Johns Hopkins Taps AI Retinal Scans to Predict Heart, Dementia Risks

Windows to the Soul: How AI in Ophthalmology Unlocks Secrets of Systemic Health — Ai4

Ai4

Deep learning extracts invisible vascular and neural biomarkers from routine retinal photos to forecast heart attacks and organ failure before symptoms appear, enabling retail-based triage if payers resolve reimbursement hurdles.

AI-assisted summary of Windows to the Soul: How AI in Ophthalmology Unlocks Secrets of Systemic Health — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 28, 2025, this presentation features Dr. Alvin Liu, Endowed Professor and Inaugural Director of the Gills Artificial Intelligence Innovation Center at Johns Hopkins Medicine. According to the page description, Dr. Liu explains how the retina provides a non-invasive window into living blood vessels and brain tissue, allowing AI models to predict age, sex, blood pressure, kidney function, cardiovascular risk, and dementia.

## TL;DR Dr. Alvin Liu introduces oculomics, an emerging discipline where deep learning models analyze high-resolution retinal photography to extract systemic health biomarkers invisible to human clinicians. By leveraging the retina's unique biological position as accessible neural and vascular tissue, AI algorithms achieve superhuman capabilities in forecasting cardiovascular events, renal damage, and neurodegenerative disorders like Alzheimer's. To maximize impact, Dr. Liu proposes deploying non-mydriatic retinal screening into community retail spaces and pharmacies to streamline patient triage, while emphasizing that realizing this future requires resolving clinical workflow integration, reimbursement models, and stakeholder alignment.

## Atmosphere & style The presentation is intellectually captivating, authoritative, and visionary, balancing advanced medical informatics with relatable clinical analogies like digital camera sensors and uncomfortable pupil dilations. Dr. Liu speaks with an engaging, lucid delivery that makes complex neuroembryology and algorithmic breakthroughs accessible to a broad audience. The emotional arc progresses from fascination with AI's superhuman predictive powers to a pragmatic, grounded discussion of the real-world healthcare infrastructure required to bring equitable screening to life.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) The retina is the only site in the human body allowing direct, non-invasive observation of living blood vessels and central nervous system tissue simultaneously at micron-level resolution. - [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) Because the retina develops as an embryonic outpouching of the diencephalon, it functions biologically as an accessible extension of the living brain. - [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) Commercial retinal imaging platforms achieve 1-micron resolution, providing up to 1,000 times higher spatial resolution of microvascular and neural structures than conventional 3-millimeter clinical brain MRI. - [02:18](https://www.youtube.com/watch?v=RguDF7OwHDI&t=138s) Oculomics leverages deep learning on standard color fundus photography to accurately predict demographic indicators, lifestyle factors, and systemic disease risks without prior patient history. - [02:18](https://www.youtube.com/watch?v=RguDF7OwHDI&t=138s) Retinal AI models possess superhuman predictive power, detecting latent textural and microvascular signatures invisible to expert clinicians to forecast long-term cardiovascular risks like myocardial infarction. - [02:18](https://www.youtube.com/watch?v=RguDF7OwHDI&t=138s) The primary clinical focus of oculomics centers on three major systemic disease domains: cardiovascular disease, neurodegenerative disorders (such as Alzheimer's and Parkinson's), and chronic kidney disease. - [06:00](https://www.youtube.com/watch?v=RguDF7OwHDI&t=360s) AI pioneer Geoffrey Hinton specifically highlighted deep learning applications in retinal imaging as a standout example of artificial intelligence delivering clear, transformative societal benefit. - [06:00](https://www.youtube.com/watch?v=RguDF7OwHDI&t=360s) AI-powered retinal screening identifies pathological biomarkers during asymptomatic windows, enabling timely interventions before severe, irreversible organ damage occurs. - [08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) Non-mydriatic retinal cameras enable rapid, drop-free imaging in community retail settings like pharmacies and supermarkets, dramatically lowering barriers to regular health screening. - [08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) Decentralized AI triage can bypass traditional primary care wait times and referral friction by directly routing flagged patients to targeted diagnostics and medical specialists. - [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) Successful clinical translation requires establishing standardized, condition-specific referral protocols tailored to existing local health delivery infrastructures. - [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) Widespread adoption depends on establishing sustainable reimbursement frameworks, validating return on investment through avoided hospitalizations, and aligning incentives across patients, providers, and payers.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) The Retina: A Direct Window into the Brain and Body

Dr. Alvin Liu opens his talk by introducing his background as an ophthalmologist and vitreoretinal surgeon—someone dedicated to treating pathologies in the back of the eye. He describes the retina as a thin, delicate tissue lining the rear of the eye, drawing an analogy to the sensor or computer chip inside a digital camera: if the chip fails or malfunctions, the camera cannot produce an image, and the patient cannot see. Retinal diseases are widespread and devastating. Diabetic retinopathy, caused by systemic diabetes damaging the delicate retinal microvasculature, stands as the leading cause of irreversible blindness among working-age individuals worldwide. Similarly, age-related macular degeneration (AMD) represents the predominant cause of severe central vision loss for adults aged 50 and older across the globe.

To engage the audience, Dr. Liu asks for a show of hands on who has visited an eye doctor over the last two years and endured pupil dilation. Dilation is an uncomfortable yet routine clinical ritual allowing specialists to inspect the fundus directly. He then poses a startling question: what if a clinician required no medical history, no demographic charts, and zero preliminary conversation, but could simply ingest a standard retinal photo into an artificial intelligence model and predict with high precision the patient’s age, biological sex, blood pressure, and smoking status?

Far from speculative fiction, Dr. Liu stresses that this is reality today, welcoming the audience to the emerging discipline of oculomics. Oculomics investigates the intricate connections between retinal biomarkers—visible features and microstructures in the eye—and systemic health conditions across the rest of the body. While medical science has long acknowledged these systemic links, the raw observational signal available to the human eye was never robust enough to yield reliable diagnostic value until the recent revolution in deep learning.

The retina holds a completely unique anatomical position: it is the only place in the entire human body where clinicians can directly and non-invasively observe both living blood vessels and central nervous system tissue simultaneously at micron-level resolution. During embryonic development, the optic cup and retina develop as direct neuroectodermal outpouchings of the diencephalon. As Dr. Liu notes, the retina is literally an accessible piece of the living brain.

- A standard millimeter contains 1,000 microns; modern off-the-shelf retinal imaging platforms achieve optical resolution down to a single micron. - In contrast, standard clinical brain magnetic resonance imaging (MRI) provides an isotropic resolution of roughly 3 millimeters. - Commercial retinal cameras provide up to 1,000 times higher spatial resolution of microvascular and neural structures than conventional cross-sectional brain imaging.

---

[02:18](https://www.youtube.com/watch?v=RguDF7OwHDI&t=138s) The Rise of Oculomics and Superhuman AI Capabilities

By pairing conventional color fundus photographs—a routine clinical modality that has existed for decades—with advanced deep learning architectures, researchers can now extract predictive features far beyond the limit of human perception. In addition to basic demographic and lifestyle indicators such as age, biological sex, body mass index (BMI), blood pressure, smoking status, and complete blood counts, neural networks can probe deeper to identify the presence, trajectory, and severity of systemic illnesses.

Dr. Liu categorizes the current clinical reach of oculomics into the "big three" disease domains: 1. **Neurological disorders:** Predicting both the presence and progression markers of neurodegenerative pathologies such as Alzheimer’s disease and Parkinson’s disease. 2. **Cardiovascular diseases:** Identifying underlying coronary artery disease, structural vessel blockages supplying the myocardium, and quantifying the imminent risk of major adverse cardiovascular events (MACE), including myocardial infarction. 3. **Renal diseases:** Uncovering chronic kidney disease (CKD) at early, asymptomatic stages before progression to end-stage renal failure, while even differentiating underlying etiologies such as diabetic nephropathy versus hypertensive damage.

Dr. Liu emphasizes that these predictive feats are explicitly "superhuman." No ophthalmologist or retina specialist in the world—neither today nor in the foreseeable future—can examine an ophthalmoscopic image and reliably tell a patient their percentage likelihood of suffering a heart attack over the next five years. AI models detect latent morphological, vascular, and textural signatures invisible to human clinical evaluation.

Connecting this breakthrough to broader technology discussions, Dr. Liu references the keynote delivered earlier that day by AI pioneer Dr. Geoffrey Hinton. While Dr. Hinton is famously known for voicing deep reservations regarding AI safety and existential risks, when asked about genuine bright spots where machine learning delivers indisputable societal benefit, he pointed specifically to healthcare—singling out deep learning on retinal imaging as a premier example of AI unlocking transformative diagnostics.

---

[06:00](https://www.youtube.com/watch?v=RguDF7OwHDI&t=360s) Tackling Chronic Disease Burden and Geoffrey Hinton’s Endorsement

The urgent motivation behind AI-driven oculomics lies in the catastrophic public health toll exacted by these target diseases. Cardiovascular disease continues to reign as the single leading cause of mortality across the globe. Concurrently, as global populations age, the incidence and socio-economic burden of Alzheimer’s disease and related dementias are projected to rise exponentially. Chronic cardiovascular, renal, and neurodegenerative illnesses inflict immense human suffering, generate staggering healthcare expenditures, and often progress silently without conspicuous symptoms until substantial organ damage has already occurred.

Deploying deep learning on retinal photography creates a scalable avenue for early detection. Identifying pathological shifts during presymptomatic windows unlocks timely lifestyle, pharmacological, and surgical interventions that substantially improve long-term outcomes at the population level.

```mermaid flowchart TD accTitle: Evolution of Oculomics Through AI

subgraph Anatomy["Biological Foundation"] A["Embryonic Brain Tissue Outgrowth"] B["Living Microvasculature"] end

subgraph Imaging["Non-Invasive Capture"] C["Color Fundus Photography<br/>(1-micron resolution)"] end

subgraph AI["Deep Learning Analysis"] D["Superhuman Feature Extraction"] end

subgraph Clinical["Systemic Risk Prediction"] E["Cardiovascular Risk<br/>(CAD, Heart Attack)"] F["Neurological Disease<br/>(Alzheimer's, Parkinson's)"] G["Renal Health<br/>(Chronic Kidney Disease)"] end

A --> C B --> C C --> D D --> E D --> F D --> G ```

---

[08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) Reimagining the Patient Journey with Community Pre-Screening

Dr. Liu outlines a transformative, near-future care delivery paradigm. Rather than navigating a labyrinthine hospital network, a person could walk down the street into a neighborhood pharmacy, supermarket, or local community center. There, the individual sits before a commercially available, non-mydriatic fundus camera that captures detailed retinal images within seconds—completely bypassing the need for dilating eye drops. The images are securely uploaded to cloud servers, processed by specialized AI pipelines, and evaluated within minutes.

Patients receive an immediate risk assessment highlighting elevated vulnerabilities to heart attacks, kidney dysfunction, or cognitive decline. An on-site health worker or allied provider, fully integrated into the local health delivery network, immediately reviews the findings and routes the patient directly to the appropriate downstream clinical step.

This framework directly attacks the chronic bottlenecks of the legacy healthcare system: - **Fragmentation and friction:** The conventional patient journey requires booking an initial primary care appointment (often requiring months of waiting), securing referral authorizations, undergoing piecemeal lab work, and enduring further delays to consult subspecialists. - **Access barriers:** A vast segment of the domestic and global population lacks an established primary care medical home altogether, remaining completely screened out until acute crises strike. - **Targeted routing:** Under an oculomics-first approach, individuals flag specific physiological hazards before seeing a physician, allowing primary care doctors to immediately order targeted diagnostics (such as specialized lipid panels) or routing patients directly to downstream cardiologists or nephrologists.

```mermaid flowchart LR accTitle: Conventional vs Oculomics-Driven Patient Triage

subgraph Traditional["Legacy Patient Pathway"] direction TB T1["Patient Symptoms / Delay"] --> T2["Primary Care Visit<br/>(Long wait times)"] T2 --> T3["Diagnostic Labs / Workup"] T3 --> T4["Specialist Referral<br/>(Secondary delays)"] end

subgraph Oculomics["AI-Driven Oculomics Pathway"] direction TB O1["Community Pre-Screening<br/>(Retail / Pharmacy / Clinic)"] --> O2["Cloud Deep Learning Triage<br/>(Real-time risk scoring)"] O2 --> O3["Targeted Primary Care"] O2 --> O4["Direct Specialist Fast-Track<br/>(Cardiology / Nephrology)"] end ```

---

[12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) Overcoming Implementation Hurdles to Transform Care

Transforming this screening paradigm into clinical standard-of-care requires overcoming non-trivial structural, operational, and financial hurdles. Dr. Liu notes that the initial scientific proof-of-concept is validated, but turning algorithms into reliable clinical outcomes demands meticulous systemic engineering:

- **Codifying clinical pathways:** Health systems must construct clear, condition-specific referral protocols. A pathway designed to fast-track an elevated myocardial infarction risk differs fundamentally from workflows managing incipient Alzheimer’s or staging diabetic nephropathy, and these pathways must adapt to local healthcare infrastructure. - **Reimbursement and economics:** Society and healthcare policymakers must reach consensus regarding payment models. Payers, health systems, and regulators must determine who finances mass community-level pre-screenings, validate the return on investment via avoided downstream hospitalizations, and establish standardized billing codes. - **Stakeholder incentive alignment:** The multi-sided healthcare ecosystem must be harmonized from day one. Patients, primary care physicians, subspecialists, hardware manufacturers, AI software developers, and commercial/public insurance payers each operate under distinct operational incentives that must align to prevent adoption gridlock.

Dr. Liu concludes by emphasizing that while AI-driven oculomics is no longer science fiction, realizing its potential depends on refining implementation best practices, constructing sound business architectures, and thoughtfully deploying these models to eliminate health disparities and improve medicine worldwide.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) The retina (back of the eye) is the only location in the body, where one can observe blood vessels and brain tissue non-invasively. — Dr. Alvin Liu

> [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) if the chip fails or malfunctions, the camera cannot produce an image, and the patient cannot see. — Dr. Alvin Liu

> [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) the retina is literally an accessible piece of the living brain. — Dr. Alvin Liu

> [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) Commercial retinal cameras provide up to 1,000 times higher spatial resolution of microvascular and neural structures than conventional cross-sectional brain imaging. — Dr. Alvin Liu

> [02:18](https://www.youtube.com/watch?v=RguDF7OwHDI&t=138s) No ophthalmologist or retina specialist in the world—neither today nor in the foreseeable future—can examine an ophthalmoscopic image and reliably tell a patient their percentage likelihood of suffering a heart attack over the next five years. — Dr. Alvin Liu

> [02:18](https://www.youtube.com/watch?v=RguDF7OwHDI&t=138s) AI models detect latent morphological, vascular, and textural signatures invisible to human clinical evaluation. — Dr. Alvin Liu

> [06:00](https://www.youtube.com/watch?v=RguDF7OwHDI&t=360s) Cardiovascular disease continues to reign as the single leading cause of mortality across the globe. — Dr. Alvin Liu

> [06:00](https://www.youtube.com/watch?v=RguDF7OwHDI&t=360s) Identifying pathological shifts during presymptomatic windows unlocks timely lifestyle, pharmacological, and surgical interventions that substantially improve long-term outcomes at the population level. — Dr. Alvin Liu

> [08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) A vast segment of the domestic and global population lacks an established primary care medical home altogether, remaining completely screened out until acute crises strike. — Dr. Alvin Liu

> [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) while AI-driven oculomics is no longer science fiction, realizing its potential depends on refining implementation best practices, constructing sound business architectures, and thoughtfully deploying these models to eliminate health disparities and improve medicine worldwide. — Dr. Alvin Liu

## Mentioned Dr. Alvin Liu, Ai4, Johns Hopkins Medicine, James P Gills Jr MD and Heather Gills Artificial Intelligence Innovation Center, Dr. Geoffrey Hinton, Diabetic retinopathy, Age-related macular degeneration (AMD), Alzheimer's disease, Parkinson's disease, Chronic kidney disease (CKD), Coronary artery disease

## Video description Alvin Liu, Endowed Professor, AI Oversight Team, and Inaugural Director of the James P Gills Jr MD and Heather Gills Artificial Intelligence Innovation Center, Johns Hopkins Medicine

Abstracts

The retina (back of the eye) is the only location in the body, where one can observe blood vessels and brain tissue non-invasively. In recent years, tremendous progress has been made in artificial intelligence (AI). Now, we can combine AI and retinal images to directly predict your age, sex, blood pressure, kidney status, future chance of heart attacks, dementia risk, and more.. If this sounds like science fiction, it is not- the future is here!

[Source](https://www.youtube.com/watch?v=RguDF7OwHDI)

中文简报

约翰斯·霍普金斯借助AI视网膜扫描预测心脏病与痴呆风险 深度学习从常规视网膜照片中提取肉眼不可见的血管与神经生物标志物，在症状出现前预测心脏病发作及器官衰竭；若支付方解决报销难题，该技术将有望实现基于零售网点的分诊。 Alvin Liu 博士介绍了眼组学（oculomics）这一新兴学科，该学科利用深度学习模型分析高分辨率视网膜照片，以提取临床医生肉眼不可见的全身健康生物标志物。借助视网膜作为可触达的神经与血管组织这一独特生物学特征，AI算法在预测心血管事件、肾损伤以及阿尔茨海默病等神经退行性疾病方面展现出超越人类的能力。为最大化其影响力，Liu 博士提议将免散瞳视网膜筛查部署至社区零售空间和药房以简化患者分诊，同时他强调，实现这一前景需要解决临床工作流整合、报销模式以及利益相关方协同等问题。 视网膜作为胚胎间脑的外突组织，是人体唯一能无创观察活体血管与中枢神经系统的部位；商用眼底成像技术可达1微米分辨率，比常规临床脑部MRI高出近1000倍。 (0:04) 新兴学科眼组学（Oculomics）利用深度学习分析普通彩色眼底照片，具备超越人类医生的预测能力，不仅能推断基础人口学特征，还能评估心肌梗死等心血管风险、阿尔茨海默病及慢性肾脏病。 (2:18) AI先驱Geoffrey Hinton特别将视网膜深度学习分析列为人工智能带来显著社会效益的典范，该技术能在无症状期捕捉病理特征，助力在严重器官损伤前实现早期干预。 (6:00) 未来该模式有望在社区药房或超市部署免散瞳眼底相机，数分钟内生成AI风险评分并实现精准分诊，但全面临床普及仍需克服转诊流程规范化、医保支付以及利益相关方协同等落地挑战。 (8:06)

Reading summary

Ophthalmologist Dr. Alvin Liu argues that applying deep learning to high-resolution retinal photography—a field termed oculomics—enables non-invasive detection of major cardiovascular, neurological, and renal diseases. Translating these predictive capabilities into routine practice, however, requires re-engineering community triage pathways, reimbursement models, and clinical workflows. The Retina as an Accessible Neural and Vascular Window (0:04) Dr. Alvin Liu highlights that the retina provides a biologically unique observation point, representing the only site in the human body where living central nervous system tissue and microvasculature can be inspected non-invasively at micron resolution. Because the retina and optic cup form during embryonic development as neuroectodermal outpouchings of the diencephalon, the tissue is an accessible extension of the living brain. Modern commercial fundus cameras achieve optical resolutions down to one micron. Compared to standard clinical brain magnetic resonance imaging, which provides an isotropic resolution of roughly 3 millimeters, off-the-shelf retinal imaging delivers up to 1,000 times higher spatial resolution of microvascular and neural architecture, providing deep learning models with rich structural data unavailable in cross-sectional brain imaging. Superhuman Feature Extraction Across Major Chronic Illnesses (2:18) By analyzing conventional color fundus images, neural networks can predict baseline patient traits—including age, biological sex, blood pressure, body mass index, and smoking status—without prior medical history. Dr. Liu categorizes the deeper clinical reach of oculomics into three major disease categories: cardiovascular disease (quantifying coronary artery disease and five-year risks of major adverse cardiovascular events), neurological conditions (identifying markers and trajectories of Alzheimer’s and Parkinson’s), and renal disease (detecting early asymptomatic chronic kidney disease and differentiating hypertensive damage from diabetic nephropathy). Dr. Liu characterizes these predictive feats as superhuman, noting that no human specialist can inspect a retinal photo and calculate a patient's five-year myocardial infarction likelihood. AI pioneer Dr. Geoffrey Hinton also singled out deep learning in retinal imaging as an indisputable societal benefit of artificial intelligence. Because cardiovascular, neurodegenerative, and renal diseases often develop silently and represent immense global mortality and economic burdens, automated retinal screening offers a scalable route to initiate interventions during presymptomatic stages. Decentralizing Screening Through Neighborhood Non-Mydriatic Imaging (8:06) To improve population-level detection, Dr. Liu proposes shifting initial risk screening from tertiary hospital centers into community retail locations, such as local pharmacies, supermarkets, or neighborhood clinics. Patients can sit for automated, non-mydriatic fundus photography that captures retinal structures within seconds without requiring pupil-dilating drops. Cloud-based neural networks process the images within minutes to deliver risk stratification scores. This community-based pathway is designed to bypass common healthcare system bottlenecks, including multi-month appointment delays, complex referral approvals, and the widespread reality that many individuals lack an established primary care medical home. On-site health personnel can immediately review automated outputs and route patients either to primary care providers for targeted diagnostic workups or directly to cardiologists or nephrologists. Operational and Economic Hurdles to Clinical Adoption (12:32) Despite validated algorithmic proofs-of-concept, Dr. Liu cautions that integrating oculomics into standard clinical care requires overcoming distinct systemic hurdles. Health delivery systems must establish standardized referral protocols, as workflows managing acute myocardial infarction risks differ fundamentally from those handling incipient cognitive decline or early diabetic nephropathy. Additionally, healthcare leaders and regulators must develop clear reimbursement frameworks, identify who finances community screening kiosks, assign dedicated billing codes, and validate economic returns via avoided downstream hospitalizations. Implementation also demands harmonizing the divergent operational incentives of patients, primary care clinicians, subspecialists, hardware manufacturers, software vendors, and insurance payers to prevent adoption gridlock. Dr. Liu concludes that while AI-driven oculomics is clinically viable today, realizing its potential to reduce chronic disease burdens and health disparities depends on establishing clear referral rules, sustainable economic models, and multi-stakeholder alignment across healthcare systems.

精炼阅读摘要

眼科玻璃体视网膜外科专家Alvin Liu博士指出，结合视网膜成像与深度学习算法的新兴“眼组学”（Oculomics），使临床医生能够突破肉眼观察的极限，通过快速、无创的眼底图像精准预测心血管、神经系统及肾脏疾病风险。该技术的全流程普及仍取决于转诊路径标准化、医保报销支付机制以及各方利益格局的系统性协调。 视网膜作为全身血管与脑组织的独特窗口 (0:04) 视网膜是贴附于眼球后壁的一层精密薄层组织，Liu博士将其比作数码相机的感光元件芯片，一旦发生病变将导致不可逆的视力丧失。诸如糖尿病视网膜病变已成为全球劳动年龄人口致盲的首要原因，而年龄相关性黄斑变性（AMD）则是50岁及以上人群严重中心视力丧失的主因。从胚胎发育学来看，视网膜视杯由间脑的神经外胚层直接外凸形成，本质上是裸露在外的活体中枢神经系统的一部分。这使得视网膜成为人体内唯一能够在活体状态下、完全无创且同时直接观察微血管与中枢神经组织的解剖部位。 常规临床脑部磁共振成像（MRI）的各向同性分辨率通常约为3毫米，而普通的商用眼底成像设备在光学层面能精确捕捉单微米级别的微细结构（1毫米等于1000微米）。这意味着常规眼底相机在观察活体神经与微循环结构时，其空间分辨率可达传统断层脑成像的1000倍，为无创探索微循环与神经病理变化提供了前所未有的解剖学基础。 眼组学赋能深度学习实现“超人”级诊断 (2:18) 传统眼科检查依赖瞳孔散瞳，观察信号往往受到人类视觉分辨率和经验的限制，无法提供可靠的全身诊断价值。而新兴的眼组学将常规彩色眼底照片与深度学习神经网络结合，能够从中提取出人类临床医生完全无法识别的微细形态特征、血管走行与纹理模式。目前，算法不仅能高精度推断患者的年龄、生理性别、体重指数（BMI）、血压、吸烟状态与血常规指标，更在三大慢病领域展现出被Liu博士定义为“超越人类能力”的预测表现：其一是在阿尔茨海默病与帕金森病等神经退行性疾病中识别早期病变与进展标志物；其二是在心血管领域检测冠状动脉粥样硬化及管腔狭窄，量化未来发生心肌梗死等主要心血管不良事件（MACE）的几率；其三是在肾脏疾病早期且无症状阶段筛查慢性肾脏病（CKD），并辅助区分糖尿病肾病与高血压肾损害。 Liu博士强调，当前及可预见未来的任何资深眼科医师，都无法仅凭肉眼检眼镜图像评估患者未来5年的心梗概率。AI先驱Geoffrey Hinton在讨论机器学习的现实价值时，亦特别将视网膜深度学习诊断列为AI造福人类社会的典范案例。 社区前置筛查重塑慢病就诊与转诊链路 (8:06) 开发AI驱动的眼组学工具，核心驱动力在于缓解全球慢性病造成的沉重公共卫生危机。心血管疾病至今仍是全球首要致死原因，而随着人口老龄化加剧，阿尔茨海默病等认知障碍的患病负担呈指数级上升。此类慢性病通常起病隐匿，在出现明显临床症状时脏器往往已发生实质性损害。现行医疗体系流程繁琐碎片化，患者通常面临漫长的初级保健预约等待、繁复的转诊授权及重复的零散化验，甚至大量人群缺乏固定的基础医疗照护。 Liu博士构想了一种去中心化的近未来筛查模式：患者无需散瞳滴眼，只需在社区药店、超市或活动中心使用免散瞳眼底相机，在数秒内完成拍摄；图像自动上传至云端由深度学习模型分析，数分钟内即可生成针对心血管、肾脏及认知障碍的风险分层报告。现场医护人员可根据评估结果，协助患者直接对接入组精准干预路径——或是协助全科医生安排特定的专项化验，或是将高危患者直接分流至心内科或肾内科专科，在临床前黄金窗口期实现早期干预。 临床工程落地面临路径规范与经济学瓶颈 (12:32) 尽管眼组学的早期科学概念验证已经确立，但Liu博士指出，将前沿算法转化为制度化的临床标准流程仍需攻克复杂的工程化、运营与政策挑战。首先是临床转诊路径的标准化建设，医疗系统必须为不同靶向疾病设计因地制宜的转诊预案，心肌梗死风险的紧急转诊流程与阿尔茨海默病早期干预或糖尿病肾病分期管理存在本质差异，必须与各区域既有的医疗资源深度接轨。 其次是医疗经济学与报销机制的确定。公共卫生决策者与保险机构需要达成明确共识，明确谁来为基层社区的大规模筛查买单，设立标准的诊疗计费代码，并通过避免下游急重症住院来测算长期投资回报率。最后，医疗生态各方的利益必须紧密协同，包括患者、全科医生、专科医生、硬件设备商、算法研发企业以及公共和商业保险机构，唯有理顺各方的运营动机，才能防止技术推广陷入体制性僵局。 Liu博士总结认为，基于视网膜的AI多系统健康预测不再是科幻构想，但要真正释放其潜能以消除医疗不平等并降低重大疾病负担，医疗界不仅需要持续验证算法效能，更必须着力打通支付报销机制、构建严密的转诊通路，并推动跨生态利益主体的深度协同。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2025 年 8 月 28 日在 YouTube 平台发布。视频记录了约翰斯·霍普金斯大学医学院 Alvin Liu 博士的主题演讲，介绍如何利用人工智能分析视网膜图像，实现对年龄、性别、血压、心血管疾病、阿尔茨海默病及肾病风险等多项全身健康指标的无创精准预测。 ## TL;DR 演讲全面介绍了前沿交叉学科“眼组学”（Oculomics），指出视网膜是人体唯一能以微米级超高分辨率直接无创观测活体血管与中枢脑神经组织的窗口。借助深度学习对经典彩色眼底照片的分析，AI 具备了超越人类医生的预测能力，不仅能推断基础生理指标，还能对神经系统、心血管和肾脏三大系统重疾进行早期风险分层。该技术有望推动社区免散瞳筛查普及，优化分诊与转诊路径，重塑传统就医流程。最后，演讲强调了推进临床落地时在标准化工作流、多方利益协同与支付商业模式上面临的现实挑战。 ## Atmosphere & style 整场演讲充满前沿科技的启迪感与医学人文关怀，Liu 博士以通俗生动的比喻开篇并与台下亲切互动，迅速拉近了专业医学与大众认知的距离。叙述节奏层层递进，从眼底解剖学基础平稳过渡到 AI 展现“超人能力”的震撼转折，并引用 Geoffrey Hinton 的评价增强技术说服力。整体氛围兼具严谨务实的临床考量与对颠覆性未来的坚定乐观，带给观众一种“未来已来、触手可及”的强烈感染力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) 视网膜是眼球后壁将光信号转化为图像的关键感光组织，其病变（如糖尿病视网膜病变和年龄相关性黄斑变性 AMD）是导致成人不可逆致盲与视力丧失的主要诱因。 - [01:53](https://www.youtube.com/watch?v=RguDF7OwHDI&t=113s) 眼组学（Oculomics）是一门结合视网膜生物标志物与深度学习算法，用于评估和预测全身系统性健康状况的前沿交叉学科。 - [01:53](https://www.youtube.com/watch?v=RguDF7OwHDI&t=113s) 视网膜是人体内唯一无需创口即可同时以微米级超高分辨率直接观察活动血管和中枢神经脑组织的解剖窗口，其分辨率比临床脑部 MRI 高近千倍。 - [04:11](https://www.youtube.com/watch?v=RguDF7OwHDI&t=251s) 深度学习算法能够通过彩色眼底图像精准预测个体的基础生理指标，包括年龄、生理性别、血压水平、BMI 指数、吸烟状况及全血细胞计数。 - [04:11](https://www.youtube.com/watch?v=RguDF7OwHDI&t=251s) AI 模型展现出超越人类医生的诊断洞察力，能够有效评估神经系统疾病（如阿尔茨海默病、帕金森病）、心血管疾病（冠心病与心梗风险）以及早期慢性肾脏病与病因分类。 - [04:11](https://www.youtube.com/watch?v=RguDF7OwHDI&t=251s) AI 先驱 Geoffrey Hinton 在关注 AI 安全风险的同时，对医疗健康应用抱有高度乐观，并特别肯定了利用深度学习分析视网膜图像预测全身健康的突破。 - [08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) 将免散瞳眼底相机部署在药房、超市或社区中心，能在数分钟内出具全身疾病风险报告，推动筛查场景下沉并解决医疗流程碎片化与准入门槛高的痛点。 - [11:33](https://www.youtube.com/watch?v=RguDF7OwHDI&t=693s) 基于眼组学的新模式通过扩大患者触达漏斗、在上游实现精准分诊并促进极早期干预，彻底重塑就医流程并显著改善长期健康预后。 - [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) 眼组学临床规模化落地仍面临挑战，亟需根据疾病与机构类型定制标准化转诊路径，并明确支付与报销共识。 - [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) 实现眼组学医疗愿景需要协同平衡患者、初级保健医生、专科医生、软硬件技术研发方及医保支付方的多方诉求与激励机制。 - [13:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=812s) 眼组学时代已经到来，未来需确立最佳操作规范、平衡生态激励并建立可持续商业模式，以全面赋能全身健康预测与疾病干预。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) 视网膜：人体内部的“数码相机芯片”与常见眼底病变 演讲伊始，约翰斯·霍普金斯大学医学院的 Alvin Liu 博士向现场听众做自我介绍。他是一名眼科医生，专业方向为视网膜专科外科医生（Retinal Surgeon），专门诊治眼底后部结构的各类病变。为了让大众理解这一复杂的眼部构造，他打了一个生动的比方：视网膜是衬于眼球后壁的一层纤薄而脆弱的组织，其功能与数字相机内部的感光芯片如出一辙——它是将光信号转化为图像的核心部件；一旦视网膜受损或发生功能失常，人眼便无法获得清晰的视力。 Liu 博士指出，视网膜相关疾病在日常生活中十分常见。他展示了典型的临床眼底图像，并着重介绍了两种代表性病症：其一是糖尿病视网膜病变（Diabetic Retinopathy），由全身性糖尿病侵害眼底微血管所致，目前已成为全球劳动适龄人口不可逆致盲的首要病因；其二是年龄相关性黄斑变性（Age-related Macular Degeneration, AMD），在美国乃至全世界范围内，它都是导致 50 岁以上人群中心视力丧失的最主要原因。 在阐述这些基础知识后，Liu 博士与现场观众展开了互动，询问大家在过去两年内有多少人看过眼科，又有多少人在就诊时接受过散瞳检查。在看到台下举起的双手后，他微笑着表示，散瞳确实是视网膜专科极为日常的标准流程，医生必须通过扩大瞳孔才能清晰观察眼底的病理变化。 - 视网膜的生理地位：眼球内壁的感光组织，结构与作用类似于数码相机中的传感器芯片。 - 糖尿病视网膜病变：由于全身血管病变损害眼底微血管，是全球工作年龄人群不可逆盲的主要元凶。 - 年龄相关性黄斑变性（AMD）：50 岁以上中老年人群发生中心视力丧失的首要诱因。 ### [01:53](https://www.youtube.com/watch?v=RguDF7OwHDI&t=113s) 眼组学（Oculomics）的兴起：从视网膜透视全身健康 随后，Liu 博士抛出了一个令人惊叹的设想：“如果我告诉你，我不需要事先了解你的任何背景信息，甚至在和你见面或交谈之前，只需要拍摄一张你的眼底视网膜照片，将其输入 AI 模型，就能高置信度地预测你的真实年龄、生理性别、血压水平以及是否吸烟，你会怎么想？”他紧接着自问自答道：“这难道是科幻小说吗？不，它绝非科幻，而是现实中已经发生的技术——欢迎来到眼组学（Oculomics）的世界。” 眼组学是一门专注于探索视网膜生物标志物与全身其他器官生理病理状态之间内在关联的学科。Liu 博士解释道，医学界数十年前就意识到视网膜微血管和神经表现与全身健康高度相关，但在过去，人类肉眼所能捕获的生物学信号十分有限，直到近年来人工智能与深度学习的爆发式发展，这些潜藏在图像深处的微弱信号才被彻底挖掘出来。 视网膜之所以具备如此独特的诊断价值，关键在于其不可替代的解剖学特质：它是人体内唯一一个无需穿刺或切口即可同时以微米级超高分辨率直接无创观测活体血管和中枢神经脑组织的窗口。从胚胎发育学角度来看，人类在母体子宫内发育时，视网膜本身就是大脑神经管向外延伸的一部分，“在本质上它就是你大脑的一部分”。在成像精度方面，1 毫米等于 1000 微米，常规临床脑部核磁共振（MRI）的空间分辨率大约在 3 毫米级别，而市面上普通的商用眼底成像设备就能达到微米级精度，这意味着普通眼底相机提供的成像分辨率比传统脑成像技术高出近 1000 倍。 - 眼组学（Oculomics）：通过分析视网膜特征来评估全身系统性健康状况的前沿交叉学科。 - 视网膜的解剖独特性：全身唯一可无创、同时直接观察活动血管及中枢脑神经组织的结构。 - 分辨率对比：临床常规脑部 MRI 分辨率约为 3 毫米，而常规眼底图像能达到微米（micron）级，分辨率相差近千倍。 ### [04:11](https://www.youtube.com/watch?v=RguDF7OwHDI&t=251s) 深度学习的超人能力与三大疾病领域的突破 步入 2025 年，将经典的彩色眼底照相（Color Fundus Photography）与前沿的深度学习算法相结合，已经能够实现诸多惊人的全身指标预测。除了推断年龄、生理性别、血压和吸烟状态外，AI 还可以准确评估体质指数（BMI）和血细胞计数。不仅如此，算法还能进一步深入疾病诊断层面，在神经系统疾病、心血管疾病以及肾脏疾病这“三大重疾门类”中展现出惊人的洞察力。 ```mermaid flowchart TD accTitle: AI结合视网膜成像预测全身疾病架构 subgraph Input["数据输入层"] A["标准彩色眼底照相\n（无创微米级成像）"] end subgraph AI["深度学习分析引擎"] B["深度学习图像模型\n（多模态生物特征提取）"] end subgraph Biomarkers["基础生理指标"] C1["年龄 / 生理性别"] C2["血压 / BMI 指数"] C3["血细胞计数 / 吸烟史"] end subgraph SystemicDiseases["三大系统重疾预测"] D1["神经系统疾病\n· 阿尔茨海默病\n· 帕金森病"] D2["心血管系统疾病\n· 冠心病与血管狭窄\n· 近期心肌梗死风险"] D3["肾脏系统疾病\n· 慢性肾脏病诊断\n· 病因分类（糖尿病等）"] end A --> B B --> Biomarkers B --> SystemicDiseases ``` 在神经系统方面，AI 能够通过眼底神经纤维层的微观变化预测阿尔茨海默病（Alzheimer's）和帕金森病（Parkinson's）的存在与否；在心血管系统方面，算法可以评估患者是否患有冠心病（冠状动脉阻塞），并预测其在近期发生急性心肌梗死等严重心血管不良事件的概率；在肾脏疾病方面，慢性肾病在早期通常毫无症状，待到患者察觉时往往已处于晚期竭竭阶段，而 AI 不仅能筛查出早期慢性肾病，甚至还能推断肾功能损害是由糖尿病并发症还是其他诱因引起的。 Liu 博士着重强调，这种预测能力是真正意义上的“超人（Superhuman）能力”。他向观众断言：“我敢保证，全世界没有任何一位眼科医生——无论是现在还是未来——能够单凭肉眼看一眼患者的眼底，就能断定此人近期患心脏病的风险有多高。”为了佐证这项技术的深远前景，Liu 博士还提及了当天早些时候人工智能先驱 Geoffrey Hinton 博士的主题演讲。尽管 Hinton 博士在安全问题上对 AI 的失控表达了诸多担忧，但在被问及对哪些 AI 应用场景仍抱有强烈乐观态度时，Hinton 特别将医疗健康列为一大亮点，并当场点名表扬了“利用深度学习挖掘视网膜图像预测全身健康”这项突破。 - 预测维度：基础生理体征（年龄、性别、血压、BMI、吸烟状况、全血细胞指标）及系统性重大疾病。 - 神经病变筛查：在临床前或发病早期识别阿尔茨海默病与帕金森病。 - 心血管预警：量化冠脉病变程度，对未来突发心肌梗死（Heart Attack）进行风险分层。 - 肾病早期溯源：在症状显现前发现慢性肾脏病（CKD），并辅助鉴别原发病因。 ### [08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) 重塑患者就医旅程：社区筛查与精准分流的未来图景 为何这项技术如此关键？Liu 博士指出，心血管疾病是全球致死率最高的头号杀手，而随着人口老龄化加剧，阿尔茨海默病和痴呆症的患病率与医疗负担正呈指数级攀升。这些慢性疾病不仅耗费高昂医疗资源，更会给患者及其家庭带来长期的病痛折磨。如果能在基层实现更早期的疾病筛查，便能触发及时的临床干预，从而在大规模人群层面彻底改善长期健康预后。 为了让大家具象化理解这一变革，Liu 博士描绘了正在成为现实的应用场景：在不远的将来，人们只需在散步时走进身边的社区超市、街角药房或社区活动中心，坐在现成的免散瞳商用眼底相机前拍张照；照片实时上传至云端并经由多种 AI 模型分析，短短几分钟之内，受检者就能获知自己未来患心脏病、痴呆症或肾病的风险评估。现场配备的基层医务人员会直接将其健康档案接入当地医疗网络，并根据 AI 的评估等级无缝衔接后续诊疗。 当前的医疗体系高度碎片化且门槛重重，患者往往因预约全科医生或专科医生周期漫长、就医流程繁琐混乱而错失慢性病的黄金干预窗口；甚至有相当一部分人群根本无法享受到初级卫生保健服务。而眼组学驱动的基层智能筛查体系彻底颠覆了传统的被动看病模式：AI 风险报告能让全科医生在患者踏入诊室之前便明确重点筛查方向（如直接开具针对性的血脂与胆固醇检验单），甚至对于极高危患者，系统还能协助跳过全科医生环节，直接精准转诊至心内科或神经科专科专家手中，从而极大缩减诊疗路径。 - 筛查场景下沉：免散瞳眼底相机入驻药房、超市与社区中心，实现数分钟级实时云端风险反馈。 - 破解医疗痛点：消除就医流程碎片化、排队预约周期长及缺乏初级保健准入的核心瓶颈。 - 诊疗路径优化：由 AI 驱动靶向转诊，全科医生可提前调阅预测指标，极高危人群可直通专科会诊。 ### [11:33](https://www.youtube.com/watch?v=RguDF7OwHDI&t=693s) 基于眼组学（Oculomics）重塑医疗旅程与核心临床优势 结合人工智能与视网膜图像分析的眼组学（Oculomics）新范式，能够精准指引患者前往肾脏科、心血管科等对应专科，从而根本性地修复目前破碎且效率低下的传统就医流程。通过将前沿 AI 技术与视网膜成像深度融合，医疗系统不仅能够打破学科壁垒，还能构建一套更加顺畅、高效的诊疗闭环。 这一基于 AI 的眼组学范式带来了三大核心层面的系统性红利： - **扩大患者触达漏斗（Patient Capture Funnel）**：大幅降低筛查门槛，拓宽早期发现患者的渠道，从源头上显著改善患者获得优质医疗服务的可及性。 - **实现精准的上游分诊（Upstream Triage）**：在诊疗最前端进行精准风险评估与人群分流，避免不必要的检查与盲目就医，使下游宝贵的医疗资源得到更高效的配置与利用。 - **推动早期干预并改善预后**：让潜在的全身性疾病在无症状或极早期阶段被捕捉并尽早干预，最终显著提升患者的长期健康预后。 ```mermaid flowchart TD accTitle: AI-based Oculomics Healthcare Paradigm subgraph Screening["前端筛查与分诊 (Upstream)"] A["视网膜图像采集 (无创/快速)"] --> B["AI 算法深度分析"] B --> C["扩宽患者触达漏斗 & 精准风险分层"] end subgraph DownstreamCare["下游医疗资源协同 (Downstream)"] C --> D["定向转诊 (肾脏科 / 心血管科 / 神经科)"] D --> E["极早期临床干预"] E --> F["显著改善长期健康结局"] end ``` --- ### [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) 落地挑战：转诊路径、支付共识与利益相关方协同 尽管底层科学与技术突破令人振奋，但要将这一范式全面推向临床实践，仍有艰巨的现实工作需要完成。首要任务是明确并标准化最优的转诊路径与临床工作流。这种工作流程并非千篇一律，必须针对具体的疾病类型（如心脑血管风险、肾脏疾病或认知障碍）以及具体的实施地点和医疗机构类型进行精细化定制。 与此同时，商业与支付机制的建立是决定其能否规模化落地的关键瓶颈。全社会必须就“谁来为此买单、为何买单以及支付多少金额”达成明确的共识与标准。在医疗生态体系中，涉及的利益主体错综复杂，从项目启动的第一天起，就必须系统性地协调并对齐各方利益诉求，这些核心利益相关方包括： - **患者（Patients）**：获得更便捷、无创且早期的疾病风险洞察。 - **初级保健医生（Primary Care Doctors）与专科医生（Specialists）**：获得精准辅助诊断工具与高质量转诊支持。 - **技术研发方**：包括开发视网膜成像设备的硬件公司与研发诊断算法的软件公司。 - **医疗保险支付方（Medical Insurance Payers）**：通过前置预防与早期干预降低整体医疗支出成本。 --- ### [13:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=812s) 总结与展望：眼组学时代的到来与未来医疗变革 AI 赋能的眼组学时代已经正式开启。通过将尖端 AI 算法与视网膜图像相结合，系统仅需短短数分钟就能准确评估一个人未来罹患痴呆症（Dementia）、心脏病发作或肾脏损伤等重大疾病的风险。"The future is here. This is no longer science fiction."（未来已来，这不再是科幻小说）。眼底血管与神经组织的无创透视，真正成为了窥探全身健康奥秘的窗口。 要彻底实现这一医疗愿景，行业接下来必须在三大现实维度深耕细作：摸索并确立最佳的操作规范（Operational Best Practices）、全面平衡生态内多方利益相关者的激励机制，以及持续打磨并开发出具有可持续性的新型商业模式。这项前沿探索向我们展示了极具变革性的医学未来，有望彻底颠覆传统诊疗模式，引领人类医疗健康事业迈向更美好的新阶段。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=RguDF7OwHDI&t=4s) 视网膜是衬于眼球后壁的一层纤薄而脆弱的组织，其功能与数字相机内部的感光芯片如出一辙——它是将光信号转化为图像的核心部件；一旦视网膜受损或发生功能失常，人眼便无法获得清晰的视力。 — Alvin Liu > [01:53](https://www.youtube.com/watch?v=RguDF7OwHDI&t=113s) 如果我告诉你，我不需要事先了解你的任何背景信息，甚至在和你见面或交谈之前，只需要拍摄一张你的眼底视网膜照片，将其输入 AI 模型，就能高置信度地预测你的真实年龄、生理性别、血压水平以及是否吸烟，你会怎么想？ — Alvin Liu > [01:53](https://www.youtube.com/watch?v=RguDF7OwHDI&t=113s) 这难道是科幻小说吗？不，它绝非科幻，而是现实中已经发生的技术——欢迎来到眼组学（Oculomics）的世界。 — Alvin Liu > [01:53](https://www.youtube.com/watch?v=RguDF7OwHDI&t=113s) 在本质上它就是你大脑的一部分。 — Alvin Liu > [04:11](https://www.youtube.com/watch?v=RguDF7OwHDI&t=251s) 我敢保证，全世界没有任何一位眼科医生——无论是现在还是未来——能够单凭肉眼看一眼患者的眼底，就能断定此人近期患心脏病的风险有多高。 — Alvin Liu > [08:06](https://www.youtube.com/watch?v=RguDF7OwHDI&t=486s) 在不远的将来，人们只需在散步时走进身边的社区超市、街角药房或社区活动中心，坐在现成的免散瞳商用眼底相机前拍张照；照片实时上传至云端并经由多种 AI 模型分析，短短几分钟之内，受检者就能获知自己未来患心脏病、痴呆症或肾病的风险评估。 — Alvin Liu > [12:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=752s) 全社会必须就‘谁来为此买单、为何买单以及支付多少金额’达成明确的共识与标准。 — Alvin Liu > [13:32](https://www.youtube.com/watch?v=RguDF7OwHDI&t=812s) 未来已来，这不再是科幻小说。 — Alvin Liu ## Mentioned Alvin Liu, Geoffrey Hinton, 约翰斯·霍普金斯大学医学院 (Johns Hopkins Medicine), James P Gills Jr MD and Heather Gills 人工智能创新中心, Ai4, 眼组学 (Oculomics), 糖尿病视网膜病变 (Diabetic Retinopathy), 年龄相关性黄斑变性 (AMD), 阿尔茨海默病 (Alzheimer's Disease), 帕金森病 (Parkinson's Disease), 慢性肾脏病 (CKD), 彩色眼底照相 (Color Fundus Photography), 磁共振成像 (MRI) ## Video description Alvin Liu, Endowed Professor, AI Oversight Team, and Inaugural Director of the James P Gills Jr MD and Heather Gills Artificial Intelligence Innovation Center, Johns Hopkins Medicine Abstracts The retina (back of the eye) is the only location in the body, where one can observe blood vessels and brain tissue non-invasively. In recent years, tremendous progress has been made in artificial intelligence (AI). Now, we can combine AI and retinal images to directly predict your age, sex, blood pressure, kidney status, future chance of heart attacks, dementia risk, and more.. If this sounds like science fiction, it is not- the future is here! [Source](https://www.youtube.com/watch?v=RguDF7OwHDI)

Source quotes

[0:04] Alvin Liu: 视网膜是衬于眼球后壁的一层纤薄而脆弱的组织，其功能与数字相机内部的感光芯片如出一辙——它是将光信号转化为图像的核心部件；一旦视网膜受损或发生功能失常，人眼便无法获得清晰的视力。 [1:53] Alvin Liu: 如果我告诉你，我不需要事先了解你的任何背景信息，甚至在和你见面或交谈之前，只需要拍摄一张你的眼底视网膜照片，将其输入 AI 模型，就能高置信度地预测你的真实年龄、生理性别、血压水平以及是否吸烟，你会怎么想？ [1:53] Alvin Liu: 这难道是科幻小说吗？不，它绝非科幻，而是现实中已经发生的技术——欢迎来到眼组学（Oculomics）的世界。 [1:53] Alvin Liu: 在本质上它就是你大脑的一部分。 [4:11] Alvin Liu: 我敢保证，全世界没有任何一位眼科医生——无论是现在还是未来——能够单凭肉眼看一眼患者的眼底，就能断定此人近期患心脏病的风险有多高。 [8:06] Alvin Liu: 在不远的将来，人们只需在散步时走进身边的社区超市、街角药房或社区活动中心，坐在现成的免散瞳商用眼底相机前拍张照；照片实时上传至云端并经由多种 AI 模型分析，短短几分钟之内，受检者就能获知自己未来患心脏病、痴呆症或肾病的风险评估。 [12:32] Alvin Liu: 全社会必须就‘谁来为此买单、为何买单以及支付多少金额’达成明确的共识与标准。 [13:32] Alvin Liu: 未来已来，这不再是科幻小说。

Source quotes

[0:04] Dr. Alvin Liu: The retina (back of the eye) is the only location in the body, where one can observe blood vessels and brain tissue non-invasively. [0:04] Dr. Alvin Liu: if the chip fails or malfunctions, the camera cannot produce an image, and the patient cannot see. [0:04] Dr. Alvin Liu: the retina is literally an accessible piece of the living brain. [0:04] Dr. Alvin Liu: Commercial retinal cameras provide up to 1,000 times higher spatial resolution of microvascular and neural structures than conventional cross-sectional brain imaging. [2:18] Dr. Alvin Liu: No ophthalmologist or retina specialist in the world—neither today nor in the foreseeable future—can examine an ophthalmoscopic image and reliably tell a patient their percentage likelihood of suffering a heart attack over the next five years. [2:18] Dr. Alvin Liu: AI models detect latent morphological, vascular, and textural signatures invisible to human clinical evaluation. [6:00] Dr. Alvin Liu: Cardiovascular disease continues to reign as the single leading cause of mortality across the globe. [6:00] Dr. Alvin Liu: Identifying pathological shifts during presymptomatic windows unlocks timely lifestyle, pharmacological, and surgical interventions that substantially improve long-term outcomes at the population level. [8:06] Dr. Alvin Liu: A vast segment of the domestic and global population lacks an established primary care medical home altogether, remaining completely screened out until acute crises strike. [12:32] Dr. Alvin Liu: while AI-driven oculomics is no longer science fiction, realizing its potential depends on refining implementation best practices, constructing sound business architectures, and thoughtfully deploying these models to eliminate health disparities and improve medicine worldwide.

Original transcript

[0:04] Hello everyone. Um, welcome. Uh, my name [0:06] is Dr. Alvin Louu. I'm an opthalmologist [0:09] by training. Specifically, a retinal [0:12] specialist, a retinal surgeon. So, a [0:14] retinal surgeon is someone who takes [0:17] care of patients with issues in the back [0:19] of the eye, the retina. The retina is [0:22] that thin, delicate tissue that lines [0:24] the back of the eye. um it's analogous [0:27] to the computer chip inside a digital [0:29] camera. So it's a very important [0:32] structure. If the retina does not work [0:34] or does not function properly, then you [0:37] will not be able to see well. [0:41] Retinal conditions are fairly common. [0:43] You may have heard some of them. Um for [0:46] example, the image on the left is [0:48] something called diabetic retinopathy. [0:50] Essentially it happens when blood [0:53] vessels in the back of the eye um gets [0:55] damaged by systemic diabetes and in fact [0:58] diabetic renopathy is the leading cause [1:01] of irreversible blindness in the working [1:04] age population around the world. The [1:06] other very common condition is called [1:09] age related macular degeneration. You [1:12] may have heard of it before. Um it is [1:14] the leading cause of central vision loss [1:17] in persons over the age of 50. um in the [1:20] United States and around the world. [1:23] So, which one of you have seen an [1:25] opthalmologist in the past let's say two [1:28] years? Let me get a show of hand. So, a [1:30] few of you have have seen one. And how [1:34] many of you were dilate dilated when you [1:37] saw the opthalmologist? [1:39] Okay. So, it is a pretty common [1:41] experience. If you were my patient and [1:44] you come to my clinic, one of the things [1:46] I would do as a retinal specialist is to [1:48] dilate your eyes so that I can see, you [1:51] know, what's happening in the retina. [1:53] What if I tell you I don't need to know [1:56] anything about you and even before [1:58] meeting you or having a conversation at [2:01] all. All I need is to take your Renault [2:04] images, run them through AI models, and [2:08] be able to predict with high confidence [2:11] your age, your biological sex, your [2:15] blood pressure, and whether you smoke or [2:16] not. [2:18] Is that science fiction? [2:22] The answer is no. It is happening [2:24] already. It's possible. And welcome to [2:27] the world of alkalomics. [2:30] So alkalomics is a field of study that [2:33] studies the association between retinal [2:36] biomarkers, meaning findings in the [2:38] retina and what's happening in the rest [2:41] of the body. This is a concept that has [2:44] been around for decades. We've always [2:46] known that there's a high degree of [2:48] correlation and association between what [2:50] we see in the retina and the rest of the [2:52] body. But the signal and the information [2:56] that we could get was never strong [2:59] enough until now with the rise of AI. [3:03] And you may ask why is this even [3:05] possible? What's so special about the [3:07] retina? So the retina is the only place [3:10] in your body where we can image and [3:13] visualize both blood vessels and brain [3:16] tissue at the same time non-invasively [3:19] at a micron level resolution. In fact, [3:23] when you were an embryo inside your [3:25] mother's womb and when your body was [3:27] developing, the retina was a direct [3:30] extension of the brain. It's literally [3:33] your brain. [3:36] And to provide a little bit more [3:38] context, there are 1,00 microns in [3:41] millimeter. And a lot of these cutting [3:44] edge retinal imaging platform can image [3:47] at the level of a micron. And for [3:49] comparison, a typical resolution of a [3:52] brain MRI is about 3 millimeter. Meaning [3:55] these are offtheshelf commonly available [3:58] reno imaging platforms. These machines [4:01] really can deliver up to 1,000 times [4:04] higher resolution as compared to [4:07] conventional brain imaging. [4:11] So let's fast forward and just talk [4:13] about what's happening today 2025. If we [4:17] combine retinal images specifically [4:20] color furnace photograph really it's [4:21] just a color photo of the back of your [4:23] eye. This technology has been around for [4:26] decades. If you combine that with deep [4:28] learning, which is the cutting edge AI [4:30] technique for image analysis, these are [4:32] the things that we can predict as I [4:35] mentioned earlier, your age, biological [4:38] sex, blood pressure, [4:41] your body mass index, your blood counts, [4:44] and whether you smoke cigarettes or not. [4:47] We can even go one step further and one [4:51] layer deeper and predict categories of [4:53] disease. And the big three that are [4:56] possible now are neurological diseases, [5:00] cardiovascular diseases, and kidney [5:01] diseases. [5:04] Let's do a turn deeper. Specifically, [5:06] when it comes to neurological diseases, [5:08] we can predict the presence or absence [5:10] of Alzheimer's and Parkinson's. [5:13] When it comes to cardiovascular [5:15] diseases, we can tell whether you have [5:17] coronary heart disease, meaning whether [5:20] the blood vessels that supply your heart [5:22] are blocked or not. and whether you're [5:24] at a higher risk for some major cardio [5:28] cardiovascular adverse event in the near [5:30] future, for example, a heart attack. [5:33] When a K when it comes to kidney [5:35] disease, oftent times you um you don't [5:38] really know you have issues with the [5:40] kidneys until it's really endstage and [5:42] severe. So again, with these techniques, [5:44] we can certainly predict the presence of [5:48] chronic kidney disease and maybe even go [5:50] one step further and try to figure out [5:52] why you're having kidney problems to [5:54] begin with, whether it's from diabetes [5:56] or something else. [5:59] I want to emphasize that all the [6:01] association and prediction that I [6:03] mentioned it's superhuman meaning I [6:07] guarantee you there's not a single [6:09] opthalmologist in the world right now on [6:11] a future ever can simply look at your [6:13] eyes and your retina and tell whether [6:16] you're at elevated risk for having a [6:18] future heart attack. This is literally [6:20] superhuman capability. And for those who [6:23] were at Dr. Jeff Hinton's keynote [6:25] address earlier today um you know [6:28] throughout the series of questions at [6:30] towards the very end he was asked you [6:32] know as you know like Dr. Hinton has a [6:35] lot of reservations about the safety of [6:37] AI. Towards the end he was asked well [6:40] are there certain bright spots in AI [6:42] that you like and he did mention [6:44] healthcare and how AI could potentially [6:46] improve health care significantly and he [6:49] used this specific example. He [6:52] specifically said he found that um by [6:54] combining deep learning AI and retinal [6:56] images we can now gain very valuable [7:00] information about the rest of body. So [7:02] that's one specific example that Dr. [7:05] Jeff Hinton liked and this is what we're [7:07] talking about here. [7:09] So you may ask great you we can do we [7:13] can detect three major categories of [7:16] diseases. So what why do we care about [7:17] it? Well we care about it because these [7:20] are important diseases. For example um [7:22] cardiovascular disease is the leading [7:24] cause of death around the world already. [7:28] When it comes to Alzheimer's and [7:30] dementia, its incidence and disease [7:32] burden is expected to increase [7:34] exponentially with population aging. [7:39] And also these conditions, I'm sure you [7:41] know of someone, friends, family who [7:44] have these conditions. They're chronic. [7:46] They're costly. They create a lot of [7:48] morbidity. And in general, by combining [7:51] AI and renal images, this should allow [7:53] us to perform earlier detection, which [7:56] will lead to timely intervention and [7:59] very likely better long-term health [8:01] outcomes for patients on a population [8:03] level. [8:06] So, imagine a world [8:08] this how this is going to play out. and [8:11] just imagine a world in a notsodistant [8:13] future. You can simply walk down your [8:16] your street to your favorite local [8:19] supermarket or pharmacy or community [8:22] center. You sit down in front of these [8:24] machines, which by the way are [8:26] offthe-shelf commercially available [8:28] retinal cameras. You sit down, you get [8:31] your retinal images um taken for the [8:34] most part without the need for dilation. [8:37] These images get sent to the cloud and [8:40] get analyzed by various AI models and [8:43] within minutes in real time you will be [8:46] notified whether you're at an elevated [8:48] risk for having heart attack or dementia [8:50] or kidney disease. Then we can have an [8:52] on-site medical personnel who who's [8:55] already plugged in with the various [8:57] healthcare providers in your area and [8:59] based on the AI analysis results you'll [9:02] be sent on your next step of your [9:03] healthcare journey. This [9:07] future is imminent. We can make it [9:11] happen. In fact, it is happening right [9:13] now. [9:15] So essentially what we're trying to [9:16] address is a fragmented healthcare [9:18] system where access to care is limited [9:21] which in turn leads to suboptimal [9:23] detection, prevention and treatment of [9:26] co costly chronic diseases. [9:29] This is schematic representation of what [9:32] a typical patient journey involves. Um [9:34] there are many stakeholders involved [9:36] there many different moving parts and [9:38] oftentimes it is fragmented and it's a [9:41] very confusing and even frustrating [9:43] journey for individual patients [9:48] and this is how neurological [9:50] cardiovascular and kidney disease are [9:53] typically true action manage right now. [9:56] If you have an issue or um you have some [9:59] symptoms, you may go to your primary [10:02] care doctor which is your first step of [10:04] the journey. You talk to your primary [10:06] care doctor uh about some of your [10:08] symptoms or maybe it's just for routine [10:09] check and then based on that interaction [10:12] um your primary care doctor will either [10:14] send you for some for some tests or [10:16] further work up uh with a specialist. [10:19] The issue with this current workflow is [10:21] as you know it takes a long time to even [10:24] see a primary care doctor. It usually [10:27] takes even longer to see a specialist [10:30] and a significant portion of the [10:34] population in US and really around the [10:36] world may not even have any access to [10:38] primary care to begin with. So that's [10:40] that's a big issue [10:43] and this is what possible with a new [10:47] paradigm that's driven by oalomics and [10:51] pre-screening. [10:53] You'll get the screening as I mentioned [10:55] in a community center and based on [10:57] results of AI model you will be sent to [11:01] a primary care doctor in a much more [11:03] targeted fashion. Meaning before you [11:05] even go to your primary doctor, he or [11:07] she may know. So with these AI results [11:11] available, you will be referred to your [11:13] primary care doctor in a much more [11:15] targeted fashion. Meaning he or she may [11:17] know that you're at an elevated risk for [11:19] having a heart attack. So he or she may, [11:21] you know, send you for additional blood [11:22] test, looking at cholesterol, maybe [11:24] something else. Or even yet even better [11:26] yet we may be able to skip this step [11:29] alto together and send you straight to a [11:31] specialist straight to like like [11:33] nefologists or card like cardiologists. [11:36] So this is what this new paradigm could [11:39] enable. [11:42] In other words [11:44] the field of economics specifically by [11:47] harnessing the incredible power of AI [11:49] and ret images could fix a very broken [11:53] healthcare journey. [11:56] And these are some of the additional [11:57] benefits of an AI based oculomics [12:00] paradigm. So first of all, as you can [12:03] imagine, it will create a much wider [12:06] patient capture funnel [12:08] um which will in turn improve patient [12:10] access. [12:12] Second, this will allow us to perform [12:14] more precise upstream triage which will [12:17] lead to more efficient utilization of [12:19] healthcare resources downstream and [12:22] ultimately um hopefully this will enable [12:25] early detection and intervention which [12:27] will lead to improved long-term patient [12:29] outcome. [12:32] Despite the excitement around the [12:34] science that I talked about, there's [12:35] still a lot of work ahead of us. So [12:38] first we need to figure out and codify [12:40] the optimal referral pathway and [12:42] workflow that just talked about which [12:44] would likely be specific to the medical [12:47] condition we are considering and also [12:49] the location. [12:51] Second as a society we need to come to [12:54] some sort of consensus in terms of who's [12:57] going to pay for this why and how much. [13:01] And third, as you can imagine, there are [13:03] a lot of stakeholders involved in [13:04] healthcare overall and definitely in [13:07] this new paradigm of healthcare [13:08] delivery. So I think it's critical from [13:11] day one that we align incentives of all [13:14] the stakeholder stakeholders involved [13:16] and some examples of these stakeholders [13:18] will be patient, primary care doctor, [13:22] specialist doctors, um software company, [13:26] hardware companies and last but not [13:28] least medical insurance payers. [13:32] So as a summary, [13:34] welcome to the age of AI based alomics. [13:38] Now by combining cutting edge AI [13:40] technique and retinal images, we can [13:42] really predict within minutes whether [13:44] someone is at risk for having dementia [13:48] like or future heart attacks or kidney [13:50] issues. So the future is here. This is [13:52] no longer science fiction. [13:55] However, there's still a lot of work [13:57] ahead of us. Specifically, we need to [13:59] figure out the optimal uh operational um [14:04] best practices. We need to align the [14:07] incentives of many stakeholders and [14:10] equally importantly we will need to [14:12] refine and develop novel business plans [14:14] to make this all work. [14:17] I hope that in the short time that we [14:19] had together, I was able to provide for [14:21] you a glimpse into an exciting future, [14:25] one that hopefully will change medicine [14:27] as we know it for the better. Thank you. [14:30] And that's all I have.