Inside Houston Methodist Hospital: Deploying AI in Practice With Houston Methodist Hospital

Ai4

AI-assisted summary of Inside Houston Methodist Hospital: Deploying AI in Practice With Houston Methodist Hospital — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by Ai4 on August 15, 2026, this fireside chat features Joshua Robert, Founding Editor-in-Chief of the Journal for Artificial Intelligence in Medicine (JAIM), in conversation with Dr. Roberta Schwartz, Executive Vice President and Chief Innovation Officer at Houston Methodist Hospital. The session outlines real-world case studies of how the health system evaluates, prioritizes, and scales artificial intelligence across both clinical and operational domains while navigating organizational and technical hurdles.

## TL;DR Dr. Roberta Schwartz explains Houston Methodist Hospital's rigorous evaluation framework, emphasizing that enterprise adoption requires broad platform capability rather than isolated point solutions. She addresses the internal tug-of-war between HR retention messaging and CFO cost-reduction expectations, showing how AI like radiology triage and scheduling agents streamlines care even without lowering per-scan reading times. Schwartz also highlights the disruption of traditional entry-level career ladders as administrative tasks are automated, differentiating mature operational bots from early-stage autonomous clinical concepts. Concluding with an analysis of demographic pressures and specialist shortages, she calls on healthcare leaders to gain hands-on technical literacy and lead workforce transformations with radical transparency and empathy.

## Atmosphere & style The session has a brisk, candid, and refreshingly unvarnished tone, driven by Dr. Schwartz’s sharp executive realism, self-deprecating wit, and refusal to sugarcoat uncomfortable organizational truths. Joshua Robert provides an engaging, peer-level sounding board, balancing clinical curiosity with lighthearted banter about specialty choices and clinician skepticism. The conversation feels less like a standard tech promotional keynote and more like an authentic, behind-the-scenes postmortem on the messy human, cultural, and financial friction of deploying AI inside a major health system.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=4s) Only about 1% of pitched healthcare AI solutions secure a hospital pilot, as health systems aggressively filter out point solutions and undifferentiated DIY agent platforms. - [01:54](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=114s) Narrow point solutions create insurmountable integration and maintenance overhead; enterprise adoption requires platforms with the architectural aperture to expand horizontally across multiple clinical and administrative workflows. - [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) AI return on investment cannot be evaluated through a single monolithic framework; for example, emergency triage AI delivers value by reprioritizing worklists and speeding downstream care rather than speeding up individual radiologist reading times. - [06:00](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=360s) Deploying generative and agentic AI creates organizational tension between HR commitments to job security and CFO demands for measurable headcount and budgetary efficiencies. - [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) Clinician adoption is being accelerated externally by a consumer AI wave, with patients arriving at appointments pre-armed with algorithmic diagnoses and specific medication requests. - [09:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=580s) Physicians will rapidly adopt point-of-care tools like OpenEvidence when they eliminate friction, even accepting novel data-sharing or advertising arrangements in exchange for high-utility decision support. - [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) Executive leadership must practice radical transparency regarding AI's disruptive impact on clinical practice instead of relying on reassuring platitudes that mask the scale of impending change. - [12:51](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=771s) Automating routine administrative functions eliminates traditional entry-level roles, posing a major structural challenge for training non-expert staff into future domain experts. - [15:00](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=900s) Routine operational workflows such as call center scheduling and back-office tasks are mature targets for bot automation, whereas autonomous clinical diagnosis and prescribing remain in their earliest experimental infancy. - [17:31](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1051s) An aging Baby Boomer demographic combined with acute specialist shortages makes AI capacity augmentation essential to prevent widening disparities in healthcare access. - [19:00](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1140s) Healthcare leaders must achieve personal technical literacy and hands-on experimentation while guiding their organizations through cultural transformation with patience and empathy.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=4s) Introduction and the 1% Threshold for Hospital AI Adoption

Joshua Robert, Founding Editor-in-Chief of the *Journal for Artificial Intelligence in Medicine* (JAIM), opens the session on deploying AI in high-stakes clinical environments where patient lives are on the line and health systems remain intensely risk-averse. He introduces Dr. Roberta Schwartz, Executive Vice President and Chief Innovation Officer at Houston Methodist Hospital—which, as of the day of the talk, marked 15 consecutive years ranked as the number one hospital in Texas. Schwartz, formerly Director of Business Development at the Icahn School of Medicine at Mount Sinai and recognized among the Top 25 Women Leaders in Modern Healthcare, founded and leads the Center for Innovation at Houston Methodist.

Joshua poses the fundamental entry question for health tech innovators: with countless vendors claiming to possess revolutionary AI that will transform medicine, what truly separates the 1% of solutions that secure a pilot from the 99% that get rejected at the door?

Schwartz reflects on the sheer volume of vendor pitches and how the landscape has shifted. At standard healthcare conferences like HLTH and ViVE, she can typically walk down the exhibit floor making rapid "yes, no, yes, no" determinations. However, the current wave of generative and agentic AI is "kicking my ass," she jokes, calling herself "self-prescribed pathetic" because she has not yet "vibecoded" anything herself. The market dynamic has radically pivoted from vendors pitching pre-built, targeted solutions to selling DIY tools where health systems are expected to build their own agents.

```mermaid flowchart TD accTitle: Hospital AI Vendor Evaluation Funnel accDescr: Visual representation of how Houston Methodist filters vendor pitches down to enterprise-ready solutions.

subgraph Inflow ["Vendor Market Inflow"] A["1,000+ Pinpoint Point Solutions (Single-use AI, DIY Agent Platforms)"] end

subgraph Filter ["Enterprise Evaluation Filters"] B{"Aperture Test: Can underlying tech expand beyond one niche?"} C{"Value & Impact Test: Is the clinical/operational pain point massive?"} D{"Internal Expert Review & RFP: IT, Clinical, & Domain Validation"} end

subgraph Outcome ["Adoption Funnel"] E["99% Rejected / Filtered Out"] F["Top 1% Approved for Hospital Pilot"] end

A --> B B -- "Narrow / Single-use" --> E B -- "Broad Platform Aperture" --> C C -- "Marginal ROI" --> E C -- "High-Impact Value" --> D D --> F ```

---

[01:54](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=114s) The Problem of Point Solutions vs. Expanding the Aperture

Schwartz details why narrow point solutions struggle to gain enterprise traction. In radiology alone, there are over a thousand different AI applications—one algorithm detects lung nodules on a chest X-ray, another finds rib fractures or signs of osteoarthritis, while others focus on retinal imaging. An enterprise health system cannot realistically procure, integrate, and maintain hundreds of distinct pinpoint tools unless a specific problem represents an overwhelmingly severe, urgent pain point.

Instead, the vendors that make it into the top 1% are those that may enter through a single use case but possess the architectural "aperture" to expand across multiple clinical and operational domains using the same underlying platform. Facing an exhibit floor with a hundred undifferentiated AI agent builders, Schwartz does not rely on vendor hype; she defers to Houston Methodist's domain experts across IT, marketing, and clinical specialties to run comprehensive, RFP-style market analyses to benchmark the full field before choosing a partner.

- **Market fragmentation:** More than 1,000 AI solutions exist in radiology alone, creating insurmountable integration and management overhead for health systems if bought piecemeal. - **The "Aperture" requirement:** Successful enterprise vendors must demonstrate that their core platform can scale horizontally into adjacent clinical and administrative workflows. - **Internal vetting:** Enterprise evaluation relies on cross-functional domain experts and rigorous market sweeps rather than single-stakeholder pilot requests.

---

[04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) Tailoring ROI Models and Navigating the CFO vs. HR Dilemma

When evaluating AI value, Schwartz emphasizes that there is no single monolithic "ROI model" across healthcare. Each use case requires its own distinct framework: - In pulmonary embolism (PE) detection, the return must be measured in true reductions in length of stay (LOS), saved brain cells, and saved lives—not merely flagging incidental findings three hours early if that time delta has no clinical consequence.

Schwartz recounts a debate with a Houston Methodist radiologist after deploying Aidoc. The radiologist argued the AI was not delivering on its promises because his per-scan reading time had not decreased. Schwartz pointed out that while his individual reading speed remained unchanged, the software fundamentally reorganized the worklist by triaging positive scans straight to the top of the queue. Consequently, the emergency department gained enormous operational certainty: if an ED physician does not hear from radiology within 30 minutes, they can be reasonably confident the scan is negative, dramatically accelerating downstream emergency care.

Establishing return on investment becomes particularly delicate with agentic and generative AI because internal stakeholders demand conflicting narratives. Human Resources leadership insists on messaging to staff that "it's okay, we won't get rid of anybody," whereas the Chief Financial Officer directly demands to know headcount reductions: "Hey, how are you getting rid of people?"

Schwartz highlights the institutional fortitude required to enforce real productivity shifts. When an internal department requested a $200,000 AI solution because "it can do so much more that we couldn't get to," Schwartz asked how many fewer staff members they would need. When the team replied they would keep the exact same staffing levels, she pushed back: "So what you're telling me is you want $200,000 more and the same amount of people? Yes. No."

| Stakeholder / Domain | Metric Focus & Strategic Expectation | Organizational Tension | |:--- |:--- |:--- | | **Emergency / Radiology (e.g., Aidoc PE)** | Worklist triage, accelerated intervention for positive scans, true Length of Stay (LOS) reduction. | Radiologist reading time per scan does not change, but ED clinical workflow speeds up significantly. | | **Human Resources (HR)** | Workforce morale, psychological safety, culture maintenance ("no one will be replaced"). | Clashes directly with financial ROI demands for measurable headcount and budget efficiencies. | | **Chief Financial Officer (CFO)** | Hard dollar savings, labor efficiency, tangible reduction in operational run-rate costs. | Creates friction when departments want expensive AI tools to expand scope without reducing headcount. |

---

[08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) Overcoming Clinician Inertia and the Consumer-Driven AI Wave

Transitioning to clinical adoption, Joshua notes that physicians are rigorously trained in evidence-based medicine and risk aversion, asking how leaders separate valid skepticism from pure resistance to change. Schwartz responds candidly: clinicians—like all humans—fundamentally dislike change. People order the same dish at restaurants despite extensive menus and take the same commute to work even when navigation apps suggest a two-minute shortcut. When told an AI will triage patients or suggest diagnoses, clinicians reflexively dismiss it with "the AI is not that good."

However, physician behavior is being forced to adapt not from top-down mandates, but from an external consumer wave: - Primary care clinics are observing measurable drops in routine visits because patients are utilizing consumer AI tools for at-home self-diagnosis. - Patients no longer arrive with a few printouts from a search engine; they walk in with their problems pre-analyzed, differential diagnoses compiled, and explicit lists of prescriptions they want approved.

At the same time, clinician-facing AI tools that remove friction are witnessing viral adoption. Schwartz points out that two years ago, almost no one had heard of OpenEvidence, whereas today many physicians feel they "cannot survive without OpenEvidence." Doctors are so eager to access streamlined decision support that they willingly accept business models where data is shared with pharmaceutical partners or community advertising is introduced—signaling a massive shift in how physicians consume technology and recommendations at the point of care.

[11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) Clinician Skepticism, Radiologist Anxiety, and Radical Transparency

When introducing AI triage tools to clinicians, enthusiasm is rarely the default reaction. Radiologists in particular voice deep anxieties about whether AI is being brought in to replace them. While Roberta Schwartz emphasizes that replacement is not the goal, she is candid about the reality: AI will fundamentally alter the way they practice medicine and conduct their daily business. Clinicians inherently approach emerging technological claims with "10 pounds of skepticism and one pound of optimism and excitement," making the rare individuals who enter the field with enthusiasm a true anomaly.

Addressing how leadership must communicate these disruptive shifts to thousands of employees across a large health system, Schwartz advocates for uncompromising honesty over comforting platitudes. She frequently observes audience members approaching her after keynotes to praise her candor, only to immediately whisper, "Did you have to say that?" Despite the discomfort it can cause, Schwartz insists that full transparency is the only viable path forward. Healthcare organizations must actively involve their workforce in designing and adopting these transformations rather than disguising the scale of impending change.

[12:51](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=771s) The Entry-Level Career Pipeline and Transitioning to Expert Teams

While mid-career and senior personnel generally possess the institutional knowledge to adapt, retire, reskill, or transition into emerging roles as natural attrition takes its course, Schwartz expresses profound concern for new graduates. Foundational entry-level positions—such as front-desk registration staff and appointment schedulers—traditionally served as the primary training grounds where young professionals gained industry literacy and initial workplace experience. As these routine administrative tasks are automated, those traditional career stepping stones are rapidly disappearing.

```mermaid flowchart TD accTitle: The AI Capability Gap and Career Ladder Disruption subgraph Traditional["Traditional Healthcare Career Pipeline"] direction TB T1["Entry-Level Roles (Schedulers, Registrars)"] --> T2["Mid-Level Staff"] --> T3["Domain Experts"] end subgraph AIEra["AI-Driven Healthcare Model"] direction TB A1["Automated Tier (AI Agents, Bots)"] -. "Missing On-Ramp".-> A2["Expert Tier (Human + AI Hybrids)"] end ```

This structural shift introduces a fundamental training dilemma: non-expert, task-based tiers are increasingly handled by autonomous systems, while high-value tiers require sophisticated human-AI collaboration. The central leadership challenge is figuring out how to develop a workforce from baseline non-expert to domain expert when the entry-level rungs of the ladder no longer exist. Schwartz stresses that leaders must communicate the arrival of AI "often, loud, again, again, in every different kind of mode, in every different kind of medium," ensuring staff are supported and educated for the next generation of hybrid jobs rather than left behind.

- **Non-Expert Layer:** Repetitive, rules-based administrative and operational tasks automated by standalone AI agents. - **Expert Layer:** Advanced diagnostic, clinical, and operational workflows managed by multidisciplinary human-AI collaborative teams. - **Communication Mandate:** Relentless, transparent, multi-channel messaging to prepare employees for reskilling rather than concealing automation roadmaps.

[15:00](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=900s) Business Automation vs. Autonomous Clinical Visions

The immediate, low-friction frontier for healthcare AI remains business operations. Schwartz recalls a telling moment when a finance team member asked her directly, "Are you trying to botize me?" to which she frankly replied, "I am actually." Today, operational automation is well integrated at Houston Methodist Hospital; callers dialing into the hospital are just as likely to interact with their conversational AI bot agent, "Zoey," as they are with a human scheduling coordinator.

```mermaid flowchart LR accTitle: Spectrum of Healthcare AI Maturity subgraph Operational["High Maturity (Operational)"] direction TB O1["Call Center Scheduling ('Zoey' AI Agent)"] O2["Routine Back-Office & Business Tasks"] end subgraph Transition["Moderate Maturity (Exception Handling)"] direction TB T1["Prescription Routing & Delivery Exceptions"] end subgraph Clinical["Early Infancy (Clinical Judgment)"] direction TB C1["Autonomous Diagnostic Packages (Full-Body MRI + Labs)"] C2["Algorithm-Generated Compounding"] end Operational --> Transition --> Clinical ```

Clinical operations represent a far more controversial frontier. Schwartz recounts an intense, two-hour executive debate at a C-suite healthcare forum sparked by a speaker claiming that annual medical exams will eventually consist entirely of blood work and a whole-body MRI scanned directly by AI. Under that proposed model, algorithms would independently determine diagnoses and formulate customized medication regimens delivered straight from an automated compounding pharmacy. While that extreme vision illustrates the radical outer edge of the automation continuum, Schwartz emphasizes that complex clinical decision-making remains in its earliest infancy and requires rigorous human judgment, standing in stark contrast to mature administrative and banking workflows.

[17:31](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1051s) Demographic Pressures, Specialist Shortages, and Cultural Transformation

The urgency to operationalize AI is compounded by an impending demographic crisis. As the massive Baby Boomer cohort ages—ranging from newly turned 65-year-olds to those in their mid-eighties—and the number of centenarians expands, the demand for medical care will drastically outstrip the available supply of physicians, nurses, and healthcare aides. Hands-on practitioners like surgeons retain high occupational security, but medical specialty access is already severely constrained. Schwartz notes that even high-profile individuals regularly reach out to her executive office begging for assistance in securing appointments with neurologists, urologists, and gastroenterologists. Without aggressive AI integration to augment clinical capacity, healthcare risks exacerbating systemic inequities, creating severe divides between those who can access timely care and those who cannot.

Joshua Robert humorously remarks that the safety of tactile fields will influence him to choose surgery for his upcoming residency specialty, while Schwartz applauds Robert's initiative in founding the *Journal for Artificial Intelligence in Medicine* (JAIM) so early in his medical career. Closing the discussion, Schwartz issues a clear call to action for healthcare leaders: they must actively "lean in," learn the underlying technologies, write and experiment with code directly, and guide their workforces through cultural change with sustained time, patience, energy, and genuine empathy.

- **Baby Boomer Surge:** A rapidly expanding aging demographic (ages 65 to 85+) and growing centenarian population placing unprecedented strain on existing clinical infrastructure. - **Provider Supply Deficit:** Critical shortages spanning medical doctors, registered nurses, and allied healthcare aides across both rural and urban health ecosystems. - **Specialty Bottlenecks:** Pronounced access barriers in high-demand specialties including neurology, urology, and gastroenterology. - **Actionable Advice:** Leaders must achieve personal technical literacy, experiment hands-on, and lead organizational culture change with empathy and commitment.

## Notable quotes > [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) it's okay, we won't get rid of anybody — Human Resources leadership

> [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) Hey, how are you getting rid of people? — Chief Financial Officer

> [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) So what you're telling me is you want $200,000 more and the same amount of people? Yes. No. — Roberta Schwartz

> [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) the AI is not that good — Clinicians

> [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) cannot survive without OpenEvidence — Roberta Schwartz

> [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 10 pounds of skepticism and one pound of optimism and excitement — Roberta Schwartz

> [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) Did you have to say that? — Audience member

> [12:51](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=771s) often, loud, again, again, in every different kind of mode, in every different kind of medium — Roberta Schwartz

> [15:00](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=900s) Are you trying to botize me? — Finance team member

> [15:00](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=900s) I am actually. — Roberta Schwartz

## Mentioned Joshua Robert, Roberta Schwartz, Journal for Artificial Intelligence in Medicine (JAIM), Houston Methodist Hospital, Center for Innovation at Houston Methodist, Icahn School of Medicine at Mount Sinai, HLTH, ViVE, Aidoc, OpenEvidence, Zoey

## Video description Joshua Robert - Founding Editor-in-Chief, Journal for Artificial Intelligence in Medicine (JAIM) Roberta Schwartz - EVP and Chief Innovation Officer, Houston Methodist Hospital

Many health systems have moved past experimentation. The challenge now is operationalizing AI responsibly at scale. This fireside chat presents a real case study of how Houston Methodist Hospital evaluates, prioritizes, and deploys artificial intelligence across clinical and operational environments.

We examine where AI has delivered measurable value, where implementation stalled, and the organizational and technical barriers encountered along the way. The conversation also covers governance, clinician adoption, and how leadership plans to expand AI use safely over time.

Attendees will leave with practical insights into what enables enterprise AI deployment and how leaders can approach their next AI initiative.

[Source](https://www.youtube.com/watch?v=XUeGwUSyKPk)

中文简报

走入休斯敦卫理公会医院：AI在医疗实践中的落地部署 Roberta Schwartz 博士阐述了休斯敦卫理公会医院严格的评估框架，强调企业级落地需要广泛的平台化能力，而非孤立的单点解决方案。她深入探讨了人力资源部门在人才留任上的宣传诉求与首席财务官（CFO）缩减成本预期之间的内部博弈，指出放射科分诊和排班智能体等 AI 即使没有缩短单次影像判读时间，也能有效精简诊疗流程。Schwartz 还指出，随着行政任务实现自动化，传统的初级职位晋升通道正在被打破，并将成熟的运营型机器人与处于早期的自主临床概念进行了区分。最后，结合人口老龄化压力和专科医生短缺的现状，她呼吁医疗行业的领导者应掌握切实的技术素养，并以极度透明和共情的态度来引领团队转型。在推介的医疗 AI 解决方案中，仅有约 1% 能获得医院的试点机会，因为医疗系统正在大力筛选淘汰单点方案以及同质化的 DIY 智能体平台。 (0:04) 狭隘的单点解决方案会造成难以逾越的集成和维护开销；企业级落地需要具备架构包容性的平台，以便在多项临床和行政工作流中进行横向扩展。 (1:54) 不能用单一铁板一块的框架来评估 AI 的投资回报率；例如，急诊分诊 AI 产生价值的途径是重新编排任务优先级并加快后续诊疗流程，而非加快放射科医生个人的单次阅片速度。 (4:33) 部署生成式和智能体 AI 会在组织内部造成紧张局面，使得人力资源部门对工作稳定性的承诺与 CFO 对编制缩减及预算效益的可量化要求产生冲突。 (6:00) 消费级 AI 浪潮正从外部加速推动临床医生的采纳，许多患者在就诊时已手握算法给出的诊断以及具体的用药诉求。 (8:40)

Reading summary

According to Dr. Roberta Schwartz, Chief Innovation Officer at Houston Methodist Hospital, enterprise adoption of artificial intelligence requires moving past single-use point solutions to scalable platforms, navigating internal conflicts over return on investment, and transparently managing workforce disruption amid worsening demographic shortages. Filtering Point Solutions for Architectural Scalability (0:04) Healthcare systems are inundated with vendor pitches, but Schwartz notes that only about 1% of solutions secure an enterprise pilot. Narrow point solutions—such as the more than 1,000 discrete AI algorithms developed for radiology alone to detect specific findings like lung nodules or rib fractures—create unsustainable procurement, integration, and maintenance overhead if acquired piecemeal. Instead of adopting standalone tools, enterprise leaders evaluate whether an AI platform possesses the broad architectural 'aperture' to expand across multiple clinical or operational domains. Measuring Value and Managing Institutional Tensions (4:33) Measuring the return on investment (ROI) for healthcare AI requires distinct frameworks tailored to specific workflows. Schwartz highlights Aidoc's pulmonary embolism tool: while it did not reduce the per-scan reading duration for individual radiologists, it prioritized positive cases to the top of the queue, giving emergency departments operational certainty that unflagged scans were likely negative within 30 minutes. When evaluated strictly by per-scan reading time, the tool appeared ineffective, yet it delivered significant emergency room workflow acceleration. Clinician Inertia and the Consumer Wave (8:40) Clinicians often exhibit natural resistance to workflow changes, typically responding to new predictive or diagnostic tools with skepticism. However, physician adoption is being accelerated by patient behavior. Primary care clinics report drops in routine visits because patients use consumer AI applications to pre-analyze symptoms, generate differential diagnoses, and bring requested drug lists directly to appointments. Simultaneously, clinician-facing decision-support tools like OpenEvidence have seen rapid uptake because they directly remove friction from daily practice, even when built on models that monetize through data sharing or pharmaceutical sponsorships. Workforce Disruption and the Vanishing Entry Level (12:51) Operational and administrative workflows, such as call center scheduling through Houston Methodist's bot 'Zoey,' represent the lowest-friction adoption areas for healthcare AI. In contrast, complex autonomous clinical diagnostics remain experimental and require deep human oversight. Automating routine administrative roles, however, introduces a training dilemma: positions like schedulers and front-desk registrars have historically functioned as the foundational training ground for junior personnel. Automating these baseline tasks removes traditional career entry rungs, obligating leaders to design alternative pathways that train non-expert staff for hybrid human-AI roles. Demographic Pressures and Specialty Care Bottlenecks (17:31) Schwartz asserts that adopting AI is urgent due to the aging Baby Boomer demographic (ages 65 to 85+) and an expanding centenarian population, which are generating clinical demand that outstrips provider capacity. Critical shortages already affect nursing, allied health, and medical specialties like neurology, urology, and gastroenterology, making timely access difficult even for high-profile patients. Without AI integration to expand throughput, Schwartz warns that systemic divides in timely care will worsen. Schwartz concludes that successfully integrating AI in healthcare requires executive leaders to achieve personal technical literacy, directly experiment with emerging tools, and actively guide staff through cultural shifts with open communication rather than concealing automation plans.

精炼阅读摘要

休斯敦卫理公会医院（Houston Methodist Hospital）执行副总裁兼首席创新官 Roberta Schwartz 指出，医疗系统采纳人工智能时仅有不到1%的方案能通过严选，医院必须优先布局具备跨领域扩展能力的平台架构，并在化解财务与人力利益冲突、打破临床惯性阻力及应对老龄化劳动力短缺中实现务实落地。筛选标准与单点方案的平台化扩展 (1:54) Schwartz 表示，目前医疗 AI 市场充斥着大量针对细分场景的单点解决方案。仅在放射科领域，就有超过一千种算法分别针对胸片肺结节、肋骨骨折或视网膜病变进行检测。对于大型医疗系统而言，逐一采购、集成与维护数百个孤立的专用工具会带来难以承受的管理开销和集成负担，除非某个工具能够解决极其紧迫的重大临床痛点，否则绝大多数都会被直接拒之门外。因此，能够进入前1%并获得试点资格的供应商，核心在于其技术是否具备跨领域延伸的“架构广度”（Aperture），即底层平台从单一临床切入点横向扩展至相邻流程的能力。面对市场上大量同质化的智能体构建工具，医院不再依赖销售推广或个别科室的单方诉求，而是由信息技术、临床专家及行政运营团队联合开展征求建议书（RFP）式的市场横向对比，对全行业进行基准评估后再择优合作。投资回报模型分歧与内部组织张力 (4:33) Schwartz 强调医疗 AI 不存在统一适用的投资回报（ROI）模型，不同场景必须制定专属衡量框架。以用于肺栓塞（PE）筛查的 Aidoc 为例，尽管放射科医生反馈单张扫描的阅片耗时并未减少，但系统自动将阳性病例置顶重排了工作流；急诊科由此获得关键确定性，若30分钟内未收到放射科通知即可大概率推断扫描为阴性，从而大幅压缩急诊决策链条并缩短住院时间（LOS）。然而，生成式与代理型 AI 的落地在医院管理层引发了显著目标冲突。人力资源部门倾向于向员工强调“不会裁员”以维系团队心理安全感，首席财务官（CFO）则明确要求通过缩减人员编制和削减硬性运行成本来兑现软件价值。Schwartz 强调领导层必须保持战略定力，例如当某业务部门申请20万美元预算采购 AI 工具以“处理更多以往顾不上的工作”，却拒绝削减任何岗位编制时，她当即否决了该预算，明确要求高额技术投入必须带来相匹配的劳动力效率转型。临床惯性阻力、患者倒逼与彻底透明 (8:40) 面对 AI 的介入，受过严格循证医学训练且高度规避风险的临床医生天然抱持抵触与怀疑态度。Schwartz 指出，医生本能地习惯固定工作路径，初期常以“算法不够成熟”为由排斥技术分诊或辅助诊断，放射科医生更对可能被技术取代抱有深层焦虑。Schwartz 明确说明 AI 的目的并非替代医生，但必然会从根本上改变日常行医流程与业务逻辑。与此同时，临床行为正受到外部消费级 AI 浪潮的强烈倒逼。初级诊疗机构观察到常规就诊量出现可测量的下滑，越来越多患者在家通过消费级 AI 完成自诊，并带着预先生成的鉴别诊断和处方清单来到诊室。在医生端，能直接消除工作摩擦的辅助工具亦迎来爆发式采用，如临床决策支持平台 OpenEvidence 在两年内便成为许多医生难以割舍的工作伴侣。Schwartz 主张领导者在向全院推进转型时必须坚持彻底坦诚的沟通原则，直面变革的震荡与焦虑，而非依赖虚妄的安抚之词。运营自动化落地与基层职业阶梯断裂 (12:51) 相较于尚处婴儿期的复杂临床决策，医院的业务与行政运营是 AI 落地阻力最小的成熟领域。休斯敦卫理公会医院已在呼叫中心等前端部门广泛部署对话式 AI 智能体“Zoey”，患者拨打医院电话时由 AI 或人工协调员处理预约挂号的概率已各占一半，后台财务等重复性工作也在加速自动化。对于全套血液化验结合全身 MRI 由算法自主诊断并自动调配药物的激进构想，Schwartz 认为这仅代表自动化远景的极端末端，复杂的医疗判断短期内仍需人类经验严格把关。然而，后台与前台任务的自主化给医疗人才梯队带来了严峻结构性断层。传统上，挂号登记与预约排班等基层入门岗位是青年从业者熟悉医疗业务、完成职业起步的基础跳板。随着这类规则化任务被智能体全面接替，传统的晋升阶梯基层断裂，而高阶岗位则直接要求人机协作的复合经验。Schwartz 警示，领导层必须解决在基础岗位消失后如何将新人培养为资深专家的培养困境，并高频、多渠道地对员工进行再培训与转岗引导。老龄化人口危机与专科医疗供给瓶颈 (17:31) 医疗系统加速拥抱 AI 的最根本现实推力在于不可逆转的人口老龄化危机。随着庞大的“婴儿潮”（Baby Boomer）一代步入65岁至85岁以上高龄，百岁老人群体不断扩大，医疗照护需求已呈爆发式增长，而注册护士、内科医生和护理人员的供给缺口持续拉大。虽然外科等依赖高触感操作的专业具备较高防护度，但神经科、泌尿科和胃肠科等高需求专业已出现严重的就医准入瓶颈。 Schwartz 指出，当前专科挂号资源极度短缺，甚至高管人脉也频繁被寻求用于协助预约稀缺专科。如果不能通过 AI 深度介入来实质性扩充临床承载力，医疗系统势必进一步分化并加剧结构性不公，使普通患者更难获得及时诊疗。Schwartz 呼吁医疗管理层必须主动跨入技术底层，亲自动手学习与尝试编程，以深度的同理心、时间投入与耐力带领组织完成文化与流程重塑。 Schwartz 论证了医院推进人工智能必须从追逐单点噱头转向评估平台级扩展能力，并通过彻底坦诚的管理沟通在行政自动化与临床辅助中实现渐进落地。然而，如何解决基层岗位自动化所导致的年轻人才晋升阶梯断裂，以及如何在 CFO 削减编制的要求与维持员工心理安全感之间达成动态平衡，依然是医院长期面临的核心制度挑战。

Detailed notes — zh

## Background 该内容由 Ai4 频道于 2026 年 8 月 15 日发布在 YouTube 平台。对话记录了《人工智能医学杂志》（JAIM）创刊主编 Joshua Robert 与休斯敦卫理公会医院执行副总裁兼首席创新官 Roberta Schwartz 博士的炉边对谈，重点探讨医疗机构如何在临床和运营环境中负责任地评估、推进与规模化落地 AI 技术。 ## TL;DR 休斯敦卫理公会医院分享了医疗 AI 选型与落地的核心经验，明确拒绝维护成本过高的孤立单点工具，优先青睐具备横向拓展能力的技术平台。在 ROI 衡量上，医院强调需紧扣临床具体用例并平衡降本与员工情绪之间的组织博弈；与此同时，患者深度自诊与新型决策支持工具正自外向内倒逼医生重构传统诊疗工作流。未来医疗 AI 将从成熟的行政运营稳步迈向临床深水区，这既带来了年轻医生培养断层的挑战，也是化解老龄化海啸与医疗资源严重短缺的唯一出路。 ## Atmosphere & style 整场对话气氛务实、幽默且极具穿透力，主宾互动融洽而直奔痛点。Schwartz 博士以罕见的坦诚撕开行业粉饰，毫不避讳地直面内部政治冲突、岗位替代以及变革阻力等敏感话题。整场讨论既有令人会心一笑的生动比喻，又始终贯穿着对人口老龄化与医疗公平性危机的深切紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=4s) 休斯敦卫理公会医院连续 15 年位列德克萨斯州最佳医院榜首，在高度风险厌恶的医疗体系中探索临床与运营 AI 的规模化落地路径。 - [01:16](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=76s) 医院拒绝维护成百上千个高度碎片化的单点算法，仅有具备底层技术延展性（Aperture）、能横向扩展多场景的 1% 供应商能够突围进入医院试点。 - [01:16](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=76s) 面对大量功能趋同的 AI Agent 搭建平台，医院需要依靠 IT、临床及业务专业专家的严格把关，并通过全市场 RFP 竞标流程进行筛选。 - [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) 医疗 AI 不存在通用 ROI 模型，价值衡量必须严格锚定具体用例（如缩短住院日或实质性降低重症死亡风险），而非仅停留在细微的时间提前量。 - [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) AI 系统（如 Aidoc）通过重构影像判读优先级列表而非缩短单张读片时长，有效优化了急诊到放射科的整体临床急救链条。 - [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) 落地生成式 AI 伴随激烈的内部利益博弈：HR 强调不裁员安抚士气，CFO 追求编制精简，科室则希望追加预算并保留原有编制，需要管理层具备定力以强制推行流程变革。 - [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) 患者自诊行为正倒逼临床工作流重塑，患者带着 AI 生成的病情分析与处方需求就医，直接导致初级保健门诊就诊量出现历史性下滑。 - [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) 以 OpenEvidence 为代表的专业临床辅助工具在医生群体中迅速普及，以病历减负和信息检索切入的工具正不可逆地重塑传统诊疗模式。 - [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 临床医生普遍带着怀疑态度审视 AI 并担忧被取代，管理层必须采取高度透明坦诚的沟通策略，明确流程重塑的必然性而非粉饰太平。 - [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) AI 正在吞噬前台接待与调度等初级入门岗位，导致医疗行业面临从零经验新人晋升为“人类+AI”协同专家的职业培养断层。 - [14:42](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=882s) 医疗 AI 的应用重心正从高度成熟的后台商业运营（如电话智能语音助手 Zoey 处理半数预约）逐步迈向探索期的临床决策与自动化处方深水区。 - [17:22](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1042s) 老龄化海啸加剧了医护人员短缺，大规模推广部署 AI 是化解优质医疗资源挤兑与两极分化的必然出路，而非可有可无的选择。 - [18:52](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1132s) 管理者推动医疗 AI 必须躬身入局尝试技术工具并弄清底层逻辑，同时投入长期的耐心与关怀以实现组织文化的深层蜕变。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=4s) 会议开场与休斯敦卫理公会医院的 AI 落地背景《人工智能医学杂志》（JAIM）主编 Joshua Robert 首先向现场观众致意，并拉开本次炉边对话的序幕。他指出，在关乎患者生死存亡的医疗健康领域，行业普遍具有极高的风险厌恶倾向，如何将 AI 真正落地到临床与运营实践中是一项极其复杂的挑战。本次对话邀请到了休斯敦卫理公会医院（Houston Methodist Hospital）执行副总裁兼首席创新官 Roberta Schwartz 博士。Schwartz 博士不仅一手创立了该院的创新中心，此前还曾担任西奈山伊坎医学院的业务拓展总监，并荣获《现代医疗》（Modern Healthcare）全美 25 位杰出女性领袖之一。Joshua 特别提到，就在活动当天，休斯敦卫理公会医院达成了连续 15 年蝉联德克萨斯州排名第一医院的里程碑成就。随后，Joshua 直奔主题，探讨医院如何对 AI 产品进行准入审批，以及在无数自称能“颠覆医疗”的 AI 初创公司中，究竟是什么特质让仅有 1% 的企业脱颖而出并成功进入医院试点。 - **主持人**：Joshua Robert，JAIM（Journal for Artificial Intelligence in Medicine）创刊主编。 - **对话嘉宾**：Roberta Schwartz 博士，休斯敦卫理公会医院 EVP 兼首席创新官。 - **医院背景**：截至对话当天，休斯敦卫理公会医院已连续 15 年稳居德克萨斯州最佳医院榜首。 --- ### [01:16](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=76s) 医疗 AI 选型准则：拒绝狭窄单点工具，看重技术延展性 Schwartz 博士幽默地坦言，以往参加各类医疗创新大会（如 HLTH、ViVE 等）时，自己在展厅里走过一排排展位，能以极快的速度迅速做出“要/不要”的判断，但这次大会却让她感到前所未有的挫败与挑战。她自嘲道：“我承认我是个自甘落后的人，到现在还没自己‘氛围编码’（vibecoded）过任何东西。”她发现当下的 AI 行业正在向所有人兜售“DIY 自主构建（Do It Yourself）”的理念，这与以往大不相同。以往厂商通常会宣称自己解决了某个具体问题，但医院面临的最大痛点在于解决方案的切入点往往过于狭窄。仅在放射科领域，市面上就有上千种针对极其细分场景的算法——有的专门看胸片找肺结节，有的专门看肋骨骨折筛查骨关节炎，视网膜读片领域也有多套独立系统。然而，像休斯敦卫理公会医院这样的大型医疗机构，根本不可能引入并维护上千个各自独立的点状工具，除非该问题造成的痛点巨大且医院急于解决。因此，能够获得青睐并成功进驻医院的 1% 的企业，往往具备技术“光圈”（aperture）的延展能力：即使切入点很小，其底层技术也足够通用，能够顺畅横向拓展至多个临床与运营场景。面对展厅中上百家功能看似大同小异的“AI Agent 搭建平台”，Schwartz 强调管理层必须依靠 IT、市场等内部专业领域专家的把关，并通过严格的全市场调研与类似 RFP（征求方案书）的竞标流程来做出最终决策。 ```mermaid flowchart TD accTitle: 医院评估 AI 解决方案的准入逻辑 subgraph S1["市场现状与痛点"] A["大量高度细分的单点方案 （如单一结节、单部位骨折筛查）"] --> B{"医院面临的运维瓶颈 （无法承载成百上千套离散工具）"} end subgraph S2["准入考量与筛选标准"] B --> C["具备底层技术延展性（Aperture） 可横向扩展多场景"] B --> D["极度迫切、非解不可的重大痛点"] end subgraph S3["落地评估流程"] C --> E["内部专家评估（IT / 临床 / 业务）"] D --> E E --> F["全市场 RFP 竞标与综合选型"] end ``` - **单点方案过载**：仅放射影像领域就有超过 1,000 种针对单一病种的细分解决方案，大型医疗系统无法单独对接并实施每一个微小方案。 - **技术延展能力**：医院更青睐具备底层技术架构可扩展性的平台型技术，能够从一个单点问题扩展覆盖到多个临床科室或业务线。 --- ### [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) 投资回报率（ROI）的多维考量与组织内部的利益博弈谈及 AI 项目的投资回报率（ROI）评估时，Schwartz 强调并不存在通用的标准 ROI 模型，每一款产品的价值衡量必须严格锚定其具体的临床或业务用例。例如在肺栓塞（Pulmonary Embolism, PE）的诊断中，评估的核心指标在于是否切实缩短了住院日（Length of Stay）。如果算法声称能“提前 3 小时发现 PE”，医院就必须甄别这 3 小时究竟只是提前发现了一个偶然病灶，还是能够真正抢救脑细胞、挽救生命。她分享了一个关于 Aidoc 在放射科上线的真实案例：放射科医生起初抱怨该 AI 系统并没有缩短自己的读片时间，质疑其宣称的“更早发现病灶”。Schwartz 则向医生解释，双方的说法其实都没有错——医生的单张影像判读时间确实没变，但 AI 彻底重构了阅片优先级列表，将判定为阳性的危急病例自动置顶。这样一来，急诊科（ED）医生如果 30 分钟内没收到放射科通知，就能大致判断影像为阴性；若 30 分钟内收到报告，则意味着存在紧急情况。系统在没有改变读片时长的情况下，实质性优化了整个临床急救链条。在生成式 AI（Generative AI）与 Agent 的 ROI 落地过程中，医院往往面临复杂的内部政治与利益平衡难题： - **人力资源负责人（HR）的诉求**：希望对外安抚员工情绪，强调“引入 AI 绝对不会裁员”。 - **首席财务官（CFO）的诉求**：直接施压询问“引入 AI 究竟能帮我们裁减多少编制”。 - **业务科室的博弈**：Schwartz 曾与医院内部某个部门沟通，该部门盛赞新方案能完成大量以往顾不上的工作，但当被问及“既然效率提升，是否可以精简人手”时，对方断然拒绝，甚至提出“既要多给 20 万美元采购系统，又必须保留原班全部人马”，这遭到了 Schwartz 的坚决驳回。要在保持员工士气、推动组织变革与达成严苛的财务回报之间取得平衡，需要领导层具备极强的定力与穿透力，明确设定预期、建立衡量标准并强制推动工作流程变革。 --- ### [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) 临床采纳的阻力：从医生的习惯抗拒到患者与新工具的倒逼针对 Joshua 提出的临床医生对证据高度依赖、天生风险厌恶且难以改变工作流的问题，Schwartz 博士直言不讳地指出：“人们自以为喜欢改变，但事实是根本没有人喜欢改变。”这就如同去最喜欢的餐厅总是点同样的老菜谱，通勤时明知导航提示有其他路线能省两分钟却依然选择习惯的道路一样，临床医生在接受变革方面往往格外迟缓。然而，临床工作模式正在被外部一股巨大的“患者消费化浪潮”所颠覆： ```mermaid flowchart LR accTitle: 医疗工作流与临床采纳的双向推动机制 subgraph Sub1["外部推动力量（患者端）"] P1["患者使用 AI 工具自诊"] --> P2["初级保健门诊就诊量下降"] P2 --> P3["患者带着诊断清单和处方需求就诊"] end subgraph Sub2["内部推动力量（工具端）"] T1["OpenEvidence 等专业 AI 工具普及"] --> T2["极高依赖度并改变信息检索方式"] T2 --> T3["AI 生成病历与文书减负"] end P3 --> C["重塑临床实践流程与医生决策模式"] T3 --> C ``` 1. **初级保健就诊模式剧变**：行业内首次出现了初级保健（Primary Care）门诊量下滑的趋势。患者不再像过去那样仅带着几个从 Google 搜来的问题问诊，而是先通过 AI 工具完成深度自诊，就诊时直接递上自己整理好的病情分析、潜在解决方案以及希望医生开具的处方清单，从外部强力挑战并重构医生的传统权威与诊疗节奏。 2. **专业临床 AI 工具的爆发式普及**：两年前几乎无人知晓的 OpenEvidence 等临床决策支持工具，如今已成为许多医生日常工作中“离不开”的必备助手。甚至为了获取这些强大工具的使用权限，医生社群开始逐渐接受在其平台上嵌入广告，或者以数据交换为代价获取服务。 3. **工作流重塑的渐进性**：从 AI 辅助文书生成到改变临床诊疗推荐，虽然让医生完全信任并采纳 AI 临床决策建议仍具挑战，但从文书减负和信息检索切入的工具，正在以不可逆的态势重塑临床实践。 ### [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 临床医生的质疑与 AI 时代的职场代际鸿沟 Roberta Schwartz 坦言，当管理层告诉放射科医生要用 AI 辅助分诊时，医生们其实并不买账。放射科医生最普遍的担忧是：“你们是不是想取代我们？”虽然 Schwartz 明确表示无意取代他们，但毫无疑问会彻底改变他们的工作流程。她风趣地形容：“临床医生带着 10 磅的怀疑和 1 磅的乐观兴奋走进来”，而能够积极拥抱技术的人是罕见的少数。谈到如何向成千上万名员工传达 AI 变革而不引发恐慌时，Schwartz 强调“坦诚（Honesty）和透明高于一切”。许多人在演讲后会跑来对她说：“天呐，我太佩服你的坦诚了”，但紧接着又心有余悸地问：“你真的非得把大实话全说出来吗？”Schwartz 认为必须直面现实，因为 AI 浪潮正在扑面而来。她最担忧的不是现有老员工——他们可以通过适应、转岗、培训或退休来平稳过渡——而是刚刚步入职场的年轻毕业生。以往诸如前台接待、预约调度等供新人积累经验的初级入门岗位正在被 AI 吞噬。 ```mermaid flowchart TD accTitle: AI时代人才层级演进与面临的挑战 subgraph Transition["技能层级与培养断层"] A["初级/非专家层级 （传统入门岗如调度、登记）"] -->|"AI 自动化替代"| B["AI 独立承担基础工作"] A -.->|"断裂的晋升通道：如何培养新人？"| C["专家层级 （人机协同决策）"] end ``` - **初级岗位萎缩**：AI 正在接管“非专家级”工作，而高价值的“专家级”工作则演变为“人类 + AI”的协同模式；现在的核心挑战在于如何把零经验的新人培养成专家。 - **高频透明的沟通**：管理层必须“大声、反复、利用一切渠道和媒介”进行坦诚沟通，不能粉饰太平；如果机构不帮助基层调度员提升技能以适应下一阶段的工作，那就是管理层的失职。 --- ### [14:42](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=882s) 未来 5 年的医疗 AI：从商业运营到临床决策的颠覆展望未来五年，Schwartz 指出当前的落地重心仍在相对容易的商业与后台运营（Business Operations）。她分享了一个趣事：曾有一位财务员工问她：“你是不是想用机器人把我取代了（botize me）？”她坦承：“其实我确实是这么打算的。”所幸该员工随后按计划退休了。但随着大数据与大模型的发展，真正的深水区是临床运营（Clinical Operations）。 Schwartz 提到了一个高管论坛上引发剧烈争议的激进构想：未来人们的年度体检可能精简为“一次全身 MRI 扫描加一套血液生化检测”。所有数据直接输入 AI 系统，由 AI 诊断、自动开具处方并直接对接药房配药寄送到家，完全省去传统的人工看诊环节。这一话题让在场领袖足足争论了两个小时。尽管临床全流程自动化目前仍处于极其早期的萌芽阶段，不可与成熟的调度机器人同日而语，但它清晰勾勒出了技术演进的方向。 | 业务领域 | 落地成熟度 | 实际应用现状 / 未来构想 | |:--- |:--- |:--- | | **商业/行政运营** | 高度成熟，全面落地 | 休斯敦卫理公会医院的 AI 智能语音助手 **Zoey**，患者拨打电话时由 Zoey 或人工坐席处理预约的概率已各占一半。 | | **临床运营与处方** | 处于探索与萌芽期 | 自动化处方纠错、全身影像 + 血液多模态 AI 联合诊断及个性化自动配药管线。 | --- ### [17:22](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1042s) 应对老龄化海啸与医疗资源匮乏的唯一出路针对主持人提到的医生职业倦怠问题，Schwartz 认为 AI 接管行政杂务固然能带来缓解，且外科医生等高度依赖手工操作的岗位目前最安全，但整个医疗体系正面临更为严峻的宏观人口危机——“婴儿潮（Baby Boomers）”一代的加速老龄化。随着最年轻的婴儿潮一代迈向 65 岁、年长者步入 85 岁以上高龄，百岁老人日益增多，现有的医生、护士和辅助护理人员数量根本无法支撑庞大的照护需求。 Schwartz 警告，如果不能快速推进 AI 技术在医疗体系中的规模化落地，社会将不可避免地陷入更加严重的医疗资源“两极分化（have and have-not）”： - **资源挤兑日常化**：即便是有良好医保覆盖或社会地位优越的人群，如今也常常需要通过私人关系辗转求助医院高管，只为了抢到一个神经科、泌尿科或胃肠科专家的号源。 - **弱势群体困境加剧**：在无保/低保人群以及偏远乡村地区，医生短缺已演变为系统性危机。大规模部署 AI 并非可选项，而是维持医疗体系运转的必然出路。 --- ### [18:52](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1132s) 寄语与总结：躬身入局与文化重塑在对谈尾声，主持人 Joshua Robert 开玩笑说鉴于就业保障，自己明年选专科时打算选外科；Schwartz 则由衷称赞 Joshua 作为医学生能够创办《人工智能医学期刊》（JAIM）的远见与魄力。最后，Schwartz 向所有与会者分享了她对推动医疗 AI 的核心建议： 1. **躬身入局（Lean in）**：管理者必须强迫自己深度参与技术实践，甚至亲自动手尝试编写代码或使用工具，先搞清楚底层逻辑，才能有效制定员工的培训与转型策略。 2. **文化重塑需要长期投入**：驱动组织和员工跨出舒适区拥抱新技术是一项极其艰巨的任务，需要领导者付出“时间、耐心、充沛的精力以及大量的爱与包容”，唯有如此才能真正完成组织文化的深刻蜕变。 ## Notable quotes > [01:16](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=76s) 我承认我是个自甘落后的人，到现在还没自己‘氛围编码’（vibecoded）过任何东西。 — Roberta Schwartz > [04:33](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=273s) 既要多给 20 万美元采购系统，又必须保留原班全部人马 — 业务科室部门负责人 > [08:40](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=520s) 人们自以为喜欢改变，但事实是根本没有人喜欢改变。 — Roberta Schwartz > [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 你们是不是想取代我们？ — 放射科医生 > [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 临床医生带着 10 磅的怀疑和 1 磅的乐观兴奋走进来 — Roberta Schwartz > [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 坦诚（Honesty）和透明高于一切 — Roberta Schwartz > [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 天呐，我太佩服你的坦诚了 — 医院员工 > [11:37](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=697s) 你真的非得把大实话全说出来吗？ — 医院员工 > [14:42](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=882s) 你是不是想用机器人把我取代了（botize me）？ — 财务部门员工 > [14:42](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=882s) 其实我确实是这么打算的。 — Roberta Schwartz > [18:52](https://www.youtube.com/watch?v=XUeGwUSyKPk&t=1132s) 时间、耐心、充沛的精力以及大量的爱与包容 — Roberta Schwartz ## Mentioned Joshua Robert, Roberta Schwartz, 休斯敦卫理公会医院 (Houston Methodist Hospital), 《人工智能医学杂志》（Journal for Artificial Intelligence in Medicine / JAIM）, 西奈山伊坎医学院, 《现代医疗》（Modern Healthcare）, HLTH, ViVE, Aidoc, OpenEvidence, Zoey, Ai4 ## Video description Joshua Robert - Founding Editor-in-Chief, Journal for Artificial Intelligence in Medicine (JAIM) Roberta Schwartz - EVP and Chief Innovation Officer, Houston Methodist Hospital Many health systems have moved past experimentation. The challenge now is operationalizing AI responsibly at scale. This fireside chat presents a real case study of how Houston Methodist Hospital evaluates, prioritizes, and deploys artificial intelligence across clinical and operational environments. We examine where AI has delivered measurable value, where implementation stalled, and the organizational and technical barriers encountered along the way. The conversation also covers governance, clinician adoption, and how leadership plans to expand AI use safely over time. Attendees will leave with practical insights into what enables enterprise AI deployment and how leaders can approach their next AI initiative. [Source](https://www.youtube.com/watch?v=XUeGwUSyKPk)

Source quotes

[1:16] Roberta Schwartz: 我承认我是个自甘落后的人，到现在还没自己‘氛围编码’（vibecoded）过任何东西。 [4:33] 业务科室部门负责人: 既要多给 20 万美元采购系统，又必须保留原班全部人马 [8:40] Roberta Schwartz: 人们自以为喜欢改变，但事实是根本没有人喜欢改变。 [11:37] 放射科医生: 你们是不是想取代我们？ [11:37] Roberta Schwartz: 临床医生带着 10 磅的怀疑和 1 磅的乐观兴奋走进来 [11:37] Roberta Schwartz: 坦诚（Honesty）和透明高于一切 [11:37] 医院员工: 天呐，我太佩服你的坦诚了 [11:37] 医院员工: 你真的非得把大实话全说出来吗？ [14:42] 财务部门员工: 你是不是想用机器人把我取代了（botize me）？ [14:42] Roberta Schwartz: 其实我确实是这么打算的。 [18:52] Roberta Schwartz: 时间、耐心、充沛的精力以及大量的爱与包容

Source quotes

[4:33] Human Resources leadership: it's okay, we won't get rid of anybody [4:33] Chief Financial Officer: Hey, how are you getting rid of people? [4:33] Roberta Schwartz: So what you're telling me is you want $200,000 more and the same amount of people? Yes. No. [8:40] Clinicians: the AI is not that good [8:40] Roberta Schwartz: cannot survive without OpenEvidence [11:37] Roberta Schwartz: 10 pounds of skepticism and one pound of optimism and excitement [11:37] Audience member: Did you have to say that? [12:51] Roberta Schwartz: often, loud, again, again, in every different kind of mode, in every different kind of medium [15:00] Finance team member: Are you trying to botize me? [15:00] Roberta Schwartz: I am actually.

Original transcript

[0:04] I hope we're all fed and we're all good to take a lot of information today. Today I'm Joshua Robert, [0:08] as you just said, editor-inchief of the journal for artificial intelligence medicine. [0:12] We'll be going over deploying AI in practice, how you can use AI in medicine, healthcare when [0:16] patients lives on the line, and real use cases of that. We'll be going over various things with [0:21] the various nuances of that, especially when in healthcare and medicine you're very risk adverse [0:26] and we want to talk about how you can tackle that. And we have the very best speaker for that. Meet [0:31] Dr. Roberto Schwarz. She is the chief innovation officer and executive vice president of Houston [0:36] Methodist Hospital. She led and started the center for innovation there. She was previously [0:41] the director of business development at Icon Med at Mount Sinai. Also awarded top 25 women leaders [0:46] in modern healthcare. And as of actually today, Houston Methodist is the number one hospital in [0:51] Texas for 15 years straight. So congrats. Good day today. Yeah, very good day today. Thank you for [0:56] being here. And let's get right into it. I want to talk about how AI actually gets approved. How [1:01] do you get your foot in the door in a hospital for this? And let's actually go very broad. You [1:06] probably have a lot of people come in. They'd say, "Oh, I have the most revolutionary AI. It's going [1:10] to change medicine today." But what separates the 1% that get a pilot from the other 99%. [1:16] Yeah. You know, there I mean, it's it's amazing. I I wrote to a friend of mine actually today um [1:22] and said to her, I am I'm officially like I can I can normally walk into a conference, health, [1:28] Vive, I mean anything you name it, and I can go down the row and I can go yes, no, yes, no, yes, [1:33] no, yes, no, like super quick. And this conference is kicking my ass. Um I'm just going to tell you [1:39] like I don't know how to build my own. I I admit I'm I'm selfprescribed pathetic. I have not yet [1:46] vibecoded my own anything. Um uh but I I think that what this conference is is so different [1:54] than any others because I think what they're saying to our people now is do it yourself [2:00] right which is very different than anything has been in the past. So normally uh people come to us [2:06] and say we have solved X problem and when they say they've solved X problem the question really [2:12] becomes how narrow is the problem that they're solving. So, as an example, there's 1,000 over [2:19] I I think it's 1,000, but there may be more now. Um, different solutions in the radiology space, [2:26] and I can solve a chest X-ray looking for a lung nodule or I can solve the rib fracture looking for [2:35] osteoarthritis or the bone, right? Whatever it is, um, I can give you six solutions in the [2:41] retinal space or whatever it is. The problem is is that a company like us cannot speak [2:47] and implement 10,000 or 1,000 different solutions that all start and solve pinpoint little problems [2:55] unless those problems have such large pain points at which point we are just desperate to solve that [3:03] that issue that problem is so big for us. So I would tell you that the companies that tend to get [3:10] in are ones that even if they solve one problem have the aperture to be able to open that aperture [3:17] and solve multiple different solutions with the underlying technology that they have. So the [3:23] technology that underlies it is interesting enough that if they start in one area can go to multiple. [3:30] Now there are probably a hundred companies on the floor where I can build my AI agents and I have [3:36] absolutely no idea having stopped at five booths this morning and listening to five pitches. I have [3:42] no idea what separates these companies. And so we really do rely on the experts in whether it's IT [3:49] or marketing or another space to separate and say that is one like I'm very good about saying like I [3:55] hear it and I actually think this is useful but we rely on our experts within Methodist to say yes or [4:02] no. And then we try to force them to do a market like look at the entire market and make sure that [4:09] like an RFP process we would understand everybody who's out there and what they do. Yeah, that's [4:15] definitely very interesting. I never realized that. I guess I haven't thought that maybe niching [4:19] yourself down could also hurt you in the foot when you're trying to grow or grow further. You know, [4:23] whenever you're actually evaluating these companies and deciding on that, do you also look [4:27] at like ROI? How do you evaluate specifically which ones are the best for your use cases? [4:33] So the the ROI is definitely that's that's the term that everyone wants to use, right? Security [4:40] ROI, like there are like 12 buzzwords and and the ROI is really it's amazing. And people will [4:46] say to me like I want your ROI model and I'm like there's not one ROI model because for each [4:52] solution that you put in, you are looking for that ROI in that use case for what you're doing. So, [4:59] if I'm looking at um PE, pulmonary ambolism, I'm going to be looking at is it a length of stay and [5:07] it is a true length of stay. If they say to me, you know, hey, we can find the PE 3 hours earlier. [5:14] What's the real impact of finding it 3 hours earlier? Is it like who cares? Like that was [5:19] a nice incidental finding or like that's going to save brain cells, that's going to save lives, [5:24] that's going to save whatever it is. Um it was interesting because I had a a disagreement. [5:29] They're just some recent examples, but I can if there's any of the spaces you want me to go into [5:34] where my radiologist was like, "This is not we we put in a do recently and he was saying to me, um, [5:43] it's not changing my reading time at all." And he said, "I don't know why a A do is claiming [5:50] to have found these things much earlier because my reading time has not changed." And I was like, [5:55] "You're right and they're right. So, I'm not going to disagree. So, you're right in that your reading [6:01] time for a scan has not changed. They're right in that they've changed the fundamental order [6:08] by which your scans are read. So, anything they've deemed positive is going to the top of the list. [6:13] So, I am getting a different product even if your reading time is not different. So if the ED doctor [6:23] basically doesn't hear from the radiologist in 30 minutes, he's pretty sure it's negative, [6:28] right? If he's heard from him in 30 minutes, he's he knows there's something going on. [6:32] So it's changing a lot of the ways that we practice. So your ROI model for each one of these [6:38] has to be very different. The question really for institutions is how diligent are they about [6:45] setting the expectation on what they're going to get out of the products and then measuring it and [6:50] making damn sure you get it. And a genic being the most difficult one that you're going to do because [7:00] depending on who you talk to, they want you to give a different set of talking points about your [7:05] ROI. So the head of HR wants you to say to your people, it's okay. We won't get rid of anybody. [7:14] And your CFO is like, hey, how you getting rid of people? Right? So each one [7:20] has a different way that they want you to spin the outcome of what you're talking about. [7:26] And so you have to be really careful to um to figure out how to thread the needle [7:34] between not freaking your people out completely and getting them to um adopt and or change. So um [7:46] so I went to one group uh within our hospital and I said to them, I think the solution meets a lot [7:52] of needs and their answer was it can do so much more that we couldn't get to. And I was like so [7:57] how many less people will you be doing that with? And they're like oh no less people. I was like so [8:05] what you're telling me is you want $200,000 more and the same amount of people? Yes. No. [8:14] Right. And so having that fortitude to force the change in the way you work, that's that's [8:21] really difficult when you're trying to keep your culture high and your people excited about this [8:26] and you want to keep them working in the way they did before. I mean being on the other side of that [8:32] and like on the other side of the table, these are great insights to learn from and how to build and [8:36] how to plan for the future too. So thank you for that. And then from there I wanted to also go into [8:41] clinical adoption. So after approval going into actually getting this into the workflow into the [8:45] process I'm in my medical training right now all physicians were trained to be very evidence-based [8:50] reasoning you do research you're very riskadverse patient lives on the line how do you distinguish [8:55] that healthy skepticism from actually maybe like being resistance to change or adverse to it [9:01] so as I talk to our employees and and spend all the time my comment to them straight out is you [9:06] think you like change and the answer is you do not like change if you go to your favorite restaurant, [9:12] you generally order the same thing, even though there's 12,000 other things on the menu, right? [9:17] You drive to work the same way every day most of the time. Even though Google sometimes tells you, [9:23] you can save 2 minutes, you're like, "Hey, I'm going to do what I know." Right? So, people don't [9:27] like that change. And let me just promise you that clinicians are really bad at adopting change. Um, [9:35] so if we basically tell them, look, the you have to convince them over time, the literature tells [9:42] you to do something different. Great. When you say to the clinician, hey, I'm going to triage this [9:48] problem for you, and the AI is going to tell you what's wrong with your patient. The first thing is [9:54] they're like, the AI is not that good. But in reality, the consumers are forcing that change. [10:03] So what you're seeing for the first time is declines in primary care visits. [10:11] You're seeing people use the AI based solutions to do self diagnostic and they are coming to [10:18] their appointments not with what used to be a few questions from Google. They're coming [10:23] in primary care doctors will tell you they're coming in with all the problems that they have, [10:28] all the solutions that they have, and all the meds they would like prescriptions for. [10:33] Right? Like it is a fundamental change. And so consumers, you have this consumer wave rushing to [10:40] like change the way the doctors are operating and question everything that the doctors are doing, [10:46] right? And you've got new tools. I mean, two years ago, who had ever heard of open evidence? Now, [10:54] if I talk to any one of my doctors, they cannot survive without open evidence. And you're like, [10:59] it's, you know, you're giving all of your data and everything that you're doing to [11:04] the pharmaceutical companies. They're like, yeah, bring it on. I talked to one company recently to [11:08] try to understand their model with it. And for the first time, they're allowing advertising within [11:13] their physician community just to get a hold of some of the tools, right? Tools are telling you, [11:17] you give us your data, we'll give you XYZ. It's fascinating to see that that as these tools [11:25] make it easier for physicians and we haven't even talked about AI generated documentation [11:30] physicians are are changing the clinical practice recommendations it's a little bit harder [11:37] me telling them that basically we're going to triage things for them not super excited [11:44] um one of the concerns of the radiologists is are you replacing us right I don't think I'm replacing [11:49] them, but I am going to change the way they do their business. And that that's that's hard. And [11:55] clinicians come with 10 pounds of skepticism, one one pound of optimism and excitement. [12:03] The ones that come the opposite way around are the rare breed. Yeah. You being one of them. Yeah. [12:08] Honestly, you segue amazing to my next question for you. It's actually about leadership and how [12:12] you tackle these hard challenges, especially with so many employees, so many beds under you. [12:17] whenever AI it essentially has some stigma as well around like how it can maybe tackle like change [12:22] some jobs employees things like that how do you communicate AI to thousands of employees without [12:27] creating that fear um a lot of times people come up to up to me after a talk and they're [12:32] like oh my god I love your honesty and then right after they say I love your honesty comes the like [12:39] did you have to say that like it's amazing to me like it's it's always stunning [12:44] and I go for honesty over anything else because I believe that transparency is the way to go. Um, [12:51] but I think we need to really get our people in healthcare, other industries to basically [12:58] accept the fact that we need them involved in these changes. The people I feel the worst for [13:04] are the young graduates. Actually, I think most of the people in our companies or a lot of the people [13:09] in our companies will adjust and or retire and or switch and or get themselves educated. But the [13:14] younger people coming out that used to have those jobs where they could be the register at the front [13:19] desk or the scheduler or the everything to get that experience, those jobs are the ones that are [13:25] disappearing. So I think it's incumbent upon us to work with that next generation of people just [13:30] coming into the workforce and figure out how to get them in the workforce and how to get them to [13:37] more expert levels because we are our non-expert levels are AI. our expert levels are going to be [13:44] human combos, right? And so as you do that, the question is how do you how do you get the human [13:51] from non-existent to expert, right? And that's that's going to be where we're all challenged [13:56] today in this this transition time. Most people are in pretty good good shape and we're tritting [14:03] out. We're tritting out at a time where other jobs are going to be created. How to create the [14:07] jobs for those younger people. That's probably the one I worry about the most. As for communication, [14:13] often loud again again in every different kind of mode, in every different kind of medium [14:19] and quite honestly with honesty. Don't don't try to lie about this stuff like it's coming. [14:26] It's waving. It's waving in different places and and if we don't help that scheduler [14:32] to get educated for the next job, then you've done them a complete disservice. [14:38] I'm glad you're in the position you are. Honesty is the number one policy. Yeah. [14:42] And then I guess it's a very important conversation when looking at the future as well. [14:46] Whenever you're talking about AI in the healthcare system, AI with doctors, whether they work [14:51] together, replace things like that. What do you see as being maybe 5 years from now? The biggest [14:56] changes in AI and healthcare. What does the future look like in your opinion that you can see? [15:00] Yeah, everyone will tell you the easiest stuff right now is the business operations, right? The [15:04] business operations are the easiest. Although I I can remember we had a finance person who um [15:10] retired, which is probably a good thing because his comment to me is, "Are you trying to botize [15:15] me?" And I was like, "I am actually, but okay." Um he retired. Uh so we're good. Thumbs up. Um [15:23] uh but but I think that that now it's business operations. The next part of this with big data [15:31] comes clinical operations. And that's where um where you put in a we had an [15:39] insanely controversial conversation at a C forum like seale forum of leaders [15:45] in healthcare and the guy came in and his comment was the future of a [15:51] your annual exam will be lab work and a whole body MRI and you will need nothing else and I was like [16:01] I'm not sure like so we we probably talked about this for 2 hours and debated whether or not [16:08] like did you just really need a whole body MRI and a lab like your blood scanned and that was [16:14] going to tell you everything you needed to know for the next year and like that that information [16:19] would automatically go into an AI which would spit out the drugs you would need that would [16:23] spit out basically the formula that would get in sent into a pharmacy that would formulate [16:28] your drugs and basically send them on and here was your packet for the year right and And I was like, [16:33] uh, wow, that's like I don't know. Um, but if you think about it as that is the most controversial [16:39] of like statements you're making, the question is where are we on that continuum and where does [16:45] judgment match with what you're doing? I don't know. But what I would say to you if that that [16:53] process is in its infancy, that's very different than my banking function. That's very different [16:59] than my scheduling function. If you call into the hospital now, you're equally as likely to get our [17:04] AI bot agent Zoey as you are to get a human to basically do your scheduling. But when it comes to [17:12] um you know uh something was sent to the wrong pharmacy, something was sent to the wrong place, [17:17] that kind of stuff is on the early end with the rest of it on the later end. Well, it's I guess [17:22] it's really good news, especially with physician burnout being a big deal right now. Administration [17:26] tasks and taking care of that with AI could help a lot with that. So, thank you so much. Yeah, [17:31] the people most protected at this point in time are your surgeons, your people who are going to [17:35] really be that hands-on, but I can tell you we are waving through the baby boomer population. [17:41] You're about to wave through, right? The youngest ones are coming up on that 65 with older ones in [17:46] their 85. We do not have enough doctors to take care of this population. Just straight up, [17:52] there are not enough doctors. There are not enough people. There are not enough nurses. There's not [17:56] enough healthcare aids. So if we basically say that we are now going to have a population it's [18:00] going to be living till they're some of them into their hundreds more and more centennians [18:05] centennials um but but if you basically look at that with a population matched with the [18:11] healthcare workforce if this stuff does come out into does not come into being very quickly [18:17] we're just not going to have enough people and then we will have a have and have not [18:22] of society in a very different way. It doesn't matter if we have enough Medicaid. You already [18:29] have your problems in the underinsured population and in the rural population. Now imagine it in, [18:35] you know, whatever it is. The amount of people that call me as the CEO to try to get into a [18:40] neurologist, a urologist, a GI physician, and basically beg me to figure out how to get them [18:46] an appointment. And those are super smart people who, you know, just need a connection to get in. [18:52] I mean, we we have to do this. Yeah, thank you so much for all of your insights and I've definitely [18:58] learned a lot. I hope you all have too. When I go to choose my specialty next year, I'll [19:01] choose surgery, I think, because there's some job security. And I do want to brag on you. I mean, to [19:05] basically be at your stage and start a journalism in AI and medicine is absolutely incredible. So, [19:11] kudos and congratulations to you. Thank you so much, Dr. Schwarz. And thank you so much [19:15] for being here today. We're definitely very lucky to have you. And I guess if you have [19:20] one last message to everybody here, what's your last message about AI and medicine? Um, I think [19:26] what I say to people all the time is which everyone's doing here, right? Lean in. Um, force [19:31] yourself to do it. Force yourself to do your own code. Force yourself to do that. Uh, and [19:36] then you figure out what you're going to have to teach your employees how to do. Um, now many of [19:41] you are sitting in the audience going, I did that last night. I I do not do that routinely. But um [19:46] but I think it's forcing your employees and the people around you to do that is really really [19:53] hard and it takes time, patience, energy and a lot of love to um really change the culture. [20:01] Definitely. Thank you so much Dr. Schwarz. Could we get a round of applause for today?