GE HealthCare's Beger Leverages AI Models to Cut Medical Error Rates

GE HealthCare的Beger借助AI模型降低医疗差错率 医疗系统必须部署人工智能以实现常规工作流程自动化，并消除全球严重的临床医生短缺问题。 Jan Beger概述了困扰全球医疗的严峻结构性危机，包括医疗数据的指数级增长、居高不下的医疗差错率，以及普遍存在的临床医生职业倦怠和劳动力短缺。他展示了AI如何显著提高诊断效率，其应用涵盖深度学习图像重建、即时移动分诊，以及LLM驱动的病历总结和自主扫描筛查等。Beger强调了GE HealthCare在监管批准方面的领先地位，并指出尽管技术能力已获证实，但实际应用普及依然滞后。最终，他主张将日常行政琐事自动化，从而让临床医生能够重新将精力专注于直接的患者护理。 全球医疗数据正以每72天翻倍的惊人速度增长，且占到全人类总数据量的30%，但医疗机构产生的数据中高达97%从未被有效利用。 (0:04) 全球每年因医疗安全问题导致约300万人可预防性死亡；同时医疗体系正面临严重的人力荒，世卫组织预计到2030年医护人员缺口将扩大至1100万。 (4:48) 尽管OpenAI的o1等顶尖AI模型在MedQA测试中以96.5%的高分超越资深医生，且FDA已批准超1000款AI医疗设备（GE HealthCare以100款领跑），但实际临床工作流的应用依然滞后。 (8:04) 移动床旁X光AI自动分诊气胸、欧盟批准首个自主排除正常胸片的AI算法以及掌上超声智能引导等落地应用，正加速推动临床流程从繁琐的EMR录入向自动化与多模态交互转型。 (12:56)

Jan Beger of GE HealthCare argues that healthcare faces severe structural crises—unmanageable data growth, high rates of patient harm, and acute workforce shortages—that make integrating validated artificial intelligence an urgent clinical and ethical priority. Data Explosion and the Utilization Paradox (0:04) Healthcare data is expanding at an unprecedented rate, doubling roughly every 72 days compared to every 3.5 years previously, and now accounting for approximately 30% of global data volume. Beger notes that keeping pace manually is impossible: a new medical paper appears every 26 seconds, requiring general practitioners to read 21 hours daily to stay current. Imaging alone constitutes 90% of healthcare data, with modern procedures like lower-body CT angiography generating up to 3,000 images, and busy radiologists reviewing up to 50,000 images in a single 12-hour shift—leaving less than one second per image. Despite this sheer volume, health systems suffer from a severe utilization paradox. An average hospital generates approximately 50 petabytes of data, yet 97% of all healthcare data remains unleveraged, leaving health systems utilizing just 3%. Facilities shoulder substantial financial and logistical burdens to store tens of petabytes of information without translating it into actionable clinical outcomes. Systemic Risks in Safety and Staffing (4:48) Beger contrasts the safety of modern healthcare with commercial aviation, noting that the risk of harm during a medical journey is 1 in 10, compared to 1 in 1,000,000 during a commercial flight. Globally, unsafe medical care results in approximately 3 million preventable deaths each year. Concurrently, health systems face compounding labor deficits and operational strain. The World Health Organization projects the current shortage of 7 million healthcare workers will increase to 11 million by 2030, while 4.5 billion people worldwide lack access to proper healthcare. In Canada, patients face average MRI wait times of 13 weeks. Among active staff, 60% of doctors spend more time dealing with electronic medical records (EMRs) than with patients, half of radiologists report professional burnout, and healthcare organizations experience a 100% staff turnover rate every five years on average. Algorithmic Capability Versus Real-World Deployment (8:04) Modern artificial intelligence demonstrates high clinical benchmark capabilities. OpenAI's o1 model, for example, achieved a 96.5% accuracy rate on the MedQA exam, surpassing experienced physicians (87%) and recent medical graduates (60%). However, Beger stresses that standardized, controlled testing differs substantially from messy real-world clinical practice, which requires thorough validation to avoid patient harm. On the regulatory side, medtech firms outpace general tech developers. The US FDA has cleared over 1,000 AI/ML-enabled medical devices, mostly in radiology, with GE HealthCare leading at 100 clearances. Yet, despite this growing catalog of authorized tools, Beger observes an adoption gap: widespread clinical utilization remains low, and many radiologists have not embedded AI into routine daily workflows. Clinical Applications Across Care Workflows (12:56) Real-world AI deployment is expanding across diagnostics, administration, and patient interaction. At the bedside, AI algorithms embedded in mobile X-ray units rapidly identify critical conditions like pneumothorax and verify endotracheal tube positioning, which is crucial where 24/7 specialist coverage is unavailable. In diagnostics, deep learning image reconstruction has scanned over 50 million patients globally, reducing scan times and enabling free-breathing cardiac MRIs, while miniaturized handheld ultrasound units employ AI to guide probe manipulation for non-expert clinicians. The European Union has also granted CE mark authorization for an autonomous AI algorithm to filter out normal chest X-rays from diagnostic queues. To counter clinician burnout, ambient scribes and natural language processing synthesize messy EMR charts—including handwritten notes and lab results—into interactive, single-page summaries with clickable source references. Natural language tools also act as pathology co-pilots and power conversational platforms that support patients, addressing the roughly 50% dissatisfaction rate regarding provider communication. Because healthcare relies on uniquely complex multimodal data—from images and vitals to historical records—Beger concludes that modern foundation models are uniquely suited to synthesize this information. Given their performance in reducing administrative overhead and catching critical conditions, Beger poses the unresolved ethical question of whether it remains defensible for health systems to delay deploying validated AI tools.

GE HealthCare的Jan Beger指出，现代医疗体系正遭遇数据指数级过载、医疗安全风险高企与全球医护资源严重匮乏的三重结构性危机；尽管AI算法在基准测试与监管获批上进展迅速，但临床应用的核心使命在于分担繁重的流程与行政负担，使医护人员能够重构诊疗专注度并弥合服务鸿沟。 海量数据激增与仅3%利用率的悖论 全球医疗数据正以惊人的速度膨胀，其翻倍周期已从数年前的3.5年急剧缩短至约72天，整体规模已占全球数据总量的30%左右。文献与知识的更新同样超越了人类承受极限：全球平均每26秒便发表一篇新医学论文，皮肤科医生若想掌握前沿需每年精读3,900篇专业文献，而全科医生每天需阅读21小时才能跟上初级保健领域的知识迭代。这种数据膨胀在医学影像领域尤为突出，影像数据占据了涵盖放射、皮肤、眼科及病理等全部医疗数据的90%。例如，单次下肢CT血管造影即可生成多达3,000张高分辨率图像，单次乳腺钼靶检查的数据量甚至超过了传统的印刷版纽约电话簿。 然而，与数据剧增形成鲜明对比的是极其低下的利用效率。普通医院平均需要长期存储约50 PB的数据，承担着沉重的存储与归档财务负担；同时，繁忙的诊断放射科医生在单个12小时班次中需审阅多达50,000张图像，平均每张图像的停留时间已不足一秒。Beger强调，目前医疗系统中高达97%的生成数据完全处于未被开发和利用的闲置状态，临床实际调用的数据仅占3%，造成了严重的资源浪费与认知过载。 医疗差错隐患与临床人力枯竭危机 (4:48) 在医疗质量与患者安全维度，医疗差错问题十分严峻。Beger援引统计对比指出，商业航空出行的受伤害概率约为百万分之一，而患者在医疗就医流程中遭受伤害的概率却高达十分之一；尽管此类伤害并不等同于致死，但全球每年因不安全医疗导致的死亡人数已达约300万人。这一惊人数字表明，临床安全监控与错误规避机制存在亟待填补的体系漏洞。 与此同时，全球医疗人力资源的短缺与从业者的职业倦怠正将系统推向极限。由于繁重的电子病历录入工作，60%的医生在面对电脑屏幕系统的时间已超过了与患者面对面沟通的时间。高压环境下，约50%的放射科医生报告存在职业倦怠，医疗机构平均每5年就会经历一次高达100%的人员流失流转。供需失衡直接导致患者等待时间被大幅拉长，例如加拿大患者预约一次MRI平均需要等待13周。根据世界卫生组织（WHO）的预测，全球医护人员缺口将从目前的700万人扩大至2030年的1,100万人，全球更有45亿人口目前仍无法获得充分的基本医疗服务。 模型评测突破、合规进展与临床采纳滞后 (8:04) 虽然医疗行业出于患者安全和审慎验证的原则，历史上采纳新技术的步伐相对缓慢，但过去四到五年中AI在计算机视觉、大语言模型和预测分析领域的应用正在加速。在标准化能力评测中，先进模型表现出了超越人类常规水准的测试成绩。例如，OpenAI的o1模型在医学测试基准MedQA中取得了96.5%的准确率，不仅显著高于医学院应届毕业生的约60%，也超过了资深从业医生约87%的平均得分。不过Beger提醒，标准化单选题测试所处的受控环境与复杂、不可预测的真实临床现场存在本质区别。 在合规上市方面，专业医疗科技企业展现了主导地位。截至目前，美国FDA已批准或授权了超过1,000款启用AI/ML的医疗器械，其中GE HealthCare以100项授权位列行业首位，获批产品绝大部分集中在放射与医学影像分析领域。但Beger指出，在大量算法通过安全验证并进入市场的情况下，实际临床日常诊疗的渗透率仍然滞后，全球仍有大量放射科医生在常规工作流中尚未正式引入任何AI辅助工具。 实战落地应用：床旁分诊、影像重构与环境智能 (12:56) AI正从边缘实验进入实质性的临床工作流。在急诊与移动床旁X光设备上，内置算法能够在拍摄瞬间自动排查气胸或评估气管插管位置等危急病症，在缺乏全天候放射专科医生的资源匮乏地区实现自动优先级分诊。在流程自动化方面，欧盟CE认证已批准首款用于自主排除正常胸部X光片的AI算法，直接将无异常影像从放射医生的手动阅读队列中剥离。针对磁共振检查，深度学习图像重建技术已应用于全球超5,000万名患者，不仅大幅压缩扫描时间，更实现了无需重复屏气的心脏MRI自由呼吸扫描；小型手持超声亦借助AI姿态引导技术，帮助非专业操作人员规范采集诊断级影像。 在文档与人机交互领域，AI正逐步化解医护人员的系统录入负担。病理学多模态副驾驶允许病理学家以自然语言直接向高分辨率切片提问；电子病历图表生成工具能够将包含手写笔记、检验表格与既往报告的杂乱记录，实时浓缩为带有一键溯源功能的单页交互式摘要；环境记录工具则允许医生通过口语化对话完成数据管理。此外，针对约50%患者对医患沟通不满意的问题，基于乳腺癌康复日记训练的患者端对话模型正在提供7x24小时的院外共情辅导。 Beger在总结中强调，医疗健康是极其典型的多模态行业，横跨影像、结构化记录、实时体征与实验室数据，现代基础模型首次提供了整合分析这些异构数据的可能。鉴于AI工具在差错预防与诊疗提效上已展现出的验证价值，他提出一项行业反思：在技术具备可行性的当下，延缓引入AI是否已构成某种伦理问题。AI在医疗中的终极价值并非替代医生，而是通过接管机械繁琐的行政杂务，将宝贵的人力与时间归还给真正需要人性关怀的患者照护过程。

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 24 日发布，主讲人为 GE HealthCare 全球 AI 倡导负责人 Jan Beger。演讲主题为《Human. Machine. Care.》，页面摘要指出该演讲探讨了 AI 如何应对医疗数据爆炸、支持复杂临床决策并缓解全球医护短缺，同时强调技术创新背后的核心在于人与信任。 ## TL;DR 演讲首先指出全球医疗数据每 72 天翻倍、97% 数据闲置以及全球医护缺口将在 2030 年达 1,100 万人等严峻行业挑战。随后展示了 AI 在基准测试中的突破表现，以及 GE HealthCare 等老牌医疗器械巨头在 FDA 获批数量上的领先地位。Jan Beger 详细介绍了移动 X 光端侧危急预警、大语言模型病历一页纸摘要与共情沟通、深度学习 MRI 自由呼吸扫描及手持超声手法引导等落地场景。最后，演讲强调了多模态医疗基础模型与环境智能的未来方向，并提出当 AI 能实质挽救生命时，不用 AI 或将引发全新伦理反思。 ## Atmosphere & style 整场演讲呈现出严谨沉稳、富有洞见且充满人文关怀的科技行业分享氛围。主讲人通过极具冲击力的数据对比建立起医疗系统深陷负荷危机的紧迫感，随后以大量扎实的软硬件落地用例将叙事引向清晰务实的破局希望。现场语调平和专业，既没有硅谷式的过度技术狂热，也没有传统医疗的保守迟疑，最终升华至探讨技术赋能医患关怀与伦理责任，带给观众一种兼具现实厚度与前瞻格局的启发感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=qIQHYeysCLU&t=4s) 全球医疗数据翻倍周期已缩减至 72 天且占全球数据总量的 30%，海量医学文献与影像信息过载已彻底突破人类医生的生理阅读极限。 - [02:00](https://www.youtube.com/watch?v=qIQHYeysCLU&t=120s) 医疗系统中 90% 的数据集中在医学影像领域，但其中 97% 的数据长期处于闲置未利用状态，形成沉重的存储合规负担与未开采价值。 - [04:48](https://www.youtube.com/watch?v=qIQHYeysCLU&t=288s) 全球医疗安全风险高企（就医受伤害概率达十分之一），加之预计到 2030 年全球将面临 1100 万医护人员短缺，临床供需失衡亟待 AI 破局。 - [08:04](https://www.youtube.com/watch?v=qIQHYeysCLU&t=484s) OpenAI o1 等前沿大模型在 MedQA 医学基准测试中取得 96.5% 的高准确率，同时美国 FDA 已累计批准超 1000 款 AI/ML 医疗设备，合规落地加速推进。 - [11:34](https://www.youtube.com/watch?v=qIQHYeysCLU&t=694s) 传统医疗器械巨头（如 GE HealthCare 斩获 100 项 FDA 认证）在医疗 AI 合规赛道领跑，但临床一线医生的日常实际采纳率仍存在明显滞后。 - [12:56](https://www.youtube.com/watch?v=qIQHYeysCLU&t=776s) 移动 X 光机端侧 AI 能在床旁曝光后数秒内自动识别气胸与插管位置异常，为缺乏 7×24 小时放射科专家的夜间与偏远基层医疗提供关键生命防线。 - [14:35](https://www.youtube.com/watch?v=qIQHYeysCLU&t=875s) 大语言模型（LLM）不仅展现出高同理心的患者心理支持能力，还可实现影像报告初稿生成、异构病历一键提炼与带溯源超链接的交互摘要。 - [17:13](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1033s) 深度学习图像重建技术彻底革新了 MRI 检查体验，不仅大幅缩短扫描耗时，还让心脏检查实现自由呼吸，全球已有超 5000 万名患者受益。 - [18:00](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1080s) 内置 AI 实时视觉引导打破了便携手持超声的高手法依赖壁垒，使非专家级基层与全科医生也能规范采集标准切面影像。 - [19:01](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1141s) 随着环境感知记录与多模态医疗基础模型重塑工作流，行业伦理转向探讨“在 AI 明确能减少漏诊时不用 AI 是否构成不道德”，其核心使命是将医生还给病床与患者。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=qIQHYeysCLU&t=4s) 医疗数据爆炸：行业面临的极限负荷与信息过载 演讲伊始，Jan Beger 首先向现场观众确认了有多少人从事广义的医疗健康领域，并指出接下来的分享将聚焦于医疗行业当前面临的几大核心难题，以及人工智能（AI）如何作为一项广义技术为这些挑战带来解决方案。他列出的第一个震撼事实是：如今全球所有医疗数据的翻倍周期已经缩短至仅仅 72 天左右，而在短短数年之前，这个周期还是三年半。目前，全球总体数据量中有高达 30% 属于医疗健康相关数据，这直观展现了该行业的庞大体量与复杂程度。 医疗知识的更新速度已经彻底突破了人类医生的生理极限。全球平均每 26 秒就有一篇新的医学论文发表；这意味着一名皮肤科医生每年需要阅读 3,900 篇科研论文才能跟上本领域的最新文献；而对于全科医生（家庭医生）来说情况更为夸张，他们每天必须不吃不喝阅读 21 个小时才能掌握初级医疗文献的更新步伐。在影像方面，一张乳腺钼靶（Mammogram）影像所包含的信息量甚至超过了整本纽约电话黄页；一次下肢 CT 血管造影即可产生多达 3,000 张断层图像。一家普通医院每年产生的数据量可达 50 PB。 由于像 GE HealthCare 这样的医疗设备制造商不断研发出分辨率更高、层厚更薄的先进影像设备，诊断放射科医生在长达 12 小时的轮班中，平均需要阅片多达 50,000 张（几年前这一数字仅为 500 张左右）。算下来，放射科医生审视单张影像的时间甚至不足一秒。然而，在医疗系统产生的海量数据中，90% 集中于医学影像（涵盖放射、皮肤、眼科、病理等），而其中竟然有 97% 的数据完全处于闲置、未被利用的状态，对改善整体医疗效果毫无贡献，实际被有效利用的数据仅占 3%。海量数据对医疗机构而言既是归档与存储的沉重合规负担，但在 AI 的加持下也是蕴含巨大价值的未开采金矿。 - **数据翻倍周期**：从数年前的 3.5 年急剧缩短至当前的 72 天。 - **全球医疗数据占比**：占全球总数据量的 30%。 - **文献产出速度**：每 26 秒发表一篇新医学论文；全科医生跟进文献需每天阅读 21 小时。 - **放射阅片负荷**：放射科医生 12 小时排班需阅片约 50,000 张，平均单张审阅时间不足 1 秒。 - **数据利用率**：医疗系统产生的数据有 97% 未被有效利用，仅 3% 真正发挥了作用。 --- ### [04:48](https://www.youtube.com/watch?v=qIQHYeysCLU&t=288s) 医疗安全与全球医护人员短缺危机 ```mermaid flowchart TD accTitle: 医疗体系危机与供需缺口放大模型 subgraph Challenges["医疗系统三大核心挑战"] D["数据过载与信息利用率低\n(97% 数据闲置 / 72天翻倍)"] S["医疗安全与差错风险\n(1/10 伤害率 / 每年300万死亡)"] H["医护资源短缺与职业倦怠\n(全球缺口700万-1100万)"] end subgraph Demographic["人口与需求结构变化"] A["全球人口老龄化加剧"] C["慢性病患病率持续攀升"] U["45亿人缺乏基本医疗保障"] end subgraph Gap["供需失衡加剧"] G["供给与需求之间形成巨大断层"] end subgraph Solution["破局路径"] AI["引入经过临床验证的 AI 技术\n赋能临床决策与流程自动化"] end Challenges --> G Demographic --> G G --> AI ``` 在讨论完数据过载后，Jan Beger 转向了医疗质量与安全问题，并给出了一个令人警醒的对比：普通人乘坐飞机遭受伤害的概率仅为百万分之一，而当一个人进入医疗救治流程时，遭受某种形式伤害的概率竟然高达十分之一（"one in 10"）。尽管“伤害”并不必然等同于死亡，但全球每年因不安全医疗而导致的死亡人数高达 300 万人。 与此同时，全球医疗体系正深陷医护人员短缺与严重职业倦怠的泥潭。数据显示，60% 的医生把更多的工作时间花在面对电子病历系统（EMR）录入数据上，而非与患者进行面对面的直接沟通；约 50% 负责审阅 MRI 和 CT 扫描的放射科医生处于严重的职业倦怠状态；医疗系统中每 5 年就会发生 100% 的全员人员流动率。在加拿大等地区，患者为了预约一次 MRI 检查平均需要等待 13 周。 根据世界卫生组织（WHO）的统计，目前全球医疗专业人员的短缺缺口已达 700 万人，预计到 2030 年这一数字将进一步恶化至 1,100 万人。雪上加霜的是，全球人口老龄化进程加速，慢性病发病率不断上升，全球还有 45 亿人无法获得应有的正规医疗保障。这种医疗“供给”与“需求”之间的鸿沟正在以惊人的速度扩大。Jan Beger 强调，唯一切实可行的破局之道就是尽快将经过严格临床验证的 AI 等前沿技术落地到医疗系统中，帮助医护人员分担工作压力。 - **医疗不良事件风险**：就医受伤害概率达 1/10（对比航空出行的 1/1,000,000）；全球每年因医疗安全问题致死约 300 万人。 - **临床精力分散**：60% 的医生在 EMR 界面前消耗的时间超过了与患者面对面交流的时间。 - **职业倦怠与流动**：50% 的放射科医生处于倦怠状态；医疗系统每 5 年员工流失轮换率达 100%。 - **全球人力短缺预测**：世卫组织预计医疗专业人员缺口将从目前的 700 万人增长到 2030 年的 1,100 万人。 --- ### [08:04](https://www.youtube.com/watch?v=qIQHYeysCLU&t=484s) AI 赋能医疗的广阔前景与基准测试表现 虽然医疗行业历来因为严格的安全验证和避免伤害患者的准则，在新技术采纳上表现得相对谨慎缓慢，但在过去四到五年间，全球医疗体系对 AI 的关注与落地速度迎来了爆发式增长。从顶层的大语言模型（LLM）驱动的临床文本处理、病历整理，到计算机视觉辅助的医学影像筛查与诊断，AI 的应用场景正在全方位铺开。 Jan Beger 以权威的医学问答测试基准 MedQA 为例展示了 AI 技术的突破：OpenAI 的 o1 模型在标准医学问答测试中取得了高达 96.5% 的准确率。为了建立直观参照，他对比了人类医生的表现：一名经验丰富的资深医生在该项基准中的平均准确率约为 87%，而刚从医学院毕业的年轻医生准确率仅在 60% 左右。尽管这种标准化测试与复杂的临床真实世界尚不能直接等同，但它无可辩驳地证明了先进 AI 模型在处理标准化临床专业知识时已经达到甚至超越了人类专家的水准。 无论是在急诊室的分诊与流程优化，还是在放射与病理的精准诊断环节，AI 在多种受控评估环境中的表现都展现出了巨大的实用价值。Jan Beger 同时也指出，真实的医疗环境极其复杂多变，基准测试的高分只是第一步。值得关注的是，医疗这一强监管行业的监管步伐也在加快，截至目前，美国 FDA 已经批准了 1,000 多项基于 AI/机器学习（AI/ML）赋能的医疗设备，标志着 AI 正在以合规、安全的方式全面融入临床实践。 - **MedQA 标准医学测试基准对比**： - OpenAI o1 模型：96.5% 准确率 - 资深临床医生（平均）：87% 准确率 - 医学院应届毕业生：约 60% 准确率 - **监管审批里程碑**：美国 FDA 迄今已获批超过 1,000 款 AI/ML 赋能的医疗器械与软件产品。 ### [11:34](https://www.youtube.com/watch?v=qIQHYeysCLU&t=694s) 传统医疗器械巨头的AI领跑与临床落地的现实差距 在医疗人工智能领域，目前绝大多数应用场景都集中在放射学以及基于计算机视觉的医学影像分析上。很多人通常会下意识地认为，引领医疗 AI 浪潮的必定是硅谷科技巨头（Big Tech）或是那些极具颠覆性的初创公司。然而事实恰恰相反，在这一高度严谨且受到严格监管的专业赛道中，真正处于领跑地位的依然是“优秀的传统医疗器械企业（the good old MedTech）”。以 GE HealthCare 为例，公司当时已斩获整整 100 项美国 FDA 医疗器械 AI/ML 认证许可，在获批的 AI/ML 赋能医疗设备榜单上位列全球第一。这意味着市场上其实已经存在大量经过严苛验证、合规批准且成熟可用的医疗 AI 产品，可供全球各大医疗系统直接采购并部署。 然而，技术储备与实际临床应用之间却存在着显著的反差。即便在 AI 应用最为成熟的放射学领域，来自全球不同地区、不同层级医院的数据均表明，绝大多数放射科医生在日常临床工作中仍未使用 AI。尽管这一统计数据可能会随时间略有推移而微调，但整体而言，医疗 AI 在临床一线的实际采纳率依然偏低且推进缓慢。 - **FDA 认证领跑**：GE HealthCare 在 AI/ML 赋能医疗设备领域拥有恰好 100 项 FDA 认证许可，位居行业首位。 - **临床采纳落差**：虽然合规商用算法数量充足，但全球范围内大部分放射科医生在日常读片中仍未常规使用 AI 工具。 --- ### [12:56](https://www.youtube.com/watch?v=qIQHYeysCLU&t=776s) 移动X光设备中的端侧AI：应对急重症与医疗资源不均 在急诊室、病房或重症监护室的病床边，常常能看到医护人员推着移动式 X 光机（Mobile X-ray）为行动不便的患者就地进行胸部摄片。为了将急救响应时间压缩到极限，工程团队在多年前便将 AI 算法直接嵌入到了这类移动摄片设备中，实现关键病症的原生端侧自动识别。当机器完成图像采集的瞬间，AI 算法便能即刻判断患者是否存在气胸（Pneumothorax，即肺萎陷）等致命危急重症。 ```mermaid flowchart LR accTitle: 移动X光端侧AI危重症筛查与分流流程 subgraph S1["床旁影像采集"] A["移动X光机床旁摄片"] --> B["端侧AI实时图像分析"] end subgraph S2["危重症即刻预警"] B --> C{"检测到危急征象？<br/>(气胸 / 插管移位)"} C -- "是" --> D["设备屏幕即刻高危警报<br/>通知临床护理团队快速处置"] C -- "否" --> E["常规归档与放射科排队读片"] end ``` 气胸属于必须由临床团队争分夺秒处置的紧急状况。对于资深放射科医生而言，在胸片上识别气胸并不困难，但问题在于现实中并非所有医疗机构都能提供 7×24 小时全天候在岗的放射科专家。当夜间或基层缺少读片专家时，由机器主动发出警报——“嘿，这里出现了危急状况，最好立即处置”——具有不可替代的挽救生命价值。同样的端侧 AI 技术也被广泛应用于气管插管（Endotracheal tube）位置的自动化评估，辅助医生判断插管深度是否恰当。在西方发达国家，放射科专家的覆盖相对完备，但在全球许多医疗资源匮乏的地区，想找到一名专家及时出具读片报告依然极其困难。 - **气胸自动识别**：移动 X 光机在床旁曝光完成后数秒内自动检测肺萎陷征象并触发预警。 - **气管插管定位评估**：通过算法自动测量与评估插管尖端位置，防止插管过深或过浅。 - **弥合资源鸿沟**：为缺乏 7×24 小时放射科专家值班的中小型医院与边远地区提供急危重症第一道防线。 --- ### [14:35](https://www.youtube.com/watch?v=qIQHYeysCLU&t=875s) 大语言模型落地：重塑患者沟通、病历摘要与病理协作 医疗不仅关乎冰冷的图像诊断，更关乎医患之间的沟通与关怀。调研显示，约有 50% 的患者对与医护人员的沟通体验感到不满。随着大语言模型（LLM）的演进，生成式 AI 不仅在专业准确性上迅速攀升，甚至在多项测评中被患者评价为“比医护人员更具同理心（more empathic）”。基于这一特性，业界诞生了许多充满温情的创新应用。例如，有团队将一位乳腺癌康复者多年抗癌历程中写下的完整心路笔记与治疗经验喂给大语言模型进行微调，打造了一款完全免费公开的心理支持对话工具。当患者或家属深处绝望、孤独无助且身边没有倾听者时，这类工具能够提供全天候的共情陪伴与经验参考。 在医生侧的工作流中，AI 也正在全方位接管机械繁重的信息整理工作： 1. **影像报告初稿生成**：AI 初步分析影像并自动生成结构化诊断报告草稿，放射科医生只需进行核对、修改并签名确认。 2. **多源异构病历一页纸汇总**：当医生接手一位新转入的复杂患者时，往往需要翻阅极其混乱的电子病历（EMR）——包括手写病程记录、非结构化文本、格式各异的化验单与检查表格。AI 能够瞬间解析整套复杂档案，提炼出关键病史、诊断要点与异常指标，生成简洁的“一页纸摘要（One-page summary）”。 3. **带超链接溯源的交互式病历**：新一代集成到临床产品中的 LLM 摘要系统在左侧提供了结构化索引超链接。医生若想深入了解患者某项特定既往症或异常化验，只需点击词条即可一键穿透跳转至原始电子病历对应的扫描件或历史文档。 4. **病理学智能副驾（Pathology Co-pilot）**：病理科医生可以直接使用自然语言与超高分辨率病理切片图像进行多轮问答与交互检索，大幅降低切片筛查耗时。 --- ### [17:13](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1033s) 影像重建与自主AI：重塑MRI检查体验与超声操作门槛 在放射影像流程优化方面，欧洲已批准了首个获得 CE 认证的“全自主 AI（Autonomous AI）”系统。该算法能够全自动筛查出完全正常的胸部 X 光平片，并直接将其从放射科医生的日常待读列表中剔除，使医生能够将全部精力聚焦于存在可疑病变的复杂病例，从而极大地释放了科室产能。 在成像硬件底层，深度学习重建算法正在彻底改变磁共振成像（MRI）的技术范式。传统的 MRI 检查耗时长、机器轰鸣刺耳且舱内空间狭窄，对受检者而言是一段极其煎熬的体验。GE HealthCare 运用深度学习图像增强与超分辨率技术，在大幅缩短扫描时间的同时显著提升了图像信噪比与解剖清晰度。尤其在心脏磁共振扫描中，AI 技术彻底免去了心脏病患者在检查过程中反复吸气、屏气的痛苦，患者如今可以在扫描舱内“自由呼吸（breathe freely）”，而系统依然能输出高质量的心脏动态影像。全球已有超过 5000 万名患者受益于这类基于深度学习的成像技术。 ```mermaid flowchart TD accTitle: 深度学习在医学成像与超声引导中的赋能维度 subgraph MRI["磁共振成像 (MRI) 体验升级"] M1["深度学习降噪与超分辨率重建"] --> M2["扫描时间大幅缩短"] M1 --> M3["心脏扫描实现全程自由呼吸"] M2 & M3 --> M4["图像清晰度与患者舒适度双重提升"] end subgraph US["便携式超声 (Ultrasound) 降门槛"] U1["探头位置与手法实时 AI 视觉引导"] --> U2["降低高难度超声手法的操作门槛"] U2 --> U3["赋能经验不足的基层与全科医生"] end ``` 针对高度依赖操作手法的手持便携超声设备，AI 同样扮演了关键导师的角色。超声检查存在极高的“操作者依赖性（operator dependent）”，唯有经过系统培训且经验丰富的医师才能精准捕捉到符合诊断标准的切面。内置于微型超声设备的 AI 算法能够实时分析超声回波，并在屏幕上以可视化轨迹引导操作者如何移动、旋转或倾斜探头，从而帮助非专家级医护人员也能快速采集到标准质量的超声影像。 - **自主正常胸片剔除**：首款获欧洲 CE 认证的自主 AI 系统，自动过滤正常胸片以实现读片分流。 - **超 5000 万患者受益**：深度学习 MRI 重建技术大幅缩短采集时间，并攻克了心脏扫描需频繁屏气的临床痛点。 - **手持超声智能运棒指引**：在微型超声仪中利用 AI 实时纠正手法，破解超声高依赖人工经验的普及瓶颈。 --- ### [19:01](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1141s) 告别繁琐系统：环境智能、多模态大模型与AI伦理新命题 在影视剧作品中，人们从未见过医生整天坐在电脑屏幕前机械地录入和翻查电子病历，但现实中的医护人员却被笨拙、老旧且复杂的医院 IT 系统牢牢捆绑。医疗信息化的未来必然要求彻底颠覆这种交互方式，转向利用“环境感知记录技术（Ambient Scribe）”与自然语言对话来实时捕捉和检索患者信息，重塑医院 IT 系统的用户界面体验。 医疗健康领域的数据形态具有极其独特且复杂的“高度多模态（Incredibly multimodal）”特征——涵盖高精影像、基因组学数据、时序生命体征、实验室指标以及非结构化临床叙述。医疗基础模型（Foundation Models）的成熟，终于为跨模态、深层次整合分析这些复杂数据集打开了大门。 随着医疗 AI 技术的可靠性与效能得到持续验证，整个行业正在迎来一场深刻的伦理视角转变：当技术已经能够明确减少漏诊、提升效率并挽救生命时，“不使用 AI 是否反倒构成了一种不道德（is it still ethical to not use it）？”技术的终极使命绝非取代医护人员，而是全面接管那些重复、繁琐且机械的日常杂务，将宝贵的时间还给临床团队，让医生真正回归病床边缘，全心专注于关怀与救治患者本身。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=qIQHYeysCLU&t=4s) 如今全球所有医疗数据的翻倍周期已经缩短至仅仅 72 天左右，而在短短数年之前，这个周期还是三年半。 — Jan Beger > [00:04](https://www.youtube.com/watch?v=qIQHYeysCLU&t=4s) 医疗系统产生的数据有 97% 完全处于闲置、未被利用的状态，对改善整体医疗效果毫无贡献，实际被有效利用的数据仅占 3%。 — Jan Beger > [04:48](https://www.youtube.com/watch?v=qIQHYeysCLU&t=288s) 普通人乘坐飞机遭受伤害的概率仅为百万分之一，而当一个人进入医疗救治流程时，遭受某种形式伤害的概率竟然高达十分之一（one in 10）。 — Jan Beger > [04:48](https://www.youtube.com/watch?v=qIQHYeysCLU&t=288s) 60% 的医生把更多的工作时间花在面对电子病历系统（EMR）录入数据上，而非与患者进行面对面的直接沟通。 — Jan Beger > [11:34](https://www.youtube.com/watch?v=qIQHYeysCLU&t=694s) 真正处于领跑地位的依然是优秀的传统医疗器械企业（the good old MedTech）。 — Jan Beger > [12:56](https://www.youtube.com/watch?v=qIQHYeysCLU&t=776s) 嘿，这里出现了危急状况，最好立即处置。 — Jan Beger > [14:35](https://www.youtube.com/watch?v=qIQHYeysCLU&t=875s) 生成式 AI 不仅在专业准确性上迅速攀升，甚至在多项测评中被患者评价为比医护人员更具同理心（more empathic）。 — Jan Beger > [17:13](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1033s) AI 技术彻底免去了心脏病患者在检查过程中反复吸气、屏气的痛苦，患者如今可以在扫描舱内自由呼吸（breathe freely）。 — Jan Beger > [19:01](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1141s) 当技术已经能够明确减少漏诊、提升效率并挽救生命时，不使用 AI 是否反倒构成了一种不道德（is it still ethical to not use it）？ — Jan Beger > [19:01](https://www.youtube.com/watch?v=qIQHYeysCLU&t=1141s) 技术的终极使命绝非取代医护人员，而是全面接管那些重复、繁琐且机械的日常杂务，将宝贵的时间还给临床团队，让医生真正回归病床边缘。 — Jan Beger ## Mentioned Jan Beger, GE HealthCare, OpenAI, OpenAI o1, MedQA, 美国食品药品监督管理局 (FDA), 世界卫生组织 (WHO), CE 认证, EMR (电子病历系统) ## Video description Jan Beger, Global Head of AI Advocacy, GE HealthCare Abstracts In the Human. Machine. Care. keynote, Jan will explore how AI confronts healthcare's critical challenges: managing exponential medical data growth, supporting complex clinical decisions, and alleviating global healthcare professional shortages. By showcasing AI's transformative potential, the talk will demonstrate how technological innovation can enhance diagnostic precision, reduce medical staff burden, and improve patient outcomes. Crucially, the discussion will highlight that AI is not just about technology. In healthcare, it is about the people, the trust. [Source](https://www.youtube.com/watch?v=qIQHYeysCLU)

[0:04] Jan Beger: 如今全球所有医疗数据的翻倍周期已经缩短至仅仅 72 天左右，而在短短数年之前，这个周期还是三年半。 [0:04] Jan Beger: 医疗系统产生的数据有 97% 完全处于闲置、未被利用的状态，对改善整体医疗效果毫无贡献，实际被有效利用的数据仅占 3%。 [4:48] Jan Beger: 普通人乘坐飞机遭受伤害的概率仅为百万分之一，而当一个人进入医疗救治流程时，遭受某种形式伤害的概率竟然高达十分之一（one in 10）。 [4:48] Jan Beger: 60% 的医生把更多的工作时间花在面对电子病历系统（EMR）录入数据上，而非与患者进行面对面的直接沟通。 [11:34] Jan Beger: 真正处于领跑地位的依然是优秀的传统医疗器械企业（the good old MedTech）。 [12:56] Jan Beger: 嘿，这里出现了危急状况，最好立即处置。 [14:35] Jan Beger: 生成式 AI 不仅在专业准确性上迅速攀升，甚至在多项测评中被患者评价为比医护人员更具同理心（more empathic）。 [17:13] Jan Beger: AI 技术彻底免去了心脏病患者在检查过程中反复吸气、屏气的痛苦，患者如今可以在扫描舱内自由呼吸（breathe freely）。 [19:01] Jan Beger: 当技术已经能够明确减少漏诊、提升效率并挽救生命时，不使用 AI 是否反倒构成了一种不道德（is it still ethical to not use it）？ [19:01] Jan Beger: 技术的终极使命绝非取代医护人员，而是全面接管那些重复、繁琐且机械的日常杂务，将宝贵的时间还给临床团队，让医生真正回归病床边缘。

[4:48] Jan Beger: When I've seen this number... the first time, I thought it cannot be right. Because it's just one in 10. [8:04] Jan Beger: Healthcare is pretty slow in adopting new technologies... for a reason because everything we implement to healthcare we need to make sure it's tested, it's validated and it's not harming the patients. [19:36] Jan Beger: is it still ethical to not use it

[0:04] Thank you very much for joining. Just a [0:06] quick uh check who of you is kind of [0:08] working in the broader healthcare [0:11] domain. [0:13] Perfect. Just a it's good to see a [0:15] couple of folks in that area. Look, what [0:18] I would like to do the next couple of [0:19] minutes, uh, talk a little bit about [0:21] like the big problems we have in [0:23] healthcare as an industry and then try [0:25] to kind of share a couple of ideas and [0:27] thoughts and use cases on how AI as a [0:30] kind of broad concept, broad technology [0:33] can make an impact to those issues and [0:35] problems. Let's get into this and let me [0:38] start with the with the first fact which [0:40] um is that today it just takes about 72 [0:44] days to double all medical data [0:46] globally. Think about this. Yeah. [0:48] Compared to three and a half years just [0:50] a couple of years back, [0:53] 30% of the um of the world's data volume [0:57] is somehow healthcare related. Another [1:00] interesting fact and you see here about [1:03] the dimension of uh of the data and the [1:06] complexity of this industry. [1:09] A few examples here. Every 26 seconds a [1:11] new medical article is published. Think [1:13] about this. Yeah. If you are a [1:15] dermatologist, for instance, like the [1:17] doctor taking care of your skin, they [1:19] would need to read 3,900 scientific [1:22] publications a year to keep up with [1:25] literature in that space. [1:27] And if you think that's impossible, I [1:29] mean, have a look to this. If you are a [1:31] family doctor, a general practitioner, [1:33] you would need to read for 21 hours each [1:35] and every day to keep up with primary [1:38] care literature. Yeah, impossible. [1:42] So, this is an uh radiology image of a [1:45] woman's press, the so-called mamogram, [1:47] and it contains more information than [1:49] the telephone book of New York, even if [1:52] I don't know if there's still like a [1:53] physical telephone book. [1:56] Another example, um from the radiology [1:58] domain, um so-called lower body um CT [2:02] and geography can quickly get to 3,000 [2:05] images. Yeah. And think about 3,000 [2:08] images. a radiologist. So the doctor has [2:11] to read through [2:14] like an average hospital produces 50 [2:16] pabytes of data. [2:21] And as I just talked about radiologists [2:23] um as you can see in a 12-hour shift [2:26] like on average a diagnostic radiologist [2:29] reads 50,000 [2:32] excuse me. Yeah. 50,000 images. 50,000. [2:36] Yeah. compared to just 500 a few years [2:39] back. And companies like G Healthcare [2:42] contribute to this massive kind of [2:44] increase in in healthcare data because [2:46] we produce better and better imaging [2:48] devices that can produce yeah thinner [2:50] and thinner slices of the human body. [2:53] And when you think about 50,000 um [2:56] images and 12 hours, you will also [2:59] realize that they actually have less [3:01] than a second time to read a single [3:04] image. Yeah. So massive amounts of data [3:06] of information they have to deal with [3:10] as I just talk about medical imaging. [3:13] Yeah, 90% of healthcare data is medical [3:17] imaging related not just radiology but [3:19] also things like dermatology, of [3:20] opthalmology, pathology and so on. [3:24] Yeah. So this is the amount of uh yeah [3:27] the the amount of uh radiology images [3:31] stored in like a typical university [3:33] hospital or in like a um radiology chain [3:36] for instance. [3:38] And actually one of the reasons why in [3:40] the radiology space especially in X-ray [3:43] we uh create so many images is the fact [3:46] that radiologists often produce kind of [3:49] images of two different orientations. [3:51] Yeah. to make sure we cannot can locate [3:54] an object in the human's body correctly. [3:57] Um, another interesting number, 97% of [4:01] the data we produce uh in healthcare is [4:04] not used. It's unleveraged, unharnessed [4:07] and does nothing to improve overall [4:09] healthcare outcomes. Yeah, we just use [4:11] 3% of that data. [4:15] So, this was the first kind of overall [4:17] theme I wanted to mention. Yeah, this [4:19] massive amounts of data we have to deal [4:21] deal with which is both like a burden [4:24] especially to healthcare providers [4:26] today. They often see this as an as a [4:28] burden as a struggle because they often [4:30] have also to kind of archive and save [4:32] this data for a certain amount of time [4:34] but on the flip side it's also a huge [4:36] opportunity and if we apply AI to those [4:39] kind of data sets um in the future uh in [4:42] the right manner we can get to better [4:44] outcomes and there is a huge opportunity [4:46] in that space as well. want to move on [4:48] and talk about another um interesting [4:50] kind of theme of kind of challenges. Uh [4:54] here you can see the risk uh of being [4:56] harmed during a flight. Yeah. Um one in [5:00] a million. So pretty okay. When you [5:03] compare this number to uh the risk of [5:05] being harmed during your healthcare [5:07] journey, when I've seen this number on [5:09] the next slide the first time, I thought [5:10] it cannot be right. Yeah. Because it's [5:13] just one in 10. Yeah. Whatever harm [5:15] exactly means. Okay. doesn't doesn't [5:18] need to mean uh to mean death [5:21] but it leads to the fact that like [5:22] globally um three million deaths occur [5:25] uh on a on an annual basis just due to [5:28] like unsafe healthcare. [5:32] Yeah. Another space so medical hours [5:34] overall quality another space where I [5:37] believe AI can make a a difference going [5:40] forward. A third example or a third big [5:43] problem I want to highlight maybe some [5:45] of you are aware of this um and this is [5:47] the um kind of the capacity the shortage [5:51] of health care professionals. First of [5:52] all as you can see 60% of doctors spend [5:55] more time in front of the EMR system the [5:58] electronic medical record instead of [6:00] spending time face to face with their [6:01] patients. Yeah. [6:04] We see about 50% of radiologists right [6:07] so those doctors reading your MRI CAT [6:09] scans and so on. uh uh being burned out. [6:14] Uh in health systems, we see uh overall [6:16] a 100% staff turnover every 5 years. [6:20] Yeah. I mean, think about this. [6:24] Looking over to Canada, uh if you need [6:26] an MRI scan, you have to wait for about [6:28] 13 weeks. Yeah. To get an appointment. [6:33] Overall, globally, uh today, we have a [6:36] shortage of um 7 million healthcare [6:38] professionals. This is data from the [6:40] World Health Organization and they [6:42] expect this number to increase to 11 [6:44] million by 2030. [6:46] Yeah. So um lots of problems and uh if [6:50] you find this funny I do as well but [6:52] actually it's not really funny. Yeah. [6:56] And this is just kind of one side of the [6:58] story because what we also see is that [7:00] it's not just we don't have enough [7:01] healthare professions today. We also [7:03] have the issue of an uh population [7:06] that's getting older and older and with [7:08] this you also know um the increase of [7:11] like chronic disease for instance um so [7:14] kind of multiple issues um together uh [7:17] which I try to describe here [7:21] not even talking about this yeah um four [7:24] uh 4.5 billion people around the world [7:27] do not have access to kind of proper um [7:29] care overall [7:31] and look what I'm trying to say here is [7:33] that um this delta between supply and [7:37] demand is just increasing. Yeah. And I [7:39] think what we need to do uh and I think [7:41] the audience here and everybody at this [7:43] conference can can significantly [7:45] contribute is yeah implementing as many [7:49] as possible useful and validated [7:52] technologies such as AI into our health [7:54] systems to really make an impact and t [7:57] tackle some of those problems I just [7:58] mentioned. There are plenty more. I just [8:00] picked three of them the last couple of [8:02] minutes. [8:04] And look um in healthcare the the [8:07] potential the opportunities for AI is [8:09] endless. Yeah. Um here just a kind of a [8:13] a brief example or like an high level [8:16] example of what is possible in this [8:18] space. What are potential use cases and [8:20] areas of opportunity. [8:23] One thing I have to say maybe the people [8:25] uh who work in healthcare for some time [8:27] know um usually healthcare is pretty [8:29] slow in adopting new technologies. Yeah. [8:32] For a reason because everything we [8:33] implement to healthcare we need to make [8:35] sure it's tested, it's validated and [8:36] it's not harming the patients. Yeah. So [8:39] therefore things are slow for a reason. [8:41] But what we have seen especially like [8:42] the last like I don't know four to five [8:44] years is um huge kind of uptake and [8:48] adoption of um AI in general um really [8:52] across the globe. And here are just a [8:54] couple of example yeah use cases if you [8:57] will of how AI is used in health systems [9:00] around the globe. Yeah. On top you can [9:02] see some of those uh LLM type of use [9:05] cases. Further down some computer vision [9:07] examples and so on and so forth. [9:10] So let me show you this um which I find [9:13] a very interesting um uh kind of example [9:16] to share. You can see and maybe you have [9:18] seen this in the past as well um the the [9:22] 01 model um reached a uh 96.5% [9:26] accuracy in a standard medical test in [9:29] the so-called met QA test. Yeah, [9:31] standard medical questions and 01 [9:34] achieved this 96.5% [9:37] accuracy. And now I mean ask yourself [9:39] would you trust your doctor if your [9:40] doctor would be accurate 96.5% [9:43] of the time [9:45] probably not right so let me show you [9:48] this and maybe that's not fair yeah to [9:50] like human experts human doctors but [9:52] what you can see on the right hand side [9:54] you can see I mean exemplary uh the 01 [9:57] model with 96.5% accuracy next to it 87% [10:02] this is like the accuracy of an [10:04] experienced medical doctor Yeah, average [10:07] of course and on the left you can see [10:10] someone who just left medical school [10:11] around 60%. [10:13] Yeah, I mean again it's not a fair apple [10:16] comparison but I just want to point out [10:18] those technologies those AI models get [10:21] really uh good at standard medical [10:24] questions. This is far from being like [10:26] the real world I have to say as well. [10:29] And here just a couple of stats on uh [10:31] how good they are today. And again, it's [10:33] just like um standard clinical [10:35] benchmarks, but here is an example of [10:37] how um AI can make an impact to like the [10:40] emergency room. Here is an example of [10:42] how AI can overall really make an impact [10:45] to uh uh the diagnostic space. And as [10:49] you can see, uh it's really uh in a [10:51] variety of different areas getting close [10:53] or like superseding already the human [10:56] expert. But again, it's all in kind of [10:58] controlled environments in kind of [11:00] narrow um uh questions, if you will, and [11:04] it's still far from the real world [11:07] because real world healthcare, some of [11:09] you know, is is messy, is complicated, [11:11] and so on. I wanted to show you this uh [11:15] right, and I mentioned this already a [11:16] few minutes ago, healthcare is a very [11:18] regulated market. Um the US FDA has um [11:23] uh authorized more than 1,000 AI and [11:26] machine learning and and machine [11:29] learning enabled medical device devices [11:31] as of today. And you can see a big [11:34] portion of this is in the radiology [11:36] space. [11:38] Yeah, it's primarily like imaging use [11:40] cases. So computer vision use cases. Um [11:43] and just to mention um a lot of people [11:46] uh in this space think that uh when it [11:49] comes to healthcare AI or medical AI [11:52] either the big tech or those great [11:54] innovative startups are leading the [11:55] space but actually it's the good old met [11:58] who is leading in this specific space. [12:00] Yeah. So um at G healthcare for instance [12:03] we have uh more than yeah actually [12:06] exactly 100 um FDA authorizations as of [12:10] today and we are leading this list um of [12:13] AI and machine learning aid medical um [12:16] devices um right now. So so what I'm [12:20] saying here is there is a lot of kind of [12:23] validated approved regulated AI out [12:26] there for health systems to kind of [12:28] purchase and implement. But this is also [12:30] interesting again I stick to radiology [12:32] here as well. You can see that a lot of [12:35] radiologists around right different [12:37] geographies but also different health [12:39] systems are not using AI today. Yeah. Um [12:44] and this uh stat is from a few months [12:46] ago. Maybe things have changed already [12:49] um slightly but um adoption is still [12:52] rather um let's say low or slow. [12:56] I wanted to show you this picture. This [12:57] is a mobile X-ray machine. Maybe some of [12:59] you seen this um in the emergency room [13:02] at the patient's bedside at a ward for [13:05] instance you can move this around and [13:07] can take an X-ray image for instance of [13:10] the chest um where the patient is [13:12] located [13:14] and look what what we have done uh quite [13:16] some time ago is we have put AI into [13:18] those machines to detect critical [13:21] conditions automatically. So in this [13:24] specific case the what the machine does [13:26] immediately after the image has been [13:27] acquired is it can detect a so-called [13:30] pumothorax a collapsed lung. Yeah. [13:32] Something that is a critical condition [13:34] where the care team has to act quickly [13:36] but also something which is very easily [13:38] to spot by a radiologist when there is a [13:42] radiologist. Yeah we have a lot of [13:45] situations around the globe where there [13:47] is no radiologist available 24 by7. uh [13:50] and therefore I think it's better that [13:52] the machine tells you hey there is [13:54] something critical you better act [13:55] quickly. Yeah I mean same things for [13:57] like uh positioning a so-called [14:00] endotracial tube to help the patient [14:02] breathe. Um also here the tube location [14:05] can be assessed um with the help of uh [14:08] of AI today and as I just said right not [14:11] everywhere um you have access 24 by7 to [14:14] radiologists I mean have a look to to [14:16] this comparison um like in the western [14:19] world everything is perfectly fine I [14:21] think with this specific uh type of role [14:24] or uh yeah the radiologist themselves [14:26] but uh when you look to other countries [14:28] it's it's pretty difficult to get an [14:31] expert read your scan [14:35] Switching gears talking to to about [14:38] something very different. Um yeah about [14:41] 50% of uh patients are not happy with [14:44] the communication uh uh with their [14:46] providers. Um what we also see now with [14:49] large language models for instance that [14:51] they are not just getting better in [14:52] terms of accuracy. You probably also [14:54] have heard or read about the fact that [14:56] they are perceived as more empathic than [14:59] the the the healthare professionals. [15:01] Yeah. And this leads to tools like this. [15:03] This is a an LLM that has been fed with [15:07] the experiences of a cancer survivor of [15:10] a breast cancer survivor who wrote down [15:12] her entire care journey over multiple [15:15] years. This has been used to train in [15:18] large language model. This is available [15:20] uh for free on the web and built for [15:23] like patients, their families, their [15:25] friends. If you have a hard time, you go [15:27] through and you just need somebody to [15:29] talk to and there isn't no human being [15:31] around, maybe you want to have a look to [15:33] to this tool. Um what we do for instance [15:37] as well is we use um uh AI uh to read [15:40] images and produce u the report. Yeah, [15:43] like preliminary reports the [15:45] radiologists then have to just review, [15:47] validate and sign. Um here's an example [15:50] of an AI tool that helps doctors who are [15:53] for instance new to an uh a patient um [15:56] to analyze their medical record which [15:59] can be very messy as you can see here. [16:01] Yeah. Structured unstructured [16:03] information, handwritten notes, um lab [16:05] results and tables and so on and so [16:07] forth. And what this AI does is it puts [16:09] it all together into like a one-page [16:11] summary. Yeah. For the new for the [16:13] doctor to get a quick overview about [16:16] that patient. I just move on here. It [16:18] just at the end this video shows this [16:20] onepage summary. [16:22] Something what we are about to uh kind [16:25] of build into our products right now is [16:27] to uh yeah using um large language [16:30] models to provide those kind of patient [16:33] summaries. Yeah. So again get a quick uh [16:36] overview at a glance of what is status [16:38] of the patient and if you as you can see [16:40] on the left hand side you see some like [16:43] hyperlinks. If you want to zoom more [16:45] into a specific condition, you can [16:47] always click to uh that item and then [16:49] you get more information and can even [16:52] get to this specific document or type of [16:54] information in the electronic medical [16:56] record. [16:58] Um this is another example of uh [17:00] something like a co-pilot for pathology. [17:03] Yeah. So like a pathologist interacting [17:06] with natural language with pathology [17:08] images. Um very powerful nice use case. [17:12] Um I also wanted to show you this. This [17:14] is the first um what we call autonomous [17:17] AI um which got uh authorized in the [17:21] European Union, a so-called CE mark [17:23] which helps with detecting um chest [17:26] X-rays which are normal. Yeah. And when [17:29] you rule them out from the radiologist [17:31] reading workflow, it's a huge efficiency [17:33] play. Look, I'm going to uh skip a [17:35] couple of things. AI also helps with um [17:38] improving um improving uh image quality. [17:41] As you can see on the left hand side, [17:43] this is like a conventional MRI image. [17:44] On the right hand side is an MRI image [17:47] that has been uh improved using uh deep [17:50] learning technologies. Um here as well [17:52] and this is also interesting if some of [17:54] you had an MRI scan yourself in the [17:57] past. Um it's not a kind of a pleasant [18:00] experience. Takes time. It's noisy. It's [18:02] not fun to be in there. Uh we use AI as [18:05] well for reducing those um scan times [18:07] and at the same time improving image [18:09] qualities. On the right you can even see [18:11] for cardiac patients we um remove the [18:14] need to stop breathing multiple times [18:16] during that scan. They can now free uh [18:19] breathe freely and it still works very [18:21] well. I mean more than 50 million uh [18:24] patients around the globe have been [18:25] scanned with those technologies. Um [18:28] actually this is another interesting use [18:29] case. We have also put AI into those [18:32] machines, ultrasound scanners, they are [18:34] that small today. Yeah. Or can be that [18:37] small today. And the issue we have with [18:39] ultrasound is it's incredibly operator [18:41] dependent. Meaning you need to have a [18:42] certain level of education and [18:44] experience to use those kind of powerful [18:46] tools properly. And uh the AI in here [18:49] helps the user and tells you how to move [18:52] this probe on the patient's body to get [18:54] to the to right quality um medical [18:58] images. Look, I skipped this example. [19:00] Um, and just want to point out that [19:02] maybe when you know some of those TV [19:04] shows, you probably know you we never [19:06] see like doctors sitting in front of a [19:09] computer and going through the EMR. I [19:11] think what we need to do is and maybe [19:13] some of you also work on like um ambient [19:16] scribe type of technologies. We need to [19:18] get to a point where we interact with [19:20] patient information in a more like [19:21] natural language way. And I think this [19:23] is really the way to go. And also on the [19:25] flip side, we need to look into how can [19:27] we improve the user interfaces of those [19:30] cumbersome, complicated, old-fashioned [19:33] um IT systems in hospitals. Look, one of [19:36] the things to wrap up here, one of the [19:38] things which excites me very uh much uh [19:40] and I focus the most of my time on is um [19:43] AI or healthcare is incredibly [19:46] multimodal. Yeah. um compared to many [19:49] other industries and AI really now with [19:51] foundation models opens the door to uh [19:54] analyzing those kind of multimodal data [19:57] sets properly and look I skipped this [19:59] use case about agents u because I just [20:02] want to point out that um there's so [20:04] much power in this technology now and um [20:07] the question now really is is it still [20:09] ethical to not use it yeah especially or [20:12] at least for a certain number of use [20:14] cases and look when I think about health [20:16] care professionals [20:18] which we want to serve with those [20:19] technologies. As you can see here, they [20:21] want to spend more time with their [20:22] patients and ultimately we want to make [20:24] sure that AI covers all those routine, [20:26] mundane, repetitive tasks so that [20:28] healthcare professionals can focus on [20:30] what is really important. And with this [20:32] um thank you very much everyone.