HackerNews中文版

大家好，我是 Jakub，我和我的联合创始人 Agnieszka 和 Nathan 一起开发了 Aidlab，这是一款可穿戴设备，为开发者提供金标准生理数据。与 API 封闭的健康追踪器不同，Aidlab 附带免费的 SDK [1]，支持 6 个以上的平台，因此您只需 pip install aidlabsdk 或 flutter pub add aidlab_sdk 或任何平台（甚至是 Unity），就可以通过简单的 didReceive*(timestamp, value) 回调函数，开始实时流式传输原始健康数据和事件。目前，我们通过 API 提供了 13 种数据类型，包括原始心电图 (ECG)、咳嗽/打鼾、运动、原始呼吸、皮肤温度、体重、重复次数、身体姿势，以及 20 种高级统计数据，如压力或准备情况。我收到的最常见的问题是：1) “它比我的智能手表好在哪里？”2) “我们为什么开发它？”胸部佩戴式可穿戴设备被认为是生理测量领域的金标准。例如，每当 Apple 验证其手表时，他们都会以胸带 [2] 为基准，因为某些信号只能在心脏附近可靠地测量（或根本无法测量！），包括连续心电图、真实呼吸（基于肺容积变化）或身体姿势/方向。至于第二个问题：对我们来说，问题在于智能手表过于简单，数据不够准确，而先进的医疗设备又过于昂贵或过于复杂。我们找到了一个兼顾准确性和可及性的平衡点——Aidlab 提供医疗级信号，而没有医院级的复杂性。由于“医疗级”是一个大胆的说法，我们已经发表了验证论文，将 Aidlab 的性能与其他经过认证的医疗设备进行了比较 [3]。如今，Aidlab 已经是一个相当成熟的概念。我们已经开发 Aidlab 两年了，于 2020 年发布了第一个版本，我们获得了第一批客户，包括来自 Kernel/Blueprint（长寿研究）的 Bryan Johnson 或来自波音/Jeppesen（在测试和培训期间监测飞行员的生物信号）。现在我们即将发布 Aidlab 2 [4]——增加了 EDA 和 GPS 等信号，以及许多新功能，包括设备端机器学习（我们已经训练了一些小型 LSTM 模型，使用 TensorFlow Lite for Micro 进行推理）。最酷的部分是，我们在 FreeRTOS 之上构建了一个自定义 shell，允许任何人直接在设备上调用类似 POSIX 的命令，例如：timeout 10 temperature --sampling-rate 1 | tee /data/temperature.csv | tail -n 5对我们来说，最大的突破是意识到基于云的计算是错误的方法。一开始，我们将大部分计算推送到云端——这看起来很自然，但结果却很慢、成本高昂，而且开发者也不想要（“嘿，有没有办法在不使用云的情况下使用你的产品？”）。例如，我们的心电图分析流程过去常常将原始数据发送到外部微服务，通过 Bull 队列以 30 分钟为单位进行处理。一个 24 小时的 Holter 分析可能会产生 10 万多个事件对象，并需要很长时间才能完成。现在，我们正在尽一切努力将计算转移到边缘。在理想情况下，云端不会存储或处理任何东西——只需直接从设备接收已经分析过的、保护隐私的结果即可。另一个教训：不要在凌晨 3 点手工焊接原型来省钱 -> 请支付专业人士组装 PCB。我们决定现在展示这一点有三个原因：- 随着长寿研究和生物黑客技术的兴起，健康问题比以往任何时候都更相关，- 我们即将完成 Aidlab 2，- 我非常想知道这里是否有人觉得它有用！如果您想亲自检查 Aidlab 的质量，我们每周都会发布不同活动期间的免费数据集 [5]。[1] <a href="https://github.com/Aidlab" rel="nofollow">https://github.com/Aidlab</a>[2] <a href="https://www.apple.com/health/pdf/Heart_Rate_Calorimetry_Activity_on_Apple_Watch_November_2024.pdf" rel="nofollow">https://www.apple.com/health/pdf/Heart_Rate_Calorimetry_Acti...</a>[3] <a href="https://aidlab.com/validation" rel="nofollow">https://aidlab.com/validation</a>[4] <a href="https://aidlab.com/aidlab-2" rel="nofollow">https://aidlab.com/aidlab-2</a>[5] <a href="https://aidlab.com/datasets" rel="nofollow">https://aidlab.com/datasets</a>

查看原文

Hey HN! I'm Jakub, and together with my co-founders Agnieszka and Nathan, we built Aidlab, a wearable that gives developers gold-standard physiological data. Unlike health trackers with locked-down APIs, Aidlab ships with a free SDK [1] across 6+ platforms so you can just pip install aidlabsdk or flutter pub add aidlab_sdk or whatever platform (even Unity), and start streaming raw health data and events in real time with simple didReceive*(timestamp, value) callbacks. Currently, we are exposing 13 data types including raw ECG, cough/snoring, motion, raw respiration, skin temperature, bodyweight reps, body position, and 20 high-level stats like stress or readiness through the API.The most common questions I got are:1) "how is it better than my smartwatch?"2) "why we built it?"Chest-mounted wearables are considered the gold standard for physiological measurements. For example, whenever Apple validates their watch, they benchmark against chest straps [2], as some signals can only be reliably measured (or measured at all!) near the heart including continuous ECG, true respiration (based on lung volume changes) or body position/orientation.As for the second question: the problem for us was that smartwatches were too simple and the data too inaccurate, while advanced medical devices were too pricey or too complicated. We found a sweet spot between accuracy and accessibility - Aidlab delivers medical-grade signals without the hospital-level complexity. As "medical-grade" is a bold statement, we’ve published validation papers comparing Aidlab’s performance with other certified medical devices [3].Today Aidlab is already a pretty mature concept. We've been building Aidlab for 2 years, shipped our first version in 2020, we got our first clients, including Bryan Johnson from Kernel/Blueprint (longevity research) or Boeing/Jeppesen (monitoring pilots’ bio-signals during tests&training).Now we're about to release Aidlab 2 [4] - with additional signals like EDA and GPS, and a bunch of new features, including on-device ML (we've trained a few small LSTM models running inference with TensorFlow Lite for Micro). The cool part is that we've built a custom shell on top of FreeRTOS, letting anyone invoke POSIX-like commands directly on the device, for example:timeout 10 temperature --sampling-rate 1 | tee /data/temperature.csv | tail -n 5The biggest breakthrough for us was realizing that cloud-based processing was the wrong approach. In the beginning, we pushed most of the computation to the cloud - it seemed natural, but turned out to be slow, costly, and devs didn't want it ("hey, is there a way to use your product without cloud?"). For example, our ECG analysis pipeline used to send raw data to an external microservice, processing it in 30-minute chunks through Bull queues. A 24-hour Holter analysis could spawn 100k+ event objects and take significant time to complete. Now we're doing everything we can to move computation to the edge. In an ideal world, the cloud wouldn't store or process anything - just receive already-analyzed, privacy-preserving results straight from the device.Another lesson: don't hand-solder prototypes at 3 a.m. to save money -> please pay professionals to assemble PCBs.We decided to showcase this now for three reasons:- health feels more relevant than ever with the rise of longevity research and biohacking,- we are close to finalizing Aidlab 2,- and I am super curious to see if anyone here finds it useful!If you'd like to check the quality of Aidlab for yourself, we are publishing free datasets every week during different activities [5].[1] <a href="https://github.com/Aidlab" rel="nofollow">https://github.com/Aidlab</a>[2] <a href="https://www.apple.com/health/pdf/Heart_Rate_Calorimetry_Activity_on_Apple_Watch_November_2024.pdf" rel="nofollow">https://www.apple.com/health/pdf/Heart_Rate_Calorimetry_Acti...</a>[3] <a href="https://aidlab.com/validation" rel="nofollow">https://aidlab.com/validation</a>[4] <a href="https://aidlab.com/aidlab-2" rel="nofollow">https://aidlab.com/aidlab-2</a>[5] <a href="https://aidlab.com/datasets" rel="nofollow">https://aidlab.com/datasets</a>

Show HN: Aidlab – 开发者适用的健康数据