Deep learning-enhanced paper-based vertical flow assay for high-sensitivity troponin detection

Abstract

A method of detecting and / or quantifying the concentration of one or more analytes in a sample uses a vertical flow assay cartridge that includes one or more cartridges having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more types of capture agent(s). A mixture of the sample and detection reagents is loaded into the sample inlet along with signal amplification reagents. The sensing membrane is imaged with a reader device to illuminate and obtain a plurality of time-lapsed images of the sensing membrane containing time-lapsed intensity signals for the spots. The time-lapsed intensity images and / or signals for the spots are processed with 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 the amount or concentration of the one or more analytes in the sample.

Description

2024-179-2 DEEP LEARNING-ENHANCED PAPER-BASED VERTICAL FLOW ASSAY FOR HIGH-SENSITIVITY TROPONIN DETECTION Related Application

[0001] This Application claims priority to U.S. Provisional Patent Application No. 63 / 554,854 filed on February 16, 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 vertical flow assay (VFA) and determine the concentration of a target analyte. 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 (cTnI). Statement Regarding Federally Sponsored Research and Development

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

[0004] Cardiovascular diseases (CVDs) present a major public health challenge worldwide, representing a leading cause of mortality, with approximately 19.1 million deaths in 2020, accounting for roughly 32% of global deaths, and imposing a significant financial burden on patients and healthcare systems. From a socioeconomic standpoint, the incidence of CVDs shows an inverse correlation with income, with more than 80% of CVD-related deaths globally reported in low- and middle-income countries, exacerbating the global burden of CVDs shifting to these regions. Similarly, within the U.S., the burden of CVDs is increased in lower-income communities. Amid the spectrum of CVDs, acute myocardial infarction (AMI), commonly referred to as a heart attack, is a particularly critical event characterized by a sudden and severe impairment of blood flow due to a blockage of the coronary artery. Chest pain that may indicate AMI is a primary reason for emergency department visits, with suspected AMI patients accounting for ~10% of all emergency2024-179-2 department visits. Furthermore, AMI stands as a paramount contributor to sudden death, with its swift onset and severe consequences, making it a primary concern in cardiovascular emergencies.

[0005] To date, numerous clinical care advancements related to AMI have benefitted from laboratory-based high-sensitivity (hs) cardiac troponin I (cTnI) assays. cTnI is a protein specific to myocardial cells and is released into the bloodstream in the case of myocardial damage, making it the gold standard biomarker for the diagnosis of AMI. The latest clinical guidelines, based on the fourth universal definition of myocardial infarction, prioritized the identification of myocardial injury by small increases in cTnI levels as an essential measure in diagnostic algorithms along with other evidence (e.g., electrocardiography or imaging), thereby reinforcing the role of cTnI measurement in the diagnosis of AMI. Currently available, clinically deployed benchtop hs-cTnI assays can quantify cTnI at low concentrations (typically a few pg / mL) at or below the 99thpercentile upper reference limit (URL) or in more than 50% of the healthy population with a high precision (coefficient of variation [CV] ≤ 10%), providing opportunities for early diagnosis, rapid rule-in / rule-out, risk categorization, and prognostic applications. However, laboratory-based hs-cTnI assays and readout instruments are relatively expensive and bulky, and require specialized / trained operators. As a result, they are mainly available in large hospitals with well-developed medical infrastructure, making these assays inaccessible to remote and low-income regions. Underserved populations in these regions have a higher prevalence of AMI, and the need for rapid and accurate diagnosis is critical. Even in cases where a patient suspected of AMI reaches the hospital within the golden hour (the first hour after the onset of symptoms), the medical protocol conventionally followed is complicated and multi-step (i.e., ordering a test from a healthcare provider, collecting and transporting a specimen to a central facility, processing the specimen, and reporting the results as seen in FIG. 1A). Turnaround times for this procedure can take over several hours, further compounded by a 2–3 h waiting period that patients with intermediate cTnI levels (between normal and abnormal ranges) must undergo for follow-up testing. This follow-up testing is necessary to monitor the potential elevation of cTnI levels and make a conclusive determination of safe rule-in or rule-out. Moreover, high-volume sample processing and batch testing done at clinical laboratories typically cause additional delays in reporting results and performing critical medical interventions, making this framework not ideal for addressing the time-sensitive nature of AMI, where rapid diagnosis and response are essential.2024-179-2

[0006] Coupled with the continuous advances in sensing, information and communication technologies, point-of-care testing (POCT) is emerging as a viable alternative approach in the diagnostic landscape of AMI, potentially transforming diagnostic pipelines by delivering rapid, affordable, and on-site diagnostic measures (FIG. 1B). POCT technologies for AMI can provide the advantage of decentralized testing, which expands testing capabilities in various healthcare settings (e.g., hospitals, clinics, nursing homes, pharmacies) and accessibility even in resource-limited areas. By circumventing the complexities and delays associated with the traditional workflow, the POCT of cTnI could expedite clinical actions to be performed within 20 min and reduce the time and cost linked to the waiting period before the follow-up testing. To address this important need, several cTnI-POCT products are currently available, including those based on microfluidics, lateral flow assays, and automated enzyme-linked immunosorbent assays. Some of the additional recent advances also involve microarray technologies, paper-based sensors, electrode-based sensors, and surface-enhanced Raman scattering coupled with luminescent, electrochemical, and colorimetric sensing modalities. However, it is important to note that because of the extremely low clinical cut-off level of cTnI compared to other biomarkers, most of these POCT assays currently have limitations, particularly in terms of their sensitivity and precision (see FIG. 1C). Additionally, some of these assays require expensive benchtop hardware and rely on AC power, which limits their portable use. Even though assay platforms can be miniaturized with hand-held analyzers, the high hardware costs exceed $15K and create another barrier to widespread implementation in clinical settings, particularly in low- and middle-income countries. These necessitate continued research and development to create high-performance cTnI-POCT solutions that meet clinical requirements for high sensitivity and precision while ensuring cost-effectiveness. Summary

[0007] A paper-based high-sensitivity vertical flow assay (hs-VFA) is disclosed for rapid and cost-effective POCT of cTnI, providing enhanced sensitivity and high precision, and meeting the clinical standards for hs-cTnI testing. This hs-VFA system consists of a paper- based sensor within a cartridge coupled with rapid nanoparticle amplification chemistry, a portable reader device with time-lapse imaging capability, and computational analysis algorithms executed on a computing device empowered by deep learning (FIG. 1D). This system enabled highly sensitive and accurate quantification of cTnI from 50 μL of serum2024-179-2 within 15 min per test. The paper-based sensor and assay cartridge was used to perform the immunoassay and signal amplification with simple operational steps, comparable to at-home rapid diagnostic tests. The performance of the hs-VFA was validated by testing it using cTnI spiked serum samples as well as clinical patient serum samples collected from healthy and diseased donors. Signals from the activated immunoassays were captured by a custom- designed, cost-effective and hand-held reader device in a time-lapse manner. The acquired data were processed by computational assay validation and outlier analysis algorithms to improve assay performance, resulting in a limit of detection (LoD) of 0.2 pg / mL with high precision (average CV < 7%). Furthermore, hs-VFA signals were used to train neural network-based models that accurately quantified cTnI levels in clinical serum samples obtained from patients. When blindly tested on a set of 46 serum samples from 23 patients, the predictions of the computational hs-VFA showed a high correlation (Pearson's r = 0.965) with the ground truth measurements obtained by an FDA-cleared benchtop analyzer, while achieving ~93.5% accuracy for classifying the concentration of the serum samples to be below / above the cut-off of the benchtop clinical analyzer (< 40 pg / mL), which increased to 100% accuracy using duplicate testing per patient. The platform is a significant leap forward in hs-VFA technology for POCT and offers an affordable, robust, and high-performance sensing platform for cTnI and potentially other critical low-abundance biomarkers used for rapid diagnoses and medical interventions.

[0008] In one embodiment, a method of detecting the presence of and / or quantifying the amount or concentration of one or more analytes in a sample using a vertical flow assay cartridge includes one or more cartridges having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more types of capture agent(s). The method includes loading a mixture of the sample and detection reagents into the sample inlet along with signal amplification reagents. The sensing membrane is imaged with a reader device configured to illuminate and obtain a plurality of time-lapsed images of the sensing membrane containing time-lapsed intensity signals for the plurality of spots. The time-lapsed intensity images and / or 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 the amount or concentration of the one or more analytes in the sample.

[0009] In another embodiment, a system for detecting the presence of and / or quantifying the amount or concentration of one or more analytes in a sample includes a vertical flow2024-179-2 assay cartridge including one or more cartridge subunits having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more types of capture agent(s). The system includes a reader device that includes a housing having one or more light sources, a camera, and a cartridge receiver that receives the bottom cartridge and places the sensing membrane along an optical path between the one or more light sources and the camera. In one embodiment, a computing device is configured to control the one or more light sources and the camera for obtaining a plurality of time-lapsed images of the sensing membrane. The computing device further includes 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 a classification of the sample and / or a quantification of the amount or concentration of the one or more analytes in the sample based on the time-lapsed intensity images for the plurality of spots.

[0010] In one embodiment, the entire vertical flow assay cartridge that includes multiple cartridge subunits is secured to reader device for the reading operation. In another embodiment, only one of the cartridge subunits (e.g., the bottom cartridge) is secured to the reader device for the reading operation. For example, the bottom cartridge can be loaded into a cartridge tray that is then inserted into the reader device in a cartridge receiver. The bottom cartridge may also be rotatably secured to the reader device using, for example, posts, detents, or boss(es) that interface with slots / recesses located in the reader device. Brief Description of the Drawings

[0011] FIGS. 1A-1D provide an overview of paper-based hs-VFA system and its application in hs-POCT of cTnI. FIG. 1A illustrates a traditional workflow (prior art) of central laboratory tests for diagnosing patients with AMI. FIG. 1B illustrates POCT workflow using hs-VFA and its mobile reader system can enhance the efficiency and quality of patient care. FIG.1C illustrates the current performance landscape of commercial cTnI assays (hs- assays and POCT assays) and the requirements for a successful hs-POCT cTnI assay. The bars show the ranges of sensitivity and precision specifications, indicated by dots and circles, respectively. FIG. 1D is a schematic summary of current approach for hs-POCT cTnI assay using the hs-VFA system, a portable reader, and computational analysis.

[0012] FIGS. 2A-2I illustrate aspects of the assay, reader, and imaging strategy of the hs- VFA system. FIG. 2A illustrates structure of the hs-VFA cartridge, including two top cartridges and a bottom cartridge with a sensing membrane disposed on the bottom cartridge.2024-179-2 FIG. 2B illustrates the portable reader device that uses a Raspberry Pi as the computing device. The inset displays the internal cartridge tray during imaging under LED illumination. FIG. 2C illustrates the assay process and timeline. FIG. 2D illustrates the principle of immunoassay and signal amplification based on the reduction of Au3+onto the surface of AuNP conjugates. FIG. 2E shows sensing membrane images at each assay step and corresponding SEM images of the test spots. Scale bar: 200 nm. FIG. 2F shows evaluation of spot uniformity in the amplification reaction. Three individual sensors treated with 1 OD AuNP-antibody conjugate in each spot were used for comparison. FIG. 2G illustrates the effect of the signal amplification on analytical sensitivity. Conjugates were serially diluted and pre-immobilized on the sensing membrane. Two top-cartridge operations were conducted. FIG. 2H illustrates a comparison of imaging methods: time-lapse vs. end-point analyses. FIG. 2I shows the impact of time-lapse imaging on assay sensitivity. Data points in FIG. 2G and FIG. 2I represent the mean of triplicates ± SD.

[0013] FIGS. 3A-3E illustrate computational data processing and the impact of assay quality check algorithms for cTnI testing using hs-VFA. FIG. 3A illustrates the data- processing pipeline; (i) Dataset and criteria to create a VCI. (ii) An example illustrating data refinement through Outlier Analysis (OA). FIG. 3B shows the effect of OA; (i) The number of remaining test spots. (ii) Exclusion rate for each test spot. (iii) Test spot uniformity within a single cassette (intra-assay). (iv) Assay repeatability (inter-assay). Data in (i)–(iii) are based on N=330 (55 spiked and 55 clinical samples with three time points). Data in (iv) are based on N=21 (7 cTnI spiked serum samples, triplicates). C stands for control, before OA. OA refers to a statistics-based method and OA-D refers to a differential-based method (see the Methods section). FIG. 3C shows test results for cTnI spiked serum sample and calibration plot comparison with and without OA (N=14, triplicates). The top line represents results with OA (R2= 0.998, ), and the bottom line represents results without OA (R2= 0.986, ). FIG. 3D is a comparison of test spot intensities for the spiked samples within the 99thpercentile range of cTnI concentration. FIG. 3E shows cTnI clinical sample test results and the impact of data-processing algorithms to exclude invalid results and improve repeatability; (i) before and (ii) after the assay validation and data refinement (Ninitial = 62, duplicated). Ground truth values of the samples with cTnI levels < 40 pg / mL are assigned to 39 pg / mL due to the clinical analyzer's cut-off level (40 pg / mL).

[0014] FIGS. 4A-4C illustrate how neural network-based analysis of cTnI concentrations in clinical serum samples is performed. FIG. 4A illustrates the neural network-based signal2024-179-2 processing pipeline. The input to the neural networks includes a time-lapse signal calculated from three time-lapse images. The neural network-based processing pipeline consists of one classification (DNNClassification) and two quantification (DNNQuantificationand DNNLow) neural networks that collaborate with each other. If a sample is blindly classified by the DNNClassification model as < 40 pg / mL (below the cut-off concentration), that is the final decision, and the quantification neural network models are not used in that case. Samples that are blindly classified as ≥ 40 pg / mL by DNNClassificationare then processed by DNNQuantificationfor quantitative analysis. Blind inference results of DNNQuantification are used as the final concentration measurements if and only if the inference result is ≥ 40 pg / mL, in agreement with the DNNClassificationblind inference. The samples predicted as < 40 pg / mL by DNNQuantification are further processed by DNNLow since this indicates a disagreement between the blind inference results of DNNClassification and DNNQuantification. If the cTnI concentration predicted by DNNLowis still < 40 pg / mL the sample is labeled as “undetermined” as the disagreement with DNNClassification is further confirmed; otherwise, the prediction from DNNLow is used as the final concentration measurement. FIG. 4B shows the classification results from the optimized classification model (DNNClassification) for forty-six (46) serum samples from twenty-three (23) patients used in the blind testing set. The 40 pg / mL threshold concentration is due to the cut-off level of the benchtop clinical instrument used to obtain ground truth concentration measurements. FIG. 4C shows the cTnI quantification results from the optimized quantification models (DNNQuantificationand DNNLow) for thirty-five (35) serum samples from eighteen (18) patients classified into ≥ 40 pg / mL range by DNNClassificationblind inference.

[0015] FIG. 5 illustrates an exemplary graphical user interface (GUI) of the Raspberry Pi- based reader device.

[0016] FIGS. 6A-6B illustrate the reagent flow mechanism during the amplification reaction. FIG. 6A illustrates reagent transport from the reagent chamber of the second top cartridge to absorption pads. FIG. 6B illustrates the reversed flow of reagent from the absorption pads to the sensing membrane after decapping the second top case.

[0017] FIGS. 7A-7B illustrates solution-based characterization of Au-ion reduction chemistry. FIG. 7A shows an analysis of the absorption peak shift. FIG. 7B is a comparison of color changes in the solutions.

[0018] FIGS. 8A-8C illustrate optimization of the assay. FIG. 8A illustrates an evaluation of the effect of cassette compressibility. FIG. 8B shows the impact of the size of AuNP on2024-179-2 assay performance. FIG. 8C illustrates selection of the running buffer combination. Buffer A contains 1% v / v Triton X-100 and 1% v / v BSA in PBS. Buffer B contains 3% v / v Tween 20, 1% v / v albumin, 0.5% w / w protein saver, and 1% w / w trehalose in PBS. Each data point represents the mean of triplicates ± SD.

[0019] FIGS. 9A-9B illustrate optimization of the reagent concentrations for Au-ion reduction-based signal amplification reaction. FIG. 9A shows signal intensity analysis. FIG. 9B shows a reaction uniformity test. A 1 ng / mL cTnI was assayed triplicate. The intensity in FIG. 9A represents the mean of the three measurements. CV(%) in FIG. 9B represents the mean of inter-assay CV(%) from the triplicates. The numbers in the bracket indicate the rank.

[0020] FIG. 10 illustrates representative images of sensing membranes after conducting assays with varying concentrations of cTnI spiked in human serum. Images were captured at t = 60s within the time-lapse sequence.

[0021] FIG. 11 illustrates the architecture of the optimal classification neural network (DNNClassification).

[0022] FIGS. 12A-12B illustrate the architectures of the optimal quantification neural networks. FIG. 12A illustrates the DNNQuantificationneural network and FIG. 12B illustrates the DNNLowneural network.

[0023] FIGS. 13A-13B illustrate the performance of the optimized neural network models on the validation sets. FIG. 13A illustrates classification predictions on the validation set. FIG. 13B shows quantification predictions on the validation set.

[0024] FIGS. 14A-14C illustrates predictions of cTnI concentrations in clinical samples from non-optimized quantification models. FIG. 14A is a power-fitting model with the time- lapse input data processed by OA. FIG. 14B shows the optimized neural network models (i.e., and ) with the time-lapse input data not processed by OA. FIG. 14C shows optimized neural network models with the end-point input data processed by OA. Dashed lines on x and y axis indicate 40 pg / mL.

[0025] FIG. 15 illustrates the evaluation of the imaging-based signal quantification capacity of the Raspberry Pi-based portable reader device. Each data point represents the mean of triplicates ± SD.

[0026] FIGS. 16A-16D illustrate the evaluation of the impact of the number of testing spots on the assay performance. FIG. 16A illustrates a spotting map. FIG. 16B illustrates sensitivity and FIG. 16C illustrates intra-assay uniformity, and FIG. 16D shows inter-assay2024-179-2 repeatability. Each data point represents the mean of triplicates ± SD without outlier exclusion. Detailed Description of Illustrated Embodiments

[0027] With reference to FIGS. 1D, 2a-2E, a platform or system 2 for detecting the presence of and / or quantifying the amount or concentration of one or more analytes in a sample is disclosed. The analyte may include cardiac troponin I (cTnI) although other analytes can be used with the platform 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 blood or serum but may also include an environmental sample. The system 2 includes a vertical flow assay cartridge 10 that, in one embodiment, uses several cartridge subunits that are combined together or separated from one another in different parts of the assay procedure although other embodiments envision a single cartridge 10. In one embodiment of a cartridge 10 that includes multiple subunits, a bottom cartridge 12 is provided that includes a plurality of absorption layers 14 (e.g., engineered paper or other material or pads) and includes a sensing membrane 16 populated with a plurality of spots 18 containing one or more types of capture agent(s) that act as reaction areas. The capture agent(s) may include antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, peptoids, etc. The capture agent in one specific embodiment includes capture antibodies are specific to cTnI as described herein but other capture agents may be used for other analytes. For example, other analytes include 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 16 may be a nitrocellulose membrane that is spotted at different locations to form the spots 18 with capture antibodies. As explained herein a wax printer may be used to compartmentalize discrete hydrophilic locations on the sensing membrane 16 that contain the capture antibodies in spots 18 that are surrounded by hydrophobic wax layer(s) 20. This focuses or concentrates fluid flow onto the spots 18 or reaction areas.

[0028] With reference to FIG. 2A, the vertical flow assay cartridge 10 includes a first top cartridge 22 having one or more engineered paper layers 24 therein and a sample inlet 26. The engineered layers 24 (e.g., made from paper or cellulose-based material) may include2024-179-2 sequentially stacked engineered paper layers. This may include an absorption layer, flow diffuser, 1stspreading layer, interpad, 2ndspreading layer, and support layer for holding the layers. For example, the first top cartridge 22 may include two asymmetric membranes. These asymmetric membranes are asymmetric in that the pore size changes in the direction of the thickness of the membrane. The top such membrane is 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). Some engineered layers 24 such as vertical flow diffuser layers promote vertical (e.g., top-to-bottom) movement of fluid through the particular engineered layer 24 and inhibit lateral flow. Still other engineered layers 24 act as supporting structures (e.g., support layer) or support lateral flow (i.e., asymmetric membranes). The various layers may be held together and in place with an external or peripheral support which in the experiments conducted herein was foam tape, although it should be understood that other materials and structures may be used.

[0029] The second top cartridge 28 includes an inlet 30 used for reagents and an outlet 32 and a fluidically coupled reagent chamber 34. Fluid entering the reagent inlet 30 leaves the outlet 32 that fills the reagent chamber 34. In some embodiments, the regent chamber 34 includes a cover that is optically transparent. In this regard, the camera 46 of the reader device 40 (discussed below) is able to image the sensing membrane 16 during the assay with the second top cartridge 28 secured to the bottom cartridge 12. Or, alternatively, the user may monitor the coverage of the reagent fluid over the sensing membrane 16 via the optically transparent cover. The fluid in the reagent chamber 34 covers the sensing membrane 16 when the second top cartridge 28 is secured to the bottom cartridge 12 and the reagent fluid is loaded therein. The first and second top cartridges 22, 28 may be detachably coupled with / from the bottom cartridge 12. For example, the first top cartridge 22 and the second top cartridge 28 may include may include one or more posts, detents, or bosses 35 that interface with a slot or recess 36 contained in the bottom cartridge 12. In this way, the first and second top cartridges 22, 28 are detachably connected to the bottom cartridge 12 by twisting the first / second top cartridges 22, 28 onto the bottom cartridge 12 (FIG. 1D).

[0030] The system 2 includes a reader device 40 (FIG. 1B) that is used to take time-lapse images of the sensing membrane 16. The reader device 40 includes a housing 42 that includes one or more light sources 44, a camera 46, and a cartridge receiver 48 that receives the vertical flow assay cartridge 10 or just the bottom cartridge 12 so as to place the sensing2024-179-2 membrane 16 within the field of view of the camera 46. In one embodiment, a cartridge tray 50 receives the bottom cartridge 12, the cartridge tray 50 being insertable into the cartridge receiver 48 which may include slots or the like in the housing 42 to place the sensing membrane 16 within the field of view of the camera 46 when fully inserted into the cartridge receiver 48. Insertion of the cartridge tray 50 with the bottom cartridge 12 into the reader device 40 also blocks ambient light from entering the interior of the reader device 40 which would interfere with the images of the sensing membrane 16. In other embodiments, the bottom cartridge 12 could be secured to the reader device 40 using, for example, one or more posts, detents, or bosses 35 that interface with a slot or recess 36 contained in the bottom cartridge 12 in a similar manner to how the first top cartridge 22 and second top cartridge 28 are secured to the bottom cartridge 12. In this embodiment, the bottom cartridge 12 may be twisted onto the reader device 40 for imaging of the sensing membrane 16. As best seen in FIG. 2B, the reader device 40 may include an LED module that acts as the light source(s) 44. An optional lens 52 or set of lenses 52 may be interposed between the light sources(s) 44 and the camera 46.

[0031] The reader device 40 further includes a computing device 54 that has software or instructions thereon that execute (or are executable) an algorithm including one or more trained neural networks 80, 82, 84 as illustrated in FIGS. 11, 12A, 12B that are configured to generate one or more outputs that classify the sample and / or quantify the amount or concentration of the one or more analytes in the sample. This output may be used to, for example, diagnose a condition of a subject that provided the sample. One trained neural network 80 may be used to classify the measured analyte above or below a target (e.g., 40 pg / mL) such as that illustrated in FIG. 11. Another trained neural network 82 (FIG. 12A) may be used to quantify the amount or concentration of the analyte if high enough or flag the sample as containing a low amount or quantification that needs further confirmation. In this regard, another trained neural network 84 (FIG. 12B) is used to classify the sample as “undetermined” if the inferred amount or concentration is below the threshold (e.g., < 40 pg / mL) or output an inferred amount or concentration if at or above the threshold (≥ 40 pg / mL).

[0032] The software may be used to acquire and store the time-lapsed images of the sensing membrane 16 including image analysis like segmentation and quantifying the intensity of the various spots 18 on the sensing membrane 16. The software can also perform the virtual control indicator (VCI) as described herein which removes invalid assay results.2024-179-2 The software may also perform the outlier analysis (OA) that excludes outliers of test spots 18. Finally, the software may execute the one or more trained neural networks 80, 82, 84 including, for example, a first neural network (DNNClassification) 80 that classifies a concentration of the one or more analytes in the sample either above or below a threshold value; a second neural network (DNNQuantification) 82 that outputs a quantification of the concentration of the one or more analytes in the sample that are above the threshold value; and a third neural network (DNNLow) 84 that receives an input from the second neural network (DNNQuantification) a classification of the concentration below the threshold and outputs one of: (1) a quantification of the concentration of the one or more analytes in the sample or (2) undetermined. In one embodiment, a Raspberry Pi device is used as the computing device 54 although the invention is not limited. The invention may use another microcontroller / computer as the computing device 54 or even a mobile device like a Smartphone is used as the computing device 54 that can be secured to the reader device 40 (and provide imaging functionality). The computing device 54 may even include a remote computing device 54 like a server or the like.

[0033] The reader device 40 further includes a display 56 for displaying time-lapse images of the sensing membrane 16 or real-time observations of the sensing membrane 16. A graphical user interface (GUI) 100 like that illustrated in FIG. 5 may be provided that allows the user to adjust imaging parameters (i.e., exposure time, total number of images, and time interval) for optimal time-lapse imaging. The display 56 may be a touch screen display such as those found on mobile phones or Smartphones. In some embodiments where the computing device 54 is a Smartphone, the display 56 can include the display of the Smartphone. The display 56 can also display quantitative assay results to the user or other qualitative information (e.g., sample is undetermined).

[0034] To use the system 2 for detecting the presence of and / or quantifying the amount or concentration of one or more analytes in a sample, according to one embodiment, the first top cartridge 22 is secured to the bottom cartridge 12 and a sample and reagent mixture are loaded into the sample inlet 26. This may include serum sample and gold nanoparticle (AuNP)-detection antibody conjugates as described herein. While AuNP detection antibody conjugates were used it should be appreciated that other metallic nanoparticles may be used and these may be conjugated to different capture agents. These may include antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, peptoids, etc. As this mixture flows through the sensing membrane 16, a secondary antigen recognition event occurs upon2024-179-2 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 pads 14, accelerating the washing process in the stacked papers. After this initial step, the first top cartridge 22 is removed from the bottom cartridge 12 and the second top cartridge 28 is secured to the bottom cartridge 12 and signal amplification reagents are loaded into the inlet 30 of the second top cartridge 28 so as to fill the reagent chamber 34. This may include amplification solution (500 μL) containing gold ions (Au3+) and a reducing agent (H3NO) is introduced to initiate the signal amplification. In other embodiments, the signal amplification reagents may include metal ions other than gold (e.g., silver ions) and one or more reducing agents such as hydroxylamine, ascorbic acid, hydroquinone, hydrogen peroxide, sodium citrate, tris(hydroxylmethyl) aminomethane), or enzymatic substrates.

[0035] Next, the vertical flow assay cartridge 10 or a portion thereof (e.g., the bottom cartridge 12) is then imaged by the reader device 40. This may be done by removing the second top cartridge 28 and securing the bottom cartridge 12 to the cartridge tray 50 and inserting the two into the cartridge receiver 48 of the reader device 40. The bottom cartridge 12 may also be directly secured to the reader device 40 as explained herein using interfacing post(s), detent(s), or boss(es) 35 that allow the bottom cartridge 12 to be twisted on the reader device 40. In yet another alternative, the second top cartridge 28 and the bottom cartridge 12 may be secured together to the reader device 40. The sensing membrane 16 is imaged through an optically transparent reagent chamber 34. Regardless of how the vertical flow assay cartridge 10 or subunits (e.g., cartridges 12, 28) are attached to the reader device 40, the sending membrane 16 is then illuminated with the one or more light sources 44 to obtain a plurality of time-lapsed images of the sensing membrane 16. The time-lapsed images are used to obtain obtaining time-lapsed intensity signals for the plurality of spots 18 in the plurality of time-lapsed images. Invalid assay spot results may be omitted (e.g., VCI) and OA may perform statistics-based refinement of test spots 18. These time-lapsed intensity signals for the plurality of spots 18 are then input to one or more trained neural networks 80, 82, 84 configured to generate one or more outputs that quantify the amount or concentration of the one or more analytes in the sample. The outputs can be displayed to the user on the display 56 of the reader device 40. For the plurality of time-lapsed images, a waiting period of typically a number of seconds between successive images and may, in one embodiment, be between about 30 to about 60 seconds between successive images is used although this may2024-179-2 vary beyond this range. In some embodiments, at least three time periods are used for the time-elapsed images.

[0036] Experimental

[0037] Design and optimization of the hs-VFA system / platform

[0038] The hs-VFA system 2 was engineered to achieve highly sensitive biomarker detection while maintaining the essential features of conventional VFAs, such as cost- effectiveness, simplicity, user-friendliness, and rapid assay time. The hs-VFA cartridges 10 includes a bottom cartridge 12 with a capture antibody-spotted sensing membrane 16, along with two top cartridges 22, 28, each serving distinct roles of the immunoassay (first top cartridge 22) and the signal amplification chemistry (second top cartridge 28) (FIG. 2A). Wax printing created multiple hydrophilic compartments on the sensing membrane 16, surrounded by hydrophobic wax layers 20, thereby concentrating fluid flow into the spots 18 or reaction areas. Among the seventeen (17) reaction regions of the hs-VFA cartridge 10, ten (10) were assigned as testing spots 18 (treated with cTnI capture antibodies), two (2) functioned as positive control spots 18 (treated with secondary antibodies that bind the detection antibodies), and the remaining five (5) were designated as negative control spots 18 (blanks). This arrangement allowed for ten (10) repeated tests within a single assay, which was important to mitigate test-to-test variations and flow non-uniformity-related imperfections usually observed in inexpensive POCT sensors.

[0039] The hs-VFA cartridge 10 utilized passive capillary flow for vertical sample transport. The first top cartridge 22 facilitated uniform three-dimensional fluid flow, ensuring an even distribution of small-volume samples (≤ 50 μL) and reagents across open compartments on the sensing membrane 16 (12 × 12 mm2). The second top cartridge 28 was fabricated by assembling a 3D-printed cartridge with a transparent acrylic cover and a gasket. An amplification reagent solution, introduced through the inlet 30, flowed directly onto the sensing membrane 16 via the enclosed reaction chamber 34, which served as a reservoir. The reagent solution evenly spread across the entire membrane 16 in < 1 s, enabling uniform reaction across all the test spots 18.

[0040] A custom-designed portable hs-VFA reader device 40 was assembled using a Raspberry Pi computing device 54, a touch display 56, and cost-effective optical components (i.e., camera module for camera 46, macro lens, and light-emitting diodes [LEDs] as the light source 44, see the “Portable reader and image analysis” subsection in the Methods). This configuration enabled high-resolution imaging of the sensing membrane 16 within a compact,2024-179-2 portable design ideal for POCT applications (FIG. 2B). The cartridge tray 50 blocked external ambient light and connected to the main housing 42 through a slide as the cartridge receiver 48. Time-lapse images of the sensing membrane 16 (for each test) were taken under 532 nm LED illumination from the light source 44. The graphical user interface 100 enabled real-time observations of the image of the sensing membrane 16 and provided adjustable imaging parameters (i.e., exposure time, total number of images, and time interval) for optimal time-lapse imaging (FIG. 5).

[0041] The entire assay process runs within 15 min per test, encompassing two successive phases of the assay (immunoassay and signal amplification) followed by the signal readout (FIGS. 2C–2D). For the immunoassay using the first top cartridge 22 secured to the bottom cartridge 12, the first top cartridge 22 is initially filled with a running buffer to activate the immobilized capture antibodies on the sensing membrane 16. Then, a mixture of the serum sample (to be tested) and 15 nm gold nanoparticle (AuNP)-detection antibody conjugates is introduced, allowing the recognition of antigens. As this mixture flows through the sensing membrane 16, 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 pads, accelerating the washing process in the stacked papers 24.

[0042] After the 10-min immunoassay step, the first top cartridge 22 is replaced with the second top cartridge 28 on the bottom cartridge 12, and the amplification solution (500 μL) containing gold ions (Au3+) and a reducing agent (H3NO) is introduced via the inlet 30 to initiate the signal amplification. This process amplifies the colorimetric response of each spot 18 or reaction area where AuNPs were previously bound. Notably, the AuNPs catalyze the reduction of the Au3+on their surfaces in the presence of H3NO, leading to gold particle growth on the paper (FIG. 2D). This results in a more intense optical signal per test spot 18, primarily due to increased absorption, further enhancing the assay’s sensitivity to <1 pg / mL. FIG. 2E shows brightfield images of the sensing membrane 16 alongside the corresponding scanning electron microscope (SEM) images of the testing spots 18 after the immunoassay and signal amplification stages. These images demonstrate a substantial increase in the original AuNP diameter (from 15 nm to 100–200 nm) and the formation of particle clusters following the signal amplification, providing direct evidence of the physical changes occurring during the Au-ion reduction strategy for assay signal amplification. The flow2024-179-2 mechanism of the amplification reagent solution (FIGS. 6A-6B) and spectroscopic analysis of the amplification reaction (FIGS. 7A-7B) are also provided in Supplementary Note 1.

[0043] The uniformity of this signal amplification step across all the reaction spots 18 was assessed, revealing a decrease in CV, from 7.9% to 4.1%, when excluding the edge spots 18 (FIG. 2F). Regarding the analytical sensitivity of the VFA system 2 (FIG. 2G), this was experimentally verified a 212.3-fold improvement in sensitivity threshold due to the Au-ion reduction, defined as , compared to that of the original AuNP conjugate (0.0009 optical density [OD] vs. 0.1911 OD, respectively). This major signal enhancement elevates the sensitivity of the hs-VFA system 2 to the sub pg / mL range, a major improvement over conventional VFAs that typically operate at ng / mL levels of LoD. Further analyses on sensor performance optimization for improved signal-to-noise ratio, cartridge compressibility (FIG. 8A), AuNP size optimization (FIG. 8B), running buffer combinations (FIG. 8C), and signal amplification reagent concentrations (FIGS. 9A-9B) are provided in Supplementary Note 2.

[0044] Time-lapse imaging using a hand-held quantitative reader

[0045] Time-lapse imaging involves the periodic capture of images over a specified time period, which can play a pivotal role in understanding the dynamic nature of bio / chemical reactions, analyzing assay kinetics, and reducing the overall measurement and sensing time. Time-lapse imaging was utilized to improve the sensitivity and precision of the hs-VFA system 2. Coupled with the signal amplification strategy, the ongoing Au-ion reduction process was periodically imaged. It was hypothesized that this approach would mitigate some of the non-specific signals typically associated with prolonged reaction times in conventional end-point measurements, thus allowing for more sensitive measurements.

[0046] To test this hypothesis, the time-lapse and end-point imaging methods were compared using the cTnI assay. In the time-lapse approach, data acquisition was restricted to three images of the sensing membrane 16 sampled at 30s intervals (t = 0, 30, 60s) to avoid the generation of excessive data with even shorter capture intervals (FIG. 2H). The time-lapse hs-VFA signal was computed that was obtained from these images to measure the intensity increase over time (see the “Limit of detection of cTnI hs-VFA” subsection for the definition of the time-lapse hs-VFA signal). In the end-point measurement method, however, the intensity was measured from a single image captured at t = 360s when the signal intensity saturated (FIG. 2H). An end-point at t = 60s provided insufficient signal intensity, preventing2024-179-2 the differentiation of low cTnI levels with acceptable statistical significance (P < 0.05). For comparison, each dataset was normalized by dividing the intensity values using the negative control results. Analyses showed that the time-lapse imaging approach significantly outperformed the end-point measurement and enhanced the assay sensitivity by 6.5-fold, also reducing the CV in low cTnI levels (≤ 10 pg / mL) from 11.0% to 4.2% and shortening the readout time to 1 min (FIG. 2I). Based on these analyses, time-lapse imaging (with t = 0, 30, 60s) was adopted as part of the hs-VFA concentration measurement strategy, which will be further detailed in the following subsections.

[0047] Computational data processing: virtual control indicator and outlier analysis

[0048] Incorporating assay quality control algorithms ensures the precision and reliability of the hs-VFA system 2. To achieve this, a two-tier computational data processing approach was developed, which included a virtual control indicator (VCI) and outlier analysis (OA) protocol (FIG. 3A). It was hypothesized that the incorporation of the VCI would act as an internal benchmark to monitor assay failure, and the OA would enable the identification and exclusion of outlier test data points per test. By strategically leveraging these digital data processing techniques, the robustness, sensitivity, and precision of the hs-VFA results was improved.

[0049] The VCI functions as the first algorithmic step that assesses the validity of the hs- VFA results (FIG. 3A, inset i). In contrast to conventional immunoassays, like lateral flow assays that rely on binary on / off responses from the control area for assay validation, a more advanced approach was employed. A VCI max / min threshold was constructed by consolidating data from 120 hs-VFA test results. To establish the max / min threshold for determining assay validity, ±1.96 SD from the mean of the intensity ratio between the positive (i.e., ) and the negative (i.e., ) control spots was used; see the Methods subsection “Computational analysis of hs-VFA signals”. This VCI signal range is effectively used to filter out any false signals that could occur due to interference or non-specific interactions during the hs-VFA operation. If anomalies arise due to assay failure or sample- related issues, the VCI signal will be out of its acceptable range, and the system will display an "invalid" message, suggesting the user perform further dilution and / or retesting. The impact and details of the application of VCI on the processing of clinical samples are detailed in the "Limit of detection of cTnI hs-VFA" subsection of the Results.

[0050] Following the VCI-based virtual quality control that utilizes signals from the positive and negative control spots 18, the second digital step, OA, is used to statistically2024-179-2 validate the quality of the assay data using the raw intensities of the ten (10) test spots 18shown in the inset ii of FIG. 3A, the test data points that lie outside of the 95% confidence interval (CI) level are excluded. This process helps mitigate spot-to-spot variations that may be introduced by potential errors, such as nonuniformity in the fluid flow within the paper matrix, non-specific binding, or irregularities in antibody spotting. As shown in FIG. 3B, the OA algorithm significantly improves intra-assay uniformity (from a CV of 4.1% to a CV of 1.1%) and inter-assay reproducibility (from a CV of 8.2% to a CV of 5.1%). Additional details on the design and performance of this OA algorithm are provided in Supplementary Note 3. The impact of the OA algorithm on improving assay sensitivity is detailed in the next subsection.

[0051] Limit of detection of cTnI hs-VFA

[0052] The LoD of the optimized hs-VFA system 2 was evaluated using titration experiments with different concentrations of cTnI spiked and serially diluted in cTnI-free human serum. The signal intensity of the optimized hs-VFA with signal amplification, time- lapse imaging, and OA was calculated according to equation (1):

[0054] whereOA-refined test signal at time point t (t=1,2,3 - corresponding to 0, 30, 60s, respectively) after subtraction of the negative spot signal (refer to “Computational analysis of hs-VFA signals” subsection in Methods for thedefinition ofNormalized time-lapse signal ( ;FIGS. 3C-3D) was obtained as:

[0056] wheretime-lapse hs-VFA signal for the negative control sample (i.e., 0 pg / mL cTnI).

[0057] The results of the titration assay revealed that the hs-VFA system 2 provided a response that was proportional to varying cTnI concentrations across the clinically relevant range (100–105pg / mL)18(FIG. 3C and FIG. 10). The detectable range spanned six orders of magnitude without signal degradation due to the hook effect. Signal saturation was observed above the cTnI level of 105pg / mL, beyond the clinically relevant range. Based on these measurements, the LoD of the hs-VFA was determined to be 0.2 pg / mL using the following equations:2024-179-2 (3) (4)

[0058] where the time-lapse hs-VFA signal ( ) of the Mean blankvalue, SD of the blank value, and the SD of the lowest measured cTnI value were 1.000, 0.029, and 0.103 pg / mL, respectively; and the lowest measured cTnI value was 1 pg / mL. The LoB value was further estimated at 0.13 pg / mL by converting the time-lapse hs-VFA signal( , y-value) to a concentration value (x-value) using the optimalcalibration curve in FIG.3C (i.e., with OA, ). Finally, the LoD value was calculated as 0.2 pg / mL using the same optimal fitting method.

[0059] In addition, as demonstrated in FIG. 3C, the assay results using OA showed very good precision (with a CV of 3.0 ± 1.6%) and a coefficient of determination (R2) of 0.998. However, without OA, a relatively worse performance was reported, with an LoD of 0.42 pg / mL, a CV of 6.1 ± 4.3%, and an R2of 0.986, confirming the impact of OA in improving the precision and robustness of the hs-VFA.

[0060] The statistical validation of the assay results for low cTnI levels at or below the 99thpercentile URL level (1–50 pg / mL) showed that all the P-values were lower than 0.001 (FIG. 3D). This demonstrates that the hs-VFA can effectively differentiate between cTnI concentrations, even at levels below the clinical cut-off.

[0061] These results and the overall performance of the hs-VFA system 2 adhere to two crucial criteria outlined in clinically recommended guidelines for hs-cTnI testing: (i) achieving a CV of ≤10% at the 99thpercentile URL, and (ii) detecting concentrations at or above the assay’s LoD in over 50% of healthy individuals (i.e., a few pg / mL to 50 pg / mL), demonstrating the approach’s alignment with the well-established clinical standards.

[0062] Next, clinical sample testing was performed using sixty-two (62) samples, including fifty-four (54) patient serum samples obtained from UCLA Health and eight (8) derived samples produced by diluting two of these patient samples with cTnI-free serum. The ground truth values were determined using an FDA-cleared benchtop analyzer, with cTnI concentrations below 40 pg / mL recorded as “< 40 pg / mL” due to the equipment’s cut-off level. The assay response plot demonstrated an increasing signal with rising cTnI concentrations (FIG. 3E, inset i); see the “Computational analysis of hs-VFA signals” subsection of the Methods. Four samples with cTnI concentrations ≥ 40 pg / mL showed significant deviations from the expected trend. However, the application of VCI and OA2024-179-2 successfully identified and excluded these four outlier samples (FIG. 3E, inset ii), leading to enhanced precision (average CV reduced from 5.4% to 2.7% after OA). The variations observed between the time-lapse assay signals (i.e., ) and the ground truth measurement values may be due to (i) potential troponin degradation during sample storage, (ii) biochemical interferents such as lipemia or elevated bilirubin, or (iii) the presence of cTnI autoantibodies or heterophile antibodies in patient samples. FIGS. 3A-3E provides detailed results corresponding to the cases before and after applying VCI and OA for each clinical serum sample.

[0063] To infer cTnI concentrations from the captured hs-VFA signals, trained neural network models 80, 82, 84 were applied for accurate quantification of cTnI in serum samples. The impact of deep learning-based cTnI quantification based on the hs-VFA signals will be detailed in the following subsection.

[0064] Neural network-based analysis of cTnI hs-VFA and blind testing with clinical samples

[0065] Neural networks consist of multiple interconnected layers with non-linear activation functions, which allow for robust quantitative analysis of biological samples with point-of-care (POC) sensors despite additional noise from the sample matrix and the low-cost nature of these sensors. Here, fully connected neural networks 80, 82, 84 were used to measure cTnI concentrations in clinical serum samples from the time-lapse hs-VFA signals (i.e., , defined by equation (1)). The neural network-based hs-VFA analysis consisted of two successive parts: (i) classification of serum samples as either ≥ 40 pg / mL or < 40 pg / mL using a trained neural network 80 since the clinical cut-off of the ground truth benchtop cTnI analyzer used was 40 pg / mL; and (ii) quantification of cTnI concentration for the samples in the ≥ 40 pg / mL range using trained neural networks 82, 84 (FIG. 4A). For this cTnI measurement / inference task, the architecture of the neural network models was first optimized using a portion of the samples (validation set). Based on this optimization (detailed in the Methods section), three (3) individual fully-connected neural networks 80, 82, 84 were used that collaborate with each other as illustrated in FIG. 4: neural network 80 is used for the classification between the two concentration ranges (i.e., ≥ 40 pg / mL or < 40 pg / mL; FIG. 11) and the other two neural networks 82, 84 are used for cTnI quantification (i.e., and ; FIGS. 12A-12B). If revealed a classification decision of < 40 pg / mL, that was the final decision, and thequantification neural networks 82, 84 ( or ) were not used. The2024-179-2 quantification stage was only used when blindly classified the sample as ≥ 40 pg / mL. In this quantification stage, the serum sample measurement was first processed by the model 82 and its inference revealed the cTnI concentration of the serum sample; this blind inference of was used as the final cTnI concentration measurement if it was ≥ 40 pg / mL - i.e., complying with the former inference(see FIG. 4A). However, if the inference of predicted < 40 pg / mL, contradicting the prediction of forsame sample, a second quantification model 84 was used, namely , which was treated as an adjudicative model. was only used if the classification and quantification neural networks disagreed with each other and the blind inference of was used as the final concentration measurement if it was ≥ 40 pg / mL - i.e., complying with the inference of (see Supplementary Note 4). If the prediction from also disagreed with the inference of , the sample was then labeled as “undetermined” and excluded from the quantification results. Therefore, the final cTnI quantification was implemented with this collaboration among the three neural networks 80, 82, 84, as illustrated in FIG. 4A.

[0066] 84 was trained separately from 82 using samples with lower cTnI concentrations (i.e., < 1,000 pg / mL), whereas 82 was trained across a larger concentration range of 40–40,000 pg / mL. Quantitative cTnI predictions by the optimized neural network models for the samples from the validation set showed a strong correlation with the ground truth concentrations measured by the benchtop device, achieving a Pearson’s r of 0.962 (see FIG. 13B, and the “Neural network-based analysis” subsection of the Methods for details of the neural network training and architectures).

[0067] After optimizing the cTnI inference models using validation serum samples, the resulting optimal neural network models 80, 82, 84 were blindly tested on a set of forty-six (46) serum samples from twenty-three (23) patients - never seen before. Serum samples from each patient were measured in duplicate with two separate hs-VFAs 10 activated per patient. Blind testing results for the classification model 80 (with a clinical cut-off of 40 pg / mL) showed a sensitivity and a specificity of 97.1% and 83.3%, respectively, with 1 false negative (FN) and 2 false positive (FP) predictions (FIG. 4B). These FN and FP predictions can be attributed to the potential degradation of the clinical samples over storage time, matrix effects, low-cost design of the VFA cartridge 10 or the small size of the model training set. It2024-179-2 should be emphasized that the performance of the classification model 80 could be further improved by duplicate sample tests – at the cost of an increase in the sample volume per test. For instance, the sample can be labeled as “undetermined” if the variation in the predicted concentrations between the duplicate tests exceeds a certain threshold. This threshold was empirically determined by analyzing the mean absolute error (MAE, see the “Statistical analysis” subsection of the Methods) between the duplicate tests of each sample. It was found that the MAE scores for all 3 falsely classified samples (i.e., 1 FN and 2 FPs) were larger (i.e., > 10%) than for the correctly labeled samples. Therefore, with a test-to-test variation threshold of ~10%, one could reduce the number of FNs and FPs to 0 and improve the accuracy of 80 to 100% by correcting the falsely classified samples as “undetermined”. Due to the requirement of duplicate testing, this strategy was not employed in the rest of the clinical testing analysis and kept the serum sample volume at 50 μL per test.

[0068] After this initial blinded cTnI classification stage, thirty-five (35) serum samples that were classified into ≥ 40 pg / mL range, including thirty-three (33) correctly classified samples (from seventeen (17) patients) and two (2) falsely classified samples (from one (1) patient), were further processed by 82. Out of these thirty-five (35) samples, twenty-one (21) were only quantified by 82 revealing the final concentration measurements; the remaining fourteen (14) had contradicting predictions between 82 and 80, and therefore were also processed by 84 - the adjudicative quantification model. Since none of the fourteen (14) predictions from 84 disagreed with 80 (i.e., all concentration predictions from 84 were in the ≥ 40 pg / mL range) no serum samples from the blind testing set were labeled as “undetermined”. The final quantitative predictions of cTnI concentrations from the two quantification neural network models on these thirty-five (35) samples had a good match with the gold standard measurements from an FDA-cleared analyzer, achieving Pearson’s r of 0.965. In addition, the mean variation between duplicate measurements of the samples was minimal, with an average CV of 6.2%, confirming a repeatable inter-sensor operation (FIG. 4C).

[0069] Importantly, cTnI quantification by the neural network-based method outperformed a standard rule-based method where cTnI concentration and the time-lapse hs- VFA signal are related to each other with an explicit equation. The optimal equation from this rule-based method for the case represented a power law fit (see Supplementary Note 5). To provide a fair comparison, the deep learning models and the power law function were created2024-179-2 using the same training set, and then tested on identical blinded serum samples. For the sametest samples, the optimized quantification network models ( 82 and84) outperformed the power-fitting model in terms of accuracy and reproducibility (i.e., Pearson’s r: 0.861 vs. 0.965 and CV: 10.5% vs. 6.2%; FIG. 14A). The advanced performance of neural network models 80, 82, 84 originates from their inherent universal function approximation power and ability to effectively learn robust quantification functions between the analyte concentration and non-linear time-lapse response of hs-VFA despite the interference of noise from clinical samples and the low-cost design of the paper-based POC sensor.

[0070] In addition, the performance of the cTnI quantification by neural network models 80, 82, 84 improved from the incorporation of the OA step and the use of the time-lapse imaging method. For instance, cTnI quantification precision achieved by the optimized deeplearning models ( 82 and 84) using the time-lapse signal inputsprocessed by OA (i.e., ) was higher compared to the precision achieved by the same models using the time-lapse inputs without OA (see FIG. 14B, a CV of 8.8% for time- lapse inputs without OA vs. a CV of 6.2% for time-lapse inputs with OA). In addition, the cTnI concentrations predicted by optimized models with the end-point inputs (i.e., at t = 60 s) had a larger deviation from the ground truth values than the predictions from the same models with the time-lapse inputs (FIG. 14C). This was especially apparent in the lower concentration range. In the models with end-point and time-lapse inputs, Pearson's r coefficients were 0.806 and 0.959, respectively, for concentrations below 1,000 pg / mL. Therefore, both the time-lapse imaging and computational assay quality check (i.e., OA) have a positive impact on the quantification performance of the neural network models, and the incorporation of these methods into the neural network-based analysis is important for more accurate and robust quantification of cTnI concentrations.

[0071] Discussion

[0072] cTnI concentrations quantified by the hs-VFA system 2 using neural network- based analysis showed a strong correlation with the ground truth measurements obtained by an FDA-cleared analyzer (achieving a Pearson’s r of 0.965) and demonstrated competitive testing precision with an average CV of 6.2%, which falls within the precision requirement of the clinical hs-cTnI assays (i.e., CV of <10%). In addition, the neural network-based algorithms 80, 82, 84 were able to quantify cTnI levels over a large concentration range and correctly classify samples above / below the clinical cut-off (40 pg / mL). This successful2024-179-2 performance of the hs-VFA system 2 can be attributed to the synergistic effects of several key innovations: (i) a signal amplification reaction which provided a 212.3-fold increase in the sensitivity threshold of the assay; (ii) time-lapse imaging which provided an additional 6.5-fold increase in the sensitivity threshold; (iii) computational assay quality check algorithms, including VCI and OA, which provided a 2.1-fold improvement in LoD also reducing the CV of the assay; and (iv) neural network-based cTnI inference approach which further helped reduce the CV. Moreover, the compact and cost-effective paper-based sensor ($3.86 per test), paired with a custom-designed Raspberry Pi-based user-friendly portable reader device 40 (priced at approximately $170 per prototype), emphasizes the suitability of this platform for POC assays and positions it as a viable alternative for standard laboratory testing. When benchmarked against commercial cTnI-POCT assays and recent research results in the literature, the hs-VFA system 2 has a competitive precision over a wide assay range, spanning six orders of magnitude in concentration, fully covering the clinically relevant cTnI range (0.01 – 100 ng / mL). It also provides better sensitivity, with an order of magnitude lower LoD. The improved performance of the hs-VFA system 2 could be leveraged to enable 0-hour rule-out with a single cTnI test in low-risk patients and improve the prehospital phase of care for high-risk patients empowered by accurate cTnI testing even in primary care or nursing home infrastructure.

[0073] The quantification of the analyte concentration in clinical samples conventionally relies on calibration curves established using, e.g., spiked samples with known antigen concentrations. However, this method has inherent limitations that may compromise the accuracy and adaptability of the assay. For example, when applying this calibration-based approach to the hs-VFA clinical sample results, a significant performance degradation was observed, as shown in FIG. 14A. This limited performance in calibration-curve-based cTnI quantification may arise due to various factors, including (i) matrix effects from the spiked and clinical samples, (ii) varying storage conditions for different sample batches, (iii) variations in the assembly between sensor batches, and (iv) limited control over environmental factors (i.e., temperature, humidity). Furthermore, ensuring the accuracy of the calibration curve over time can be challenging, necessitating frequent recalibration and quality control measures. In contrast, neural network-based concentration inference can adapt to more diverse sample matrices and variations in sample / sensor batches better than a fixed function determined with a single explicit rule.2024-179-2

[0074] It should also be noted that the introduction of the Raspberry Pi computing device 54 for constructing the portable reader device 40 yields multiple benefits compared to using a smartphone-based reader device. The custom-designed Raspberry Pi-based reader device 40 showed equivalent quantification performance to a smartphone-based reader device (FIG. 15). The Raspberry Pi’s open-source ecosystem delivered higher flexibility and customization options in hardware and software, facilitating the operation of the device for time-lapse imaging-based analysis. The standalone nature of the Raspberry Pi-based reader device 40 can help unify the design to a fixed footprint and camera model, contrary to smartphone-based reader devices 40, which are sensitive to the model brand and camera optics, which frequently change as new models are introduced in the market. Furthermore, the Raspberry Pi reader device 40 offers cost benefits compared to high-end smartphones. The Raspberry Pi-based custom-designed hs-VFA reader device 40 can be easily equipped with batteries and communication modules (3G / 4G / 5G / Wi-Fi), fulfilling the real-time connectivity requirements of POCT. This could help transform the hs-VFA platform into a more advanced POCT assay that is not only used in traditional medical facilities but also enables cTnI measurements to be performed during patient transport (e.g., ambulance-based testing), thus expanding the potential diagnostic use cases.

[0075] One of the major advances behind the presented hs-VFA results is the highly sensitive detection of a single biomarker with an LoD of < 1 pg / mL. To increase the analytical sensitivity, the sensing membrane 16 was designed to improve the overall surface area for capturing the target analyte, coupled with a signal amplification reaction. With this strategy, the sensing membrane 16 was configured to conduct ten (10) repeated tests under equivalent conditions across seventeen (17) multiple spots in a single assay run (all within 15 min), which is not achievable with conventional rapid diagnostic tests such as lateral flow assays. When investigating the impact of the number of test spots 18 on the assay sensitivity without data refinement, it was observed that increasing the number of test spots 18 enhanced the signal intensity and assay sensitivity (FIGS. 16A-16B). Nevertheless, the assay uniformity (FIG. 16C) and repeatability (FIG. 16D) were reduced due to spot-to-spot variability between the inner (closer to the center) and outer (closer to the edges) reaction spots 18 on the sensing membrane 16. These spot-to-spot and test-to-test variation issues were resolved through VCI and OA by taking advantage of the statistical benefits derived from the redundancy of multiple test spots 18 on the test membrane, similar in concept to “swarm sensing”. The experiments on cTnI spiked / clinical sample assays showed that2024-179-2 excluding the compromised spots 18 ensures data integrity, achieving more reliable and consistent assay results without a trade-off in assay sensitivity (FIGS. 3B-3D). This approach mitigates errors inherently associated with paper-based sensors (due to the use of inexpensive components such as paper membranes, the possibility of nonuniformity in fluid dispersion, and potential defects during the antibody spotting and sensor assembly processes), demonstrating that high-sensitivity analyte measurements at the pg / mL level can be reliably achieved using the low-cost assay platform of the hs-VFA system 2. Additionally, incorporating OA during the assay development and validation stages has the potential to reduce sensor production costs by leveraging the spot position-dependent exclusion rate (FIG. 3, insets i–ii). For example, by selecting six (6) test spot positions that consistently exhibit a < 30% exclusion rate, a 36% reduction in the cost of antibodies (equivalent to $0.7 per test) was anticipated. This reduction is significant, as antibodies purchased at low volumes account for > 50% of the unit sensor cost; it should also be emphasized that over large-volume manufacturing practices, the overall cost per test can be reduced to < $1 per test.

[0076] In summary, a deep learning-enhanced paper-based hs-VFA system 2 was developed that achieved highly sensitive (LoD, 0.2 pg / mL) and precise (average CV < 7%) quantification of cTnI in serum samples within 15 min per test, showcasing significant progress in advancing high-performance POC sensors. Notably, the performance enhancement of hs-VFA system 2 was achieved without compromising key attributes inherent to POCT, such as low cost, simple operation, rapid assay times, and digital connectivity, which underscore the significance of this platform. The results highlight the potential to democratize diagnostic testing by demonstrating that high-quality assays, traditionally limited to high-end clinical analyzers in central laboratories, can be performed on cost-effective POC platforms. This technology can be widely used to expedite global diagnostic equity for testing challenging disease biomarkers.

[0077] Methods

[0078] Antibody conjugation to AuNP: Antibody conjugation to 15 nm AuNP is based on physical adsorption. In brief, the anti-cTnI detection antibody (10 μL, 1 mg / mL; 19C7, Hytest) was suspended in 15 nm AuNPs (1 mL; BBI Solutions) mixed with 100 mM borate buffer (100 μL, pH 8.5; Thermo Scientific). Following a 1 h incubation using a rotary mixer (20 rpm) at room temperature (RT), the conjugate was blocked for 2 h by adding bovine serum albumin (BSA; 10 μL, 10% w / w; Thermo Scientific). Then, the conjugate underwent three rounds of centrifugal washing (21380g, 25 min, 4 ℃) using a 10 mM borate buffer (pH2024-179-2 8.5). Next, the final conjugate pellet was resuspended in the storage buffer (100 μL) containing 5% w / w trehalose (Sigma), 0.5% w / w protein saver (Toyobo), 0.2% v / v Tween 20 (Sigma), and 1% v / v Triton X-100 (Sigma) in 10 mM PBS (pH 7.2; Thermo Scientific). The absorption spectra and concentration of the conjugate were analyzed using a microplate reader (Synergy H1; BioTek). The conjugates were stored at 4 ℃ at 9 OD concentrations until ready for use.

[0079] Sensing membrane preparation and hs-VFA assembly: The sensing membrane 16 fabrication involves four steps: printing, heating, antibody spotting, and blocking. A wax printer (Xerox) was used to print and compartmentalize seventeen (17) spots 18 or reaction areas outlined by a black background onto a nitrocellulose membrane (NC, 0.2 μm; Bio- Rad). The wax-printed membranes or layers 20 were baked (120 ℃, 55s) in a forced air convection oven (Across International) to melt the wax on the membrane’s top side. This created a three-dimensional compartment with hydrophilic reaction areas or spots 18 against a hydrophobic background. Considering heat convection inside the baking chamber, up to thirty (30) sensing membranes 16 (arranged in a 6 × 5 array, with 1 mm gaps between membranes) were uniformly prepared in a single baking batch. The anti-cTnI capture antibody (0.8 μL, 1 mg / mL; 560 Hytest) and goat anti-mouse IgG (0.8 μL, 0.1 mg / mL; Southern Biotech) were respectively dispensed onto the test and positive control spots. The batch membrane sheet was subsequently dried in the oven (37 ℃, 15 min). The membrane sheet was then immersed in 1% w / w BSA solution for blocking (30 min, at RT). After another round of drying the membrane sheet (37 ℃, 15 min), it was divided into individual sensing membranes 16 using a razor.

[0080] All the paper materials for hs-VFA were prepared as previously outlined in Goncharov, A. et al. Deep Learning-Enabled Multiplexed Point-of-Care Sensor using a Paper-Based Fluorescence Vertical Flow Assay. Small 19, e2300617 (2023), which is incorporated herein by reference. Briefly, raw paper materials were precisely cut using a CO2laser cutting system (Trotec). The first top cartridge 22 for the immunoassay portion of the assay was assembled by sequentially stacking engineered paper layers 24 (absorption layer, flow diffuser 25, 1stspreading layer, interpad, 2ndspreading layer, and supporting layer) with a double-sided adhesive foam tape as an assembly frame. A concentric circular pattern in the flow diffuser 25 and outer contour 27 in the supporting layer were prepared using the same wax-printing / baking process as the sensing membrane 16. The flow diffuser, interpad, and supporting layer were treated with 1% w / w BSA solution for blocking. The bottom cartridge2024-179-2 12 was prepared by stacking five absorption layers / pads 14 and affixing the sensing membrane 16 on top using adhesive foam tape.

[0081] hs-VFA cartridges: The plastic VFA cartridges 10 were 3D-printed using Form 3 (gray resin, Formlabs) with 100 μm resolution mode. The design of the bottom cartridge 12 and firs top cartridge 22 aimed to compress the stacked paper materials by 25% of the initial total paper layers’ thickness, enhancing flow diffusion, assay uniformity, and efficiency (FIG. 13A). For the second top cartridge 28 used for signal amplification, an optically transparent window 29 (e.g., acrylic 16.3 mm × 16.3 mm × 1 mm, laser-cut) was affixed to the 3D- printed cartridge using a clear acrylic adhesive. Foam tape was employed as a gasket 31 to prevent reagent leakage in the second top case. The transparent window 29 has four ventilation outlets (0.8 mm in diameter) at each corner. These outlets are designed to remove air from the reaction chamber 34, facilitating the inflow of reagent solution. The reagent inlet 30 of the second top cartridge 28 was designed to hold a volume of up to 1 mL of solution.

[0082] Portable reader and image analysis: A custom-designed portable reader device 40 was assembled using 3D-printed parts (i.e., housing 42, LED holder, and cartridge tray 50) and low-cost off-the-shelf components (i.e., LEDs [DigiKey] for light source(s) 44, macro lens [Edmund Optics], Raspberry Pi for computing device 54, camera module V2-8- megapixel [Adafruit] as camera 46, and touch screen display [Elecrow] 56). The 3D-printed custom parts were produced by Object 30 (Stratasys) and Ultimaker 3 (Ultimaker) 3D printers. The LED module contained four green LEDs (532 nm) for time-lapse imaging and two white LEDs for brightfield imaging arranged in a circular shape. All LEDs were polished from the front to provide even light distribution across the sensing membrane 16. The reader device 40 was designed for easy pedestal installation with four optical posts, enabling both hand-held and benchtop use. The user interface of the image capture software consisted of a real-time camera screen and input fields with user-adjustable parameters (i.e., exposure time, the number of images to capture, and capture interval, FIG. 5) for automated time-lapse imaging. All images were captured in raw format under consistent imaging conditions (100 μs exposure time, and 30 s interval).

[0083] After the end of the hs-VFA operation (per test), captured time-lapse images (N=3) were processed by an automated image segmentation code that extracted the green channel from the RGB images, segmented all seventeen (17) immunoreaction spots 18 from the activated sensing membrane 16 and averaged pixels within each segmented spot to generateseventeen (17) intensity values per image ( , i {Test, Pos, Neg}– type of2024-179-2 immunoreaction, j– spot repeat within the given type, t– time point). These intensity values oftime-lapse images were further normalized by the corresponding background intensities ( )from the sensing membrane image captured before the start of the hs-VFA operation and raw absorption signals for each time-lapse image were calculated as:

[0084] = 1 – . (5)

[0085] Alike spots 18 within each immunoreaction type were averaged, resulting in a totalof 3×N raw absorption signals ( ) per assay, where N=3 is the total number of capturedimages during time-lapse operation.

[0086] Assay operation: Assay operation has two main steps: immunoassay and signal amplification. For the immunoassay, the bottom cartridge 12 is assembled with the first top cartridge 22. The process begins by activating the VFA cartridge 10 with the addition of 1strunning buffer (200 μL), which consists of 1% v / v Triton X-100 and 1% v / v BSA in PBS (10 mM, pH 7.2). After allowing 30 s for complete absorption to paper layers 24, a mixture (100 μL) of the sample (serum, 50 μL) and conjugate (2.5 OD, 50 μL) is added following immediate mixing. This mixture is left for 1 min to ensure complete absorption. Next, 2ndrunning buffer (300 μL) is introduced. This buffer contains 3% v / v Tween 20, 1% v / v albumin, 0.5% w / w protein saver, and 1% w / w trehalose in PBS (10 mM, pH 7.2). Its purpose is to maintain the flow of solutions within the hs-VFA structure and to wash away any unbound conjugate and target molecules from the sensing membrane 16, thereby minimizing non-specific binding. The first top cartridge 22 is removed from the bottom cartridge 12 after 8.5 minutes and replaced with the second top cartridge 28 for Au-ion reduction-based signal amplification. In this step, 500 μL of a reagent solution containing 10 mM HAuCl4 (Sigma) and 10 mM H3NO (Sigma) in PBS is added to the inlet 30 to supply the reagent to the sensing membrane 16 and absorption pads 14. After 3 min, when the reagent feeding is complete, the second top cartridge 28 is removed, and the cartridge 12 is transferred to the reader device 40 for time-lapse imaging.

[0087] cTnI spiked and clinical serum samples: The cTnI spiked serum samples were prepared for assay optimization and validation by spiking cTnI standard antigen (I-T-C complex purified from the human heart; Lee Biosolutions) into cTnI-free serum (Hytest). To obtain various concentrations of cTnI, serial dilutions were performed using the same serum. The cTnI antigen was stored at -80°C after being aliquoted into 1 μL portions. A fresh solution was prepared immediately before each assay to prevent potential antigen2024-179-2 degradation. Clinical serum samples containing cTnI were provided by UCLA Health. This study was approved by the UCLA Institutional Review Board (IRB no. 20-002084). Patient consent was waived since these specimens were pre-existing remnant samples collected independent of this research project. The ground truth values of the clinical samples were determined using a standard analyzer (Access, Beckman Coulter) at UCLA Health, immediately after collection. The analyzer had a cut-off level of 40 pg / mL, which quantified samples with cTnI levels ≥ 40 pg / mL and displayed results below the cut-off as < 40 pg / mL. fifty-four (54) clinical samples were tested, including thirty-five (35) samples with cTnI levels ≥ 40 pg / mL and 19 samples with cTnI levels < 40 pg / mL. Additionally, 8 serially diluted clinical samples derived from two stocks were tested. Clinical samples were stored at -80°C and were thawed immediately before the assay at 4°C for 1 h.

[0088] Computational analysis of hs-VFA signals: Prior to neural network-based analysis, all the assays activated during clinical sample testing underwent computational assay validation, which consisted of two steps: VCI and OA. At the VCI step, individual assays were excluded from the clinical sample dataset based on the ratio between rawintensities from positive ( ) and negative () control spots. An assay was excluded ifthe ratio did not fall into the CI calculated according to:the mean of the distribution of the positive and negative control ratios of one-hundred-twenty (120) assays (i.e., eighty (80) spiked samples + forty (40) clinical samples), N=3 is the total number of captured images during time-lapse operation.

[0091] Next, during the OA step, individual test spots from each assay were excluded based on 95% acceptance range from the statistical distribution of the test spots:

[0093] j=1… , t = 1…N – time point, 2.262 is the t-score for a two-tailed test with 9degrees of freedom (i.e., ). Averages (were calculatedover the test spot repeats on the same paper-based test membrane (i.e., j=1…10) independently for each of the 3 membrane images captured during the time-lapse operation.2024-179-2

[0094] In addition to OA based on absolute test spot values (termed statistical OA model), OA was also performed based on differential spot values (i.e., OA-D model). The 95% acceptance range for OA-D was defined as:

[0096] t = 1…N-1. Uniformity and reproducibilityof hs-VFA after applying the statistical OA model showed superior improvement compared to CV before any OA and CV after applying the OA-D model (FIG. 3B, Supplementary Note 3); therefore, the statistical OA method was selected for data refinement of the clinical samples. Refined hs-VFA signals after OA are defined as.

[0097] Time-lapse response from hs-VFA using the OA-refined signals ( ) wasfurther calculated according to equation (1) (i.e., ). During the sensor optimizationand testing on spiked samples (FIGS. 3C-3D), time-lapse signals ( ) werenormalized by the time-lapse signal from the negative control sample generating normalizedtime-lapse signals ( , see equation (2)). For the clinical sample test(FIG. 3E), time-lapse signals were used without normalization ( ) in order to minimize variations between different testing batches.

[0098] Neural network-based analysis: The neural network-based analysis consisted of classification and quantification parts and a total of three independent shallow fully- connected neural networks 80, 82, 84, including one network for classification 80 between samples from ≥ 40 pg / mL and < 40 pg / mL concentration ranges (i.e., ), and two networks 82, 84 for quantification of cTnI in ≥ 40 pg / mL range (i.e. and ) – see FIG. 4A. Inputs to all three neural networks 80, 82, 84 represented the time-lapse hs-VFA signals calculated from the three time-lapse images according to equationinput signal was standardized according to the formula:

[0100] where < > and σ( ) are the mean and the SD of the time-lapsesignal, respectively, calculated over the training dataset.

[0101] The classification network ( ) 80 consisted of three (3) hiddenlayers (512, 256, 128 units) with ‘ReLU’ activation functions and L2 regularization for all layers (see FIG. 11). Each hidden layer was followed by a batch standardization layer. The2024-179-2 output layer had 1 unit with a ‘sigmoid’ activation function. The loss function that was used was binary cross-entropy compiled with Adam optimizer, a learning rate of 1e-4 and a batchsize of 5. Binary cross-entropy loss ( is defined as:

[0102]

[0103] where are the ground truth labels (i.e., “1” for samples from ≥ 40 pg / mL concentration range and “0” for samples from < 40 pg / mL concentration range), are the predicted labels and is the batch size.

[0104] The quantification part consisted of two independent fully-connected neural networks (namely and ) 82, 84. 82 contained three (3) hidden layers (256, 128 and 64 units), each with ‘ReLU’ activation functions and L2 regularization. 84 consisted of two (2) hidden layers (128 and 64 units), each with ‘ReLU’ activation functions and no regularization (FIGS. 12A-12B). Each hidden layer in both models was followed by a batch standardization layer. The loss function for both models was the mean squared logarithmic error (MSLE) compiled with Adam optimizer, a learning rate of 1e-4 and a batch size of 5. MSLE loss is defined as:

[0106] where are the ground truth cTnI concentrations, are the predictedconcentrations. Both of these quantification models ( and )predicted cTnI concentrations ( ) in pg / mL. Architectures of all three neural networks wereoptimized through a 4-fold cross-validation on the validation sets.

[0107] The classification model ( ) 80 was trained on sixty-four (64)serum samples from thirty-two (32) patients and validated on forty-two (42) samples from twenty-one (21) patients (i.e., serum from each patient was tested in duplicates with two activated hs-VFAs per patient). Out of sixty-four (64) training samples, fifty-four (54) were clinical samples and the remaining ten (10) were diluted samples. The optimal classificationnetwork ( ) 80 achieved 84.6% sensitivity and 93.8% specificity on thevalidation set (FIG. 13A). The optimized model was further blindly tested on forty-six (46) additional serum samples from twenty-three (23) patients, achieving a sensitivity and specificity of 97.1% and 83.3%, respectively, as reported in the Results section.

[0108] The quantification stage was only used when 80 blindly classified the sample as ≥ 40 pg / mL. 82 was trained on sixty-four (64)2024-179-2 samples from thirty-two (32) patients and validated on twenty-two (22) samples from eleven (11) patients. cTnI concentrations in the training samples were in 40–40,000 pg / mL range, while cTnI concentration range for the validation set was 50–5,000 pg / mL. Quantification results from 82 revealed the final cTnI concentration measurement per sample if and only if its blind inference agreed with 84 inference (see FIG. 4A). For ten (10) out of twenty-two (22) validation samples, the blind inference results of 82 were in < 40 pg / mL range, contradicting the predictions from the classification model (i.e., all of these samples were classified into ≥ 40 pg / mL range by 80). As illustrated in FIG. 4, such samples were further processed by a second quantification network, , which was used as an adjudicative model. If the blind inference from 84 still disagreed with , the sample was labeled as undetermined; otherwise the prediction from 84 was used as the final concentration measurement result for the cases where 80 and 82 disagreed with each other. 84 was trained separately from 82 on thirty-four (34) serum samples (from seventeen (17) patients) with cTnI ground truth levels below 1,000 pg / mL and validated on ten (10) samples (from five (5) patients) with cTnI concentrations in 50–200 pg / mL range.

[0109] One out of ten (10) validation samples processed by 84 had < 40 pg / mL concentration prediction and it was labeled as “undetermined”. Predictions of the optimized quantification models on the rest of the validation samples (i.e., twenty-one (21) samples) had a high correlation with the ground truth cTnI concentrations quantified by an FDA-cleared analyzer with a Pearson’s r of 0.962 (FIG. 13B). For the blind testing set composed of eighteen (18) new serum samples, these networks 80, 82, 84 achieved a Pearson’s r of 0.965 as reported in the Results section. Training sets for both 82 and 84 included samples in < 40 pg / mL range and during the training process, MSLE loss for such samples was increasing only if the predicted concentration was ≥ 40 pg / mL. Incorporation of the samples under 40 pg / mL cut-off into the training sets helped to create more robust quantification models and achieve reliable cTnI quantification despite a limited number of training samples.

[0110] Statistical analysis: For assay validation and cTnI spiked serum sample tests, all experimental data were presented as the mean of at least three measurements ± SD. Clinical sample test results were derived from the mean of duplicate measurements ± SD. Further2024-179-2 details regarding experimental replicates are provided in the corresponding FIG. legends. The CV was calculated by dividing the SD by the mean (%). Group differences were assessed using an unpaired two-sample t-test, with statistical significance set at P < 0.05. In addition to CV, MAE was used to estimate repeatability between duplicated samples, defined as:

[0112] where Rep1 and Rep2 are first and second repeats of the serum sample respectively.

[0113] Supplementary Note 1 | Signal amplification reaction.

[0114] The colorimetric signal amplification reaction occurred as a result of the chemical reduction of Au3+on the surface of the AuNP conjugates. The reagent solution, which is a mixture of Au3+and H3NO, is injected into the second top cartridge 28 and initially flows downward into the absorption pads 14 through the sensing membrane 16. After removing second top cartridge 28 after 3 min, the amplification reaction continued with the reagent flowing back up from the absorption pads 14 to the sensing membrane 16 by evaporation (FIGS. 6A-6B). As spectroscopic evidence of the amplification reaction, a significant redshift was observed (from 520 nm to 560 nm) of the absorption peak, broadening of the peak, and darkening of the aqueous solution, which supports signal amplification by particle growth (see FIGS. 7A-7B).

[0115] Supplementary Note 2 | Assay optimization.

[0116] Assay steps were optimized to enhance the signal-to-noise ratio in the hs-VFA system 2. Tests regarding cartridge 10 compressibility, which determines the degree of paper layer compression when assembled, showed that the best signal-to-noise ratio occurred at 25% compressibility, where signal intensity increased, and non-specific intensity decreased (FIG. 8A). This was attributed to improved contact efficiency between paper layers 24, enhancing the washing of remaining conjugates in the first top cartridge 22 and free conjugates in the sensing membrane 16. Compressibility beyond 25% was not considered due to the relatively thick paper layers, making it difficult to assemble the top and bottom cartridges 22, 12. Using 15 nm AuNPs led to a 50% increase in signal intensity compared to 40 nm AuNPs (FIG. 8B). This enhancement is due to the superior membrane permeability ofthe 15 nm AuNPs and the higher particle quantity (1.40 1012 / mL) at the same optical density(OD) concentration, approximately 15.5 times more than 40 nm AuNPs (9.00 1010 / mL). Inthe optimized running buffer composition and combination, an average improvement in signal intensity of approximately 44–68% (FIG. 8C) was observed. For amplification, signal2024-179-2 intensity and inter-assay uniformity were evaluated at the testing spots 18 with various combinations of Au3+and H3NO concentrations and selected the condition that ranked highest in both criteria (Au3+: 10 mM and H3NO: 10 mM, FIGS. 9A-9B).

[0117] Supplementary Note 3 | Comparison of two outlier analysis (OA) models.

[0118] While integrating the OA algorithm into the workflow, the statistics-driven approach (statistical OA model) was compared with a differential OA method (OA-D), which utilized differential changes in raw intensities between the time-lapse images (see the “Computational analysis of hs-VFA signals” subsection in the Methods for more details about OA and OA-D). This comparison enabled the impact of different OA methods on assay uniformity (intra-assay) and reproducibility (inter-assay). When analyzing assay data sets (N=330) using OA algorithms, the statistics-based analysis retained an average of 6.5 ± 1.5 out of the original ten (10) test spots 18, whereas the OA-D method retained 4.8 ± 2.1 spots 18 (FIG. 3B, inset i). In the statistical OA model, higher exclusion rates were observed for spots 18 located towards the outer edges of the sensing membrane 16 (i.e., No. 1, 2, 5, 10 in FIG. 3B, inset ii). This is mainly due to minor differences in flow homogeneity between the inner and outer spots 18, within a CV of 5%. When comparing the intra-assay CV before and after implementing the OA, a substantial improvement in spot-to-spot assay uniformity was observed (FIG. 3B, inset iii) and assay reproducibility (FIG. 3B, inset iv). Specifically, with the use of the statistical OA model, the assay uniformity improved from a CV of 4.1% to a CV of 1.1%, and assay reproducibility improved from a CV of 8.2% to a CV of 5.1%, which was superior to the differential model (OA-D). Therefore, the statistical OA model was chosen for subsequent analysis as it offered better generalizability (by retaining more test spots 18) and achieved superior data refinement compared to other exclusion methods.

[0119] Supplementary Note 4 | Limitations of cTnI quantification with a single neural network model.

[0120] Quantitative predictions of cTnI concentrations from DNNQuantification82 for fourteen (14) blind tested samples were < 40 pg / mL, contradicting the predictions from DNNClassification 80 (i.e., these samples were predicted as ≥ 40 pg / mL by DNNClassification 80). Ground truth cTnI concentrations in all of these samples were in 50–200 pg / mL range, whereas quantitative outputs from DNNQuantification 82 were equal to 0 pg / mL. One of the reasons for such false negative outputs from DNNQuantification 82 for the lower concentration samples can be the limited amount of training samples. Only twelve (12) out of sixty-four (64) training samples fell into 50–200 pg / mL concentration range and DNNQuantification822024-179-2 might be overfitting to higher concentrations considering the large total dynamic range of the samples used to train the model (i.e., 40–40,000 pg / mL). The overfitting issue can be minimized by including more samples from all relevant clinical ranges into the training set. However, even with large training sets and minimal model overfitting, a single quantification network (i.e., DNNQuantification 82) might still return contradicting results with respect to DNNClassification80, especially for the borderline samples (i.e., samples close to the cut-off level of the equipment used to measure ground truth concentrations, 40 pg / mL in this work). Quantification errors on such samples can be attributed to the proximity between the assay signals for borderline samples and samples from the < 40 pg / mL range, as well as the interference of the noise factors from the sample matrix and the low-cost nature of hs-VFA cartridges 10. Keeping an adjudicative neural network (i.e., 84) to process samples with contradicting predictions between DNNClassification 80 and DNNQuantification 82 would still be beneficial to achieve optimal cTnI concentration accuracy in the low concentration ranges and improve sensitivity for more reliable high-sensitivity cTnI detection.

[0121] Supplementary Note 5 | Power-fitting function.

[0122] The performance of the optimized quantification neural networks was compared with a rule-based method where the time-lapse response of the hs-VFA system 2 was related to cTnI concentration through a power-fitting function. The input to the function was time-lapse signal from hs-VFA ( defined by equation (1). The power-fitting functionform is defined by the power law and the explicit function form is outlined below:

[0123] , a = 4.359, b = 3.636, (S1)

[0124] is the predicted concentration. The performance of this power-fitting model on blindly tested clinical samples is shown on FIG. 14A.

[0125] 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. The invention, therefore, should not be limited except to the following claims and their equivalents.

Claims

2024-179-2 What is claimed is:

1. A method of detecting the presence of and / or quantifying the amount or concentration of one or more analytes in a sample using a vertical flow assay cartridge comprising one or more cartridges having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more types of capture agent(s), the method comprising: loading a mixture of the sample and detection reagents into the sample inlet along with signal amplification reagents; imaging the sensing membrane with a reader device configured to illuminate and obtain a plurality of time-lapsed images of the sensing membrane containing time-lapsed intensity signals for the plurality of spots; and processing the time-lapsed intensity images and / or signals for the plurality of spots 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 in the sample.

2. The method of claim 1, wherein the one or more cartridges comprises a bottom cartridge, a first top cartridge, and a second top cartridge, wherein the bottom cartridge comprises the sensing membrane, the first top cartridge comprising one or more paper layers and a sample inlet, the second top cartridge comprising an inlet and a fluidically coupled reagent chamber.

3. The method of claim 2, further comprising: securing the first top cartridge to the bottom cartridge and loading the sample and a reagent mixture into the sample inlet; removing the first top cartridge from the bottom cartridge and securing the second top cartridge to the bottom cartridge and loading the signal amplification reagents into the inlet of the second top cartridge so as to fill the reagent chamber.

4. The method of claim 1, wherein the capture agent type of each spot comprises one or more antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, or peptoids,2024-179-2 5. The method of claim 1, wherein the algorithm or machine learning comprises a first neural network that classifies the concentration(s) of the one or more analytes in the sample as either above or below a threshold value and a second computing algorithm involving one or more neural networks configured to output the concentration(s) of the one or more analytes in the sample that are above the threshold value.

6. The method of claim 5, wherein the algorithm or machine learning comprises a third neural network (DNNLow) that outputs one of: (1) a quantification of the concentration of the one or more analytes in the sample or (2) an undetermined output.

7. The method of claim 1, wherein the one or more analytes comprise proteins, including cardiac troponin I (cTnI), 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.

8. The method of claim 1, wherein the sensing membrane is populated with one or more negative control spots containing no capture antibodies.

9. The method of claim 1, wherein the plurality of time-lapsed images comprise images obtained over at least three time periods.

10. The method of claim 1, wherein the mixture of the sample and detection reagents comprises serum, plasma, or whole blood, one or more buffers, and one or more capture agent(s) conjugated to metallic nanoparticles.

11. The method of claim 10, wherein the one or more capture agent(s) conjugated to the metallic nanoparticle comprises one or more antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, or peptoids, 12. The method of claim 1, wherein the signal amplification reagents comprises one or more metal ions and one or more reducing agents, or enzymatic substrates.2024-179-2 13. 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.

14. The method of claim 2, wherein the bottom cartridge is secured to a cartridge tray and the cartridge tray and bottom cartridge are inserted into the reader device for imaging the sensing membrane.

15. A system for detecting the presence of and / or quantifying the amount or concentration of one or more analytes in a sample comprising: a vertical flow assay cartridge comprising one or more cartridge subunits having a sample inlet and a sensing membrane populated with a plurality of spots containing one or more types of capture agent(s); a reader device comprising a housing that includes one or more light sources, a camera, and a cartridge receiver that receives the bottom cartridge and places the sensing membrane along an optical path between the one or more light sources and the camera; and a computing device configured to control the one or more light sources and the camera for obtaining a plurality of time-lapsed images of the sensing membrane, the computing device further comprising 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 a classification of the sample and / or a quantification of the amount or concentration of the one or more analytes in the sample based on the time-lapsed intensity images for the plurality of spots.

16. The system of claim 15, wherein the one or more cartridge subunits comprises a bottom cartridge, a first top cartridge, and a second top cartridge, wherein the bottom cartridge comprises the sensing membrane, the first top cartridge comprises one or more paper layers and a sample inlet, the second top cartridge comprises an inlet and a fluidically coupled reagent chamber.

17. The system of claim 16, further comprising a cartridge tray that receives the bottom cartridge, the cartridge tray being insertable into the cartridge receiver.2024-179-2 18. The system of claim 16, wherein the reagent chamber of the second top cartridge is optically transparent and wherein the cartridge receiver receives the bottom cartridge and the second top cartridge secured to one another and images the sensing membrane through the optically transparent reagent chamber.

19. The system of claim 16, the reader device further comprising a display.

20. The system of claim 16, wherein the capture agent type of each spot comprises one or more antibodies, enzymes, proteins, nucleic acids, aptamers, peptides, or peptoids, 21. The system of claim 16, wherein the algorithm or machine learning comprises a first neural network that classifies the concentration(s) of the one or more analytes in the sample as either above or below a threshold value and one or more additional neural networks configured to output the concentration(s) of the one or more analytes in the sample that are above the threshold value.

22. The system of claim 15, wherein the one or more analytes comprise proteins, including cardiac troponin I (cTnI), 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