Resistive skin biosensors refer to a class of imperceptible wearable devices for health monitoring and human–machine interfacing, in which conductive materials are deposited onto or incorporated into an elastomeric polymeric sheet. A wide range of resistive skins has been developed so far to detect a wide variety of biometric signals including blood pressure, skin strain, body temperature and acoustic vibrations; however, they are typically non-specific, with one resistive signal corresponding to a single type of biometric data (one-mode sensors). Here we show a hierarchically resistive skin sensor made of a laminated cracked platinum film, vertically aligned gold nanowires and a percolated gold nanowire film, all integrated into a single sensor. As a result, hierarchically resistive skin displays a staircase-shaped resistive response to tensile strain, with distinct sensing regimes associated to a specific active material. We show that we can, through one resistive signal, identify up to five physical or physiological activities associated with the human throat speech: heartbeats, breathing, touch and neck movement (that is, a multimodal sensor). We develop a frequency/amplitude-based neural network, Deep Hybrid-Spectro, that can automatically disentangle multiple biometrics from a single resistive signal. This system can classify 11 activities—with different combinations of speech, neck movement and touch—with an accuracy of 92.73 ± 0.82% while simultaneously measuring respiration and heart rates. We validated the classification accuracy of several biometrics with an overall accuracy of >82%, demonstrating the generality of our concept.

电阻式皮肤生物传感器是指用于健康监测和人机接口的一类难以察觉的可穿戴设备，其中导电材料沉积在或纳入弹性聚合物片材中。迄今为止，已经开发出多种电阻皮肤来检测各种生物特征信号，包括血压、皮肤应变、体温和声振动；然而，它们通常是非特定的，一个电阻信号对应于一种类型的生物识别数据（单模式传感器）。在这里，我们展示了一种分层电阻皮肤传感器，由层压裂纹铂薄膜、垂直排列的金纳米线和渗透金纳米线薄膜制成，所有这些都集成到单个传感器中。因此，分层电阻皮肤对拉伸应变表现出阶梯形电阻响应，并具有与特定活性材料相关的不同传感机制。我们证明，通过一个电阻信号，我们可以识别与人类喉咙言语相关的多达五种物理或生理活动：心跳、呼吸、触摸和颈部运动（即多模态传感器）。我们开发了一种基于频率/幅度的神经网络，即深度混合光谱，它可以自动从单个电阻信号中分离出多个生物特征。该系统可以对 11 种活动进行分类，包括语音、颈部运动和触摸的不同组合，准确度为 92.73 ± 0.82%，同时测量呼吸和心率。我们验证了多种生物识别技术的分类准确性，总体准确度>82%，证明了我们概念的通用性。