Oura CMO Says Commodity AI Models Shift Value to Hardware Ecosystems

Oura首席医疗官称商品化AI模型正将价值转向硬件生态 Oura首席医疗官Ricky Bloomfield预测，基础AI模型将进入平台期并沦为本地商品化工具，这使得高质量传感器硬件与用户参与度成为预防性医疗的核心驱动力。 Ricky Bloomfield博士阐释了通过ŌURA Ring而非传统腕戴设备追踪生物特征指标在工程与生理学上的优势，强调了其全天候可穿戴性以及紧邻动脉血流的特点。他详细介绍了ŌURA依托内部32位博士团队建立的严苛临床验证框架，以及与Medicare Advantage等支付方的整合，以减少再入院率并降低成本。讨论还重点提及了AI驱动的ŌURA Advisor、通过整合CGM实现的上下文工程等关键创新，并阐明了为何用户参与度对于推动医学从被动救治转向主动、预防性的健康寿命延长至关重要。 ŌURA 创立约 13 年，最初专攻睡眠监测并采用戒指形态，其光学传感器距手指动脉仅 1 至 2 毫米，配合长达一周的续航能捕获极其精准的夜间生理基线。 (0:04) ŌURA Ring 搭载 50Hz 双色 PPG 传感器、NTC 温度传感器与加速度计，可追踪逾 50 项生理指标，旨在推动医疗从被动治疗转向具备早期预警的主动预防体系。 (4:17) 由 32 位博士组成的研发团队正在评估约 95 项检测与预测算法概念，并通过与联邦医疗保险优势计划 Essence Healthcare 合作，验证连续监测降低再入院率与医疗支出的效果。 (7:28) 借助基于生成式 AI 的 ŌURA Advisor 及与 Dexcom 血糖仪的整合，ŌURA 强调以高质量生物特征数据为核心的上下文工程（Context Engineering），为用户提供深度个性化健康指导。 (15:38)

ŌURA Chief Medical Officer Ricky Bloomfield argues that continuous, high-fidelity physiological data from ring-based sensors provides the foundation needed to shift medicine from reactive treatment to proactive prevention. He contends that as artificial intelligence models become commoditized, competitive differentiation will depend on reliable sensor context, clinical validation, and sustained user engagement. Physiological Advantages of Finger-Based Biometric Sensing (0:04) ŌURA was founded roughly 13 years ago to address sleep tracking, choosing a screenless ring form factor to maximize comfort and avoid night-time disruptions. Battery life has expanded from two to three days in early iterations to seven days or more. Bloomfield explains that sleep offers an optimal baseline for measuring heart rate and heart rate variability (HRV) because ambient conditions remain stable and the body is still. The hardware architecture combines dual-color photoplethysmography (PPG) sensors flashing red and green lights 50 times per second with negative temperature coefficient (NTC) thermistors and multi-axis accelerometers. According to Bloomfield, the finger offers a distinct anatomical advantage over the wrist: optical sensors sit within 1 to 2 millimeters of the palmar digital arteries, capturing cleaner arterial waveforms than wrist-worn devices that position sensors on top of the wrist. This sensor suite derives over 50 continuous biometrics, including blood oxygen saturation (SpO2), respiratory rate, and sleep architecture stages. Clinical Outcome Validation and Payor Adoption (7:28) Bloomfield emphasizes that translating continuous biometric signals into preventive tools requires formal clinical validation rather than consumer novelty. Supported by an in-house team of 32 PhD researchers, ŌURA maintains an internal backlog of approximately 95 "detect or predict" algorithmic concepts under scientific review. Bloomfield notes that while many consumer wearable companies bypass formal clinical trials due to expense or the risk of unfavorable results, ŌURA established a dedicated Director of Clinical Outcomes Research to quantify real-world medical efficacy. This evidence-based approach underpins commercial partnerships with institutional health plans. For example, Essence Healthcare, a Midwest Medicare Advantage plan, covers the ring as a benefit to evaluate whether continuous biometric monitoring can reduce hospital readmissions and decrease healthcare expenditures. Bloomfield states that payers such as Aetna are attracted by user engagement metrics uncommon in clinical tools, including an average daily wear time of roughly 23.5 hours and companion application access frequencies of two to three times per day. Kinetic Edge Cases in Movement Detection (10:47) Despite advances in physiological sensing, fine-motor kinetics present persistent engineering hurdles for finger-worn devices. Bloomfield acknowledges that interpreting motion solely through finger-based accelerometry can lead to classification errors. Routine household tasks involving complex hand movements, such as folding laundry or preparing meals, are occasionally misidentified by automated algorithms as structured sports or workouts. Similarly, tracking swimming mechanics remains difficult because finger movement produces complex hydrodynamic motion artifacts, contrasting with the relative stability of head-mounted devices like smart goggles. To minimize user friction across varied lifestyles, Bloomfield notes that ŌURA engineers its core biometric algorithms to maintain data reliability regardless of which hand or specific finger the user selects. Context Engineering Over AI Model Differentiation (15:38) As ŌURA integrates generative artificial intelligence through its Oura Advisor tool, Bloomfield contends that the industry focus on prompt engineering is shifting toward "context engineering." Because large language models (LLMs) depend entirely on data quality, feeding high-precision, peer-reviewed sensor inputs into models matters far more than phrasing prompts. Generative AI allows the system to move beyond rigid, legacy decision trees to deliver contextual health advice based on longitudinal biometrics. To enrich this contextual picture, ŌURA integrates external data streams, such as Dexcom’s Stelo over-the-counter continuous glucose monitor. Cross-referencing metabolic readings with overnight vitals illustrates practical health patterns; Bloomfield observed during personal testing that late-night dining elevated sleeping heart rate and worsened blood glucose time-in-range metrics. Looking ahead five to ten years, Bloomfield forecasts that foundational LLMs will commoditize, plateau in capability, and run locally on consumer devices, making proprietary hardware ecosystems and user engagement the primary sources of value. Bloomfield maintains that shifting healthcare from a reactive "break-fix" model to proactive management requires pairing continuous 24/7 biometric tracking with validated early-warning algorithms. However, the broader clinical success of wearable ecosystems will ultimately depend on proving that continuous data collection translates into sustained behavioral change and reduced systemic healthcare costs across commercial payor populations.

ŌURA 首席医疗官 Ricky Bloomfield 医学博士指出，依托指环形态获取高精度连续生理数据，并结合临床结果验证与上下文工程，能够推动医疗体系从被动的“故障维修”模式转向以延长健康寿命为核心的预防性干预模式。 指环形态的生理监测优势与传感器架构 (0:04) ŌURA 创立约 13 年前最初专注于解决睡眠追踪问题。由于睡眠监测需要无屏幕干扰、无强光通知且无需夜间充电的设备形态，指环成为优于腕表或动态心电图（Holter）监测仪的选择；其单次充电续航已从早期的 2 至 3 天提升至现代机型的 7 天以上。夜间人体处于静止且环境稳定的状态，能够为心率、心率变异性（HRV）以及长期健康基线提供高质量的低噪声信号。在解剖学层面，手指掌侧的两条指固有动脉距离指环光学传感器仅有 1 至 2 毫米，相比佩戴于手腕背部的传统手表传感器，能够更直接地捕获动脉血流信号。 在硬件配置上，设备采用以 50 Hz（每秒 50 次）频率交替闪烁的红光与绿光光电容积脉搏波（PPG）传感器，用以测量心率、HRV 和血氧饱和度（SpO2），并辅以负温度系数（NTC）热敏电阻监测皮肤温度波动，以及多轴加速度计捕获人体运动。通过提取单次心跳的完整波形并进行多周至数月的长期趋势分析，设备衍生的健康监测指标已从早期的 40 项扩展至 50 余项，覆盖呼吸频率、呼吸扰动与睡眠结构分期等维度，为构建连续的早期预警系统提供底层生理数据支撑。 临床结果验证与医疗支付方整合 (7:28) 针对部分消费级可穿戴设备因研发成本高昂或担忧负面数据而规避正式临床试验的行业现状，Bloomfield 强调 ŌURA 坚持以科研为基础，组建了包含 32 位博士的研究团队，并专设临床结果研究总监一职，以确保各项功能具备可验证的行为干预和临床价值。目前，ŌURA 产品团队内部维护着一个包含约 95 项基于现有传感器硬件的“检测或预测”算法候选库，这些算法必须经过严格的临床与科学验证后才会正式推向用户端。 在医疗支付方合作方面，ŌURA 正在探索将可穿戴监测接入公共医保体系。公司已与美国中西部的联邦医疗保险优势计划（Medicare Advantage）Essence Healthcare 达成合作，由该保险计划为参保受益人全额提供指环福利。该项目的核心目标是通过 24 小时不间断的生物识别数据追踪，评估连续体征监测在减少老年患者非计划性再入院率、实现早期医疗干预以及降低系统性医疗总支出方面的实际效益。 运动算法边界与多源数据上下文整合 (10:47) 在日常动作识别的算法工程中，基于单根手指的动能信号处理存在客观局限。Bloomfield 指出，由于日常动作的多样性，算法有时会将折叠衣服或烹饪等精细手部动作误分类为体育锻炼，而像游泳这类涉及复杂流体力学水流干扰的运动项目，手指端的精准划水与趟数统计依然面临算法挑战；不过，设备算法在设计与验证中已做到对手指选择与左右利手无关。为解决用户“如何根据数据采取行动”的痛点，ŌURA 推出了基于生成式 AI 的健康顾问功能（Oura Advisor），取代了过去在复杂个性化场景中容易失效的刚性规则决策树。 Bloomfield 认为，当前 AI 应用的重心正从“提示词工程”转向“上下文工程”。大语言模型（LLM）的输出严格遵循“输入劣质、输出劣质”的原则，因此喂入模型的输入数据精度远比提示词表述更为关键。为建立全面的个人健康全景视图，ŌURA 开始接入第三方生物传感器数据，例如集成 Dexcom 的 Stelo 非处方连续血糖监测仪（CGM）。根据 Bloomfield 分享的个人测试案例，度假期间的夜间进食会导致血糖处于达标时间（time-in-range）显著缩短，并因持续消化作用导致夜间静息心率长时间处于偏高状态，证实了多模态生理数据交叉分析对行为干预的指导价值。 模型商品化趋势与主动预防医疗愿景 (18:48) 对于未来五至十年的技术演进，Bloomfield 提出基础大模型将迅速走向标准化与商品化。随着大语言模型的参数暴力缩放出现边际效用递减并进入性能瓶颈期，模型将逐步轻量化并以足够的性能直接在消费级端侧设备本地运行。因此，未来的核心商业壁垒将不再是底层的 LLM 本身，而是围绕模型构建的软硬件一体化生态系统以及用户的实际依从度。 实现以较低成本延长“健康寿命（healthspan）”而非仅仅延长“预期寿命（lifespan）”，依赖于具备高黏性的硬件交互触点与经过验证的预警算法。目前 ŌURA 用户平均每天佩戴指环约 23.5 小时，移动端应用每日打开频次为 2 至 3 次，这种高频交互构成了健康险公司（例如 Aetna）开展战略合作的核心考量。Bloomfield 表示，该技术路线的终极目标是通过全天候生理监测提供类似于汽车“发动机检查灯”的守护功能，推动现代医学从被动的“生病救治”体系转型为主动的预防性医疗模式。 ŌURA 展示了通过高黏性硬件采集连续指端动脉数据、并借助上下文工程与商业保险生态落地预防性医疗的路径；然而，单指动能传感器在复杂日常活动自动分类上的固有算法噪声，以及大模型向端侧迁移过程中多源临床数据的实际转化效果，仍是有待长期观察与验证的核心环节。

## Background 本视频由 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台上。内容为《时代》周刊（TIME）资深健康记者 Alice Park 与 ŌURA 首席医疗官 Ricky Bloomfield 博士的炉边对话，重点探讨人工智能如何赋能可穿戴硬件与软件创新，以及如何通过挖掘高精度健康数据推动医疗体系向主动预防转型。 ## TL;DR 对话深入探讨了 Ōura Ring 依托手指解剖优势与全天候无感佩戴采集高信噪比生理数据的底层逻辑，强调了公司坚持硬核临床结局验证、拒绝行业浮躁“鸵鸟策略”的研发哲学。Ricky 阐述了生成式 AI 从“提示词工程”向“上下文工程”的演变，指出高质量连续生理数据是打破传统脆弱规则树、实现超级个性化健康指导的关键。最后，他提出基础大模型将迅速商品化，未来的核心护城河在于融合硬件、软件与严谨算法的高粘性生态，以此推动用户持久的行为改变，重塑从被动救治走向 24/7 主动预防的医疗新范式。 ## Atmosphere & style 现场氛围轻松热烈且充满专业思辨，主持人与嘉宾之间互动默契、妙语连珠，开场的幽默调侃迅速拉近了与观众的距离。Ricky 兼具资深临床医生与硅谷顶尖科技从业者的双重视角，既有对严谨科学证据的恪守，又有对颠覆传统医疗模式的澎湃激情与“辛辣”洞见，展现出极高的公信力与前瞻性感染力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=A8g6AHo5agA&t=4s) ŌURA 选择戒指形态的核心初衷是优化夜间睡眠监测，通过无屏幕防打扰设计与一周以上的长续航，消除频繁充电和光线对睡眠的干扰。 - [02:41](https://www.youtube.com/watch?v=A8g6AHo5agA&t=161s) 从人体解剖学来看，手指微小动脉距离戒指传感器仅 1 至 2 毫米，夜间静止状态可提供极高信噪比的原生生理信号，在测量基线上优于传统手腕外侧设备。 - [04:16](https://www.youtube.com/watch?v=A8g6AHo5agA&t=256s) 戒指内部集成了 50Hz PPG、高精度温度传感器及加速度计等多模态组件，监测 50 余项指标，充当人体全天候的“发动机故障预警指示灯”。 - [06:30](https://www.youtube.com/watch?v=A8g6AHo5agA&t=390s) ŌURA 坚持基于真实临床结局（Health Outcomes）的研发哲学，拒绝行业常见的“鸵鸟策略”，若验证某项功能对健康无实质改善即果断叫停。 - [09:40](https://www.youtube.com/watch?v=A8g6AHo5agA&t=580s) 公司设立首位临床成果研究总监，并与 Medicare Advantage 机构（Essence）深度合作，将智能戒指纳入全额医保福利以降低再入院率。 - [10:47](https://www.youtube.com/watch?v=A8g6AHo5agA&t=647s) 单指活动识别在工程上极具挑战，需从微小运动轨迹中过滤日常琐碎动作以分类庞大的活动类型，算法支持任意手指佩戴且保持高信号质量。 - [13:04](https://www.youtube.com/watch?v=A8g6AHo5agA&t=784s) 游泳等复杂运动由于手臂在水中剧烈运动导致信号嘈杂，精准识别泳姿、打腿分解练习和圈数仍是手部穿戴设备持续攻关的技术难点。 - [14:26](https://www.youtube.com/watch?v=A8g6AHo5agA&t=866s) AI 健康建议的演进正从“提示词工程”转向“上下文工程”，LLM 的推理价值完全依赖于底层高精度、连续生理数据的输入质量。 - [16:20](https://www.youtube.com/watch?v=A8g6AHo5agA&t=980s) 结合 Dexcom Stelo CGM 血糖监测仪的多维数据验证了深夜进餐会显著拉高夜间心率并加剧血糖波动，体现了跨设备生物标志物融合的临床价值。 - [19:34](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1174s) 基础大模型（LLM）正迅速走向通用大宗商品化，未来的核心壁垒在于将硬件载体、高品质软件体验与临床验证算法紧密结合的一体化生态。 - [20:38](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1238s) ŌURA 用户日均佩戴 23.5 小时的高参与度为驱动长期行为改变奠定基础，助力医疗体系从被动“坏了才修”转向主动预防并提升健康寿命。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=A8g6AHo5agA&t=4s) 现场互动与选择戒指形态的初衷 对话在轻松的现场互动中展开。主持人 Alice Park 首先邀请台下的听众举手示意有多少人佩戴了 Ōura Ring，面对台下的反应，Ricky Bloomfield 幽默地回应“不予置评（No comment）”。随后 Alice 进一步追问，在佩戴戒指的人群中，有多少人真正根据戒指提供的数据改变了自己的睡眠、饮食或运动等健康行为。看到现场不少人举手，Alice 笑着对 Ricky 说这对 Ōura 而言是个绝佳的肯定。 随后讨论切入正题：在智能手环、智能手表乃至医用动态心电图仪（Holter monitor）等各类健康监测设备林立的市场中，Ōura 为何坚定选择“戒指”这一产品形态？Ricky Bloomfield 解释道，Ōura 大约在 13 年前创立，最初的核心聚焦点就是睡眠。戒指形态并非偶然，而是为了夜间佩戴体验做出的针对性设计：它体积小巧、贴合无感，更不会像带有屏幕的设备那样在夜间突然亮起刺眼的强光打扰睡眠。此外，电池续航是夜间监测设备的关键命脉——如果设备每隔 12 小时或一天就必须充一次电，用户被迫在夜间摘下充电，就彻底违背了夜间连续追踪的初衷。早期 Ōura 戒指的单次充电续航为 2 至 3 天，而如今已提升至一周以上，真正做到了让用户在入睡时“彻底忘记它的存在”。 --- ### [02:41](https://www.youtube.com/watch?v=A8g6AHo5agA&t=161s) 夜间生理信号与手指解剖结构的测量优势 除了极致的佩戴舒适度，选择戒指形态更有着深层的生理学与信号采集考量。Ricky 指出，夜间实际上是测量几乎所有人体生理指标的“黄金窗口期”。当人体处于静止且环境高度恒定的睡眠状态时，诸如心率、心率变异性（HRV）等长期生理指标能够摆脱日间运动和环境噪音的干扰，输出极为干净的基线信号。这为研发团队利用机器学习（ML）与人工智能（AI）训练并验证传感器算法提供了坚实的数据基础。 ```mermaid flowchart TD accTitle: 戒指与手表在生理信号采集上的解剖学对比 subgraph RingMechanism["Ōura 戒指测量路径"] A["手指侧下方固有动脉"] -->|"间距仅 1-2 mm"| B["高信噪比生理数据"] B --> C["夜间静止基线 / 纯净信号"] end subgraph WristMechanism["传统手表测量路径"] D["手腕内侧主动脉"] -->|"表盘佩戴于手腕外侧"| E["距离远 / 存在空间偏置"] E --> F["需复杂算法补偿运动干扰"] end ``` 从人体解剖学结构来看，手指两侧和底部走行着两条微小动脉，Ōura 戒指内侧的传感器距离这些动脉仅有 1 至 2 毫米的物理距离，能捕获极高质量的脉搏波形。相比之下，手腕处的主要动脉虽粗，但大多位于手腕内侧，而数百年来的钟表佩戴习惯促使人们将手表的传感主体置于手腕外侧。尽管智能手表通过算法迭代在一定程度上弥补了这种生理距离，但在生物传感领域，“没有任何算法能完全替代更近距离、更纯净的原生信号”。 --- ### [04:16](https://www.youtube.com/watch?v=A8g6AHo5agA&t=256s) 传感器架构与全天候主动健康“故障灯”机制 针对目前设备所能采集的指标维度，Ricky 提到官方口径已从最初的 40 余项扩展到 50 余项，且内部数据还在持续高速增长。戒指内部集成了微型光电容积脉搏波描记（PPG）传感器，以每秒 50 次的频率发射红光与绿光脉冲来监测心率并测算血氧饱和度（SpO2）；同时搭载了高精度温度传感器和加速度计。通过这些多模态传感器的协同运作，系统不仅能推导心率、心率变异性（HRV）、细分睡眠分期，还能精准识别夜间呼吸紊乱事件及实时呼吸率。 - **多模态传感器套件**：红绿双色 PPG 传感器（50Hz 采样率）、高精度温度传感器、三轴加速度计。 - **核心监测维度**：静态心率、HRV 趋势、血氧饱和度（SpO2）、睡眠阶段分期、夜间呼吸紊乱指数、体表温度波动。 - **波形信息密度**：单次心跳的脉搏波形蕴含着庞大的心血管生理信息，跨越数天、数周乃至数月的连续波形趋势分析能挖掘出更深层的健康规律。 Ricky 强调，这类可穿戴设备堪称用户所能拥有的“最亲密的硬件设备（the most intimate piece of hardware you can have）”，因为它 24 小时紧密贴合皮肤。传统医疗体系多属于“故障报修式（break-fix）”的被动响应机制，患者只有在生病后才去医院；而要实现从被动治疗向主动预防的范式转变，24/7 连续、高质的数据流是不可或缺的底层支撑，它能够充当人体健康的“发动机故障指示灯（check engine lights）”或全天候的“守护天使（guardian angel）”。 --- ### [06:30](https://www.youtube.com/watch?v=A8g6AHo5agA&t=390s) 医疗级验证体系与拒绝“鸵鸟策略”的研发哲学 当 Alice 谈及如何将庞大的原始数据提炼为可指导行动的“信息”，并询问首席医疗官（CMO）如何在未被满足的医疗需求（如老人跌倒监测、感染早期预警）中布局时，Ricky 明确表示，Ōura 是一家完全建立在科学基石上的公司，团队目前拥有约 32 位博士（PhDs），任何创新都是跨学科团队协作的结晶。仅在产品团队规划的“检测与预测（detect or predict）”功能储备库中，基于现有硬件技术可行且待推进的构想就有大约 95 项。 ```mermaid flowchart LR accTitle: Ōura 临床验证与功能落地流程 subgraph Pipeline["研发决策闭环"] A["95 项功能构想池"] --> B["严苛科学临床试验"] B --> C{"是否带来真实健康改善?"} C -- "验证有效" --> D["产品化推送并加大投入"] C -- "无效 / 假象" --> E["果断叫停研发"] end ``` Ricky 将自己作为 CMO 的首要使命总结为“在公司内外建立信任与公信力”，推动 Ōura 从泛健康管理（Wellness）纵深迈向严肃医疗健康（Healthcare）领域。他直言不讳地指出，行业内许多可穿戴设备公司习惯采取“把头埋在沙子里的鸵鸟策略（stick your head in the sand approach）”，只顾砸钱做产品功能包装，却害怕投入时间与资金进行严格的临床研究，唯恐试验结果证明自己的设备无效。Ōura 的研发原则则是必须对真实健康结局（Health Outcomes）进行硬核测量： > “如果验证表明某项功能确实改变了用户行为并改善了具体健康指标，我们就全力加大投入；但如果测量发现它对健康无益，我们就会立刻叫停。向用户提供一种‘自以为有用但实际无效’的干预措施，其危害甚至可能大过什么都不做。” ### [09:40](https://www.youtube.com/watch?v=A8g6AHo5agA&t=580s) 设立临床成果研究与医疗保险合作 ŌURA 正在全力确保其推出的每一项功能都能带来实质性的临床意义并切实改善健康结局。就在本周（访谈当天为周二，即周一），ŌURA 迎来了公司首位临床成果研究总监（Director of Clinical Outcomes Research）入职。公司正重金投入这一领域，以确保能够准确捕捉并衡量关键的健康指标。在商业落地方面，ŌURA 与美国中西部的 Medicare Advantage（联邦医疗保险优势计划）机构 Essence 展开了深度合作。Essence 将 Oura Ring 作为一项全额覆盖的医保福利直接发放给其所有参保受益人，并与 ŌURA 共同探索这些生理指标如何帮助改善用户健康、降低医疗开支、预防患者再入院，以及在疾病恶化前实现早期筛查与干预。 Ricky Bloomfield 动情地表示，作为一名临床医生，这种预防与干预前景让他倍感振奋：“你不会只想卖一个闪闪发光的小玩意儿（shiny trinket），而是真正想在医疗健康领域带来实质性改变（move the needle）。”这也正是他在苹果公司（Apple）工作了八年半后，毅然选择加入 ŌURA 的核心原因，他坚信现在的 ŌURA 拥有推动医疗变革的绝佳契机。 - **高管变动**：公司新设立并入职了首位“临床成果研究总监”，全面发力健康结果的量化验证。 - **医保合作模式**：Essence Healthcare（中西部 Medicare Advantage 计划）将智能戒指列为全额报销福利，重点探索降低再入院率与早期预警。 - **职业转型**：Ricky Bloomfield 在苹果深耕 8.5 年后加入 ŌURA，致力于将消费级穿戴转化为真正的医疗干预工具。 --- ### [10:47](https://www.youtube.com/watch?v=A8g6AHo5agA&t=647s) 算法验证与单指活动识别的工程挑战 当主持人 Alice Park 问及是否存在某些原本寄予厚望但最终效果不佳的“负面/意外结果”时，Ricky 幽默地回答说他倒希望手头能有一个失败案例拿来分享，因为那样故事会更具戏剧性。但现实是，ŌURA 仍处于相对早期的阶段，且算法研发团队在功能正式上线前进行了极其严苛的临床与数据验证，因此并未出现彻底失效的重大翻车事故。 不过，早期在“活动检测”（Activity Detection）——即识别用户正在进行的具体运动类型——方面确实遇到过技术痛点。例如许多老用户曾经历过戒指将叠衣服、做饭等日常琐碎动作误判为某种剧烈运动。Ricky 指出，仅凭一根手指的运动轨迹去分类近乎无限种人类活动，在工程上是一个极难解决的问题。针对佩戴在不同手指或利手（惯用手 vs. 非惯用手）是否会影响数据质量的问题，Ricky 表示 ŌURA 在算法设计时默认用户可能佩戴在任意手指上，且不会强制询问是否为主导手；经过团队的对比验证，戒指在不同佩戴位置上均未出现明显的信号质量缺陷。 ```mermaid flowchart TD accTitle: Single Finger Activity Detection Challenges subgraph Input["数据输入与约束"] A["单指微小运动信号\n(One-finger kinematics)"] B["任意手指佩戴 / 惯用手不限\n(Non-dominant / Dominant Hand)"] end subgraph Challenge["工程识别难点"] C["日常微动作干扰\n(如叠衣服、做饭备餐)"] D["近乎无限的活动类别\n(Infinite Action Space)"] end subgraph Output["算法目标与演进"] E["高精度运动类型判定\n(精确识别运动与时长)"] F["扩大用户样本与包容性优化\n(More inclusive & nuanced)"] end A --> C B --> C C --> D D --> E E --> F ``` --- ### [13:04](https://www.youtube.com/watch?v=A8g6AHo5agA&t=784s) 愿望清单与游泳监测的软硬件限制 在谈及内部包含 95 个待办项目的“功能愿望清单”时，Alice 询问是否有某项功能让他个人格外兴奋。作为一名每天清晨坚持游泳锻炼的爱好者，Ricky 提到了“游泳检测”（Swimming Detection）。他解释道，游泳检测绝非单纯判断“是否在游泳”的二元问题，而是极其复杂的子任务集合：识别具体泳姿（自由泳、蛙泳等）、辨识打腿训练（kicking）或分解练习（drills），以及精准统计泳道圈数。 Alice 随即共鸣道，市面上现有的许多穿戴设备在游泳监测上“表现并不尽如人意”。Ricky 坦言赞同，并透露自己目前佩戴 Form 智能泳镜进行训练（泳镜内置 AR 抬头显示器）。他分析指出，泳镜之所以准确是因为头部在水中的运动轨迹远比手臂稳定，而手腕或手指设备会随着手臂在水中剧烈划水翻飞，导致信号极为嘈杂。虽然 ŌURA 当前聚焦于手部形态并未涉足头部硬件，但愿望清单中还包含诸多更为激动人心、目前尚处于保密阶段的健康监测能力。 --- ### [14:26](https://www.youtube.com/watch?v=A8g6AHo5agA&t=866s) ŌURA Advisor、上下文工程与跨设备血糖闭环 随着今年早些时候 ŌURA Advisor（一款常驻口袋的 AI 健康顾问）的推出，Alice 引导话题转向生成式 AI 如何帮助用户将数据转化为可执行的行动。Ricky 指出，用户最常提出的问题是：“面对这些生理数据，我个人究竟该如何针对性改善？”在生成式 AI 出现前，行业依赖由海量规则构成的复杂“决策树”，这种方式极其脆弱（brittle），极易漏掉未预料到的长尾边缘场景。 虽然 ŌURA 在底层算法中应用传统机器学习/AI 已有十余年，但当前大语言模型（LLM）的核心价值在于实现真正的“超级个性化”。在合规层面，Ricky 幽默地补充道：“我不会把它称作‘医生’，FDA 大概也绝不会允许我们叫它医生。”针对当下 AI 技术的演进路径，Ricky 提出了从“提示词工程”向“上下文工程”的关键转变： > “过去 18 个月左右，提升 AI 推理输出的流行词一直是‘提示词工程’（Prompt Engineering），但最近大家开始转向讨论‘上下文工程’（Context Engineering）——这意味着，向 LLM 喂入高质量的数据，其重要性远远超过你如何去雕琢 Prompt。这就是‘垃圾进，垃圾出’（Garbage in, garbage out）的道理。如果输入端具备极高精度的传感器数据，AI 推理输出的质量就会成倍提升。” 为了构建全方位的 360 度个人健康画像，ŌURA 正积极引入外部生物标志物数据。Ricky 分享了他体验 Dexcom Stelo 非处方持续葡萄糖监测仪（CGM）的亲身经历：他恰好在出发参加为期两周的邮轮旅行当天戴上了该设备。邮轮上的饮食节奏被打乱，监测数据生动揭示了深夜进食对糖代谢与身体负荷的破坏——由于夜间持续消化食物，心率持续偏高且下降缓慢，葡萄糖在目标范围外的时间（Time out of range）大幅恶化；而当他回到家恢复规律作息、确保睡前数小时不再进食后，指标立刻得到了显著改善。这一案例强力证明了将戒指传感器与外部数据在安全私密前提下进行多维融合的巨大临床价值。 - **AI 范式转移**：从僵硬脆弱的传统决策树规则，演进为基于 LLM 的高自由度个性化对话建议。 - **Context Engineering 优于 Prompt Engineering**：底层硬件采集的连续生物数据精度，直接决定了生成式 AI 建议的可靠性与针对性。 - **Dexcom Stelo CGM 联动实证**：深夜进餐不仅导致血糖波动脱轨（Time out of range 增加），还会使夜间静息心率长时间维持高位，直观印证消化对睡眠与代谢的双重负担。 ### [18:38](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1118s) 展望未来：Agentic AI 与医疗不平等的弥合 访谈临近尾声，主持人 Alice Park 提出了最后一个展望未来的宏观问题。她指出，当前医疗体系中存在着巨大的缺口与资源分配不平等，许多偏远或特定群体因就医渠道受限，与正规医疗系统几乎处于脱节状态。面对这些挑战以及 Agentic AI（智能体 AI）技术的兴起，Alice 询问 Ricky Bloomfield：在与医疗服务提供方（providers）展开深度合作的背景下，ŌURA 平台在未来 5 到 10 年的“蓝天愿景”（blue sky vision）究竟是什么？ ### [19:34](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1174s) 基础模型商品化与软硬件生态系统的核心价值 Ricky 在给出回答前幽默地预警，在一个 AI 峰会上讲出这番话可能会显得“有点辛辣”（spicy），Alice 则笑称“辛辣才好”。Ricky 直言，基础大模型（Foundation Models）在不久的将来很快就会变成通用的大宗商品（commodities）。他认为单靠暴力堆算力（brute force）提升模型能力的路径正在触及瓶颈期（plateau），而在不久的将来，这些足够优秀的模型都将直接部署在用户的本地设备端运行。 因此，未来的核心壁垒绝不在于底层模型本身，而在于整体生态系统。医疗健康最核心的目标是推动“行为改变”（behavior change）——即以更低的成本帮助人们保持健康，并切实提升“健康寿命”（healthspan，即处于健康状态的生命周期），而不仅是单纯延长生理寿命（lifespan）。Ricky 强调，孤立的 LLM 根本无法独立促成这种改变，它必须依托于一个能够持续吸引用户高频参与的完整平台：将有形的硬件载体、高品质的软件交互体验、经过严格临床验证的算法，以及类似汽车“发动机故障灯”（check engine light）般直观的预警功能紧密结合在一起，这种软硬件与算法的一体化体验是极难被复制的。 ```mermaid flowchart TD accTitle: 从被动应对到主动预防的健康生态闭环 subgraph WearablePlatform["软硬件一体化生态"] A["全天候佩戴智能硬件<br/>(Ōura Ring 每日佩戴 23.5 小时)"] --> B["高品质软件交互与<br/>临床验证算法"] end subgraph BehaviorChange["高粘性参与与预警机制"] B --> C["发动机故障灯式预警<br/>(Check Engine Light)"] C --> D["驱动持久行为改变<br/>(每日主动打开 App 2-3 次)"] end subgraph HealthcareShift["医疗模式根本转变"] D --> E["摆脱被动维修模式<br/>(Break-Fix Reactive System)"] E --> F["构建主动预防体系<br/>(Proactive & Preventive System)"] end ``` ### [20:38](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1238s) 从被动修复走向主动预防：高粘性驱动健康变革 谈及与健康险机构的合作时，Ricky 指出保险支付方极其看重用户参与度（engagement），因为唯有当参保人保持高度参与时，行为改变才会真正发生。ŌURA 展现出了远超其他可穿戴设备的用户粘性：用户平均每天佩戴戒指长达 23.5 个小时，且每天主动打开 App 达 2 至 3 次。这种“突破天际”（through the roof）的参与度，让保险机构看到了引导大众更积极管理自身健康的契机。 Ricky 总结了他对未来的核心论点：整个医疗体系必须摆脱传统的“坏了才修”（break-fix）的被动应对模式，加速迈向基于科学证据、全天候 24/7 守护的主动预防体系（proactive and preventative system）。Alice 对此表示高度赞同，认为通过持续提升人群的参与感与内在动力，医疗健康的天平必将实质性地向预防端倾斜。对谈在全场的热烈掌声中圆满结束。 - **核心观点与数据支持**： - **基础模型趋势**：LLM 正在迅速走向商品化，未来的差异化核心在于融合了硬件、算法与软件的高粘性生态平台。 - **用户参与度指标**：Ōura 用户日均佩戴时长达 23.5 小时，日均主动打开 App 2 至 3 次，为行为干预提供了极佳触点。 - **医疗范式转移**：从被动的“坏了再修”（break-fix）模式，转型为基于科学实证的 24/7 主动健康预防体系，核心目标是拓展公众的健康寿命（healthspan）。 ## Notable quotes > [02:41](https://www.youtube.com/watch?v=A8g6AHo5agA&t=161s) 没有任何算法能完全替代更近距离、更纯净的原生信号。 — Ricky Bloomfield > [04:16](https://www.youtube.com/watch?v=A8g6AHo5agA&t=256s) 这类可穿戴设备堪称用户所能拥有的“最亲密的硬件设备（the most intimate piece of hardware you can have）”，因为它 24 小时紧密贴合皮肤。 — Ricky Bloomfield > [06:30](https://www.youtube.com/watch?v=A8g6AHo5agA&t=390s) 如果验证表明某项功能确实改变了用户行为并改善了具体健康指标，我们就全力加大投入；但如果测量发现它对健康无益，我们就会立刻叫停。向用户提供一种‘自以为有用但实际无效’的干预措施，其危害甚至可能大过什么都不做。 — Ricky Bloomfield > [09:40](https://www.youtube.com/watch?v=A8g6AHo5agA&t=580s) 你不会只想卖一个闪闪发光的小玩意儿（shiny trinket），而是真正想在医疗健康领域带来实质性改变（move the needle）。 — Ricky Bloomfield > [14:26](https://www.youtube.com/watch?v=A8g6AHo5agA&t=866s) 我不会把它称作‘医生’，FDA 大概也绝不会允许我们叫它医生。 — Ricky Bloomfield > [14:26](https://www.youtube.com/watch?v=A8g6AHo5agA&t=866s) 过去 18 个月左右，提升 AI 推理输出的流行词一直是‘提示词工程’（Prompt Engineering），但最近大家开始转向讨论‘上下文工程’（Context Engineering）——这意味着，向 LLM 喂入高质量的数据，其重要性远远超过你如何去雕琢 Prompt。这就是‘垃圾进，垃圾出’（Garbage in, garbage out）的道理。如果输入端具备极高精度的传感器数据，AI 推理输出的质量就会成倍提升。 — Ricky Bloomfield > [19:34](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1174s) 基础大模型（Foundation Models）在不久的将来很快就会变成通用的大宗商品（commodities）。单靠暴力堆算力（brute force）提升模型能力的路径正在触及瓶颈期（plateau）。 — Ricky Bloomfield > [20:38](https://www.youtube.com/watch?v=A8g6AHo5agA&t=1238s) 整个医疗体系必须摆脱传统的‘坏了才修’（break-fix）的被动应对模式，加速迈向基于科学证据、全天候 24/7 守护的主动预防体系（proactive and preventative system）。 — Ricky Bloomfield ## Mentioned Ricky Bloomfield, Alice Park, ŌURA, Ōura Ring, ŌURA Advisor, TIME, Apple, Essence Healthcare, Medicare Advantage, Dexcom Stelo, Form 智能泳镜, FDA ## Video description Ricky Bloomfield, Chief Medical Officer, ŌURA Alice Park, Senior Health Correspondent, TIME Abstracts Artificial Intelligence is driving the development of innovative wearable devices. With this approach, users are now more mindful of their health. This discussion will explore the role artificial intelligence plays in hardware and software features used to scale health tracking devices, and the future of AI’s role in healthcare. [Source](https://www.youtube.com/watch?v=A8g6AHo5agA)

[2:41] Ricky Bloomfield: 没有任何算法能完全替代更近距离、更纯净的原生信号。 [4:16] Ricky Bloomfield: 这类可穿戴设备堪称用户所能拥有的“最亲密的硬件设备（the most intimate piece of hardware you can have）”，因为它 24 小时紧密贴合皮肤。 [6:30] Ricky Bloomfield: 如果验证表明某项功能确实改变了用户行为并改善了具体健康指标，我们就全力加大投入；但如果测量发现它对健康无益，我们就会立刻叫停。向用户提供一种‘自以为有用但实际无效’的干预措施，其危害甚至可能大过什么都不做。 [9:40] Ricky Bloomfield: 你不会只想卖一个闪闪发光的小玩意儿（shiny trinket），而是真正想在医疗健康领域带来实质性改变（move the needle）。 [14:26] Ricky Bloomfield: 我不会把它称作‘医生’，FDA 大概也绝不会允许我们叫它医生。 [14:26] Ricky Bloomfield: 过去 18 个月左右，提升 AI 推理输出的流行词一直是‘提示词工程’（Prompt Engineering），但最近大家开始转向讨论‘上下文工程’（Context Engineering）——这意味着，向 LLM 喂入高质量的数据，其重要性远远超过你如何去雕琢 Prompt。这就是‘垃圾进，垃圾出’（Garbage in, garbage out）的道理。如果输入端具备极高精度的传感器数据，AI 推理输出的质量就会成倍提升。 [19:34] Ricky Bloomfield: 基础大模型（Foundation Models）在不久的将来很快就会变成通用的大宗商品（commodities）。单靠暴力堆算力（brute force）提升模型能力的路径正在触及瓶颈期（plateau）。 [20:38] Ricky Bloomfield: 整个医疗体系必须摆脱传统的‘坏了才修’（break-fix）的被动应对模式，加速迈向基于科学证据、全天候 24/7 守护的主动预防体系（proactive and preventative system）。

[4:17] Ricky Bloomfield: the most intimate piece of hardware you can have [4:17] Ricky Bloomfield: break-fix [4:17] Ricky Bloomfield: check engine lights [4:17] Ricky Bloomfield: guardian angel [7:28] Ricky Bloomfield: detect or predict [13:04] Alice Park: Oura Advisor [15:38] Ricky Bloomfield: context engineering [15:38] Ricky Bloomfield: garbage in, garbage out [18:48] Alice Park: blue sky

[0:04] And thank you for being here, Rick. I'm [0:05] looking forward to the discussion. [0:07] >> I'm excited. It's great to be here. [0:08] >> Yeah. Okay. Before we start, show of [0:11] hands. How many of you are wearing an [0:13] aura ring? [0:16] >> Okay. Well, we'll talk about that. [0:18] >> No comment. [0:20] >> We'll we'll talk about that. Um All [0:22] right. Keep your hands up. And among [0:25] those of you with your hands up, how [0:27] many of you have actually changed your [0:30] behavior, sleep, diet, exercise, [0:33] whatever it may be, based on what you've [0:35] seen from your ring? [0:37] >> Oh, good. Wow. [0:40] All right. Interesting. [0:42] >> Good news for you, right, Ricky? Okay. [0:45] Um, so I wanted to start just by laying [0:48] a little bit of groundwork and um [0:51] talking about the ring. There's so many [0:55] devices now that can, you know, track [0:57] various things, various aspects of our [0:59] health. So, um, walk me through a little [1:02] bit why the ring, um, and are there any [1:06] advantages to that platform versus [1:11] having something on your wrist or the, [1:14] you know, the granddaddy like the halter [1:16] monitor that would, you know, monitor [1:18] um, heart metrics? Um, are there things [1:21] that you can do on the ring? Things you [1:22] can pick up, um, measures you can pick [1:24] up that you can't with others? What are [1:26] the advantages of having the Ring? [1:28] >> Yeah, it's a great question. Um, Aura [1:30] got started, um, about 13 years ago with [1:33] a focus on sleep. Um, and so the Ring [1:35] form factor was chosen intentionally um, [1:39] as the device that would work best and [1:41] be the easiest for you when you're [1:43] trying to sleep. You want something that [1:44] is not bulky, gets out of your way, it's [1:47] not going to shine a, you know, a bright [1:48] light at you through a screen, those [1:50] types of things. And so that's why the [1:51] ring was chosen and that's why Aura was [1:53] sort of first in this area looking at [1:55] how do we create a better sleep tracking [1:57] device. Um, and another key point was [2:00] that the device needed to last um, more [2:03] than just 12 hours or a day on a charge [2:05] because if you have to charge your [2:06] device at night, it kind of, you know, [2:08] defeats the point of having a a device [2:10] you can wear at night. Um, and so the [2:12] initial ring, I think it lasted 2 to 3 [2:15] days on a charge. Um, now it's it's it's [2:18] up to a week or more on a charge. And [2:22] that's critical if you want something [2:23] that actually works at night. So that's [2:25] that's the that's the biggest reason I [2:26] think is that it's something that you [2:28] can wear. It gets out of your way. Um, [2:30] however, it's also important to note [2:32] that many people just don't want to wear [2:34] something bulky on their wrist at night. [2:35] And so the ring, you know, you forget [2:37] it's there. Um, and but it actually goes [2:41] a little bit further than that, which is [2:43] at night you actually don't just measure [2:46] your sleep. Um, the nighttime is [2:49] actually the best time to measure almost [2:51] anything when it comes to physiologic [2:54] metrics. Why? Because night is when [2:56] you're still and there's a consistent [2:59] environment. So, if you want to measure, [3:00] you know, heart rate, you want to look [3:01] at your HRV over time, you want to look [3:03] at longer term metrics, you just get a [3:05] much cleaner signal at night. And so [3:08] when you're doing the science to [3:10] actually validate the sensor, you know, [3:12] based algorithms that you're developing [3:14] using, you know, machine learning AI, of [3:16] course, um you just get a much better [3:18] signal. Um and then I think finally, [3:22] um for those that are, you know, [3:23] familiar with anatomy, you actually have [3:25] two little arteries that are running on [3:27] the bottom of your finger right here. So [3:30] the sensors on the ring are within like [3:33] one or two millimeters away from those [3:35] arteries. And so you actually get a [3:37] better signal because it's just closer [3:39] on your wrist. Of course, your major [3:41] arteries are actually on the bottom of [3:43] your wrist, but of course we wear [3:44] watches with the, you know, the big [3:46] part, the face of the watch on the top [3:48] just for convenience because it, you [3:49] know, over, you know, hundreds of years, [3:51] it's been used for time. Um, and so, you [3:54] know, watch algorithms have been [3:55] developed in a way to kind of overcome [3:57] that to some degree, but nothing can [4:00] replace a cleaner, closer signal, and [4:03] that's what the ring gives you. hadn't [4:05] really thought about that in terms of [4:07] the you know activity and so you're [4:09] really eliminating as many confounding [4:11] factors exactly right [4:12] >> as possible if you're if you're going to [4:14] study it. Um so correct me if I'm wrong [4:17] it now measures 40 metrics and can you [4:19] just sort of give us the broad [4:20] categories that you're looking at right [4:23] now that it can detect? [4:24] >> Sure. Yeah. there. I think I think our [4:26] official marketing um you know position [4:28] is is 50 plus now or 40 plus I think [4:30] we're not even sure internally there it [4:32] just keeps growing so quickly how many [4:34] things. Um so the the actual um you know [4:37] sensors on the device include your PPG [4:40] which is what um you know flashes little [4:42] light 50 times a second and um actually [4:44] checks your your heart rate. Um and so [4:46] we have two different colors of that [4:48] both red light and green light. Um and [4:51] so we can also measure your your um you [4:52] know blood oxygen with that. Then [4:54] there's also temperature. Um you have [4:56] accelerometer you know there as well. [4:58] And with with these sensors you know [5:01] combined you can derive many many [5:03] metrics from that. Of course heart rate [5:05] HRV all the sleep metrics all the [5:08] metrics you can derive from sleep [5:09] including sleep stages um breathing [5:12] disturbances at night when you're [5:13] sleeping. You can measure your [5:14] respiratory rate because of the [5:15] accelerometer and movement. It just goes [5:17] on and on and on. Um, and it's I think [5:21] we're still not close to sort of [5:23] finishing all the, you know, work that [5:25] you can do just with these sensors. Um, [5:27] and it's kind of mind-blowing when you [5:29] look at if you can reliably check the, [5:32] you know, waveform of your heartbeat, [5:35] how much information is embedded in that [5:37] waveform is kind of mind-blowing. And [5:39] that's just one waveform. when you look [5:41] at all of the waveforms over time, you [5:44] know, for days or weeks or months, when [5:46] you look at trends, you get even more [5:48] information. And and I think that's the [5:51] the interesting thing about a device [5:52] like this. You know, I say that um these [5:55] types of devices are the most intimate [5:57] piece of hardware you can have because [5:59] it's in contact with your skin 247. Um, [6:02] and if you want to move from this sort [6:04] of reactive break fix, you know, um, [6:06] health care system that we have today to [6:08] something more proactive, having that [6:10] 247 information about you is is key if [6:14] we want to create these features that [6:15] are, you know, these check engine lights [6:16] for your health or, you know, guardian [6:18] angel type features. Um, and you you [6:21] have to have that consistent [6:22] highquality, you know, data 24/7 which [6:25] you just don't get um with other devices [6:26] or with your interaction with a general [6:28] healthare system. [6:30] We're here to talk about how technology [6:32] like this, right, and and data and not [6:35] just raw data, but information, which to [6:38] me is different from data because [6:39] information is is something that's [6:41] useful. It's knowledge that you can act [6:42] on. [6:43] >> Um, yeah, that will change your [6:45] behavior. And we're talking about how [6:46] that can empower patients. um as chief [6:50] medical officer, how how does it work at [6:53] Aura in terms of are you looking at gaps [6:56] in terms of our understanding or [6:58] problems that we haven't been able to [7:00] solve? Um you know whether it's [7:03] monitoring elderly patients for falls or [7:06] monitoring for temperature changes for [7:09] signs of infection or things like that. [7:11] as CMO are you sort of contributing and [7:14] and thinking up ways you know that this [7:17] because it's it's brand new territory [7:18] right where where you can go and say can [7:21] we address this with our platform you [7:24] know can we solve these medical problems [7:26] health issues with our platform [7:28] >> right the first thing I would say is [7:30] that we the company was founded on on a [7:33] basis of science we have I think 32 PhDs [7:35] right now so I certainly don't want to [7:37] overstate my own importance to the [7:39] company we have an incredibly talented [7:41] ented team and everyone contributes to [7:44] you know coming up with these amazing [7:45] ideas. I think the last time I checked [7:47] with our product team around what we [7:50] call our detect or predict features [7:52] meaning this check engine light type [7:53] idea um our list of ideas that we wanted [7:57] to pursue was something like 95 [8:00] different things that we know we can do [8:01] with the current technology. Um it just [8:04] takes time because we're not going to [8:05] release a feature unless we've done the [8:07] science and we've validated the accuracy [8:09] of that feature. like that is that is [8:10] core to how we work. Um and that just [8:13] that takes time um in order to um to [8:16] build that. So I would say that if I [8:19] could summarize my role, it is building [8:21] trust and credibility um you know [8:24] through the company through engagements [8:26] like this through interaction with with [8:27] healthcare organizations as we build [8:29] those relationships as we you know move [8:31] meaningfully into not just wellness but [8:34] also you know health and health care. [8:36] Um, part of that trust and credibility [8:38] is actually making sure that we are [8:41] honest with ourselves and measuring um, [8:44] everything that we're doing, that we're [8:45] actually looking for changes to [8:48] meaningful health outcomes and that [8:50] we're investing in those studies that [8:52] show changes. Um, and this is something [8:54] that not all wearable companies do [8:56] because it takes time, it takes money. [8:58] Um many companies would rather just [8:59] invest in product development and they [9:02] might be afraid that if they study it [9:04] they would find that it's that it [9:05] doesn't work and so they don't they [9:07] don't want to know they it's sort of the [9:08] you know stick your head in the sand [9:09] type approach. But for us measuring [9:12] outcomes is so important because if we [9:15] can verify that something is actually [9:17] changing behavior and improving you know [9:19] concrete metrics in your health we want [9:21] to do more of that. If at the same time [9:24] we measure those things and find that it [9:26] is not helping or not improving health, [9:29] we want to stop doing that. Like giving [9:32] an intervention to someone that they [9:34] think is helping but isn't might [9:36] actually be worse than doing nothing at [9:38] all. So it's really important to you [9:40] know take a look inside and make sure [9:42] that we are releasing features that are [9:44] actually meaningful that are that are [9:46] changing you know outcomes. And so uh [9:48] this week actually um what is today? [9:50] Today's Tuesday I think. Um yesterday um [9:54] our first director of clinical outcomes [9:56] research started at the company. So we [9:58] hired this role um and we are investing [10:01] heavily in this area to make sure that [10:03] we are we are um you know capturing the [10:05] right metrics as a as a company and we [10:07] have partnerships with organizations [10:09] like Essence which is a Medicare [10:10] advantage plan in the Midwest US. um and [10:14] they are actually providing the ring as [10:16] a covered benefit to all of their [10:17] beneficiaries and they're very [10:19] interested in engaging with us on how [10:21] these metrics can help um improve health [10:24] and reduce cost. How do you prevent [10:25] readmissions? How do you um you detect [10:27] something earlier so you can intervene [10:29] before something gets serious? So that [10:31] gets me excited as a clinician. That's [10:33] what you want to do. You don't want to [10:35] just sell a you know a shiny trinket, [10:37] but you want to actually move the needle [10:39] in healthcare. And that's why I moved [10:41] from from Apple where I worked for eight [10:43] and a half years over to Aura because I [10:44] think we have a great opportunity to to [10:46] do that now. [10:47] >> Could you give us an example of some of [10:49] the surprising um negative results where [10:52] you thought that this would be, you [10:54] know, a good feature, something that [10:56] Aura would be good at or would lead to [10:58] good outcomes and it turns out that it [11:00] didn't. [11:01] >> So I actually wish I had one because I [11:03] think that would be more compelling if [11:04] if I had it. I think Aura is still a [11:06] relatively young company and we're in [11:09] the sort of infancy of that data [11:11] collection process and I mean to be [11:13] honest the teams have done such a great [11:14] job validating um these algorithms [11:17] before they release them that we haven't [11:19] seen anything that's just doesn't work. [11:22] Um but one thing that I would say is [11:24] early on in the company we we did see [11:27] that there were uh things like activity [11:29] detection in order in other words [11:30] detecting the type of activity that [11:32] you're doing. Um, many of you if you [11:34] have Aura, you've probably seen that it [11:36] it identified activity as like something [11:38] totally random when you were just like [11:39] folding the laundry or something or like [11:41] preparing your meal. Um, so there's [11:43] definitely work that can be done for [11:45] some of that. Activity detection is a [11:47] particularly hard thing to do because [11:48] there's almost an infinite number of [11:50] different activities you can have and [11:52] when you're trying to use um information [11:55] coming literally from one finger, the [11:56] movement of one finger to figure out [11:58] what activity that is, it's a really [12:00] hard problem. But it it just goes to [12:02] show that there there are always [12:03] opportunities to improve and that [12:04] improvement can happen when you expand [12:06] um you know the number of people that [12:08] use the product and you can try to get [12:10] more nuance and you're more inclusive in [12:12] terms of uh you know what you can what [12:14] you can cover. [12:16] >> Are you finding differences in the data [12:18] that you get or the quality of it [12:19] depending on which finger people use and [12:21] whether they use a dominant or [12:22] non-dominant hand? [12:24] >> We so the team has looked at that a [12:26] little bit and we haven't seen a huge [12:28] difference in that. Um, I also would say [12:31] that I think there's probably more work [12:33] that could be done there, but as part of [12:35] our validation, we have to assume that [12:38] people could wear the device on any [12:40] finger. We and and and we don't ask is [12:42] it dominant hand or not. Um, and so so [12:46] so so you have to kind of uh, you know, [12:48] make that assumption and build your [12:49] product knowing that it needs to work [12:51] for anyone regardless of of, you know, [12:54] of what finger they wear it on. Um, and [12:56] so this is something that um, in our in [12:58] our validation, we haven't seen any [13:00] glaring issues regardless of where, you [13:02] know, people people wear the ring. [13:04] >> Um, and in that list of 95 wish list, [13:08] >> yeah, [13:08] >> things that you would like to do. Um, [13:10] anything that particularly excite you? I [13:13] mean, doesn't even have to be started, [13:14] but something you're like, "Yes, this [13:15] would be really cool to do." [13:18] >> Yeah, one of the um, one of the things [13:21] on the list I remember because one of my [13:22] colleagues called it out. So, I'm I'm a [13:24] swimmer. I swim every morning um just [13:26] you know for for health and exercise and [13:28] uh swimming detection was on that list [13:30] and and not just swimming detection is [13:31] actually a little more complex than are [13:33] you swimming or not it's not a binary [13:34] thing it's what stroke are you doing are [13:36] you swimming or doing drills or doing [13:38] kicking or like how many lengths of the [13:40] pool and there are lots of devices today [13:42] that that do that um and so probably not [13:45] the highest priority for us because [13:46] there are a lot of other things [13:47] >> though I have to say what's a fellow [13:49] swimmer they're not very good [13:50] >> they're not very good I actually wear [13:52] the this is an advertisement I wear the [13:53] form goggles. It like actually has the [13:55] little display in the goggles. And those [13:57] have actually been pretty good, I think, [14:00] because it's measuring movement on your [14:02] head rather than your hand. And like a [14:04] wristworn device is much harder for [14:06] swimming because your arms are just [14:07] flapping all around, whereas your head [14:09] is a much more stable place to measure. [14:10] But um we don't have any head devices [14:12] today. We just focus on the hand. So [14:14] that does make it a harder problem. But [14:16] I think that's just one example. There [14:18] are many many others that are far more [14:19] exciting than that. unfortunately I [14:21] can't talk about um but uh you might [14:24] imagine what some of those could be. [14:26] >> Okay, so we're at an AI conference. So [14:28] um earlier this year Aura released the [14:31] Aura Advisor um and this is kind of like [14:34] an ombbudzman [14:36] dock in your pocket safe to say can you [14:38] um tell us sort of the the leadup to [14:41] that and you know as we were saying [14:43] there's one thing to have data and [14:45] another thing to have information right [14:48] that people can act on. And so it seems [14:50] like um you know Aura has spent a few [14:52] years collecting that data and you know [14:55] people can act on it to say oh I need [14:57] more sleep but the advisor is something [15:00] that goes a little deeper right um tell [15:02] us about the thinking to do that behind [15:05] doing that. Yeah, I think one of the [15:07] most common pieces of feedback we get [15:10] and I actually just did a Q&A an hour [15:12] ago and this question came up there as [15:14] well which is a great question which is [15:16] I have this data I want to know [15:19] personally for me how can I improve this [15:23] this this thing um and of course the the [15:27] pre-generative AI approach for that is [15:29] you would have like a complex decision [15:31] tree and you would try to you know based [15:33] on various inputs you would try to you [15:35] know pick a different part of that tree [15:36] and sort of get to something that's sort [15:38] of personalized. Um, that approach is [15:42] very brittle. You it's you're just going [15:44] to miss out on a lot of, you know, cases [15:46] that you either didn't anticipate or, [15:49] you know, you just u are not worth [15:52] spending time on. Um, and so when we [15:55] look at the value of generative AI, [15:57] because we we've been using AI for over [15:59] a decade in our tools, all the machine [16:00] learning and all the algorithms that are [16:02] developed, all AI. When I think we talk [16:04] about AI today, we're mostly referring [16:07] to generative AI. And I know that's at [16:08] this conference a lot of what is being [16:10] discussed. Um, I think a lot of the [16:12] value comes from being able to use that [16:13] to really personalize. Um, and that's [16:16] that's what we care about as well. And [16:17] so you can ask advisor almost [16:19] everything. I don't know that I'd call [16:21] it a doctor. The FDA probably doesn't [16:22] want us to call it a doctor. Um but the [16:25] reality is for these types of tools um [16:29] you know the the buzzword for improving [16:32] your inference or your output of AI has [16:34] always been um prompt engineering I [16:37] think over the last you know 18 months [16:38] or so but more recently we're starting [16:41] to talk about context engineering [16:44] meaning uh it's far more important to [16:46] have high quality data being fed into [16:49] the LLM than it is to figure out what [16:52] your prompt is going to be. It's still [16:53] important to figure out prompts, don't [16:54] get me wrong. Um, but it's far more [16:56] important to have high quality data. [16:58] It's the idea of garbage in, garbage [16:59] out, and that's where I think Aura is in [17:01] this really unique place because we [17:03] believe we have the, you know, the most [17:05] accurate device out there. Um, and we've [17:07] published, you know, that information. [17:10] And if you have really good highquality [17:12] input, you're going to have much higher [17:13] quality output when it comes to [17:15] inference. Um, I still think there's a [17:17] lot more that we can do. There are a lot [17:18] more data types that you can bring [17:20] together. One example is we partnered [17:21] with Dexcom for their Stella [17:22] over-the-counter glucose bio sensor. Um [17:25] it's what they call it, a continuous [17:27] glucometer. Um I wore one as well. I was [17:29] actually um I put it on the morning that [17:32] we left for a twoe cruise. And if you [17:34] all have done cruises before, um you [17:38] know, your your dietary habits are [17:39] probably a little different than they [17:40] would otherwise be if you're not on a [17:42] cruise. And it was fascinating to see [17:46] what my glucose did in that situation. [17:49] And one of the things that I learned was [17:52] um eating late is not good for your [17:55] glucose metabolism. And I think that's [17:56] probably true for most people um because [17:59] your heart rate is higher and it stays [18:01] higher for longer because you're [18:02] digesting the food. My time out of range [18:04] was much worse um when I ate really [18:06] really late. Um, and so I was very [18:08] excited to get back home after that and [18:10] go to my normal eating um, schedule [18:13] where I have, you know, a few hours [18:14] after my dinner before I go to sleep and [18:17] looked totally different and much much [18:19] better. So I think that's an example [18:21] where you need to start bringing in data [18:23] from external sources to complement the [18:25] sensor data to have that 360 degree view [18:28] of a person. And we have a lot of work [18:30] underway to bring even more data types [18:32] together, you know, securely and [18:33] privately for that user. And I think [18:35] that's going to be I don't want to use [18:37] I'm not going to say game changer I just [18:38] did but um it's going to really help us [18:41] be uh you know produce better results [18:44] when it comes to the the the inference [18:45] that we do. I think we only have about a [18:48] minute left. Unfortunately, I could keep [18:50] talking, but um uh when you look ahead [18:54] and as we look at what Aentic AI can do, [18:58] um what role do you see something like [19:00] the Aura Ring playing [19:03] in a partnership with a provider, [19:05] especially addressing, you know, the [19:08] huge healthcare gaps and inequities that [19:10] we currently see in the country where [19:11] access is really an issue and there, you [19:13] know, are just pockets of people who [19:16] don't have any connection with the [19:17] health care system. So as you look at [19:20] what Agentic AI and and a technology [19:23] platform like Aura Ring can do, what is [19:28] your sort of blue sky vision for five 10 [19:31] years down the road for the ring? [19:34] >> Yeah, this is an AI conference. This [19:36] might be a little bit spicy. Um but I [19:38] think that [19:39] >> spicy is good. [19:40] >> Spicy is good. Maybe it's not that [19:41] spicy. Um I think foundation models [19:44] generally are going to be commodities [19:46] within I don't know very soon. Um, and [19:48] they're all going to be very good. We're [19:49] start starting to reach a plateau. We [19:50] see this with chat GPT5 as well. Um, and [19:54] you you can't just keep making it better [19:55] through brute force. And um, performance [19:58] is going to improve a little bit. [19:59] Eventually, these are all going to run [20:00] locally on your device within the not [20:02] too distant future. And performance is [20:03] going to be good enough in that context. [20:05] And so, so I think it's not about the [20:07] model, it's about the ecosystem. And [20:09] when it comes to behavior change, which [20:11] is what we really care about, how do you [20:13] help someone be healthier um ideally at [20:15] lower cost and stay healthier, have an [20:18] improved health span, not just lifespan, [20:20] but more healthy years, that's going to [20:22] require an engaging platform. An LLM is [20:25] not going to bring that by itself, but a [20:27] piece of hardware, something tangible [20:29] that is paired with a high quality [20:31] software experience and validated [20:33] algorithms and these check engine light [20:36] type features, that's really hard to [20:37] replicate. Um, and one of the things [20:39] that we've seen as we've worked with, [20:40] you know, the essences of the world, um, [20:42] these these insurance plans is they care [20:45] about engagement because if a patient is [20:49] engaged, they're more likely to change [20:51] their behavior. And so they partnered [20:53] with us because they see your platform [20:55] is, you know, people wear the ring 2 [20:57] three and a half hours a day. They open [20:58] the app two to three times per day. They [21:00] are engaged. Like engagement is through [21:02] the roof compared to other devices. And [21:04] if they can, you know, get some of that [21:06] magic and help people engage more with [21:08] their health, then everyone's a winner. [21:11] And so I I think my short-term thesis is [21:13] we need to move from this, you know, [21:15] break fix reactive health system to a [21:17] more proactive preventative system. And [21:20] you can do that with a device you're [21:21] wearing 247 if it is evidence-based, if [21:24] it's based in science, and if it does [21:26] what you say it's going to do. I think [21:28] that's a great point to end on that I [21:30] think as we get a more engaged and [21:32] motivated population, we can really [21:35] start shifting that needle toward [21:37] prevention rather than reaction. Ricky, [21:39] thank you so much. [21:40] >> Thanks so much, Alice. Yeah. [21:43] [Applause]