Here, a microfluidic paper-based analytical device (μPAD) was first combined with a deep learning-based smartphone app called “DeepLactate” and then applied for quantitative and selective determination of lactate concentration in sweat. The μPAD was made using wax printing protocol and the detection area was modified with horse radish peroxidase, lactate oxidase and the chromogenic agent 3,3′,5,5′-tetramethylbenzidine for enzymatic detection. The images of μPADs taken by smartphones of several brands in different lighting conditions were used to train various deep learning models to make the system more robust and adaptable to lighting changes. The top-performing model, Inception-v3, was then embedded into a smartphone app, offering easy-operation for non-expert users. Deep learning models, unlike machine learning classifiers, can automatically extract features and be embedded in a smartphone app, enabling analysis without internet access. According to the results, the current system showed a classification accuracy of 99.9 % with phone-independent repeatability and a processing time of less than 1 sec. It also showed excellent selectivity towards lactate over different interfering species. Finally, μPAD was turned into a patch to determine the level of sweat lactate in two volunteers after resting and 15 min of jogging. The system successfully detected lactate in human sweat and confirmed that the level of lactate in sweat increased after jogging. Since the μPAD was designed to first absorb a sample and then transfer it to the detection area, avoiding direct contact with the skin, the system reduces the possibility of skin irritation and has great potential for practical use in a variety of fields including self-health monitoring and sports medicine.

在这里，基于微流控纸的分析装置 ( μ PAD) 首先与基于深度学习的智能手机应用程序“ DeepLactate ”相结合，然后应用于汗液中乳酸浓度的定量和选择性测定。μ PAD采用蜡印协议制作，检测区域用辣根过氧化物酶、乳酸氧化酶和显色剂 3,3',5,5'-四甲基联苯胺进行酶检测。μ的图像多个品牌的智能手机在不同光照条件下拍摄的 PAD 用于训练各种深度学习模型，使系统更加健壮和适应光照变化。然后将性能最佳的模型 Inception-v3 嵌入到智能手机应用程序中，为非专业用户提供简单的操作。与机器学习分类器不同，深度学习模型可以自动提取特征并嵌入智能手机应用程序中，无需访问互联网即可进行分析。根据结果​​，当前系统显示出 99.9% 的分类准确率，具有与电话无关的可重复性和小于 1 秒的处理时间。与不同的干扰物质相比，它还显示出对乳酸的出色选择性。最后，μ将 PAD 变成一个贴片，以测定两名志愿者在休息和慢跑 15 分钟后的汗液乳酸水平。该系统成功检测出人体汗液中的乳酸，并确认慢跑后汗液中的乳酸水平升高。由于μ PAD 的设计目的是先吸收样品，然后将其转移到检测区域，避免与皮肤直接接触，因此该系统降低了皮肤刺激的可能性，在包括自健康监测和运动医学。