INTRODUCTION
Vaccines are exceptionally safe and effective, saving an estimated 2 million to 3 million lives annually (
1). However, each year, 1.5 million vaccine-preventable deaths occur due to undervaccination—primarily in areas of the developing world with poor health care infrastructure (
2). One key barrier to improving vaccination coverage in these regions is the inability to accurately identify the immunization status of infants given resource constraints, which can affect the quality of care provided (
3,
4). These areas often lack accurate medical recordkeeping systems and rely on vaccination campaigns to distribute vaccines. However, investigations in response to recent outbreaks of measles and mumps in the United States (
5), Australia (
6), and Italy (
7) have highlighted that poor immunization recordkeeping is not unique to developing nations.
Paper vaccination cards or certificates are the most widely used records in the developing world but are subject to error (
8) and possessed by only 60% of all households in low- and middle-income countries (
9). Without accurate vaccination records, health care professionals lack the data needed to make informed decisions about administering vaccines, often relying on parental recall (
10). This may result in the application of additional, unnecessary vaccine doses and therefore undue cost or, more problematically, missed opportunities to vaccinate, which leave the child at risk for contracting infectious diseases (
8,
10). As many as two-thirds of opportunities to vaccinate may be missed in some areas (
11), leading to a potential 30% drop in vaccination coverage (
3). Several solutions have been proposed including smartphone-based database applications (
12), fingerprinting (
13), and near-field communication chips (
14); however, these more technologically advanced methods have yet to achieve widespread adoption due to difficulty of implementation.
Therefore, we aimed to develop a platform that is robust, inexpensive, and easy to use to overcome the primary obstacles to implementation in the developing world. We hypothesized that information, such as vaccination history, could be encoded invisibly in the skin by applying a distinct pattern of near-infrared (NIR) fluorescent microparticles using a microneedle patch. By providing all of the requisite information on the patients themselves and delivering microparticles in the same microneedles as the vaccine, this platform offers several key advantages compared to traditional paper or electronic medical records including the following: (i) the elimination of reliance on widely accessible, yet secure database of patient information; (ii) lack of reliance on accurate patient identification and data entry by medical professionals; (iii) ability to make rapid determinations of vaccination status; (iv) elimination of vaccination fraud; (v) capacity for population-level assessment of vaccine coverage; and (vi) minimal cost that is feasible to implement in low-resource settings. By using a microneedle form factor, this platform should easily assimilate into the future vaccination landscape because microneedles are currently in development for several vaccines (
15) and have shown advantages such as antigen dose sparing, improved antigen stability, and ease of (self-)administration compared to traditional soluble injections (
16–
18). Microneedle patches for vaccination recordkeeping may not require cold chain storage and could potentially have long shelf life, which could greatly enhance feasibility of implementation in low-resource settings.
DISCUSSION
To maximize the utility of this technology for vaccination campaigns, we aimed to create a platform compatible with microneedle-delivered vaccines that could reliably encode data on an individual for at least 5 years after administration. In addition, this system also needed to be highly biocompatible, deliver a sufficient amount of dye after an application time of 2 min or less, and be detectable using a minimally adapted smartphone. Given the limitations of organic dyes (photobleaching and water solubility) and inorganic dyes [rare earth metal toxicity (
22), low fluorescence, and processing difficulty], QDs were an attractive option as a potentially bright, photostable, and tunable alternative. However, clinical implementation of QDs has been stymied by the toxicity of their core elements, such as cadmium and lead (
23). To overcome these safety concerns—which are especially important because these materials would be given to healthy children—we chose to synthesize custom QDs composed of more well-tolerated elements. We identified an optimal fluorescent emission range of 850 to 1100 nm. Operating in this range mitigates the potential cultural opposition to visible skin markings, reduces background from ambient light, minimizes light absorption by tissue at wavelengths below 850 nm (
24), and maximizes signal detection by avoiding the poor sensitivity of inexpensive silicon-based detectors to light above 1100 nm (
25). Although the use of fluorophores at the higher end of the NIR-II window (1000 to 1700 nm) would further improve tissue transparency and reduce background (
26), appropriate detectors for these wavelengths (indium gallium arsenide) are typically considered cost prohibitive (
27), which could render this approach infeasible for widespread implementation.
Fluorophore photobleaching under accelerated solar light exposure enabled us to predict signal loss associated with sun exposure over long periods of time. As expected, organic fluorophores degraded relatively quickly despite protection by heavily pigmented skin, which absorbs a large fraction of ultraviolet (UV) and visible light. These dyes also exhibited strong self-quenching in the dry state and therefore had to be studied in a dispersed, hydrated state. This self-quenching property is problematic for this microneedle-delivered recording system because it limits the packing density of the dye and thus its brightness. Alternatively, QDs performed very well under light exposure and have a larger Stokes shift, which enables them to avoid substantial reabsorption and therefore be used in a densely packed format.
Although the accelerated photobleaching assay does not perfectly mimic the real-world use case, it does provide evidence supporting the longevity of QDs compared to organic dyes. Further, this experimental setup may overestimate photobleaching because it assumes that skin is in direct sunlight every hour of the day that the sun is out. If administered at the thigh, where current vaccines are administered to infants, this area may be covered by clothing and/or in an area that is not exposed to direct sunlight for some portion of the daytime. In addition, photobleaching may also be enhanced in this experiment because sevenfold more intense light results in greater energy transfer per time, potentially enabling QD degradation to proceed more quickly.
PMMA was used as a nondegrading, encapsulating material for QDs to improve biocompatibility and enhance tissue permanence. The large size of microparticles (15.7 ± 5.3 μm) was also hypothesized to minimize clearance because previous studies have shown that larger particles are more resistant to clearance by macrophages, the most relevant cell type for clearing foreign material (
28). The flexibility of QD surface coating also enabled us to make QDs soluble in the same organic solvent as PMMA, which helped markedly enhance loading given that a traditional double emulsion technique (water-in-oil-in-water) cannot achieve 60% (w/w) loading.
We hypothesized that, unlike microneedle-delivered vaccines, which must simply break the skin’s water barrier to deliver their payload effectively (
29), these fluorescent microparticles must be delivered below the shed layers of skin to ensure their long-term residence in tissue. Therefore, we aimed to create needles 1500 μm in length. To successfully penetrate to this depth, the microneedles must resist mechanical stress, deflection, and displacement upon insertion into skin. To effectively deliver PMMA-encapsulated QDs into the skin, the microneedles must pierce the skin without fracturing. The critical load factor of 0.014557 was higher than the values reported for carboxymethyl cellulose microneedles using the same simulation procedures (
30). The final selection of α = 0.5, height = 1500 μm, and width = 300 μm avoids critically low mechanical robustness against all the major mechanisms of mechanical failure. Using a larger diameter and larger α increased the volume of material and thereby the number of QD-loaded particles that are available for transfer into the skin. The drawback of this is the potential increase in pain and higher penetration force requirements. However, although pain perception has been shown to increase with the diameter of hypodermic needles (
31–
33), the amount of pain induced by our microneedle patch is likely less than for traditional needles, as has been shown elsewhere (
34). Further, a previous study has shown that tip angle does not significantly affect pain upon needle insertion (
35), so the effect of α (effective tip angle) on pain perception was not considered. These needles are slightly smaller in diameter than a 30-gauge needle and therefore should minimize both pain and the force required for skin penetration (
36). Last, the lateral microneedle spacing results in a patch, and subsequently intradermal pattern, with a footprint of 0.25 cm
2. This spacing was sufficiently large to allow each needle to act independently and thereby allow force requirements to scale linearly with needle number (
37).
Over the past decade, smartphones have become ubiquitous in many areas of the world, including the developing world despite limited infrastructure (
38). Because these phones offer on-board processing power, camera applications, and inexpensive consumer-grade camera modules, we chose to adapt an existing smartphone to enable NIR imaging rather than build a completely new imaging system. In addition, we believe that familiarity with the function of these devices will lessen the learning curve for NIR imaging in a field setting. Whereas a stock Nexus 5X camera is built with a short-pass filter to prevent NIR light from affecting images, we wanted the exact opposite—the elimination of light in the visible range and passage of NIR light. In addition, we needed the new filters to block reflected light from LED illumination. Therefore, after stripping the stock short-pass filter, we added a pair of 850-nm long-pass filters, which would block both environmental light and LED illumination. A dielectric filter was used to impose a sharp cutoff at 850 nm and paired with an 850-nm color glass filter to eliminate the passage of visible light entering the filter at very small angles, which we observed to be problematic with the dielectric filter alone (
39). These optics were fit into a 3D-printed phone case customized for SM1-threaded components (1.035″-40) to fit with commercially available optical components.
Because both excitation and emission light can be absorbed by the body, we needed to create a system where both stages were within the optical imaging window. Although QDs are broadly excitable and exhibit higher quantum yield when exited at lower wavelengths (UV and visible light), high absorption from tissue components such as hemoglobin, water, and melanin at these wavelengths greatly attenuates the signal. Given previous studies examining the relative absorption of these components, excitation at higher wavelengths was deemed worth the trade-off. For example, our leading candidate, S10C5H, exhibited a PL QY of 43.6 ± 0.1% upon excitation with a 405-nm laser and a PL QY of 31.3 ± 0.1% when excited with an 808-nm laser, but would be subjected to an increase in absorption from hemoglobin and melanin on the order of one log unit each (
40). We chose to use a 780-nm LED rather than a laser because of the reduced cost, safety concerns, and maintenance combined with improved portability. Adapting a laser to this setup would have required a diffusor or beam splitter to spread the light, safety mechanisms, and potentially some type of cooling apparatus. These components would add cost, complexity, and size, which are undesirable for distributed mobile use. Alternatively, an LED offered low cost, inherently diffuse illumination, and safety/power advantages. The only consideration necessary to use an LED was the considerably wider emission profile (28-nm full width at half maximum); however, this was solved using a simple 800-nm short-pass filter to ensure a 50-nm gap between excitation light and light collected by the adapted smartphone.
NIR images collected with the smartphone at various ISO and shutter speed ratings were manually cropped to enable the comparison of NIR signal intensity over time. Manual ISO and shutter speeds were necessary to enable longitudinal comparison. The autoexposure setting was not well adapted for NIR imaging, as might be expected, and would often select higher exposure settings to make the field brighter to the point where the background was bright and the signal was highly saturated. Selection of the correct ISO for a pattern was critical for accurate analysis of both signal-to-noise ratio (SNR) as well as machine learning classification. To enable accurate SNR quantification, consistent imaging settings that produced nonsaturated pixels at all time points were selected. In the future, custom autoexposure settings could be implemented to automate this process.
The greatest reduction in user-identifiable signal occurred soon after administration, suggesting that some microneedles had not delivered QD-loaded microparticles to a sufficient depth—a phenomenon possibly exacerbated by the high elasticity of rodent skin (
41). A portion of particles may have been deposited on the skin rather than in the skin and thereafter cleared by external perturbation. The QD signal stabilizes rather quickly and then presents a fairly consistent SNR at 3, 6, and 9 months, indicating that microparticle clearance likely primarily occurs at early time points (less than 3 months). This is further supported by the imaging of patterns in vivo at short time points, which showed considerable loss in the days after application before stabilizing.
AlexNet was chosen for its classification accuracy, having been originally trained on more than 1 million images and having demonstrated the ability to classify images into 1000 categories (
20). Despite relatively little training data (30 images per pattern), AlexNet classified images with high accuracy. Overall, 80 images were tested for the circular pattern, 70 for the cross pattern, and 60 for the rectangular pattern originating from five distinct applications per pattern. The modified transfer-based convolutional neural network using AlexNet provided accurate detection confidence. Further, because there was no trend toward lower machine learning classification probability at 3, 6, and 9 months, it appears that the patterns are stable after an acute period of signal loss, which is also supported by short-term quantitative data and long-term qualitative data. On the basis of optical images of microneedles before and after administration, in vivo NIR imaging, and projections for photobleaching from in vitro experiment, we estimate that <1 μg of particles (<600 ng of QDs) is required to retain a detectable signal for 5 years.
Regarding accuracy, machine learning algorithms for image classification match and typically exceed manual inspection (
42,
43). Whereas the human eye is ill equipped to distinguish low grayscale value differences in images that are dim or have poor contrast (for example, grayscale values of 1 versus 10), AlexNet’s ability to use quantitative rather than qualitative information seems to make this distinction trivial. This effect is readily apparent when the grayscale photo values are represented as a rainbow heat map. AlexNet was particularly valuable when markings were missing, dim, or imaged at an unusual angle. Given the minimal training data used, it was important that images were collected using settings that roughly approximated the range of brightness/signal spread of images used for training. More training data and/or consistent autoexposure would mitigate this potential issue long term. Similarly, more patterns could be incorporated into the model to expand the variety of data that can be encoded. In the real-world use case, this automated classification will eliminate classification subjectivity that could interfere with the accuracy of this approach. Synthetic data could also be produced using image augmentation to train the algorithm without the need for actual patch applications. However, the accuracy of this method could be lower if the variability in application exhibited in the synthetic data did not replicate the features of real microneedle patch applications in the test data. In our experience, training AlexNet with 5000 synthetic images yielded a real-world data classification accuracy of 92% for the circle and 97% for both the cross and the rectangle. Future training data could also include a fourth classification category for all images that do not closely resemble any of the patterns. This would be essential for real-world use when the presence of a pattern is uncertain.
For this encoding system to be clinically useful, microneedles and NIR microparticles must be biocompatible and, if applicable, maintain the utility of any codelivered vaccines. Histology showed that the particles were well tolerated in the body, similar to largely inert PMMA-based tattoo dyes (
44), although some macrophages were observed. The accumulation of macrophages observed around the particles was less than other studies have shown for poly(lactic-
co-glycolic acid) microparticles (
45), for example, which are present in many U.S. Food and Drug Administration–approved drug delivery systems (
46). In addition, histological images demonstrated that microneedle-delivered particles were delivered to a depth similar to professional tattoos (
47), which supports their potential for long-term NIR detection.
In addition to the stand-alone value of an intradermal information encoding and detection platform, this system may offer greater advantages when codelivered with vaccines. By delivering both agents in the same microneedle patch, there is the potential to realize production cost advantages and eliminate the possibility of misuse (such as applying the encoding patch without the vaccine). However, to function in combination with microneedle vaccines currently under development, PMMA-encapsulated NIR QDs must not interfere with the robust immune response generated by intradermal antigen delivery. To ensure sufficient QD delivery and vaccine stability after fabrication (a process that includes multiple days under high vacuum), the amount of sucrose in the microneedle formulation and application time were both increased. Despite these modifications, a decrease in vaccine stability (IPV2 D-to-C conversion) was still observed and a considerable microneedle volume was not transferred to the skin during application, which would be expected to reduce the magnitude of the immune response. Nevertheless, microneedle-treated groups yielded total and neutralizing antibody titers against IPV2 that were similar with theoretical dose-matched subcutaneous injections and were above the threshold known to confer protection in humans, regardless of the presence of encapsulated QDs. This was observed despite substantial losses during formulation and administration and is likely due to the dose-sparing effects of intradermal delivery (
16,
17). Although the addition of polymer could have had an adjuvant effect, this was not observed, likely because of the well-tolerated, largely inert nature of our PMMA microparticles, as demonstrated in our histological data. These results support the clinical potential for using these microneedles to coadminister an invisible marking agent and a vaccine in one application.
This study provides support for long-lived, reliable pattern detection using QDs; however, like many preclinical studies, it is limited in duration and relies on small animal models. We present data characterizing signal expression over 9 months in animals and in vitro accelerated solar light photobleaching, which suggest that detectable markings would persist to the target time point of 5 years. On the basis of histology and longitudinal imaging studies, we would not anticipate a substantial decrease in signal within 5 years after application; however, we did not explicitly demonstrate this because of time constraints, which exceed the typical life span of our animal model.
In summary, we have demonstrated proof of concept for a platform capable of invisibly recording data in the skin. To establish this system, we synthesized, encapsulated, and delivered cadmium-free and lead-free photostable QDs into the skin using a custom microneedle array, adapted a smartphone to create an inexpensive NIR imaging system, and demonstrated the ability to accurately detect patterns in vivo for a period of 9 months using a semiautomated machine learning algorithm. Transferring this technology to the clinic will require several additional steps including preclinical safety and toxicology studies, manufacturing scale-up, and first-in-human studies. Additional in vivo small animal studies will aid in the robust characterization of local and systemic responses to QD-loaded PMMA microparticles and microneedle patches to ensure quality, safety, and reliability. Formative studies in which users test the devices, including the packaging and labeling, will need to be conducted to improve components for commercialization and to manufacture devices for testing in human studies. Ultimately, we believe that this invisible, “on-body” technology opens up new avenues for decentralized data storage and biosensing applications that could influence the way medical care is provided, especially in the developing world.
MATERIALS AND METHODS
Study design
These studies were designed to evaluate the suitability of intradermal NIR fluorescent microparticles as a reliable, long-term detection method. To evaluate this technology under typical field conditions, we used an adapted smartphone in ambient light to determine success criteria, except in circumstances where spectral and/or highly quantitative data were necessary to characterize our materials. In vivo studies were generally preferred to assess the performance of our delivery and detection systems and animals were randomly assigned to different experimental groups at the onset of each study. The use of machine learning for detection eliminated the need for blinding. In vitro studies were performed with three replicates and in vivo studies with five replicates, unless otherwise noted. Individual subject-level data are reported in data file S1.
QD synthesis
Stoichiometric CuInSe
2 cores were synthesized by mixing 1.5 mmol of copper(I) iodide and 1.5 mmol of indium(III) acetate in 1.5 ml of 1-dodecanethiol (DDT) and 45 ml of 1-octadecene (ODE) based on a modified procedure (
48). The reaction mixture was then degassed under vacuum for 20 min, purged with nitrogen for 20 min, and degassed for an additional 45 min at 120°C. Next, 1.5 ml of oleic acid was added into the mixture, and the solution was degassed for 20 min. After another 20-min nitrogen purge, the solution was heated to 175°C before the injection of selenium stock solution. A selenium stock solution was first prepared by mixing 3 mmol of selenium powder and 3 ml of oleylamine (OLA) in 3 ml of DDT (
49). The selenium solution was degassed under vacuum for at least 30 min at 60°C before injected into the reaction mixture. The reaction mixture was heated to 200°C and maintained at this temperature for 30 min under the protection of nitrogen.
ZnS:Al shells were formed around CuInSe
2 cores through the dropwise addition of a stock solution containing zinc and aluminum. The zinc precursor of the stock solution was prepared by mixing 30 mmol of zinc acetate in 30 ml of OLA and 30 ml of ODE, degassing for 20 min, and purging with nitrogen for 20 min. Next, the solution was heated to 120°C and degassed under vacuum. The Al precursor mixture consisting of 9 mmol of Al(IPA)
3, 5.4 ml of DDT, and 36 ml of ODE was degassed for 20 min and purged with nitrogen for 20 min. The vessel was then sealed and sonicated for 1 hour at 60°C (
50). The aluminum stock solution was then added to the zinc stock solution using a glass syringe and long needle. After mixing the zinc and aluminum precursors, the resulting shell stock solution was added dropwise to the reaction mixture of CuInSe
2 cores at 0.1 ml/min using a pump. At the same time, 15 ml of DDT was added at a rate of 0.5 ml/min to thermally trigger the release of sulfur (
51). The reaction was allowed to proceed for varying amounts of time, depending on the formulation. Afterward, the reaction solution was cooled to room temperature and precipitated twice into acetone, once into a 50:50 solution of acetone and methanol, and twice more into methanol. Between each round of precipitation, QDs were resuspended in a minimal amount of toluene and oleic acid that was added dropwise until the solution turned transparent. After the final precipitation step, QDs were dispersed in toluene.
Photobleaching analysis
Organic dyes required different imaging conditions than inorganic dyes and QDs because of their small Stokes shift and therefore high propensity to self-quench in the dry state. Organic fluorophores were dissolved in water at a concentration of 10 μg/ml, which was found to be within the linear absorbance range for all dyes. Fifty microliters of each dye was added to a black-walled 384-well plate, covered with a 1-mm-thick quartz slide, and sealed with parafilm at the edges to prevent evaporation during light exposure. The plate was then read using a Tecan Infinite M200 Spectrophotometer to determine the fluorescence intensities of Alexa Fluor 790 [784/814-nm excitation/emission (ex/em)], DyLight 800 (760/810-nm ex/em), IRDye 800CW (768/798-nm ex/em), IR-820 (710/820-nm ex/em), Sulfo-Cyanine7 (750/773-nm ex/em), VivoTag 800 (785/815-nm ex/em), and indocyanine green (805/835-nm ex/em).
Samples were then placed under a 300W PV Cell Testing Solar Simulator Model 16S-300-002 (Solar Light). The ozone-free short-arc xenon lamp was used with an air mass 1.5 filter and focusing lens to expose samples to conditions mimicking the spectral distribution and power density of sunlight at sevenfold the intensity of the sun (695 mW/cm2). A 2-mm-thick piece of pigmented human cadaver skin from donors age 21 to 68 years who self-identified as African American was overlaid onto the sample to recapitulate light absorption by tissue above an intradermal fluorophore. Skin was kept hydrated by a continuous flow of water over the surface and chilled from below using a CP-200HT-TT Peltier-Thermoelectric Cold Plate Cooler (TE Technology) to prevent damage to the skin. Samples were periodically removed from the solar simulator and imaged using the spectrophotometer. This experiment was repeated three times with one replicate each.
Inorganic dyes and QDs were capable of being imaged in their dry state, so a modified protocol was used. Holes were punched in a black silicone sheet using a 2-mm stainless steel biopsy punch to create space for dyes. The silicone sheet was then treated with air plasma for 1 min on high power at 500 mtorr. The sheet was then placed on a 2.54 cm by 2.54 cm quartz slide, which allowed it to adhere in a water-tight manner. Suspensions of Infrared Down-Conversion 2 (IRDC2) and IRDC3 at 10 mg/ml in water were deposited 2 μl at a time and allowed to dry until a total of 10 μl had been deposited. The same steps was performed for ZnS:Al-coated CuInSe
2 QD solutions at 10 mg/ml in toluene. Double-sided tape and parafilm were then used to seal a glass slide to the other side of the black silicone. The sample was then placed quartz slide-up beneath a piece of pigmented cadaveric skin in a metal block under continuous hydration on a cold plate cooler and exposed to light at sevenfold solar intensity. At predetermined time points, samples were removed from light and imaged using a custom NIR-imaging platform consisting of a 500-mW laser emitting at 808 nm and thermoelectric-cooled mount powered by a LDC210C Laser Diode Controller and TED200C Temperature Controller, a 15× achromatic Galilean beam expander, protected silver mirror, 850-nm color glass long-pass emission filter, and a high-sensitivity USB 3.0 complementary metal-oxide semiconductor camera (DCC3240N) affixed to an optical breadboard (Thor Labs) inside a dark work enclosure (U.S. Laser). Images were collected at different exposure lengths for different dyes ranging from 10 to 2000 ms. The longest exposure length that did not saturate the 10-bit pixel depth was used for each dye at all time points. Owing to the length of this experiment, all samples (
n = 3) were run simultaneously. For all photobleaching experiments, skin was replaced every 2 to 4 days to reduce the effects of melanin degradation. Intensity data were quantified using ImageJ (National Institutes of Health) and normalized by selecting a region of interest (ROI) around a well and comparing the intensity above background in that image to the intensity above background of the corresponding ROI at the beginning of the experiment (
52). All photobleaching data are reported as simulated days of exposure based on a 12-hour light/dark cycle.
QD encapsulation
QD-loaded PMMA microparticles were formed using an oil-in-water emulsion/solvent evaporation technique. Briefly, 150 mg of QDs and 100 mg of PMMA were dissolved in 2 ml of dichloromethane. This solution was then added to 50 ml of 1% poly(vinyl alcohol) (PVA) [88% hydrolyzed, Mw (weight-average molecular weight), 31,000] solution in water. The resulting solution was emulsified at 5000 revolutions per minute (RPM) for 1 min using a T 18 digital ULTRA-TURRAX homogenizer (IKA Works) and subsequently stirred at 250 RPM for 3 hours to allow solvent to evaporate. Particles were the centrifuged at 1000 relative centrifugal force (RCF) after which the supernatant was removed. Particles were then washed four times by adding deionized water, centrifugation at 1000 RCF, and supernatant removal. The resulting particles were resuspended in deionized water and measured using a Multisizer 3 (Beckman Coulter) with a 100-μm aperture. Data were smoothed using an 11-frame moving window and plotted as a histogram with 300 equal-sized bins ranging from 2.1 to 59.9 μm. QD-loaded microparticles were then imaged using optical microscopy and scanning electron microscopy (SEM) to observe their shape. Optical imaging was performed using an Olympus MX40 inspection microscope with a ToupCam industrial digital camera (ToupTek Photonics). In preparation for SEM, samples were deposited on double-sided carbon tape and coated with a thin layer of Au/Pd using a Hummer 6.2 Sputtering System (Anatech) to prevent charging. Imaging was then performed using a JEOL JSM-5600LV scanning electron microscope with an acceleration voltage of 5 kV.
The pH stability of S10C5H QDs in PMMA was determined by casting a solution onto the bottom of a plate to prevent agitation upon addition of the buffer. QDs (150 mg) and 100 mg of PMMA were added to 2 ml of dichloromethane and sonicated for 5 min. Using a glass pipette, about 5 μl of the QD solution was added to the bottom of a 96-well black-walled glass-bottom plate and dried overnight. PBS (200 μl) was then added to each well, and the samples were incubated at 37°C for 24 hours. At this stage, the initial intensity of the QD-PMMA samples were analyzed using a custom NIR-imaging platform as described above. To investigate the effect of pH on the intensity of the QDs, the PBS was replaced by a buffer solution at pH 4, 5, 6, 7.4 (PBS), or 10. The plate was then sealed to minimize solvent loss and incubated at 37°C. After 1, 4, and 22 hours, samples were removed from the incubator, and QD intensity was measured using the custom NIR-imaging system with a 40-ms collection time. Intensity data were quantified in ImageJ and normalized by selecting an ROI around a well and comparing the intensity above background in that image to the intensity above background of the corresponding ROI at the beginning of the experiment.
Dissolvable microneedle fabrication and characterization
Water-soluble microneedle patches containing QD-loaded microparticles at the needle tips were fabricated using a solvent casting process. First, microparticles were resuspended in water, pipetted onto the top of the PDMS mold (4 μl per microneedle), and centrifuged at 3234 RCF for 5 min. Excess solution was cleared from the top of the mold, and the solution was allowed to dry leaving behind particles in the tips of the mold. This process was repeated when loading was lower than desired or unevenly distributed. About 300 μl of a 17% (w/v) sucrose and 17% (w/v) PVA solution in water was then dispensed on top of the mold and centrifuged at 3234 RCF for 5 min. Molds were then left at room temperature overnight in a laminar flow hood as an initial drying stage. Laser-cut acrylic discs were then affixed to double-sided tape and attached to the back side of the solidified microneedle patch. The patch was then carefully removed from the mold and stored for an additional 72 hours under vacuum desiccation.
Smartphone modifications
A smartphone camera was adapted with commercially available optical components purchased from Thorlabs to enhance NIR QD detection in the NIR. To enable NIR detection, the stock short-pass IR filter was removed from a Google Nexus 5X smartphone. A smartphone case was then designed and 3D-printed to interface tightly with optical components having SM1 threading directly in front of the rear-facing camera. An 850-nm long-pass dielectric filter (FEL0850) and 850-nm long-pass color glass filter (FGL850) were placed in parallel in a lens tube attached to the smartphone case and held in place with a retaining ring. For QD illumination, a 780-nm, 200-mW mounted LED (M780L3) powered by a T-Cube LED Driver (LEDD1B) and 15-V, 2.4-A power supply (KPS101) was used. To augment the shape and spectrum of emission, an 800-nm dielectric short-pass filter (FEL0800) and aspheric condenser lens with diffuser (ACL2520U-DG6-B) were used in an adjustable lens tube. All imaging was performed using the Camera FV-5 Lite app (FGAE Studios).
Biocompatibility of encapsulated QDs
Microneedles containing encapsulated QDs were administered to Wistar rats (Charles River Laboratories) 8 to 12 weeks of age weighing about 250 g to assess in vivo biocompatibility. Rats were anesthetized via continuous inhalation of 2.5% isoflurane throughout the administration and imaging process. Hair removal was performed before microneedle administration by shaving the rear flank with an electric razor and applying depilatory cream for about 2 min. The area was then rinsed to remove excess hair, sterilized using an ethanol swab, and allowed to dry. Microneedles were applied using an MPatch Mini spring-loaded applicator (Micropoint Technologies) for 2 min. Rats were returned to their cages and housed until terminal time points at 1 day, 2 weeks, and 4 weeks (n = 4 per time point). At the time of sacrifice, rats were euthanized via CO2 asphyxiation. Skin was explanted and fixed in formalin-free tissue fixative (Sigma-Aldrich) for 24 to 72 hours. The relevant portion of the tissue was identified using the adapted smartphone, transferred to 70% ethanol, and embedded in paraffin wax. Samples were sectioned and stained with hematoxylin and eosin or Masson’s trichrome. Interpretation of the foreign body reaction was performed under the guidance of an experienced veterinary pathologist.
Longitudinal in vivo imaging of NIR patterns
Two longitudinal imaging experiments were performed to assess the short-term loss of signal after administration and long-term residence of microneedle-delivered particles. In both cases, QDs were administered to rats using dissolvable microneedle patches as described above and imaged periodically to observe signal intensity over time. At each time point, rats were anesthetized under continuous inhalation of 2.5% isoflurane and imaged with the 780-nm LED and NIR-adapted smartphone, maintaining a consistent imaging distance.
In the first experiment, microneedles containing PMMA-encapsulated and free (unencapsulated) ZnS:Al-coated CuInSe2 QDs were compared to evaluate the potential benefits of encapsulation on initial signal intensity and signal retention. Encapsulated QDs were embedded in dissolvable microneedles using the protocol detailed above. Unencapsulated QDs were not readily dispersible in water, so the loading procedure was slightly modified. Before loading, PDMS molds were treated with air plasma using a Harrick Plasma Cleaner PDC-091-HP on high power at 500 mtorr for 1 min. QDs in toluene (4 μl per microneedle) were then deposited into the microneedle molds and centrifuged at 3234 RCF for 5 min. Toluene was then allowed to dry. A solution of sucrose and PVA was then applied to the molds and centrifuged into the needles and allowed to dry using the fabrication steps outlined above.
After applying 16-microneedle patches, the highest nonsaturating imaging settings were identified and used at all subsequent time points for that group. For the unencapsulated group, these settings were ISO 100 with a shutter speed of
s. For the encapsulated group, these settings were ISO 100 with a shutter speed of
or
s. For each animal and time point, about 20 images were collected with slightly different aim of the LED, which was controlled independently from the camera that remained in a fixed location. An ImageJ macro was then used to split images into their three constituent red, green, and blue channels, save the red channel in grayscale format, and identify the images with the maximum intensity. This image with the best aim was then used for subsequent quantification of signal-above background. Briefly, 80-pixel circular ROIs were applied to each of the 16 dots in the 4 × 4 array to quantify the average signal. The average grayscale intensity of the remaining pixels was subtracted to generate longitudinally comparable values for signal above background.
On the basis of this short-term study, a second long-term study was performed using encapsulated QDs in eight-needle microneedle patterns (circle, cross, or rectangle) for image detection and machine learning analysis. Arrays were administered at 0 (circle), 4 (cross), and 8 (rectangle) weeks to mimic a common vaccination schedule in the developing world. A separate group was started in parallel to generate an image dataset for training the machine learning algorithm. Images were collected every 2 weeks at a variety of ISO and shutter speeds, processed, and used for both SNR analysis and machine learning classification.
Microneedle-based vaccine delivery
The solution used to cast microneedles was altered in favor of additional sucrose (2:1 w/w sucrose:PVA) to potentially improve vaccine stability during processing. A solution of 34% (w/v) sucrose and 17% (w/v) PVA solution containing 3.2 DU of Salk IPV2 (MEF-1 strain; Statens Serum Institut) was centrifuged into PDMS microneedle molds using the procedures detailed above. After processing, some microneedle patches were collected, resuspended in assay buffer, and evaluated using a D-antigen–specific monoclonal sandwich enzyme-linked immunosorbent assay (ELISA) for IPV2 to determine the amount of the vaccine still in its immunity-conferring state. A monoclonal antibody against IPV2 poliovirus (HYB 294-06-02, Thermo Fisher Scientific) was used as both the capture and detection antibody, which was made possible because of the multiple identical binding sites on IPV. To avoid species cross-reactivity, a Lightning-Link Horseradish Peroxidase Kit (Novus Biologicals) was used to prebind the detection epitope. Briefly, 100 μl of antibody diluted 1:1500 in carbonate buffer (pH 9.6) was added to each well of a Nunc Maxisorp LockWell 96-well plate and incubated overnight at 4°C on an orbital shaker. Plates were then washed with PBS containing 0.05% Tween 20 three times and incubated in 300 μl of blocking buffer containing the wash buffer and 5% (w/v) nonfat milk for 1 hour at 37°C. Plates were again washed and then loaded with 50 μl of samples. After 2 hours of incubation at 37°C, plates were washed five times and the Lightning Link–modified antibody was added at 1:833 in PBS. After 1-hour incubation, plates were washed another five times and then 100 μl of a SIGMAFAST OPD substrate (Sigma-Aldrich) was added after resuspension according to the manufacturer’s protocol. After a color change, 150 μl of 1 M sulfuric acid was added to each well to stop the reaction, and absorbance values were read at 490 nm with a background reference of 630 nm using a spectrophotometer.
Microneedle patches not used for ELISA were applied to rats for a duration of 5 min using the protocol described above. After application, the portion of the microneedles remaining on the patch was collected, resuspended in assay buffer, and measured via the IPV2 D-antigen ELISA. A control group of rats received a dose-matched subcutaneous injection in the rear flank. These procedures were repeated 4 and 8 weeks later to provide rats with second and third doses. Rats were bled vial tail vein every 2 weeks after administration to collect serum for immunological analysis. Blood was collected in BD SST tubes (product no. 365967, Becton Dickinson) and used according to the manufacturer’s protocol to obtain serum, which was subsequently stored at −20°C until use. A total anti-IPV2 antibody ELISA was performed in house using an indirect ELISA, as previously described (
53). IPV2 neutralizing antibody titers were determined by the U.S. Centers for Disease Control and Prevention using a protocol previously published (
21).
Statistical analysis
Statistics were performed in GraphPad Prism using Student’s t test for pairwise comparisons and one-way analysis of variance (ANOVA) with Tukey’s multiple comparisons test for comparing multiple groups at a significance level of α = 0.05. In vitro toxicity experiments evaluating both QD type and concentration were analyzed using two-way ANOVA at a significance level of α = 0.05. All in vitro experiments were performed in experimental triplicate unless otherwise noted. All in vivo experiments were performed with five experimental replicates unless otherwise noted. Data are reported in the text as means ± SD.