展示HN：Aidlab – 开发者的健康数据
大家好！我是Jakub，和我的联合创始人Agnieszka和Nathan一起，我们创建了Aidlab，这是一款可穿戴设备，能够为开发者提供黄金标准的生理数据。与那些API受限的健康追踪器不同，Aidlab提供了一个免费的SDK，支持6个以上的平台，因此你只需使用<i>pip install aidlabsdk</i>或<i>flutter pub add aidlab_sdk</i>，甚至在Unity等平台上，也可以开始实时流式传输原始健康数据和事件，使用简单的<i>didReceive*(timestamp, value)</i>回调。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、身体位置，以及20个高层统计数据，如压力或准备状态。
我收到的最常见问题是：
1) “它比我的智能手表好在哪里？”
2) “我们为什么要开发它？”
胸部佩戴的设备被认为是生理测量的黄金标准。例如，每当苹果验证他们的手表时，他们都会以胸带为基准，因为某些信号只能在靠近心脏的地方可靠测量（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）或身体位置/方向。
至于第二个问题：对我们来说，智能手表太简单，数据也不够准确，而先进的医疗设备则价格昂贵或过于复杂。我们找到了准确性和可及性之间的最佳平衡——Aidlab提供医疗级信号，而不需要医院级的复杂性。由于“医疗级”是一个大胆的声明，我们已经发布了验证论文，将Aidlab的性能与其他认证医疗设备进行了比较。
如今，Aidlab已经是一个相当成熟的概念。我们已经花了两年时间来构建Aidlab，在2020年发布了我们的第一个版本，获得了我们的第一批客户，包括来自Kernel/Blueprint（长寿研究）的Bryan Johnson和波音/杰普森（在测试和训练中监测飞行员生物信号）。
现在我们即将发布Aidlab 2，新增了EDA和GPS等信号，以及一系列新功能，包括设备上的机器学习（我们训练了一些小型LSTM模型，使用TensorFlow Lite for Micro进行推理）。令人兴奋的是，我们在FreeRTOS之上构建了一个自定义的shell，允许任何人在设备上直接调用类似POSIX的命令，例如：
<i>timeout 10 temperature --sampling-rate 1 | tee /data/temperature.csv | tail -n 5</i>
对我们来说，最大的突破是意识到基于云的处理是错误的方法。起初，我们将大部分计算推送到云端——这看起来很自然，但结果却是缓慢、昂贵，开发者也不想要这种方式（“嘿，有没有办法在没有云的情况下使用你的产品？”）。例如，我们的心电图分析管道曾经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。24小时的Holter分析可能会生成超过10万个事件对象，并需要相当长的时间才能完成。现在，我们正在尽一切可能将计算移至边缘。在理想的情况下，云端不应存储或处理任何内容——只需直接接收来自设备的已分析、保护隐私的结果。
另一个教训是：不要在凌晨3点手动焊接原型以节省成本——请支付专业人士来组装电路板。
我们决定现在展示这个项目有三个原因：
- 健康问题在长寿研究和生物黑客的兴起下显得尤为重要，
- 我们接近完成Aidlab 2的最终版本，
- 我非常好奇这里是否有人会觉得它有用！
如果你想亲自检查Aidlab的质量，我们每周都会发布不同活动的免费数据集。
[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_Activity_on_Apple_Watch_November_2024.pdf</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 <i>pip install aidlabsdk</i> or <i>flutter pub add aidlab_sdk</i> or whatever platform (even Unity), and start streaming raw health data and events in real time with simple <i>didReceive*(timestamp, value)</i> 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.<p>The most common questions I got are:<p>1) "how is it better than my smartwatch?"<p>2) "why we built it?"<p>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.<p>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].<p>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).<p>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:<p><i>timeout 10 temperature --sampling-rate 1 | tee /data/temperature.csv | tail -n 5</i><p>The 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.<p>Another lesson: don't hand-solder prototypes at 3 a.m. to save money -> please pay professionals to assemble PCBs.<p>We decided to showcase this now for three reasons:<p>- health feels more relevant than ever with the rise of longevity research and biohacking,<p>- we are close to finalizing Aidlab 2,<p>- and I am super curious to see if anyone here finds it useful!<p>If you'd like to check the quality of Aidlab for yourself, we are publishing free datasets every week during different activities [5].<p>[1] <a href="https://github.com/Aidlab" rel="nofollow">https://github.com/Aidlab</a><p>[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><p>[3] <a href="https://aidlab.com/validation" rel="nofollow">https://aidlab.com/validation</a><p>[4] <a href="https://aidlab.com/aidlab-2" rel="nofollow">https://aidlab.com/aidlab-2</a><p>[5] <a href="https://aidlab.com/datasets" rel="nofollow">https://aidlab.com/datasets</a>