Glucose sensor using a compact and cost-effective phosphorescence imager and machine learning

Abstract

A computational continuous glucose monitoring system is disclosed which integrates a biocompatible phosphorescence-based insertable sensor and a custom-designed phosphorescence imager (PI). This PI captures phosphorescence lifetime images of an insertable sensor through the skin where the lifetime of the emitted phosphorescence signal is modulated by the local concentration of glucose. The lifetime images acquired through the skin are processed by neural network-based models for misalignment-tolerant inference of glucose levels, accurately revealing normal, low (hypoglycemia) and high (hyperglycemia) concentration ranges. Using a 1-mm thick skin phantom mimicking the optical properties of human skin, in vitro testing of the PI using glucose-spiked samples yielded 88.8% inference accuracy, also showing resilience to random and unknown misalignments within a lateral distance of ~4.7 mm. Lateral misalignments beyond 5 mm prompted user intervention for re-alignment.

Description

2024-260-2 GLUCOSE SENSOR USING A COMPACT AND COST-EFFECTIVE PHOSPHORESCENCE IMAGER AND MACHINE LEARNING Related Application

[0001] This Application claims priority to U.S. Provisional Patent Application No. 63 / 657,743 filed on June 7, 2024, which is hereby incorporated by reference in its entirety. Priority is claimed pursuant to 35 U.S.C. § 119 and any other applicable statute. Technical Field

[0002] The technical field generally relates to continuous glucose monitoring (CGM) systems. More particularly, the technical field relates to a computational CGM system, that integrates a biocompatible phosphorescence-based insertable sensor and a custom-designed phosphorescence imager (PI). Statement Regarding Federally Sponsored Research and Development

[0003] This invention was made with government support under 1648451 awarded by National Science Foundation. The government has certain rights in the invention. Background

[0004] Wearable sensors have emerged as a minimally invasive monitoring platform for real-time detection of key physiological and healthcare biomarkers, including blood pressure, oxygen level, temperature, pH, glucose level, electrolytes, hormones, and metabolites. Recent advances in hardware miniaturization, material science and data analysis techniques have allowed for robust in vivo sensing technologies using various samples such as sweat, saliva and interstitial fluid (ISF). Continuous glucose monitoring (CGM) devices stand out as an important class of wearable sensors utilized for real-time tracking of glucose levels. Glucose is a biomarker for diabetes, a chronic metabolic disorder, affecting over 450 million individuals globally, with projected patient numbers expected to increase to over 550 million by 2030. Daily monitoring of glucose levels is required for the proper management of diabetes to prevent hypo- and hyperglycemic events, which may cause severe health complications, such as cardiovascular disease (CVD), kidney failure, and neuropathy. Conventional glucose monitoring methods rely on highly invasive and laborious finger2024-260-2 pricking procedures; CGM systems have evolved as a user-friendly alternative approach, facilitating real-time tracking of patients’ glucose levels.

[0005] Typical CGMs employ subcutaneous electrodes connected to a wearable transmitter to collect glucose information from ISF through enzymatic electrochemical reactions. Several CGMs based on this technology have received approvals from the Food and Drug Administration (FDA) for human use. However, despite their widespread adoption, these CGMs are associated with relatively high costs (i.e., ~$450 / month) due to the short lifetime of the sensing electrodes (~1-2 weeks) and cause patient discomfort during the electrode implantation and operation. In addition, the insertion of the electrodes under the skin may damage surrounding cells and potentially cause tissue inflammation, leading to skin irritation and reduced testing accuracy.

[0006] To address some of these challenges, alternative methods based on optical signal transduction have been proposed. In particular, fluorescence, colorimetric detection, near- infrared spectroscopy, and surface-enhanced Raman spectroscopy (SERS), are some of the emerging optical modalities for noninvasive and minimally invasive monitoring of glucose levels. Among these methods, fluorescence-based minimally invasive CGMs have been successfully tested in vivo and, to date, represent the only FDA-approved optical CGM platform. This CGM system includes an implantable glucose-specific fluorescent sensor, which measures glucose concentration in ISF and sends the data to a removable wearable transmitter, connected to a smartphone for data display. The sensor is implanted subcutaneously into the upper arm and can operate for up to 90 days post-implantation, offering a longer-term and more cost-effective solution (i.e., ~$120 / month) compared to electrochemical CGMs. However, this platform has lower testing accuracy and requires additional calibration through finger pricking ~2 times per day, hindering the minimally invasive nature of the technology. In addition, the relatively large size of the implant (i.e., 3.3 mm x 15 mm) may lead to tissue inflammation, causing patient discomfort and potentially reducing sensor accuracy. Therefore, further advancements in optical CGMs are necessary to enhance the usability, stability, and overall robustness of these devices for their widespread adoption in POC settings.

[0007] As an alternative to existing optical sensor technologies, phosphorescence is an appealing optical technique for CGM systems due to a substantially longer lifetime of phosphorescence emission (>10 µs) compared to human tissue autofluorescence (≤10 ns), allowing for a superior temporal separation between native fluorophores and phosphorescent2024-260-2 sensing signals. Consequently, phosphorescence-based sensors provide a better signal-to- noise ratio (SNR) in comparison with fluorescence-based designs, leading to a higher signal transduction efficiency. Palladium and platinum porphyrins form a major class of oxygen- sensitive phosphorescent probes reported in the literature for biosensing applications. Both the emission intensities and lifetimes of these phosphors vary in response to differences in the local oxygen concentrations, and with the addition of enzymes consuming the analyte of interest, phosphor emissions can become sensitive to the analyte concentrations. Therefore, phosphorescence-based sensors utilizing this “indirect” sensing approach have been developed for continuous glucose monitoring in ISF. These sensors are typically smaller in size (i.e., 1.2 mm x 6.5 mm) compared to the conventional fluorescence-based sensors, enabling improved biocompatibility and easier subcutaneous insertion.

[0008] With these advances in phosphorescence-based biosensor fabrication, the development of appropriate readout hardware that is both compact / wearable and cost- effective becomes essential for accurate and real-time inference of glucose levels through the skin. Conventional wearable phosphorescence readers deliver excitation light to phosphors and use a single photodetector to capture the signal intensity and / or lifetime of phosphorescent emissions. However, this single pixel-based readout approach encounters challenges in uncontrolled real-life testing environments, mostly related to reader misalignments due to patient movement. These misalignments can significantly alter the effective signal captured by the photodetector, leading to high variability in the output glucose levels. Some optical CGMs overcome this issue by integrating both the sensing assay and the miniaturized readout module within the implanted sensor, utilizing an external wearable reader solely for data transfer and display. A major drawback of this method is the relatively large size of the sensor (i.e., 3.3 mm x 15 mm), which lowers its biocompatibility and requires surgical operation for implantation and removal, thereby limiting its usability. Furthermore, owing to its large size, this sensor causes inflammation in the surrounding tissue, leading to variability in glucose diffusion rate and necessitating invasive re- calibration. In contrast, a passive phosphorescence-based insertable sensor that solely contains the sensing assay would be significantly more compact, allowing for easier insertion with less tissue inflammation; however, it would necessitate an external wearable and cost- effective hardware for the signal readout. Therefore, efficient alignment control of the external reader positioning is particularly important to ensure accurate sensor readout through the skin and minimize misalignment-induced glucose concentration inference errors. Single2024-260-2 photodetector-based readers, in general, lack spatial information about the relative sensor location, preventing effective control over reader alignment. Imaging-based readout configurations present a promising alternative owing to their ability to capture emitted signals in a highly multiplexed manner by capturing sensor images. Such imaging-based solutions can provide spatial information about the sensor’s location, enabling alignment control of the reader. Summary

[0009] Here, a compact and cost-effective phosphorescence imager (PI) is disclosed for misalignment-tolerant inference of glucose levels using an insertable passive hydrogel sensor. The PI continuously captures both phosphorescence intensity and lifetime images of an insertable sensor through the skin and automatically processes the acquired image data using trained neural network models to: (1) assess the alignment of the PI’s field-of-view (FOV) with respect to the implanted sensor location and inform the user in case of misalignments; and (2) classify the glucose concentration levels as low (<70 mg / dL), normal (70-125 mg / dL), and high (>125 mg / dL). It should be noted that glucose concentrations in the 125- 180 mg / dL range are associated with prediabetes, another widespread health condition linked to a higher risk of developing type 2 diabetes and CVD. Just in the US, over 90 million patients are affected by this disease, with 80% being unaware of their condition. Hence, early detection of both diabetes and prediabetes through CGM systems holds promise to lower diabetes prevalence and improve the quality of glucose management.

[0010] The performance of the PI system was tested in vitro using glucose-spiked deionized (DI) water samples using a 1-mm thick skin phantom to simulate the optical properties of the skin tissue. The results demonstrated consistent sensor response for glucose concentrations in the range of 0-250 mg / dL and a good inter-sensor repeatability with a coefficient of variation (CV) of <15% regardless of the glucose level. In addition, the PI was integrated with a neural network-based framework for misalignment resilient inference of glucose levels and showed an accuracy of 88.8% for the classification of glucose concentration levels (low / normal / high) while being resilient to random / unknown misalignments within a lateral distance of ~4.7 mm. Furthermore, this compact and cost- effective PI reader achieved 100% accuracy in autonomously identifying larger physical misalignments beyond 5 mm, prompting the users to re-align the reader and minimizing misalignment-induced glucose level measurement inaccuracies. This misalignment-resilient2024-260-2 glucose level inference capability of the PI system, coupled with the biocompatibility and small size of the insertable phosphorescence biosensor, make the platform an attractive CGM system for personalized monitoring of ISF glucose levels. The same imaging platform also permits multiplexed sensing capabilities to implement parallel monitoring of multiple biomarkers through the skin, thereby expanding the application scope of this technology beyond glucose monitoring, potentially making it a versatile wearable sensing platform.

[0011] In one embodiment, a method of sensing glucose levels in a subject using an implantable sensor includes implanting the sensor in the subject. The sensor includes a substrate having one or more test regions, wherein the one or more test regions comprise phosphorescent nanoparticles, glucose oxidase, and catalase. As explained herein, these may be contained in microparticles within a hydrogel. The implanted sensor is imaged with a phosphorescence imager that acquires a plurality of raw phosphorescence intensity images of the one or more test regions (and optionally one or more control regions). One or more phosphorescence intensity images of the one or more test regions is generated from the plurality of raw phosphorescence intensity images. One or more phosphorescence lifetime images of the one or more test regions is / are also generated from the plurality of raw phosphorescence intensity images. The generated phosphorescence intensity image(s) and / or the raw phosphorescence intensity images are input to a first trained neural network that is trained to classify the implantable sensor (or the phosphorescence imager) as properly aligned or misaligned. The one or more generated phosphorescence intensity image(s) and / or the raw phosphorescence intensity images and / or the one or more phosphorescence lifetime image(s) is / are input to a second trained neural network that is trained to generate a qualitative or quantitative output of the glucose level of the subject. In a preferred embodiment, the phosphorescence lifetime image is used as an input to the second trained neural network.

[0012] In another embodiment, a system for sensing glucose levels in a subject is provided. The system includes an implantable sensor comprising a substrate having one or more test regions (and optionally one or more control regions), wherein the one or more test regions comprise phosphorescent nanoparticles, glucose oxidase, and catalase (which may be in microparticles contained within a hydrogel). The system further includes a phosphorescence imager that comprises one or more light sources, a camera or image sensor, and light source driving circuitry configured to pulse the one or more light sources, wherein the phosphorescence imager is configured to capture a plurality of raw phosphorescence intensity images of the implantable sensor. A computing device separate from or part of the2024-260-2 phosphorescence imager contains image processing software configured to generate one or more phosphorescence intensity images and one or more phosphorescence lifetime images from a plurality of raw phosphorescence intensity images of the implantable sensor. The computing device further executes a first trained neural network that is trained to classify the implantable sensor (or the phosphorescence imager) as properly aligned or misaligned based on the one or more generated phosphorescence intensity images and / or the raw phosphorescence intensity images and a second trained neural network that is trained to generate a qualitative or quantitative output of the glucose level of the subject based on the one or more of the raw phosphorescence intensity images, the generated phosphorescence intensity images and / or the generated phosphorescence lifetime images. Brief Description of the Drawings

[0013] FIG.1A illustrates top and side views of an insertable / implantable sensor according to one embodiment.

[0014] FIG.1B illustrates a system for sensing glucose levels in a subject. The system includes a phosphorescence imager or reader device for imaging phosphorescence light emitted from an illuminated sensor that is implanted on or within the skin tissue of a living subject. A computing device is provided that includes image processing software.

[0015] FIG.2 illustrates a schematic overview of optical continuous glucose monitoring using the sensor, featuring an insertable / implantable passive sensor and a wearable phosphorescence imager or reader device.

[0016] FIG.3A illustrates the operation of the phosphorescence imager (PI) which is used to monitor glucose levels in a subject.

[0017] FIG.3B illustrates the configuration of the test regions and control regions (e.g., channels) in the sensor.

[0018] FIG.3C illustrates the neural network-based processing of phosphorescence images for misalignment-resilient classification of glucose concentration levels. The CNNAlignment network is used to check for alignment or misalignment of the sensor relative of the phosphorescence imager or reader device. The CNNClassnetwork is used to classify the measured glucose level using the sensor into one of three ranges (low / normal / high).

[0019] FIG.4A illustrates the experimental setup for in vitro testing of the PI reader.2024-260-2

[0020] FIG.4B shows the reconstruction of the phosphorescence intensity and phosphorescence lifetime images from the raw phosphorescence images. Also illustrated are the timing events for the one or more light sources, emission intensity, and camera activation.

[0021] FIG.5 shows the PI reader field-of-view (FOV) with aligned and misaligned regions. Aligned region, represented by a 3.2 mm x 3.4 mm rectangle, is a misalignment- tolerant zone. When the implanted sensor location is within the aligned region, the PI reader and the trained neural network accurately perform glucose level inference despite random reader misalignments from the ideal location (i.e., FOV center). If the implanted sensor falls into the misalignment region, the reader will prompt the user for re-alignment. Also illustrated are the lifetime responses of the insertable sensors for 0-250 mg / dL glucose concentration range captured by the PI reader for three representative sensor locations withing the FOV, including the ideal location (i.e., the center of the aligned region), a non- ideal location (i.e., the border of the aligned region), and a misaligned location within the misaligned region.

[0022] FIG.6A illustrates neural network-based misalignment-resilient glucose level inference framework. The framework consists of two separate convolutional neural networks, namely the alignment network (CNNAlignment) and the classification network (CNNClass). First, phosphorescence intensity images acquired by the PI are processed by CNNAlignment, which classifies the reader location as aligned or misaligned. If the PI reader position is identified as misaligned, the reader prompts the user for re-alignment and CNNClassis not used in this case. If the reader location is identified as aligned, the PI image data are further processed by the second neural network, CNNClass, which classifies the glucose concentration levels (low / normal / high) using phosphorescence lifetime measurements as input data;

[0023] FIG. 6B illustrates blind testing results for CNNAlignment and forCNNClass.

[0024] FIGS.7A-7B illustrate timelapseresponses. Shown are timelapse phosphorescence intensity responses for the three representative sensor locations depicted in FIG.5 (ideal (aligned), ideal (non-aligned), misaligned).

[0025] FIGS.8A-8D illustrates inter-sensor repeatability comparison between the phosphorescence intensity and phosphorescence lifetime responses. FIG.8A shows phosphorescence intensity and phosphorescence lifetime images for deionized (DI) water samples for 8 tested insertable sensors; FIG.8B intensity responses, FIG.8C signal-to-noise2024-260-2 ratios (SNRs), and FIG.8D lifetime responses along with the corresponding coefficient of variation (CV) values for the DI water samples for the 8 tested sensors.

[0026] FIGS.9A-9C illustrates intensity responses for a glucose concentration sweep. FIG.9A shows intensity responses to a glucose concentration sweep (0-250 mg / dL) for the three locations depicted in FIG.5. Comparison of FIG.9B inter-sensor CVs and FIG.9C shows intra-sensor CVs for phosphorescence lifetime and intensity responses.

[0027] FIGS.10A-10C illustrates the impact of different pixel binning sizes on the sensor 2 lifetime response. FIG.10A shows coefficients of determination (R ), FIG.10B shows intra- sensor CVs, and FIG.10C shows inter-sensor CVs for the phosphorescence sensor lifetime responses to a glucose concentration sweep for different pixel binning sizes (N).

[0028] FIG.11 illustrates a schematic of the alginate microparticles coating with polyelectrolyte layers (CX layers).

[0029] FIGS.12A and 12B illustrate sensor responses for different numbers of CX layers at two different temperatures (23°C and 37°C).

[0030] FIGS.13A-13B illustrates classification network (CNNClass) performance for lower resolution images. Blind testing results for CNNClassfor pixel binning sizes (N) of 20 (FIG. 13A) and 50 (FIG.13B).

[0031] FIGS.14A-14C illustrates CNNClassperformance for extended aligned regions. FIGS.14A and 14B show the PI reader FOV with aligned, misaligned and extended aligned regions for extended aligned region #1 and extended aligned region #2. FIGS.14C and 14D show blind testing results for CNNClassfor extended aligned region #1 (FIG.14C) and extended aligned region #2 (FIG.14D).

[0032] FIGS.15A-15D illustrates LED driver circuit. FIG.15A is a top-level schematic of the LED driver circuit. The circuit consists of a constant current LED driver (FIG.15C), powered by voltage from an external supercapacitor (FIG.15D). Pulse operation of the LED is controlled by the pulse width modulation (PWM) pin, connected to an external PWM timer circuit (FIG.15B).

[0033] FIGS.16A-16E illustrates sensor responses for different LED excitation times. FIG.16A illustrates phosphorescence intensity images. FIGS.16B and 16C shows timelapse phosphorescence intensity responses. FIG.16D shows phosphorescence lifetime responses, and FIG.16E shows SNR values for a DI water sample and different LED excitation times (texcitation).2024-260-2

[0034] FIGS.17A-17B illustrates the graphical user interface (GUI) of the PI reader. FIG. 17A shows phosphorescence intensity image and FIG.17B shows phosphorescence lifetime image.

[0035] FIGS.18A-18E illustrates the impact of using different numbers of phosphorescence intensity images to generate a phosphorescence lifetime image. FIG.18A 2 R , FIG.18B intra-sensor CVs, FIG.18C inter-sensor CVs, FIG.18D lifetime dynamic ranges between 0 mg / dL and 250 mg / dL, and FIG.18E lifetime responses for 0 mg / dL and 250 mg / dL for different numbers of phosphorescence intensity images used to generate a phosphorescence lifetime image.

[0036] FIG.19 illustrates the PI field of view (FOV) with sensor locations used for each sensor / glucose concentration. Dimensions are illustrative. Detailed Description of Illustrated Embodiments

[0037] With reference to FIGS.1A, 1B, and 2, a system 10 for sensing glucose levels in a living subject (e.g., mammal) includes an implantable or insertable sensor 20 along with a compact a phosphorescence imager 40. The implantable sensor 20 is, in one preferred embodiment, formed from a hydrogel substrate 22 having one or more test regions 24 and one or more control regions 26 formed therein. In some embodiments, the one or more control regions 26 may be omitted entirely. The one or more test regions 24 and the one or more control regions 26 may include channels 28 that are formed in the substrate 22. These channels 28 are then filled with the appropriate test or control reagents as discussed herein. The one or more test regions 24 and optional control regions 26 form a “barcode” structure that include spatially separated discrete compartments holding the test or control reagents. In a preferred embodiment, the one or more test regions 24 includes phosphorescent nanoparticles, glucose oxidase, and catalase (FIG.3B). In one embodiment, the phosphorescent nanoparticles, glucose oxidase, and catalase are contained in microparticles made from a biocompatible material (e.g., made from alginate) that are mixed with a hydrogel solution (e.g., same material as substrate 22) which then fills the one or more test regions 24 by pipetting or other loading procedure. The phosphorescent nanoparticles may include ethyl cellulose nanoparticles (ECNP) containing palladium (II) meso-tetra(4- carboxyphenyl)tetrabenzo-porphyrin) phosphorescent dye (PdBP). The one or more control regions 26, if used, include a phosphorescent dye (i.e., just the phosphors). The dye may be2024-260-2 mixed with the hydrogel and loaded into the one or more control regions 26 similar to how the one or more test regions 24 are loaded.

[0038] The sensor 20 is preferably implanted or inserted into the skin 110 of the subject to be tested as seen in FIGS.1B and 2. The sensor 20 is located beneath the surface of the skin 110 and preferably in the subcutaneous layer as this has ample ISF that can contact the test / control regions 24, 26. The sensor 20 may be implanted or inserted in any number of ways including ejection from a delivery device like a needle from a syringe 112 or the like. An incision may also be made in skin 110 where the sensor 20 is then inserted into the tissue. The substrate 22 is made of a biocompatible material and in one embodiment includes a hydrogel such as polyethylene glycol diacrylate (PEGDA) but other hydrogel or biocompatible materials may also be used. In the illustrated embodiment, there are two test regions 24 and two control regions 26 but it should be appreciated that different numbers of these test / control regions 24, 26 may be used with the sensor 20. In another embodiment, the sensor 20 may include a microneedle patch which is placed on the skin 110 and the microneedles penetrate the skin 110. The microneedles of the patch form the test / control regions 24, 26.

[0039] With reference to FIG.1B, the system 10 further includes a phosphorescence imager 40 that has one or more light sources 42, a camera or image sensor 44, and light source driving circuitry 46 configured to pulse the one or more light sources 42. The light source driving circuitry 46 and associated electrical components may be powered by a battery (not shown) in the phosphorescence imager 40 or through a power source that operates the camera 44 as disclosed herein. The one or more light sources 42 may include light emitting diodes (LEDs) or laser diodes. As described herein, one particular implementation used a red LED (633 nm) as the one or more light sources 42. With reference to FIG.1B, light from the one or more light sources 42 acts as excitation light that enters a collimating lens 48, a bandpass filter 50, and a focusing lens 52. Light then passes to a cold mirror 54 that directs light through a sample lens 56 and onto the skin 110 which contains the sensor 20. Excited (phosphorescent) light from the test regions 24 (and optional control regions 26) reflects back through the sample lens 56 and onto the cold mirror 54 and through a long-pass filter 58 and an imager lens 60 and onto the camera or image sensor 44. Illustrative dimensions are shown in FIG.4A for an embodiment of the phosphorescence imager 40 used for experiments described herein, but it should be appreciated that the dimensions are not limited. In some embodiments, the phosphorescence imager 40 may be made small such that it can be worn by2024-260-2 the user like a watch or the like as illustrated in FIG.2. In such an embodiment, the phosphorescence imager 40 may include a display 82 such as illustrated in FIG.2. In the embodiment described herein, the size was ~5 cm x 5 cm but it could be made smaller or larger. It should be appreciated that the phosphorescence imager 40 may utilize lens-less or lens-free designs to further reduce the size and cost.

[0040] The light source driving circuitry 46 is used to operate the one or more light sources 42 in pulsed mode. FIGS.15A-15D illustrates one exemplary light source driving circuitry 46 that was designed to operate the excitation LED in a pulse mode. The light source driving circuitry 46 includes an LED driver 70, PWM timer circuitry 72, and supercapacitor charger 74 and is based on a 1 A constant-current LED driver 70 which contains internal pulse width modulation PWM timer circuitry 72 allowing for pulsed operation with custom pulse on / off times. The operation of the PMW pin was set by a timer circuitry consisting of an 8 MHz oscillator, a 14-stage shift register, and a quadruple NAND gate. The oscillator generated square pulses with 125 ns (50% duty cycle) baseline period, and the shift register further divided baseline oscillation frequency, yielding square pulses with 2,000 µs and 500 µs periods. Aa quadruple NAND gate combined these two square pulses to generate the desired waveform with 250 µs (ton) and 1000 µs (toff) on and off times, respectively. The waveform shape can be adjusted to accommodate different lifetime ranges by changing the baseline oscillation frequency or the output pins of the shift register. The constant-current LED driver 70 was powered by an external supercapacitor charger 74 to supply the required input voltage (i.e., VIn= 6V). The LED light source 42 was connected to GND through an N-channel MOSFET, enabling fast switching between on / off cycles. The LED driver circuit 46 was powered by the output supply voltage pin (5 V, 100 mA) of the PI camera 44.

[0041] FIG.1B illustrates a computing device 80 that interfaces with the phosphorescence imager 40. The computing device 80 may communicate via a wired or wireless connection (e.g., WiFi or Bluetooth). The computing device 80 may include a personal computer, laptop, tablet PC, smartphone, or the like. The computing device 80 includes a display 82 that displays a Graphical User Interface (GUI) 84 which is used to review images generated with the phosphorescence imager 40. The GUI 84 can also be used to adjust or specify the operation parameters of the phosphorescence imager 40, including shutter parameters, exposure time and gain. The GUI 84 also displays the results of the measurements made of the sensor 20 with the phosphorescence imager 40 which includes a qualitative or2024-260-2 quantitative output of the glucose level of the subject. The GUI 84 has two displaying modes, namely Camera mode and File mode. In the Camera mode, the GUI 84 displayed real-time images directly captured by the camera 44, while in the File mode, the displayed images were transferred from a user-specified path. For each successive phosphorescence lifetime image 94 measurement, the camera shutter can be programmed to capture a plurality of timelapse phosphorescence images and the background image after the LED on cycle. For each of these images, the camera shutter is activated periodically, and the user is able to specify both the shutter activation times and the shutter on / off time periods with the GUI 84.

[0042] The computing device 80 executes image processing software 86 via one or more processors 88 (FIG.1B). As explained herein, the image processing software 86 may be used to store raw phosphorescence intensity images 90 of the sensor 20 acquired with the camera or image sensor 44. The image processing software 86 also generates one or more phosphorescence intensity images 92 from the raw phosphorescence intensity images 90 and / or generate one or more phosphorescence lifetime images 94 from the raw phosphorescence images 90. In one embodiment or aspect of the invention, a plurality of raw phosphorescence intensity images 90 are used to generate a single phosphorescence intensity image 92 at a given measurement point. Of course, a plurality of phosphorescence intensity images 92 may also be generated by the image processing software 86. This may be the case, for example, if the phosphorescence imager 40 operates in a continuous mode or makes multiple measures over a period of elapsed time. Similarly, one embodiment or aspect of the invention, a plurality of raw phosphorescence intensity images 90 are used to generate a single phosphorescence lifetime intensity image 94 at a given measurement point. Of course, a plurality of phosphorescence lifetime intensity images 94 may also be generated by the image processing software 86. This may be the case, for example, if the phosphorescence imager 40 operates in a continuous mode or makes multiple measures over a period of elapsed time. As an alternative to having the image processing software 86 executed using a separate computing device 80, the image processing software 86 and the one or more processors 88 may be integrated with the phosphorescence imager 40.

[0043] The phosphorescence imager 40 is configured to capture a plurality of raw phosphorescence intensity images 90 of the implantable sensor 20. With each pulse of the one or more light sources 42, a plurality of raw phosphorescence intensity images 90 are obtained. As explained herein, in the tested phosphorescence imager 40 this includes raw phosphorescence intensity images 90 taken at times t1-t11. This process is repeated for each2024-260-2 pulse of the one or more light sources 42. The image processing software 86 is configured to generate one or more phosphorescence intensity images 92 of the one or more test regions 24 and the one or more control regions 26 as applicable. The phosphorescence intensity image 92 at any one measurement point is generated using a plurality of images taken at time t=0 immediately after the light source 42 is switched to the OFF state (i.e., within 1 microsecond). From the plurality of t=0 images (e.g., ten such images in one embodiment), the corresponding background image for each image (which is taken along each set of raw phosphorescence images 90) is subtracted with the image processing software 86. This is followed by the image processing software 86 performing pixel binning to adjust resolution and pixelwise averaging of the plurality of images to generate the final phosphorescence intensity image 92 that is used by the first neural network 96 as disclosed herein. This process is illustrated in FIG.4B. As explained herein, a plurality of phosphorescence intensity images 92 may also be generated when the phosphorescence imager 40 takes images over a period of time.

[0044] In some embodiments, the image processing software 86 is also configured to generate one or more phosphorescence lifetime images 94. The image processing software 86 further includes a first neural network 96 (also referred to herein as CNNAlignment) that is trained to classify the implantable sensor 20 as properly aligned or misaligned with respect to the phosphorescence imager 40 (or vice versa). During reading, the phosphorescence imager 40 is placed over the location where the implantable sensor 20 is located within the tissue 110. There may be instances where the phosphorescence imager 40 is not properly aligned relative to the implantable sensor 20 and this first neural network 96 is used to determine whether or not the alignment is within and acceptable zone or range needed for accurate measurements.

[0045] In some embodiments, the one or more generated phosphorescence intensity images 92 and / or the raw phosphorescence intensity images 90 are input to the first trained neural network to classify the implantable sensor 20 as properly aligned or misaligned. In one preferred embodiment, the one or more generated phosphorescence intensity images 92 are used for this alignment but it could be just the raw phosphorescence intensity images 90 or a combination of both.

[0046] The image processing software 86 further includes a second neural network 98 (also referred to herein as CNNClass) that is trained to generate a qualitative or quantitative output of the glucose level of the subject in response to inputs of one or more of the raw2024-260-2 phosphorescence intensity images 90, the one or more generated phosphorescence intensity images 92 and / or the one or more generated phosphorescence lifetime images 94. In a preferred embodiment, the one or more generated phosphorescence lifetime images 94 are used as the input to the second neural network 98 as seen in FIGS.3C and 6A. Examples of qualitative outputs generated by the second neural network 98 include glucose readings of low, normal, or high. Examples of quantitative outputs generated by the second neural network 98 include actual measured glucose levels or ranges of levels (e.g., 150 mg / dL).

[0047] In one embodiment, a Sweep Delta T button as seen in FIGS.17A, 17B is included in the GUI 84 that, when activated, the continuous operation of the phosphorescence imager 40, sequentially capturing sets of raw timelapse phosphorescence images 90 (e.g., a plurality of timelapse images + 1 background image) with user-specified time intervals in between captures. Furthermore, the GUI 84 may contain a screen to display phosphorescence intensity and lifetime image measurements.

[0048] The system 10 described herein may be used to sense glucose levels in a subject using an implantable sensor 20. The sensor 20 is implanted or inserted in the skin tissue 110 of the subject and the implanted / inserted sensor 20 is then imaged with the phosphorescence imager 40 where a plurality of raw phosphorescence intensity images 90 of the one or more test regions 24 and, optionally, the one or more control regions 26 are acquired. A phosphorescence intensity image 92 is generated from the plurality of raw phosphorescence intensity images 90 using the image processing software 86. A phosphorescence lifetime image 94 of the one or more test regions 24 and the one or more control regions 26 of the sensor 20 is generated from the plurality of raw phosphorescence intensity images 90 using image processing software 86. The generated phosphorescence intensity image 92 is input to a first neural network 96 that is trained to classify the implantable sensor 20 as properly aligned or misaligned. Alternative, in other embodiments the plurality of raw phosphorescence intensity images 90 alone or with the generated phosphorescence intensity image 92 are input to a first neural network 96. The user may be notified whether the implantable sensor 20 is properly aligned or misaligned with the phosphorescence imager 40. For example, different tones or sounds may be emitted by the phosphorescence imager 40 or the computing device 80 that correspond to whether the sensor 20 is properly alignment or misaligned. Visual cues could also be given to the user via the GUI 84 or lights on the phosphorescence imager 40. Haptic responses (e.g., vibrations) of the phosphorescence imager 40 may also indicate whether the sensor 20 is properly alignment or misaligned.2024-260-2

[0049] One or more of the raw phosphorescence intensity images 90, the generated phosphorescence intensity image 92 and / or the generated phosphorescence lifetime image 94 are input to a second neural network 98 that are trained to generate a qualitative or quantitative output of the glucose level of the subject. In one preferred embodiment, the just the phosphorescence lifetime image 94 is used to generate a qualitative or quantitative output of the glucose level of the subject. For example, this may include a qualitative output such as low, normal, and high for glucose concentrations (e.g. FIG.3C). Alternatively, or in addition to, the output may include a numerical indication of glucose concentration. This may be expressed as a single number or a range. The qualitative and / or quantitative output form the second neural network may be displayed on the GUI 84 on the display 82.

[0050] Experimental

[0051] Results and Discussion

[0052] Design of the insertable phosphorescence-based glucose sensor

[0053] The insertable phosphorescence-based sensor 20 was fabricated from a hydrogel (PEGDA) and had a “barcode” structure, featuring four spatially separated discrete compartments filled with glucose-sensitive test regions 24 and glucose-insensitive control regions 26 formed as channels 28 and positioned in duplicates. These assays utilize palladium-based phosphorescent dyes (i.e., palladium benzoporphyrins [PdBP]) with typical emission lifetimes in the 50-250 µs range (FIG.4B), three orders of magnitude longer than the maximal autofluorescence lifetime of endogenous fluorophores like melanin and collagen, which do not exceed 10 ns. This enables superior temporal separation of the sensor responses from the native fluorescence, leading to better SNR of these phosphorescence- based sensors 20 compared to other luminescence-based designs, and therefore allowing for deeper sensor operation from subcutaneous tissue layers where ISF is most abundant.

[0054] In addition to the phosphor dyes, the test regions 24 also contain glucose-specific enzymes (i.e., glucose oxidase and catalase), making phosphorescence intensities and lifetimes affected by the local changes in the glucose level. In contrast, the control regions 26 only contain phosphors and provide constant signals, regardless of the glucose concentration in the local environment. The insertable sensors 20 demonstrated stable phosphorescence lifetime response for up to 12 weeks, with enzyme activity maintained at over >80% for up to 4 weeks. In addition, it showed good biocompatibility for up to 7 months in real-world environments within the animal (i.e., pig) body. The combination of high SNR, small2024-260-2 footprint and biocompatibility renders these insertable phosphorescence-based sensors 20 a promising signal transduction probe for next-generation CGM systems.

[0055] Design of the phosphorescence imager (PI) and image processing pipeline

[0056] Along with the development of phosphorescence-based insertable sensors 20, appropriate readout hardware is also needed for a robust glucose concentration inference with resilience against potential misalignments of the wearable reader under real-life operating conditions. Toward this goal, a compact and cost-effective phosphorescence imager 40 was created, which captures phosphorescence intensity images 92 and phosphorescence lifetime images 94 of an insertable / implantable glucose sensor 20 and utilizes these images for misalignment-resilient inference of glucose levels through the skin 110. The phosphorescence imager 40 has a compact footprint (~5 cm x 5 cm) and includes an imaging system with a red excitation LED (633 nm) as the one or more light sources 42 operating in a pulse mode and optical filters (50, 58) and optics as described herein, accommodating the spectral properties of the phosphors (630 nm excitation / 810 nm emission). The imaging system has a sensing FOV of ~9.6 mm x 7.2 mm, allowing the capture of the whole sensor image. Excitation intensity was adjusted to ~10-fold below the American National Standards Institute (ANSI) safety exposure limit for safe operation. For each measurement, the phosphorescence imager 40 captured ten (10) successive phosphorescence images, utilized to generate the phosphorescence intensity and lifetime images 92, 94 of the sensing FOV, which were further processed by a first neural network 96 for misalignment-resilient glucose level inference with a repetition rate of ~1 min. For each phosphorescence lifetime image 94, the reader acquires 11 timelapse phosphorescence images, evenly distributed within 200 µs time interval, starting immediately after the excitation LED light source 42 turns off, to capture the decay of the sensor emission through the skin. See FIG.4B. These 11 timelapse images are processed by image processing software 86 to generate the phosphorescence intensity image 92 and phosphorescence lifetime image 94 of the sensing FOV from the raw phosphorescence images 90. Lifetime responses of each implanted sensor 20 are derived by averaging the pixels within rectangular masks superimposed on each of the four (4) sensor channels (two (2) test channels and two (2) control channels), yielding four (4) lifetime values per implanted sensor barcode behind the skin phantom. Typical lifetime response values for the test regions 24 ranged from ~80 µs for DI water (i.e., 0 mg / dL glucose concentration) to ~170 µs for high glucose concentration levels (≥150 mg / dL) and stayed around 180-190 µs for the2024-260-2 control regions 26. Details of the in vitro testing of insertable glucose sensor responses to different glucose concentrations are reported below.

[0057] In vitro testing of insertable glucose sensors quantified using PI

[0058] The in vitro performance of the insertable sensors 20 was characterized and the custom-designed PI using glucose-spiked DI water samples as described in the Methods section. For these experiments, a skin phantom was utilized (1 mm, Type 1-2) to simulate the optical properties of a bulk skin tissue 110 (FIG.3A). Leveraging the imaging-based design of the PI system 10, sensor responses were measured through the skin phantom at different locations within the reader FOV. For the analysis of the misalignment tolerance of the system 10, two regions of interest (ROIs) were defined, namely aligned and misaligned. The aligned region was selected as a 3.2 mm x 3.4 mm rectangle centered around the center of the FOV of the phosphorescence imager 40, while the misaligned region was set as the remaining area, outside of the aligned region (FIG.5).

[0059] In an ideal PI operation scenario, the insertable sensor 20 would maintain a fixed position at the center of the FOV of the phosphorescence imager 40 (i.e., the ideal location), exhibiting stable excitation power and constant imaging condition. However, real-life operation scenarios may involve random wearable reader misalignments, leading to fluctuations in the effective sensor excitation profile and variability in the sensor positioning within the PI reader FOV, which may alter the captured sensor measurements. In this context, the aligned region represents an acceptable misalignment-tolerant zone: as long as the sensor 20 remains within this zone, the phosphorescence imager 40 captures sensor measurements for accurate inference of glucose levels despite random deviations from the ideal location (i.e., FOV center). Otherwise, if the sensor 20 falls outside of the aligned region, i.e., into the misaligned area, using a trained neural network 96, the phosphorescence imager 40 detects it and prompts the user for re-alignment to mitigate potential readout variabilities induced by large misalignments.

[0060] Prior to the development of the neural network-based analysis for glucose level inference, the responses of the insertable sensor 20 were first tested through a 1-mm thick skin phantom at different locations within the phosphorescence imager 40 FOV. FIG.5 depicts sensor lifetime responses from three (3) representative locations within the reader FOV, including 1 ideal location (i.e., the center of the aligned region), 1 non-ideal location (i.e., at the border of the aligned region) and 1 misaligned location within the misaligned region. The phosphorescence lifetime responses of the test regions 24 at both the ideal and2024-260-2 non-ideal locations within the aligned zone exhibit a strong correlation with glucose concentrations in the 0-150 mg / dL range. In the meantime, the control regions 26 at both locations demonstrate stable responses, independent of the glucose concentration as expected, with less than 5% CV (i.e., intra-sensor CV), validating consistent sensor operation (FIGS.5, 7A and 7B). In addition, inter-sensor variability, assessed between the lifetime responses of 8 (Nsens) tested implantable sensors 20, shows less than 15% CV (i.e., inter-sensor CV) for all three glucose levels (low / normal / high), indicating a repeatable PI operation within the aligned region. The comparable performances of the sensor 20 responses from both the ideal and non-ideal locations within the aligned region further justify the definition of the alignment zone.

[0061] In contrast to the aligned region, only a single test region 24 (Test 1) from the misaligned sensor shows a strong correlation with the spiked glucose concentration levels (FIG.5). The second test region 24 (Test 2) behaves similarly to the control regions 26, providing a constant response regardless of the glucose level. This undesired behavior can be attributed to the low SNR of this misaligned test region 24 and the interference of the emissions from the control regions 26, scattered by the skin phantom layer. Additionally, inter-sensor repeatability for the misaligned location is lower compared to aligned locations, with inter-sensor CV exceeding 15%. This inferior performance of the insertable sensor 20 from the misaligned location underscores the need for alignment control for the wearable phosphorescence imager 40 to achieve robust inference of glucose levels.

[0062] Both the lifetimes and intensities of the phosphorescence emission are influenced by the local changes in the glucose concentration. Therefore, in addition to looking at the lifetime responses, the intensity responses were explored as an alternative detection modality. It was observed that lifetime signals exhibit significantly lower inter-sensor variability compared to the intensity signals, making phosphorescence lifetime the preferred modality for consistent and accurate inference of glucose levels through the skin. For instance, for the DI water samples (i.e., 0 mg / dL glucose), the inter-sensor CV for the intensity responses exceeded 25% and 15% for the test and control channels, respectively, while the CVs of the lifetime responses remained under 11% for all the channels (FIGS.8A-8D). Moreover, the average CVs of the intensity responses to a glucose concentration sweep for the locations depicted in FIG.5 exceeded 50% for all three locations, including the aligned cases (i.e., ideal and non-ideal) (FIGS.9A-9C). This observed higher variability in the phosphorescence intensity responses may originate from the manual fabrication of the insertable sensors 20 or2024-260-2 variations in the ambient environment (i.e., the oxygen concentration in DI water samples), which primarily affect phosphorescence intensities rather than lifetimes. However, it should be appreciated that in some embodiments, the phosphorescence intensity image 92 may be used instead of the phosphorescence lifetime image 94.

[0063] The linearity and repeatability of the sensor responses were also tested under different levels of pixel binning. Captured phosphorescence lifetime images 94 without pixel binning consisted of 800x600 pixels with a pixel size of 9 µm, resulting in ~110x90 pixels per channel of each insertable sensor 20. A baseline pixel binning size of N=8 was used (i.e., ~14x11 pixels per channel) for the in vitro testing results reported herein. However, sensor responses were investigated for much larger pixel binning sizes and observed a consistent sensor performance with < 15% inter-sensor CV up to N=30, which corresponds to ~4x3 pixels per test channel (FIGS.10A-10C). Therefore, lower-resolution imaging systems with ~26x20 pixels and larger photodetector sizes (i.e., 270 µm) could potentially be utilized for the PI design.

[0064] As illustrated in FIG.5, the insertable sensors 20 showed a good linearity up to a glucose concentration of ~150 mg / dL; however, for higher glucose concentrations, the response was saturated, likely indicating an enzyme-limited behavior of the sensors 20. A higher glucose diffusion rate at higher glucose concentrations might be saturating the enzyme activity; therefore, increasing the enzyme concentration may help to extend the linear range of the glucose sensors 20. Alternatively, employing a larger number of diffusion control crosslinked polyelectrolyte layers (i.e., CX layers) can slow down the glucose diffusion into the sensor matrix, consequently improving the sensor linearity. Due to a tradeoff between the number of CX layers and the oxygen concentration in the environment, the number of CX layers was not increased beyond 7 to maintain optimal testing conditions for the ambient oxygen environment (~21%) used in this study (see FIGS.11, 12A, 12B). Nevertheless, for testing at lower oxygen environments (e.g., ~5%), increasing the number of CX layers can yield a more linear sensor response across a larger glucose concentration range.

[0065] Neural network-based analysis of PI data for inferring glucose concentration levels

[0066] The PI system 10 used deep learning and convolutional neural network (CNN) models for the misalignment-resilient inference of glucose concentration levels from the phosphorescence intensity and phosphorescence lifetime images 92, 94. The neural network- based analysis consisted of two separate CNN models or neural networks 96, 98: the2024-260-2 alignment model 96 (i.e., CNNAlignment), which was used to assess the alignment of the phosphorescence imager 40 with respect to the position of the sensor 20 underneath the skin phantom, and the classification model 98 (i.e., CNNClass) utilized to classify glucose concentrations between low, normal, and high ranges (FIG.6A). CNNAlignmentautomatically assessed the alignment of the phosphorescence imager 40 location by analyzing phosphorescence intensity images 92 and identifying the sensor location within the image. If CNNAlignment located the sensor 20 within the aligned region (i.e., 3.2 mm x 3.4 mm, defining the misalignment-tolerant zone), it predicted that the phosphorescence imager 40 was properly aligned; conversely, if the sensor location fell into the misaligned region, the network 96 predicted misalignment of the phosphorescence imager 40. In misaligned cases, the phosphorescence imager 40 prompted the user for re-alignment, and the CNNClass was not used for such measurements. If and only if CNNAlignmentidentified that the phosphorescence imager 40 was properly aligned, the measured sample was further processed by the CNNClass, which utilized phosphorescence lifetime image 94 measurements for glucose level classification. In some other embodiments, the CNNClass may also receive the generated phosphorescence intensity image 92. This may be separate from the phosphorescence lifetime image 94 or in addition to the phosphorescence lifetime image 94. Based on this information processing pipeline, neural network-based glucose level inference accommodated up to 4.7 mm reader misalignment.

[0067] These alignment and classification network models were trained and optimized separately using a total of 680 samples comprising 200 aligned and 480 misaligned samples from 4 different insertable sensors, each tested on 10 glucose concentrations in the range of 0-250 mg / dL. To generate a diverse dataset representing various reader alignment scenarios, each glucose concentration was measured at 17 different sensor locations within the PI reader FOV, including 5 aligned (1 ideal + 4 non-ideal) and 12 misaligned locations. The CNNAlignment network 96 was trained and optimized using all these 680 measurements, achieving 100% classification accuracy on the validation set. CNNAlignmentaccurately identified aligned and misaligned locations, benefitting from the spatial features present in the phosphorescence intensity images. At the same time, the CNNClassnetwork 98 was solely trained on 200 aligned samples since it would only be used if CNNAlignment identified that the reader was properly aligned; the optimal CNNClassmodel exhibited 89.3% accuracy on the validation set for the glucose level classification, including an accuracy of 93.7% and 87.5% for the ideal and non-ideal sensor locations, respectively.2024-260-2

[0068] After this training stage, the optimized models were blindly tested on 672 new measurements from 4 additional implantable sensors 20, never used during the training stage. CNNAlignment network 96 showed the same 100% accuracy on the blind samples, correctly identifying 198 aligned and 472 misaligned measurements (FIG.6B). For all measurements identified as misaligned by CNNAlignment, re-alignment of the phosphorescence imager 40 would be needed, and CNNClassnetwork 98 would not be applied to any of these samples. Consequently, only the measurements falling within the aligned region (i.e., 3.2 mm x 3.4 mm, misalignment-tolerant zone), as determined by the CNNAlignment, were further processed by the CNNClass network 98. The performance of CNNClass network 98 on these samples showed an overall accuracy of 88.8% for the classification of low, normal, and high glucose levels, revealing 92.5% accuracy for the ideal sensor locations and 87.3% accuracy for the non-ideal sensor locations, as illustrated by confusion matrices in FIG.6B.

[0069] To further explore the implications of some of the false predictions by CNNClass, these confusion matrices were analyzed and false predictions categorized into three classes based on the severity of their impact. In the context of the false model predictions on patient health, a misclassification between hypoglycemic (low glucose level) and hyperglycemic (high glucose level) events represents a severe error, leading to opposite patient treatment and exacerbating the patient's condition. Fortunately, the CNNClass did not confuse hypoglycemic and hyperglycemic events for any of the blinded test samples, as illustrated by the zero- containing cells of the confusion matrix. The second most critical error involves the misclassification of low or high glucose levels as normal, leaving abnormal glucose levels undetected. These cases are represented in the confusion matrices with the cells in the Prediction Norm row and Low and High (Ground Truth) columns. The optimal CNNClassmodel showed 93.1% sensitivity (i.e., 54 / 58 true positives) for detecting low glucose levels and 96.7% sensitivity (i.e., 58 / 60 true positives) for detecting high glucose levels, which confirms the competitive performance of the phosphorescence imager 40. Finally, the misclassification of normal glucose levels as low or high is another type of error. These predictions are depicted with cells Norm ground truth column and Low and High Prediction rows in the confusion matrices. During the model optimization, CNNClassnetwork models 98 were specifically prioritized with lower error rates for the two more severe error cases in order to minimize the negative impact of false predictions from the phosphorescence imager 40.2024-260-2

[0070] The presented CNN models benefit from the collective behavior of the highly multiplexed signals at the captured phosphorescence images 90 to achieve competitive accuracy for glucose level inference despite physical misalignments of the phosphorescence imager 40. To better assess the impact of multiplexing on the model performance, glucose level classification accuracy was compared using images with lower resolution, controlled by the pixel binning size (N). CNNClassdemonstrated accuracies of 84.8% and 83.8% when increasing N to 20 (40x30 pixels per sensor image) and 50 (16x12 pixels per image), respectively (FIGS.13A-13B); CNNAlignmentmaintained 100% accuracy in both cases. These glucose concentration classification accuracies for large N suggest that lower-resolution imaging systems can be employed to reduce the cost of the phosphorescence imager 40 without a significant impact on its performance. These findings are also in line with earlier results reported in FIGS.10A-10C, demonstrating a consistent sensor performance with < 15% inter-sensor CV up to N=30 pixel binning.

[0071] The size of the aligned region creates a tradeoff between the usability of the phosphorescence imager 40 and the accuracy of glucose testing. A smaller aligned region may increase testing accuracy. However, it requires a more precise positioning of the phosphorescence imager 40, leading to tedious manual alignment and reduced usability of the device. At the same time, large misalignments, particularly near the edge of the FOV of the phosphorescence imager 40, introduce additional excitation non-uniformities and aberrations, resulting in higher variabilities in captured sensor measurements and lower glucose testing accuracy. Evaluation of the same CNN models on locations beyond the baseline aligned zone outlined in FIG.5, revealed that the CNNClass accuracy did not exceed 70% for such regions, with the highest error rate occurring for locations outside of the aligned zone (see FIGS.14A- 14D). Therefore, the 3.2 mm x 3.4 mm aligned zone accommodating up to ~5 mm lateral misalignments presents a balanced compromise between the ease-of-use of the phosphorescence imager 40 and glucose testing performance, with a glucose concentration classification accuracy of 88.8%. In the future, the size of the aligned region can be extended by further increasing the imaging FOV and including more samples from different locations within the misaligned region in the training of the CNN model 96.

[0072] The usability of the sensor 20 can be improved by further miniaturization of the phosphorescence imager 40 to a watch-size footprint, which can be achieved by utilizing lens-less imaging designs or miniaturized systems. Finally, one can benefit from the multiplexing capabilities of the phosphorescence imager 40 and the multiplexed “barcode”2024-260-2 structure of the insertable sensors 20 to extend the platform to parallel monitoring of different analytes, making it a versatile wearable sensor for various biomedical sensing / monitoring applications.

[0073] A CGM system 10 was developed based on insertable phosphorescence-based sensors 20 paired with a compact and cost-effective phosphorescence imager 40 for robust measurement of glucose levels with resilience to random reader misalignments. In vitro testing of the CGM system 10 using a 1-mm thick skin phantom as a proxy for skin tissue 110 revealed 88.8% accuracy for the classification of glucose levels covering low, normal, and high concentration ranges accommodating up to 4.7 mm misalignment. Furthermore, the phosphorescence imager 40 exhibited 100% accuracy in identifying larger misalignments beyond the misalignment-tolerant region, prompting user intervention for reader re- alignment, and ensuring appropriate reader positioning. The misalignment resilient glucose level inference capability of the phosphorescence imager 40 operating through the skin 110, coupled with the small size and biocompatibility of the insertable sensors 20, make this approach an appealing candidate platform for continuous ISF-based glucose monitoring.

[0074] Methods

[0075] Fabrication of glucose-sensing alginate microparticles

[0076] Phosphorescent signals from glucose sensing alginate microparticles were generated by oxygen-sensitive ethyl cellulose nanoparticles (ECNP) containing palladium (II) meso-tetra(4-carboxyphenyl)tetrabenzo-porphyrin) (Frontier Specialty Chemicals) phosphorescent dye (PdBP). PdBP-ECNPs were fabricated using the previously developed nano-emulsion method. Glucose specificity in alginate microparticles was enabled by encapsulating PdBP-ECNP with glucose oxidase (GOx, Tokyo Chemical Industries) and catalase (Cat, Sigma-Aldrich, Inc.) using an emulsion technique. In brief, a 4% (w / v) sodium alginate aqueous solution (3.75 mL) was mixed with PdBP-ECNP (1.25 mL) suspension for 30 minutes by a nutating mixer. Separately, GOx (58.5 mg) and Cat (54.9 mg) were dissolved in 50 mM TRIS buffer (2.5 mL, pH 7.2, Sigma-Aldrich) by nutation. Further, the alginate mixture and the enzyme solution were mixed, generating a precursor solution. This solution was dropwise added and emulsified in a mixture of isooctane (10.8 mL, Avantor performance materials) with SPAN 85 (322 µL) using a homogenizer (8000 rpm). Further, isooctane (1.5 mL) and TWEEN 85 (175 µL) were added to the above mixture and stirred at the same speed for 15 seconds. During the last 50 seconds of the emulsification process, 10% (w / v) CaCl2 (Sigma-Aldrich) solution (4 mL) was added to induce external gelation of the alginate2024-260-2 microparticles. Next, the emulsion was transferred to a round bottom flask and gently stirred in a magnetic stirrer for 20 minutes. The microparticles were then centrifuged at 2000 g for two minutes and washed twice with DI water. Finally, the microparticles were coated with surface nanofilms of polyelectrolytes deposited using the layer-by-layer (LbL) assembly. The nanofilms consisted of crosslinked poly(allylamine hydrochloride) (PAH, Sigma-Aldrich) and poly(sodium-4-styrenesulfonate) (PSS, Sigma-Aldrich) polyelectrolyte bilayers (i.e., CX layers) and were designed for controlled diffusion of glucose inside the microparticles. The number of CX layers was optimized to 7 for the optimal sensor operation at the ambient oxygen environment (i.e., ~21%) used herein (FIG.11). The number of CX layers can be increased for testing at low-oxygen environments, which will reduce the glucose diffusion rate and increase the linear range of sensor responses.

[0077] Fabrication of insertable / implantable glucose sensors

[0078] The hydrogel-based sensor 20 was fabricated from PEGDA (i.e., polyethylene glycol diacrylate, Alfa Aesar) hydrogel using a previously reported soft lithography process. In short, first, the 20% (w / v) PEGDA was mixed with 2% (v / v) photo-initiator solution and dispensed into a PDMS bottom mold. Further, a PDMS top mold with four discrete compartments for the test regions 24 and control regions 26 was added squeezing the hydrogel solution between the PDMS top and bottom molds. Next, the squeezed hydrogel solution was placed under a UV lamp (360 nm, 10-15 mW / cm²) for 5 min to crosslink PEGDA monomers. The formed hydrogel was peeled off from PDMS molds and rinsed with DI water. PDMS top and bottom molds were fabricated by replica molding from 3D-printed master molds.

[0079] Two (test regions 24) out of the four hydrogel compartments were pipetted with test channels (0.64 µL each) and two others (control regions 26) with control channels (0.64 µL each). Test channel solution was prepared by mixing glucose-sensing alginate microparticles (see the Fabrication of glucose-sensing alginate microparticles section) with the hydrogel solution. The control channel solution was prepared by mixing oxygen- insensitive nanoparticles with the hydrogel solution. Oxygen-insensitive nanoparticles were fabricated by encapsulating palladium(II) tetramethacrylated benzoporphyrin (PdBMAPP, PROFUSA) within poly(vinylidene chloride-co-acrylonitrile) hydrogel. The hydrogel with pipetted test and control channels was polymerized under a UV lamp for 5 minutes. The total cost for one insertable sensor 20 was < $0.3 as seen in Table 1 below.2024-260-2 Table 1 Category Name Cost / 10 sensors ($) Phosphorescent PdBP dyes 0.0257

[0081] A custom LED driver circuit 46 was designed to operate the excitation LED in a pulse mode (FIGS.15A-15D). The circuit is based on a 1 A constant-current LED driver (Analog Devices Inc.) which contains internal pulse width modulation (PWM) circuitry 72 allowing for pulsed operation with custom pulse on / off times. The operation of the PMW pin was set by a timer circuitry consisting of an 8 MHz oscillator (Analog Devices Inc.), a 14- stage shift register (Onsemi), and a quadruple NAND gate (Texas Instruments). The oscillator generated square pulses with 125 ns (50% duty cycle) baseline period, and the shift register further divided baseline oscillation frequency, yielding square pulses with 2,000 µs and 500 µs periods. Finally, a quadruple NAND gate combined these two square pulses to generate the desired waveform with 250 µs (ton) and 1000 µs (toff) on and off times, respectively. The waveform shape can be adjusted to accommodate different lifetime ranges by changing the baseline oscillation frequency or the output pins of the shift register. The constant-current driver 70 was powered by an external supercapacitor 74 to supply the required input voltage (i.e., VIn = 6V). The LED light source 42 was connected to GND through an N-channel MOSFET (Vishay Siliconix), enabling fast switching between on / off cycles. The LED driver circuit 46 was powered by the output supply voltage pin (5 V, 100 mA) of the PI camera 44. The circuit was designed using Altium circuit design software.2024-260-2

[0082] The excitation time (i.e., texcitation), determined as the total LED activation time used for excitation of the phosphors to generate timelapse phosphorescence images 90, was optimized to texcitation = 70 ms (FIGS.16A-16E). During this time, the LED light source 42 was operated in a pulse mode, and the phosphorescence imager 40 periodically captured phosphorescent responses, accumulating emissions over multiple excitation and emission cycles, to achieve reliable SNR. Shorter excitation times resulted in a low SNR of the captured phosphorescence intensities, while texcitation > 70 ms depleted the voltage at the constant-current driver, leading to fluctuations in the LED output power.

[0083] PI operation

[0084] A custom-designed graphical user interface (GUI) 84 was developed to control the operation of the phosphorescence imager 40 (FIGS.17A-17B). The GUI 84 contained input fields to specify the operation parameters of the PI camera 44, including shutter parameters, exposure time and gain. For each successive phosphorescence lifetime image 94 measurement, the camera shutter was programmed to capture up to 11 timelapse phosphorescence images 90 and the background image after the LED “on” cycle. For each of these images, the camera shutter was activated periodically, and the user could specify both the shutter activation times and the shutter on / off time periods. The shutter activation times for the 11 images were adjusted to cover a 200 µs time interval after the LED on cycle with a 20 µs gap between subsequent activations, i.e., t1= 0 µs, t2= 20 µs, t3= 40 µs, … t11= 200 µs. The shutter activation time for the background image was set to tb = 650 µs after the LED on cycle. The shutter on / off times were further tailored to match the LED duty cycle (FIG. 4B). Additionally, the camera exposure time for each image was set to match the excitation time (i.e., texcitation= 70 ms). Camera gain was set to 1. The operation parameters of the phosphorescence imager 40 can be customized to accommodate different lifetime ranges, depending on the emission decay properties of the phosphors.

[0085] The Sweep Delta T button in the GUI 84 activated the continuous operation of the phosphorescence imager 40, sequentially capturing sets of timelapse phosphorescence images 90 (i.e., 11 timelapse images + 1 background image) with user-specified time intervals in between captures. Furthermore, the GUI 84 contained a display or screen 82 to display phosphorescence intensity and lifetime image measurements. The GUI 84 had two displaying modes, namely Camera mode and File model. In the Camera mode, the GUI 84 displayed real-time images directly captured by the camera, while in the File mode, the displayed images were transferred from a user-specified path. Continuous operation of the reader was2024-260-2 set to capture 10 repeated lifetime images within ~45 s time interval and 1 s time gap (Δt) between subsequent images. These images were automatically processed on a benchtop computer to generate final phosphorescence intensity and lifetime images 92, 94, which were used in the neural network-based glucose level analysis. The number of phosphorescence intensity images 90 used to generate the final phosphorescence lifetime image was optimized to 6 out of 11, corresponding to optimal linearity, repeatability, and dynamic range of the sensor responses (FIGS.18A-18E). Total glucose inference time, including image data capture, image processing and deep learning-based analysis, was ~1 min per measurement.

[0086] PI design

[0087] The phosphorescence imager 40 consists of the LED driver circuit 46, camera 44 (UI-3250CP-M-GL Rev.2, IDS) and off-the-shelf optical components arranged within a 3D printed case which operated as a housing (~5 cm x 5 cm footprint). The 3D-printed case / housing was printed by the Object 30 printer (Stratasys). The reader contains excitation and emission channels, accommodating spectral properties of the PdBP dyes (630 nm excitation / 810 nm emission). The excitation channel contains an LED light source 42 (633 nm peak wavelength, Osram) mounted to the custom driver circuit 70, collimating lens 48 (5.0 mm Dia. X 5.0 mm FL, Edmund Optics), excitation filter 50 (632 nm BP, Edmund Optics), and focusing lens 52 (12.0 mm Dia. X 12.0 mm FL, Edmund Optics). The emission channel consists of a camera 44, imager lens 60, emission filter 58 (725 nm LP, Edmund Optics), cold mirror 54 (45° AOI, 12.5 mm Square, Edmund Optics), and sample lens 56 (15.0 mm Dia. X 20.0 mm FL, Edmund Optics). The cold mirror 54 directs collimated light from the LED 42 towards the sample plane, generating uniform illumination over a circular area with ~1 cm diameter. Excited phosphorescence emission from the insertable sensor 20 is collected by the emission channel, passing through a two-lens (56, 60) imaging system to generate phosphorescence images of the sensor 20 (FIGS.1B and 4A). The two-lens system has ~0.75x demagnification, generating 9.6 mm x 7.2 mm FOV. The phosphorescence imager 40 was designed for a simple mount on the Z-axis stage for in vitro testing using glucose-spiked solutions. The cost of an assembled phosphorescence imager 40 was ~$1170, including the camera 44 as seen in Table 2.2024-260-2 Table 2 Category Name Cost ($) LED Driver circuit Electronic components 53.1 Printed Circuit Board 0.4

[0088] Generation of phosphorescence intensity and lifetime images

[0089] Image processing software 88 executed an automated image analysis algorithm generated the phosphorescence intensity 94 and lifetime images 96 from 10 successive images 90 (FIG.4B). At first, for each of these 10 images, the background image was subtracted from each of the 11 timelapse intensity images 90, and further performed pixel binning to adjust the image resolution. Next, the pixel-binned images were averaged pixel- wise over the 10 successive measurements, yielding a final set of phosphorescence intensity images 92. These images were used to construct the phosphorescence lifetime image 94: ^^^^,^^^^^^^^ = ^^0 ^^,^^^^െ^^^^ / ^^^^,^^,are the camera shutter activation times for the phosphorescence intensity images (i.e., t1 = 0 µs, …, t11 = 200 µs), ^^0 ^^,^^is the maximal phosphorescence intensity of the pixel located at index n, m, and ^^^^,^^is the phosphorescence lifetime of the same pixel. Pixelwise lifetime values at the phosphorescence lifetime image were calculated by solving a least squares optimization problem for ^^^^,^^: argmin‖ ^^^^^^^^^^^, ^ – ^^^^^^^^^ ^ – 1 / ^^^^,^^t‖,^^0 ^^,^^, ^^^^,^^ ^^ ^^ ^^,^^

[0091] where ^^^^^^^^^^^,^^^ = ^log^^^^,^^^^^^) log(^^^^,^^^^^2^) … log(^^^^,^^^^^11^)] are the pixel-wise phosphorescence intensities, ^^^^^^^^^^^ ^^,^^^ = log^^^0 ^^,^^^[1 … 1]1x11 and t =[^^1 … ^^11].2024-260-2

[0092] In vitro data collection procedures

[0093] The phosphorescence imager 40 and neural-network-based analysis of glucose levels were tested in vitro on glucose-spiked DI water samples, utilizing a skin phantom (1 mm, Skin Type 1-2) to simulate the optical properties of human skin tissue 110. During in vitro testing procedures, a syringe pump drove spiked glucose samples from a stock solution through a fluidic channel containing the sensor 20 at a constant flow rate of 20 µL / min. The skin phantom was located on top of the fluidic channel, fully covering the sensor 20 (inset on FIG.4A). The fluidic channel was mounted on a custom stage for precise XY alignment between the sensor 20 and the phosphorescence imager 40. The phosphorescence imager 40 was mounted on a Z-axis optical stage, and the height was adjusted to focus the phosphorescence imager 40 on the sensor plane. The phantom experiments were conducted at room temperature (i.e., 22.8±0.7 °C). Note that the stage and pump was only used as part of the in vitro experiments. These components are not part of the system 10.

[0094] A total of eight (8) insertable sensors 20 were tested in vitro at ten (10) different glucose concentrations in the 0-250 mg / dL range (i.e., 0, 50, 65, 80, 95, 110, 125, 150, 200, 250 mg / dL).250 mg / dL glucose sample was prepared by adding 125 mg of glucose (Sigma- Aldrich) into 50 mL DI water. Other glucose samples were prepared by ratiometric mixing of 250 mg / dL sample with DI water. A 15-minute waiting period was used to allow for enzyme activation and signal ramp-up after switching between different glucose levels.

[0095] To collect a comprehensive dataset for the neural network-based analysis, each sensor / glucose concentration was tested at seventeen (17) discrete locations within the imager FOV, including five (5) aligned locations (1 ideal + 4 non-ideal) and twelve (12) misaligned locations arranged into a 4x4 grid (FIG.19). In between subsequent measurements, the custom XY axis stage automatically translated the fluidic channel with the sensor 20 to different locations. The time delay between subsequent data measurements was set to 30 s to allow for the full channel refill with fresh glucose solution. Each data measurement involved 10 phosphorescence lifetime image measurements over ~ 45 s time interval with Δt = 1 s time gap between subsequent measurements.

[0096] Fluidic channel design and assembly

[0097] The fluidic channel was assembled from acrylic sheets (1.5 mm [SOURCEONE] and 1 mm [SimbaLux]), transparency (Apollo) and double-sided tape (3M). The channel shape was precisely cut by a laser cutter (Trotec), and the channel was assembled by stacking 1.5 mm acrylic (bottom part, i.e., floor), 2 transparency sheets, 1 mm acrylic (main part, i.e.,2024-260-2 channel) and 1 transparency sheet (top part, i.e., ceiling). Double-sided tape was used to glue together the adjacent channel layers. The assembled channel had a size of 75x25x2.8 mm3.

[0098] XY axis stage design

[0099] The XY axis stage was designed to automate sensor misalignments during the in vitro testing of the phosphorescence imager 40. The stage contained 2 stepper motors connected to lead screws, enabling X and Y translation in discrete steps with a 10 µm step size. Stepper motors were controlled by Arduino through stepper motor driver modules. Lead screws were connected to a fluidic channel holder allowing for lateral translation of the channel up to 5 cm along each axis (FIG.4A).

[0100] A custom GUI was developed to facilitate the operation of the XY axis stage. The GUI showed the current stage location within the operation range and allowed the user to manually move the stage along the X and Y axes with 2 mm discrete steps. The GUI also enabled the user to align the stage to its origin location (i.e., the default origin location was adjusted to the minimal translation value) and update the origin location with user-controlled values. Finally, the GUI allowed to move the stage based on imported shift values from an external file source. Stage operation code and GUI were developed in Python, using the Tkinter package to build the GUI and the Serial package to communicate with Arduino. For the automated misalignment data collection with the phosphorescence imager 40, the stage operation code was integrated into the PI reader operation code.

[0101] Neural network-based analysis of PI data

[0102] Neural network-based sensor analysis was implemented to (1) assess proper alignment of the phosphorescence imager 40 relative to the sensor 20 and (2) classify glucose concentrations between low, medium, and high ranges for the properly aligned samples. The neural network-based analysis consisted of two separate CNNs for the alignment (CNNAlignment) and the classification (CNNClass) tasks (FIG.6A). The alignment network 96 was trained using phosphorescence intensity images 92, while the classification network 98 was trained utilizing phosphorescence lifetime images 94. Phosphorescence intensity images 92 were not used to train the classification network / model 98 due to the high inter-sensor variability of phosphorescence intensity signals although it should be appreciated that in other embodiments phosphorescence intensity images 92 may be used for training. Prior to input into the CNNClass98, the phosphorescence lifetime images 94 were cropped to center the sensor location within the image and standardized by the maximal lifetime value (i.e., 2502024-260-2 µs); higher lifetime values at phosphorescence lifetime images 94 were set to 250 µs to reduce background noise.

[0103] The CNNAlignment 96 contained two (2) convolutional layers (32 units with 3x3 kernel and 16 units with 2x2 kernel) followed by a flattening layer and two fully-connected layers (128 and 32 units) as illustrated in FIG.6A. Both convolutional and fully-connected layers had ‘ReLU’ activation functions and L2 regularization. The output layer had 1 unit with a ‘sigmoid’ activation function. The network used a binary cross-entropy loss complied with the Adam optimizer, a learning rate of e-3and a batch size of 5. Binary cross-entropy loss (^^^^^^^^) is defined as:^^^^^^^^ ൫^^, ^^′൯ ൌ െ1 ^^^^ ∑^^^^^^ൌ1൫^^^^ log൫^^′ ^^൯ െ ൫1 െ ^^^^൯ log൫1 െ ^^′ ^^൯൯ , samples located within aligned region, and “0” for samples located within the misaligned region), ^^′^^are the predicted labels and ^^^^is the batch size.

[0105] The CNN 98 consisted of three (3) convolutional layers (512 units withkernel, 256 units 3x3 kernel and 128 units with 3x3 kernel) followed by a flattening layer and two fully-connected layers (128 and 32 units). All layers had ‘ReLU’ activation functions and used L2 regularization. The output layer had three (3) units with ‘sigmoid’ activation functions. The loss function for this network was the categorical cross-entropy loss compiled with Adam optimizer, a learning rate of e-4and a batch size of 5. Categorical cross- entropy loss (^^CC^^) is defined as:^^^^^^^^ ൫^^, ^^′൯ ൌ െ1 ^^^^ ∑^^^^^^ ൌ 1 ∑3 ′ ^^ ൌ 1^^^^^,^^ log ^^^ ^^,^^^^ ,class (i.e., 1 – low, 2 – normal, and 3 – high), ^^ are ′ ^^,^^the ground truth classification labels, ^^^^,^^are the predicted classification labels and ^^^^is the batch size.

[0107] The alignment neural network 96 was optimized and trained on a training / validation set with a total of 680 measurements, including four (4) different sensors 20, ten (10) glucose concentrations per sensor and seventeen (17) locations per concentration. CNNAlignment 96 achieved 100% accuracy on the validation data through 4-fold cross- validation, correctly identifying all the aligned and misaligned locations. The network 96 was further blindly tested on a testing set composed of 672 samples (i.e., 4 sensors x 10 glucose concentrations x 17 locations; the number is <680 due to the unexpected experiment2024-260-2 interruptions / failures for a few glucose concentrations) and achieved a blind testing accuracy of 100%.

[0108] CNNClass 98 was solely trained on sensors 20 located within the alignment region, and the total number of training / validation samples was 200 (i.e., 4 sensors x 10 glucose concentrations x 5 aligned locations). CNNClass 98 achieved 89.3% accuracy on the validation data, including an accuracy of 93.7% and 87.5% for the ideal and non-ideal locations, respectively. The blind testing set included 198 samples, and CNNClass 98 achieved 88.8% blind testing accuracy, including accuracies of 92.5% and 87.3% for the ideal and non-ideal locations, respectively.

[0109] Training times for CNNAlignment96 and CNNClass98 were 5 min and 30 min respectively. Blind testing times of trained CNNAlignment 96 and CNNClass 98 models were considerably lower, with ~20 ms per measurement. Data preprocessing to generate phosphorescence intensity and lifetime image sets was done in MATLAB 2023b, and the training / testing of the neural networks was performed in Python, using OpenCV and TensorFlow libraries. Training / testing of the neural networks was done on a desktop computer with a GeForce GT 1080 Ti (NVIDIA).

[0110] Skin phantom fabrication

[0111] To create tissue-simulating optical phantoms, PDMS (SylgardTM 184 Silicon Elastomer Kit) was used as the base, and TiO2and carbon black (Carbon, mesoporous nanopowder, Millipore Sigma) were added as the scattering and absorbing agents, respectively. To fabricate a 1-mm thick phantom mimicking lighter skin type (i.e., Type 1-2, Fitzpatrick scale has an absorption coefficient [µa] of 0.04-0.14 cm-1and a reduced scattering coefficient ^^^′-1^^] of 10-15 cm ), 2 grams of PDMS were weighed onto a 2-inch diameter aluminum mold. A 0.0025% (w / w) carbon black to PDMS and 0.05% (w / w) of TiO2 to PDMS were then added to 0.2 grams of curing agent in a separate glass beaker. This beaker was sonicated for 30 min to break up TiO2 powder clumps and mixed using a glass rod to create a homogeneous mixture. The TiO2 and carbon black mixture was then transferred to the PDMS in the aluminum mold and mixed thoroughly using a glass rod to ensure a homogenous mixture. Next, a vacuum chamber was used to degas the mixture in the mold for 30 min and remove any air bubbles trapped within the viscous media. The mold was then placed in a toaster oven and heated at 150 ºC for 4 min to get a completely cured phantom. The silicon phantom was carefully removed from the mold using a lab spatula. The finished disc phantoms were optically characterized using a laser source at 633 nm, a single2024-260-2 integrating sphere, and an inverse adding-doubling algorithm, yielding an absorption coefficient of 0.05 cm-1and a reduced scattering coefficient of 13.9 cm-1with anisotropy, g = 0.9. A rectangular slab was then cut out of this disc phantom for the in vitro testing of the insertable glucose sensors (FIG.4A).

[0112] While embodiments of the present invention have been shown and described, various modifications may be made without departing from the scope of the present invention. For example, different materials beyond alginate may be used for the microparticles. In addition, while the sensor 20 described herein is implantable or insertable, in other embodiments, the sensor 20 may be a wearable sensor such as a patch. For example, a patch with microneedles that function as the test / control regions 24, 26 may also be used as the sensor 20. The invention, therefore, should not be limited except to the following claims and their equivalents.

Claims

2024-260-2 What is claimed is:

1. A method of sensing glucose levels in a subject using an implantable sensor, the method comprising: implanting the sensor in the subject, the sensor comprising a substrate having one or more test regions, wherein the one or more test regions comprise phosphorescent nanoparticles, glucose oxidase, and catalase; imaging the implanted sensor with a phosphorescence imager that acquires a plurality of raw phosphorescence intensity images of the one or more test regions; generating one or more phosphorescence intensity images and one or more phosphorescence lifetime images of the one or more test regions from the plurality of raw phosphorescence intensity images; inputting the one or more generated phosphorescence intensity images and / or the raw phosphorescence intensity images to a first trained neural network that is trained to classify the implantable sensor as properly aligned or misaligned; and inputting one or more of the raw phosphorescence intensity images, the generated phosphorescence intensity images and / or the generated phosphorescence lifetime images to a second trained neural network that is trained to generate a qualitative or quantitative output of the glucose level of the subject.

2. The method of claim 1, wherein the sensor comprises one or more control regions comprising a phosphorescent dye and wherein the phosphorescence imager acquires a plurality of raw phosphorescence intensity images of the one or more test regions and the one or more control regions.

3. The method of claim 2, wherein the sensor includes a plurality of test regions and a plurality of control regions.

4. The method of claim 1, wherein the phosphorescent nanoparticles comprise ethyl cellulose nanoparticles (ECNP) containing palladium (II) meso-tetra(4- carboxyphenyl)tetrabenzo-porphyrin) phosphorescent dye (PdBP).

5. The method of claim 1, wherein the substrate comprises a hydrogel.2024-260-2 6. The method of claim 4, wherein the hydrogel comprises polyethylene glycol diacrylate (PEGDA).

7. The method of claim 1, wherein the one or more generated phosphorescence intensity images and / or the one or more phosphorescence lifetime images are generated by image processing software.

8. The method of claim 7, wherein the one or more generated phosphorescence intensity images are generated by image processing software that performs background subtraction, pixel binning, and / or image averaging.

9. The method of claim 1, wherein the user is prompted to adjust the position of the phosphorescence imager in response to a misaligned classification by the first trained neural network.

10. The method of claim 1, wherein the plurality of raw phosphorescence intensity images of the one or more test regions is obtained over a plurality of excitation and emission cycles.

11. A system for sensing glucose levels in a subject comprising: an implantable sensor comprising a substrate having one or more test regions, wherein the one or more test regions comprise phosphorescent nanoparticles, glucose oxidase, and catalase; and a phosphorescence imager that comprises one or more light sources, a camera or image sensor, and light source driving circuitry configured to pulse the one or more light sources, wherein the phosphorescence imager is configured to capture a plurality of raw phosphorescence intensity images of the implantable sensor in response to the pulsing of the one or more light sources.

12. The system of claim 11, wherein the sensor comprises one or more control regions comprising a phosphorescent dye.2024-260-2 13. The system of claim 11, wherein the phosphorescence imager or a computing device communicating therewith comprises image processing software configured to generate one or more phosphorescence intensity images and one or more phosphorescence lifetime images from the plurality of raw phosphorescence intensity images of the implantable sensor.

14. The system of claim 13, wherein the computing device executes a first trained neural network that is trained to classify the implantable sensor as properly aligned or misaligned based on the one or more generated phosphorescence intensity images and / or the raw phosphorescence intensity images and a second trained neural network that is trained to generate a qualitative or quantitative output of the glucose level of the subject based on one or more of the raw phosphorescence intensity images, the generated phosphorescence intensity images and / or the generated phosphorescence lifetime images.

15. The system of claim 11, wherein the sensor includes a plurality of test regions and a plurality of control regions.

16. The system of claim 11, wherein the phosphorescent nanoparticles comprise ethyl cellulose nanoparticles (ECNP) containing palladium (II) meso-tetra(4- carboxyphenyl)tetrabenzo-porphyrin) phosphorescent dye (PdBP).

17. The system of claim 11, wherein the substrate comprises a hydrogel.

18. The system of claim 17, wherein the hydrogel comprises polyethylene glycol diacrylate (PEGDA).

19. The system of claim 12, wherein the one or more test regions comprise separate test channels and the one or more control regions comprise separate control channels.

20. The system of any of claims 11-19, further comprising a display associated with the phosphorescence imager or the computing device having a graphical user interface (GUI) configured to control one or more parameters of the phosphorescence imager and / or2024-260-2 display the generated phosphorescence intensity image and the generated phosphorescence lifetime image or a qualitative or quantitative output of the glucose level of the subject.

Images