you want but what do you  hate? Yeah. First of all in the pit they're overt [26:42] taxed by everything just so you know but but but  rightfully so. Um I I think the model you know [26:48] when we were preparing the the the model that you  just said is very similar to the rice model that [26:53] I was thinking of where where you think about you  know how how badly do you need something? How bad [27:00] is the problem? Um and uh I I think that there  there are mathematical equations for change. Um [27:07] you think about you know how how how burning  is the platform, how badly is change needed, [27:13] um how much resistance is there? And of  course all that goes over effort, you know, [27:18] how much effort is it going to take to be able  to get that change. Um, but I I think that any [27:25] change management has to be led with empathy and  you have to realize that change is hard and and [27:33] there's no way not to make it hard. We're we're  rolling out a an unnamed epically large EMR, [27:41] you know, to right well it's it's epically large  to um to you know to to 110,000 employees and [27:49] it's it that change is hard but if the outcome is  worth it and the process is done properly and it's [27:56] led with empathy I think that that it can work. I  couldn't agree more. So uh I would say that coming [28:03] back to the topic of burnout who has heard the  phrase that physicians are resistant to change I [28:11] mean I have heard it number of times but I think  that's a fallacy if you look at the COVID period [28:16] how fast did we change we changed at every turn  sometimes we were incorrect but we were always [28:21] following the guidelines we changed at every  turn so give us a proper reason for change right [28:26] um so that then then people will change there are  as you mentioned there mathematical models to see [28:32] that okay how which change should be implemented  which AI tool in this case should be implemented [28:37] there are rule of thumbs there are rules of thumb  too for example for any clinician if this tool is [28:44] allowing the clinician to practice at the top of  their license be at the bedside do the human work [28:49] of like talking with the patient working with the  family then that is a rule of thumb that we follow [28:55] um and then of course we do pre and post survey  not not similar to what you I mentioned that [29:01] uh controlling for all variables we don't for  general operational implementation not not [29:07] talking about research we don't go to that extent  but we do some pre and post survey and measure [29:11] some process metrics I don't think clinicians are  resistant to change per se they're resistant to [29:18] extra clicks they're resistant to extra complexity  they're resistant to extra risk exactly and the [29:24] problem is that so much change comes with that  as part and parcel and that's Why part of the [29:30] equation is also receptivity and timing because  those things really matter particularly when [29:39] uh you know there are different cadences across  any given year and any given organization and you [29:46] have to be very careful about when you implement  a change to ensure that the organization has [29:54] the capacity to to manage it. you know, I'm  curious about the the changes that you all [30:00] uh have been through. Um who was involved at the  onset like was it the physicians, engineering, [30:13] uh uh the actual users who was engaged in the  process? Let's let's push on that a little bit. [30:23] say I'll put two cases where I've seen su success  and failure. Every failure change, every failure [30:32] imple failed implementation we've had was pushed  by somebody else's or some different organization [30:38] into nurses, clinicians, providers, back office,  anybody. Somebody said, "I know your pain. I know [30:46] how I'm going to solve it. Here, here it is. Start  using it." Didn't ask. didn't get the context, [30:53] didn't immerse, there's no empathy. Where we saw  the most success is getting people from day one, [31:00] not not an afterthought, not like I should consult  or I'm going to involve you to test come to the [31:06] table, co-create with us what the solution  is going to be because that brings people on [31:11] board and they're part of the solution. At that  point, the the minds the the the the mindset [31:17] changes from somebody's pushing something to me.  I'm a resist. I'm resisting. I'm actually part [31:24] of this. I'm embracing everybody. We're going to  march and push this together. And you're speaking [31:28] the language. Like if you have a physician, I  don't again I'm continue to make fun of me. I'm [31:35] not the doctor. I been around the systems and all  that. I understand the the language, but I'm not a [31:41] doctor. I don't have the credentials. I can't talk  until I'm blue in the face. If you don't want to, [31:46] any one of you want to adopt it, there's no  argument I can use. If you want to adopt, you [31:52] were skeptical. You will listen to him. You not  listen to me because you're you're part of that [31:57] same construct. You you have the same background.  You have that he understand or you understand his [32:04] thinking process. I don't. It's different. So, it  removes some of the barriers further down the road [32:10] too. It's not just the initial uh uh champions,  but also down the road when you're scaling your [32:16] solution, people seeing that being vetted by  others become a much easier pill to swallow than [32:23] uh completely out of the blue. I I think it's  about having a coalition because I I've seen many [32:29] physician-led initiatives that fall flat because  they didn't involve nursing and vice versa. [32:36] So I think really and and we should even take it  out of the hospital context but having a coalition [32:42] of people so that you have all the voices at  the table all the stakeholders and maybe not [32:47] all the stakeholders have the same belief that  the problem needs to be solved but at least if [32:52] they're there to have their voices heard you'll  have a much higher chance of success. I've seen [32:58] this working like I I heard this many many years  throughout my career is legal is where ideas go [33:04] to die. Every time you involve legal, legal will  put the the lid on it and it's going to be over. [33:10] I've had many very successful implementations  where we brought legal not as a validator, [33:15] not as a rubber stamp. They were sitting on day  one. They we made a point getting them early, [33:20] getting them all of their concerns early so  that we design around that and they became huge [33:26] supporters. They actually became promoters.  They were the first ones say this is going [33:30] to be great. This is going to help us. this is  going to actually uh um re potentially reduce [33:35] compliance risk and all of this. They became  the promoters, not the people to say, "Nope, [33:40] put the kibash on this and kill it." This is  a real drop the mic moment. This cocreation, [33:46] this coalition to drive the uh transformation um  you know, the more integrated the thinking is, [33:54] the better. So, I don't want to step on your point  because I'm sure it's going to be a great share. [34:01] So I completely for I completely agree with  the is this working? Yes. Yes. Completely agree [34:08] with the coal. So yeah, now it's working. Okay.  Completely agree with like all getting all the [34:13] voices and then informatics projects especially  in the healthcare setting is 90% about people 10% [34:19] about technology. What suggest to make a counter  not a counter point but one word of caution and [34:26] I'm sure everybody has experienced this. When  we get a coalition sometimes we are not clear [34:31] who is making the ultimate decision and who is  responsible for implementing the project and [34:36] make sure it's implemented in the correct way and  who is responsible for failure. I think that's the [34:41] flip side when we build wide coalition we have  to absolutely keep in mind. So we actually use [34:46] the AC matrix responsible who's that accountable  consultant and this matrix for all of our project [34:53] or most of our projects to make sure that even  when we are getting feedback from everybody and [34:58] taking everybody into consideration we have a  clear line. Very good point. So, we've talked [35:04] about sort of what you've done up until now and  we have a sort of some understanding of what's, [35:10] you know, you're trying to scale um current  efforts. I wonder if there's anything that you [35:14] have in mind. You haven't really started it yet,  but you have this sort of project in mind. Here's [35:19] here's something that we could really do that  would make a big difference. Is there that sort of [35:24] AI use case that you see coming down the line or  that you want to come down the line, I don't know, [35:28] six months from now, maybe even 12 months from  now. And if there is that sort of project that you [35:32] have in mind, are there any particular blockers  that you see that are particularly troublesome? [35:41] I think robots are going to take over the world.  Yeah, for sure. That's going to happen. Robots are [35:45] going to take over the world. Yeah. Yeah. Yeah.  Um, no. I I think we keep talking about human [35:50] in the loop. Human in the loop. And it's almost as  if you're not allowed to say, "Well, what if there [35:56] wasn't a human in the loop?" But I think we're  we're getting awfully close to the day where, [36:01] you know, they don't need us anymore. Um, and so I  I do I do worry about that. But at the same time, [36:09] agentic AI is so close to not needing a human  in the loop. And there are ways that we can take [36:17] advantage of that in a in a in a beneficial way  um with guard rails, with governance, you know, [36:23] with ethics. Um but I I feel like we're playing  catch-up and we really need to get in front of [36:30] instead of instead of chasing um this technology.  Yeah. Um something we are excited about and I [36:40] think that's coming down the road is uh I mean  if you go back a few years AI was mostly doing [36:47] prediction. This is the patient at risk. Now with  generative AI and other language tools we are now [36:52] saying because this patient is a risk risk these  are the things you should do and then we have [36:57] agents which will say that okay now I'm helping  you to do this right where the physician or the [37:02] clinician is still important is to elicit the  evaluation of the patient um touching the patient [37:09] asking what's going on and and taking all of that  information and and then putting it into a note or [37:16] whatever that then the AI can understand reason  I'm saying that so and physicians I mean AI has [37:23] already surpassed us in terms of the memory  that it can retain in terms of how much data [37:28] it can go through and pattern recognition but the  evaluation part it is still lacking we don't have [37:32] an AI that can take all of that but with evolving  technology in in in processing of pictures and [37:41] videos in in in room camera technology um we are  able to do much more now we are able to get vital [37:48] science we are able to actually we are working  on something that will try to predict that the [37:52] patient is likely to get agitated in the next  few hours while maintaining the patient privacy [37:58] I think that will be a gamecher because then the  sensory capture part is also taken over by AI my [38:07] my vision my hope is that just I see most on the  panel are wearables multiple wearables smart rings [38:15] and all of it It's getting to see and getting  to look at the patient not just on the side [38:23] of care but across your reality your dayto-day.  You can talk about white coat syndrome. You can [38:30] talk about everything else but you don't really  look at the patient across the stress of the day. [38:36] The physician understanding that the patient  has a sick kid at home, sick parent at home [38:41] that they're caring for that's influencing their  blood pressure and all of that. and you can see [38:46] what the moments are and start that correlation.  That's where I think AI unlocks or will unlock a [38:53] lot of the ability to see and care for people  as people not as a as a picture in time. Um, [39:02] but to get there to your to your question, I think  the the big block is the interoperability data. [39:08] In order for you to be able to care for it, you  need to get data from all of the different devices [39:13] from the different exams, different systems,  different EHRs, the ones of epic proportion or [39:19] smaller proportion and be able to coordinate and  analyze that and provide that back, not just hoard [39:24] the data, but actually generate insights and get  that back to the physician because ultimately the [39:29] decision is on the physician side. So, we're Oh,  sorry. Just want it sounds like there might be a [39:36] little bit of tension because it sounds like some  of you think the physician at the end of the day [39:40] has got to make the call and some of you think  the physicians are going to get left behind and [39:45] they're going to serve more as a system to the  AI as opposed to the final decision maker. So, [39:50] this goes back to the is the AI going to put me  out of a job question and or at least drastically [39:56] change it. And what I always say is, you know,  when people come to me and ask me that question, [40:00] I say the AI is not going to put you out of a job,  but someone who knows how to work with AI is going [40:07] to put you out of a job. And so I don't think that  the AI is is a risk to physician in and of itself, [40:13] but I think that learning to consider yourself  a partner to AI is I it may sound crazy today, [40:22] but let's watch this tape back uh in a few years.  uh because I I don't think it's that crazy. So [40:28] again just just to carry this conversation forward  I think for many physician specialties the only [40:36] thing that is that is going to stand in the way  of completely taking over AI completely taking [40:40] over are regulations. Um in some physician  specialties as we know in radiology there [40:46] are number of studies that they have become  smarter. AI has become smarter than physicians. [40:51] The only thing that is protecting them right now  or protecting us right now is regulations. Not in [40:55] all disciplines. Don't envision that in the next 5  years complete AI robots will be doing complicated [41:01] surgeries. Uh but some disciplines, yes, it's  is the regulation that is that's the barrier. [41:08] So, imagine this. We're at a white coat  ceremony and uh you know these brighteyed [41:16] bushytailed clinicians are about to go  in and start their clinicals, right? [41:23] Um what would you say to them particularly in  this new era knowing what you know and how you [41:32] came up and learned how to be clinicians and then  what's available to them now what would you all [41:40] what advice would you give them today you start  uh one advice would be learn AI that's absolutely [41:47] 100% learn AI how to use it how to interpret it  even sometimes how to make it because unless you [41:53] know a little bit of the nuts and bolts, you  wouldn't know why hallucinations occur or how [41:58] does it become a psychopan to you and things like  that. So we are actually starting our our program [42:05] in our medical school now where they will have an  AI for about 2 months of train. It's not gone live [42:12] yet but that's our plan. They still have a little  bit but that's going to be part of the curriculum [42:16] and we are going to continue through residency as  well as as as as uh train our faculty and similar [42:22] to your organization we are also doing a lot in  terms of our workforce literacy non nonclinical [42:28] workforce literacy but one answer would be learn  about AI what would you add to that I'd say not [42:35] just learn but work with it it's like I I said it  before it's not an artificial intelligence it's an [42:42] augmented intelligence Use it to your benefit.  Use it to help with your blind spots. Use the [42:49] um people use AI today as a glorified Google  search, right? You put you paste a halfbaked [42:58] query into the into the chat and you expect the  the AI the model to understand parse it and give [43:04] you a complete answer. Learn to ask better  questions to get better results to criticize [43:11] better to continue iterating the loop. Don't be  just a in gesture or or don't just ingest whatever [43:20] is coming out. Criticize, ask better questions  and continue iterating. Yeah. Yeah. So, at my [43:29] white coat ceremony, um the dean said something to  us that I think is still very applicable. He said [43:35] when you put this coat on your life will never  be the same. And I think it's true today just for [43:40] different reasons. For different reasons. I love  that. Well, I will just share a framework around [43:47] uh burnout uh from a human resources perspective  and then afterwards we'll talk about how HR was [43:54] in involved in all these initiatives. Um and then  give uh my co-odderator the the final words here. [44:02] uh but when you think about u uh capacity  there's a human side the organization side and [44:09] then the financial side so the human side is the  burnout that's where the erosion happens and then [44:17] it begins to affect the organization because of  the loss of engagement um and all of the factors [44:25] associated with that with potentially uh issues  with patient care and then ultim Ultimately, [44:32] it will go to the financial side uh where you  see increased turnover. Um individuals like [44:41] myself then have to go and spend about $500,000 to  get a new clinician in and ultimately it it lends [44:51] itself to revenue and other financial metrics.  So the burnout is real and so the more we can [44:57] use AI to support and help that certainly uh  certainly the better because the clinician [45:03] experience equates to the patient experience and  uh please close us out. I don't have I'm not going [45:10] to you know pratt on we we're out of time but  I do just want to thank you know our panelists [45:14] and my co-odderator and thank you everyone  for coming. Thank you. Thank you. Thank you.