Introduction: : Continuous glucose level tracking is required for diabetes patients to prevent deviations of glucose concentrations in blood from the normal physiological range. Continuous glucose monitoring (CGM) systems have emerged for real-time tracking of glucose with minimal user intervention, facilitating better diabetes management for patients. Conventional CGMs measure glucose in the interstitial fluid (ISF) using subcutaneously inserted electrodes functionalized with glucose-specific enzymes. Even though these electrochemical CGMs are commercially available, they have a limited lifetime (10-14 days), increasing the operational costs. Alternative CGMs based on optical modalities have been proposed as a longer-term alternative to standard electrochemical CGMs. Insertable biocompatible phosphorescent sensors have been developed for glucose level tracking in ISF; however, appropriate readout hardware is still needed for reliable processing of sensor emissions while being resilient to potential misalignments between the sensor and the external reader. We developed a compact phosphorescence lifetime imager (PLI) for misalignment-tolerant inference of glucose levels using phosphorescent sensor emissions (Fig. 1a) [1]. The PLI captures phosphorescence intensity and lifetime images of the sensor, which are processed by convolutional neural network (CNN) models for misalignment-tolerant glucose level classification between low ( < 70 mg/dL), normal (70–125 mg/dL) and high (>125 mg/dL) ranges. The PLI reader achieved 88.8% accuracy in glucose level classification, accommodating up to ~4.7 mm lateral misalignment of the reader [1]. The reader also accurately identified (with 100% accuracy) larger misalignments beyond 5 mm. Biocompatibility of the sensor, and the misalignment-tolerant glucose level inference capability of the reader, make our platform an appealing next-generation CGM system.
Materials and
Methods: : The insertable sensor is made of a hydrogel (PEGDA) and contains 4 spatially separated channels, including 2 test channels and 2 control channels, all filled with phosphorescent dyes. In addition to the dyes, test channels also incorporate glucose-specific enzymes, which make them responsive to variations in the local glucose levels. The sensor has a compact rectangular shape with a size of ~1x1x6.5 mm. For each measurement, the PLI reader captures 11 images of the sensor evenly distributed within 200 µs interval after the excitation LED turns off. These images are utilized to obtain phosphorescence intensity and lifetime images of the sensor, which are processed by CNN models for the misalignment tolerant glucose level classification between low, normal, and high concentration ranges [1]. We tested our system on glucose-spiked DI water samples using skin phantom (1 mm thickness, Type 1-2) to simulate optical properties of human skin. To test misalignment tolerance, we split the reader FOV into two regions, namely the aligned zone and the misaligned zone, where the aligned zone represented a 3.2 × 3.4 mm2 rectangle at the center of the reader FOV, while the misaligned zone covered the remaining area (Fig. 1b) [1]. All captured images were first processed by the alignment model (CNNAlignment), which validated the alignment between the sensor and the reader by identifying whether the sensor stayed within the aligned zone. All sensors located within the aligned zone were further processed by the classification model (CNNClass), which classified glucose levels as low, normal and high.
Results, Conclusions, and Discussions:: In vitro testing of our platform revealed that phosphorescence lifetime responses from the insertable sensors captured by the PLI reader through the skin phantom had a strong correlation with glucose concentrations as long as the sensor stayed within the aligned zone. In addition, inter-sensor repeatability for sensor locations within the aligned zone was reliable with a coefficient-of-variation (CV) of ~10-15% (Fig. 1d-e). In contrast, phosphorescent responses from the sensors located in the misaligned zone had a substantially lower correlation with glucose concentrations, also showing inferior inter-sensor repeatability (i.e., CV > 15%, Fig. 1f) [1]. These results highlight the need for the incorporation of the misalignment control step to minimize the negative impact of the reader misalignments on the accuracy of glucose level inference. Blind testing of our neural network-based misalignment-tolerant glucose level classification pipeline showed that CNNAlignment achieved 100% accuracy in differentiating sensor locations between the aligned and misaligned zones (Fig. 1c [left]). CNNAlignment utilized phosphorescence intensity images of the sensors as input, benefitting from spatial features present in the acquired images for accurate classification of the sensor locations underneath the skin phantom. In cases when CNNAlignment located the sensor outside of the aligned zone, i.e., in the misaligned zone, the reader asked the user for realignment, and no glucose level inference took place. Only when CNNAlignment located the sensor within the aligned zone, it was further processed by CNNClass, which utilized phosphorescence lifetime images of the sensors for glucose level classification between the 3 concentration ranges (low, normal, high). CNNClass achieved a classification accuracy of 88.8% when blindly tested on sensors located in the aligned zone by CNNAlignment (Fig. 1c [right]). Synergistic implementation of both CNNAlignment and CNNClass accommodated accurate glucose level inference with up to ~4.7 mm lateral misalignment of the reader. In conclusion, we demonstrated a minimally invasive CGM system featuring a compact biocompatible sensor and a phosphorescence lifetime imaging-based wearable reader [1]. The compact and biocompatible nature of the sensor, along with the misalignment-tolerant glucose level inference capability of the PLI reader, make this platform appealing for next-generation CGM platforms.
