Liquid Medicines Identification with mmWave Sensing: From Theory to Practice

Intravenous infusion plays a key role in delivering rapid and effective treatment for patients in critical care settings. However, many intravenous fluids are visually indistinguishable, and their labeling still depends on manual processes performed by nurses. This raises a critical issue: human labeling is prone to errors, and mislabeling of intravenous medications contributes to millions of medical accidents each year. On the other hand, differentiating between visually similar intravenous fluids is non-trivial, requiring intrusive sampling and sophisticated high-end equipment, which is time-consuming, costly, and can lead to medication contamination. In recent years, wireless sensing has gained attention for its non-intrusive capability in liquid identification. However, while it can effectively differentiate between similar liquids such as whole milk and skimmed milk, it still faces challenges in recognizing liquids with low concentrations (e.g., as low as 0.45% in intravenous medications). The problem is further complicated by the diversity of container materials, geometries, and placements. In this paper, we propose MedID, a millimeter wave based solution for intravenous medicine identification. We formulate a double-layer medium reflection model based on Fresnel equations, which analytically describes how the container's permittivity and wall thickness induce periodic spectral variations, while the liquid's permittivity modulates the signal amplitude. Building on this physical foundation, we designed a GAN-based representation learning model that decouples container and liquid features mixed within the radio-frequency signals. We evaluated MedID in a major hospital, encompassing five types of intravenous medications, various infusion bottle materials, and common deviations in bottle placement. All samples were prepared on-site by nurses according to clinical prescriptions and corresponded to the medications actually administered to patients. The results show that the MedID system can non-intrusively identify medications with a recall of 95.42%.