Hospitals Cut Claims Cost 98% with AI Agents, Hackensack Executive Says

Hackensack高管称医院利用AI Agent将理赔成本削减98% 自主AI工作流将拒赔申诉成本从35美元降至60美分，而缺乏引导的聊天界面未能带来运营价值。 与会嘉宾探讨了人工智能在医疗健康领域的当前影响及未来五年预测，重点强调了从单纯的信息检索向诸如收入周期管理等自动化Agent工作流的转变。核心议题包括通过结构化的AI治理避免碎片化的零散项目、对照临床和运营指标衡量ROI，以及利用具备上下文感知能力的AI工具替代缺乏引导的在线搜索。此外，专家组还重点指出了围绕数据隐私的关键挑战、HIPAA安全港去标识化的风险、患者知情同意书的现代化，以及设计能无缝融入现有临床工作流的用户界面。 AI 正在通过释放临床医生面对电子病历（EMR）的时间来改善医患交流，同时利用对话式 AI 开展术后 30 天随访以填补人力短缺，但未经审查的生成式搜索建议仍存在严重临床风险。 (2:43) 医疗机构在投资 AI 时应避免盲目追求新奇技术的“百花齐放”陷阱，而需建立明确核心支柱，并以降低 30 天再入院率、缩短住院时间及减少运营成本等硬性指标评估投资回报率（ROI）。 (8:38) 在收入周期管理（RCM）等后台任务中，结合生成式与规则模型的智能体工作流（Agentic Workflows）成功将单笔拒赔申诉处理成本从 35 美元降至 0.60 美元。 (17:05) 业内专家呼吁摒弃流于形式的“免责式”长篇患者知情同意书，并指出 HIPAA 传统的“安全港”去标识化方法在现代算力下容易被逆向破解，企业应采用专家判定法（Expert Determination）以兼顾隐私与数据实用性。 (29:25) 针对 AI 工具界面设计，专家建议将 ChatGPT 式的空白对话框比作“高果糖玉米糖浆”，主张将预设提示词封装在单键按钮中，以避免自由输入带来的低效与推广阻碍。 (41:08)

Deploying healthcare AI effectively requires shifting focus from novelty to disciplined governance, targeting concrete operational bottlenecks, and protecting patient trust through transparent privacy standards and clinician-led implementation. Automating Administrative Burdens and Clinical Workflows (17:05) Panelists emphasize that healthcare AI delivers its most immediate value by reducing operational and administrative friction rather than replacing clinical decision-making. Sameer Sethi projects that 40% to 60% of routine analytical and operational healthcare tasks can eventually be handled by non-human agents. For example, Hackensack Meridian Health introduced an agentic workflow to appeal denied insurance claims—using generative AI to synthesize denial letters and rule-based models to retrieve missing modifiers from electronic health records (EHRs)—which dropped processing costs from roughly $35.00 to $0.60 per claim. Clinically, Dr. Margaret Lozovatsky advocates for "augmented intelligence" that liberates physicians from administrative reporting and manual narrative documentation. By automating note-taking and capturing incident data directly from EHR audit trails, ambient AI allows providers to re-engage in direct eye contact with patients. Sethi also notes that conversational AI agents now handle post-operative follow-up calls within 30 days of discharge, bridging staffing shortages without sacrificing touchpoints. Governance Frameworks and Hard ROI Metrics (8:38) Health systems face significant risk when adopting AI reactively without clear strategic priorities. Luiz Cieslak cautions against a "garden of a thousand flowers," where disconnected departmental experiments rely on superficial vanity metrics like login counts or active users. Cieslak argues initiatives must directly reduce lengths of stay, lower operational expenses, or decrease triage escalations to survive budget cuts, noting that tools disrupting established clinical workflows will simply be rejected by medical staff. To avoid chasing shiny objects, Hackensack Meridian Health consolidated its AI roadmap into six board-approved pillars, evaluating post-discharge AI strictly by whether it lowers 30-day hospital readmissions. Similarly, the American Medical Association established two resources: the AI Governance Toolkit to help organizations evaluate use cases and integrate tools safely, and the Return on Health Framework, which measures both tangible financial offsets and broader quintuple-aim benefits such as clinician well-being and care quality. Transitioning from Unguided Search to Contextual Care (21:28) Panelists anticipate consumer-facing AI will replace chaotic internet health searches, which frequently generate false alarms and heighten patient anxiety. Dr. Lozovatsky highlights the real dangers of unguided search summaries, citing an instance where a live AI query advised drinking urine to treat a urinary tract infection. She and Cieslak explain that open-ended web queries often lead patients to suspect extreme conditions like brain tumors, prompting unnecessary clinical visits. Cieslak expects contextual AI to largely replace generic search behavior by 2030. Rather than providing alarming possibilities, context-aware tools can combine personal health history with localized micro-environment data (such as active regional flu outbreaks) to recommend appropriate over-the-counter interventions or direct patients to suitable care levels. However, Dr. Lozovatsky stresses that patient-facing tools must always maintain explicit escalation paths to human clinicians whenever clinical judgment is required. Privacy Vulnerabilities and Flawed Patient Consent (29:25) Data governance remains a critical vulnerability in healthcare AI deployment. Adam Kamor cautions software builders to formulate privacy strategies before training models, warning that embedding protected health information (PHI) directly into model weights can invalidate months of engineering work. Sethi warns that health systems often enter risky agreements allowing technology vendors to train commercial models on patient data, exposing health systems to public backlash if vendors suffer breaches. Sethi also points out that standard HIPAA "Safe Harbor" de-identification—which removes 18 specific identifiers—is outdated against modern computational methods capable of re-identifying records in milliseconds, making the underutilized "Expert Determination" statistical standard essential. Concurrently, Cieslak and Lozovatsky critique the current state of patient consent, comparing dense digital terms to unread software agreements. Lozovatsky notes that true de-identification is largely impossible for rare medical conditions, urging health systems to use clinic wait times to educate patients realistically about data security. While the panel anticipates substantial efficiency gains from agentic workflows and ambient documentation, full diagnostic autonomy remains constrained by regulatory boundaries, technical limitations, and the ongoing necessity of human clinical judgment.

行业专家指出，医疗人工智能的发展重心正从概念探索转向解决临床疲劳、优化患者体验以及量化实际投资回报；尽管生成式 AI 在非临床行政流程中展现出显著的降本增效潜力，但在临床决策与患者直接交互层面，仍必须依赖严格的治理框架、隐私合规保障以及人机协同边界。 临床减负与患者交互的即时转变 (2:43) Margaret Lozovatsky 博士指出，AI 极大提升了医疗信息的获取便利性，但未经验证的输出潜藏严重的临床风险；例如在线 AI 检索工具曾给出饮尿治疗尿路感染（UTI）的离诞建议，这表明面向患者的 AI 必须进行专业引导。在积极层面，环境式 AI 工具能够使医生摆脱电子病历（EMR）屏幕的束缚，重新恢复诊疗过程中与患者的面对面眼神交流。Hackensack Meridian Health 首席 AI 与洞察官 Sameer Sethi 补充指出，面对术后 30 天随访的人力短缺，该机构已通过对话式语音 AI 定期外呼出院患者，有效弥补了人工随访的盲区。 CI&T 的 Luiz Cieslak 表示，全美约有三分之一的受众使用生物监测设备，依赖可穿戴设备的数据个性化正在加速。CI&T 与癌症机构合作开发的个性化方案结合了癌症亚型、合并症及人口学因素，直接提升了患者生存率与长期照护质量。Tonic.ai 联合创始人 Adam Kamor 补充强调，大型保险机构与 AI 初创企业对后台自动化的迫切需求，催生了医疗行业对高质量、符合隐私规范数据的巨大需求。 投资回报评估与治理框架的建立 (8:38) 面对医疗系统盲目跟风引入新技术的现象，Sameer Sethi 警告称，缺乏系统性战略会导致机构陷入无休止追逐热点的困境。Hackensack 经董事会确立了六大核心投资支柱，并要求所有 AI 项目必须与实际临床指标绑定，例如其出院随访系统严格以能否降低术后 30 天再入院率作为衡量投资回报的准则。Luiz Cieslak 则提醒应规避“百花齐放的陷阱”，各部门各自推进的孤立项目往往依赖日活跃用户或登录次数等虚荣指标，若无法切实缩短住院时间（LOS）、降低运营成本、减少护理升级或提升接诊吞吐量，最终必遭撤资。 针对落地挑战，美国医学会（AMA）推出了两项指导框架：其一是指导医疗机构组建多学科团队并安全融入临床工作流的“AI 治理工具包”；其二是围绕“五重目标”（quintuple aim）建立的“健康回报框架”，用以评估改善医患体验与医疗质量带来的综合收益。Lozovatsky 博士与 Cieslak 均强调，工作流契合度是决定工具存亡的关键非财务指标，任何打乱既有临床流程的产品都难以获得医生的实际采纳。 代理工作流在非临床运营中的成本突破 (17:05) 在技术演进路径上，Adam Kamor 区分了技术能力与监管合规的差异；他回忆 25 年前牙科已运用机器学习辅助诊断，未来五年基础模型将稳步演进并辅助行政与诊断，但在严格监管下完全自主诊断仍将受限。Sameer Sethi 则对代理工作流持乐观态度，他预测医疗系统中 40% 到 60% 的日常分析与运营任务可交付机器代理自主处理，系统可自主过滤哪些信号交由机器执行、哪些需人类介入，从而解决过去因人工交接造成的效率瓶颈。 Sethi 以 Hackensack 的收入周期管理（RCM）为例，展示了非临床流程的自动化收益。传统人工处理保险拒赔申诉需耗费人力核对代码并调取 Epic 系统记录，单笔申诉成本约为 35 美元；通过结合生成式 AI 与规则模型，系统能自动解析拒赔信、调取预授权信息并起草申诉文本，经人工最终审核发出后，单笔处理成本骤降至 0.60 美元，且能实现全天候无间断运转。Lozovatsky 博士对此倡导“增强智能”理念，认为医生的直觉与临床判断不可替代，但应通过技术解放医生，消除繁杂的叙述性病历书写以及 37 项安全事故表格的重复录入，依托后台数据实现自动化呈报。 隐私合规短板与知情同意界面的重构 (29:25) 在数据合规与隐私领域，Sameer Sethi 警示医疗机构必须严加区分临床诊疗同意与用于商业产品开发的数据授权。他批评部分技术供应商在推广环境录音软件等工具时施压获取不受限的数据权利，一旦第三方发生数据泄露，医疗机构将承受严重的声誉打击。针对脱敏技术，Sethi 指出传统 HIPAA “安全港”标准（Safe Harbor，剔除 18 项个人标识符）在现代算力下可在数毫秒内被重新识别，因此应采用更科学的“专家测定法”（Expert Determination）评估统计风险；然而 Kamor 与 Sethi 发现，多数商业供应商对这一合规路径的细节缺乏基本了解。 聚焦于交互体验，Luiz Cieslak 批评当前医疗知情同意书已退化为冗长晦涩的法律免责文本，用户往往直接滑动到底部点击同意，未能实现真正的知情透明。在人机交互方面，Sethi 将缺乏限制的自由聊天输入框比作“高果糖玉米糖浆”，看似便捷但容易产出低质结果；他建议软件设计团队对用户进行系统性提示词工程培训，或者将预设提示词封装于单键按钮之后，以防无指导的自由输入破坏系统落地的预期回报。 该讨论表明，医疗 AI 的核心定位应为支持临床与运营决策的“增强智能”而非独立替代者；其长期价值的释放，取决于行业能否在保护患者隐私、重塑知情同意体验以及构建贴合临床工作流的严谨治理体系之间取得平衡。

## Background 本视频由行业平台 Ai4 于 2025 年 8 月 28 日发布在 YouTube 上。视频围绕“利用 AI 端到端优化患者体验”的主题展开，汇集了来自投资机构、美国医学会（AMA）、大型医疗系统与科技企业的行业领袖，共同探讨 AI 在个性化诊疗、虚拟助手、预后预测分析、跨科室护理协调以及提升临床与财务效益中的落地实践。 ## TL;DR 圆桌嘉宾围绕医疗 AI 如何贯穿诊前交互、院内诊疗到院后管理的全流程展开了深度交锋，强调必须拒绝盲目跟风，以降低 30 天再入院率和缩减运营成本等硬性指标衡量 ROI，例如在拒赔申诉中利用智能体将单笔成本从 35 美元压缩至 60 美分。讨论重点剖析了美国医学会的“健康回报框架”与环境智能对医患关系的重塑，并尖锐揭示了供应商数据霸王条款、HIPAA 安全港失效与知情同意形式化等合规痛点。最后，专家们以“高果糖玉米糖浆”为喻解构生成式 AI 迷思，建议通过按钮化封装与提示词工程推动精准落地。 ## Atmosphere & style 整场圆桌讨论节奏明快、专业务实且不乏犀利的幽默感，散发着资深从业者历经行业热潮后的清醒与厚重。主持人的敏锐穿引与五位横跨医疗、科技、投资领域的专家形成高密度的观点共振，席间既有分享离谱诊断案例的诙谐自嘲，也有直击数据主权与商业博弈的针锋相对，在极具现场感与启发性的探讨中，将冰冷的工程指标与充满人情味的医疗伦理紧密交织在一起。 ## Key takeaways - [02:43](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=163s) 环境智能（Ambient AI）通过自动化病历记录大幅减轻医护人员行政负担，让医生能够从电脑屏幕前解放，重塑与患者的面对面眼神交流与人情味。 - [02:43](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=163s) 公开消费级 AI 在医疗自查中存在给出荒谬危险建议的严重隐患，凸显了在患者体验中引入临床级 AI 监管与专业审查的必要性。 - [08:38](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=518s) AI 既不免费也绝不便宜，医疗机构应确立清晰的战略聚焦领域，拒绝盲目跟风追逐热点，并在启动前制定严格的投入产出账。 - [10:35](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=635s) 医疗 AI 项目必须与降低 30 天再入院率、缩短平均住院日等硬性业务与财务损益（P&L）指标挂钩，仅关注使用率或活跃度的实验难以持续。 - [12:53](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=773s) 美国医学会（AMA）推出治理工具包与“健康回报”框架，将医疗质量提升、患者与医护体验等无形价值系统性折算为财务效益以评估高昂的 AI 成本。 - [15:11](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=911s) 未来五年 AI 将从模型副驾驶（Co-pilots）迈向自主 Agent，在无需直接临床干预的后台与运营场景中承接 40% 到 60% 的多步骤自动化流程。 - [19:08](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1148s) 在收入周期管理（RCM）等非致命运营场景中，自主 Agent 编排将单份拒赔申诉处理成本从 35 美元骤降至 60 美分，显著加速资金回笼。 - [21:02](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1262s) 结合实时上下文与情境数据的 AI 前端精准初筛，有望在未来终结因传统搜索引擎搜索症状引发的“谷歌医生（Dr. Google）”恐慌与医疗资源挤兑。 - [22:50](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1370s) AI 在临床中应定位为“增强智能（Augmented Intelligence）”，其核心是解放数据并重构行政流程，而复杂临床决策权与医患信任纽带仍须由医生把关。 - [25:12](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1512s) 医疗 AI 的透明披露是建立患者信任的基础，系统必须在明确告知 AI 介入的同时，配备随时无缝转交人类临床专家的健全升级机制。 - [29:25](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1765s) 知情同意需严格区分临床诊疗辅助与算法研发训练：将患者数据用于第三方商业模型开发必须获取显式授权（Explicit Consent）。 - [31:30](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1890s) 医疗系统正面临与技术供应商在患者数据主权上的严峻博弈，需警惕厂商通过向一线医生推销便捷工具而变相迫使医院在数据商用条款上妥协。 - [33:16](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1996s) 在强大算力下 HIPAA 传统的 Safe Harbor 脱敏方式极易被逆向重识别，机构与初创团队应更规范地运用专家裁定法（Expert Determination）以兼顾数据效用与合规。 - [35:00](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2100s) 冗长的法律免责条款已沦为形式主义，医疗机构应利用患者就诊候诊等碎片化闲暇时间，借助交互式数字化工具开展场景化数据安全透明教育。 - [41:27](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2487s) 通用大模型如同“高果糖玉米糖浆”，医疗产品落地应避免滥用无引导的空白对话框，而应通过系统性培训或将 Prompt 封装在功能按钮后来确保产出业务价值。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=4s) 开场介绍与圆桌嘉宾背景 Sound Ventures 医疗健康投资负责人 Juliette Bolea 担任本场圆桌的主持人。她在开场时幽默地询问大家在拉斯维加斯是否已经碰到了好运，并强调 AI 改善患者体验这个主题与在场每个人都息息相关，因为无论从事什么职业，每个人在生活中都扮演着“患者”的角色。随后，各位嘉宾依次介绍了各自的背景： 美国医学会（AMA）数字健康创新副总裁兼首席医疗信息官（CMIO）Margaret Lozovatsky 博士首先发言。她拥有儿科住院医生的临床背景，在一年半前加入 AMA 之前，曾在多家医疗系统中担任过 CMIO 和 CHIO 职务，积累了丰富的医疗数字化与技术落地经验。CI&T 医疗与生命科学高级副总裁 Luiz Cieslak 介绍称，CI&T 是一家专注于数字化互动系统（Systems of Engagement）的全球服务商，致力于为患者和医生打造提升生活质量的数字解决方案与体验。 新泽西州 Hackensack Meridian Health（拥有 18 家医院和近 500 个护理网点）的首席人工智能与洞察官（Chief AI & Insights Officer）Sameer Sethi 介绍了该医疗系统的庞大规模。合成数据平台 Tonic.ai 的联合创始人兼工程主管 Adam Kamor 则阐述了其公司的业务核心：确保患者的受保护健康信息（PHI）不会在应用测试开发阶段泄漏，更不会混入 AI 模型的训练权重中。主持人 Juliette Bolea 最后补充道，Sound Ventures 在医疗健康和 AI 交叉领域投资极为活跃，不仅投资了 Oscar Health、Capsule、Hims & Hers 等医疗创新企业，同时也是 OpenAI 和 Anthropic 的早期核心投资方。 ### [02:43](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=163s) AI 如何全面重塑患者体验与临床流程 针对主持人提出的“AI 如何重塑患者体验并带来实质改善”的问题，嘉宾们从信息获取、临床减负、个性化治疗、院后跟进及底层数据等多维度展开了剖析。 ```mermaid flowchart TD accTitle: AI in Patient Experience Workflow accDescription: 医疗健康体系中 AI 赋能患者体验与临床流程的关键环节 subgraph FrontEnd["患者交互与数据感知"] A["公开信息获取与患者自查"] B["可穿戴设备健康数据采集"] end subgraph Clinic["院内诊疗与临床赋能"] C["环境智能（Ambient AI）转录"] D["电子病历（EMR）海量数据整合"] E["个性化癌症治疗方案定制"] end subgraph PostCare["院后管理与底层支持"] F["对话式 AI 术后随访与关怀"] G["Tonic.ai PHI 隐私脱敏与合成数据支撑"] end A -.-> Clinic B --> E C --> D D --> E E --> F G -.-> FrontEnd G -.-> Clinic G -.-> PostCare ``` Margaret Lozovatsky 博士指出，当前患者获取健康信息的便利程度前所未有，但这既带来了便利也伴随着巨大风险。她分享了一个令人啼笑皆非的亲身体验：在准备采访时测试 AI 工具询问“如何处理尿路感染（UTI）”，AI 居然给出了“饮用尿液”的离谱回答。这直观反映出未经筛选的公开 AI 摘要存在严重误导的隐患。但从积极面来看，环境智能 AI（Ambient AI）等技术正在成为行业标配（table stakes），通过大幅减少医护人员的文书行政负担，让医生在接诊时重新把视线投向患者，找回至关重要的面对面交流时间。 Luiz Cieslak 将重点放在了个性化医疗上。他引用数据指出，目前约有三分之一的美国人佩戴至少一种可穿戴设备（他自己甚至戴了两部），这些多维数据为深度定制诊疗提供了可能。他分享了 CI&T 与一家癌症研究所合作的标志性成果：通过 AI 综合考量患者的具体癌症类型、身体状况与人口统计学特征，制定个性化治疗方案，显著提升了患者的长期生存率。 Sameer Sethi 则深入剖析了 Hackensack Meridian Health 从“痛点反推”的实践路径。他坦言，过去海量的患者数据堆积在电子病历（EMR）系统中，形成了难以消化的“信息汪洋”，导致医生在诊疗过程中不得不一直盯着屏幕打字，丧失了医患之间本应有的眼神交流。医疗系统利用 AI 辅助医生高效提炼信息，让诊疗回归人情味。此外，针对以往由于人力不足而无法覆盖的“患者出院后随访盲区”，他们上线了对话式 AI 与聊天机器人，主动致电出院或术后患者，跟进恢复情况并安排复诊，填补了院后护理断层的短板。 Adam Kamor 从底层技术基础设施视角补充说明，Tonic.ai 的客户从大型保险集团到小型医疗 AI 初创公司均有覆盖，AI 在减轻医护人员负荷和自动化后台流程方面发挥了巨大作用。由于大语言模型极度“饥渴于数据”（data-hungry），市场上对高质量、合规数据的需求呈现爆发式增长，这也推动了合成数据技术的飞速演进。 - **患者信息自查风险**：公开 AI 工具可能给出如“饮用尿液治疗 UTI”等荒谬且危险的医疗建议，凸显临床级 AI 监管的必要性。 - **可穿戴设备普及率**：三分之一的美国成年人至少佩戴一款健康可穿戴设备。 - **眼神交流的回归**：通过环境智能转录，医生得以从 EMR 录入中解放，重塑“医患面对面交流”的温度。 - **对话式 AI 院后管理**：利用智能外呼与 Chatbot 弥补人工护士不足，实现全天候的术后随访与关怀。 ### [08:38](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=518s) 拒绝追逐虚火：制定 AI 战略与衡量 ROI 当主持人将话题引向医疗机构如何衡量 AI 项目的投资回报率（ROI）及设定关键绩效指标（KPI）时，Sameer Sethi 强调了清晰战略在抵御行业盲目跟风中的决定性作用。 Sameer Sethi 直言，如果没有清晰的战略牵引，机构很容易被市场上“每天每周都在层出不穷的新鲜事物”牵着鼻子走，陷入被动追赶热点的混乱状态。他回忆道，Hackensack 在探索 AI 的第一年里也曾走过弯路，频繁在各种供应商和用例之间疲于奔命。为了摆脱这种内耗，他们将董事会成员召集在一起进行战略定调，最终确立了六大核心聚焦领域，并立下铁律：“凡是不属于这六大业务范畴的 AI 提案，一律不予立项”。 关于财务与运营回报，Sameer Sethi 明确指出：“AI 既不免费，也绝不便宜（AI is not free, and AI is not cheap）”。在启动任何具体项目之前，必须在事前就算清投入产出账。正因如此，他们在推行出院后患者 AI 自动化随访等具体用例时，都会将严谨的 ROI 测算作为立项前置条件。 ### [10:35](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=635s) 以硬性业务指标衡量 AI 投资回报（ROI） Sameer Sethi 强调，构建任何医疗 AI 用例之前，必须确立明确、无可争议的衡量指标，而最重要的核心就是**降低 30 天内再入院率**（readmissions）。他直言，很多 AI 功能看似酷炫、能提升机构声誉，导致整个领域充斥着噪音，但实际应用必须保持极致专注。例如构建电话回访聊天机器人（Chatbot），目标就是明确追踪：它拨打了多少出院患者的电话，以及在随访人群中究竟有多少人在 30 天内没有再次入院。 ```mermaid flowchart LR accTitle: AI Chatbot Readmission Measurement subgraph OUTREACH["触达阶段"] A["出院患者清单"] --> B["AI 聊天机器人<br/>自动化随访拨测"] end subgraph METRIC["效果验证"] B --> C["追踪 30 天状态"] C --> D{"是否发生<br/>30天内再入院?"} D -- "否" --> E["降低医疗成本<br/>直接改善财务底线"] D -- "是" --> F["复盘触发原因与干预失败点"] end ``` Luiz Cieslak 补充指出，许多医院系统在推进创新时常陷入“百花齐放的花园”（garden of a thousand flowers）的陷阱——每个人都有自己的绝妙点子，并各自挑选一套有利于自己的指标来辩护。然而，如果指标无法直接与医院的损益表（P&L）挂钩，这些创新终将消亡。日活用户数、功能采用率或转化率再高，若未能实质减少运营成本、缩短住院天数（length of stay）、减少护士呼叫升级，或者提升门诊吞吐量带来更多收入，就无法创造实际价值。探索性实验固然有其阶段意义，但最终必须跨入产生底线财务效益的成熟期。 - 衡量 AI 的核心硬性业务结果包括：30 天再入院率、平均住院日（Length of Stay）、升级至护士处理的来电量。 - 仅关注采用率（Adoption）或日活跃用户（DAU）而脱离财务损益（P&L）的项目，往往在实验期后无以为继。 --- ### [12:53](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=773s) AMA 治理工具包与“健康回报”框架 Margaret Lozovatsky 结合自身担任首席医疗信息官（CMIO）的从业经验表示，面对每天接踵而至的各类技术方案，最关键的追问永远是：“我们试图解决的根本问题是什么？”（"What is the problem we're trying to solve?"）。将讨论拉回痛点本身，是项目优先级排序的关键准绳。 为了帮助医疗机构在瞬息万变的技术浪潮中建立秩序，美国医学会（AMA）推出了两项关键举措： 1. **AI 治理工具包（AI Governance Toolkit）**：提供逐步落地指引，指导组织如何优先排列用例、组建跨职能团队、科学评估工具，并确保临床医生全程参与决策，实现技术在临床工作流中的审慎整合。 2. **健康回报框架（Return on Health Framework）**：鉴于 AI 工具极其昂贵，传统的纯财务 ROI 难以全面衡量其价值。该框架依据“医疗五重目标”（Quintuple Aim），将医疗质量提升、患者就医体验改善以及临床医护体验优化等非结构化/无形价值（intangibles），系统性地折算并映射到财务效益中，以对冲部署成本。 Luiz Cieslak 接着提醒，除财务回报外，AI 工具在跨科室扩展时最大的非资金障碍是**工作流整合与采用率**。许多方案最初仅是某个部门负责人的“宠物项目”（pet project），一旦推广到隔壁科室，就会因为打破现有工作流而遭遇抵触。Margaret Lozovatsky 幽默地借用一句名言进行了总结：“我最喜欢的一句话是——‘只要建好它，他们绝不会自动找上门’”（"If you build it, they will not come"）。 --- ### [15:11](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=911s) 展望未来五年：从模型辅助到自主 Agent 的进化 当主持人 Juliette Bolea 询问未来五年的格局是否会从“副驾驶”（Co-pilots）全面走向“全自主”（Fully Autonomous）时，Adam Kamor 表示技术与监管是两大决定性维度。他分享了一段个人经历：早在 20 到 25 年前，他在看牙医做 X 光拍片时，诊所就已经在利用早期机器学习算法检测龋齿。AI 在医疗领域的应用由来已久，并非由 OpenAI 开启。虽然大语言模型带来了跨越式突破，但技术在临床场景中的落地仍然会呈现阶段性渐进演进（incremental improvements），并在医护人员监督下提供更深度的支持。 Sameer Sethi 则给出了更加激进（bullish）的判断。过去 20 年他致力于利用数据构建决策信号，而实际应用中最大的阻碍恰恰是“人”——将数据信号交给人工去判断执行时，往往因为人为疏忽或延误导致价值落空。他认为医疗行业绝不能取代一线临床医生，但在后台与运营场景中，当前由算法输出的信号有大约 60% 完全不需要人工介入： - **Agent 的本质**：Agent 不仅是一个流行词，其核心在于自动化编排协调（orchestrating）各类异构技术，代表非人类系统自动执行多步骤操作。 - **未来五年演进路线**：从单纯由人工处理信号，转变为由自主 Agent 承接 40% 到 60% 的中间流程与数据流转，大幅释放被繁琐步骤消耗的人力。 --- ### [19:08](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1148s) 落地实例：收入周期管理（RCM）中 Agent 的降本实绩 Sameer Sethi 分享了目前在其医疗系统内真实运行的 Agent 实践。在探索早期 Agent 技术时，团队确立了一条首要红线：“选择绝对不会出人命的场景做试验”（"Let's work on something that doesn't kill anybody"）。他们将目标聚焦在收入周期管理（RCM）中的保险理赔拒付申诉流程上。 以往商业保险公司对索赔进行拒付时，整套申诉极为依赖人工，单份拒赔申诉处理成本高达约 35 美元。借助自主 Agent 流程编排后，机器处理环节的单笔成本直线骤降至 60 美分。 ```mermaid flowchart TD accTitle: RCM Denial Appeal Workflow Automation subgraph INTAKE["1. 拒赔受理"] A["收到保险拒赔信函"] --> B["生成式 AI: 解析与结构化提取"] end subgraph RESOLUTION["2. 自动诊断与取证 (Agent 编排)"] B --> C{"基于规则的模型匹配"} C -->|"缺失修饰符"| D["自动补齐对应 Modifier 编码"] C -->|"缺失预授权"| E["自动调取 Epic / EMR 抓取凭证"] end subgraph FULFILLMENT["3. 申诉与出件"] D --> F["生成式 AI: 起草专业申诉信函"] E --> F F --> G["人工审核员最终确认 (Human-in-the-loop)"] G --> H["正式发出申诉信函"] end ``` 该流程的经济效益与运营提升表现为： - **成本断崖式下降**：机器自动化部分将原先人工所需的每单约 35 美元成本压缩至仅 60 美分（$0.60）。 - **优化现金流周转**：人工团队难免面临员工请假、生病、离职与交接等波动，而软件 Agent 能够 7×24 小时全天候无间断执行，显著加速了被拒款项的追讨周期与资金回笼。 ### [21:02](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1262s) 智能体工作流与“终结谷歌医生（Dr. Google）” 未来的医疗体系将依赖多项技术的深度编排。嘉宾指出，未来五年的核心在于构建“智能体工作流”（agentic workflows），将生成式 AI、传统机器学习以及数据分析技术有机结合，在所有可行的环节中最大程度减少人工干预，从而提升保险申诉等后台流程的效率与回款能力。 ```mermaid flowchart LR accTitle: 医疗工作流演进与前端患者分流 subgraph BackEnd["后台：智能体工作流编排"] direction TB A["传统机器学习 / 数据分析"] --> C["Agentic 流程编排"] B["生成式 AI 模型"] --> C C --> D["减少人工干预 / 自动申诉"] end subgraph FrontEnd["前端：消灭 Dr. Google"] direction TB E["患者输入症状与上下文"] --> F["AI 结合时令/背景精准评估"] F --> G["消除不必要恐慌 / 阻断过度就医"] end BackEnd -.-> FrontEnd ``` 与此同时，技术变革不仅发生在后台（收益周期管理），更将深刻重塑前端患者体验。Luiz Cieslak 提出了一个大胆的预测：到 2030 年左右，行业将彻底“终结谷歌医生（Dr. Google）”。过去患者身体不适时自行搜索 WebMD 或 Google，往往会得到最极端可怕的诊断（例如将普通发热误判为脑瘤），导致恐慌并给儿科医生和全科诊所带来沉重负担。AI 能够结合具体上下文（如当前正值过敏高发季、患者仅有流涕与低热）在患者触点端提供精准分析，直接推荐过敏药（如 Flonase），从源头消除过度焦虑与不必要的医疗资源挤兑。 - **Dr. Google 的弊端**：传统搜索引擎缺乏患者上下文，导致患者产生过度焦虑并滥用医疗资源； - **AI 前端干预的优势**：基于实时情境进行个性化初筛，在就诊流程最前端精简不必要的步骤。 --- ### [22:50](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1370s) 增强智能（Augmented Intelligence）与临床工作流重构 来自美国医学会（AMA）的 Margaret Lozovatsky 从临床角度回应了这一观点。医生群体非常欢迎“终结谷歌医生”，虽然赋能患者了解健康信息至关重要，但错误信息引发的恐慌往往导致医疗资源的非理性利用。在 AMA 的理念中，AI 更应被定义为“增强智能”（Augmented Intelligence）——技术的核心使命是自动化繁琐的行政流程，而真正的临床决策权和医患关系必须牢固保留在医护人员手中。因为“现实中的患者永远不会完全按照医学教科书生病”，复杂情况必须依赖临床医生的专业判断。 在未来的工作流中，技术将“解放数据并将其转化为真正的信息”。Margaret 质疑了传统医疗记录的必要性：当所有数据都实时存在于电子健康档案（EHR）并可即时调用时，医生为何还需要耗时撰写病历？当系统拥有完整的操作审计数据并能提供实时决策支持时，为何还需要人工填写多达 37 个字段的不良事件报告？未来行政事务将被彻底重塑，但医患之间的信任纽带与专业裁量权仍将是医疗的核心。 --- ### [25:12](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1512s) 患者知情权、透明度与信任机制 讨论转向患者端的知情权与透明度问题：当 AI 参与到预约外呼或诊疗流程中时，医疗机构应当披露到何种程度？Adam Kamor 坦言，作为普通人，接到明显由大语言模型（LLM）驱动的语音电话会令人反感；但若医生在后台使用 AI 辅助诊断，患者其实并不需要事事知情，正如医生一直以来使用 Google 或传统算法辅助决策一样，AI 只是工具箱中的一件新工具。 ```mermaid flowchart TD accTitle: 医疗 AI 的透明度与信任闭环 A["主动告知与透明披露<br>(短信/电话/交互渠道)"] --> B["建立患者预期与信任"] B --> C["AI 高效响应与实时干预"] C --> D{"遇到复杂/高危问题?"} D -- 是 --> E["顺畅升级至人工临床专家"] D -- 否 --> F["完成日常服务/闭环"] E --> G["保障医疗质量与安全"] ``` Luiz Cieslak 则强调，透明度是建立信任的基础。无论通过电话还是短信交互，机构都应在用户授权（Opt-in）的前提下明确告知 AI 的介入。他以美联航（United Airlines）的延误通知为例：当航空公司明确标注“本信息由 AI 自动生成”时，旅客不仅能理解信息能被更早推送的逻辑，还能基于信任及时改签更早的航班。Margaret 补充指出，医疗领域应用 AI 已有数十年历史，透明度本身永不过时；只要系统具备健全的升级机制（Escalation path），在关键时刻能将问题无缝转交至人类医生把关，透明化披露反而会进一步强化患者对医疗体系的信任。 --- ### [29:25](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1765s) 知情同意的双重维度：诊疗应用 vs. 数据研发 Sameer Sethi 剖析了“知情同意（Consent）”在医疗 AI 落地中的两个截然不同的维度： | 知情同意维度 | 现状与主要挑战 | 解决路径与趋势 | |:--- |:--- |:--- | | **临床诊疗中使用 AI**<br>*(Use of AI for treatment)* | • 机构通常将其置于通用签署协议中，患者往往不细读<br>• **核心短板**：缺乏对一线医生的培训，医生无法清晰向患者解释 AI 的原理与作用 | • 社会大众接受度正快速提升（类似疫情后对远程医疗 Telehealth 从抗拒到主动接受的转变）<br>• 需加强临床医生的 AI 沟通与解释能力培训 | | **数据用于 AI 产品研发**<br>*(Use of patient data for AI development)* | • 行业目前普遍忽视了该维度的独立授权<br>• 存在严重的合规与伦理隐私隐患 | • **必须获取显式授权（Explicit Consent）**<br>• 建立严格的数据合规治理与权限管理体系 | Sameer 指出，社会大众对新技术的接受度正在显著提升。例如在疫情前，许多老年患者坚决抗拒远程问诊，坚持去医院看相识 20 年的老医生；而疫情后，同一批患者已经开始主动要求使用远程诊疗。因此，患者对于在诊疗流程中使用 AI 的接受度会逐步提高。然而，真正的严峻挑战在于**将患者数据用于算法与产品研发**，医疗机构绝不能将其混为一谈，必须建立严格、透明且独立的显式知情同意机制。 ### [31:30](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1890s) 医疗数据主权、厂商霸王条款与患者知情同意困境 小组讨论切入了整个医疗 AI 落地过程中最尖锐的问题之一：医疗系统与技术供应商（Vendors）之间关于患者数据所有权与知情同意的博弈。Sameer Sethi 率先指出，全社会普遍在回避这个问题，尽管像 Hackensack Meridian Health 这样的大型机构选择正面应对，但 99% 的医疗系统至今仍视而不见。在采购软件时，医疗机构往往因为供应商的功能强大、商业报价合理就将数据交付出去。然而，一旦供应商发生数据泄露，或者患者发现自己的敏感数据被转交给商业公司用于“产品研发以谋取商业利益”时，医疗机构就会深陷被动，遭到患者“我根本没有同意你们把数据拿去商业化”的直接问责。 现实中的博弈更为残酷：许多强势的 AI 软件大厂会公然向医院施压——“如果你不接受我们的数据条款，就别用我们的产品”。他们甚至绕过管理层向临床医生直接施压，迫使医院为了临床需求而被迫妥协。Sameer Sethi 举了一个不点名的真实案例：他们目前使用的一家环境音/临床听写（ambient note-taking）供应商就直言，只要医院不反对，他们就会拿临床交互数据去迭代自己的产品模型。对此，Adam Kamor 幽默地插话推荐可以采用 Tonic.ai 的去标识化方案，但各方也坦言，在大型厂商的强硬态度面前，很多时候连推进保护技术的谈判空间都没有。 ```mermaid flowchart TD accTitle: 医疗数据使用争议与知情同意链条 subgraph S1["临床数据产生与流转"] P["患者 (Patient)"] -->|"产生临床就诊数据"| H["医疗机构 (Health System)"] H -->|"采购引入AI辅助系统"| V["AI/软件供应商 (Vendor)"] end subgraph S2["数据风险与合规博弈"] V -->|"未充分去标识化/商用训练"| M["AI模型研发与训练 (Model Training)"] V -.->|"数据泄露风险 (Breach Risk)"| X["泄露隐患"] M -.->|"未明确知情同意"| P end ``` - **知情同意真空**：绝大多数医疗系统在与供应商签订合约时，并未就患者数据用于第三方商业模型训练获得明确的患者知情授权。 - **临床端与合规端的脱节**：技术厂商通过向一线医生推销便捷工具，自下而上迫使医疗系统在数据主权条款上妥协。 --- ### [33:16](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1996s) HIPAA 法规门槛：安全港协议的失效与专家裁定法的落地难 随后，Adam Kamor 将话题引向监管底线——即使合同条款有漏洞，供应商依然绕不开 HIPAA 法规的要求，在拿数据训练模型时必须严格去标识化（De-identification）。但 Sameer Sethi 随即指出了残酷的现实：HIPAA 传统的“安全港方法”（Safe Harbor）在当今强大的算力环境下早已形同虚设，利用海量算力对脱敏数据进行重新识别（Re-identification）只需要“几毫秒的时间”。 当 Safe Harbor 方案导致数据因删减过度而丧失临床效用（Utility）时，HIPAA 提供的另一条路径是“专家裁定法”（Expert Determination）。然而当 Sameer Sethi 向众多 AI 初创团队建议走专家裁定时，对方甚至反问“那是什么”。Adam Kamor 对此深有感触地附和，许多创业团队只要多读两句法规指南就能看到专家裁定条款；虽然获得一份专家裁定评估信可能需要耗费数周时间（而非数月），但急于快速迭代的初创公司往往将其视作绊脚石，忽视了这是保护患者和自身合规的关键屏障。 - **Safe Harbor（安全港法）的局限**：在现代高性能算力与多源数据交叉碰撞下，传统剥离特定标识符的方法极易被暴力逆向重识别。 - **Expert Determination（专家裁定法）**：通过统计学和科学评估证明数据泄露风险极小，耗时通常为数周（weeks, not months），但常被追求速度的初创公司忽略。 --- ### [35:00](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2100s) 冗长法务条款的“免责骗局”与患者教育场景化重构 Luiz Cieslak 将讨论拉回到了患者体验的本质。他指出，目前的知情同意流程和大众日常在网络上签署软件服务条款（Terms and Conditions）毫无二致——数十页晦涩的法文术语，用户唯一做的就是迅速滑动到页面底端点击“我同意”，这本质上只是一种机构单方面的“免责声明（CYA - Cover Your A**）”。Luiz Cieslak 现场询问观众有谁在注册医院患者门户（EMR portal）时认真读过数据披露条款，全场仅有一人举手；Adam Kamor 也附和吐槽自己两天前刚买了一台新电视，开机引导里居然塞了 50 多份法律协议要求阅读，简直成了笑话。 针对这种形式主义对患者体验与产品研发的双重伤害，Margaret Lozovatsky 提出了极具建设性的解决方案。她强调，不能把解释数据隐私和安全条款的重担甩给临床医生（这不是医生时间的最佳用途），而是应当充分利用患者在就医过程中的碎片化“死时间”（Dead Time）——比如候诊室等待、病房输液卧床的空闲时段，利用交互式数字化工具向患者普及他们的医疗数据安全究竟意味着什么。Margaret 特别警示，由于现代数据重识别技术的发达，尤其是具有罕见病（unique conditions）特征的患者，其数据一旦泄露几乎无法被真正匿名，医疗机构的领导者必须主动建立透明沟通机制。 - **知情同意的形式化**：患者端通过滚动条强制阅读 JavaScript 的交互设计未能提供真正的知情权，反而在医患间筑起了法律壁垒。 - **利用就医碎片时间进行患者教育**：利用门诊等候与病房留观的闲暇时段，通过平板或交互屏幕向患者透明展示其数据的使用边界与保护机制。 --- ### [38:20](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2300s) 闭幕建议：隐私规划、临床参与、定制化微生态与认知普及 在对谈的尾声，主持人 Juliette Bolea 给在场观众布置了一道趣味“课后作业”——尝试去询问常用的大语言模型“它到底掌握了你多少个人信息”，其结果绝对会令人瞠目结舌。随后，主持人邀请每位嘉宾为医疗 AI 开发者和从业者留下最后一句核心建议： 1. **Adam Kamor（技术与合规视角）**： > "Make sure you have a plan in place ahead of time for how you're going to deal with the privacy aspect."（一定要提前制定好隐私应对方案。） 绝不能在模型训练完成、权重中已经固化了受保护健康信息（PHI）甚至环境违规之后才发现无法商用，必须未雨绸缪地保护数据隐私。 2. **Margaret Lozovatsky（临床融合视角）**： 无论开发的是否是 AI 软件，都必须让距离医疗最近的群体参与决策。“技术专家懂软件，但临床医生才真正懂得日常医疗照护的交付实质。” 3. **Luiz Cieslak（数据与业务定制视角）**： “通用大模型并不是万能的（One size does not fit all）。”基础模型在大语境和通用知识检索上表现优异，但无法直接匹配具体医院、特定人群微生态（Microcosmos）和地域特征的诊疗现实，必须将本地真实医疗数据分层构建在通用大模型之上。 4. **Sameer Sethi（行业与认知视角）**： 面对扑面而来的技术浪潮，医疗机构必须持续开展教育，清晰界定 AI 技术的“能与不能”，帮助行业与消费者在狂热中保持清醒认知。 ### [41:27](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2487s) 解构 AI 迷思：“高果糖玉米糖浆”比喻与产品交互设计 对于医疗机构而言，引导管理层和组织成员正确认识 AI 至关重要。嘉宾指出，必须向组织内部反复澄清“AI 到底是什么，不是什么”——它本质上是一项高度依赖底层数据的技术工具，而不是无所不能的“魔法”。当前大众认知中存在一个巨大误区，即把 AI 简化为一个“在对话框里输入文字、然后等待回答”的黑盒界面，这种将 AI 等同于输入框的交互模式是对 AI 技术非常糟糕且狭隘的呈现。 发言者幽默地打了个比方，甚至开玩笑说自己可能会因为这个言论“惹上麻烦”：像 ChatGPT 这类大语言模型对话工具，本质上就像食品工业里的“高果糖玉米糖浆”（High Fructose Corn Syrup）等人工甜味剂。这个比喻的原因在于——它确实能快速解决问题，尝起来也足够甜美诱人，但它绝不适合所有人，也无法适配每一个具体的医疗业务场景。 ```mermaid flowchart TD accTitle: 生成式 AI 在组织落地的两条路径 subgraph TechnologistStrategy["技术与产品设计策略"] A["生成式 AI 落地挑战：避免单纯提供空白输入框"] --> B{"二选一产品路径"} B -->|"路径一：赋能用户"| C["系统性培训提示词工程<br/>(Teach Prompt Engineering)"] B -->|"路径二：产品化封装"| D["将优化好的 Prompt 隐藏在按钮之后<br/>(Hide Prompt Behind a Button)"] C --> E["产出稳定、准确的业务价值与 ROI"] D --> E end ``` 针对生成式 AI（Generative AI）的落地，技术专家提出了两条清晰的实践路径：要么系统性培训终端用户掌握提示词工程（Prompt Engineering），要么将精心调试好的 Prompt 直接封装在产品功能按钮之后。如果只是丢给用户一堆空白输入框让他们随意提问，他们往往无法以正确的方式组织问题，最终只能得到错误或低质量的回答。如果继续给员工提供更多无引导的聊天框，企业将永远无法实现预期的投资回报率（ROI）。 - **破除魔法论**：AI 是建立在数据基础之上的技术架构，而非凭空产生奇迹的魔法。 - **甜味剂隐喻**：类 ChatGPT 工具如同“高果糖玉米糖浆”，虽立竿见影、见效极快，但并非所有场景的健康解法。 - **交互落地准则**：避免滥用自由对话框，应通过“深度培训 Prompt 技巧”或“按钮化封装预设 Prompt”来确保输出质量与业务回报。 --- ### [43:32](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2612s) 讨论收尾与圆桌致谢 在精彩的观点碰撞之后，主持人对“将提示词封装在按钮后”以及“警惕把 AI 当作玉米糖浆”的独到见解表示高度赞同，称要将这一观点牢记带走。随着会议时间临近尾声，主持人向全体与会者致谢，并号召全场观众为台上来自投资界、医疗协会、数字化科技公司以及医疗系统的各位专家嘉宾报以热烈的掌声，本次围绕“利用 AI 端到端优化患者体验”的深度圆桌研讨至此圆满落幕。 ## Notable quotes > [08:38](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=518s) 凡是不属于这六大业务范畴的 AI 提案，一律不予立项 — Sameer Sethi > [08:38](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=518s) AI is not free, and AI is not cheap — Sameer Sethi > [12:53](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=773s) What is the problem we're trying to solve? — Margaret Lozovatsky > [12:53](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=773s) If you build it, they will not come — Margaret Lozovatsky > [19:08](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1148s) Let's work on something that doesn't kill anybody — Sameer Sethi > [22:50](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=1370s) 现实中的患者永远不会完全按照医学教科书生病 — Margaret Lozovatsky > [38:20](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2300s) Make sure you have a plan in place ahead of time for how you're going to deal with the privacy aspect. — Adam Kamor > [38:20](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2300s) 技术专家懂软件，但临床医生才真正懂得日常医疗照护的交付实质。 — Margaret Lozovatsky > [38:20](https://www.youtube.com/watch?v=ZbPppJ-ZUqs&t=2300s) One size does not fit all — Luiz Cieslak ## Mentioned Juliette Bolea, Sound Ventures, Margaret Lozovatsky, American Medical Association (AMA), Luiz Cieslak, CI&T, Sameer Sethi, Hackensack Meridian Health, Adam Kamor, Tonic.ai, Oscar Health, Capsule, Hims & Hers, OpenAI, Anthropic, Epic, ChatGPT, Google, WebMD, United Airlines, Flonase ## Video description Juliette Bolea, Investor, Sound Ventures Margaret Lozovatsky, VP of Digital Health Innovations, American Medical Association Luiz Cieslak, SVP of Healthcare & Life Sciences, CI&T Sameer Sethi, SVP, Chief AI & Insights Officer, Hackensack Meridian Health Adam Kamor, Co-founder, Head of Engineering, Tonic.ai Abstracts This session will focus on how AI is enhancing the patient experience from initial engagement to post-treatment care. Panelists will discuss the application of AI in areas such as personalized treatment plans, virtual health assistants, predictive analytics for patient outcomes, and seamless care coordination. Learn how AI is improving patient satisfaction by enabling more proactive, personalized, and efficient care. Discover best practices for integrating AI solutions that not only optimize clinical outcomes but also create a patient-centered healthcare experience. [Source](https://www.youtube.com/watch?v=ZbPppJ-ZUqs)

[8:38] Sameer Sethi: 凡是不属于这六大业务范畴的 AI 提案，一律不予立项 [8:38] Sameer Sethi: AI is not free, and AI is not cheap [12:53] Margaret Lozovatsky: What is the problem we're trying to solve? [12:53] Margaret Lozovatsky: If you build it, they will not come [19:08] Sameer Sethi: Let's work on something that doesn't kill anybody [22:50] Margaret Lozovatsky: 现实中的患者永远不会完全按照医学教科书生病 [38:20] Adam Kamor: Make sure you have a plan in place ahead of time for how you're going to deal with the privacy aspect. [38:20] Margaret Lozovatsky: 技术专家懂软件，但临床医生才真正懂得日常医疗照护的交付实质。 [38:20] Luiz Cieslak: One size does not fit all

[8:38] Dr. Margaret Lozovatsky: If you build it, they will not come. [8:38] Sameer Sethi: chasing the shiny object [8:38] Luiz Cieslak: garden of a thousand flowers [17:05] Sameer Sethi: it doesn't kill anybody [21:28] Luiz Cieslak: turn off your router [34:38] Luiz Cieslak: cover your ass [34:38] Luiz Cieslak: a joke [41:08] Sameer Sethi: high-fructose corn syrup

[0:04] Good afternoon. Thank you all for [0:05] joining us here in Vegas. Has anyone [0:08] gotten lucky thus far? [0:11] Um, thank you so much for coming to this [0:13] panel. I would say the the reason I'm [0:15] personally very excited for this [0:17] conversation is every single one of you [0:20] is a patient. And so the impact that AI [0:24] is going to have on the patient [0:26] experience is something you may have [0:28] already personally experienced or will [0:30] experience over the next few years [0:32] regardless of what anyone does here in [0:35] their professional careers. Um so we are [0:38] just delighted to have an amazing panel [0:41] of folks here. So why don't we kick off [0:43] with some introductions if you can each [0:45] just give some background on yourself, [0:47] your organization, and then we can kick [0:49] it off from there. [0:52] All right. Thank you. Uh, my name is Dr. [0:55] Margaret Lossky. I am a pediatrician. I [0:58] take care of kids in the hospital. So, [1:00] I'm a pediatric hospitalist and I am the [1:03] CMIO and vice president of digital [1:06] innovation at the American Medical [1:08] Association. Um, prior to joining the [1:10] AMA about uh a year and a half ago, I [1:13] have had multiple CMIO CHIO roles at [1:16] health systems implementing technology. [1:20] >> All right. So uh I'm Louis Chislick U [1:22] head of healthcare life science for CNT. [1:25] We're a global service provider focus on [1:27] systems of engagement. So a lot of what [1:29] we're talking about here is going to be [1:31] is part of what we focus on building [1:33] solutions and experiences for patients [1:35] physicians that really help them get to [1:38] uh live a better life. [1:41] >> Uh my name is Samir Sati. I function as [1:43] the AI officer for Hackinack Marian [1:45] Health. We're a health system based out [1:46] of New Jersey which consists of uh 18 [1:49] hospitals close to uh 500 care sites. [1:53] >> Hey, my name is Adam. I'm the head of [1:55] engineering at Tonic AI. Uh Tonic AI is [1:57] a synthetic data platform. We help [1:59] ensure that your customers or your [2:01] patients PHI data doesn't make its way [2:04] into environments where it shouldn't. Uh [2:06] both for application testing and [2:07] development and also ensuring that you [2:09] know customer PHI doesn't make its way [2:11] into the weights of models that you're [2:12] training when you're building AI models. [2:16] Uh, thank you all and I'll just be [2:18] facilitating the conversation, but as a [2:20] bit of background, I'm an investor at [2:22] Sound Ventures. I lead up all our entire [2:24] healthcare practice. We've invested in [2:26] everything from Oscar Health to uh [2:29] Capsule to HIMS and hers n and a bunch [2:31] of other companies in the healthcare [2:33] space. We're also very active investors [2:34] in AI and are one of the earliest and [2:36] largest backers of OpenAI and Anthropic. [2:39] So, spend a ton of time at the [2:40] intersection of both. Um, but to kick us [2:43] off, uh, we're gonna hit everyone first [2:46] and could you please kind of share with [2:48] us how AI is impacting the patient [2:50] experience and any specific examples you [2:53] can give would be best. [2:57] All right, I'll jump in. Um, I think [3:00] that there are multiple aspects of AI's [3:04] impact on our patients. One of which is [3:07] that we now have information that is [3:09] readily available in a way it has never [3:12] been and it is very tempting I think to [3:15] all of us um because I'm a patient as [3:18] well as a physician to quickly get [3:21] access to that information and as a [3:24] healthc care industry we have to think [3:26] about what that looks like when our [3:29] patients are looking at their little [3:31] Google summaries which is driving some [3:34] of the decisions that they're making And [3:36] a great example of that is I recently um [3:39] just as I was prepping for some of the [3:42] interviews I was doing asked a few [3:44] questions. One of which was how to [3:46] handle a UTI and it told me to drink [3:48] urine. And [3:51] while that may be comical and most of us [3:53] would not proceed with that plan, I [3:55] think it highlights some of the risks [3:58] and benefits of where we are today with [4:01] these tools being readily available. And [4:04] the other aspect of it that I want to [4:06] make sure that I mention is that a lot [4:09] of these tools are being used to [4:11] decrease administrative burden in our [4:13] healthcare environment which is allowing [4:15] us to have much more facetime with our [4:18] patients. So that is a very positive [4:20] impact. A lot of the ambient AI tools [4:23] that are now being implemented across [4:25] our industry and we're asking ourselves [4:26] is a table stakes at this point are [4:29] allowing us to have more time to answer [4:32] questions to look at our patients when [4:34] we're seeing them. And I think that [4:36] these positive aspects are also really [4:38] important for us to highlight and [4:40] continue to build on. [4:43] >> Can jump in. I think you covered two of [4:45] the main points I think fascinating. The [4:47] third one that I believe is the ability [4:49] to personalize the care. Um I was [4:52] reading some statistics recently. It's [4:53] like one out of three Americans now wear [4:56] one at least one wearable the exception [4:58] I wear two not content with one set of [5:01] data but the ability now to crunch that [5:04] data and take my personal aspects into [5:07] care into into consideration for care. [5:09] Um we I think one of the moment the [5:13] proudest moments we had in CINT with one [5:15] of the projects helping a cancer [5:17] institute was creating personalized [5:20] treatment protocols for cancer patient [5:22] that took into consideration their [5:24] treatment their specific type of cancer [5:26] their uh uh conditions their their also [5:31] um demographics in being able to provide [5:34] them with a better care in the long run [5:36] and higher survival rates. [5:40] At HackinSack, our our approach is [5:42] slightly different. Uh when we started [5:44] down the AI journey, we first looked at [5:47] what we don't do well today and see how [5:50] AI can help with that. Uh burnout is a [5:52] real thing. Uh right. So what we [5:55] realized is that as more and more data [5:58] is getting generated about patients, we [6:00] haven't really provided our clinicians [6:03] with the right tools to consume that [6:05] data and do something with it. it's ex [6:06] it stays in the EMR and uh to swim in [6:11] that sea of information is really hard. [6:14] So what I I still remember when we [6:16] talked about this you know we actually [6:19] looked at some footage of of an [6:21] appointment and we noticed that the [6:23] patient while the patient was next to [6:25] the doctor the the the clinician was [6:28] staring at the screen while talking to [6:29] the patient. That shouldn't be the case, [6:31] right? They should be an eye to eye [6:33] contact, right? That's what makes [6:34] healthcare personal and healthcare is [6:36] personal. So, we focused on building or [6:39] using AI to help clinicians consume this [6:43] information because we were not doing [6:44] that well. So, I think that's one aspect [6:46] of this. The other aspect is was we [6:49] weren't really doing a good job after [6:53] sending the patient home. So, we did a [6:55] really good job, I think, taking care of [6:57] the patient while they're in the [6:58] hospitals, but once they went home, they [7:00] were they weren't cared for. Uh, we [7:03] didn't have enough humans to make these [7:06] post-operative phone calls, right? So, [7:08] we invested in that. We invested in [7:10] conversational AI where because we [7:12] didn't have humans to do it, in most [7:14] cases, wherever it made sense, we would [7:17] have use AI or chat bots to actually [7:20] call these patients and ask them how are [7:22] they doing? do they need follow-up care? [7:24] So that's been our approach at [7:26] HackinSack is we have focused on what we [7:28] have not been able to do well for [7:30] various reasons and see how AI can help [7:32] with that. [7:34] >> So atonic we don't deal directly with [7:36] patients but our our customers do. So [7:39] you know our customers can range from [7:40] you know very large insurance companies [7:42] to you know small startups that are [7:44] using LLMs to you know solve problems in [7:46] the patient experience. And I I think [7:48] it's really mostly what what you know [7:49] the doctor said. Um like you know we see [7:52] our customers using AI to either you [7:55] know relieve the workload or relax the [7:57] workload a bit on the providers [7:58] themselves or to enable like automation [8:00] of back office workflows that just kind [8:03] of like let humans focus more on like [8:06] the human aspect of things uh and less [8:08] on kind of the the wrote work that must [8:11] be done. uh and and you know an [8:14] interesting side effect of all that [8:15] actually has been there is now a a [8:18] massive demand by by many companies for [8:20] data and just more and more data beyond [8:23] what they have. Sometimes like we have [8:24] customers that come that they have no [8:25] data or maybe they have a lot of data [8:27] but they need more. Everyone needs data [8:29] to start building these experiences [8:30] because these models are very data [8:32] hungry. Uh and it's it's it's very been [8:34] a very interesting time I think. [8:38] >> Thank you. Um, this next one's for Samir [8:41] and Lee. [8:42] Right now, everyone needs an AI strategy [8:45] and no one's really asking what is your [8:48] ROI. Maybe they are, but that will [8:50] become a more serious conversation over [8:53] the next couple of years. How do you [8:55] think about measuring success? What is [8:58] the right KPIs? What are the metrics in [9:02] order to ensure that there is success [9:03] with these programs? And maybe how are [9:05] you tracking that today? How are your [9:07] clients tracking that today or into the [9:09] future? [9:12] >> Yeah. So, I think AI strategy is really [9:14] important. Otherwise, you will chase the [9:17] shiny object and there's there's stuff [9:20] coming out every day, every week. Uh I [9:22] can't tell you how many times there's a [9:24] conversation like this or a show like [9:26] this and our leaders come in and say, [9:28] "Why aren't we doing this?" Uh without a [9:30] strategy in place, you're going to be [9:31] all over the place. [9:33] uh we uh struggled with this like I [9:35] think all organizations do. I still [9:37] remember the first year or so was was [9:39] just doing that which is just running [9:41] from one place to another one use case [9:42] to another one vendor to another and [9:45] then eventually we act uh decided to get [9:48] our our board in the room and have a [9:51] conversation what are the areas that we [9:53] need to focus on. uh we selected six [9:55] areas which are broad in nature but [9:58] enough for us to say if it doesn't fall [10:01] into these buckets or these areas we [10:03] will not do it right uh so so I think I [10:07] think that helped us quite a bit from an [10:09] ROI perspective I think that's a [10:12] very growing conversation uh AI is not [10:15] free uh and AI is not cheap uh and [10:20] understanding what the ROI is is really [10:23] important before you do any use case. So [10:26] I talked about uh the the use case of [10:29] where we were uh building AI capability [10:33] to call patients post discharge. Well, [10:35] the ROI for that is readmissions, right? [10:38] It's plain and simple. Our intention is [10:41] to make sure that a patient does not end [10:43] up in the hospital within 30 days of [10:45] discharge. It's as simple as that, [10:48] right? So I think setting those KPIs or [10:52] measures is really important before you [10:55] start down the path of building any use [10:57] case. And it's really hard to do that by [10:58] the way, right? Because these a lot of [11:01] these capabilities seem cool and nice to [11:03] do and reputationally aligned and a [11:05] bunch of other things come into the mix. [11:07] So it's a noisy space, but being very [11:09] clear, for example, I'm going to build a [11:11] chatbot to reduce readmissions is going [11:14] to be really important and you measure [11:15] against it. That's what we do now, [11:17] right? We we we look at how many [11:18] patients this chatbot called and then of [11:22] these patients how many of them actually [11:24] ended up in the hospital within 30 days [11:26] of readmissions and that I think [11:27] structured and focused approach and [11:29] helped us quite a bit. Uh what we see a [11:32] lot with different hustle systems is [11:34] that like the garden of a thousand [11:36] flowers. Everybody has an idea. [11:38] Everybody's idea is great, right? And [11:41] everybody starts creating those metrics [11:44] that um they want to defend and they [11:46] want to promote. And what we see across [11:49] different clients has been everybody [11:51] adopts a metric that is not necessarily [11:54] tied to the business outcome. You [11:55] mentioned readmission that's direct [11:57] impact to the bottom line. length of [11:59] stay and escalations to the nurse to the [12:02] call. You mentioned the chatbot is a [12:05] great one. Um if you don't tie up your m [12:08] your metrics back to dollars that ties [12:11] back to the piano of the hostel, you're [12:14] you're not going to be able to succeed. [12:15] can pick the adoption, you can pick uh [12:18] yes daily users, great. You can have [12:20] 100% of the usages, the the the the [12:24] access their conversion, but if at the [12:26] end of the day you're not reducing your [12:29] operational cost or improving the [12:31] ability to see more patients and [12:32] therefore more revenue, the ideas will [12:34] die essentially. So, it's focusing less [12:37] on all those thousand flowers that you [12:39] have on the garden and picking up what [12:41] is the the result. experimentation is [12:44] great but there's a phase a time and for [12:45] that and then start moving on to what [12:48] does generate that impact with the in [12:50] the in the in the bottom line [12:53] >> I was going to jump in um briefly to [12:56] highlight a couple things that we've [12:58] done at the AMA on this topic um and I [13:00] love the garden of a thousand flowers [13:02] because when I have worn my CMIO hat [13:04] over the years people come to you every [13:06] day with solutions and I always ask what [13:09] is the problem we're trying to solve and [13:11] when you bring it back to a you're [13:13] trying to solve that really helps with [13:15] prioritization of all of the different [13:17] ideas that are coming at us. So, one [13:20] thing that I wanted to highlight is that [13:21] we uh released this year the AI [13:24] governance toolkit and it is a [13:26] stepbystep guidance and I know this uh [13:30] industry is changing and it'll probably [13:32] continue to evolve by the day but with [13:34] what we know today of how organizations [13:36] can start to think about prioritizing [13:39] these use cases, setting up the right [13:41] teams, thinking through how to evaluate [13:44] these tools, making sure that they are [13:47] going to be thoughtfully integrated ated [13:49] in the clinical environment where you [13:50] have clinicians at the table for some of [13:52] these discussions. And then we also [13:54] developed what we call the return on [13:56] health framework which looks at ROI and [13:59] it looks at ROI a little bit differently [14:01] because we recognize it's a huge [14:03] investment. I appreciate you saying that [14:05] because it these tools are not cheap and [14:08] yet a lot of the ROI comes in [14:10] non-tangles such as improvements in [14:13] quality, the patient experience which we [14:15] talked about the clinician experience. [14:17] So the framework looks at what I think [14:19] of as the quintuple aim. All of these [14:21] different aspects of delivering health [14:24] care and how we can tie those to dollars [14:28] to offset the costs. [14:32] >> You touch on a point that I actually [14:34] forgot to mention. I think it's the [14:35] adoption too. A lot of the tools they [14:38] grow because it's somebody's pet project [14:40] and you can never get to the next [14:42] department because it's going to affect [14:45] my workflow. So adoption that the the [14:48] one metric that I would say it's [14:49] important that's not tied to the money [14:51] and the it's still tied to the problem [14:53] is how can I scale this app? How can I [14:56] get more people to adopt my my tool or [14:58] my solution that we're deploying without [15:00] affecting the workflow because then we [15:01] go back to [15:04] >> my favorite quote is if you build it [15:06] they will not come. [15:11] >> That's a good one. Um perhaps fast [15:13] forwarding in some time and Adam maybe [15:15] you can kick us off on this one and then [15:17] anybody else that would want to hop in. [15:19] Uh let's fast forward five years from [15:21] now. Um many of your clients are on the [15:23] cutting edge of what's happening with [15:25] these technologies. What what are maybe [15:27] some predictions you have as to what AI [15:30] and the patient experience will look [15:31] like? Will we go from co-pilots to fully [15:35] autonomous? Um any other kind of takes [15:38] as to where we're headed? [15:40] >> That's a tough one. Um I I think there's [15:42] like a there's a few angles by which you [15:44] could try to answer that. One would be [15:45] like will the technology be capable of [15:48] doing those things. Uh another you know [15:51] question that you could another approach [15:52] you could take is well will will the [15:54] regulatory framework be in place to [15:56] allow just like widespread adoption of [15:58] AI in you know diagnosing healthcare [16:01] conditions. Um I don't know the answer [16:03] to either. Um but those are the [16:05] questions I would ask. Um but I I could [16:07] probably touch at least a bit on the [16:08] first one. Um, and you know, because we [16:11] we we talked about this a little bit [16:12] before and I remember like probably 20 [16:16] years ago. I was no even longer than [16:17] that. Probably like 25 years ago, I'm in [16:19] the dentist office and they're using [16:20] machine learning models already to like [16:22] detect cavities in my mouth like when I [16:24] get an X-ray. And I think everyone here [16:26] or many of you have probably seen that [16:28] similar experience where, you know, AI's [16:29] actually been helping in healthcare for [16:31] a very long time already um before open [16:34] AI was a household name. And and I think [16:36] the progress that we see in, you know, [16:38] helping diagnose conditions is just [16:39] going to continue to incrementally [16:41] improve. And it's not always [16:42] incremental. There's going to be leaps [16:44] and bounds. Like when a large language [16:45] model comes out, that's a that's a leap, [16:47] right? And then after that, there will [16:48] be like incremental improvements. Um, [16:51] and I I think that we're going to [16:52] continue to see just incremental [16:54] improvements in in large language models [16:55] over the next few years that are going [16:56] to enable more and more um just like [17:00] workflows and interactions to be [17:02] supported by AI and not by and with with [17:05] the help of providers of course. [17:08] >> Uh I I think I have a bit of a bullish [17:11] vision around this uh or at least an [17:13] opinion. Um I think so I just just my [17:17] background I have been helping build [17:19] signals from data for almost 20 years [17:22] now. Uh the my barrier has been humans [17:28] people because what I'm doing is I'm [17:31] handing over these signals to a human to [17:34] do something with. [17:36] uh and that's what has I think led to uh [17:40] lack of adoption or or or or not usage [17:44] of the model or the the signal at the [17:46] right time. I think the next five years [17:49] are about how you can filter those [17:52] actions and then make a decision on [17:55] which of these signals require human [17:57] intervention and which do not and some [17:59] totally do especially in healthcare uh [18:01] cannot replace a clinician right uh I [18:04] think that's that's a given but of the [18:08] maybe 100% of the signals that people [18:11] like my teams provide maybe 60 of them [18:14] do not require human intervention and a [18:17] machine can take that on and take the [18:19] next step. So I think the next five [18:22] years is about agents to me and I know [18:25] there's a that's a buzz word right but [18:28] to me agents are about orchestrating [18:32] different technologies to work together [18:34] to take more steps and I think the next [18:38] 5 years we will see AI signals analytic [18:41] signals data signals being actioned by [18:44] nonhumans more and more and more and [18:46] that's what agents are to me like agents [18:48] could mean a lot of different [18:50] So we will see better use of data. We [18:53] will see better use of analytics because [18:55] steps that we rely on humans to do do [18:59] today maybe 50 40 60% of that will start [19:03] to get done by agents. So I think that's [19:06] what's going to happen here. I'll uh [19:08] I'll give you guys an example and this [19:10] is a nonclinical example. You know when [19:11] we're playing with agents today our our [19:14] rule is let's work on something that [19:16] doesn't kill anybody, right? [19:19] uh and uh yeah it's worked so far uh and [19:22] what we have done is we have focused in [19:24] the RCM space revenue cycle management [19:26] so we as a health system insurance [19:28] companies love denying our claims right [19:31] and when we get a denial uh we have a [19:34] human that looks at the letter uh makes [19:36] sense of it could be a modifier that's [19:38] missing it could be a pre-authorization [19:40] that we didn't attach right uh looks at [19:43] it go fetches that information generates [19:46] the the new letter and then sends it [19:48] out. Right? Uh that cost us somewhere [19:51] around $35ish dollars, right? Now, some [19:56] of that does not require a human to do, [19:58] right? So, the letter comes in, we use [20:00] generative AI to synthesize that letter. [20:03] We then tap into a a a [20:07] rule-based model to to identify, okay, [20:09] this for for this denial code against [20:12] this claim, this is the modifier that's [20:14] missing missing. attach that or go into [20:17] Epic or EMR, find the pre-authorization, [20:20] attach that. Next, use generative AI to [20:23] generate the appeals letter, then hand [20:25] it over to a human, right? And then the [20:29] human says okay, and the letter goes [20:31] out. The parts of the this piece that uh [20:36] that I described that was that can now [20:38] be done by a machine cost us 60 cents. [20:43] So that's the impact right and by the [20:46] way this isn't just about the cost right [20:47] it's also about cash flow you know we [20:50] rely on humans we we got to find a [20:52] person and you know humans get sick and [20:54] they quit and you know and machines [20:56] don't right so we have this thing [20:58] working day in and day out so our cash [21:00] flow gets better as well because we're [21:02] sending more appeals out and we're [21:03] getting more money in the bank so I [21:05] think that's what the next 5 years is [21:07] about it's about agentic workflows where [21:10] we are orchestrating different [21:11] technology ies including AI, including [21:14] traditional machine learning, including [21:16] data and analytics to all work together [21:18] to remove humans out of the the the [21:22] process wherever it seems possible. [21:28] >> I would pick back on your idea, not [21:30] looking from the rival cycle, but [21:31] actually the front of the house. My [21:33] extremely bullish in the five years that [21:35] by 2030ish, we're going to kill Dr. [21:38] Google. We everybody here as a patient [21:41] went there it's like hey I have this and [21:44] you come out with all the worst possible [21:47] conditions. I got a laugh from the the [21:50] doctor. So she has to deal with that [21:52] with all the parents going back to the [21:53] office. [21:54] >> It's like WebMD like I saw on WebMD but [21:56] now it's chatbt. [21:57] >> Oh I got a told my my by my p pediatric [21:59] my kids pediatrician to say go back home [22:01] and you have to do one thing. Turn off [22:03] your router. Don't look at anything on [22:04] the internet. [22:05] >> First- time parents that's what happens. [22:07] Um, but I'm bullish on killing Dr. [22:10] Google because the AI can get and [22:12] process that information at the point of [22:14] the patient. You can get that [22:15] information in the context, not just [22:18] because I have a fever. I have a fever. [22:20] I have it's fever season. I have runny [22:23] nose. Hey, you have an allergy. Go take [22:26] chances are you have flown or or you [22:28] need a flon or whatever medication is. [22:31] you don't need to go and build boil the [22:33] ocean for that problem and come in and [22:35] say hey I have a brain tumor and now [22:37] everything's coming out um and you get [22:40] that care more specific and you start [22:43] the process early on and you're ready [22:45] eliminate some of those steps so that's [22:47] my bullish prediction on the five years [22:50] >> well thank you for that um I think [22:52] you'll get a cheer from every physician [22:54] if you say we're going to kill Dr. [22:55] Google. Um, we want our patients to be [22:58] empowered and we want people to have [23:00] information and yet it often creates [23:03] angst and it creates utilization of the [23:05] health care system that's not always [23:07] appropriate. And so it'll be great to [23:09] get to a place where people are getting [23:11] information at their fingertips but in a [23:13] productive manner. Um, if I can provide [23:17] a clinical perspective from where we're [23:21] going and what I see happening in five [23:23] years. Um at the AMA we talk a lot about [23:26] this concept of augmented intelligence [23:29] and it's very aligned with what's been [23:30] said is we envision that a lot of things [23:33] will get automated but we really see [23:35] that as more of the administrative [23:37] functions understanding that the [23:39] clinical judgment that remains with all [23:42] of our clinicians is still going to be [23:43] at the center of how we deliver care and [23:46] the relationship between the physicians, [23:49] the nurses, all of our care providers [23:52] and the patient is at the center of how [23:54] we deliver care. I often tell parents [23:57] that I wish patients follow textbooks, [23:59] but they don't. And because they don't, [24:02] you need the clinical judgment to be [24:03] able to care for them appropriately. But [24:06] what I do see changing significantly is [24:10] we talk a lot about workflow, right, and [24:11] adoption. I see that changing. The [24:15] technology is going to liberate the [24:18] data, as I like to say, and truly [24:20] translate it into information. And so a [24:23] lot of the administrative functions that [24:25] we have today in healthcare like [24:27] notewriting, we're going to need to [24:28] start to question why do we do that? Why [24:31] do we need notes in a future world when [24:34] all of that data lives in our EHRs and [24:37] can be presented to us in real time? Why [24:40] do we have to enter 37 fields in a [24:43] report of something that went wrong when [24:46] you know event reporting is how we do it [24:48] manually now when now there's audit data [24:50] of everything that happened and we can [24:52] provide decision support in real time. [24:54] So those are some of the changes that I [24:56] think will happen rapidly and the way [25:00] that we work today will change [25:02] completely. What won't change is that [25:05] human relationship and the clinical [25:07] judgment that goes along with delivering [25:09] health care. [25:12] >> Thank you, Margaret. And maybe shifting [25:15] to the patient and whoever wants to take [25:17] this one first. How do we think around [25:20] permissioning? How do we think around [25:22] letting patients know that AI is part of [25:25] what their journey is? Um, do we let [25:28] them know that it's an AI that's calling [25:29] them when it's to book their [25:30] appointment? Um, what what's the right [25:32] amount of disclosure and permissions? [25:38] >> As an individual, when I get calls from [25:40] that are clearly like LLM voice systems, [25:43] I am infuriated. [25:45] I really really don't like it. Maybe [25:48] that's just a personal thing. Uh but [25:51] again as a as a patient if if doctors [25:54] are using or clinicians are using tools [25:56] like you know AI to like help them make [25:59] a diagnosis of whatever I've come into [26:01] the hospital or doctor's office for I I [26:04] don't necessarily feel like I need to [26:05] know about that. I I think that already [26:08] happens to some extent. Doctors before [26:10] LLM existed were already using machine [26:12] learning algorithms to help. They're [26:13] already going to Google to look things [26:15] up. They don't keep all of it in their [26:17] brain. They can't. um it just seems like [26:19] another tool in their tool belt to help [26:21] them make a decision. [26:25] >> I would say that fundamentally it goes [26:28] back to trust. [26:30] If you don't trust as a patient, you're [26:33] the whole process is skewed. You're not [26:34] going to trust. You're not going to want [26:36] to take the call. First of all, nobody [26:37] wants to take the call. So, I don't [26:39] think you're on your own. [26:40] >> Also true. [26:42] Um, but I think that trust and it to [26:46] gain to obtain that trust, you need the [26:48] transparency. So, I'm in favor of [26:50] awareness, knowledge. Yes, you know, you [26:53] need to know that you're going to get a [26:54] follow-up call, you're going to get the [26:56] chatbot texting you, whatever channel is [26:59] is you opt in, but at least you already [27:02] know and you're already prepared. Um I I [27:05] mentioned during when we met at first to [27:08] me I realized um like I fly United I [27:12] know I fly United I suffer a lot. So I [27:15] got a lot of texts say hey your flight [27:16] is delayed. Hey your flight is delayed. [27:18] I realized less two weeks ago, three [27:21] weeks ago that now they started putting, [27:23] hey, this message was generated by an [27:25] AI. Exact same message probably they [27:29] were using an AI. It just now it's been [27:31] disclosed, which is, hey, now you can [27:34] actually give me alerts a little bit [27:35] sooner because there's no manual [27:37] triggering. I can get a better [27:39] experience. Actually, because the AI [27:40] triggered, I could actually change a [27:42] flight that I got it was getting [27:43] delayed. So, I was able to move into an [27:45] earlier flight. there is an element of [27:47] okay now I know that it's I trust it's [27:50] the system yes it's not human but gives [27:52] me the enough information to take action [27:55] on it I I had the exact opposite [27:57] experience flying they sent me the [27:59] notification that the flight was [28:01] cancelled way too early it was not [28:03] canceled and then they had to resend [28:05] that like an hour later and it led to [28:07] massive confusion so I guess it's just a [28:09] double-edged sword [28:10] >> yeah it's humans humans are prone to it [28:13] too right so of first [28:16] to you. [28:17] >> Um, I think that transparency is never [28:21] wrong. If we really had to think about [28:24] how we do this, if we're going to be [28:26] open about which tools we're utilizing, [28:29] and I 100% agree, we've been using AI in [28:31] healthcare for decades now, right? This [28:33] is not new. We haven't always told our [28:36] patients every technology that's used. [28:39] And yet now that it's touching them much [28:41] more directly, I think being transparent [28:43] about it makes a lot of sense. And back [28:46] to our conversation about responsible [28:49] utilization of AI tools. If we do this [28:52] well and the questions get escalated [28:56] appropriately to clinicians and clinical [28:58] judgment remains with our clinical [29:02] experts then you will gain the trust of [29:05] these tools because at the end of the [29:08] day the responses that our patients will [29:10] be receiving will be the ones that are [29:12] appropriate for their healthcare. So I [29:14] think it's about the how we do this, not [29:17] necessarily a yes or no about the [29:20] release of whether we're using the [29:23] tools. [29:25] >> So there there are two aspects to [29:28] consent. Uh one is the use of AI for [29:32] treatment, right? The other is use of [29:36] data for development of AI. And I think [29:41] we're missing the conversation of [29:43] consent for the latter. Right? The [29:46] former is required and I think us as an [29:49] organization we'll get past that past [29:51] that very quickly. Right? Uh we'll [29:53] provide a consent in the form that you [29:55] signed a year ago and which you didn't [29:57] read and uh and and and uh and I think [30:01] we'll be okay and organizations will be [30:02] safe. Honestly, I think that's less of [30:04] an issue. There is a small issue there [30:06] around uh uh training our clinicians to [30:10] explain what that means. We've missed [30:12] completely missed out on that. Right? Uh [30:15] you know we are uh the legal teams are [30:17] getting together and they're saying [30:18] we're going to have a form and people [30:19] are going to sign it or uh the physician [30:21] at the time of the treatment has to say [30:23] that AI is going to be used but we [30:25] haven't taught the clinician how to [30:26] explain what AI is. Right? So there's a [30:28] missing we have missed that boat there. [30:30] So I think I think that but we'll get [30:32] past that. I think uh and also I think I [30:34] think we're we're uh I think us as a [30:37] society we're getting a lot better in in [30:40] accepting this technology for use for [30:42] care right uh you know I I I have a I [30:46] have a elderly relative that I that I [30:48] care for and I still remember that you [30:50] know precoid I I used to ask her if she [30:53] wants to go to the do she wants to take [30:54] a tele visit and she would say that I [30:56] want to see my own doctor like I've been [30:58] seeing it for 20 years and that's what I [30:59] want to do postco the same person the [31:03] same patient says can I take a teley [31:05] visit instead of going to the doctor [31:07] that's a good tell telling that you know [31:09] people are getting comfortable with [31:11] technology so I think we'll get past [31:12] that the harder I think where consent is [31:15] absolutely required and my background is [31:17] data officer and where I cringe is is [31:21] when we start to use [31:23] patient data for product development [31:27] that consent has to be explicit and that [31:30] is really important as as a society are [31:33] are shying away from that problem. [31:34] Hackinsack is not, right? But other [31:37] organizations are looking the other way. [31:39] Uh and that's a really hard conversation [31:40] to have. Why is that important? Because [31:43] you are going to end up giving your data [31:45] to a vendor that that you got because [31:48] the capability was good and the numbers [31:49] made sense. That organization is going [31:52] to have a breach, [31:54] right? And we as us as health systems [31:58] that are providing that data to that [32:00] vendor are going to get caught and you [32:02] as a patient are going to come in and [32:03] say I did not consent to you [32:06] you know giving my data for product [32:08] development for profit. That's that's [32:10] the consent piece which people aren't [32:12] talking about and I think that's what [32:14] needs to be talked about right and that [32:16] I mean I'll tell you that that [32:17] conversation doesn't exist in 99% of [32:19] health systems today right and vendors [32:21] are stealing our data right uh and large [32:25] v vendors who who who have that muscle [32:27] are coming to say you don't like you [32:29] know you don't like it don't use our [32:30] product and our and they're putting [32:32] pressure on clinicians to sign up for it [32:34] and then we end up signing up for it [32:35] with a risk to the organization so I [32:37] think that consent is going to take some [32:40] In terms of that, I've heard of a great [32:42] vendor called Tonic AI which can help [32:44] you with that [32:46] you know you know the problem is is yes [32:48] you can but it's it's the vendor right [32:50] it is I mean I mean let's without [32:53] without putting any names out there you [32:55] know you know we we have a vendor today [32:58] that we are using for am ambience right [33:00] so note takingaking and the vendor says [33:03] fine you don't want to use us don't use [33:04] us if you don't use us we're going to [33:05] use your data for product development [33:06] Right. [33:09] So there's I mean I wish they would use [33:10] tonic [33:12] >> but they they but but there's no [33:13] conversation using tonic. Right. [33:16] >> There's the legal aspect of like does [33:17] has the pat is is the paperwork in place [33:19] for the patient to have given consent to [33:20] use the data. But then there's the other [33:22] thing which even the vendor cannot [33:23] escape which is HIPPA. So if they're [33:26] going to use this data for example for [33:28] model training they absolutely have to [33:30] adhere to the HIPPA guidelines for [33:31] deidentification. Um if they're going to [33:33] use that data. Now do all vendors do [33:35] this to the tea? Uh no, maybe not. Uh do [33:38] you know most I think make a good effort [33:40] at least. But it's a very tricky [33:42] problem. Um and I am skirting the actual [33:44] consent part and talking just about like [33:46] the regulatory aspect of HIPPA. But at [33:48] least HIPPA is there to you know force [33:52] companies to behave to some extent. [33:54] >> So I mean if you if vendors do HIPPA [33:57] they follow safe harbor right? Safe [33:59] harbor is no longer relevant right. I [34:02] mean in today's compute world there's so [34:04] much comput available reidentification [34:06] of data is like milliseconds away right [34:10] >> they can do an expert determination at [34:12] least [34:13] >> yeah and no one does that right that's [34:15] the issue I go to vendors and I h I tell [34:17] them are are if if and and they by the [34:19] way have the vendors say uh you know if [34:21] I do this using safe harbor uh the data [34:24] is no longer usable right no utility [34:26] left [34:27] >> I tell them okay again go for expert [34:28] determination and they tell me what is [34:31] that that's right. If they had just kept [34:33] reading the guidelines, two more [34:34] sentences, they'd see the expert [34:36] determination line. Um, and look, it's [34:38] more painful than a a safe harbor, but [34:40] it does um it takes a little more time, [34:42] but honestly, it's never as much as you [34:44] think. You can typically get an expert [34:45] determination letter in in weeks, not [34:47] months. Um, but you're right, it it [34:49] takes effort and for a company that's [34:51] trying to develop a product or a startup [34:52] that's trying to iterate as quickly as [34:54] possible, they might see it as a [34:55] roadblock and not as something that's [34:56] actually there to help them and to [34:58] protect their patients. [35:00] It's going to bring this back to the [35:01] patient experience because you talk [35:02] about the consent like I think we the [35:06] whole industry is losing that [35:07] opportunity to have a better approach to [35:10] the consent. It gets it's the same thing [35:13] the terms and conditions of all of the [35:14] softwares we signed up online, right? [35:15] It's pages and pages everybody just [35:17] scrolls to the end and then nobody read [35:19] and then they made you actually the [35:21] JavaScript that you have to scroll to [35:22] the end still don't read but you need to [35:24] scroll down to be able to say okay I [35:26] read it. So now the CYA it's you read it [35:30] you accept it and there's a way there [35:33] the needs to be a way of changing this [35:35] right the patient experience of hey [35:38] changing training the staff not the just [35:40] the clinicians but the people at the [35:41] gathering the forms at the beginning and [35:44] making it easier even at the your pre [35:46] previsit pre-authorization you go on to [35:49] name your EMR portal right most of us [35:52] use one around the US and it's exactly [35:56] the same thing you pages. You have no [35:58] idea. Nobody who just out of curiosity [36:01] who's ever had read their disclosure of [36:04] data here in the room when they were [36:05] signing up for the portal online. [36:09] >> Someone over there has. [36:11] >> All right, [36:13] sir. I'm impressed. [36:16] I I bought a new TV uh just two days ago [36:18] and when you when you turn it on out of [36:20] the box experience, there's like 50 [36:22] legal documents that you have to scroll [36:24] through and read. It's a joke. Yeah, [36:26] you're completely right. [36:28] >> But everybody become that because of [36:30] that lack of experience and how we just [36:33] keep the legal ease out there, it [36:35] becomes a compounding problem over and [36:37] over as we go through the process, have [36:39] the tools and nobody knows and then [36:41] because you don't have that clarity even [36:44] people that are familiar or involved not [36:47] familiar but involved in the process [36:48] they might not be familiar with the [36:49] documentation that the boundaries just [36:51] like you said they didn't read all of [36:53] the documentation. You can't do this. So [36:55] you don't do it and then you you stop [36:58] your the product development, you stop [37:00] insights, [37:01] >> right? [37:02] >> I think it's in all of us to educate the [37:06] patient community, the consumers, that [37:08] is all of us on what this means to them. [37:12] And I do think there are health systems [37:14] that have done this well. Um I don't [37:17] know if I would have, for example, the [37:19] physicians be the ones that explain this [37:21] to the patient. that's probably not the [37:23] best use of their time. And there's [37:25] often a lot of dead time when people are [37:28] coming into the clinic or even in the [37:30] hospital when people are sitting in [37:31] their beds and they don't have a lot to [37:33] do. [37:35] A lot of organizations are using that [37:37] time and using technology in an [37:40] interactive way to help educate the [37:42] patients about some of these really [37:44] important topics of what their [37:46] healthcare data security means to them. [37:49] And I the comments about the fact that [37:52] there's really no more deidentified data [37:54] is right. We have to help patients [37:56] understand that particularly those with [37:58] unique conditions that when their data [38:01] is out there, it is not going to be [38:04] deidentified in a way that will not be [38:06] tied back to them. So I it we as leaders [38:10] in this space um need to be very [38:12] proactive in making sure that we are [38:14] thinking through this and creating these [38:16] tools for our patients. [38:20] Thank you. Uh, on a related note, if [38:22] you've never asked your LLM of choice to [38:26] give you a detailed breakdown of [38:28] everything it knows about you, homework [38:30] for the audience, it is eyeopening um, [38:34] in many ways. So, um, I know we we only [38:36] have 5 minutes left, so would love to [38:39] open it up to you all. And maybe just [38:41] can you if there's one piece of advice [38:44] you want the crowd to leave with, what [38:46] would that be? anything we didn't touch [38:48] on that you think is really important. [38:50] If you guys could each touch on that [38:51] would be great. [38:55] >> Oh man, I wasn't ready for that. Um, if [38:58] I could leave the audience with one [39:01] piece of information, I think I would [39:03] say that for those in the audience that [39:06] are trying to train models on customer [39:10] data or patient information, um, make [39:13] sure you have a plan in place ahead of [39:16] time for how you're going to deal with [39:17] the privacy aspect. You you don't want [39:20] to get caught later, you know, down the [39:23] the the the creek without a paddle. I [39:25] forgot the expression, but I think [39:26] everyone knows what I mean. You don't [39:28] want to be caught unaware later when you [39:29] realize, oh my god, everything we've [39:31] just done, we can't actually use. [39:32] There's customer PHI and the weights of [39:34] our model or in this environment where [39:36] it shouldn't be. Just make sure you have [39:37] a plan in place that involves some way [39:40] of retaining the privacy of your [39:42] patients or your customers. [39:46] If I could give one piece of advice to [39:47] anyone that is developing software in [39:50] healthcare, whether it's AI or not, I [39:53] would say make sure that those closest [39:56] to the care are part of the [39:57] conversation. We have to have clinicians [40:00] at the table. We have to make sure we [40:02] understand what their needs are. What [40:03] are the problems you're trying to solve [40:05] and partner with them to create [40:08] solutions that will be useful to them. [40:12] The technology experts understand [40:15] software, but the clinicians really know [40:17] the day-to-day of delivering care. [40:22] >> I would say we'll leave everybody with [40:24] the one size does not fit all. Um the [40:28] solutions, the standards, the large [40:30] language models out there, they're great [40:32] generalistics. They're great in the [40:35] broader bigger context. They're great to [40:38] search information like you mentioned, [40:39] Adam, but the context of your patient, [40:42] your demographics, your reality, your [40:44] microcosmos of patients, your reality [40:48] that your patients are inserted in, they [40:50] will not be specific to that. So data is [40:53] a key component to it that you need to [40:57] factor in and layer on top of the [40:59] generalist the general models so that it [41:02] becomes useful meaningful to your [41:04] physicians to the treatment and to the [41:06] patients you're seeing. [41:08] >> Um my advice is I think it's important [41:11] to continue educating organizations of [41:15] what this is and what this isn't. Uh you [41:18] know I think this stuff is coming so [41:20] fast at us. It's really hard for I think [41:22] and I sometimes feel bad for the [41:24] consumers to really appreciate what tool [41:27] they should be using and how they should [41:28] be using it right uh so I think [41:30] educating organizations your leaders of [41:33] what this is and what this isn't is [41:35] really important grounding on the fact [41:36] that this is tech it relies on data it [41:40] is not magic right is is really [41:42] important um you know I you know there's [41:45] a lot of conversation about you know the [41:47] likes of you know Chad GBT and putting [41:50] you know your information in there and [41:52] seeing what you get out of it, right? I [41:54] think that's one of the biggest problems [41:57] is is the concept of AI has become uh a [42:01] box that you can you can type into and [42:04] you get a response in that's a really [42:07] bad representation of AI. You know, I uh [42:10] I uh you know, I I've done this in the [42:13] past where I have said the likes of Chad [42:15] GPT and I'm at some point I want to get [42:17] in trouble for this, but uh you know, [42:19] the likes of Chad GBT to me is is like [42:23] the high fructose of corn syrup like of [42:25] sweeteners, right? That's what the likes [42:28] of Chad GBTR are. Why? Because the [42:29] reason I use that analogy that likes of [42:32] Chad GBTR, high fructose corn syrup or [42:34] sweeteners is because it gets the job [42:35] done. It's really sweet, but it's not [42:36] good for everybody. So it's not good for [42:38] every use case. So us as technologists, [42:42] we have to focus on making sure we have [42:45] two options and I'm just talking [42:46] generalizing for generative AI, right? [42:49] Is either teach your users prompt [42:51] engineering or hide a prompt behind a [42:54] button. Please do this, right? Otherwise [42:58] you if you allow people to just type a [43:00] question, they will type the question [43:03] not in the right way and they're not [43:04] going to get the right answer. Right? So [43:07] please educate your your organizations [43:09] of what AI is and what AI isn't and how [43:11] to use it and prompt either train them [43:15] to to to write better prompts or write [43:18] prompts for them and hide it behind the [43:20] button and generate and that'll get you. [43:22] I think that's that's going to lead us [43:24] to success. Us us giving them yet more [43:27] boxes that they can type something into [43:29] is not going to get us ROI. [43:32] >> I'll take that one with me. [43:36] Well, I think we are just running out of [43:38] time. Thank you so much. Um, could [43:41] everyone please give a round of applause [43:42] for our wonderful panelists? [43:44] [Music]