Deep learning-enhanced chemiluminescence vertical flow assay for high-sensitivity analyte detection

Abstract

A method of measuring analyte(s) or biomarker(s) in a sample uses a cartridge-based vertical flow assay with a sensing membrane located within one or more cartridges and populated with a plurality of spots containing one or more capture agent(s). A mixture of the sample and detection reagents along with a chemiluminescence (CL) reagent solution are input into the one or more cartridges and the sensing membrane is imaged with a reader device configured to obtain one or more CL images and / or CL signals for the plurality of spots. The one or more images and / or CL signals for the plurality of spots are processed with an algorithm including machine learning or one or more trained neural networks configured to generate one or more outputs that include a classification of the sample and / or a quantification of the amount or concentration of the analyte(s) or biomarker(s) in the sample.

Description

2025-128-2DEEP LEARNING-ENHANCED CHEMILUMINESCENCE VERTICAL FLOW ASSAY FOR HIGH-SENSITIVITY ANALYTE DETECTIONRelated Application

[0001] This Application claims priority to U. S. Provisional Patent Application No.63 / 730,460 filed on December 11, 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 a deep learning-based system that is used to read immunoreaction spots of a chemiluminescence vertical flow assay (CL-VFA) and determine the concentration of a target analyte or biomarker. In particular, the technical field relates to a system or platform that uses the machine / deep learning-based framework to determine the concentration of the target analyte cardiac troponin I (cTnl).Statement Regarding Federally SponsoredResearch and Development

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

[0004] Cardiovascular diseases (CVDs) remain a significant global health issue, responsible for around 19.1 million annual deaths, which accounts for 32% of all global deaths. This significant mortality, coupled with a considerable socioeconomic burden, resulted in CVD-related healthcare costs exceeding $250 billion annually in the United States alone. Notably, CVDs disproportionately impact underserved populations, especially in resource-limited settings where quality preventive and diagnostic services are less accessible. Among the various types of CVDs, acute myocardial infarction (AMI), or heart attack, remains a leading cause of emergency visits and deaths, highlighting the need for prompt and accurate diagnostic strategies due to its time-critical nature. Diagnosing AMI relies heavily on detecting elevated cardiac troponin I (cTnl) levels, a highly specific and sensitive biomarker of myocardial injury. Blood cTnl levels are often used in conjunction with other diagnostic methods, including patient-reported symptoms, electrocardiograms, and imaging2025-128-2studies, as part of a comprehensive diagnostic framework. Clinical guidelines emphasize the significance of elevated cTnl concentrations beyond established cut-off levels (approximately 10-40 pg / mL) as a necessary criterion in diagnosing AMI. In clinical practice, AMI diagnosis is confirmed when an elevated cTnl level is identified alongside at least one abnormal finding among other supportive diagnostic methods. This approach emphasizes the critical role of accurate and sensitive cTnl quantification as a prioritized factor in guiding AMI diagnosis and management.

[0005] Given the critical role of cTnl measurement, numerous cTnl detection methods have been developed, ranging from sophisticated laboratory-based assays to portable point-of-care testing (POCT) platforms. Laboratory-based high-sensitivity cTnl (hs-cTnl) assays have set the gold standard in clinical diagnostics, as they enable the detection of cTnl at trace levels (down to a few pg / mL) in the bloodstream, facilitating early identification of AMI and accurate patient risk stratification. However, these hs-cTnl assays are mainly confined to centralized hospital laboratories where specialized equipment, skilled personnel, and dedicated bench space are available. The high maintenance and operating costs of these systems can be particularly burdensome for healthcare facilities with limited financial resources. Further, reliance on trained personnel restricts their use in resource-limited settings. Additionally, longer turnaround times associated with the multistep administrative procedures of central laboratory testing can hinder prompt decision-making in critical patient management. These limitations often make hs-cTnl testing inaccessible to patients in underserved areas, where the prevalence of CVDs is higher, or in high-demand emergency settings, where prompt diagnostic decisions are vital. Consequently, the current diagnostic paradigm in AMI care, dominated by laboratory-based testing, faces inherent limitations related to processing time, expense, and accessibility.

[0006] Driven by the systemic limitations of laboratory-based cTnl testing, there has been a shift toward patient-centered POCT systems in AMI diagnostics. POCT enables diagnostic evaluation without requiring extensive infrastructure or specialized equipment, making it valuable in rural, low-resource settings, and in emergency environments, where rapid diagnosis is crucial. There are currently several cTnl-POCT products on the market utilizing technologies such as microfluidics, lateral flow tests, and automated immunoassays. Some of the recent advancements in cTnl biosensing also include paper-based tests, digital microfluidics, microparticle assays, electrochemical methods, and array-based sensing.Although these technologies expand the capabilities of cTnl-POCT, achieving the high72025-128-2sensitivity needed for hs-cTnl detection remains challenging for most POCT platforms. The exceptionally low cTnl cut-off levels required to differentiate AMI from non-AMI cases are difficult for most POCT systems to meet, particularly those designed with low-cost sensors. Despite advances in miniaturization, many POCT systems still retain benchtop-sized footprints and remain costly, limiting their use in smaller, resource-constrained clinical environments. Thus, developing a POCT platform that combines affordability, true portability, and high sensitivity remains a critical goal, as overcoming these barriers could greatly improve diagnostic access and support the shift toward more patient-centered care.

[0007] Although current POCT systems have limitations in achieving the necessary sensitivity required for clinical use, chemiluminescence (CL)-based detection has proven effective for highly sensitive biomarker assays. Among the 14 FDA-, CE-, or equivalent regulatory-cleared commercial hs-cTnl benchtop analyzers devised for clinical laboratory use, 12 of them utilize CL-based technologies. CL has a distinctive advantage in generating strong, high-contrast signals without the need for an external light source, unlike fluorescence-based methods that require complex excitation optics. Its high signal-to-noise (S / N) ratio, achieved by low background noise, significantly enhances detection sensitivity, while the broad calibration range of CL enables the accurate measurement of cTnl across clinically relevant concentrations — from trace levels (a few pg / mL) to highly elevated levels (several tens of ng / mL). Nevertheless, integrating CL technology into a POCT platform poses significant challenges. Traditional CL assays rely on precise liquid handling, multiple washing steps, and a large / expensive imaging system, which complicate miniaturization efforts. As a result, most CL-based assay platforms have remained in centralized laboratories where the technical demands can be fully met. Efforts to incorporate CL into POCT systems have also faced limitations, including reliance on single-molecule enzyme conjugates with limited sensitivity, dependence on benchtop imaging stations that constrain portability and affordability, and inadequate performance of smartphone-based readers for CL imaging due to their optical / software limitations. To address these issues, advances in conjugate sensitivity, reader design, and assay platform miniaturization are required. This will ensure the implementation of a clinical laboratory-grade hs-cTnl assay in a fully portable and accessible format.J2025-128-2Summary

[0008] Here, a paper-based CL vertical flow assay (CL-VFA) is disclosed for highly sensitive, precise, rapid, and affordable cTnl quantification, achieving clinically relevant sensitivity and precision comparable to laboratory -based standard hs-cTnl assays. The CL-VFA system integrates several innovations designed to address the critical limitations of existing CL-based POCT assays: (i) a polymerized enzyme-based conjugate for enhanced CL signal intensity and sensitivity, (ii) a cost-effective, portable Raspberry Pi -based CL reader with robust imaging capabilities, eliminating the need for a complex benchtop readout system, (iii) a user-friendly, streamlined tray -based VFA cartridge that ensures stable and consistent CL imaging, and (iv) a neural network-driven computational pipeline for accurate classification and quantification of cTnl levels. This platform measures cTnl from a 50 pL serum sample within 25 min per test, combining the assay performance of a laboratory -based hs-cTnl analyzer with the simplicity of point-of-care diagnostics. For validation, the detection sensitivity and precision were assessed using both control serum samples spiked with cTnl and clinical patient serum samples. The initial validation using cTnl-spiked human serum resulted in a detection limit of 0.16 pg / mL and demonstrated high reproducibility, with an average coefficient of variation (CV) of 5.4%. In a blinded test involving 66 clinical samples from 34 patients, the predicted cTnl concentrations from the computational CL-VFA system strongly correlated (Pearson’s correlation coefficient [r] of 0.979) with the ground truth values measured by an FDA-cleared laboratory instrument, with an average CV of 14.3%. These findings highlight the transformative potential of this deep learning-enhanced CL-VFA platform, offering a rapid, low-cost, and high-performance solution for accurate cTnl testing at the point of care. This approach represents a significant step forward in making high- sensitivity cardiac diagnostics more accessible across diverse healthcare settings.

[0009] In one embodiment, a method of detecting the presence of and / or quantifying the amount or concentration of one or more analytes or biomarkers in a sample using a vertical flow assay includes one or more cartridges having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more capture agent(s). The method includes loading a mixture of the sample and detection reagents including a detection antibody conjugate that includes a plurality of peroxidase molecules associated with an affinity agent and conjugated to detection antibodies along with a chemiluminescence (CL) reagent solution into the sample inlet. The sensing membrane is imaged with a reader device configured to obtain one or more CL images and / or CL signals for the plurality of spots. The2025-128-2one or more images and / or CL signals for the plurality of spots are then processed with a computing device with an algorithm including machine learning or one or more trained neural networks configured to generate one or more outputs that include a classification of the sample and / or a quantification of the amount or concentration of the one or more analytes or biomarkers in the sample.

[0010] In another embodiment, a system for detecting the presence of and / or quantifying the amount or concentration of one or more analytes or a biomarker in a sample includes a vertical flow assay that includes one or more cartridges having a sample inlet and a removable sensing membrane populated with a plurality of spots containing one or more capture agent(s). The system includes a reader device that includes a housing that has a camera and a cartridge tray that receives the one or more cartridges and places the removable sensing membrane along an optical path of the camera. The system further includes a computing device configured to obtain chemiluminescence (CL) images and / or CL signals of the sensing membrane after exposure to a CL reagent solution, the computing device further including software or instructions that execute an algorithm including machine learning or one or more trained neural networks configured to generate one or more outputs that include classification of the sample and / or a quantification of the amount or concentration of the one or more analytes or biomarkers in the sample based on the CL images and / or CL signals for the plurality of spots.Brief Description of the Drawings

[0011] FIG. 1A illustrates the CL-VFA system, including assay cartridges, a portable reader device, and computing device for rapid and accessible POCT.

[0012] FIG. 1B illustrates an exploded view of the reader device.

[0013] FIG. 1C illustrates the location of the cartridge inserted into the cartridge tray when advanced in the optical path of the camera during the CL imaging process.

[0014] FIG. 1D illustrates the assay workflow of CL-VFA with schemes of immunoassay and CL reaction steps.

[0015] FIG. 1E illustrates the computational analysis workflow powered by deep learning, detailing the processing and analysis of assay data.

[0016] FIG. 2A is schematic illustration of the conjugation process involving 15 nm AuNP, polyHRP, and biotinylated antibody.2025-128-2

[0017] FIG. 2B is graph showing the comparison of sensitivity between CL signals generated by AuNP-polyHRP-Ab and AuNP-standard HRP-Ab conjugates, as well as the colorimetric signal of AuNP-polyHRP-Ab conjugate. The conjugates were serially diluted and pre-applied onto the sensing membrane.

[0018] FIG. 2C is a signal-to-noise (S / N) ratio comparison of CL signals across different reader systems: the Raspberry Pi-based portable system, a smartphone-based reader, and a traditional benchtop system.

[0019] FIG. 2D illustrates a comparative analysis of the key specifications of benchtop, smartphone, and Raspberry Pi -based systems. The specifications of the benchtop imaging system refer to Chemidoc MP (Bio-Rad), which was used for the comparative study.

[0020] FIG. 2E illustrates the design comparison, highlighting the differences between first bottom cartridge and second bottom cartridges, with the latter designed (right side - final design) to minimize fluid flow during the CL imaging step for improved signal stability.

[0021] FIG. 2F is a graph showing the influence of the reagent flow within the bottom cartridge on CL imaging stability and signal saturation.

[0022] FIG. 2G illustrates a comparison of CL assay results using the original VFA cartridge without the tray and the modified tray-based case, demonstrating identical assay performance despite the design modification. The numbers in FIGS. 2B and 2C are detection cut-offs. Data points in FIGS. 2B, 2C, and 2D represent the mean of triplicates ± SD. Data points in FIG. 2F represent the mean of duplicates ± SD.

[0023] FIG. 3 A is a spotting map and representative sensing membrane images after assaying various concentrations of cTnl and capturing activated assays by the Raspberry Pi-based portable reader device.

[0024] FIG. 3B is a calibration plot (solid line; R2= 0.9929, y = 0.0147x0.4043), demonstrating the assay's strong correlation and sensitivity across the clinically relevant range.

[0025] FIG. 3C is a comparison of test spot intensities for cTnl concentrations within clinically relevant cut-off levels, highlighting statistically significant differences at various cTnl concentrations.

[0026] FIG. 3D illustrates cTnl clinical sample test results.

[0027] FIG. 3E is an expanded view of the clinical sample test results below the 100 pg / mL range. Data points in FIGS. 3B and 3C represent the mean of triplicates ± SD. Data points in FIGS. 3D and 3E represent the mean of duplicates ± SD.2025-128-2

[0028] FIG. 4A illustrates a neural network-based cTnl quantification workflow pipeline. The neural network pipeline consists of a classification network (DNN classification) and three (3) quantification networks (DNNQ<40, DNNQ40–1000and DNNQ IOOO). DNNclassification classifies all samples into one of the three pre-determined concentration ranges (<40 pg / mL, 40-1000 pg / mL, and >1000 pg / mL) based on the cTnl concentration in the serum sample. Then three (3) separate quantification network models (DNNQ<40, DNNQ40–1000, and DNNQ>1000) quantify cTnl concentration in samples from each range.

[0029] FIG. 4B shows classification results of the DNNClassificationnetwork for 66 serum samples from 34 patients used in the blind testing set.

[0030] FIG. 4C illustrates combined quantification results from the three quantification models for the 66 serum samples from 34 patients used in the blind testing set. Samples with cTnl concentration in <4 pg / mL range do not have quantitative ground truth labels due to the limitations of the clinically used measurement system, and therefore x-axis does not have quantitative values in that range. All samples with cTnl concentration in <4 pg / mL range used in the blind testing set are stacked above each other within the <4 pg / mL interval.

[0031] FIGS. 5A-5H is a table showing detailed information of the clinical sample test data.

[0032] FIG. 6 shows a detailed layout of the paper layers in CL-VFA including the components of the first top cartridge and vertical flow path. The layers in the first bottom cartridge are illustrated for immunoassay and washing.

[0033] FIG. 7 A illustrates a comparison of CL signals between cTnl-negative (0 ng / mL) and cTnl-positive (10 ng / mL) samples under different assay conditions. The control conditions resulted in lower signal -to-noise ratios, while the optimized conditions in Test 3 optimized the signal for positive samples and minimized the background signals for negative samples. Data points represent the mean of triplicates ± SD.

[0034] FIG. 7B illustrates bright-field and CL images of the absorbent layers used to examine the distribution of unbound conjugates across the absorbent layers after the assay and washing steps. The absorbent stack consists of five layers, with layer 1 in direct contact with the sensing membrane and layer 5 at the bottom of the stack. Bright-field and CL images reveal the localization of the conjugates within the layers. Layer 1 shows no detectable CL signal, indicating effective washing of unbound conjugates in the CL-VFA structure. In contrast, layer 5 displays a strong CL signal, corresponding to the red coloration observed in2025-128-2the bright-field image, demonstrating that unbound conjugates are efficiently transported and concentrated in the lower absorbent layers, as desired.

[0035] FIG. 8 illustrates calibration curves of the CL-VFA for cTnl-spiked serum samples were obtained using different exposure times (10 s, 30 s, and 60 s) for CL imaging. The inset highlights the signal responses in the low concentration range (1—10 pg / mL), demonstrating the impact of the exposure time on detection sensitivity and dynamic range. Data points represent the mean of triplicates ± SD.

[0036] FIG. 9 illustrates an illustrative graphical user interface of Raspberry Pi-based portable CL reader device.

[0037] FIGS. 10A illustrates an optimal setup for chemiluminescence (CL) imaging to minimize fluid motion. In this embodiment, a flat-bottom container along with its cross-sectional diagram for CL imaging on the excised sensing membrane, which is covered by a glass slide with a CL reagent solution.

[0038] FIG. 10B illustrates a time-dependent intensity profile of the CL signal demonstrates a stable condition for CL imaging,

[0039] FIG. 11 is a graph illustrating the results of sample revalidation for outlier serum samples. The numbers above or below each dot indicate the cTnl concentrations measured using an FDA-cleared benchtop analyzer.

[0040] FIGS. 12A-12B illustrate neural netw'ork-based analysis on the validation set. FIG.12A shows the classification results from DNNclassification for 62 serum samples from 32 patients used in the validation set. FIG. 12B illustrates combined quantification results from the three optimized quantification models (DNN<40, DNN40-1000, and DNN>1000) for the 60 serum samples from 31 patients used in the validation set. Two samples were labeled "indeterminate" and excluded from the quantification dataset due to contradicting predictions between the classification and quantification neural network models.

[0041] FIG. 13 illustrates quantification results from a single neural network model for 66 serum samples from 34 patients used in the blind testing set.

[0042] FIGS. 14A-14C illustrate the results of neural network-based cTnl quantification pipeline based on three-model configuration. FIG. 14A illustrates that the neural network pipeline consists of 1 classification (DNN0100) and 2 quantification (DNN<ioo and DNN>ioo) models. DNN0100 classifies all samples between two concentration ranges (<100 pg / mL and >100 pg / mL) based on the cTnl concentration in the sample. Then 2 quantification models (DNNcioo and DNN>ioo) quantify cTnl concentration in the test samples from each range. FIG. 14B shows the classification results from DNN0100 for 66 serum samples from 342025-128-2patients used in the blind testing set. FIG. 14C shows the combined quantification results from the two quantification models for the 66 serum samples from 34 patients used in the blind testing set.

[0043] FIGS. 15A and 15B illustrate the results of neural network-based cTnl quantification pipeline based on a singleplexed test. FIG. 15A shows the classification results from DNNclassification for the blind testing set using a single spot for the test, positive control, and negative control signals of each CL-VFA FIG. 15B shows the combined quantification results from the three optimized quantification models (DNN<40, DNN40-1000, and DNN>iooo) for the blind testing set using a single spot for the test, positive control, and negative control signals of each CL-VFA.

[0044] FIG. 16 illustrates classification results from DNNclassification for the blind testing set using CL-VFA signals without the outlier exclusion step.

[0045] FIG. 17 illustrates combined quantification predictions from the three power-fitting models for the 66 serum samples from 34 patients used in the blind testing set.

[0046] FIGS. 18A-18B illustrate classification predictions on the blind testing set using other machine learning models. FIG. 18A illustrates classification predictions on the blind testing set using a random forests model. FIG. 18B illustrates classification predictions on the blind testing set using a logistic regression model.

[0047] FIG. 19 illustrates analysis of the absorption peak shift before and after the synthesis of the conjugate.

[0048] FIG. 20A illustrates the training of the DNNclassification network. DNNclassification was trained on 68 CL-VFA samples activated with serum samples from 35 different patients and validated on 62 CL-VFA samples from 32 patients. cTnl concentrations in the training and validation sets for DNNclassification varied between <4 and -10 000 pg mL ’, covering typical clinical ranges of hs-cTnl testing.

[0049] FIG. 20B illustrates the training of the DNNQ<40 network. DNNQ<40 was trained on 47 CL-VFA samples activated with serum samples from 24 different patients and validated on 34 CL-VFA samples activated with serum samples from 17 patients. The training dataset for DNNQ<40 had cTnl concentrations between <4 and -100 pg mL4

[0050] FIG. 20C illustrates the training of the DNNQ40–1000network. DNNQ40–1000was trained on 18 samples from 10 patients and further validated on 15 samples from 8 patients. Training samples for DNNQ40-1000 had cTnl concentrations from -20 to -1500 pg mL1.2025-128-2

[0051] FIG. 20D illustrates the training of the DNNQ>1000network. DNNQ>1000was trained on 19 samples from 10 patients and then validated on 13 samples from 7 patients. cTnl concentrations for DNNQ>1000varied between -500 and -10000 pg mV'.Detailed Description of Illustrated Embodiments

[0052] With reference to FIGS. 1 A-1E a platform or system 10 for detecting the presence of and / or quantifying the amount or concentration of one or more analytes or biomarkers in a sample is disclosed. The analyte may include cardiac troponin I (cTnl) (FIG. ID) although other analytes can be used with the system 10 which, of course, requires different capture antibodies specific to the analyte of interest. The sample that is tested may include a biological fluid such as a body fluid, including but not limited to whole blood or serum, saliva, urine, cerebrospinal fluid, sweat, amniotic fluid, or interstitial fluid but may also include an environmental sample. The system 10 includes a vertical flow assay 12 that, in one embodiment, uses several cartridge subunits 14 / X, 14B, 16A, 16B that are combined together in different parts of the assay procedure although other embodiments envision a single cartridge. The cartridge subunits 14A, 14B, 16 / X, 16B may be made from any number of materials but preferably may include a polymer or plastic material. With reference to FIGS. 1 A, 2E and 6 a first bottom cartridge 14A is provided that includes a plurality of absorption layers 18 (e.g., engineered paper or other material) and accommodates a sensing membrane 22 populated with a plurality of spots 24 containing one or more types of capture agent(s) 26 (FIG. ID). The capture agent(s) 26 may include antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, peptoids, etc. The capture agent 26 in one specific embodiment includes capture antibodies are specific to cTnl as described herein but other capture agents 26 may be used for other analytes. For example, other analytes include, but are not limited to, cardiac troponin T, myoglobin, creatine kinase-MB, B-type natriuretic peptide (BNP), N-terminal proBNP, C -reactive protein, interleukin-3, interleukin-8, albumin, or glycated albumin, as well as virus biomarkers such as Influenza A / B, HIV, Coronavirus, Cytomegalovirus, Hepatitis, and bacterial biomarkers including bacterial lipopolysaccharide, bacterial antigens, bacterial enzymes, or toxins. The sensing membrane 22 may be a nitrocellulose membrane that is spotted at different locations (i.e., spots 24) with capture antibodies as the capture agent 26. As explained herein, a wax printer may optionally be used to compartmentalize discrete hydrophilic locations on the sensing membrane 22 that contain2025-128-2the capture agents 26 (e.g., capture antibodies) that are surrounded by hydrophobic wax layer(s). This focuses or concentrates fluid flow onto the spots 24.

[0053] The vertical flow assay 12 further includes a first top cartridge 16A having one or more paper layers 28 (e.g., paper) therein and a sample inlet 30 as seen in FIG. 6. The paper layers 28 may include sequentially stacked engineered paper layers. In one embodiment, the paper layers 28 may include, in one embodiment, an absorption layer 28a, flow diffuser 28b, 1stspreading layer 28c, interpad 28d, 2ndspreading layer 28e, and support layer 28f for holding the layers 28a-28e. In some embodiments, the first top cartridge 16A may include asymmetric membranes as one or more of the layers 28-28f. These asymmetric membranes are asymmetric in that the pore size changes in the direction of the thickness of the membrane. For example, a top such membrane may be oriented to place the side with the larger pores at the top while the bottom membrane is oriented to place the side with the larger pores at the bottom. These asymmetric membranes aid in lateral spreading (e.g., spreading layers 28e). Some layers such as vertical flow diffuser layers 28b promote vertical (e.g., top- to-bottom) movement of fluid through the layer and inhibit lateral flow. Still other layers act as supporting structures (e.g., support layer 28f) or support lateral flow (i.e., asymmetric membranes such as spreading layers 28e). The various layers 28a-28f may be held together and in place with an external or peripheral support in the first top cartridge 16A which in the experiments conducted herein was foam tape, although it should be understood that other materials and structures may be used.

[0054] The first bottom cartridge 14A is provided that carries a removable sensing membrane tray 32 that holds the sensing membrane 22. In this regard, the sensing membrane 22 can be removed from one cartridge and then transferred to another cartridge (e.g., from cartridge 14A / cartridge 16A to cartridge 14B / cartridge 16B). A series of absorption layers or pads 18 are located beneath the sensing membrane 22 that is held in the membrane tray 32. A second bottom cartridge 14B is provided that allows for the transfer of the sensing membrane tray 32 and sensing membrane 22 (removable) from the first bottom cartridge 14A to the second bottom cartridge 14B for the CL reaction. Here the second bottom cartridge 14B has a support stage 34 (made of non-absorbent plastic) that supports the sensing membrane 22. In this regard, the absorption layers or pads 18 are removed in this second bottom cartridge 14B as these adversely impacts results in this cartridge subunit.

[0055] The second top cartridge 16B is provided that mates with the second bottom 14B cartridge. The second top cartridge 16B includes a reagent inlet 36 and a fluidically coupled2025-128-2reagent chamber 38 that overlies the sensing membrane 22 when the sensing membrane tray 32 is loaded in second bottom cartridge 14B and is secured thereto. Fluid reagents that enter the reagent inlet 36 fills the reagent chamber 38. The regent chamber 38 includes a reagent chamber window 40 that is optically transparent (e.g., acrylic). In this regard, the camera 56 of the reader device 50 (discussed below) is able to image the sensing membrane 22 during the assay with the second top cartridge 16B secured to the second bottom cartridge 14B. Or, alternatively, the user may monitor the coverage of the reagent fluid over the sensing membrane 22. The fluid in the reagent chamber 38 covers the sensing membrane 22 when the second top cartridge 16B is secured to the second bottom cartridge 14B and the reagent fluid is loaded therein. Advantageously, the first and second top cartridges 16A, 16B may be detachably coupled with their respective first and second bottom cartridges 14A, 14B. For example, the first top cartridge 16A and the second top cartridge 16B may include one or more posts, detents, or bosses 42 that interface with a slot or recess 44 contained in the first and second bottom cartridges 14A. 14B. In this way, the first and second top cartridges 16A.16B are detachably connected to the first and second bottom cartridges 14A, 14B by twisting the first / second top cartridges 16A, 16B onto the first and second bottom cartridges 14A, 14B.

[0056] The system 10 includes a reader device 50 that is used to take images of the sensing membrane 22. The reader device 50 includes a housing 52 that has stands 54 that allow the housing 52 to be elevated off of the surface that the housing 52 stands on. The housing 52 that includes a camera 56 as part of a camera module 57 that is configured to obtain image(s) of the sensing membrane 22 when the same is positioned in the housing 52 as explained herein. A cover 59 is located above the camera module 57 on the housing 52 and keeps the interior of the housing 52 dark and substantially free of ambient light. A lens or lens set 58 that is interposed along the optical path between the camera 56 and the sensing membrane 22 when positioned in the housing 52. The lens or lens set 58 may, in some embodiments, be adjustable so as to adjust the focus or the li e. The reader device 50 include a cartridge tray 60 that slides into and out of the housing 52 in the direction of arrow A of FIG. 1 A. The cartridge tray 60 is dimensioned to hold or accommodates the cartridge subunits 14 A, 14B, 16 A, 16B. In one preferred embodiment, the cartridge tray 60 holds or accommodates the second bottom cartridge 14B that holds the sensing membrane 22. Thus, the cartridge tray 60 receives the vertical flow assay cartridges which in a preferred embodiment includes just the second bottom cartridge 14B so as to place the sensing2025-128-2membrane 22 within the field of view of the camera 56 of the reader device 50.Alternatively, in another embodiment, the slidable or insertable cartridge tray 60 receives the assembled cartridge that includes the second top cartridge 16B combined with the second bottom cartridge 14B. Insertion of the cartridge tray 60 with the cartridge 14B (or assembled cartridge) into the reader device 50 also blocks ambient light from entering the interior of the housing 52 of the reader device 50 which would interfere with the images of the sensing membrane 22. In other embodiments, the second bottom cartridge 14B could be secured to the reader device 50 using, for example, one or more posts, detents, or bosses 42 located in the housing 52 that interface with the slot or recess 44 contained in the second bottom cartridge 14B. In this embodiment, the second bottom cartridge 14B may be twisted onto the reader device 50 for imaging.

[0057] The reader device 50 further includes a microcontroller 62 such as a Raspberry Pi. In one embodiment, a Raspberry Pi device is used as the microcontroller 62 although the invention is not limited. The invention may use another microcontroller 62. The microcontroller 62 is used to control parameters of the camera 56, acquire images of the sensing membrane 22, as well as interface with a graphical user interface (GUI) 66 on a display 64 that is also associated with the reader device 50. The microcontroller 62 may include one or more metallic heat sinks (not illustrated) located on the heat-emitting components (e.g., CPU / GPU core, microchips, etc.) thereof for thermal management. The display 64 may include a touch screen in which the user may interface with the device using the GUI 66. For example, this may be used to adjust or alter parameters of the camera 56 (i.e., exposure time, total number of images, and time interval), view images of the sensing membrane 22 and data related thereto (e.g., intensity or other data), patient information, sample information, and sample results. The GUI 66 may be provided that allows the user to adjust imaging parameters for optimal imaging. The display 64 can also display quantitative assay results to the user or other qualitative information (e.g., sample is indeterminate).

[0058] In one embodiment, the reader device 50 is associated with a computing device 70 with one or more processors 72 that execute software 74 or instructions that execute an algorithm including machine learning which may include one or more trained neural networks configured to generate one or more outputs that diagnose / classify the sample and / or quantify the amount or concentration of the one or more analytes or biomarkers in the sample. In one embodiment, the computing device 70 communicates with the microcontroller 62 via a wired or wireless connection. Images obtained with the camera 562025-128-2may be transmitted to the computing device 70 for further processing and analysis. Results obtained by the computing device 70 may be communicated or transmitted back to the reader device 50 via the microcontroller 62. In another embodiment, the computing device 70 may be associated with or integrated with the reader device 50. In still another embodiment, the functionality of the computing device 70 may be carried out the by microcontroller 62 or even a mobile device like a Smartphone that can be secured to the reader device 50 (and provide imaging functionality).

[0059] The software 74 may be used to process or analyze the images of the sensing membrane 22 including image analysis like segmentation and quantifying the intensity of the various spots 24 located on the sensing membrane 22. The software 74 may also perform the quality control / assurance operations on the test spots 24. Finally, the software 74 may execute the algorithm, program, or sequence of instructions that includes machine learning or the one or more trained neural networks 80 including, for example, a first neural network 80 (DNNClassification) that classifies a concentration of the one or more analytes in the sample either above or below one or more threshold values. As explained herein, the DNNClassificationnetwork 80 classifies the sample of interest into three cTnl concentration ranges: <40 pg / mL, 40-1000 pg / mL, and >1000 pg / mL.

[0060] One or more additional neural networks 80 are executed by the software 74 for analyte or biomarker quantification. As explained herein, in one embodiment, there are three cTnl quantification networks 80: DNNQ<40, DNNQ40–1000, or DNNQ>1000, each independently optimized and trained to accurately quantify cTnl concentrations within their respective concentration ranges. The computing device 70 may include a desktop computer, laptop computer, remote computer (e.g., server), or one or more processors contained or associated with the reader device 50.

[0061] To use the system 10 for detecting the presence of and / or quantifying the amount or concentration of one or more analytes or biomarkers in a sample, the first top cartridge 16A is secured to the first bottom cartridge 14A and after wetting the stacked membrane layers 18 and the capture antibodies on the sensing membrane 22 by adding 200 pL of running buffer, a mixture of 50 uL serum sample and 50 pL detection conjugate is added. As noted herein, to achieve high-sensitivity detection of cTnl using the VFA platform, a detection conjugate consisting of 15 nm AuNPs conjugated with PolyFIRP-Streptavidin (PolyHRP) and biotinylated detection antibodies was used (i.e., detection antibody conjugate). While AuNP detection antibody conjugates were used it should be appreciated2025-128-2that other particles or beads may be used and these may be conjugated to different capture agents. The particles may include nanometer-sized particles (NPs) but larger particles may also be used. These may include antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, peptoids, etc. It should be appreciated that the detection antibody conjugate may include, in a preferred embodiment, a plurality of peroxidase molecules associated with an affinity agent and conjugated to detection antibodies.

[0062] As this mixture flows through the sensing membrane 22, a secondary antigen recognition event occurs upon binding to capture antibodies, resulting in the accumulation of the detection antibody conjugate as a function of antigen levels. Finally, a running buffer is injected to expedite the downward flow of the sample-conjugate mixture toward the absorption layers or pads 18, accelerating the washing process in the stacked papers. During the washing phase, the first top cartridge 16A is switched to a new one (a second first top cartridge 16A that is identical), and then an additional 500 pL of running buffer is added to thoroughly wash away any unbound detection conjugates and serum components from the sensor membrane 22.

[0063] For the CL reaction and imaging phases, the sensing membrane tray 32 that contains the sensing membrane 22 is transferred to the CL reaction / imaging setup utilizing the second top cartridge 16B and the second bottom cartridge 14B. The twisting cartridge assembly mechanism remains the same as in the previous phase (e.g., post, detent, or boss 42 engage with slot 44), ensuring user convenience for operation. Upon adding CL reagent solution (440 uL) into the second top cartridge 16B, the solution is rapidly delivered (<1 s) to the sensing membrane 22, where it uniformly incubates within the reagent chamber 38 that is enclosed by a support stage 34, foam tape gasket, and top window 40. Subsequently, the VFA cartridge, namely the cartridge assembly that includes the second top cartridge 16B and the second bottom cartridge 14B secured to one another, is inserted into the reader device 50 using the cartridge tray 60 and is incubated for 4 min to allow the CL reaction to reach saturation. The CL reaction involves luminol, an enhancer, and hydrogen peroxide (H2O2), catalyzed by horseradish peroxidase (HRP) enzymes present in the detection antibody conjugate. The blue visible light (λmax= 425 nm) generated by this reaction on the spots 24 is transmitted through the transparent acrylic window 40 in the second top cartridge 16B and captured by the camera 56 of the portable reader 50 as the CL signal. The CL signal is imaged over 30 s in one embodiment, and the images are computationally analyzed via neural networks 80 described herein to generate one or more outputs that include a classification of2025-128-2the sample and / or a quantification of the amount or concentration of the one or more analytes or biomarkers in the sample.

[0064] Experimental

[0065] Results

[0066] Design and workflow of the CL-VFA

[0067] The CL-VFA system 10 includes modular top and bottom cartridge subunits 16A, 16B, 14A, 14B that can be joined to provide custom formats for the assay and CL imaging phases. This preserves the key features of traditional VFA systems, such as simplicity, ease of use, and rapid assay times, while enabling a CL-based high- sensitivity assay in the POCT platform.

[0068] A paper-based sensing membrane 22 (12 mm * 12 mm), secured to a tray 32, is designed to be easily moved to the reader device 50 right after the assay (see FIG. 1 A). The sensing membrane 22 includes hydrophilic compartments on the paper surrounded by hydrophobic wax barriers that form spots 24. These barriers help direct and concentrate fluid flow into the reaction zones. There are a total of nine reaction spots 24 on the sensing membrane 22: two (2) are used for cTnl detection (coated with anti-cTnl capture antibodies), one (1) serves as a positive control (coated with secondary antibodies that bind anti-cTnl detection antibodies), and five (5) are negative controls (treated with buffer). The remaining central spot 24 acts as a blank fluid channel to reduce non-specific interactions. To prevent potential interference between the cTnl testing spots 24 and positive control spots 24 during the assay and imaging stages, these spots 24 are positioned at the farthest distance from each other on the sensing membrane 22. It should be appreciated that different numbers and configurations of spots 24 may be used other than the specific implementation discussed above.

[0069] In the VFA system 10, the wicking from the absorption pads 18, combined with the capillary action of engineered paper layers in the first top cartridge 16 A, ensures the uniform distribution of the sample and reagents as they pass through the stacked layers of paper 28 through the sensing membrane 22 (see FIG. 6). The configurations of the paper layers 28 in the first top cartridge 16A are specifically designed and optimized to enhance this uniformity. While vertical flow dominates the fluid dynamics, the flow diffuser 28b facilitates lateral diffusion during the rapid vertical flow process, enabling even distribution of the sample across the relatively large surface of the sensing membrane 22 (12 mm * 12 mm) despite the short vertical flow path length (a few millimeters). This optimized flow2025-128-2mechanism minimizes variability in signal intensity across the sensing membrane 22, thereby improving the assay reproducibility. The full assay process is completed within 25 min, including 10 min for the immunoassay, 10 min for washing, and 5 min for incubating the CL reaction and capturing the signal (see FIG. ID). For the immunoassay phase, a combination of the first top cartridge 16A and first bottom cartridge 14A using a twisting lock is employed. This includes engaging the post, detent, or boss 42 of the first top cartridge 16A into the slot 44 of the first bottom cartridge 14A. After wetting the stacked paper layers 28 and the capture antibodies on the sensing membrane 22 by adding 200 pL of running buffer, a mixture of 50 pL serum sample and 50 pL detection conjugate is added. This allows the cTnl antigen to bind to the capture antibodies in a sandwich assay format (capture antibody- cTnl-detection antibody), resulting in the concentration of the detection antibody bound to the sensing membrane 22 being dependent on the cTnl levels in the sample. After that, a running buffer (350 pL) is added to help the sample-detection conjugate mixture move downwards and wash away non-specific detection antibodies. During the washing phase, the first top cartridge 16A is switched to a new one, and then an additional 500 pL of running buffer is added to thoroughly wash away any unbound detection conjugates and serum components from the sensor membrane. This is done to prevent non-specific signals and potential interference during the sensitive CL reaction. In some implementations there is no need to switch to a second first top cartridge 16A for subsequent washing.

[0070] For the CL reaction and imaging phases, the sensing membrane tray 32 (with the sensing membrane 22) is transferred to the CL reaction / iniaging setup, utilizing the cartridge assembly that includes the second top cartridge 16B and second bottom cartridge 14B. The twisting cartridge assembly mechanism remains the same as in the previous phase, ensuring user convenience for operation. Upon adding CL reagent solution (440 pL) into the second top cartridge 16B, the solution is rapidly delivered (<1 s) to the sensing membrane 22, where it uniformly incubates within the reagent chamber 38 that is enclosed by the support stage 34, foam tape gasket 35, and top window 40 (FIG. 1 C). Subsequently, the VFA cartridge assembly is inserted into the reader device 50 (see FIG. ID and IB) and incubated for 4 min to allow the CL reaction to reach saturation. The CL reaction involves luminol, an enhancer, and hydrogen peroxide (H2O2), catalyzed by horseradish peroxidase (HRP) enzymes present in the detection antibody conjugate. The blue visible light (λmax= 425 nm) generated by this reaction is transmitted through the transparent acrylic window 40 in the second top cartridge 16B and captured by the camera 56 of the reader device 50. The CL signal is imaged over 302025-128-2s, and the images of the sensing membrane 22 are then processed by software 74 in computing device 70 to segment the test, positive control and negative control spot signals, which are computationally processed via neural networks 80 to determine the cTnl concentration in the sample (see FIG. IE). This setup provides stable CL reaction and imaging conditions, thereby enabling precise and reproducible detection of cTnl at the pg / mL level.

[0071] PolyHRP-based high-sensitivity detection conjugates for CL-VFA

[0072] To achieve high-sensitivity detection of cTnl using the VFA system 10, a detection conjugate was employed consisting of 15 nm AuNPs conjugated with PolyHRP-Streptavidin (PolyHRP) and biotinylated antibodies (FIG. 2A). Other types of particles (e.g., NPs) may also be used in other embodiments. PolyHRP, an engineered polymer containing multiple HRP and streptavidin (StA) molecules attached to a hydrophilic backbone, greatly enhances the CL signal generation compared to conventional single-molecule HRP -based conjugates. It was hypothesized that PolyHRP-based conjugates, with their higher enzyme density per conjugate, would provide a significant sensitivity advantage over traditional HRP conjugates, where only a limited number of enzymes can be bound to a single nanoparticle or antibody. By increasing the number of active enzymes per conjugate, a stronger CL signal was achieved, resulting in improved detection limits for cTnl.

[0073] The synthesis of the conjugate follows a two-step process. First, PolyHRP is adsorbed onto the surface of 15 nm AuNPs. In the second step, the AuNP -PolyHRP complex is coupled with a biotinylated detection antibody through the strong affinity binding (with a dissociation constant, Kd~ 10−15M) between StA and biotin. The remaining biotin binding sites in StA are then blocked with biotin-BSA, which is smaller in size than the biotinylated antibody. While PolyHRP has traditionally been used as a sole assay label in standard laboratory techniques like enzyme-linked immunosorbent assay and western blot, the conjugation strategy uniquely incorporates 15 nm AuNPs as the core material. This approach offers several key advantages: (i) a compact conjugate size suitable for paper-based assays, allowing for easy transport through the porous paper matrix; (ii) a larger surface area that promotes efficient PolyHRP adsorption; (iii) a simplified conjugate synthesis process that relies on centrifugal washing rather than dialysis, reducing both time and the risk of protein loss, and (iv) AuNP -binding based analysis for more effective quality control, enabling simple spectral scanning and absorption peak measurements to assess the conjugation results and concentration titers.2025-128-2

[0074] As shown in FIG. 2B, the CL signal of AuNP-PolyHRP-antibody conjugate (detection cut-off: 6.1 × 10−5OD) demonstrated significantly superior analytical performance compared to both its native colorimetric signal and the CL signal of the standard HRP conjugate (AuNP-standard HRP-antibody). Compared to the conventional colorimetric signal from AuNPs (detection cut-off: 2.1 × 10−1OD), the CL signal of AuNP-PolyHRP-antibody conjugate achieved a -3400-fold increase in detection sensitivity. Similarly, it exhibited a -210-fold improvement over the CL signal of the standard HRP conjugate (detection cut-off 1.3 × 10−2OD), supporting the hypothesis.

[0075] Although the AuNP conjugate has a relatively small size of approximately 60 nm, optimizing its application in the VFA for cTnl detection required ensuring that it effectively passes through the paper layers 28 in the first top cartridge 16 and sensing membrane 22 while minimizing non-specific signals. To achieve this, various assay conditions were systematically tested and optimized, including the assay protocol, running buffer composition, types of paper materials, and membrane-blocking reagents / concentrations. Three key factors were identified that maximized the signal in positive samples (10 ng / mL of cTnl spiked in cTnl-free serum) while minimizing non-specific signals in negative samples (cTnl-free serum): (i) loading the serum-conjugate mixture in the first top cartridge 16A to increase immunoassay efficiency, and using a separate top cartridge 16 solely for the washing step (Test 1, FIG. 7A and Table 1), which reduced non-specific binding and signal variations by effectively removing unbound conjugates, (ii) incorporating Triton X-I00, a nonionic surfactant, into the running buffer (Test 2, FIG. 7A and Table 1) to enhance flow rate and washing efficiency, and (iii) using a 0.45 pm pore size nitrocellulose (NC) membrane as the material for the sensing membrane 22 (Test 3, FIG. 7A and Table 1), which provided better permeability for the conjugate compared to the 0.22 pm membrane. These optimized conditions led to the effective washing of the unbound conjugate within the CL-VFA structure (FIG. 7B), establishing them as the standard setting for the CL -VFA assay.2025-128-2Table 1No. Manipulated Control Test 1 Test 2 Test 3 Variables1 Assay Protocol Top 1: serum sample Top 1: serum + conjugate mixture loading loadingTop 2: conjugate Top 2: running buffer only (for washing) loading2 Running buffer 1% (v / v) BSA + 1% (v / v) Tween 0.5% (v / v) Triton X-100 + composition in lx PBS 1% (v / v) BSA + 1% (v / v) Tween in lx PBS3 Sensing 0.22 pm nitrocellulose 0.45 pm membrane nitrocellulosematerial

[0076] High-sensitivity CL imaging using a portable quantitative reader

[0077] The CL-VFA system 10 uses a custom-designed portable reader device 50 with a Raspberry Pi microcontroller 62 (see FIGS. 1A-1B). It incorporates cost-effective parts such as a CMOS camera module as part of the camera 56, lens 58, touchscreen display 64, 3D printed housing 52, and cartridge tray 60. Unlike previous colorimetric or fluorescent readers, which require reflectance imaging or fluorophore excitation through external light sources, the CL-VFA portable reader design benefits from the nature of CL, which generates light as a result of a chemical reaction with a high signal-to-noise (S / N) ratio. This eliminates the need for light-emitting diode-related components and simplifies the overall configuration of the reader device 50, reducing complexity in the layout.

[0078] One key advantage of the CL modality in high-sensitivity biosensing is the ability to easily balance sensitivity and dynamic range by controlling exposure time and camera gain during the readout of the image sensor in the camera 56. For example, longer exposure times can accumulate faint CL signals from low antigen concentrations, while shorter exposures help avoid signal saturation when measuring higher antigen concentrations, enabling accurate quantification over a larger dynamic range (FIG. 8 and Table 2).2025-128-2Table 2No. Exposure Time (s) LoD(pg / mL)1 10 10.7 2 30 0.163 60 1.4

[0079] The GUI 66 is designed to enable users to adjust the imaging exposure time (300 ps - 239 s) (see FIG. 12), providing flexibility in optimizing sensitivity. Additionally, it supports time-lapse imaging by enabling the user to set both the time intervals between image capture of the camera 56 and the total number of images, making it suitable for tracking CL signal dynamics over time. To minimize the impact of environmental light during these extended exposures, the reader device 10 uses an enclosed cartridge tray 60 positioned within the housing 52 that offers dark conditions for imaging the VFA cartridge. The cartridge tray 60 also minimizes potential rotational and positional variabilities of the cartridge location during the operation of the test by providing a stable fixation of the CL-VFA cartridge throughout the imaging process. Additionally, to mitigate heat generation from prolonged exposure, which could lead to interference or image distortion, a thermal pad / sheet (e.g., made from a metal such as copper or aluminum) that functions as a heatsink was integrated into the board that holds the Raspberry Pi microcontroller 62 and the CMOS image sensor of the camera 52, thus enabling passive but efficient thermal management throughout the readout process. Furthermore, the Raspberry Pi-based reader device 50 can integrate with a battery pack (not shown), potentially enabling operation in diverse settings without relying on a power grid. The reader can also be equipped with a wireless communication module to rapidly transfer the captured CL images to a computing device 70 such as a central lab server, enabling real-time connectivity with the laboratory infrastructure. Therefore, testing results can be centrally stored on a remote server along with other patient -related information, facilitating access and management of medical data.

[0080] To assess the performance of the Raspberry Pi-based CL portable reader device 50, the S / N ratio was measured of the CL signals captured by the camera 56 from the serially diluted / immobilized detection conjugate (AuNP-PolyHRP-antibody conjugate) on the sensing membrane 22. The S / N ratio of the Raspberry Pi-based reader device 50 was compared with a conventional benchtop system (cooled CCD-based) and a smartphone-based reader (CMOS-based, used for conventional colorimetric setting for VFA) for CL imaging (see FIG. 2C). The Raspberry Pi-based CL portable reader device 50 demonstrated superior2025-128-2performance, achieving high sensitivity in detecting CL signals generated by the conjugate as low as 6.1×10−5OD. This sensitivity is 2.1-fold higher than the detection cut-off of the benchtop system (1.3 * 10‘4OD) and 26.2-fold higher than that of the smartphone-based reader (1.6×10−3OD), highlighting its effectiveness in low-concentration analyte detection and suitability for high-sensitivity applications in point-of-care testing.

[0081] The Raspberry Pi-based CL portable reader device 50 has a compact design that places the sensing membrane 22 at a distance of -2 cm from the aperture of the camera 56, significantly closer than the tens of centimeters typical in benchtop systems (see FIG. 2D). This proximity enhances the sensitivity by reducing light loss associated with greater distances, resulting in more efficient CL signal capture. As a result, the Raspberry’ Pi-based reader devices 50, which are compact, lightweight, and cost-effective, provide field-portable alternatives to larger, more complex benchtop systems, demonstrating competitive sensitivity for the detection of low analyte concentrations. In addition, the Raspberry' Pi-based CL portable reader 50 achieved a better S / N ratio compared to a Smartphone-based reader with similar cost, size, and image sensor type. This improved performance can be attributed to several factors, such as enhanced sensor control options (i.e., optical and digital gain), larger pixel size (2.4 times larger), and better thermal management than the smartphone-based setup.

[0082] Tray-based assay cartridge for stable CL imaging

[0083] The tray-based assay cartridge was engineered to address signal stability issues observed when the first bottom cartridge 14A, initially used during the assay and washing steps, was reused for CL imaging. In the original setup, the sensing membrane 22 was placed above stacked absorbent pads 18 (see FIG. 2E, left), which worked well for the assay but led to unstable CL signals during imaging. Specifically, it was found that the CL intensity failed to reach saturation and decreased over time with increased signal variation (average C V of 22.9%; see FIG. 2F, w / Absorption pad 18). It was hypothesized that this issue arose because, despite the absorbent pads 18 reaching their maximum capacity (~1 mL) during the assay and washing steps, evaporation through the ventilation hole continued, causing the CL reagent on the sensing membrane 22 to continue to flow dynamically. As a result, the reagent was gradually reabsorbed into the pads 18, destabilizing the CL signal.

[0084] To validate this hypothesis, an experiment was conducted where the sensing membrane 22 was carefully excised from the absorbent-based first bottom cartridge 14A following the assay / washing step and transferred it to a flat-bottom cartridge without2025-128-2absorbent material (see FIG. 10A). The CL reagent was then added and the sensing membrane 22 was covered with a glass cover to prevent reagent flow or evaporation, and conducted CL imaging. Under these stabilized conditions, it was observed that the CL signal reached saturation in 4 min with minimal signal deviation (average CV of 4.1%), confirming that stabilizing the reagent without flow on the sensing membrane 22 is a critical parameter for consistent CL imaging (see FIG. 10B).

[0085] To maintain the efficient assay / washing functionality of the original VFA cartridge while enabling stable CL imaging, a tray -based interface was designed that allows users to seamlessly transfer the sensing membrane 22 for imaging. The new imaging cartridge which is the second bottom cartridge 14B incorporates plastic support instead of absorbent pads, preventing reagent movement and evaporation (see FIG. 2E, right). This setup preserves the original assay performance (see FIG. 2G) while providing a stable, optimized environment for CL imaging (average CV of 4.6%; see FIG. 2F, w / Support stage). By combining the stability of the flat-bottom cartridge conditions with enhanced usability, this tray-based cartridge enables hs-cTnl measurement in a user-friendly, practical format, as further described in the following section.

[0086] High-sensitivity assay performance of cTnl CL- VFA

[0087] The limit of detection (LoD) is a critical parameter for assessing the sensitivity of biomarker assays. To determine the LoD for the optimized CL- VFA system 10, titration experiments were conducted using cTnl spiked in cTnl-free human serum. Serial dilutions were performed using the same serum medium, and the corresponding CL signals were captured using the portable reader. The signal intensity (INormalized) was calculated using the averaged test and negative control spot signals. The details are described in the Computational analysis of CL-VFA signals in the Methods section.

[0088] The detection performance of this system 10 is illustrated in FIGS. 3A-3E. Using the Raspberry Pi-based portable reader device 50, CL images were acquired at varying cTnl concentrations, resulting in apparent differences in concentration-dependent test spot signals (see FIG. 3 A). The calibration plot in FIG. 3B demonstrates a strong correlation (R2= 0.99, y=0.0147x0.4043) between the signal intensity and cTnl concentrations in the serum matrix, ranging from a few pg / mL to 105pg / mL (matching the clinically relevant range). The assay also exhibited excellent precision, showing an average CV of 5.4% between triplicate testing repeats of each sample. These results highlight both a broad sensing range of over 6 orders of magnitude and the reliability of this assay system.2025-128-2

[0089] To further assess the assay's sensitivity, its performance was evaluated in detecting lower cTnl concentrations (0-10 pg / mL) within the clinical cut-off level range (-40 pg / mL), as shown in FIG. 3C. The -values derived from Ltests confirmed statistically significant differences (P <0.05) between the different cTnl levels under 10 pg / mL; for example, negative vs. 0.5 pg / mL (P = 0.0019) comparison underscores the platform's outstanding sensitivity and sensing resolution at or below the pg / mL level for high-sensitivity detection of cTnl.

[0090] The analysis indicates that the limit of detection (LoD) for the cTnl CL-VFA is 0.16 pg / mL, defined as LoD = Limit of Blank (LoB) + 1.645 * Standard Deviation (SD) of the lowest cTnl measurement. LoB is calculated as the Mean blank value + 1.645 * SD of the blank. The values measured were a mean blank of 0.0046, an SD of the blank equal to 0.00071, and an SD for the lowest cTnl concentration of 0.00073, with this lowest concentration being 0.5 pg / mL. The LoB was estimated to be 0.101 pg / mL by converting the CL intensity (y -value) into concentration using the optimal fitting curve shown in FIG. 3B (y=0.0147x0.4043). Subsequently, the LoD calculation resulted in 0.16 pg / mL through the same calibration curve (see Table 2).

[0091] These findings, along with the CL-VFA system’s performance, meet two critical criteria set by clinical guidelines for hs-cTnl testing: (i) a CV of <10% at the clinical cut-off, namely 99th percentile upper reference limit, and (ii) the capacity to identify cTnl levels at or over the assay's LoD in more than 50% of healthy patients (i.e., - 40 pg / mL). This confirms that the methodology conforms to established clinical standards for hs-cTnl assays.

[0092] 72 serum samples were measured obtained from patients tested for cTnl levels at UCLA Health for practical validation using the CL-VFA system 10(see FIGS. 3D and 3E). The ground truth cTnl concentrations for these samples were measured and provided by UCLA Health using an FDA-cleared benchtop analyzer, enabling a hs-cTnl assay. This clinically used benchtop system has an analytical cut-off value of 4 pg / mL, where samples with cTnl levels above 4 pg / mL are quantified, and those below this level are reported as <4 pg / mL. In the following sub-section, these ground truth levels will be compared to the cTnl concentrations inferred by the neural network-based analysis approach of the system 10, allowing the assessment of the accuracy and reliability of the CL-VFA system 10 in detecting clinically relevant cTnl levels.

[0093] The assay plots in FIGS. 3D and 3E demonstrate a strong correlation (R2= 0.99, y=0.0048x0.5) between the ground truth cTnl concentrations and CL intensity measured by2025-128-2the CL-VFA system 10 across the entire quantifiable concentration range (>4 pg / mL), with an average CV of 6.7%, demonstrating consistent inter-sensor measurement repeatability. Although the ground truth values of several samples were labeled as <4 pg / mL, minute differences in CL intensities were detected between these samples using the CL-VFA system 10, revealing subtle variations in cTnl levels that were otherwise indistinguishable by the benchtop analyzer. During the analysis, three samples within the 102–103ground truth level range were identified as outliers, as indicated in FIG. 3D. Outliers were defined as samples displaying a signal intensity difference corresponding to a concentration deviation of at least one order of magnitude from the expected trendline. These outliers and three randomly selected normal samples that followed the expected trend line were reanalyzed using the same clinical benchtop analyzer to ensure accurate validation and inquire if samples degraded over time after their initial clinical measurements. Upon revalidation (see FIG. 11), normal samples exhibited minimal changes in cTnl levels, ranging from 3.5% to 11%, consistent with their initial measurements. In contrast, outlier sample 1 showed an extreme degradation rate exceeding 99%, with its concentration decreasing to <2 pg / mL from an initial 512 pg / mL, which agrees with the <4 pg / mL intensity range detected by the CL-VFA—strongly indicating significant degradation of the sample. Outlier sample 2 exhibited a distinct discrepancy: despite showing a moderate degradation rate of 13.2%, its CL-VFA intensity fell within the <4 pg / mL range. Notably, the serum appeared darker red and more turbid compared to other samples in the same collection batch (typically yellowish), suggesting elevated levels of hemolysis and lipemia. These characteristics may have introduced assay interference, with a possible additional factor being the presence of autoantibodies against cTnl, which may mask target epitopes and interfere with the assay's ability to detect cTnl. Outlier sample 3 presented a more complex case, displaying characteristics seen in both samples 1 and 2. Outlier sample 3 showed a substantial degradation rate of 40%, indicating a notable decrease in troponin concentration. However, as with sample 2, additional assay interferents likely affected the VFA reading, categorizing it within the <4 pg / mL range.

[0094] These findings suggest that these three (3) outlier samples, affected by either excessive degradation or potential assay interference, showed cTnl measurement inconsistencies compared to the stable readings in normal samples. Due to these discrepancies, the outliers were excluded from further neural network analysis to maintain the reliability of the model's predictions. The trained neural network models 80 were utilized on2025-128-2the validated CL-VFA signals to accurately quantify cTnl levels, as discussed in the next subsection.

[0095] Neural network-based cTnl quantification using CL-VFA data

[0096] Deep learning and neural networks 80 enhance the performance of point-of-care sensors by effectively learning complex relationships between sensor output patterns and diagnostic outcomes, outperforming standard rule-based methods (e.g., linear regression, power fitting), which lack the complexity needed to approximate intricate functional relationships. A neural network-based pipeline was developed to accurately quantify cTnl concentrations in clinical serum samples using the signals captured by the image sensor in the camera 56 of the CL-VFA system 10. This pipeline consists of four fully connected neural networks (see FIG. 4A): one for classification (DNNClassification) and three for biomarker quantification (DNNQ<40, DNNQ40–1000, and DNNQ>1000). These neural networks 80 work together to quantify cTnl over a large dynamic range of the clinical sample cohort, which spans from a few pg / mL to -18,000 pg / mL, as shown in FIG. 3D.

[0097] First, the DNNClassificationnetwork 80 classifies the sample of interest into three cTnl concentration ranges: <40 pg / mL, 40-1000 pg / mL, and >1000 pg / mL. This classification is based on the averaged test, negative control, and positive control signals (inputs) generated by the CL-VFA. The inputs correspond to measured intensity levels of the various spots 24 which in some cases may be averaged or normalized values as explained herein. Following this initial classification, the sample of interest is processed by one of the three cTnl quantification networks 80: DNNQ<40, DNNQ40–1000, or DNNQ>1000, each independently optimized and trained to accurately quantify cTnl concentrations within their respective concentration ranges. In the <40 pg / mL (DNNQ<40) and 40-1000 pg / mL (DNNQ40–1000) categories, these networks 80 utilize the normalized CL signal (INormalized) as input, while the >1000 pg / mL network (DNNQ>1000) takes the averaged test, negative control, and positive control signals for quantification (see Computational analysis of CL-VFA signals in Methods section for details). [Normalized incorporates test and negative control spots 24; however, for higher concentration ranges (i.e., >1000 pg / mL), the test spots 24 exhibit more interference with negative controls due to the strong CL. intensity, which causes the light to spread and reach adjacent areas. To account for this, DNNClassificationand DNNQ>1000also take positive control spot 24 signals for additional signal normalization.

[0098] In case of discrepancy between the concentration prediction from the quantification network 80 and the output from DNNClassificationnetwork 80 (i.e., if the predicted2025-128-2concentration falls more than 10% outside the borders of the range predicted by DNNClassification), the sample is marked as "indeterminate" due to the conflicting decisions of successive neural networks 80, and no cTnl concentration inference is performed.Quantification occurs only when predictions from the DNNClassificationand the corresponding cTnl quantification network 80 align with each other. As a result, cTnl quantification measurement is done through this collaboration between the four neural networks 80 (see Neural network-based cTnl quantification pipeline in Methods for details about these models).

[0099] Each of the four neural networks 80 in the cTnl quantification pipeline was individually trained and optimized using distinct training and validation datasets tailored to their target cTnl concentration ranges. FIGS. 5A-5H is a table showing detailed information of the clinical sample test data. FIGS. 20A-20D illustrate how the four different networks 80 were trained using ground truth values to predict classification or quantification results. The parameters of the networks or models 80 were trained using backpropagation on training sample sets.

[0100] When tested on the validation set, DNNClassificationachieved 93.5% accuracy, and the optimized quantification models 80 resulted in a strong correlation (Pearson's r of 0.993) with the ground truth values from an FDA-approved clinical laboratory analyzer, with an average CV of <10%, demonstrating the reliable performance of all four models and good inter-sensor repeatability (see FIGS. 12A-12B). Additional details about the neural network’s 80 performance on the validation set are available in the Neural network-based cTnl quantification pipeline sub-section of the Methods section.

[0101] Next, this neural networks-based pipeline was blindly tested using 66 new serum samples from 34 patients, none of which were used during the training and optimization phases (FIG. 11). As shown in FIG. 4B, DNNClassification80 achieved 95.5% accuracy on this blind testing set, correctly classifying all samples in the 40-1000 pg / mL and >1000 pg / mL ranges. However, three samples with ground truth cTnl concentrations of 32 pg / mL (2 samples) and 37 pg / mL (1 sample) were misclassified from the <40 pg / mL range into the 40-1000 pg / mL range.

[0102] Following the classification stage, samples from the <40 pg / mL, 40-1000 pg / mL, and >1000 pg / mL ranges were further quantified by the respective quantification neural networks 80: DNNQ 40, DNNQ40–1000, and DNNQ>1000. The combined cTnl concentration predictions from these neural networks 80 showed a good match with the ground truth cTnl2025-128-2levels from an FDA-approved analyzer, with a Pearson's r of 0.984 (see FIG. 4C). In addition, cTnl concentration inferences showed good precision with an average CV of 14.3% between duplicate tests of patient samples. Some of the samples quantified by DNNQ<4080 had ground truth cTnl concentrations below 4 pg / mL and did not have quantitative concentration labels due to the cut-off level of the FDA-approved clinical analyzer.Quantitative predictions for such samples were plotted within the <4 pg / mL region of the ground truth axis (i.e., x-axis), as shown in FIG. 4C. No samples in the blind testing set were labeled as "indeterminate" as all quantification predictions aligned with the classification network results. However, three samples with ground truth concentrations in the <40 pg / mL range were misclassified into the 40-1000 pg / mL range by DNNciassifieation 80, which were successively quantified by DNNQ40–100080, resulting in the predicted cTnl concentrations of 71 pg / mL, 105 pg / mL, and 72 pg / mL; for these samples, the ground truth concentrations were 32 pg / mL (2 samples) and 37 pg / mL (1 sample), respectively. Erroneous predictions by the classification and quantification models for these samples can be attributed to increased non-specific binding relative to the other clinical serum samples, which elevated the test spot signals, creating confusion with samples of higher concentrations.

[0103] Note that the decision to implement a multi-network architecture for the cTnl quantification pipeline, incorporating four neural networks 80, was based on evaluating several strategies with varying numbers of neural network models 80. For example, a single quantification model 80 proved inadequate for reliably quantifying cTnl concentrations over the entire clinically relevant range due to the large dynamic range of cTnl levels observed in patient samples (from a few pg / mL to over 104pg / mL) and the limited number of training samples (see FIG. 13). Therefore, to improve performance, a cascaded approach was used, where all samples are first classified into narrower concentration subranges, and each subrange is further quantified by a dedicated cTnl quantification model. Two configurations were compared: one using three models 80 (1 classification and 2 quantification - FIGS. 14A-14C) and another using four models (1 classification and 3 quantification). Ultimately, as shown in FIG. 4C, the four-model configuration was selected for its superior quantification accuracy (Pearson's r of 0.984) compared to the three-model configuration (Pearson's r of 0.846 - FIG. 14C). Additionally, this four-model configuration can minimize the risk of overfitting, which can occur with limited data, while also increasing flexibility for incorporating new data in the future. The borders for the concentration subranges were selected to ensure a balanced representation of the samples within all ranges.2025-128-2

[0104] Importantly, the results highlight that systematic error mitigation steps, such as digital signal averaging and spot quality control, are critical for enhancing the analysis performance of CL-VFA system 10. For example, averaging signals across similar spots 24 within the test, negative control, and positive control conditions improved quantification precision, minimizing variability between individual spots caused by sensor fabrication errors, nonuniform flow, or CL signal leakage between adjacent spots. Using single-spot 24 inputs instead of averaged signals increased the CV from 14.3% to 16.7% across the quantification range, with a notable rise in the lower concentration range (<1000 pg / mL), where the CV increased from 15.3% to -20%, indicating reduced precision (see FIGS. 15A-15B). In addition, negative control spots 24 underwent a digital quality control step, which excluded individual spots 24 (among the five original spots) from averaging if they fell outside the 95% confidence interval (see Computational analysis of CL-VFA signals in the Methods section for more details). Omitting this quality control step lowered the classification accuracy of DNNClassificationto 94%, primarily due to the additional misclassification of one borderline sample from the 40-1000 pg / mL range (with a ground truth cTnl concentration of 51 pg / mL) into the <40 pg / mL range (see FIG. 16). Therefore, incorporating signal averaging and digital quality control enhanced the robustness of CL-VFA, ensuring accurate diagnostics outcomes despite the low-cost nature of the VFA system 10.

[0105] It is also noted that neural network-based cTnl quantification outperforms conventional deterministic methods in accuracy and precision. In this case, the optimal rule¬ based method consists of three independently optimized power-fitting curves, each tailored to a specific cTnl concentration range (<40 pg / mL, 40-1000 pg / mL, or >1000 pg / mL). For a fair comparison, these power-fitting models were trained using the same dataset as the quantification neural networks 80 and were blindly tested on identical blind testing sets.?\s shown in FIG. 17, cTnl concentration predictions by the rule-based models (i.e., powerfitting curves) revealed a weaker correlation with the ground truth concentrations (Pearson's r of 0.912) compared to the optimal neural network models 80 (Pearson's r of 0.984).Moreover, the CV between the duplicate tests was >21% for the power-fitting models, substantially higher than that observed with the performance of the neural networks 80. This inferior performance likely stems from the simplicity of the power-fitting models, which lack the large number of trainable parameters available in shallow neural networks 80, limiting2025-128-2their ability to capture complex functional relationships between input signals and target cTnl concentrations.

[0106] Similarly, DNNClassification80 outperforms other machine learning algorithms, such as random forests and logistic regression. For instance, on the same blind testing set, the random forest model achieved only 91.0% accuracy, showing inferior performance (see FIG.18 A). Furthermore, while the logistic regression model matched the overall accuracy of DNNClassification80, it misclassified some samples from the 40-1000 pg / mL range into the <40 pg / mL range, generating false negative outcomes that could overlook an ongoing MI (see FIG. 18B). Such errors pose a greater health risk to patients than false positives when samples with <40 pg / mL ground truth are classified into higher ranges. DNNClassification80 demonstrated superior accuracy for higher concentration samples, correctly classifying all samples in the 40-1000 pg / mL and >1000 pg / mL ranges and mitigating false negative predictions. Therefore, the DNNClassificationneural network model 80 was selected for the classification step of the neural network-based cTnl quantification pipeline.

[0107] Here, a hs-cTnl clinical assay for the ground truth measurements allowed one to obtain accurate cTnl levels even below 40 pg / mL, enabling neural network-based quantification within the 4-40 pg / mL range. This enabled the system 10 to accurately quantify cTnl concentrations below the 40 pg / mL threshold, which is a commonly recognized cut-off for myocardial infarction risk assessment. By facilitating reliable detection at these lower levels, the CL-VFA system 10 has the potential to improve early cardiac event detection and monitoring. For example, it could shorten the interval for confirmatory testing in suspected AMI patients, reducing the wait from 2-3 hours after initial measurement to potentially under 1 hour, for quicker observation of troponin level increases. Additionally, for patients without initial signs of elevated risk, it could support rapid discharge under the 0-hour rule-out strategy, enhancing clinical efficiency and patient throughput.

[0108] Cost analysis

[0109] At the current laboratory scale, the cost per test for the CL-VFA assay, including reagents, paper materials, and the plastic cartridge, is estimated to be $4.25. However, substantial cost reductions are expected with large-scale production practices. For example, cartridge housing, which currently represents about 32% of the total test cost, could transition from 3D printing to injection molding, bringing housing costs close to a few cents. Similarly, antibody, chemical reagent, and raw material costs could be significantly reduced through economies of scale, potentially enabling a per-test cost of under $1—2. This scalability could2025-128-2provide a feasible pathway to making hs-cTnl testing more affordable, allowing for widespread clinical and point-of-care applications, especially in resource-limited settings.

[0110] Additionally, the cost-effectiveness of the Raspberry Pi-based portable reader device 10 further enhances the practical appeal of this system for point-of-care settings.Unlike traditional benchtop imaging systems that cost upwards of $30,000, the Raspberry Pi- based reader device 10, combined with a streamlined CL-VFA cartridges 14A, 14B, 16A, 16B, offers a robust yet affordable alternative (see FIG. 2D). The reader device 10 uses cost-effective components, resulting in a total system cost of ~$222. This affordability does not compromise performance, making it suitable for high-sensitivity diagnostic applications, even in resource-limited settings. The low production costs of the cartridges 14A, 14B, 16A, 16B and the reader device 50 suggest that this platform could be widely deployed for routine testing of cTnl, particularly in decentralized healthcare settings.

[0111] A neural network-enhanced paper-based high-sensitivity CL-VFA 12 was developed for quantitative cTnl detection. The CL-VFA 12 enabled high sensitivity detection of cTnl (LoD of 0.16 pg / mL) with good precision (average CV of 14.3%, over the entire quantification range) within 25 min per test using human serum samples. The design integrates multiple innovative components to achieve hs-cTnl detection, including (i) a PolyHRP -based AuNP detection conjugates that improve detection sensitivity by more than 2 orders of magnitude over conventional HRP -based CL methods; (ii) a cost-effective, handheld Raspberry Pi-based portable CL reader device 50 that outperforms a benchtop CL imaging station despite having a significant cost reduction; (iii) a streamlined tray-based VFA assay cartridge assembly using connectable cartridges 14, 14B, 16A, 16B ensures user-friendly assay operation and stable CL imaging; and (iv) a robust neural network-based concentration inference algorithm that leverages deep learning for accurate quantification across a wide dynamic range (over 6 orders of magnitude), enhancing both the precision and reliability of cTnl measurements.

[0112] Compared to both commercially available benchtop and point-of-care cTnl assays, as well as recent advancements reported in the literature using CL-related sensing modalities, the CL-VFA system 10 demonstrates competitive performance in sensitivity, dynamic range, precision, and portability surpassing some of the traditional benchtop analyzers by an order of magnitude in sensitivity. This work paves the way for low-cost, high-performance point-of-care diagnostics, addressing the critical need for highly sensitive cardiac biomarker testing that has traditionally been limited to central laboratory -based settings. Furthermore, the CL-2025-128-2VFA system 10 holds promise for broad applications across various analytes or biomarkers beyond cTnl. Looking forward, leveraging the inherent parallel-flow capabilities of the VFA, multiplexed CL-VFAs capable of simultaneously detecting cTnl and other key cardiovascular disease biomarkers can be developed. Such an enhancement would enable comprehensive cardiovascular risk assessment in a single test, improving diagnostic precision and patient outcomes through rapid and cost-effective point-of-care testing.

[0113] Given that the current CL-VFA 12 requires blood collection to obtain serum specimens and incorporates manual liquid handling, the system 10, in this specific embodiment, is intended for use by medical professionals in facilities such as rural clinics, nursing homes, and pharmacies. Additionally, the system 10 holds potential applications in hospital clinical laboratories and emergency departments where urgent diagnostic needs arise. In these settings, staff are familiar with handling liquids and reagents as part of their routine tasks. To further simplify the test workflow and enhance its usability, reagent packaging such as premeasured volumes of conjugates or running buffers in disposable containers can be used to make the system more user-friendly. Furthermore, integrating rapid plasma or serum extraction technology for whole-blood testing will support the streamlined implementation of CL-VFA 12 in distributed clinics or other point-of-care settings. These advancements would further extend the CL-VFA’ s practical applications, facilitating rapid and reliable POCT in various healthcare environments.

[0114] Methods

[0115] Conjugate preparation: The conjugation of PolyHRP-Streptavidin (PolyHRP) and biotinylated antibody to 15 nm AuNP was performed through adsorption and affinity binding, respectively. In the first step, for the AuNP conjugation with PolyHRP, 15 nm AuNP (1 mL; BBI Solutions) was mixed with 100 mM borate buffer (100 pL, pH 8.5; Thermo Scientific) and stabilized for 5 min at room temperature (RT). Subsequently, PolyHRP (11 pL, 1 mg / mL; streptavidin-peroxidase polymer, Sigma) was added to the AuNP mixture, followed by incubation for 1 h at RT using a rotary mixer (20 rpm). After incubation, bovine serum albumin (BSA; 10 pL, 10% w / w; Thermo Scientific) was added to block the AuNP surface, and the mixture was incubated for 1 h at RT. The conjugate was then washed three times via centrifugation (25,000g, 30 min, 4 °C), using 10 mM borate buffer (pH 8.5) as the washing medium. After the washing steps, the conjugate pellet was resuspended in 10 mM PBS (100 pL, pH 7.2; Thermo Scientific) for storage as a 10× concentrated conjugate solution, which was subsequently transferred to a new reaction tube.2025-128-2

[0116] In the second step, the conjugation of AuNP-PolyHRP with a biotinylated antibody was performed. Biotin was labeled to the anti-cTnl detection antibody (19C7, Hytest) following the manufacturer's protocol (EZ-Link™ Sulfo NHS-LC-LC -Biotin, Thermo Scientific), and the concentration of the purified biotinylated antibody was quantified using a Nanodrop spectrophotometer (Thermo Scientific). Briefly, 12 pg of biotinylated antibody was added to the AuNP-PolyHRP solution (e.g., the addition of 4.8 pL of 2.5 mg / mL biotinylated antibody), followed by gentle mixing and incubation for 2 h at RT using an orbital shaker (55 rpm). Next, 5 pL of biotin-BSA (5 mg / mL; Thermo Scientific) was added and incubated for 1 h at RT (at 55 rpm on an orbital shaker) to block any remaining biotin binding sites on the streptavidin. Finally, the conjugate was washed three times with 10 mM borate buffer (pH 8.5) by centrifugation (25,000g, 30 min, 4 °C) to remove unbound components. The final pellet was resuspended in 100 pL of storage buffer containing 10 mM Fe(II)-EDTA, 4% (w / w) trehalose, 0.1% (w / w) BSA, and 1% (v / v) Triton X-100 in PBS (10 mM, pH 7.2). Absorption spectra and conjugate concentration were analyzed using a microplate reader (Synergy Hl; BioTek), resulting in a redshift of the peak wavelength after conjugation (from 520 nm to 526 nm, see FIG. 19). Conjugates were stored at 4 °C at 8 OD concentrations until use. For preparing the 15 nm AuNP-standard HRP conjugate, StA-peroxidase conjugate (Sigma) was used instead of Poly HRP, but the conjugation process remains the same.

[0117] Preparation of sensing membrane and assembly of assay cartridges. The sensing membrane 22 was fabricated following a sequential process that included wax printing, heat treatment, antibody deposition, and blocking. Initially, a wax printer (Xerox) was used to print nine active zones (eight reaction areas plus a central flow zone) on an NC membrane (0.45 pm, Sartorius), surrounded by a black background. These wax-printed NC membranes were then baked at 120 °C for 55 s in a forced-air convection oven (Across International), creating defined compartments for reaction sites and fluid flow paths. Post¬ heat treatment, the waxed areas turned hydrophobic, blocking the NC membrane pores, while the unwaxed regions remained hydrophilic to allow samples and aqueous solutions to flow. Up to 30 sensing membranes 22 were arranged in a 6 × 5 grid with 1 mm spacing and processed simultaneously within a single batch.

[0118] To apply antibodies, anti-cTnl capture antibody (0.8 pL, 1 mg / mL; 560, Hytest), goat anti-mouse IgG (0.8 pL, 20 pg / mL; Southern Biotech), and 10 mM PBS (0.8 pL, pH 7.2) were spotted onto the designated test, positive control, and negative control regions,2025-128-2respectively. Sensing membranes 22 were air-dried at 37 °C for 15 min before immersion in a 1% (w / v) BSA blocking solution for 30 min at room temperature (RT). Following blocking, the membranes were dried again at 37 °C for 15 min. Each large membrane sheet was then carefully divided into individual sensing membranes 22 for the assay using a razor and attached to the sensing membrane tray 32 with double-sided adhesive foam tape.

[0119] To assemble the CL-VFA cartridge subunits, the paper materials were prepared according to established methods. Raw materials were precisely cut to the required dimensions using a CO2 laser cutter (Trotec). For the first top cartridge 16A (used for immunoassay and washing steps) assembly, the cartridge 16 A was constructed by layering several paper layers 28, including an absorption layer 28a, flow diffuser 28b, primary spreading layer 28c, interpad 28d, secondary spreading layer 28e, and supporting layer 28f, which were bonded with adhesive foam tape for stability. Wax printing and baking processes similar to those used for the sensing membrane 12 were applied to create concentric circular patterns on both the flow diffuser 28b and the outer frame of the supporting layer 28f. To minimize non-specific binding, the flow diffuser 28b, interpad 28d, and supporting layers 28f were treated with a 1% (w / v) BSA solution. For the first bottom cartridge 14 A, five absorption pads 18 were positioned at the cartridge's center, and the assembled trays were stacked, as shown in FIG. 6.

[0120] CL-VFA cartridges: The sensing membrane tray 32 was 3D-printed using an Ultimaker S3 (PETG filament, Ultimaker) with a layer thickness of 150 pm. The top and bottom cartridges 16 A, 16B, 14 A, 14B for the CL-VFA 12 were fabricated using a Form 3 printer (Formlabs) with gray resin at a resolution of 100 pm. When the first bottom cartridge 14A, sensing membrane tray 32, and first top cartridge 16A are assembled, the design compresses the paper layers 28 to 25% of their original thickness, optimizing flow uniformity and enhancing both assay and washing efficiency. For the CL reaction and imaging, the second top cartridge 16B was created from a transparent acrylic sheet, laser-cut to precise dimensions (16.3 mm × 16.3 mm × 1 mm), and attached to the 3D-printed body using clear acrylic glue. The acrylic window 40 includes a ventilation port (0.7 mm in diameter) positioned at the lower left corner, facilitating air displacement from the reagent chamber 38 as CL reagent solutions are introduced. Upon assembly, the foam tape attaching the sensing membrane 22 to the tray 32 aligns with the extruded edges of the second top cartridge 16B, creating a secure seal. This configuration effectively prevents leakage, evaporation, and displacement of the injected CL reagent solution, eliminating the need for an additional2025-128-2gasket. The reagent inlet 36 of the second top cartridge 16B is designed to hold up to 520 pL of solution.

[0121] Raspberry Pi-based portable CL reader: The custom-designed CL reader device 50 used a Raspberry Pi (model 3b) microcontroller 62 board to control the camera 56 in the camera module through a user-friendly GUI 66. The camera module consisted of a Raspberry Pi HQ camera 56 (Adafruit) coupled to a C-mount lens 58 (Adafruit), enabling the capture of CL reactions over long exposure times of up to 239 s, achieving high sensitivity. The camera 56 was connected to Raspberry Pi microcontroller 62 via a dedicated camera serial interface inlet on the board, eliminating the need for additional software drivers. The GUI 66 displayed a real-time camera feed and enabled users to set the exposure time, number of time-lapse images, and time intervals between subsequent images (see FIG. 9). The reader device 50 was assembled with a Raspberry Pi microcontroller 62 board, camera module (with camera 56), and touchscreen display 64 (Elecrow), enclosed in a 3D-printed case, consisting of the housing 52, camera holder or mount, and cartridge tray 60. The 3D-printed parts were produced by Object 30 (Stratasys) 3D printer. The reader-supported pedestal installation used optical posts for stands 54, allowing both handheld and benchtop use (see FIG. 1A and IB).

[0122] Assay operation: The assay operation involves two primary stages: the immunoassay and the CL reaction / imaging. For the immunoassay stage, the device is assembled by combining the first bottom cartridge 14A, the sensing membrane tray 32, and the first top cartridge 16A. Initially, the combined cartridge assembly (with cartridges 14A and 16A) is activated by introducing 200 pL of running buffer containing 1% (v / v) Tween-20, 0.5% (v / v) Triton X-100, and 1% (v / v) BSA in 10 mM PBS (pH 7.2). After 30 s, when the buffer is fully absorbed, a mixture of 50 pL of serum sample and 50 pL of conjugate (1 OD) is applied and allowed to incubate for 1 min. Following this, 350 pL of running buffer is added to facilitate the flow of solutions through the CL-VFA device, removing any unbound conjugates and target molecules from the sensing membrane to minimize non-specific binding. After 8.5 min, the first top cartridge 16A is removed and replaced with a new top cartridge 16A. An additional 550 uL of running buffer is added for a secondary washing step, which takes 10 min to complete. After the washing step is complete, the sensing membrane tray 32 (with sensing membrane 22) is transferred to the second bottom cartridge 14B, which includes a support stage 34, and then assembled with the second top cartridge 16B. Next, 440 pL of CL reagent solution (SuperSignal™ ELISA Pico Chemiluminescent Substrate, 37069, Thermo Scientific) is added to the assembled cartridge (cartridges 14B and 16B). The2025-128-2assembled cartridge is then placed into a portable reader device 50 using the cartridge tray 60 for 4 min to allow the CL signal to reach saturation. Imaging is then performed for 30 s to capture the CL signal for analysis.

[0123] Through extensive timed experiments with over 300 activated cartridges, the total assay time was established at 25 min. Each step was optimized, with the immunoassay incubation period tailored for complete binding and efficient washing within 20 min, followed by a 5-min phase dedicated to the CL reaction and signalacquisition. Computational analysis was completed in less than 1 s per sample, with trained neural network models 80 inferring cTnl concentrations in under 0.5 s, a negligible duration relative to the overall assay time.

[0124] cTnl spiked and clinical serum samples: To optimize and validate the assay 12, cTnl-spiked serum samples were prepared by adding human heart-derived cTnl standard antigen (I-T-C complex; Lee Biosolutions) into cTnl-free serum (Hytest). Serial dilutions were performed with the same serum to achieve a range of cTnl concentrations. The cTnl antigen w'as dispensed in 1 pL aliquots and stored at -80 °C for stability. Fresh aliquots were thawed and used immediately before each experiment to avoid degradation. Clinical serum samples, obtained from UCLA Health under IRB protocol # 20-002084, involved remnant / exi sting specimens collected separately. The ground truth cTnl levels in clinical samples were measured shortly after sample collection using a hs-cTnl assay (Access 2; Beckman Coulter) at UCLA Health, This clinical analyzer has a detection threshold of 4 pg / mL and reports values below this threshold as <4 pg / mL while accurately quantifying levels >4 pg / mL. In total, 72 clinical serum samples were analyzed with the CL-VFA, encompassing 58 samples with cTnl concentrations of 4 pg / mL or above, and 14 samples with levels below 4 pg / mL (see FIGS. 5A-5H). The clinical samples were stored at -80°C and thawed at 4°C prior to testing.

[0125] Computational analysis of CL-VFA signals: Immunoreaction spots 24 were segmented from the captured images of the sensing membrane 22 obtained with the camera 56 using ImageJ software 74, and the same type of spots 24 within a test, negative control and positive control conditions were averaged, generating 3 output signals per CL-VFA:

[0126] IRaw = [X̄Test, X̄(−), X(+)], (1)

[0127] where XTestand X(_) refer to the averaged test and negative control signals, respectively. X(+) refers to the positive control signal. Normalized signals were calculated as:2025-128-2Ct _ -a _ y

[0128] Ixormalized = 1 - (2)

[0129] Before any averaging, negative control spots underwent digital quality control, eliminating any spots 24 outside the 95% confidence interval. Spots 24 were excluded from X(_) based on the 95% acceptance range from the statistical distribution of all negative control spots 24 from the CL-VFAs activated during the clinical study, excluding the outliers reported in the Results section, i.e.:

[0130] Xfff < X(-Xi< X™f, i = 1 ■ ■ vVNeg, (3)

[0131] where X^°^ and x” ®hare calculated as ±1.96 SD from the mean of the distribution of all negative control spots 24 from the clinical study. X(_yj denotes the signal of the negative control spot 24 with index i and AVeg = 5 is the total number of negative control spots 24 per CL-VFA. X(_j is calculated over the negative control spots 24 that remained valid after this digital quality control step. If all negative control spots 24 in a CL- VFA sample were excluded, the entire sample was removed from the dataset. As a result, 4 CL-VFAs were excluded due to negative control failure, leaving 134 CL-VFAs, which were split between training, validation, and testing datasets for neural network-based cTnl quantification. Digital quality control was not applied to the test and positive control spots 24 since each CL-VFA contained only 2 test spots and 1 positive control spot. Precise spot segmentation through the ImageJ software 70, averaging of the alike spots 24 and digital quality control measures collectively help to minimize inter-sensor variations that may arise due to spotting errors, non-uniformity of the sample flow through paper layers and potential positional misalignments of the CL-VFA cartridge during imaging.

[0132] Neural network-based cTnl quantification pipeline: The neural network-based pipeline for quantification of cTnl in clinical samples consisted of four separate neural network models 80, including one classification (DNN classification) network or model 80 and three quantification (DNNQ<40, DNNQ40–1000, DNNQ>1000) networks or models 80 (FIGS. 4A and 21 A-21D). All networks or models 80 represented shallow fully connected neural networks, which were trained and optimized separately on different portions of the clinical dataset. Inputs to DNNClassification and DNNQ>1000 networks or models 80 represented averaged raw intensity signals from the CL-VFA 12, i.e., Xix = Law. For DNNQ<40 and DNNQ40–1000networks or models 80, input signals represented normalized intensities, i.e., XIN = INormalized.

[0133] Input signals to all networks 80 were standardized according to the equation below:2025-128-2

[0134] (4)

[0135] where <*> and < (*) denote elementwise mean and SD of the input signals, respectively, calculated over the training data.

[0136] DNNClassification80 contained three (3) hidden layers (128, 64, 32 units), each with " ReLU" activations and L2 regularization (a=le-3). All hidden layers were complemented by a batch standardization layer and a 0.5 dropout. DNNClassificationwas compiled using a categorical cross-entropy loss function (ZCCE), a learning rate of le-3, and a batch size (Nb) of 5. LCCE is expressed by:

[0137] LCCE(y,ŷ) = −(1 / N)∑i ∑x yn,x log ŷn,x (5)

[0138] where y7l xrepresent the gold standard binary labels of the C = 3 concentration range classes per sample, i.e., x E |cTnI < 40 pg / mL; cTnI = 40–1000 pg / mL; cTnI >1000 pg / mL], and ynxare the model's predicted probabilities for these classes, calculated using the sigmoid activation function:

[0139] (6)

[0140] where yn xrefer to the model’s inference before the sigmoid function for the C = 3 cTnl concentration range classes. The final predicted cTnl concentration range class per sample was determined as an output node with the highest probability from the sigmoid activation.

[0141] DNNQ<40 and DNNQ40-1000 networks 80 consisted of three (3) hidden layers with 256, 128, and 64 units, while DNNQ>1000network 80 contained three (3) hidden layers with 128, 64, and 32 units. All hidden layers in all three quantification networks had 'ReLU' activations and L2 regularization (a=le-3). Furthermore, all hidden layers in all three models were supplemented by batch standardization and a 0.25 dropout. All quantification models were trained with a mean squared logarithmic error (MSLE) loss function, a learning rate of le-3 and Nb = 5. MSLE loss is expressed by:

[0142] MSLE = (1 / N)∑i (log(yi + 1) − log(y′i + 1))², (7)

[0143] where yzare the gold standard cTnl concentrations and y'tare the model-predicted concentrations. The unit of ground truth concentrations for DNNQ<40 and DNNQ40–1000networks 80 was in pg / mL, while that of the ground truth concentrations for DNNQ>100080 was in ng / mL. Ground truth labels for DNNQ40–100080 were normalized by 40 pg / mL.2025-128-2Configurations of the four models were optimized using a 4-fold cross-validation procedure on the validation sets.

[0144] The DNNClassificationnetwork 80 was trained on 68 CL-VFA samples activated with serum samples from 35 different patients and validated on 62 CL-VFA samples from 32 patients. cTnl concentrations in the training and validation sets for DNNClassification80 varied between <4 pg / mL and -10,000 pg / mL, covering typical clinical ranges of hs-cTnl testing. Optimized DNNClassification80 demonstrated 93.5% accuracy on the validation samples, correctly classifying all samples in <40 pg / mL and >1000 pg / mL ranges (see FIG. 12A). Four out of 15 samples from 40-1000 pg / mL range were misclassified into the <40 pg / mL range. Ground truth cTnl concentrations in these samples were 46 pg / mL (2 samples), 56 pg / mL (1 sample), and 139 pg / mL (1 sample). Blind testing of DNNClassification80 on 66 additional serum samples from 34 patients showed 95.5% accuracy, as reported in the Results section (FIG. 12B).

[0145] The DNNQ 40 network 80 was trained on 47 CL-VFA samples activated with serum samples from 24 different patients and validated on 34 CL-VFA samples activated with serum samples from 17 patients. The training dataset for the DNNQ<40 network 80 had cTnl concentrations between <4 pg / mL and -100 pg / mL. Ground truth concentrations in samples from the <4 pg / mL range were assigned to 4 pg / mL, and MSLE for these samples only increased if the model prediction exceeded 4 pg / mL. The DNNcyio-iooo network 80 was trained on 18 samples from 10 patients and further validated on 15 samples from 8 patients. Training samples for the DNNQ40–1000network 80 had cTnl concentrations from -20 pg / mL to -1,500 pg / mL. Finally, the DNNQ>1000network 80 was trained on 19 samples from 10 patients and then validated on 13 samples from 7 patients. cTnl concentrations for DNNQ>1000varied between -500 pg / mL and -10,000 pg / mL. Validation samples for the three quantification models were selected from 62 validation samples used for the DNNClassificationnetwork 80 according to the target quantification ranges (i.e., <40 pg / mL, 40-1000 pg / mL, or >1000 pg / mL, see FIGS. 5A-5H for more details). All of the samples used to train, validate and blindly test DNNClassification, DNNQ 40, DNNQ40–1000, and DNNQ>1000networks 80 represented real samples obtained by activating CL-VFA cartridges with patient serum samples, and no data augmentation methods were applied to any of these samples.

[0146] cTnl quantification was reported if and only if the concentration prediction by the quantification model 80 aligned with the concentration range predicted by DNNClassificationnetwork or model 80, i.e., the predicted concentration value stayed within the range predicted2025-128-2by DNNClassification80 or fell within 10% from the range borders. Otherwise, the sample was labeled as "indeterminate" and eliminated from quantification. Based on this rule, two validation samples were labeled as "indeterminate". One sample from the validation set, with a ground truth concentration measurement of 139 pg / mL, was classified into the <40 pg / mL range by the DNNClassificationnetwork 80; however, the DNNQ<40 network 80 predicted a concentration value of -60 pg / mL, exceeding the 40 pg / mL border by 50%. Similarly, another serum sample with 1,613 pg / mL ground truth was correctly classified as >1000 pg / mL by the DNNClassificationnetwork 80; however, prediction from the DNNQ > IOOO network 80 was -590 pg / mL, -40% lower than the 1000 pg / mL threshold. These 2 samples were excluded from quantification and labeled as indeterminate due to the inconsistency between the quantification and classification neural networks; the quantification models were validated on 60 samples from 31 patients (see FIG. 12B). None of the samples from the blind testing dataset had contradicting predictions between the classification and quantification models. Therefore, none of the blind testing samples were labeled as "indeterminate."

[0147] Predicted cTnl concentrations for the 60 validation samples demonstrated a good correlation with the gold standard values from an FDA-approved analyzer, with a Pearson’s r of 0.993 and a CV of 8.7% between duplicate testing repeats (see FIG. 12B). Blind testing of DNNQ<40, DNNQ40–1000, and DNNQ>1000 networks 80 on 66 new samples, which were not used in training and optimization, also demonstrated reliable quantitative performance, achieving a Pearson's r of 0.984 (FIG. 4C). Clinical samples from the <4 pg / mL concentration range were excluded from the calculation of Pearson's r and CV for both validation and blind testing sets since they did not have quantitative ground truth concentration measurements. Further details about the quantification models' performance on the blind testing dataset are reported in the Results and Discussion section, under the “Neural network-based cTnl quantification using CL-VFA data” sub-section.

[0148] Training times for DNNClassification, DNNQ<40, DNNQ40–1000, and DNNQ>1000 networks 80 were 6.4, 15.8, 14.6, and 17.0 min respectively. Blind testing times were considerably lower, not exceeding 0.5 s for any of the four models. The neural networks or models 80 were trained and tested on a desktop computer with a GeForce GTX 1080 Ti (NVIDIA). The neural network-based cTnl quantification pipeline was developed in Python using NumPy, TensorFlow, and Keras libraries.

[0149] Statistical analysis: For assay validation and tests involving cTnl-spiked serum samples, data were presented as the mean ± SD, derived from a minimum of three2025-128-2independent measurements. Results for clinical sample tests were reported as the mean of two replicate measurements ± SD. Detailed information on the number of experimental replicates is available in each FIG. caption. The CV was computed as the ratio of the SD to the mean, expressed as a percentage. Group differences were evaluated using an unpaired two-sample t-test in GraphPad Prism, with statistical significance set at P <0.05.

[0150] 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, while first and second cartridges 14A, 16A, 14B, 16B (top and bottom) are described herein, in some embodiments first and second cartridges may be combined to a single cartridge. However, as noted herein, for the CL reaction and imaging phases, the tray 32 containing the sensing membrane is preferably transferred to the second top cartridge 16B and second bottom cartridge 14B which is then read using the reader device 50. In addition, while trained neural networks 80 provided the optimal results in the cTnl quantification pipeline, it should be appreciated that other embodiments may use other machine learning algorithms like random forest and logistic regression as examples. The following publication is incorporated by reference herein including all accompanying Supporting Information: Gyeo-Re Han et al., Deep Learning-Enhanced Chemiluminescence Vertical Flow Assay for High-Sensitivity Cardiac Troponin I Testing, Nano Micro Small, Vol. 21, Issue 11, February 5, 2025. The invention, therefore, should not be limited, except to the following claims, and their equivalents.

Claims

2025-128-2What is claimed is:

1. A method of detecting the presence of and / or quantifying the amount or concentration of one or more analytes or biomarkers in a sample using a vertical flow assay¬ comprising one or more cartridges having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more capture agent(s), the method comprising: loading a mixture of the sample and detection reagents including a detection antibody conjugate comprising a plurality of peroxidase molecules associated with an affinity agent and conjugated to detection antibodies along with a chemiluminescence (CL) reagent solution into the sample inlet;imaging the sensing membrane with a reader device configured to obtain one or more CL images and / or CL signals for the plurality of spots; andprocessing the one or more images and / or CL signals for the plurality of spots with an algorithm comprising machine learning or one or more trained neural networks configured to generate one or more outputs that include a classification of the sample and / or a quantification of the amount or concentration of the one or more analytes or biomarkers in the sample.

2. The method of claim 1, wherein the plurality of peroxidase molecules comprise PolyHRP-streptavidin (PolyHRP) and the affinity agent comprises a particle or bead.

3. The method of claim 2, wherein the particle or bead comprises a nanoparticle.

4. The method of claim 1, wherein the one or more cartridges comprises a first bottom cartridge, a first top cartridge, a second top cartridge, and a second bottom cartridge, wherein the sensing membrane is disposed on a removable tray insertable into the first and second bottom cartridges and wherein the first top cartridge comprises one or more paper layers and a sample inlet, the second top cartridge comprises a reagent inlet and a fluidically coupled reagent chamber when secured to the second bottom cartridge.2025-128-25. The method of claim 4, further comprising:inserting the removable tray with the sensing membrane into the first bottom cartridge and securing the first top cartridge to the first bottom cartridge and loading the sample and the detection reagents into the sample inlet; andremoving the first top cartridge from the first bottom cartridge and placing the removable tray with the sensing membrane into the second bottom cartridge and securing the second top cartridge to the same and loading the CL reagent solution into the reagent inlet of the second top cartridge so as to fill the reagent chamber.

6. The method of claim 1, wherein the capture agent(s) comprise one or more antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, or peptoids,7. The method of claim 1, wherein the algorithm comprises a first neural network that classifies the concentration(s) of the one or more analytes or biomarkers in the sample as within a range of one or more threshold value(s) and a plurality of additional algorithms involving one or more neural networks to output the concentration(s) of the one or more analytes or biomarkers in the sample.

8. The method of claim 1, wherein the one or more analytes or biomarkers comprise proteins, cardiac troponin 1 (cTnl), cardiac troponin T, myoglobin, creatine kinase-MB, B-type natriuretic peptide (BNP), N-terminal proBNP, C-reactive protein, interleukin-3, interleukin-8, albumin, or glycated albumin, virus biomarkers including Influenza A / B, HIV, Coronavirus, Cytomegalovirus, Hepatitis, and bacterial biomarkers including bacterial lipopolysaccharide, bacterial antigens, bacterial enzymes, or toxins.

9. The method of claim 1, wherein the sensing membrane is populated with one or more of negative control spots containing no capture agent(s) and positive control spots containing a detection antibody.

10. The method of claim 1, wherein the one or more CL images comprise time- lapsed images.2025-128-211. The method of claim 1, wherein the sample comprises a body fluid, serum, plasma, whole blood, saliva, urine, cerebrospinal fluid, sweat, amniotic fluid, or interstitial fluid.

12. The method of claim 1, further comprising displaying the amount or concentration of one or more analytes on a display associated with the reader device.

13. The method of claim 4, wherein the second bottom cartridge is inserted into a cartridge tray in the reader device for imaging the sensing membrane.

14. A system for detecting the presence of and / or quantifying the amount or concentration of one or more analytes or a biomarker in a sample comprising:a vertical flow assay comprising one or more cartridges having a sample inlet and a removable sensing membrane populated with a plurality of spots containing one or more capture agent(s);a reader device comprising a housing that includes a camera and a cartridge tray that receives the one or more cartridges and places the removable sensing membrane along an optical path of the camera, the camera configured to obtain chemiluminescence (CL) images and / or CL signals of the plurality of spots; anda computing device configured to obtaining the CL images and / or CL signals of the removable sensing membrane after exposure to a CL reagent solution, the computing device further comprising software or instructions that execute an algorithm comprising machine learning or one or more trained neural networks configured to generate one or more outputs that include classification of the sample and / or a quantification of the amount or concentration of the one or more analytes or biomarkers in the sample based on the CL images and / or CL signals for the plurality of spots.

15. The system of claim 14, wherein the one or more cartridges comprises a first bottom cartridge, a first top cartri dge, a second top cartridge, and a second bottom cartridge, wherein the removable sensing membrane is disposed on a removable tray insertable into the first and second bottom cartridges and wherein the first top cartridge comprises one or more paper layers and a sample inlet, the second top cartridge comprises a reagent inlet and a fluidically coupled reagent chamber when secured to the second bottom cartridge.2025-128-216. The system of claim 14, further comprising a lens or lens set disposed in the housing along the optical path.

17. The system of claim 15, wherein the reagent chamber is optically transparent and wherein the removable tray containing the removable sensing membrane is disposed in the second bottom cartridge that is secured to the second top cartridge.

18. The system of claim 14, wherein the reader device further comprising a display.

19. The system of claim 14, wherein the one or more capture agent(s) comprises one or more antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, or peptoids,20. The system of claim 14, wherein the algorithm comprises a first neural network that classifies the concentration(s) of the one or more analytes or biomarkers in the sample as within a range of one or more threshold value(s) and a plurality of additional algorithms involving one or more neural networks to output the concentration(s) of the one or more analytes or biomarkers in the sample.

21. The system of claim 14, wherein the one or more analytes or biomarkers comprise proteins, cardiac troponin I (cTnl), cardiac troponin T, myoglobin, creatine kinase-MB, B-type natriuretic peptide (BNP), N-terminal proBNP, C-reactive protein, interleukin-3, interleukin-8, albumin, or glycated albumin, virus biomarkers including Influenza A / B, HIV, Coronavirus, Cytomegalovirus, Hepatitis, and bacterial biomarkers including bacterial lipopolysaccharide, bacterial antigens, bacterial enzymes, or toxins.

Images