Izzo Urges Hospitals to Sandbox Clinician AI Tools to Stop Shadow IT

When Everyone Can Build: Securing Healthcare in the Age of AI With San Joaquin General Hospital — Ai4

Ai4

Banning clinician-built software drives medical staff toward unmonitored shadow AI, requiring health systems to establish tiered synthetic sandboxes to safely operationalize frontline prototypes.

AI-assisted summary of When Everyone Can Build: Securing Healthcare in the Age of AI With San Joaquin General Hospital — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background Published by Ai4 on YouTube on August 19, 2026, this presentation features Dr. Joseph Izzo, Chief Medical Information Officer at San Joaquin General Hospital. The talk explores how modern AI coding agents empower clinicians to build domain-specific software and outlines how healthcare organizations can safely balance rapid frontline innovation with enterprise cybersecurity and governance.

## TL;DR Joseph Izzo explains that while AI coding agents allow non-technical clinicians to build functional clinical prototypes in hours, visual polish often conceals severe security, access management, and HIPAA compliance deficiencies. He argues that traditional gatekeeping and rigid waterfall reviews either extinguish valuable domain-driven ideas or drive clinicians toward unmonitored shadow AI. To resolve this dilemma, Izzo proposes a tiered governance framework featuring exploratory 'Lane Zero' synthetic sandboxes, EHR build environments, and technical pairing between clinicians and informatics leaders to securely shepherd prototypes into clinical production.

## Atmosphere & style The talk carries a candid, pragmatic, and constructive tone, shaped by Izzo's rare dual perspective as both an active clinician and a computer scientist. Moving engagingly from lighthearted coding mishaps—like buggy Pong paddles and tabletop dice—to high-stakes clinical informatics, the narrative delivers a reassuring yet clear-eyed message. Watching it feels like a collaborative strategy session that transforms institutional anxiety into an actionable, bridge-building roadmap for healthcare innovation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=4s) AI coding agents have democratized rapid prototyping for clinicians, but functional-looking prototypes frequently fail on edge cases, workflow ergonomics, and security perimeters. - [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) Domain experts with zero formal programming background can now build working clinical applications in a matter of hours, solving nuanced frontline problems. - [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) Healthcare faces an illusion of readiness where software functionality is immediately visible but production deficiencies—like missing authentication and audit logging—remain hidden. - [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) AI coding agents perform near 100% on syntax benchmarks but struggle significantly with context-dependent security flaws and standardized OWASP benchmarks. - [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) Generic AI prompts cannot replace organizational context, leading coding agents to miss least-privilege access rules, HIPAA audit requirements, and Business Associate Agreements. - [07:01](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=421s) Total IT blockades of clinician-built software fail to stop adoption, instead driving frontline workers toward unmonitored shadow AI and data exfiltration. - [07:01](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=421s) Subjecting software built in five hours to legacy eight-month waterfall governance cycles creates an unsustainable bottleneck for healthcare innovation. - [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) With only 18% to 25% of health systems possessing mature AI governance, CMIOs and informaticians must evolve into technical bridges between clinician builders and cybersecurity teams. - [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) Health systems should establish a tiered evaluation framework starting with 'Lane Zero,' an exploratory sandbox utilizing high-fidelity synthetic EHR data and vendor APIs. - [14:37](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=877s) EHR build domains with obscured data allow builders to iterate safely while leveraging native audit logging and progressive, proportional security reviews. - [14:37](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=877s) Advancing a tool to production ('Lane Two') requires hands-on technical pairing, named lifecycle ownership, active monitoring plans, and multi-domain clinical validation. - [16:00](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=960s) Rather than suppressing grassroots AI development out of fear, healthcare leadership must construct safe, well-governed pathways to channel clinical domain expertise into patient care.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=4s) From Toy Projects to Healthcare: The Rise of Coding Agents and the Illusion of Readiness

Joseph Izzo opens by bridging his distinct worlds: he is an addiction medicine physician, an emergency doctor, and a hospital administrator in California, but his foundational background was in computer science and programming before entering medicine. Moving deeper into medical administration traditionally meant letting programming skills atrophy due to time constraints, but the rapid emergence of AI coding agents over the past six months has completely transformed the landscape, making rapid prototyping accessible once again. To illustrate this shift, Izzo showcases three personal side projects, each highlighting a distinct failure mode of rapid prototyping:

First is a nostalgic recreation of the classic game Pong with a few visual upgrades; while charming, the ball behaves erratically when striking the paddle at specific angles because nobody tested the edge cases. Second is a 3D digital dice roller built for tabletop role-playing games when someone forgets their physical dice at home. While mathematically sound and geometrically polished in a way that would previously have taken hours to hand-code, tabletop gamers universally disliked it because they crave the physical, tactile experience of holding and rolling real dice—meaning the application failed on fundamental user workflows. Finally, bridging into healthcare, he built an educational HL7 converter designed to parse the non-human-readable, pipe-delimited data standard into structured buckets for medical students and residents. While perfectly adequate on `localhost` with synthetic examples, pointing such a prototype at real clinical systems immediately exposes severe deficiencies around identity access management, privacy boundaries, and security perimeters.

```mermaid flowchart LR accTitle: Prototype Capabilities versus Production Deficiencies subgraph S1["What Coding Agents Build Quickly"] A["Convincing UI & Demos"] --> B["Synthetic Data Pipelines"] B --> C["Localhost FHIR Endpoints"] end subgraph S2["What Prototypes Lack for Production"] D["OAuth 2.0 & Token Gating"] E["Enterprise Secrets &.env Hygiene"] F["HIPAA Audit Trails & SDLC Ownership"] G["Multi-Domain Clinical Validation"] end S1 -.->|"Fails Under Real-World Conditions"| S2 ```

- **Pong Prototype:** Demonstrates that rapid builds look functional on the surface but frequently fail on untested edge cases. - **TTRPG Dice Roller:** Proves that technically sound software can completely miss human and cultural workflow requirements. - **HL7 Interoperability Converter:** Illustrates how localhost tools crumble when exposed to real-world healthcare security and identity perimeters.

---

[01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) The Healthcare AI Hackathon and the Stigmatizing Language Demo

At a recent healthcare AI hackathon in San Francisco featuring approximately 150 invited participants, Izzo observed a phenomenon that would have been impossible six months prior: the vast majority of attendees were frontline clinician builders—including bedside nurses, physicians, physical therapists, and respiratory therapists—who had zero formal programming background. Using modern coding agents, these domain experts created and submitted working, convincing software prototypes within five hours, solving nuanced clinical problems that previously would never have progressed past an idea in someone's head.

To anchor the discussion, Izzo presents a research tool he built for an addiction medicine collaboration rather than a commercial product. Using a coding agent, he generated convincing synthetic patient records and stood up a Fast Healthcare Interoperability Resources (FHIR) server—which translates legacy pipe-delimited HL7 into modern JSON-based APIs—in less than ten minutes. The application ingests clinical notes and scans for stigmatizing phrasing that reduces patients to their conditions, flagging terms like "known alcoholic" and recommending patient-first alternatives such as "patient with alcohol use disorder," or highlighting judgmental charting like "known addict" and "he chose to drink again."

When shown to clinical colleagues and research collaborators across multiple institutions, the immediate, enthusiastic reaction was: "Let's roll this out. Let's put this into production. Let's get it in my clinic." Yet despite its polish, Izzo emphasizes that he would never deploy it in his own hospital in its current state. The prototype relied on a development FHIR server with no production authentication model, no OAuth 2.0 implementation, no token-based access gating, exposed secrets stored insecurely in `.env` files, zero production monitoring, no assigned software lifecycle owner, unvetted cross-site scripting (XSS) risks, and no clinical validation across other vulnerable specialties such as obstetrics or trauma. Healthcare now faces a pervasive dilemma: "function is visible but readiness is not," creating intense pressure to deploy tools that appear complete but harbor hidden systemic risks.

---

[05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) The Gap Between Syntax and Security: Why Coding Agents Fail Healthcare Context

Empirical research and academic benchmarks evaluating modern coding agents reveal a stark divergence between code appearance and code security. While LLM coding benchmarks demonstrate that agents approach nearly 100% accuracy in producing syntactically correct, highly functional, and visually polished code, their performance against standardized OWASP security benchmarks remains alarmingly low. Furthermore, security performance deteriorates significantly as the necessary architectural context expands; coding agents perform reasonably well on isolated, localized flaws like basic SQL injection, but struggle dramatically with context-heavy vulnerabilities like cross-site scripting.

When asked why developers cannot simply prompt the coding agent to "make the application secure," Izzo explains that genuine security depends entirely on organizational and regulatory context that an external model cannot infer. In the stigmatizing language prototype, for example, the coding agent defaulted to giving any logged-in physician global access to every patient record across the system.

```mermaid flowchart TD accTitle: The Context-Dependent Security Requirements in Healthcare AI subgraph Context["Organizational & Regulatory Context"] R1["Identity & Role-Based Access Control"] R2["Data Exfiltration Controls & BAA Agreements"] R3["HIPAA Audit Logging Requirements"] R4["SDLC Ownership & Clinical Failure Modes"] end subgraph Blindspot["AI Coding Agent Blindspot"] B1["Generates Valid Syntax & Polished UI"] B2["Lacks Knowledge of Institutional Risk Boundary"] end Context -->|"Must Be Explicitly Governed"| Blindspot ```

- **Identity and Access Governance:** Agents cannot determine the appropriate least-privilege access boundaries without explicit institutional scoping. - **Data Exfiltration and BAA Compliance:** Off-the-shelf integration with commercial frontier models routes protected health information (PHI) outside organizational perimeters without a Business Associate Agreement. - **HIPAA Audit Trails:** Federal privacy statutes require comprehensive, tamper-evident audit logging for every PHI access event. - **Clinical Validation & SDLC:** Models cannot self-assign maintenance lifecycles, monitor clinical drift, or establish failure modes across diverse clinical specialties.

---

[07:01](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=421s) The Pitfalls of Total Blockades and Outdated Governance

Healthcare organizations have historically exercised rigid gatekeeping over electronic medical record (EMR) systems, allowing IT leadership to block unapproved software completely. When health systems apply legacy, role-based heuristics—declaring that clinicians simply have no business building software—two damaging outcomes emerge:

1. **Lost Innovation:** Clinicians with decades of lived domain experience abandon novel ideas that could genuinely resolve frontline workflow bottlenecks. 2. **Proliferation of Shadow AI:** Driven not by malice, but by the reality that enterprise-approved tools fail to solve immediate workflow pain, clinicians route around IT controls by pasting PHI into personal frontier AI accounts, utilizing unsanctioned third-party EMS charting plugins, or spinning up unmonitored cloud deployments.

```mermaid flowchart TD accTitle: Consequences of Healthcare AI Governance Pathways A["Clinician Builds Working Prototype"] --> B{"Institutional Response"} B -->|"Total Blockade / No Pathway"| C["Idea Abandoned (Lost Innovation)"] B -->|"Total Blockade / High Friction"| D["Shadow AI & Data Exfiltration"] B -->|"Supported Evaluation Pathway"| E["Safe Production Integration"] ```

The alternative traditional approach—forcing fast prototypes through standard waterfall governance—is equally dysfunctional. Recalling a federal grant project he coded five years ago, Izzo describes an eight-month bureaucratic slog involving weekly IT alignment, granular FHIR resource justifications, endless Microsoft Visio architecture diagrams, and extensive security reviews. While thorough, that model was designed for an era when hand-coding was slow and iterative; applying an eight-month evaluation process to an application generated in five hours creates an unsustainable bottleneck.

To highlight the need for flexible governance, Izzo contrasts two recent security reviews: - **Team A (Mature Clinician Builders):** A small team that pre-mapped their security controls, utilized autonomous compliance tooling, and arrived with thorough, pre-documented answers for every standard architecture question. - **Team B (The Supported Evaluation Candidate):** A solo cardiologist who discovered that feeding specific cardiac arrhythmia data into an algorithm yielded unprecedented diagnostic insights. The cardiologist possessed a brilliant, clinically validated prototype but had zero understanding of enterprise deployment, OAuth, or secrets management. Under legacy rules, this clinician would be turned away, yet their clinical innovation represents immense patient value that demands an institutionally supported bridge to production.

---

[11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) Transforming Healthcare Governance: The CMIO as Translator and "Lane Zero" Sandboxing

Addressing this paradigm shift requires evolving the governance model from a static policy checklist into an adaptable operating culture. Today, only 18% to 25% of healthcare organizations possess what industry surveys classify as a mature AI governance model, primarily because legacy governance operates as rigid bureaucratic box-checking rather than a dynamic operational framework. Chief Medical Information Officers (CMIOs), Chief Information Officers (CIOs), CISOs, and legal teams must expand their traditional roles: where clinical informaticians previously acted as translators between clinicians and EMR configurations (optimizing order sets, shortcuts, and favorites), they must now serve as technical and regulatory bridges between frontline clinician builders and enterprise security teams.

The ultimate objective of modern governance is to make sanctioned progress easier than working around the system. Rather than funneling all incoming software into a single high-friction compliance pipeline, health systems should establish tiered operational tracks, beginning with an exploratory "Lane Zero."

```mermaid flowchart LR accTitle: Tiered Evaluation Framework for Healthcare AI subgraph L0["Lane 0: Exploration & Education"] E1["Synthetic EHR Data"] E2["Approved Agent Tooling"] E3["Vendor API Sandboxes"] end subgraph L1["Lane 1: Supported Evaluation"] V1["Architectural Scoping"] V2["Security & Privacy Reviews"] V3["Multi-Domain Clinical Testing"] end subgraph L2["Lane 2: Production Deployment"] P1["OAuth 2.0 Authentication"] P2["HIPAA Audit Logging"] P3["Active SDLC Ownership"] end L0 --> L1 --> L2 ```

- **Lane Zero Sandboxing:** Provides a safe, isolated self-service environment where clinicians experiment and learn without production credentials or access to real patient data. - **High-Fidelity Synthetic Data:** Capitalizes on recent breakthroughs in coding agents and generative tooling to supply realistic, rubric-driven clinical notes and synthetic FHIR endpoints. - **Vendor Sandbox Integration:** Leverages existing EMR vendor API sandboxes to let clinicians test workflows against structured schemas before initiating formal organizational security reviews.

[14:37](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=877s) Safe Evaluation Environments and Bridging Clinician Builders to Deployment

Navigating the pathway from an initial prototype to a full deployment requires structured governance rather than an immediate, heavy-handed security roadblock. In contrast to early-stage experiments, solutions entering "lane two" arrive ready to deploy and operate. These mature implementations have often been rolled out in other organizations, possess a named owner, and have clearly articulated controls for clinical safety, security, privacy, and continuous monitoring, making them fully prepared for formal security reviews. However, bridging the gap between initial ideation and production readiness requires hands-on institutional support during the evaluation phase.

```mermaid flowchart TD accTitle: Staged Pathway for Clinician-Built AI Tools subgraph Ideation["Phase 1: Safe Experimentation"] A["Clinician Concept"] --> B["EMR Build Domain\n(Obscured / Retrospective Data)"] B --> C["Scope Control & Native Logging"] end subgraph Transition["Phase 2: Collaborative Review"] C --> D["Technical Pairing\n(CMIO & IT Department)"] D --> E["Focused Security & Safety Review"] end subgraph Production["Phase 3: Lane Two Deployment"] E --> F["Full Review & Named Ownership"] F --> G["Live Clinical Operation"] end ```

To support clinician builders without exposing real patient records to premature risk, organizations can leverage limited, retrospective data or synthetic environments provided by electronic health record (EHR) vendors. Many medical record systems offer a dedicated build level configured with obscured or non-real patient data, allowing builders to practice and iterate directly inside the environment they will eventually target. Working within this build domain provides major governance advantages: it enables tight scoping of data access, activates the platform's native audit and logging tools from day one, and allows for focused, staged reviews. Clinicians are paired directly with technical partners—such as the Chief Medical Information Officer's team—ensuring that those who understand clinical workflows and those who manage technical infrastructure work in tandem. Full enterprise security reviews are reserved for when a project has genuinely earned them, avoiding unnecessary bureaucratic friction early on.

- **Lane Two Criteria**: Requires demonstrated operational readiness, named ownership, established clinical safety, defined privacy controls, and monitoring plans. - **Build Domain Sandboxing**: EHR-provided build environments with obscured data allow realistic development while enforcing boundaries and scope constraints. - **Native Auditing**: Logging and audit tracking tools integrated into the build domain establish visibility long before live deployment. - **Technical Pairing**: Clinician innovators partner directly with the CMIO and informatics teams to translate working concepts into enterprise-grade tools. - **Proportional Governance**: Comprehensive security and compliance reviews are triggered progressively as tools advance in maturity rather than at the initial concept stage.

---

[16:00](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=960s) The Five-Hour Hackathon and Building Safe Lanes for AI Innovation

Reflecting on the transformative power of AI coding tools, a five-hour hackathon demonstrated how rapidly domain experts can now produce functional software. In just five hours, participants built working clinical applications; while most were minimum viable products (MVPs), each tool was genuinely functional and directly encoded the unique clinical experience and perspective of its creator.

Because artificial intelligence has become an accessible "engine for ideas" across healthcare, the responsibility of health system leadership and informatics teams is not to shut down grassroots development out of fear. Instead, the mandate is to construct safe, well-governed lanes that allow clinicians and technical teams to evaluate, refine, and deploy innovation securely.

- **Rapid Prototyping**: Clinicians were able to build functioning clinical MVPs in a single five-hour hackathon session using AI assistance. - **Domain-Driven Software**: Tools created directly by frontline clinicians embed nuanced workflow insights that traditional software specifications often miss. - **Institutional Role**: Healthcare organizations must build the organizational and security infrastructure ("the lanes") to safely guide clinician-led AI innovation from concept to patient care.

## Notable quotes > [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) function is visible but readiness is not — Joseph Izzo

> [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) Let's roll this out. Let's put this into production. Let's get it in my clinic. — Joseph Izzo

> [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) known alcoholic — Joseph Izzo

> [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) patient with alcohol use disorder — Joseph Izzo

> [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) known addict — Joseph Izzo

> [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) he chose to drink again — Joseph Izzo

> [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) make the application secure — Joseph Izzo

> [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) Lane Zero — Joseph Izzo

## Mentioned Joseph Izzo, San Joaquin General Hospital, Ai4, Pong, HL7, FHIR, JSON, OAuth 2.0, HIPAA, OWASP, Microsoft Visio

## Video description Joseph Izzo - Chief Medical Information Officer, San Joaquin General Hospital

Artificial intelligence has dramatically lowered the barriers to building useful software, empowering clinicians and other domain experts to solve problems that once required specialized technical expertise and dedicated development teams. Drawing on my experience building and governing healthcare software and AI tools before and after the rise of coding agents, I will explore how this transformation changed the way I think about innovation, cybersecurity, and organizational risk. The talk will explore how the AI era challenges the traditional assumption that innovation and cybersecurity must compete with one another. Rather than preventing clinicians and other domain experts from building, healthcare organizations have an opportunity to create pathways where new ideas are encouraged, safely evaluated, and responsibly integrated into patient care. Through practical examples, attendees will learn practical principles for balancing innovation with governance, cybersecurity, and patient safety without creating unnecessary barriers to progress. The goal is not to slow the next generation of builders, but to create healthcare organizations where innovation and security can succeed together.

[Source](https://www.youtube.com/watch?v=Ib-RTE1rGXg)

中文简报

Izzo敦促医院对临床医生AI工具实施沙箱化以遏制影子IT 禁止临床医生自建软件会迫使医务人员转向不受监控的影子AI，这需要医疗系统建立分级合成沙盒，以安全地将一线原型投入实际应用。 Joseph Izzo解释称，虽然AI编码智能体让没有技术背景的临床医生能在数小时内构建出可运行的临床原型，但外观上的完善往往掩盖了严重的安全、访问管理以及HIPAA合规缺陷。他认为，传统的把关模式和僵化的瀑布式审查要么扼杀了极具价值的领域驱动创意，要么促使临床医生转向不受监管的影子AI。为了解决这一困境，Izzo提出了一个分级治理框架，该框架包含探索性的“零号通道”（Lane Zero）合成沙盒、EHR构建环境，以及让临床医生与信息学主管进行技术结对，从而安全地引导原型转化为临床实际生产力。 AI编码智能体让临床医生的快速原型构建得以普及，但表面上可运行的原型往往在极端情况处理、工作流人机工效以及安全边界方面表现不佳。 (0:04) 完全没有正式编程背景的领域专家现在可以在数小时内构建出可用的临床应用程序，从而解决一线细微具体的问题。 (1:53) 医疗行业正面临一种“就绪假象”，即软件功能清晰可见，但生产环境层面的缺陷（如缺少身份验证和审计日志）却隐而不现。 (1:53) AI编码智能体在语法基准测试中的得分接近100%，但在面对依赖具体上下文的安全漏洞以及标准化OWASP基准测试时却表现得相当吃力。 (5:08) 通用的AI提示词无法替代组织上下文，这导致编码智能体往往会遗漏最小权限访问规则、HIPAA审计要求以及商业伙伴协议（BAA）。 (5:08)

Reading summary

Dr. Joseph Izzo argues that while AI coding agents empower frontline clinicians to build working software in hours, these rapid prototypes create an illusion of readiness that masks severe architectural, privacy, and regulatory vulnerabilities. To prevent clinicians from abandoning novel solutions or resorting to unmonitored shadow AI, healthcare organizations must replace rigid gatekeeping with tiered governance and secure sandboxing. The Illusion of Production Readiness (1:53) Modern AI coding agents allow domain experts—such as bedside nurses, physicians, and physical therapists with no programming background—to generate functional clinical applications in mere hours, as seen at a recent 150-participant healthcare hackathon. However, Izzo warns that visual polish often conceals critical structural deficiencies. Prototypes frequently collapse when confronted with untested edge cases, mismatched clinical workflows, and enterprise infrastructure requirements. To illustrate this gap, Izzo points to a prototype he built to flag stigmatizing clinical phrasing (such as substituting 'patient with alcohol use disorder' for 'known alcoholic'). Although generated in under ten minutes using synthetic FHIR endpoints, the application lacked OAuth 2.0 authentication, relied on exposed credentials in local environment files, possessed no token-based access gating, missed HIPAA-compliant audit logging, and lacked software lifecycle ownership. While observers urged an immediate clinical rollout, deploying such unhardened localhost prototypes into live hospital systems introduces profound security liabilities. Security Blindspots in AI Coding Agents (5:08) Empirical evaluations demonstrate that while AI coding agents achieve near-100% accuracy in generating syntactically valid code, their performance drops significantly against standard OWASP security benchmarks. Izzo notes that coding assistants handle localized flaws like basic SQL injection reasonably well, but struggle with context-dependent vulnerabilities such as cross-site scripting. Crucially, developers cannot resolve this issue by simply prompting an agent to make the application secure. Genuine enterprise security depends on institutional and regulatory context that external models cannot infer. Without explicit guardrails, coding agents default to insecure configurations, such as granting a logged-in clinician global access to all hospital records or routing protected health information to commercial APIs without an underlying Business Associate Agreement. The Pitfalls of Blanket IT Blockades (7:01) Historically, healthcare IT leadership has maintained strict control over electronic medical records by imposing total bans or subjecting new software to months-long waterfall reviews involving endless architecture diagrams. Izzo argues that applying an eight-month bureaucratic evaluation to software built in five hours creates an unsustainable bottleneck that drives two distinct failure modes: lost clinical innovation and the growth of shadow AI. When legitimate ideas are blocked, clinicians seeking to solve immediate operational friction circumvent IT entirely by pasting patient information into personal AI accounts or utilizing unsanctioned third-party plugins. Izzo contrasts a technically mature developer team with a solo cardiologist who discovered valuable diagnostic insights from arrhythmia data but lacked knowledge of enterprise secrets management. Turning away clinician innovators who lack computer science credentials discards immense clinical value, underscoring the need for an institutional bridge to production. Tiered Governance and Lane Zero Sandboxes (11:41) With industry surveys indicating that only 18% to 25% of healthcare organizations possess mature AI governance models, Izzo calls for clinical informaticians and Chief Medical Information Officers to serve as technical translators between clinician builders and cybersecurity teams. Instead of funneling all incoming tools into an identical high-friction compliance pipeline, health systems should establish tiered operational pathways. This framework begins with 'Lane Zero,' an exploratory self-service sandbox that pairs generative coding tools with synthetic patient data and EHR vendor API sandboxes. Clinicians can experiment and refine workflows safely without exposing real patient records. Promising projects then advance to supported evaluation lanes where clinicians partner directly with informatics teams to implement native logging, role-based scoping, and clinical validation before graduating to fully monitored, production-ready status. Izzo concludes that because artificial intelligence has democratized software creation, healthcare leadership can no longer rely on outright prohibition. The modern imperative for health systems is to build proactive, governed pathways that guide clinician-led breakthroughs safely from experimental code to compliant, live clinical operations.

精炼阅读摘要

加州成瘾医学与急诊医生兼医院管理者Joseph Izzo指出，AI编码智能体让一线医护人员能在数小时内构建功能完备的临床原型，但“功能可见性”掩盖了生产级合规与安全的缺失；医疗系统必须摒弃一刀切的封堵或僵化的瀑布式审批，转向建设分层沙箱与协同推进机制。编码智能体带来的生产就绪假象 (0:04) 在旧金山举办的一场约150名受邀者参加的医疗AI黑客松中，完全没有编程背景的临床医生、护士及治疗师借助现代编码智能体，在5小时内便交付了能够解决实际工作流痛点的可用原型。Izzo以自身开发的去污名化语言检测工具为例，该程序利用智能体在10分钟内搭建了快速医疗互操作性资源（FHIR）服务器与合成病例管线，能敏锐标出病历中如“已知酗酒者”或“选择再次饮酒”等具有偏见的陈述，并建议替换为客观中立的临床表述。这一原型在展示给各机构同行后获得了极高评价，临床人员纷纷呼吁立刻投入生产部署。然而Izzo强调，该项目表面上的成熟具有高度误导性，它依赖缺乏生产级身份认证与OAuth 2.0机制的开发端服务器，秘钥直接明文存放在环境配置文件中，且完全缺失跨站脚本攻击（XSS）防护、HIPAA合规审计跟踪、全生命周期维护责任人以及跨专科的临床有效性验证。医疗软件由此陷入“功能直观可见，但就绪状态隐蔽不可见”的系统性风险中。语法完备与医疗安全上下文断层 (5:08) 学术与行业基准测试表明，大语言模型生成的代码在语法正确性与界面交互上已接近100%的高水准，但在标准化OWASP安全基准测试中的表现却依然堪忧，尤其是涉及深层系统架构上下文的漏洞时，其防护能力急剧恶化。针对为何无法通过提示词让智能体直接生成安全应用的疑问，Izzo解释称真正的系统安全完全取决于机构特有的组织和监管上下文，外部模型根本无法凭空推断。例如在病历分析原型中，编码智能体在未获显式限制时，默认给所有登录医生开放了全院所有患者档案的全局访问权限，违背了最小权限原则。此外，直接调用商业大模型API会导致受保护健康信息（PHI）在缺乏商业伙伴协议（BAA）的情况下向外渗漏，且模型无法自动满足联邦隐私法案对于不可篡改审计日志的要求，更无法预见不同临床专科之间潜在的模型漂移与失效模式。传统封堵与官僚审查的双重失灵 (7:01) 面对临床自建工具的激增，医疗机构传统上往往采取基于角色的封锁策略，认定临床医生不应介入软件开发。Izzo指出这种简单粗暴的封堵会诱发两种负面后果：一方面，具备数十年临床一线经验的医护人员所构思的创新方案被迫搁浅；另一方面，迫于日常工作流程的效率痛点，医护人员会绕过IT监管使用“影子AI”，包括将PHI复制粘贴到个人 frontier AI 账户中，或私自挂载未经验证的第三方插件，造成更严重的合规漏洞。然而，传统的瀑布式合规治理在当下同样失效。Izzo回顾其五年前参与的一项联邦资助项目，审批流程经历了长达8个月的技术对齐、图表绘制与层层签字；将这种漫长流程套用在5小时内生成的AI工具上会造成无法承受的治理拥堵。例如一位心脏科医生独立研发出能从心律失常数据中提取开创性诊断见解的高价值模型，却完全不懂底层OAuth或秘钥管理；若按旧规直接驳回，医院将蒙受巨大的临床价值损失。渐进式分层治理与新型沙箱机制 (11:41) 行业调查显示，目前仅有18%至25%的医疗机构建立了成熟的AI治理体系，其余机构大多仍停留在机械的静态合规审查阶段。Izzo主张首席医疗信息官（CMIO）和信息团队应转变职能，从单纯调整医嘱模板与快捷操作的配置人员，转变为一线临床开发者与企业安全团队之间的技术与合规桥梁。现代治理的核心目标是让合规创新的阻力小于违规绕行的阻力，而非设立单一高门槛关卡。为此，医疗系统应推行分层管理框架：设立“零号通道”（Lane 0），向临床人员提供由合成EHR数据和供应商API沙箱构成的安全自服务环境，允许其在不触碰真实患者数据和生产凭证的前提下自由试验；中间阶段则利用电子病历厂商提供的混淆数据构建域（Build Domain）推进评估，并在技术团队协助下前置启用原生日志审计与边界管控；当应用满足临床安全性、隐私保护、OAuth 2.0认证并明确了运维责任人后，再顺利进入“二号通道”（Lane 2）正式投产。 Izzo总结认为，AI编码工具已经成为临床专家将深层业务认知转化为可用软件的核心引擎，医疗机构管理层的使命绝非出于风险恐惧而一律压制，而是必须建立安全可控的制度轨道；如何在保障隐私合规的前提下为临床自主创新提供持续的技术配对与生命周期维护，仍是全行业治理转型的关键课题。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 19 日发布在 YouTube 平台。视频记录了圣华金总医院（San Joaquin General Hospital）首席医疗信息官 Joseph Izzo 的主题演讲，探讨了 AI 编程智能体极大降低软件开发门槛后，医疗机构如何在鼓励临床专家自主创新的同时，重塑网络安全、合规审查与组织治理路径。 ## TL;DR 随着 AI 编程智能体的爆发，非技术背景的一线医护人员能够在数小时内迅速构建出直击临床痛点的功能原型，但这些看似完善的应用往往在鉴权、数据隔离与安全审计等深层“生产就绪”指标上存在严重盲区。面对这一技术变革，医院传统的“一刀切封堵”不仅扼杀创新还会催生危险的“影子 AI”，而耗时数月的传统瀑布式审批又已脱节。医疗机构亟需重构治理体系，通过设立基于合成数据的“第零车道”（Lane Zero）沙盒以及 CMIO 与临床构建者结对的支持型评估机制，在保障患者安全底线的同时为临床创新保驾护航。 ## Atmosphere & style 演讲兼具务实洞察与建设性前瞻，语调沉稳、理性且极具说服力。演讲者依托急诊医生、成瘾医学专家和前程序员的跨界视角，通过自嘲开发经历与鲜活的黑客松案例，生动勾勒出一线临床创新冲动与医院严苛安全底线之间的激烈张力。整场分享既避免了盲目乐观的技术崇拜，也摒弃了教条保守的官僚思维，为医疗数字化决策者带来了一种务实且富有启发性的治理破局感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=4s) 编程智能体极大地降低了软件开发门槛，但快速构建的高保真原型往往掩盖了边界用例测试不足、工作流脱节以及生产安全缺位的问题。 - [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) 临床医护构建的原型呈现出“功能显而易见，但生产就绪隐形于表象之下”的特征，通常缺失生产级 OAuth 鉴权、企业密钥管理和跨学科临床验证。 - [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) 代码智能体生成的代码在语法外观上接近完美，但在面对 OWASP 等网络安全基准测试时失分严重，且对上下文依赖越深防御能力越差。 - [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) AI 无法感知医疗机构特有的组织上下文，无法自主处理细粒度访问权限、HIPAA 合规审计追踪、PHI 数据防泄露及系统全生命周期责任归属。 - [07:11](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=431s) 医院若采取一刀切的封堵路线，不仅会扼杀宝贵的一线临床创新，还会迫使医护人员绕道使用未经合规审查的“影子 AI”（Shadow AI）。 - [07:11](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=431s) 动辄数月的手工时代瀑布式 IT 治理流程已与技术现实脱节，医疗机构亟需针对不同工程成熟度的临床团队建立“支持型评估”通道。 - [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) CMIO 与临床信息学专家的角色正从传统系统界面配置员，转变为连接一线医护、AI 前沿技术与网络安全监管的核心“三方翻译官”。 - [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) 推行“第零车道”（Lane Zero）探索沙盒，依托大模型生成的高质量合成数据与去凭据化防护，让医护在触碰真实风险前自由验证构想。 - [14:37](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=877s) 利用 EHR 构建域（Build Domain）中的脱敏与回顾性数据，结合内置审计日志和技术团队结对协作，为原型提供渐进式安全演进路径。 - [16:00](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=960s) AI 已成为医疗行业的创意引擎，管理者的核心责任不是遏制创新萌发，而是主动搭建分级安全跑道，实现创新活力与患者安全的共赢。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=4s) 编程智能体的兴起与原型验证的幻象：从街机游戏到医疗格式转换器作为一名成瘾医学医师、急诊科医生以及加州一家医院的管理者，演讲者在步入医学殿堂之前曾是一名计算机背景的程序员。然而，随着在医疗与管理层级中越走越深，亲自动手写代码变得愈发困难。但在过去六个月里，随着编程智能体（Coding Agents）的爆发，整个原型开发的门槛被彻底重塑。为了展示这种变化，演讲者首先拿出了几个自己用智能体随手写出的小项目：第一个是为了满足童年情怀复刻的经典《Pong》（乒乓球对战游戏），尽管加入了现代化改进，但只要小球以特定角度击打球拍，系统就会报错——这是一个在开发时完全未被测试的边界用例（edge cases），在游戏里虽然无伤大雅，但在关键系统中却致命；第二个是一款为桌上角色扮演游戏（TTRPG）设计的虚拟掷骰子工具，智能体仅用了几分钟就精准搞定了原本需要人工耗费数小时调试的几何图形运算，但在实际跑团中玩家却依然嫌弃它，因为玩家追求的是把骰子握在手心里的真实触感与物理体验——它在工作流匹配上彻底失败了。随后演讲者将目光引向医疗领域的尝试。他为自己的医学生和住院医师制作过一个 HL7 消息转换器。对于不熟悉医疗信息化的人来说，HL7 是医疗互操作性的经典标准，它采用竖线（`|`）管道符进行字段分隔，充斥着密密麻麻的系统标识符，对人类极其不友好。这个小工具可以帮助学生理解医疗数据是如何分门别类归入不同数据桶中的。在本地主机（Localhost）上跑测试用例时，它运行得非常完美，然而一旦触及真实的患者数据，身份识别、访问控制和数据隐私等安全边界问题便接踵而至。医疗行业正进入一个“最小可行产品（MVP）”泛滥的时代：AI 编程工具让任何人都能迅速攒出一个令人信服的高保真功能原型，开发速度快得惊人，但那些真正艰深繁重的工作要么被表象掩盖，要么要在后期才会暴露出来。 - **Pong 街机小游戏**：暴露出快速生成的原型普遍缺乏对极端边界用例（Edge Cases）的充分测试。 - **跑团虚拟掷骰器**：证明了几何计算虽然完美，但因为忽视了真实用户的心理与物理习惯，无法融入实际操作工作流。 - **HL7 协议转换工具**：展示了本地调试与进入临床生产环境之间的鸿沟，真实医疗数据必须面对身份验证、访问控制和隐私等严格的安全边界。 --- ### [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) 旧金山黑客松与去污名化用语 Demo：功能可见，就绪不可见为了展现这一现实的冲击力，演讲者讲述了自己两周前在旧金山参加的一场医疗 AI 黑客松。这场活动定向邀请了约 150 人，绝大多数参与者是真正的“临床工作者构建者”（Clinician Builders）——包括护士、医生、物理治疗师、呼吸治疗师等各专业一线人员，同时配备了少量的工程与安全专家。最令人震惊的现象在半年前几乎是不可想象的：许多从未写过一行代码的医护人员，仅凭各自挑选的 AI 编程智能体，在短短 5 个小时内就提交了完全可交互的功能性应用。更重要的是，整个会场里每个人都在解决那些如果不是他们亲自动手、可能永远都走不到原型阶段的临床一线真实痛点。 ```mermaid flowchart TD accTitle: 医疗原型从开发到生产就绪的断层 subgraph Prototype["表象：5小时构建的原型（可见的功能）"] A["临床人员提出痛点"] --> B["AI 编程智能体快速生成代码"] B --> C["合成数据写入测试 FHIR 服务器"] C --> D["完美呈现去污名化高亮界面"] end subgraph RealityGap["鸿沟：生产环境缺失的关键支柱（不可见的就绪）"] E["无生产级 OAuth 2 / Token 鉴权"] F["密钥明文裸奔于.env 文件"] G["缺乏生产监控与生命周期责任人"] H["未跨学科临床验证（仅限成瘾科）"] I["未做安全审查（存在 XSS 等风险）"] end Prototype -.->|"看似完全可用，实则不可部署"| RealityGap ``` 作为成瘾医学专家，演讲者也现场展示了自己构建的原型：一个识别病历中“污名化用语”的纠错工具。利用编程智能体，他仅用 5 到 10 分钟就生成了高度逼真的合成患者数据，并无缝写入了 FHIR（快速医疗互操作资源）服务器。在演示病历中，系统精准抓取并高亮了诸如“known alcoholic”（知名酒鬼）、“known addict”（已知瘾君子）以及“he chose to drink again”（他自己选择再次饮酒）等将患者物化为其病症的冷漠表述，并提示替换为更规范的中性诊断词汇（如“酒精使用障碍患者”）。其他机构的同行和科研伙伴看后激动不已，纷纷要求立即在各自的诊所上线试用。但演讲者清醒地意识到，这款工具绝对不能以当前状态直接部署到自己的医院中：开发环境下的 FHIR 服务器根本没有生产级别的认证模型，既无 OAuth 2 也无 Token 访问网关；API 密钥直接裸奔在 `.env` 文件中；系统没有任何生产运维监控，也没有指定生命周期责任人；而且临床有效性验证仅停留在成瘾医学，完全未经产科或创伤科等其他场景的检验；代码更未经历过任何安全审计，可能潜藏着跨站脚本攻击（XSS）等漏洞。这正是当下医疗界最严峻的张力："功能是显而易见的，但就绪状态却隐形于表象之下"（"Function is visible but readiness is not."）。 - **旧金山黑客松规模**：约 150 名受邀者，涵盖各科室一线医护，在无编程背景下凭借编程智能体在 5 小时内实现应用落地。 - **合成数据生成突破**：以往生成逼真患者数据极其困难，如今借助智能体能在 5 至 10 分钟内生成合规且具有高度真实感的合成病历并写入 FHIR API。 - **阻碍上线的生产就绪缺陷**：缺乏企业级密钥管理、缺失 OAuth 2 鉴权门禁、无系统审计日志追踪、缺乏临床跨科室边界验证、未实施网络安全审查。 --- ### [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) 代码智能体的安全盲区与传统角色启发式的失效为了进一步说明潜在的技术风险，演讲者引用了一项学术界对主流编程智能体代码质量的研究分析。该研究呈现出两条极为割裂的曲线：蓝线代表代码的语法正确性与外观完整度，随着智能体模型的演进，该指标已无限逼近 100%，这也解释了为何如今的原型制作看起来越来越精美；然而代表网络安全基准测试的红线却表现糟糕，即便针对的是业界早已熟知的 OWASP 常见漏洞，智能体的抵御表现依然欠佳。更致命的是，大型语言模型在生成代码时对“上下文”（Context）的依赖越深，其在安全基准测试中的得分就越低——智能体处理纯语法层面的 SQL 注入尚可，但在处理依赖复杂系统上下文的跨站脚本（XSS）攻击时漏洞百出。很多医疗同行会天真地反问：“那为什么不直接提示编程智能体把代码写得安全一点？”演讲者指出，AI 智能体根本无法掌握由具体医疗机构和业务环境所定义的“组织上下文”。回到去污名化工具的案例中，智能体根本不知道谁该拥有何种级别的访问权限，现状是使用该工具的医生能轻易越权调取所有患者的记录；在调用前沿商业闭源大模型时，若无企业级签署协议，这本质上就是患者受保护健康信息（PHI）的外部数据泄露；此外，遵循 HIPAA 法规底线所需的完整审计追踪（Audit Trail）由谁负责记录、出现临床故障时的失效保护模式是什么、谁来承担软件开发生命周期（SDLC）的监控与退役责任，这些全局治理逻辑都是编程智能体在代码行之间无法凭空臆断的。这就宣告了过去医院管理启发式规则的破产：传统上管理层把“职务角色”作为技术准备度的捷径，认为只有 IT 或软件工程师递交的东西才算数；现在人人都成了构建者，华丽的代码界面再也不能作为系统安全可靠的代名词。 - **代码外观 vs 安全基准差距**：大模型生成代码的语法完成度与表层功能已接近 100%，但面对 OWASP 安全基准测试时失分严重。 - **上下文依赖悖论**：代码逻辑需要的上下文越广阔，模型在安全审计上的表现反而越差（SQL 注入防御尚可，跨站脚本防御极弱）。 - **智能体无法感知的组织上下文**： - 角色级数据最小化访问权限（RBAC / ABAC 规则）。 - 避免商业闭源模型导致未经授权的 PHI 数据外泄。 - 满足 HIPAA 强制要求的合规访问审计追踪日志。 - 明确系统全生命周期的维护、临床纠偏与退役负责人。 --- ### [07:11](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=431s) 两种制度困境：完全封堵引发的“影子 AI”与过时僵化的治理流程面对手握完整原型的临床构建者，医疗行业目前的反应模式暴露出严重的体制困境。与外部行业不同，医院 IT 对电子病历系统（EMR）拥有极强的行政与网络准入把关权。基于这种绝对的封锁能力，目前业内最常见的做法就是“一刀切”地全面否决——直接向医护人员宣告“这里不接受任何个人搭建的应用”。这种无通路状态导致了双重恶果：其一是极具价值的创新在萌芽中夭折。黑客松上的那些原型凝聚了一线医护数年乃至数十年的临床痛点洞察，一旦缺乏通往现实的制度路径，这些本可优化流程的构想便彻底流失；其二则是催生了极其危险的“影子 AI”（Shadow AI）。当正规通路被堵死，想要解决实际工作痛点的医护人员便会绕开 IT 部门：他们会直接复制病历贴进未经签署 BAA 协议的个人前沿模型账户；使用未经安全批准的第三方急救或随访小工具；甚至私自将原型托管在外部公共云服务器上。调研表明，医护人员使用影子 AI 绝非出于恶意，而是因为被批准的现有工具根本无法解决其工作流痛点，或者体制内根本没有递交新想法的出口。"缺乏正规通道，既是一场创新的重大损失，也是一个严重的安全隐患。" ```mermaid flowchart LR accTitle: 传统治理模式与影子 AI 的因果链路 subgraph InstitutionalResponse["传统医院的两种极端响应"] R1["完全封堵路线 （基于角色的武断拒绝）"] R2["僵化的瀑布式审查 （长达8个月的漫长审批）"] end subgraph BadOutcomes["负面衍生后果"] O1["创新构想夭折 （一线临床价值永久流失）"] O2["滋生影子 AI （绕过内网使用未授权个人模型）"] O3["数据外泄风险 （PHI 泄漏至未经合规审查的云端）"] end R1 --> O1 R1 --> O2 O2 --> O3 R2 --> O2 ``` 另一种极端响应则是拒绝改变传统的 IT 治理模型，仍用老一套僵化流程去套飞速迭代的 AI 应用。演讲者回忆起大约五年前在另一家机构为联邦科研项目开发软件的经历：每周与 IT 部门开协调会、与供应商多轮沟通、将 FHIR 标准拆解为患者、病情、文书等极细颗粒度的资源并逐一撰写权限论证，再加上无数边界用例测试，以及绘制了大量架构图和数据流向图——"密集的流程甚至让我患上了 Microsoft Visio 创伤后应激障碍，再也不想看到这款软件"。那套正规流程历时整整八个月才走到安全审查阶段。在手动逐行敲代码的慢节奏时代，这种瀑布式往复反馈是合理的，但在智能体以指数级速度产出原型的今天，八个月的审批机制完全与技术现实脱节。为了说明这一点，演讲者列举了近期审核的两个极客团队：团队 A 拥有完备的技术支撑，不仅做好了控制映射，甚至用上了自动化合规工具，提前答复了所有安全预审问题；而团队 B 是一个由心脏科医生带队的小组，他们敏锐地发现某种心律失常算法可以挖掘出前所未见的诊断洞见，并快速搭出了一个惊艳的原型，但他们完全不知道如何达到生产级合规。如果采用旧模式，缺乏企业工程经验的团队 B 必死无疑，但他们的临床价值丝毫不逊于团队 A，他们真正需要的是一种全新的“支持型评估”（Supported Evaluation）通道。 - **影子 AI 的催生动力**：并非医护恶意违规，而是现存获批系统无法满足细分工作流，且组织内部缺乏顺畅的提报通道。 - **传统治理周期**：以传统手工编程时代为基准设计，依赖周例会、细粒度资源论证与庞杂的 Visio 架构图，单单安全预审便长达 8 个月。 - **审查团队鲜明对比**： - **团队 A**：具备商业化与自主合规工具支持，准备极其完备，无需组织花费额外精力辅导。 - **团队 B**：来自心脏科等前沿临床一线，手握革命性算法雏形却缺乏企业工程规范，若无扶持性路径就会被旧机制扼杀。 --- ### [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) 治理重构与“第零车道”：支持型评估与合成数据沙盒面对日益庞大的临床构建者群体，医疗组织的治理架构必须进行底层重构。以首席医疗信息官（CMIO）和临床信息学专家为代表的角色，过去长期被视为临床医护与 EMR 系统之间的传声筒，主要精力花在为医生配置系统快捷方式或优化操作模板上。然而近一两年来，全美各地的 CIO、CISO、法务及合规人员都在被迫打破专业壁垒、身兼数职。在当前的转型期，CMIO 的职责已转变为充当临床人员、AI 前沿技术与医院网络安全防线之间的“三方翻译官”。医院的治理绝不能仅仅等同于印在纸上的静态政策或按月开会的委员会；调查显示，目前全美仅有 18% 至 25% 的医疗机构建立了成熟的 AI 治理模型，大部分机构依然将合规当作机械式的勾选任务（Check-the-box），未能建立起一种能随技术演进而自适应进化的组织文化。要建立健康有效的构建者治理体系，核心在于让临床医护从心底理解技术安全底线，使良好治理的目标变成“让遵循安全规范的前进路径，比绕道使用影子 AI 更加轻便省力”。为此，演讲者提出必须建立差异化的分级通道。其中最关键、却也是目前行业讨论最少的便是“第零车道”（Lane Zero）——一个纯粹供临床医护探索、学习与原型验证的零门槛沙盒。以往医院无法开放此类环境，主要症结在于高质量合成数据的获取成本过高；而最近三到六个月里，编程智能体已经能够根据医学准则快速生成极为逼真的模拟病历，消除了侵犯患者隐私的风险。主流 EMR 厂商也普遍提供了具备合成数据的 API 测试沙盒。通过“第零车道”，医护人员无需接触任何生产级别的真实凭据或真实数据，即可在官方批准的工具和带有护栏的自助体系下自由构建概念验证。这种早期介入模式，让医院不仅保护了自身数据资产的安全底线，更保护了临床一线那份弥足珍贵的创新热情。 - **CMIO 职能的范式转移**：从过去负责“优化 EMR 菜单与配置收藏夹”的系统操作员，演变为连接临床工作者、AI 底层技术与数据安全监管的核心翻译官。 - **医疗 AI 治理现状数据**：权威调研显示仅有 18%–25% 的医疗系统拥有公认成熟的治理模型，其余大多数仍停留在纸面政策与官僚式合规清单阶段。 - **“第零车道”（Lane Zero）的核心支柱**： - **高质量合成数据**：依托现代大语言模型与智能体，依照严谨临床规则实时生成逼真的非真实病历。 - **隔离的官方沙盒**：充分利用主流 EMR 厂商现成的开放 API 沙盒与合成数据库进行原型演练。 - **去凭据化防护**：严格杜绝向早期探索者发放真实生产环境凭证，在安全红线内实现完全的自主开发与早期风险评估介入。 ### [14:37](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=877s) 搭建从概念到落地的桥梁：沙盒环境与技术支持在讨论医疗机构的技术评审路径时，通常存在一种面向成熟团队的“快速通道”（Lane 2）。走这条通道的团队往往万事俱备，随时准备部署和运行：他们可能已经在其他医疗机构成功落地过该系统，所有用于安全审查的材料一应俱全，并且清晰阐明了临床安全（clinical safety）、网络安全与隐私保护、持续监控机制，甚至连指定的系统负责人（named owner）都已明确。然而，真正的核心难题在于：一个初期的构想究竟如何从最基础的原型跨越中间层层阻碍，最终达到这种成熟的第二阶段状态？关键在于在评估阶段为开发者提供切实的支撑。其中一个非常有效但行业内普遍未充分利用的手段，就是提供受限或回顾性的真实数据。许多医疗信息系统（EHR）供应商都提供构建级别（build level）的测试环境，里面包含脱敏或非真实患者的数据。这类环境让临床开发者能够在未来实际部署的系统架构内安全地进行演练。这种方式的巨大优势在于，安全和技术团队可以严格控制开发者能够访问的数据范围；同时，审计和监控工具本身就内置在这一构建域（build domain）中，团队可以尽早启动日志记录与审计追踪，并进行针对性的聚焦审查。更重要的是，必须为临床开发者配备技术合作伙伴进行结对协作。在圣华金总医院（San Joaquin General Hospital），承担这一角色的正是首席医疗信息官（CMIO）所领导的部门。在技术团队中，完全不缺乏渴望与一线临床医护合作、将好想法推向实际应用的人才；临床端同样有强烈的合作意愿。医疗机构有责任积极鼓励这种协同机制，划分不同的演进阶段与路径，唯有当系统真正“赢得资格”或确有必要时，才触发全面严苛的安全审查，避免在早期就将创新扼杀。 ```mermaid flowchart TD accTitle: 医疗AI从构建到落地的演进路径 subgraph S1["阶段一：安全评估与沙盒验证"] A["临床构建者构想"] --> B["构建域沙盒 (Build Domain)"] B --> C["脱敏/回顾性数据测试"] B --> D["范围控制与内置审计日志"] E["技术团队/CMIO部门结对支持"] -.-> B end subgraph S2["阶段二：成熟部署通道 (Lane 2)"] F["临床安全与隐私审查"] G["明确系统负责人与监控机制"] H["正式上线与生产运行"] F --> G --> H end C --> F D --> F ``` - 供应商提供的构建环境（Build Level）能够隔离真实患者风险，允许开发者在贴近生产的环境下打磨产品。 - 内置审计工具不仅能保障合规，还让技术治理与日志追踪从最初的开发阶段自然介入。 - “全量严格审查”（Full review）应当是一种按需触发或在项目证明自身可行性之后才进行的机制，而非从第一天起就设立的硬性路障。 --- ### [16:00](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=960s) 为创意引擎铺平跑道：医疗AI的安全共赢回到之前提到的黑客松经历，演讲者感慨道，参与者仅仅在短短五个小时内，就各自构建出了具备临床应用形态的软件。每一个工具都能实际运行，虽然大部分还只是最小可行产品（MVP），但每一款产品都凝聚了创造者作为一线医护的独特经验与切身洞察，具备实打实的临床价值。这生动证明了人工智能已经为整个医疗行业装上了一台强大的“创意引擎”（an engine for ideas）。既然人人皆可构建的时代已经到来，医疗机构的管理者和技术决策者所肩负的责任，就不应是去阻碍这些创新的萌发，而是要主动搭建起坚实、清晰且安全的“跑道”（lanes），让这些具有临床价值的探索能够在兼顾安全的前提下稳步向前推进。 - 5小时的黑客松充分展示了AI编码工具如何将临床医护的个人工作经验快速转化为可运行的MVP应用。 - 组织治理的核心职责是建立分级分类的安全通道，让创新活力与网络安全、患者安全达成共赢。 ## Notable quotes > [01:53](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=113s) 功能是显而易见的，但就绪状态却隐形于表象之下。 — Joseph Izzo > [05:08](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=308s) 那为什么不直接提示编程智能体把代码写得安全一点？ — Joseph Izzo > [07:11](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=431s) 这里不接受任何个人搭建的应用。 — Joseph Izzo > [07:11](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=431s) 缺乏正规通道，既是一场创新的重大损失，也是一个严重的安全隐患。 — Joseph Izzo > [07:11](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=431s) 密集的流程甚至让我患上了 Microsoft Visio 创伤后应激障碍，再也不想看到这款软件。 — Joseph Izzo > [11:41](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=701s) 让遵循安全规范的前进路径，比绕道使用影子 AI 更加轻便省力。 — Joseph Izzo > [14:37](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=877s) 全量严格审查应当是一种按需触发或在项目证明自身可行性之后才进行的机制，而非从第一天起就设立的硬性路障。 — Joseph Izzo > [16:00](https://www.youtube.com/watch?v=Ib-RTE1rGXg&t=960s) 既然人人皆可构建的时代已经到来，医疗机构的管理者和技术决策者所肩负的责任，就不应是去阻碍这些创新的萌发，而是要主动搭建起坚实、清晰且安全的‘跑道’，让这些具有临床价值的探索能够在兼顾安全的前提下稳步向前推进。 — Joseph Izzo ## Mentioned Joseph Izzo, 圣华金总医院 (San Joaquin General Hospital), Ai4, Pong, 桌上角色扮演游戏 (TTRPG), HL7, FHIR (快速医疗互操作资源), OAuth 2, OWASP, HIPAA, Microsoft Visio, 电子病历系统 (EMR / EHR), 商业合作协议 (BAA), 影子 AI (Shadow AI), 跨站脚本攻击 (XSS) ## Video description Joseph Izzo - Chief Medical Information Officer, San Joaquin General Hospital Artificial intelligence has dramatically lowered the barriers to building useful software, empowering clinicians and other domain experts to solve problems that once required specialized technical expertise and dedicated development teams. Drawing on my experience building and governing healthcare software and AI tools before and after the rise of coding agents, I will explore how this transformation changed the way I think about innovation, cybersecurity, and organizational risk. The talk will explore how the AI era challenges the traditional assumption that innovation and cybersecurity must compete with one another. Rather than preventing clinicians and other domain experts from building, healthcare organizations have an opportunity to create pathways where new ideas are encouraged, safely evaluated, and responsibly integrated into patient care. Through practical examples, attendees will learn practical principles for balancing innovation with governance, cybersecurity, and patient safety without creating unnecessary barriers to progress. The goal is not to slow the next generation of builders, but to create healthcare organizations where innovation and security can succeed together. [Source](https://www.youtube.com/watch?v=Ib-RTE1rGXg)

Source quotes

[1:53] Joseph Izzo: 功能是显而易见的，但就绪状态却隐形于表象之下。 [5:08] Joseph Izzo: 那为什么不直接提示编程智能体把代码写得安全一点？ [7:11] Joseph Izzo: 这里不接受任何个人搭建的应用。 [7:11] Joseph Izzo: 缺乏正规通道，既是一场创新的重大损失，也是一个严重的安全隐患。 [7:11] Joseph Izzo: 密集的流程甚至让我患上了 Microsoft Visio 创伤后应激障碍，再也不想看到这款软件。 [11:41] Joseph Izzo: 让遵循安全规范的前进路径，比绕道使用影子 AI 更加轻便省力。 [14:37] Joseph Izzo: 全量严格审查应当是一种按需触发或在项目证明自身可行性之后才进行的机制，而非从第一天起就设立的硬性路障。 [16:00] Joseph Izzo: 既然人人皆可构建的时代已经到来，医疗机构的管理者和技术决策者所肩负的责任，就不应是去阻碍这些创新的萌发，而是要主动搭建起坚实、清晰且安全的‘跑道’，让这些具有临床价值的探索能够在兼顾安全的前提下稳步向前推进。

Source quotes

[1:53] Joseph Izzo: function is visible but readiness is not [1:53] Joseph Izzo: Let's roll this out. Let's put this into production. Let's get it in my clinic. [1:53] Joseph Izzo: known alcoholic [1:53] Joseph Izzo: patient with alcohol use disorder [1:53] Joseph Izzo: known addict [1:53] Joseph Izzo: he chose to drink again [5:08] Joseph Izzo: make the application secure [11:41] Joseph Izzo: Lane Zero

Original transcript

[0:04] I'm an addiction physician, an emergency doctor and an administrator for a hospital in California, [0:08] but actually was a programmer and did computer science before medicine. And I want to talk about [0:12] how the landscape has changed for us, particularly in the last six months. Um, when coding agents [0:17] has made prototyping very possible. So, um, to start off, I will actually give you some tools [0:22] that I've made in the with the caveat that as you go deeper into administration into medicine, it [0:26] becomes harder to program uh, which these coding agents have changed. So, this is Pong. It fulfills [0:31] childhood nostalgia. Had a couple of upgrades to it. And you'll notice that when the ball comes and [0:35] hits the paddle at a certain angle. Um, it doesn't work as well and it's a something that nobody [0:40] tested. So, what we're saying is that it doesn't work on edge cases, but it's pawn, so that's fine. [0:45] And then, um, this is a tabletop role playing game, dice roller for when you forget your dice at [0:49] home. It works well. The geometry is perfect. Uh, it would have taken me hours to do this by myself, [0:54] but you'll find that the practitioners of tabletop role playing games actually like having the dice [0:59] in their hand and feeling that that texture and that experience. So, it doesn't work on [1:03] workflows. And now, let's move into healthcare. That's why you're here. Um, this is something I [1:07] made for my students and my residents. It's an HL7 converter. HL7 is our interoperability standard [1:12] for healthcare if you're not familiar. And it's very not uh it's not human readable. Uh, it's pipe [1:17] delimited. It has a lot of um identifiers and it's for my students this helps them to understand how [1:23] different healthcare data falls into different buckets. It's fine for local host. It's fine [1:27] if I'm using it on examples. But when we start to talk about real patient data then you worry about [1:32] identity access privacy. So security boundaries and what I'm noticing especially in healthcare [1:38] is that we're in an era of minimum viable products where these coding tools have made [1:41] it possible to make functional demos that work well and are convincing. and the build is fast [1:47] whereas the harder parts are even either hidden or manifest later on. So to really drive this home, [1:53] I'm going to tell you about a recent healthcare AI hackathon I was in in San Francisco. It was two [1:57] weeks ago and um it was kind of perfect timing for this talk in that it was an invitation for [2:03] about 150 folks and the majority were actually clinician builders. Um there were engineers and [2:08] uh security staff there. There were folks who were willing to partner. But what surprised me [2:12] about this and probably wouldn't have been true 6 months ago is that many clinicians who built their [2:17] own app and submitted it within 5 hours had never had any programming experience. They simply used [2:22] a coding agent of their choice and had a product that looked functional within 5 hours. And what I [2:27] noticed when looking around the room is everyone was solving a problem that otherwise may not have [2:31] actually made it to the prototype stage at all. It's folks from various varied uh backgrounds in [2:36] medicine. You had nurses, doctors, uh physical therapists, you had respiratory therapists and [2:40] so on. Each bringing their own unique experience. So I'm going to show you uh what I made to kind [2:44] of frame this talk. I'm an addiction doctor. As I mentioned, this is not a real patient. In fact, [2:49] the coding agent made it possible to generate synthetic data and put it into a fire server [2:54] um within 5 to 10 minutes and convincing synthetic data. [snorts] And fire server is just an [2:59] extension of HL7 if you're not familiar. It moves that um non-readable text I showed you before into [3:04] an API standard with JSON. And the other thing I want to mention is that I'm not a vendor and I'm [3:08] not trying to sell this product. This is actually something that was for a research collaboration I [3:12] have with some folks in the addiction medicine space. Um and [snorts] I showed it so I'll show [3:17] you the demo. Basically, it has a patient with a note kind of peppered with stigmatizing language. [3:22] Language that describes a patient as if they are their disease and not a person. So here in this [3:27] note, you'll see the agent pulling together all the details and highlighting a couple of phrases [3:31] like known alcoholic. We don't say that. We say patient with alcohol use disorder. Same thing for [3:36] known addict. He chose to drink again. These are not phrases that we want to put into our chart. [3:41] And was interesting about this is like I showed it to several folks uh colleagues across other [3:45] organizations and the immediate response including my research collaborators is let's roll this out. [3:49] Let's put this into production. Let's get it in my clinic. Our students need this. uh some of [3:53] our our more um [snorts] uh our doctors who aren't familiar with this will probably benefit from it. [3:58] But in thinking about it, you know, my reality is I um wouldn't deploy it in my own hospital [4:03] at least as it is. And the common reasons that I think a lot of you in the room are familiar with, [4:07] you know, I showed you a development fire server, right? It doesn't have a production authentication [4:11] model. Um you really should be using something like ooth 2 and I have no tokens, so no access uh [4:16] uh gating. So then there's no enterprise secrets managements. They were inn files. [4:20] um good for a hackathon, not good for this. There's no production monitoring. Um there's [4:25] no life cycle owner. If if we were to put into production, it would have to be me. And like I [4:28] mentioned, that's not something I can do. And then no clinical validation. Uh so we didn't really [4:32] check this in any other domain except addiction medicine. What about stigma in OB? What about [4:38] stigma in trauma? That's completely different and hasn't been validated. And we didn't do a security [4:42] review. Perhaps there's like a cross-sight scripting vulnerability or something like that. [4:45] And what's interesting is this pull in healthcare that we feel like we're needing to deploy tools [4:51] that may not be ready uh but look like they are and competing against what that pathway is to get [4:57] to readiness. So again, you know, function is visible but readiness is not. [snorts] And the [5:03] question that I'm I'm seeing across the board is what happens after it works. So just taking one [5:08] step back to look at the very real risk that is there. Um this is a slide from one analysis of the [5:13] landscape. A couple of other academic publications and similar analysis have shown that coding agents [5:19] um have gotten better and better at producing code that looks good and has good syntax as I'm sure a [5:23] lot of you are familiar with. But that's the blue line approaching almost 100%. It also explains why [5:29] um these tools look more and more functional in less and less time. The red line on the [5:33] other hand is the response to certain security benchmarks. And what was alarming to me is that [5:38] these benchmarks were actually on very well publicized OAS vulnerabilities. And moreover, [5:43] the context, the more context the large language model and by extension the coding agent needed, [5:48] the worse it did on the security benchmark. So it does better on SQL injection than it does on [5:52] cross-ite scripting. And I get the answer when I talk to others um in healthcare like why not [5:57] ask the coding agent to make it secure? Well, context is a big important part of that. But [6:01] more importantly, context also defines is defined by you and your organization. So if we go back to [6:05] that stigmatizing language um application that coding agent can't tell me the right identity [6:10] or level of access I have to set that. So if right now a physician using that tool has access to all [6:15] the patients. Same idea for um what patient data may be used or leave the organization. [6:19] In this example I use the frontier model. So that's exfiltration of our patient information [6:24] without an agreement. Um and then determining on what has to be logged or audited. Well, one [6:28] of the fundamentals of HIPPA are the privacy law that governs us is a need to establish an audit [6:32] trail [snorts] and then of course a set acceptable clinical failure modes. We kind of talked about [6:37] that as well as a validation and then assign who monitors, maintains and retires the application [6:41] the software development life cycle. [snorts] And so what I've come to realize especially [6:46] in healthcare is that when everyone can build our heristics don't work anymore. We used to say your [6:50] role defined how ready you were or at least use that as a shortcut. And then polish and how well [6:54] the application functions isn't a surrogate for what's safe. And so I looking out there's actually [7:01] surprisingly little guidance on what to do with this particular scenario with clinician builders [7:05] who are coming to you with products that look complete and prototyped. I've seen two responses [7:11] and they're kind of shaped by what um what control we have in healthcare. So unlike some other [7:16] industries, we actually have very significant gatekeeping over the medical record system. And [7:20] so we are able to actually be that gatekeep or we're actually able to block folks if we don't [7:25] want them to access the production level. So the responses I see across you know nationally when [7:29] talking to colleagues is more um on things like what happens if we provide no pathway. So we just [7:34] tell clinician builders kind of anchoring to the role based heristics we talk about well you just [7:39] can't do it. You're we don't accept that here. We're not going to use a clinician built app. [7:43] And that comes with two problems uh that I I've seen. The first is that the idea will die. Um when [7:47] I was talking about the h the hackathon, what I mentioned is that the folks who were participating [7:52] were coming from a very niche background with lived experience building a prototype that that [7:56] probably was in their head for years, decades, and being able to make it to a point where they [8:01] can demonstrate something that they feel improves their workflow. Without a work a clear pathway, [8:05] then the project's going to stall and then the opportunity is lost. And then the other risk that [8:10] I'm worried about is that the idea routes around us. So say I am a builder who makes an application [8:16] like the one I had the the stigmatizing language. I can do close to what I achieved with that by [8:21] using a frontier model and a personal account which has the risk of expiltration. Right? That's [8:25] the same problem. I'm sending protected health information potentially to an outside third party. [8:29] Same idea with unapproved tools and endpoints. There are other apps that do this. There's one [8:32] for emergency medical services. It has the same kind of um effect where it looks at your note and [8:37] make sure that it doesn't have harmful language, but that's again a third party. The same thing [8:41] for unsanctioned deployment. If I really wanted to use this in my clinic, I could spin it up in the [8:46] cloud, but that has again the very same risk. And the question is like why is this happening? When [8:50] you survey physicians and other healthcare folks across the country about this idea that they're [8:54] introducing shadow AI, it's not to be malicious. The answers are either they feel that the own [8:59] tools that have already been approved don't solve their workflow problems or they don't have a path [9:04] forward. In other words, the absence of a pathway is both an innovation loss and a security risk. [9:12] The second thing I've seen and I think it's rooted in the traditional way we approach [9:16] these in healthcare is to not change the way we do governance at all. Um, so for this example, [9:21] about 5 years ago at another organization, I was responsible for programming a tool for a federal [9:26] grant. And we went through the process that I think you'd all expect is weekly IT coordination, [9:32] conversations with the vendor, making sure that we had everything scoped out, the dependencies, had [9:37] access to only the right tools or right resources in the fire standard that we needed. If you're not [9:42] familiar, the fire standard is broken down into things like patient condition, documentation, [9:46] and justifying each step of the way why I needed that. We had plenty of testing evidence, [9:51] edge cases like we were talking about and then of course um architecture and data flows so much that [9:56] like I can't look at Microsoft Vizio again. And the thing is that this was all realistic and this [10:00] was all reasonable and all correct and it took eight months to get to the security review. But [10:05] what's interesting about this is what was baked into this process is the idea that it took so [10:10] long to code. Um it was an iterative process that required feedback and then correction, [10:15] feedback and correction and that doesn't match reality. things are now exponentially faster as [10:19] we're all seeing. And so it's very difficult to apply this format unchanged across what [10:24] we're seeing in the landscape. And so let me give you two examples, very obscure, but from recent [10:29] security reviews we've had that kind of shape where I want to go with this and what I think [10:32] is will work for healthcare. So the first company comes to us. They're both clinician builders. They [10:37] have small teams and they've already um kind of scoped everything out. They have their controls [10:41] mapped. They use an autonomous compliance tool. And to be perfectly honest, you don't have to. But [10:45] the fact that they thought about it is something pretty relatively rare in the clinician builder [10:48] space. And then I had answers to almost all the questions applied in advance. Um we had a large [10:53] document with pretty much everything I would have asked in a review. I would suspect that they had [10:57] rolled this out of other organizations. They had gone through the process and these are the folks [11:00] who we probably don't have to change anything of what we do. It's these folks in team B that [11:04] I'm more worried about. This is someone who um is ready for what I think is supported evaluation. [11:08] It's someone who comes with that unique clinical need. So to for purposes of this demonstration, [11:13] this discussion, say there a cardiologist who recognizes that certain arhythmias if plugged [11:17] into a model or an algorithm produces insights that have never before seen. They have a working [11:22] prototype that is convincing and looks like it will benefit their patients, but they don't quite [11:26] know how to get to production. In the old model, these folks would have been turned away. Either [11:30] the first methodology I told you or the second one. But both of these are are valuable. Both of [11:35] these are worth advancing. It's just that they needed different steps. So in order to get to [11:41] that point, the organization has to adapt. Um I'm a CMIO, so my background is clinical informatics. [11:45] Traditionally, uh we were seen as the folks who translate between the clinicians and the EMR. We [11:50] did so much more, but what our point of contact with clinicians was if they're having trouble [11:54] or they didn't have ways of optimizing their workflows, we gave them shortcuts, favorites, [11:57] things like that. In the same way that I've been talking to tons of CIO, CISOs, folks in regulatory [12:02] and compliance and legal, everyone is feeling this this being compelled to wear different hats and [12:07] grow particularly in the last one to two years. So what at least my specialty does now is a lot [12:12] more translation between the clinicians and the other medical staff and the actual tech and the [12:17] AI itself. And so by extension for this talk, we also become those folks who translate between the [12:21] clinician builders and the folks from the various uh disciplines that have a stake in making sure [12:26] that the products we roll out are safe. And that's actually why governance is more than policy in a [12:30] committee. And if you look at some of the recent um surveys across maybe the last six months, [12:35] they tend to estimate that maybe only 18 to 25% of healthcare organizations actually have what's [12:40] defined as a mature governance model. And part of that is because a lot of it has been implemented [12:45] as almost like checklists or checking a box and not as organizational culture and an operating [12:50] model that allows us to adapt. Um part of that culture means that the clinicians who are part [12:54] of our structure actually know the last 14 slides and understand why this all these concerns are [13:00] in place and then permeated throughout their department. So we have a culture change going [13:03] forward. And then if you have that then you can go into what's the specific governance for clinician [13:07] builders and it is a shared understanding of readiness early engagement normal not waiting [13:12] until the product is complete but actually work engaging earlier at the point of the prototype or [13:16] when the clinician expresses the idea and then matching the pathway to the work and the time [13:20] it's going to take. And so back to that idea of like why shadow exi exists, why folks route the [13:26] organization, it's typically because you want good governance to make progress easier than working [13:32] around us. [snorts] And with that now we can talk about how there's ways we can approach different [13:37] projects in different stage of um of development. I don't want to say that there's only three lanes. [13:41] I think there's a lot of in between. But the one I don't see talked about a lot in healthcare is the [13:45] lane zero. Like the one where we're just letting clinicians and clinical staff explore and learn. [13:50] And I think part of that is because synthetic data has actually been challenging to make. Um [13:54] I would say only in the last 3 to six months do you get reliable synthetic data whether it's with [13:58] coding agents um with some of the other tools that are out there. And for the stigmatizing language [14:03] example that was based on just a rubric that is part of my other research that the model was able [14:07] to go with and run with and make very convincing um notes that actually mirror some of the examples [14:12] I've seen in the medical record and with some of my patients. Uh approved tools. Obviously that [14:16] makes sense. And then self-service within guard rails. Um, we certainly want to make sure we know [14:20] what you're doing, that you engage the official channels, but um, we don't want production [14:23] credentials, and that's okay because most medical records companies actually provide a sandbox, [14:28] um, for their APIs that has another layer of synthetic data. It's how I learned. It's how [14:33] folks can get really comfortable without actually even touching our data at all. I'm going to skip [14:37] ahead to lane two where that's those folks who come ready. They're ready to deploy and [14:40] operate. They have probably rolled out in another organization. They have everything, you know, [14:44] ready to for the the security review. They keep have articulated clinician safety or clinical [14:49] safety, security and privacy, monitoring and even have a named owner. But the question that becomes [14:54] and I mentioned there's many steps in between. How do you get from zero to one or sorry to two [14:58] I actually gave you that. It's having support when you're doing your evaluation. So one thing [15:03] we can do and again I don't see this a lot across the board is limited or retrospective real data. [15:09] Um, a [snorts] lot of the vendors provide a build level of their medical record system [15:13] that has obscure data or data that is not based on real that is not um, ostensibly real patient [15:18] data. And it allows you to practice within the environment you're going to be building in. What's [15:23] advantageous about that is we also can control the scope of what you're able to access. We can start [15:27] the logging process there because the product the tools that do the audit are actually built into [15:32] the build domain as well. And then we can do that focus review and also have a technical partner [15:36] to pair with this uh clinician builder. And the question might be who is that? Well, at least for [15:40] us, it's my department. It's me. Um there's no der of folks on the technical side who are wanting to [15:45] work with clinicians and vice versa to get these ideas into to real practice. And it owns on us to [15:51] really encourage this for all the reasons I've said. Um essentially there's different stages, [15:56] different pathways, but a full review only really when it's earned or necessary. So um [16:00] kind of wrapping it up in [snorts] five hours back to that hackathon we each built something that [16:04] looked like a clinical application. Um each one of them worked most were minimum viable products [16:08] but had something valuable that was a unique experience of the person creating the tool. So [16:13] if we can see that AI has given healthcare an engine for ideas then it's on us to build the [16:17] lanes to move forward safely. Thank you so much. I really appreciate you. [snorts] Yeah. Yeah.