ABO blood group rapid on-site typing method and kit based on rpa-lfd technology and artificial intelligence aided interpretation

Abstract

This invention discloses a rapid on-site ABO blood typing method and kit based on RPA-LFD technology and artificial intelligence-assisted interpretation. The method includes: performing RPA amplification on the sample to be tested to obtain amplification products containing characteristic SNP sites of the ABO gene; introducing the amplification products into a CRISPR / Cas12a trans-cutting system to generate detectable report signals corresponding to each allele through single-base specific recognition; loading the sample onto a test strip, which is then separated by chromatography to form a test line; continuously acquiring images of the test area from the time of sample addition to obtain a dynamic image sequence; inputting the image sequence into a deep learning interpretation model to extract the spatiotemporal characteristics of the color development dynamics of each line and outputting the ABO blood typing results and confidence levels. By utilizing dynamic time-series analysis, the accuracy and reliability of weakly positive sample interpretation are improved, and the problem of subjective interpretation of the endpoint is solved.

Description

Technical Field

[0001] This invention relates to the field of DNA testing technology, and in particular to a rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation. Background Technology

[0002] The ABO blood group system is one of the most important blood group systems in humans, playing an irreplaceable role in forensic individual identification, paternity testing, and tracing evidence at crime scenes. Traditional ABO blood typing relies on serological methods, using antigen-antibody agglutination reactions for phenotypic identification. However, this method requires extremely high sample integrity and activity, which is often unsuitable for the trace amounts, degradation, or mixed bodily fluids commonly found at crime scenes. With the development of molecular biology techniques, genotyping methods based on characteristic single nucleotide polymorphisms (SNPs) in the ABO gene have gradually become an alternative. By detecting decisive sites such as c.261delG, c.803G>C, and c.526C>T, an individual's ABO blood type can be directly inferred, fundamentally overcoming the limitation of serological methods that rely on protein expression.

[0003] Current molecular typing methods are mostly based on polymerase chain reaction (PCR), combined with restriction fragment length polymorphism analysis, real-time quantitative PCR, or sequencing technology for SNP identification. However, PCR methods require sophisticated thermal cycling equipment and a stable laboratory environment, and the operation process is complex and time-consuming, making it unsuitable for the needs of immediate detection at crime scenes. To address this, isothermal amplification techniques, such as recombinase polymerase isothermal amplification (RPA), have been introduced. These techniques can achieve rapid nucleic acid amplification at around 39°C and, when combined with lateral flow (LFD) test strips, form a visual interpretation scheme, greatly reducing dependence on instrumentation. Such RPA-LFD schemes typically use specifically designed primers and probes targeting SNP sites to form visible band signals on the test strip detection line, achieving qualitative typing of ABO alleles.

[0004] However, conventional RPA-LFD has several shortcomings in on-site ABO blood typing. First, its interpretation method heavily relies on manual visual judgment of the endpoint color development results, only providing a qualitative conclusion of "wired / wireless". Due to the extremely low target concentration and severe background interference in forensic samples, the identification of weak positive test lines is easily affected by subjective factors, resulting in a high misjudgment rate and failing to provide reliable evidence for court acceptance. Second, the color development process of the test strip is itself a dynamic chromatographic reaction process. Different genotypes have subtle differences in the color development rate and saturation characteristics of the test lines, but existing methods completely discard this process information, only acquiring a single frame of the endpoint image, resulting in the waste of a large amount of valuable kinetic signals.

[0005] Furthermore, the matrix of bodily fluid samples from crime scenes is extremely complex. Bloodstains, saliva stains, and semen stains commonly contain strong amplification inhibitors such as heme and humic acid, and the samples themselves exhibit high background fluorescence and color interference. Current LFD protocols mostly use colloidal gold or common fluorescent dyes as markers, but these markers suffer a significant decrease in signal-to-noise ratio against complex matrix backgrounds, leading to impaired detection sensitivity. Although novel marker materials such as upconversion nanoparticles have been developed that can reduce background autofluorescence to some extent, an effective solution has yet to be found that integrates them with LFD detection and intelligent interpretation to systematically address matrix effects and ambient light interference.

[0006] Another significant technical challenge lies in the frequent encounter with mixed bodily fluid samples at forensic scenes, such as bloodstains from victims and suspects mixed together. Existing RPA-LFD methods can only output single blood type results, failing to identify whether mixed samples contain a second blood type, let alone estimate the mixing ratio. This can lead to the omission of crucial physical evidence information, severely limiting the application value of LFD methods in complex forensic scenarios.

[0007] In summary, existing technologies still lack a comprehensive technical solution that can address dynamic interpretation, background compensation, mixed sample analysis, and integrated reagent kit integration in order to achieve truly rapid and intelligent on-site ABO blood typing. Summary of the Invention

[0008] The purpose of this invention is to provide a rapid on-site ABO blood typing method and kit based on RPA-LFD technology and artificial intelligence-assisted interpretation. By continuously acquiring dynamic colorimetric image sequences of the detection area of ​​the lateral flow test strip at preset time intervals, and extracting the spatiotemporal characteristics of the colorimetric dynamics of each detection line by a deep learning interpretation model, the typing results and confidence levels are output. This improves the accuracy and reliability of interpretation of weakly positive samples and solves the problem that existing RPA-LFD methods rely on subjective interpretation of the endpoint and cannot utilize information from the colorimetric process.

[0009] To address the aforementioned technical problems, a first aspect of this invention provides a rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation, comprising the following steps: Step S1: Perform isothermal amplification of the sample to be tested using recombinase polymerase to obtain an amplification product containing the characteristic single nucleotide polymorphism site of the ABO blood group gene; Step S2: The amplification product is contacted with the CRISPR / Cas12a trans-cutting system to obtain a reaction mixture. The CRISPR / Cas12a trans-cutting system is used to perform single-base specific recognition of the characteristic single nucleotide polymorphism site and generate a detectable report signal corresponding to each allele. Step S3: The reaction mixture is loaded onto a lateral flow test strip, and after chromatographic separation, a colorimetric image sequence of detection lines indicating different alleles is formed in the detection area of ​​the lateral flow test strip; Step S4: During the preset color development acquisition period starting from the sampling time, continuously acquire the color development image of the detection line at preset time intervals to obtain a dynamic color development image sequence; Step S5: Input the dynamic color development image sequence into the pre-constructed deep learning interpretation model. The deep learning interpretation model extracts the spatiotemporal features of the color development dynamics of each detection line and outputs the ABO blood type classification result of the sample to be tested and the confidence level of the classification result based on the features.

[0010] Further, in step S3, the colorimetric image of the detection line is generated by a detection antibody labeled with upconversion nanoparticles, which emit dual-color visible light of 540nm and 660nm under 980nm near-infrared excitation light; In step S4, the continuous acquisition is multispectral image acquisition, and ambient light sensor data is recorded synchronously during the acquisition process.

[0011] Furthermore, the deep learning interpretation model includes a background compensation module; The background compensation module includes a U-Net-based background semantic segmentation sub-network, used to segment the matrix region and detection line region of the dynamic color image sequence; The background compensation module also combines a multispectral physical model and uses the segmentation results and ambient light sensor data to remove sample matrix absorption attenuation and ambient light reflection interference. The background area of ​​the lateral flow test strip is pre-printed with a rare earth fluorescent reference strip. The background compensation module performs radiometric calibration using the rare earth fluorescent reference strip and outputs the absolute fluorescence intensity ratio of each detection line. The spatiotemporal characteristics of the color development kinetics are constructed based on the absolute fluorescence intensity ratio.

[0012] Furthermore, the sample to be tested is a whole blood spot, a semen spot, or a saliva spot; The deep learning interpretation model eliminates sample matrix effects through the background compensation module, and the dynamic colorimetric image sequence is constructed based on the absolute fluorescence intensity ratio, so that the ABO blood typing results and confidence levels are consistent with the results obtained using purified DNA samples.

[0013] Furthermore, in step S5, the deep learning interpretation model includes a multi-component signal decoupling module based on a multi-head attention mechanism, which is used to perform deconvolution integral decomposition on the spatiotemporal features of the colorimetric dynamics of each detection line, output at least one primary blood type dynamic curve and one secondary blood type dynamic curve, and output the primary blood type classification result and the secondary blood type classification result, as well as the estimated ratio of primary and secondary blood type components, based on the primary blood type dynamic curve and the secondary blood type dynamic curve.

[0014] Furthermore, the input to the multi-component signal decoupling module is the gray value-time series of each detection line, which is then input to the multi-head attention layer after position encoding. The multi-head attention layer uses each detection line channel as an independent key-value pair and learns the contribution weight of different blood type components to the color development rate of each detection line through query vectors, thereby realizing the deconvolution integral solution.

[0015] Furthermore, the primary blood group kinetic curve and the secondary blood group kinetic curve correspond to the kinetic characteristics of blood group antigen-antibody reactions from different individuals, and the estimated ratio is calculated based on the ratio of the area under the curve of the primary blood group kinetic curve to that of the secondary blood group kinetic curve.

[0016] Furthermore, when the deconvolution integral solution can only separate one effective dynamic curve, the deep learning interpretation model determines that the sample to be tested is a single human source sample, outputs only a unique blood type classification result, and marks the mixing ratio as a single source.

[0017] Furthermore, in step S5, before performing the deconvolution integral solution, the multi-component signal decoupling module also performs sample environmental factor decoupling preprocessing on the spatiotemporal characteristics of the colorimetric dynamics of each detection line. The sample environmental factor decoupling preprocessing includes: inputting the chromogenic kinetic spatiotemporal features of each detection line into a pre-trained variational autoencoder to extract sample environmental latent variables independent of blood type components, wherein the sample environmental latent variables characterize the degradation degree and / or inhibitor residue level of the sample to be tested; and separating the sample environmental latent variables from the chromogenic kinetic spatiotemporal features to obtain environmentally corrected chromogenic kinetic spatiotemporal features of each detection line, which are used as inputs to the deconvolution integral solution.

[0018] Accordingly, a second aspect of the present invention provides a rapid on-site ABO blood typing kit based on RPA-LFD technology and artificial intelligence-assisted interpretation. The kit performs ABO blood typing processing based on the aforementioned rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation, including: The recombinase polymerase isothermal amplification system contains a specific primer set for amplifying characteristic single nucleotide polymorphism sites of the ABO blood group gene; The CRISPR / Cas12a trans-cleavage system comprises crRNA designed for the characteristic single nucleotide polymorphism site, Cas12a protein, and a reporter probe labeled with a detectable hapten. The CRISPR / Cas12a trans-cleavage system is used to activate the Cas12a trans-cleavage activity after the crRNA specifically binds to the amplification product, cleaving the reporter probe and releasing a detectable reporter signal. A lateral flow test strip, wherein a sample pad, a conjugate pad, a detection area and an absorption pad are sequentially arranged along the chromatography direction on the lateral flow test strip, and multiple detection lines and a control line are fixed in the detection area, the multiple detection lines being used to capture the detectable report signals corresponding to different alleles; An image acquisition device, comprising a multispectral CMOS camera and an ambient light sensor, is used to continuously acquire a dynamic color development image sequence of the detection area at preset time intervals within a preset color development acquisition period calculated from the time of sample addition. The image interpretation device has a pre-trained deep learning interpretation model built in. The deep learning interpretation model is used to extract the spatiotemporal features of the color development dynamics of each detection line in the dynamic color image sequence, and outputs the ABO blood type classification result of the test sample and the confidence level of the classification result based on the features.

[0019] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects: 1. By organically integrating upconversion nanoparticle labeling, multispectral dynamic image acquisition, and background compensation module in the deep learning interpretation model, near-infrared excitation is used at the physical layer to completely avoid the interference of autofluorescence in biological matrix. At the algorithm layer, semantic segmentation, multispectral physical modeling, and rare earth reference radiation calibration are used to transform the original colorimetric signals of each detection line into physically comparable absolute fluorescence intensity ratios. Based on this, the spatiotemporal characteristics of colorimetric dynamics are constructed, enabling ABO blood typing results and confidence levels in complex forensic sample matrices such as blood spots and saliva spots to achieve consistency with those using purified DNA samples. This fundamentally solves the inherent defects of existing lateral flow test strips, such as low signal-to-noise ratio and unreliable interpretation under strong background interference. 2. By using the multi-component signal decoupling module based on the multi-head attention mechanism in the deep learning interpretation model, the superimposed chromogenic kinetic curves of mixed samples on multiple detection lines are deconvoluted and integrally solved. The contribution weight of different blood type components to the chromogenic rate of each detection line is automatically learned, and the primary blood type kinetic curves and secondary blood type kinetic curves corresponding to different individual sources are separated. Based on the ratio of the area under the curve, the ratio of primary and secondary blood type components is output as an estimated value. Thus, a single lateral flow test strip can simultaneously complete the qualitative and quantitative analysis of ABO blood types from multiple sources in mixed spots, effectively making up for the technical gap of existing methods that can only output a single blood type conclusion and cannot identify blood type components from a second source when dealing with mixed body fluid samples. 3. Before performing deconvolution integral decomposition on mixed samples using a deep learning interpretation model, a variational autoencoder is used to decouple the spatiotemporal characteristics of the chromogenic dynamics of each detection line from the sample environmental factors. This extracts latent environmental variables that are independent of blood type components and only characterize the degree of sample degradation and the level of inhibitor residues. These latent variables are then extracted from the features and used as input for deconvolution integral decomposition. This achieves a two-stage analysis strategy of correction followed by decoupling. As a result, the estimation of the proportion of each blood type component in the mixed sample is no longer affected by the non-uniform distortion of the kinetic curve caused by uneven degradation or different levels of inhibitor interference from each source component. This significantly improves the quantitative accuracy of mixed spot typing under harsh forensic conditions. Attached Figure Description

[0020] Figure 1 This is a flowchart of the rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation provided in this embodiment of the invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0022] This invention establishes a complete detection process from sample nucleic acid amplification to intelligent interpretation, specifically designed for real-time ABO blood typing needs in non-laboratory environments such as crime scenes. After rapid processing, the sample is amplified at an isothermal temperature using recombinase polymerase to obtain target fragments containing characteristic single nucleotide polymorphisms (SNPs) of the ABO blood group gene. Subsequently, a CRISPR / Cas12a trans-cutting system is used to specifically identify these SNPs at the single-base level, converting the identification event into a detectable report signal corresponding to each allele. After the reaction mixture is transferred to a lateral flow test strip, chromatographic separation forms a colorimetric image of detection lines indicating different alleles in the detection area. Unlike traditional endpoint imaging methods, this invention continuously acquires images of the detection area at fixed time intervals within a preset time period from the moment of sample application, obtaining a dynamic image sequence recording the entire colorimetric process. The sequence is then analyzed by a pre-built deep learning interpretation model. The model extracts dynamic features containing temporal and spatial information from the color development process of each detection line, and finally outputs the ABO blood typing result of the sample to be tested and the confidence level of the result, providing a highly intelligent solution for rapid, objective and reliable on-site evidence screening.

[0023] Please refer to Figure 1 The first aspect of this invention provides a rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation, comprising the following steps: Step S1: The sample to be tested is subjected to recombinase polymerase isothermal amplification to obtain an amplification product containing the characteristic single nucleotide polymorphism site of the ABO blood group gene.

[0024] Recombinase polymerase amplification is performed on the test samples at an isothermal temperature to obtain amplification products containing characteristic SNP sites of the ABO blood group gene. ABO blood type is determined by specific SNPs on the ABO gene, such as c.261delG (common in the O allele), c.526C>T (related to the B allele), and c.803G>C (distinguishing site between A and B alleles). Recombinase polymerase amplification is an isothermal nucleic acid amplification technique that can be efficiently completed at around 39°C, without the need for a thermal cycler. In practice, a lyophilized reaction system containing recombinase, single-stranded binding protein, polymerase, and specific primers is pre-placed in a reaction tube. After adding the nucleic acid template and initiator from the test sample, the reaction is initiated by maintaining a constant temperature using a personal heating device or a simple isothermal module. The specific primers are designed to target both sides of the aforementioned SNP sites, with their 3′ ends extending to the mutant bases to initially distinguish alleles, ensuring that different alleles at each characteristic site can be amplified simultaneously. Because of its mild reaction conditions and short reaction time, RPA is very suitable for on-site operation and has good compatibility with subsequent CRISPR recognition and strip chromatography. The amplified products can be directly introduced to the next stage without purification.

[0025] Step S2: The amplification product is contacted with the CRISPR / Cas12a trans-cutting system to obtain a reaction mixture. The CRISPR / Cas12a trans-cutting system is used to perform single-base specific recognition of characteristic single nucleotide polymorphism sites and generate detectable report signals corresponding to each allele.

[0026] The amplification product obtained in step S1 is contacted with a CRISPR / Cas12a trans-cleavage system to obtain a reaction mixture. The core components of this system include the Cas12a protein and crRNAs specifically designed for each characteristic SNP site of the ABO genome. The spacer sequence of the crRNA is perfectly complementary to the specific amplicon sequence of the allele to be distinguished. The Cas12a protein is only activated and exhibits non-specific single-stranded DNA trans-cleavage activity when the amplicon and crRNA are precisely paired. The system also contains reporter probes linked to detectable haptens, such as short single-stranded DNA molecules labeled with a fluorescent group FAM at one end and biotin at the other. Upon activation, Cas12a efficiently cleaves these reporter probes, causing FAM and biotin to separate, releasing a large number of fragments carrying only a single label group. Since each crRNA targets a specific allele, the recognition of different alleles can activate the corresponding reporter probes to generate signals, achieving a one-to-one correspondence between allele information and detectable reporter signals. This signal generation mechanism based on Cas12a trans-cleavage is equivalent to adding an independent molecular proofreading step after RPA amplification, which greatly improves the specificity of SNP typing.

[0027] In step S2, the specific construction method of the CRISPR / Cas12a trans-cutting system directly determines the correspondence between the specificity of allele recognition and the signal output. For the characteristic SNP sites on which ABO blood typing depends, different recognition-specific crRNAs are configured in the system. For example, for the c.261delG deletion mutation specific to the O allele, a crRNA is designed whose spacer sequence is completely complementary to the amplicon sequence containing this deletion site; for the c.803G>C site that distinguishes between the A and B alleles, crRNAs matching the A-type or B-type specific amplicon are designed respectively, so that the corresponding crRNA can mediate the specific activation of the Cas12a protein only when the corresponding allele is present in the sample being tested. Regarding the reporter probe labeling scheme, different crRNAs employ differentiated hapten labeling combinations for their corresponding reporter probes. For example, the reporter probe associated with A allele detection is labeled with the fluorescent group FAM at one end and biotin at the other; the reporter probe associated with B allele detection is labeled with DIG at one end and biotin at the other; and the reporter probe associated with O allele detection is labeled with TAMRA at one end and biotin at the other. After trans-cleavage by Cas12a, the released fragments carrying single labeling groups such as FAM, DIG, or TAMRA migrate to their respective detection lines on the lateral flow test strip. Pre-fixed capture molecules in the detection area have a clear correspondence with the aforementioned labeling groups: the anti-FAM antibody capture line corresponds to the A allele signal, the anti-DIG antibody capture line corresponds to the B allele signal, and the anti-TAMRA antibody capture line corresponds to the O allele signal, thus achieving parallel differentiation of each allele on the test strip through a spatially separated colorimetric band pattern.

[0028] Step S3: Load the reaction mixture onto a lateral flow test strip. After chromatographic separation, a colorimetric image sequence of detection lines indicating different alleles is formed in the detection area of ​​the lateral flow test strip.

[0029] Lateral flow test strips typically consist of a sample pad, a conjugate pad, a chromatography membrane, and an absorbent pad stacked sequentially. Cleaved reporter probe fragments and uncleaved intact probes in the reaction mixture migrate forward under capillary force. In the detection area, multiple test lines are pre-immobilized with specific capture molecules for different haptens; for example, one test line may be coated with anti-FAM antibody, another with anti-DIG antibody, etc., each test line corresponding to a reporter signal of one allele. Cleavage fragments carrying free haptens are trapped by their respective capture lines and react with a universally labeled gold or fluorescent conjugate to develop color, forming visible or instrument-readable bands. Uncleaved intact probes cannot be trapped by the capture lines, while control lines indicate whether the chromatography process was completed correctly. Thus, the presence or absence of different alleles is converted into color signals on the test lines, presented spatially separated on the same test strip, forming a clearly corresponding colorimetric image of the test lines.

[0030] Step S4: During the preset color development acquisition period starting from the sampling time, continuously acquire the color development images of the detection line at preset time intervals to obtain a dynamic color development image sequence.

[0031] The acquisition program is initiated the instant the sample is added to the test strip to begin chromatography. Throughout the preset colorimetric acquisition period, such as 0 to 15 minutes, images of the detection area are continuously captured at preset time intervals of 2 frames per second or higher, forming a complete dynamic colorimetric image sequence. This sequence fully records the entire colorimetric dynamics of each detection line, from its gradual appearance at the background level to its gradual increase in intensity and eventual stabilization. Compared to traditional methods that only acquire a single image at the reaction endpoint, the dynamic sequence retains rich temporal information, providing a more comprehensive signal basis for subsequent analysis. The acquisition process can be completed under constant illumination using a portable reader's built-in CMOS camera module to avoid interference from changes in external light.

[0032] Step S5: Input the dynamic color development image sequence into the pre-built deep learning interpretation model. The deep learning interpretation model extracts the spatiotemporal features of the color development dynamics of each detection line and outputs the ABO blood type classification result of the sample to be tested and the confidence level of the classification result based on the features.

[0033] The dynamic colorimetric image sequence obtained in step S4 is input into a pre-trained deep learning interpretation model. This model can simultaneously extract the spatial and temporal evolution features of the colorimetric behavior of each detection line from the image sequence, i.e., the spatiotemporal features of colorimetric dynamics. For example, the model can capture the change pattern of gray values ​​over time within the detection line region through a three-dimensional convolutional neural network, and learn subtle signal differences such as the relative colorimetric rate, saturation plateau height, and time lag relationship between each detection line by combining attention mechanisms and other structures. Based on the extracted spatiotemporal dynamic features, the model not only outputs the ABO blood type classification result of the sample to be tested, such as type A, type B, type AB, or type O, but also provides the confidence level of the classification result, for example, representing the reliability of the judgment in the form of a probability value. This confidence level can provide objective decision-making basis for on-site operators and subsequent judicial review, significantly reducing the risk of subjective misjudgment caused by weak positive signals, inconsistent operation time windows, and other factors during manual visual interpretation.

[0034] This deep learning interpretation model is specifically designed for the intelligent interpretation of dynamic colorimetric image sequences from lateral flow test strips. Its training process requires the construction of a large-scale labeled dataset. This dataset encompasses dynamic colorimetric image sequences generated by standard samples of various known ABO blood type genotypes under different environmental conditions and concentration gradients. Each sequence carries a real ABO blood type label and signal quality annotations such as weak positive / strong positive. The model's network structure employs an architecture that integrates a 3D convolutional neural network and a Transformer encoder. The 3D convolutional neural network is responsible for synchronously capturing short-range spatiotemporal change patterns of each detection line region from multiple consecutive frames of images, extracting low-level dynamic features such as grayscale growth rate and local texture changes. The Transformer encoder takes the feature sequences of different detection lines as input and uses a self-attention mechanism to globally model long-range dependencies such as the relative colorimetric rate differences between detection lines, the timing of saturation plateau arrival, and the interrelationships of signal intensity between lines. The model input is a fixed-frame sequence of dynamic color images, arranged chronologically and accompanied by timestamps after sampling. The model output layer contains a genotyping header and a confidence assessment header, which output the ABO blood type classification result and the classification confidence score expressed as a probability value, respectively. During training, a strategy of jointly optimizing cross-entropy loss and confidence calibration loss is employed to ensure that while providing correct genotyping, the model's confidence output accurately reflects the reliability of the judgment.

[0035] Chromogenic dynamics spatiotemporal features refer to comprehensive information extracted from dynamic chromogenic image sequences that characterizes the evolution of the chromogenic signal of each detection line over time and its relative behavior pattern in the spatial layout of the test strip. The temporal dimension features are manifested as a complete curve showing the gradual increase in the mean gray value within each detection line region from the background baseline to the saturation plateau. This curve includes quantifiable parameters such as the initial chromogenic time, the chromogenic rise slope, the duration of the full width at half maximum (FWHM), and the gray value of the saturation plateau. The spatial dimension features are manifested as the gray-level difference pattern between detection lines at the same time cross-section, and the temporal order in which each detection line reaches the same gray-level threshold. For example, the chromogenic development of a detection line corresponding to a certain allele may lag behind that of another allele by several seconds, or the gray value of the saturation plateau of a certain detection line may be significantly lower than that of other detection lines. The extraction of these spatiotemporal features can be accomplished by synchronously sliding a three-dimensional convolutional kernel along the time and spatial axes in a deep learning model, or by combining an attention mechanism to perform weighted analysis of the gray-time series of each detection line, enabling the model to focus on key time windows and detection line combinations that have discriminative power for genotyping. Compared to traditional endpoint interpretation which relies solely on whether each detection line reaches a visually visible threshold at a single moment, chromogenic kinetic spatiotemporal features can capture subtle kinetic differences caused by variations in allele amplification efficiency, Cas12a cleavage activity, and antigen-antibody binding affinity, thus providing richer and more reliable genotyping evidence for deep learning interpretation models.

[0036] The above technical solution integrates isothermal amplification, precise CRISPR identification, spatial coding of lateral flow chromatography, and intelligent analysis of dynamic time-series images, breaking through the limitations of existing RPA-LFD methods that rely solely on static endpoint interpretation. By mining dynamic information in the color development process and assigning confidence measures to the results, the on-site rapid typing of ABO blood types has been fundamentally improved in terms of accuracy, objectivity, and legal evidentiary value.

[0037] Furthermore, in step S3, the colorimetric image of the detection line is generated by a detection antibody labeled with upconversion nanoparticles, which emit dual-color visible light of 540nm and 660nm under 980nm near-infrared excitation light.

[0038] The detection antibody labeled with upconversion nanoparticles is used as the signal generation unit of the lateral flow test strip. The upconversion nanoparticles emit dual-color visible light at 540nm and 660nm under 980nm near-infrared excitation light. This labeling method completely eliminates the absorption and autofluorescence bands of common fluorescent interfering substances in biological samples (such as heme, proteins, and organic residues in forensic sample matrices), thus virtually eliminating the interference of background fluorescence from complex samples at the physical level. The resulting colorimetric image of the detection line has a high signal-to-noise ratio, providing a clean signal foundation for the multispectral dynamic image acquisition in step S4. This allows the subsequent deep learning interpretation model to directly extract and analyze the spatiotemporal features of colorimetric dynamics based on the clear, high-contrast dynamic colorimetric sequence without additional processing of background noise caused by matrix autofluorescence, significantly ensuring the accuracy and stability of ABO blood typing under complex on-site sample conditions.

[0039] In the specific implementation of antibody labeling using upconversion nanoparticles, the upconversion nanoparticles typically achieve directional coupling with the detection antibody through the specific binding of streptavidin-biotin to their surface. This involves incubating the biotin-modified detection antibody and the streptavidin-coated upconversion nanoparticles in a buffer system to form a stable non-covalently coupled complex. Alternatively, a glutaraldehyde cross-linking method can be used to covalently link the amino groups of the antibody to the amino groups modified on the particle surface. The selected upconversion nanoparticles are generally controlled within the range of 20 to 80 nanometers. This size ensures smooth migration within the pores of the lateral flow test strip's chromatographic membrane while providing sufficient specific surface area to accommodate a high density of detection antibodies, thus guaranteeing detection sensitivity. The target identified by the detection antibody is the free hapten group released after the reporter probe is trans-cleaved by Cas12a in step S2. For example, it is a monoclonal antibody that recognizes different haptens such as FAM, DIG or TAMRA. The detection antibody specifically binds to its respective target hapten, so that the cleavage fragment carrying the specific hapten is efficiently captured and enriched when it flows through the detection line, and then a highly sensitive visualization report is achieved by the upconversion luminescence signal under near-infrared excitation.

[0040] The 980nm near-infrared excitation light is located outside the absorption spectrum of common interfering substances in biological samples (such as heme, humic acid, and residual proteins in forensic body fluid matrix). When the excitation light penetrates the nitrocellulose chromatography membrane, it produces almost no matrix autofluorescence, reducing background noise by more than an order of magnitude compared to traditional visible light excitation (such as 488nm or 545nm). Regarding the emission channel, the 540nm emission peak is located in the green visible light range, where both the human eye and CMOS sensors have high sensitivity, which is beneficial for achieving high-gain signal acquisition in portable readers. The 660nm emission peak is in the red-to-near-infrared transition region. In this band, the brownish-red absorption interference of residual blood in forensic samples is significantly weaker than in the shorter wavelength region, resulting in higher optical transmittance of the signal in the chromatography membrane. This can serve as an auxiliary signal channel to provide additional signal-to-noise ratio redundancy. In comparative experiments, organic fluorescent dye-labeled systems with 980nm excitation and 540nm single-channel acquisition, and conventional 545nm excitation and 570nm emission, were subjected to lateral flow chromatography under the same spiked whole blood lysis buffer matrix. The results showed that the background fluorescence intensity of the 540nm channel was reduced by more than 85% in the 980nm excitation system, and the ratio of the detection line signal to the background was several times that of the conventional system. Simultaneously, when the 660nm channel experienced local attenuation due to matrix inhomogeneity in the 540nm channel, the signal ratio change was less than 10%, indicating that the ratio between the two emission channels can compensate for local optical differences in the matrix. This provides a stable signal source for multispectral image acquisition in step S4 and for the extraction of chromogenic dynamics features based on the absolute fluorescence intensity ratio in step S5. The specific wavelength combination described above is not a simple replacement of common knowledge in the field, but rather an optimized parameter scheme established through systematic screening and experimental verification to meet the chromatographic detection needs of complex samples from forensic scenes.

[0041] Accordingly, in step S4, continuous acquisition is multispectral image acquisition, with ambient light sensor data recorded synchronously during the acquisition process. Multispectral image acquisition obtains independent grayscale image frames for each detection line from both the 540nm and 660nm emission channels, thus fully preserving the different evolution information of the colorimetric signals at the two wavelengths, providing multidimensional raw data for subsequent physical model-driven background stripping. The synchronously recorded ambient light sensor data is used to monitor real-time illumination changes at the acquisition site. In the subsequent background compensation module, it is combined with the multispectral physical model to perform frame-by-frame correction of ambient light reflection interference, ensuring that in uncontrolled lighting environments such as crime scenes lacking darkroom conditions, each frame in the dynamic colorimetric image sequence can accurately reflect the actual fluorescence signal intensity of each detection line, unaffected by ambient light fluctuations.

[0042] Furthermore, the deep learning interpretation model includes a background compensation module. This module comprises a U-Net-based background semantic segmentation subnetwork for segmenting the matrix region and detection line region of the dynamic colorimetric image sequence. The background compensation module also incorporates a multispectral physical model, utilizing the segmentation results and ambient light sensor data to remove matrix absorption attenuation and ambient light reflection interference. The background area of ​​the lateral flow test strip is pre-printed with rare-earth fluorescent reference bands. The background compensation module uses these rare-earth fluorescent reference bands for radiometric calibration, outputting the absolute fluorescence intensity ratio of each detection line. The spatiotemporal characteristics of the colorimetric dynamics are constructed based on these absolute fluorescence intensity ratios.

[0043] Before the dynamic colorimetric image sequence enters the genotyping stage, the background compensation module integrated in the deep learning interpretation model plays a crucial role in physically correcting the original image signal. This module first calls a background semantic segmentation sub-network based on the U-Net architecture to perform pixel-level classification of each frame in the dynamic colorimetric image sequence, dividing the image into two main categories: matrix regions and detection line regions. Matrix regions include non-specific staining areas formed on the nitrocellulose membrane, such as blood spot residue, salivary protein deposition, or semen stain background. Detection line regions are the specific spatial locations of the capture lines and control lines corresponding to each allele. After being trained on a large number of labeled images with complex matrix interference, the semantic segmentation sub-network can accurately identify the spatial boundaries of these two types of regions under different matrix types and levels of contamination. Even when the detection line signal is as weak as background noise, it can still distinguish it from the surrounding matrix, preventing weak positive signals from being misjudged as background and removed during subsequent background stripping.

[0044] After image region segmentation, the background compensation module inputs the segmentation results along with the ambient light sensor data recorded in step S4 into the multispectral physical model. This multispectral physical model is based on the radiative transfer theory of light propagation in multilayer media, and can quantitatively describe the reflection component of ambient light on the test strip surface, the absorption and attenuation of excitation and emission light by the matrix layer, and the scattering characteristics of the chromatography film itself. Using the matrix region provided by semantic segmentation as a sampling window, the model calculates the absorption and scattering coefficients of the matrix in each spectral channel, thereby inferring the degree of attenuation of the fluorescence signal in the detection line region by the matrix. Simultaneously, combined with the ambient light intensity and color temperature data recorded by the ambient light sensor, the additive interference component introduced by ambient light reflection is extracted from the total signal of the detection line region. After the above frame-by-frame correction, the true fluorescence emission intensity of each detection line region at each time point is restored, no longer containing artifacts introduced by differences in sample matrix and fluctuations in ambient light.

[0045] To ensure lateral comparability of calibrated signals across different test strips and batches, the background compensation module further utilizes a rare-earth fluorescent reference band pre-printed in the background area of ​​the lateral flow test strip for radiometric calibration. This rare-earth fluorescent reference band contains rare-earth complexes with known fluorescence quantum yields, exhibiting stable emission intensity under the same near-infrared excitation conditions and unaffected by the sample matrix. Using the fluorescence intensity of this reference band as an internal standard, the module calculates the ratio of the calibrated fluorescence intensity of each detection line to this internal standard, outputting a dimensionless absolute fluorescence intensity ratio. Subsequently, the deep learning interpretation model uses this absolute fluorescence intensity ratio as a basis to construct the gray-time series of each detection line throughout the entire color development process, serving as the spatiotemporal characteristics of the color development dynamics. Since this feature has been double-corrected by background stripping and radiometric calibration, the differences in the matrix of the field samples, the interference of ambient light, and the optical response deviations between different readers have all been effectively eliminated. This allows the model to establish a unified and physically comparable signal benchmark for subsequent analysis of color development rate, saturation plateau, and relative behavior between detection lines. This fundamentally ensures the consistency and reliability of typing determination in complex forensic samples such as blood spots and saliva spots.

[0046] Specifically, the overall architecture of the aforementioned background compensation module consists of three functional units connected in series: a semantic segmentation sub-network, a multispectral physical model, and a radiometric calibration unit, forming a complete data flow path from image preprocessing to signal normalization. The dynamic colorimetric image sequence first enters the semantic segmentation sub-network, outputting a frame-by-frame binarized segmentation mask for the matrix region and the detection line region. The segmentation mask, along with synchronously recorded ambient light sensor data, is input into the multispectral physical model. After the model strips away matrix absorption attenuation and ambient light reflection interference frame by frame, it outputs the corrected fluorescence intensity of each detection line region. The corrected fluorescence intensity then enters the radiometric calibration unit, where a ratio calculation is performed based on the signal of the rare-earth fluorescence reference band within the same frame. Finally, the absolute fluorescence intensity ratio of each detection line is output, which is the direct input for constructing the spatiotemporal characteristics of colorimetric dynamics.

[0047] The mathematical foundation of the multispectral physical model is built upon a framework combining the Beer-Lambert law and the Kubelka-Munke light scattering theory. The Beer-Lambert law describes the exponential decay of excitation light as it passes through the matrix layer due to light-absorbing substances such as hemoglobin and melanin, as well as the attenuation effect of emitted fluorescence being reabsorbed by the matrix on its return journey to the acquisition lens. The Kubelka-Munke theory describes the scattering and diffuse reflection behavior of the nitrocellulose chromatographic membrane and the dried matrix layer, establishing a quantitative relationship between reflectivity, absorption coefficient, and scattering coefficient. In the actual calibration process, reflectivity measurements of the matrix region under two emission channels (540 nm and 660 nm) are selected, and the absorption and scattering coefficients of the matrix at the two wavelengths are solved using the Kubelka-Munke equation. Since hemoglobin has a strong absorption peak at 540 nm but significantly reduced absorption at 660 nm, and melanin shows relatively gradual absorption changes at both wavelengths, by combining the differences in absorption coefficients at the two wavelengths, the interference contributions of hemoglobin and melanin can be separated and quantitatively subtracted, thereby accurately restoring the true fluorescence intensity of the detection line area.

[0048] The rare-earth fluorescent reference band pre-printed in the background area of ​​the lateral flow test strip uses europium or terbium complexes as fluorescent materials. These rare-earth complexes are characterized by broad excitation spectra, narrow emission peaks, and long fluorescence lifetimes. The main emission peak of the selected europium complex is located near 615 nm, and that of the terbium complex is located near 545 nm. Their emission wavelengths are effectively distinguishable from the 540 nm and 660 nm bicolor emission peaks of the upconversion nanoparticles through a bandpass filter, and they do not overlap. During the test strip preparation process, a mixed solution of the rare-earth complex and the film-forming resin is sprayed onto the background area of ​​the chromatography membrane using a precision spotting device. This location is outside the sample migration path and maintains a fixed spatial distance from the detection line area. After low-temperature drying, a stable solid fluorescent reference band is formed. Since rare earth complexes hardly produce fluorescence under 980nm near-infrared excitation, the upconversion luminescence signal excited by 980nm is used as the detection channel, while the fluorescence emission of the rare earth reference strip is obtained under separate ultraviolet or short-wavelength visible light excitation. The excitation and acquisition sequences of the two can be alternated to avoid crosstalk.

[0049] The absolute fluorescence intensity ratio is defined as the ratio of the fluorescence intensity of each detection line region after background compensation correction to the fluorescence intensity of the rare-earth fluorescence reference strip in an image acquired at the same time point. The mechanism by which radiometric calibration imparts physical comparability to this ratio lies in the fact that the fluorescence quantum yield of the rare-earth fluorescence reference strip is known and stable. Under different test strip batches, different ambient temperatures, and different reader optical responses, the variation in its emission intensity only originates from common-mode factors such as excitation source power fluctuations and detection system gain drift. These common-mode factors also affect the fluorescence intensity measurement of each detection line. By calculating the ratio between the detection line signal and the reference strip signal, common-mode fluctuations are automatically canceled out. The resulting absolute fluorescence intensity ratio no longer depends on specific detection equipment and operating conditions, thus enabling direct lateral comparison between samples tested from different test strips and at different times. It also provides a standardized input quantity for deep learning interpretation models that is independent of the detection platform.

[0050] Furthermore, the samples to be tested were whole blood spots, semen spots, or saliva spots. The deep learning interpretation model eliminated the sample matrix effect through a background compensation module, and the dynamic colorimetric image sequence was constructed based on the absolute fluorescence intensity ratio, making the ABO blood typing results and confidence levels consistent with those obtained using purified DNA samples.

[0051] When the sample to be tested is a complex bodily fluid specimen commonly found at forensic scenes, such as whole blood spots, semen spots, or saliva spots, the heme, protein, humic acid, and other organic residues contained in the sample matrix will cause significant absorption attenuation and non-specific fluorescence interference to the detection signal. At the same time, the differences in the physicochemical properties of different specimen matrices will also lead to deviations in the colorimetric kinetics. In this case, the background compensation module in the deep learning interpretation model initiates a complete correction process for each frame of dynamic colorimetric image: the semantic segmentation subnetwork first identifies and divides the matrix region and the detection line region. The multispectral physical model calculates the absorption coefficient and scattering coefficient of the matrix in the two emission channels of 540nm and 660nm based on the segmentation results and ambient light sensor data. It then removes the attenuation of excitation light and emission fluorescence caused by the matrix and the additive interference introduced by ambient light reflection frame by frame, restoring the true fluorescence intensity of each detection line. Subsequently, radiometric calibration is performed using rare earth fluorescence reference strips as internal standards, converting the corrected fluorescence intensity into an absolute fluorescence intensity ratio that is independent of the detection equipment and operating conditions. The spatiotemporal characteristics of chromogenic kinetics constructed based on this absolute fluorescence intensity ratio no longer include signal distortions introduced by sample matrix type, concentration, and batch-to-batch differences. Therefore, the ABO blood typing results and confidence levels output by the deep learning interpretation model for whole blood spots, semen spots, or saliva spots can reach the same level of consistency as the results obtained using purified DNA samples under the same detection conditions.

[0052] Furthermore, in step S5, the deep learning interpretation model includes a multi-component signal decoupling module based on a multi-head attention mechanism, which is used to perform deconvolution integral decomposition on the spatiotemporal characteristics of the colorimetric dynamics of each detection line, outputting at least one primary blood type dynamic curve and one secondary blood type dynamic curve, and based on the primary blood type dynamic curve and the secondary blood type dynamic curve, outputting the primary blood type classification result and the secondary blood type classification result, as well as the estimated ratio of the primary and secondary blood type components, respectively.

[0053] When the sample to be tested is a mixture of bodily fluids from multiple individuals, the chromogenic kinetic curves presented on each test line are actually the combined result of the superimposed reactions of blood type components from different individuals under the same chromatographic conditions. The differences in antigen-antibody binding rates, Cas12a cleavage efficiency, and chromatographic migration characteristics of each component will leave specific kinetic traces in the superimposed curves. However, these superimposed information cannot be separated by the human eye or traditional endpoint interpretation methods alone. The multi-component signal decoupling module integrated in the deep learning interpretation model is designed to solve this problem. This module takes the absolute fluorescence intensity ratio-time series obtained after background compensation and radiometric calibration of each test line as input, and automatically learns the independent contribution weights of different blood type components to the chromogenic rate of each test line through a multi-head attention mechanism. In its implementation, the multi-head attention layer treats each detection line channel as an independent key-value pair. Multiple attention heads analyze in parallel the relative color development delay, saturation rate differences, and cross-reaction patterns between each detection line from different time scales and feature subspaces. After deconvolution integral decomposition, at least one primary blood type kinetic curve corresponding to the primary blood type source in the sample and one secondary blood type kinetic curve corresponding to the secondary blood type source are reconstructed from the original superimposed kinetic curves. Based on the separated primary and secondary blood type kinetic curves, the model can independently determine the ABO typing results of the primary and secondary blood types, and by comparing the ratio of the areas under the curves of the two curves within the same time window, it provides an estimated value of the proportion of primary and secondary blood type components in the mixed sample. This enables qualitative and quantitative analysis of multiple blood types in mixed body fluid samples using a single test strip.

[0054] The network structure of the multi-component signal decoupling module consists of a position encoding layer, a multi-head attention layer, and a decoupling output layer connected in series. The position encoding layer first performs sinusoidal position encoding on the absolute fluorescence intensity ratio-time series of each detection line in the time dimension. It concatenates the absolute fluorescence intensity ratio of each time step with an encoding vector representing the position of that time step within the entire color development sequence, forming an embedded representation of the detection line channel carrying temporal position information. This embedded representation is then input to the multi-head attention layer, which is configured with several parallel attention heads. Each attention head maps each detection line channel to a query vector, a key vector, and a value vector, respectively. Each detection line channel is treated as an independent key-value pair, and the query vector is used to learn the color development rate dependence between the current detection line and other detection lines. Multiple attention heads compute the attention weight distribution among the detection line channels in parallel in different representation subspaces. The outputs of each head are then concatenated and fused via linear projection to obtain the context-enhanced features of each detection line channel after global interaction. The decoupled output layer comprises principal component output branches and secondary component output branches. Each branch is composed of stacked fully connected layers, taking the output of the multi-head attention layer as input and outputting discrete-time sampling sequence of the primary and secondary blood type dynamics curves after nonlinear transformation. Regarding the training strategy, the training data for this module consists of dynamic colorimetric image sequences collected through a complete detection process using mixed samples with known primary and secondary blood type categories and known mixing ratios. Each training sample carries triple supervision signals: primary blood type category label, secondary blood type category label, and primary / secondary component ratio label. During training, the total loss function consists of a weighted sum of three parts: the first part is the curve reconstruction loss, which calculates the mean square error between the sum of the primary and secondary blood type dynamics curves output by the model and the original input superimposed curve, ensuring that the decomposition result can numerically reconstruct the original signal; the second part is the genotyping loss, which calculates the cross-entropy loss between the principal component output branch and the true primary and secondary blood type categories for the secondary component output branch; the third part is the proportional regression loss, which calculates the mean square error between the estimated primary and secondary component ratios output by the model and the true mixing ratio. Through joint training with the aforementioned triple supervision signals, the model can simultaneously learn accurate multi-component dynamic curve separation capabilities and reliable blood type classification capabilities.

[0055] Furthermore, the input to the multi-component signal decoupling module is the grayscale value-time series of each detection line, which is then input to the multi-head attention layer after position encoding. The multi-head attention layer uses each detection line channel as an independent key-value pair and learns the contribution weight of different blood type components to the color development rate of each detection line through query vectors, thereby achieving deconvolution integral solution.

[0056] The multi-component signal decoupling module receives the absolute fluorescence intensity ratio-time series of each detection line after background compensation and radiometric calibration as input data. Before entering the attention calculation, this input series is first assigned a unique position encoding vector to each sampling moment in the time series through a position encoding layer. This position encoding vector is then concatenated or added to the absolute fluorescence intensity ratio of each detection line at the corresponding moment, thereby embedding temporal structure information into the input representation, enabling the model to perceive the sequential order of each time point during the color development process. Subsequently, the position-enhanced feature vector is fed into the multi-head attention layer. The multi-head attention layer treats each detection line channel as an independent key-value pair, meaning that the overall color development dynamic behavior of each detection line serves as the source of both the key and value vectors. The query vector is generated by aggregating the global context of all detection line channels. When calculating the attention weights, the query vector is multiplied by the key vector of each detection line channel, and after normalization, the attention weight distribution of each detection line channel in the current context is obtained. This weight distribution essentially quantifies the strength of the contribution of different blood type components to the color development rate of each detection line. After the value vectors of each detection line are weighted and summed according to their corresponding attention weights, the color development patterns driven by different blood type components are naturally separated into different feature subspaces, thereby realizing the deconvolution integral solution from the superimposed original dynamic signals to obtain the independent dynamic curves corresponding to the primary blood type component and the secondary blood type component respectively.

[0057] Furthermore, the primary blood group kinetic curve and the secondary blood group kinetic curve correspond to the kinetic characteristics of blood group antigen-antibody reactions from different individuals, and the ratio estimate is calculated based on the ratio of the area under the curve of the primary blood group kinetic curve to that of the secondary blood group kinetic curve.

[0058] The primary and secondary blood group kinetic curves, obtained by deconvolution integration using a multi-component signal decoupling module, reflect the independent binding-dissociation dynamics between blood group antigens from two different individuals and their corresponding capture antibodies in a mixed sample. Due to variations in erythrocyte surface antigen density, secretory versus non-secretory morphology, and the initial concentration of target nucleic acids in the samples, the kinetic parameters of primary and secondary blood group components on multiple detection lines, such as color development rate, saturation plateau height, and peak time, exhibit discernible individual specificity. The separated primary blood group kinetic curve fully preserves the aforementioned reaction characteristics of the dominant blood group component in the sample, while the secondary blood group kinetic curve clearly records the corresponding characteristics of the less dominant blood group component. The independent evolution trajectories of the two curves on the time axis constitute the basis for the separate qualitative identification of blood types from multiple sources. At the quantitative level, the estimated ratio of primary and secondary blood type components is obtained by calculating the ratio of the area under the curve of the primary blood type kinetic curve and the area under the curve of the secondary blood type kinetic curve within the same preset time window. This area ratio directly reflects the cumulative contribution of the primary and secondary blood type components to the total signal intensity in the mixed sample and is linearly related to the relative concentration of the two components in the original mixed body fluid. This provides a simple method for the quantitative analysis of multiple blood types in mixed field samples that can be directly estimated without the need for an external standard curve.

[0059] Furthermore, when the deconvolution integral solution can only separate one effective dynamic curve, the deep learning interpretation model determines that the sample to be tested is a single human source sample, outputs only a unique blood type classification result, and marks the mixing ratio as a single source.

[0060] When the multi-component signal decoupling module in the deep learning interpretation model performs deconvolution integral decomposition on the spatiotemporal characteristics of the chromogenic dynamics of the test sample, if only one effective curve with a complete dynamic morphology can be separated from the superimposed signals of each detection line, while the other channels only exhibit random fluctuations at the baseline level, do not show identifiable chromogenic start times or saturation plateaus, etc., the model automatically determines that the test sample contains only blood type components from a single individual. At this point, the deep learning interpretation model no longer executes secondary blood typing logic, but directly outputs the corresponding ABO blood typing result based on this single effective dynamic curve, and marks the mixing ratio as a single source in the proportion labeling field. This ensures that a single-source sample is not misclassified as a mixed sample, while maintaining the consistency of the output format and the traceability of the interpretation results.

[0061] Furthermore, in step S5, before performing the deconvolution integral solution, the multi-component signal decoupling module performs sample environmental factor decoupling preprocessing on the chromogenic spatiotemporal features of each detection line. This sample environmental factor decoupling preprocessing includes: inputting the chromogenic spatiotemporal features of each detection line into a pre-trained variational autoencoder to extract latent variables of the sample environment independent of blood type components. These latent variables characterize the degree of degradation and / or inhibitor residue level of the sample under test. The latent variables of the sample environment are then separated from the chromogenic spatiotemporal features to obtain the environmentally corrected chromogenic spatiotemporal features of each detection line, which are used as input for the deconvolution integral solution.

[0062] When the test sample contains varying degrees of nucleic acid degradation or residual amplification inhibitors such as heme and humic acid, the chromogenic kinetic curves of each blood type component, in addition to reflecting the differences in the alleles themselves, will also be superimposed with non-specific distortions such as delayed color development and reduced saturation plateau caused by decreased template integrity or inhibited enzyme activity. In mixed samples, the degree of influence of these distortions on components from different sources is often inconsistent, and direct deconvolution integration may lead to systematic bias in the proportion estimation. To address this, the multi-component signal decoupling module first performs sample environmental factor decoupling preprocessing before entering the deconvolution integration. This preprocessing step calls a pre-trained variational autoencoder to map the spatiotemporal characteristics of the chromogenic kinetics of each detection line to the latent variable space, and forcibly separates two independent information dimensions in the latent variables: one is blood type component-related latent variables, encoding specific color development patterns driven by different alleles; the other is sample environmental latent variables, specifically capturing global kinetic distortion features caused by the degree of nucleic acid degradation and the level of inhibitor residues. During the training phase, the variational autoencoder uses a mixed sample of known blood type component proportions collected under different degradation conditions and inhibitor concentrations as the training set. By introducing mutual information minimization constraints or adversarial training strategies, it forces statistical independence between the latent variables of the sample environment and the blood type component labels, ensuring that the latent variables only encode environmental factors related to sample quality. After latent variable extraction, the latent variables of the sample environment are fixed to a preset standard environmental vector, and decoding and reconstruction are performed only using latent variables related to blood type components. The resulting reconstructed sequence is the spatiotemporal feature of the chromogenic dynamics of each detection line after environmental correction, stripped of degradation and inhibition interference. This corrected feature is then used as the formal input to the multi-head attention layer in the multi-component signal decoupling module, and deconvolve integral decomposition is performed under conditions unaffected by sample environmental factors, thus ensuring the accuracy of primary and secondary blood type classification and proportion estimation even in harsh forensic sample environments.

[0063] The core of the training scheme for decoupling preprocessing of sample environmental factors lies in ensuring that the latent variables of the sample environment extracted by the variational autoencoder only encode environmental factors related to sample quality, without containing any blood type component information. The construction of the training dataset is a crucial step: first, mixed sample standards with known primary and secondary blood type categories and known mixing ratios are prepared. Then, these standards are treated under different simulated environmental conditions, including artificial nuclease digestion for different durations to simulate different degrees of DNA degradation, and the addition of different concentrations of heme or humic acid to simulate different levels of amplification inhibitor residues, forming a multi-gradient combination of environmental conditions covering fresh samples to severely degraded samples, and from no inhibition to strong inhibition. Each sample treated as described above undergoes steps S1 to S4 of the method of this invention, acquiring its dynamic colorimetric image sequence and extracting the absolute fluorescence intensity ratio-time series of each detection line as an environmental condition variant for the training sample. To achieve latent variable decoupling, an adversarial training strategy is introduced during training: a blood type component discriminator is connected in parallel on the latent variable space of the variational autoencoder. This discriminator attempts to predict the blood type component category of the sample from the latent variables of the sample environment. The encoder part of the variational autoencoder aims to maximize the prediction error of this discriminator and performs adversarial optimization through a gradient inversion layer, forcing statistical independence between the latent variables of the sample environment generated by the encoder and the blood type component information. Alternatively, mutual information minimization constraints can be used to replace adversarial training. That is, a mutual information estimation term between the latent variables of the sample environment and the blood type component label is added to the loss function as a penalty term, and decoupling is explicitly forced by minimizing this mutual information. The latent variables of the sample environment are comprehensive quantitative indicators characterizing the degree of degradation of nucleic acid templates and the level of inhibitor residues in the test sample. Their values ​​can be normalized to the interval [0, 1], where 0 corresponds to the ideal sample state with no degradation and no inhibition, 1 corresponds to the extreme deterioration state with severe degradation or strong inhibition, and intermediate values ​​continuously reflect the gradual process of sample quality change.

[0064] In terms of the variational autoencoder network structure design, the encoder consists of multiple stacked fully connected layers or one-dimensional convolutional layers, typically configured with four fully connected layers. The number of neurons in each layer decreases progressively. For example, the input layer receives the chromogenic dynamics spatiotemporal feature vectors after flattening and splicing each detection line, which are then progressively compressed through hidden layers with 256, 128, and 64 neurons. The final output layer simultaneously outputs the mean vector and log-variance vector of the latent variable distribution. The latent variable dimension is set to 16 dimensions, with the first 8 dimensions assigned to blood type component-related latent variables and the last 8 dimensions assigned to sample environment latent variables. The independence between the two sets of latent variables is ensured by decoupling constraints in the loss function. The decoder adopts a mirror-symmetric structure with the encoder, taking the spliced ​​latent variable vector as input, which is progressively restored through hidden layers with 64, 128, and 256 neurons. The final output layer reconstructs the chromogenic dynamics spatiotemporal feature vector with the same dimension as the original input. In addition to the aforementioned environmental condition variant samples, the training dataset also needs to include original mixed sample standards without any simulated environmental treatment as a benchmark. All samples are labeled with triple tags: major blood type category, minor blood type category, and the proportion of major and minor components, and are divided into training, validation, and test sets according to the proportions. The total loss function during training consists of three terms: reconstruction loss, which calculates the mean squared error between the decoder output and the original input; Kullback-Leibler divergence loss, which constrains the latent variable distribution to approximate a standard normal distribution; and decoupling loss, which separates the latent variables of blood type components from the latent variables of the sample environment through the aforementioned mutual information minimization term or adversarial training term.

[0065] The ABO blood type rapid on-site typing method of the present invention will be further explained and illustrated below with examples 1-2.

[0066] Example 1: Rapid typing of a single sample and intelligent interpretation of weak positive signals This embodiment simulates a typical crime scene, using a human oral swab with known blood type A as the sample to be tested, to demonstrate the complete process of the method of the present invention from sample processing to intelligent interpretation.

[0067] After swabbing the inside of the subject's mouth with a sterile cotton swab, the operator immersed the swab tip into a sample processing tube containing 300 μL of lysis buffer, stirred thoroughly for 10 seconds, and then removed the swab to obtain a crude lysis buffer containing genomic DNA. Two μL of this crude lysis buffer was directly added to a pre-lyophilized RPA reaction tube. The lyophilized system contained recombinase, single-stranded binding protein, polymerase, and a specific primer set designed for the three characteristic SNP sites of the ABO gene: c.261delG, c.803G>C, and c.526C>T. The primer was then added to reconstitute the mixture and mixed thoroughly. The reaction tube was then incubated in a portable thermostat at 39°C for 12 minutes. After the reaction, the RPA amplification product was mixed with Cas12a reaction buffer at a volume ratio of 1:5 and then transferred to reaction wells containing a CRISPR / Cas12a trans-cleavage lyophilized system. This lyophilized system contained Cas12a protein, a mixture of crRNAs targeting the three SNP sites mentioned above, and corresponding reporter probes. The reporter probes for the specific crRNAs targeting the A allele were FAM-ssDNA-Biotin, for the B allele they were targeting DIG-ssDNA-Biotin, and for the O allele they were targeting TAMRA-ssDNA-Biotin. The reaction wells were placed in adjacent positions on the same heating module and incubated at 39°C for 8 minutes without temperature switching. This allowed Cas12a to specifically recognize the matching amplicons under the guidance of crRNA and initiate trans-cleavage activity, releasing cleavage fragments carrying free FAM, DIG, or TAMRA groups, resulting in the reaction mixture.

[0068] The entire reaction mixture was transferred to the sample pad of the lateral flow test strip. The test strip was sequentially set along the chromatography direction with the anti-FAM antibody capture line corresponding to the A allele signal, the anti-DIG antibody capture line corresponding to the B allele signal, the anti-TAMRA antibody capture line corresponding to the O allele signal, and a control line. The multispectral image acquisition program of the portable reader was activated immediately upon sample addition. The reader's built-in CMOS camera continuously acquired images of the detection area at a frequency of 2 frames per second through two emission channels at 540nm and 660nm, for 15 minutes. Simultaneously, ambient light sensor data was recorded, resulting in a multispectral dynamic colorimetric image sequence containing approximately 1800 frames. A clear band appeared on the control line within approximately one minute after sample addition, indicating that the chromatography process was completed normally.

[0069] After the dynamic colorimetric image sequence is input into a pre-trained deep learning interpretation model, the model first identifies the matrix region and detection line region in each frame through the semantic segmentation sub-network of the background compensation module. It then uses a multispectral physical model to remove matrix interference such as oral cavity exfoliated cell residues and performs radiometric calibration using rare-earth fluorescence reference bands in the background region, converting the original grayscale signal into an absolute fluorescence intensity ratio. Subsequently, the model uses a three-dimensional convolutional neural network to simultaneously extract the spatiotemporal characteristics of the colorimetric dynamics of each detection line within 0 to 15 minutes along the time and spatial axes. In the test of this embodiment, the anti-FAM antibody capture line began to show a identifiable increase in grayscale approximately 2 minutes and 30 seconds after sample addition, and the colorimetric rate maintained a high linear increase between 4 and 8 minutes, reaching a saturation plateau at approximately 11 minutes. The anti-DIG antibody capture line and the anti-TAMRA antibody capture line showed no significant grayscale changes throughout the entire process, consistently fluctuating at the baseline level. After comprehensively comparing global features such as the relative color development delay, saturation plateau height, and inter-line signal ratio of the three detection lines through a self-attention mechanism, the model outputs a genotyping result of type A with a genotyping confidence level of 0.987. Since the anti-FAM capture line shows stable growth in the early stages while the other two capture lines show no response throughout, the model classifies this kinetic pattern as a typical monoallelic positive sample. Its high confidence output indicates that the interpretation result is highly reliable and can be directly used as a reference for on-site evidence screening.

[0070] Example 2: Decoupling of blood types from multiple sources and correction for environmental factors in mixed samples This embodiment simulates a common mixed blood stain sample scenario in forensic cases. Two venous blood samples, one known to be type A and the other type O, are mixed in a 3:1 volume ratio and then dripped onto a cotton cloth carrier to dry naturally, thus creating a mixed blood stain sample. This verifies the multi-group decomposition coupling ability of the method of the present invention under mixed body fluid conditions and the environmental factor correction effect under partial sample degradation.

[0071] Cut the above-mentioned mixed blood-stain cotton cloth into small pieces of approximately 5 mm × 5 mm and soak them in 500 μL of lysis buffer at room temperature for 15 minutes, intermittently agitating to promote cell lysis and DNA release. Take 2 μL of the soaking supernatant as a template and perform RPA isothermal amplification and CRISPR / Cas12a trans-cutting reaction sequentially under the same conditions as in Example 2. To simulate possible degradation of field samples, another portion of the same mixed blood-stain soaking supernatant was sealed and placed in a 37°C incubator for 48 hours to induce partial nucleic acid degradation, while avoiding the introduction of additional variables through evaporation and concentration. This served as an environmental interference control group for simultaneous detection.

[0072] In the test of undegraded mixed samples, after the deep learning interpretation model analyzed the dynamic colorimetric image sequence, the latent variable value of the sample environment output by the variational autoencoder after sample environmental factor decoupling preprocessing by the multi-component signal decoupling module was 0.07, close to the ideal state of no degradation and no inhibition, indicating good sample quality. The feature sequence after environmental correction entered the multi-head attention layer for deconvolution integral decomposition. The model successfully separated the primary blood type kinetic curve and the secondary blood type kinetic curve from the superimposed kinetic curve: the primary blood type curve showed colorimetric characteristics highly consistent with pure type A samples, with the anti-FAM capture line signal starting at about 3 minutes and saturating at about 10 minutes, and its integral value of absolute fluorescence intensity over time was 0.843; the secondary blood type curve showed a colorimetric pattern consistent with pure type O samples, with the anti-TAMRA capture line signal starting at about 4 minutes and 30 seconds and saturating at about 12 minutes, and an integral value of 0.286. The ratio of the integral values ​​of the two curves was approximately 2.95:1, which is basically consistent with the actual A:O mixing ratio of 3:1. The model's final output is that the primary blood type is A and the secondary blood type is O, with an estimated ratio of 2.95:1. The confidence level for the primary blood type classification is 0.973, and the confidence level for the secondary blood type classification is 0.941.

[0073] In the control group test after 48 hours of pre-degradation, the latent variable value of the sample environment output by the variational autoencoder after sample environmental factor decoupling preprocessing was 0.68, indicating a moderate degree of nucleic acid degradation in the sample. If the environmental factor correction was skipped and the deconvolution integral was performed directly, the saturation plateau heights of the primary and secondary blood type kinetic curves decreased by approximately 19% and 24% respectively compared to the undegraded sample, and the color development start time of both curves was delayed by approximately 1 to 2 minutes, with the ratio estimation deviation widening to approximately 2.3:1. However, after a complete environmental factor correction process, the model fixed the latent variables of the sample environment to standard environmental vectors and reconstructed the kinetic characteristics using only the latent variables related to blood type components. The saturation plateau heights, color development start times, and curve shapes of the corrected primary and secondary blood type curves were restored to levels comparable to those of the undegraded sample, with an integral ratio of 2.89:1, reducing the deviation from the true value of 3:1 to an acceptable range. The model's final output of the typing conclusions and confidence levels were completely consistent with those of the undegraded samples, indicating that the decoupling pretreatment of environmental factors effectively eliminated the interference of degradation-induced kinetic distortion on the quantitative analysis of mixed plaques, ensuring the accuracy of multi-source blood type typing and proportion estimation under harsh sample conditions.

[0074] Accordingly, a second aspect of the present invention provides a rapid on-site ABO blood typing kit based on RPA-LFD technology and artificial intelligence-assisted interpretation. The kit performs ABO blood typing processing based on the aforementioned rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation, including: The recombinase polymerase isothermal amplification system contains a specific primer set for amplifying characteristic single nucleotide polymorphism sites of the ABO blood group gene; The CRISPR / Cas12a trans-cleavage system comprises crRNA designed for characteristic single nucleotide polymorphism sites, Cas12a protein, and a reporter probe labeled with a detectable hapten. The CRISPR / Cas12a trans-cleavage system is used to activate the Cas12a trans-cleavage activity after the crRNA specifically binds to the amplification product, cleaving the reporter probe and releasing a detectable reporter signal. Lateral flow test strips are arranged sequentially along the chromatography direction, including a sample pad, a conjugate pad, a detection area, and an absorption pad. Multiple detection lines and a control line are fixed within the detection area. The multiple detection lines are used to capture detectable report signals corresponding to different alleles. The image acquisition device includes a multispectral CMOS camera and an ambient light sensor, used to continuously acquire dynamic color development image sequences of the detection area at preset time intervals within a preset color development acquisition period calculated from the time of sample addition; The image interpretation device has a pre-trained deep learning interpretation model built in. The deep learning interpretation model is used to extract the spatiotemporal features of the color development dynamics of each detection line in the dynamic color development image sequence, and outputs the ABO blood type classification result of the sample to be tested and the confidence level of the classification result based on the features.

[0075] The embodiments of this invention aim to protect a rapid on-site ABO blood typing method and kit based on RPA-LFD technology and artificial intelligence-assisted interpretation, which has the following effects: 1. By organically integrating upconversion nanoparticle labeling, multispectral dynamic image acquisition, and background compensation module in the deep learning interpretation model, near-infrared excitation is used at the physical layer to completely avoid the interference of autofluorescence in biological matrix. At the algorithm layer, semantic segmentation, multispectral physical modeling, and rare earth reference radiation calibration are used to transform the original colorimetric signals of each detection line into physically comparable absolute fluorescence intensity ratios. Based on this, the spatiotemporal characteristics of colorimetric dynamics are constructed, enabling ABO blood typing results and confidence levels in complex forensic sample matrices such as blood spots and saliva spots to achieve consistency with those using purified DNA samples. This fundamentally solves the inherent defects of existing lateral flow test strips, such as low signal-to-noise ratio and unreliable interpretation under strong background interference. 2. By using the multi-component signal decoupling module based on the multi-head attention mechanism in the deep learning interpretation model, the superimposed chromogenic kinetic curves of mixed samples on multiple detection lines are deconvoluted and integrally solved. The contribution weight of different blood type components to the chromogenic rate of each detection line is automatically learned, and the primary blood type kinetic curves and secondary blood type kinetic curves corresponding to different individual sources are separated. Based on the ratio of the area under the curve, the ratio of primary and secondary blood type components is output as an estimated value. Thus, a single lateral flow test strip can simultaneously complete the qualitative and quantitative analysis of ABO blood types from multiple sources in mixed spots, effectively making up for the technical gap of existing methods that can only output a single blood type conclusion and cannot identify blood type components from a second source when dealing with mixed body fluid samples. 3. Before performing deconvolution integral decomposition on mixed samples using a deep learning interpretation model, a variational autoencoder is used to decouple the spatiotemporal characteristics of the chromogenic dynamics of each detection line from the sample environmental factors. This extracts latent environmental variables that are independent of blood type components and only characterize the degree of sample degradation and the level of inhibitor residues. These latent variables are then extracted from the features and used as input for deconvolution integral decomposition. This achieves a two-stage analysis strategy of correction followed by decoupling. As a result, the estimation of the proportion of each blood type component in the mixed sample is no longer affected by the non-uniform distortion of the kinetic curve caused by uneven degradation or different levels of inhibitor interference from each source component. This significantly improves the quantitative accuracy of mixed spot typing under harsh forensic conditions.

[0076] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation, characterized in that, Includes the following steps: Step S1: Perform isothermal amplification of the sample to be tested using recombinase polymerase to obtain an amplification product containing the characteristic single nucleotide polymorphism site of the ABO blood group gene; Step S2: The amplification product is contacted with the CRISPR / Cas12a trans-cutting system to obtain a reaction mixture. The CRISPR / Cas12a trans-cutting system is used to perform single-base specific recognition of the characteristic single nucleotide polymorphism site and generate a detectable report signal corresponding to each allele. Step S3: The reaction mixture is loaded onto a lateral flow test strip, and after chromatographic separation, a colorimetric image sequence of detection lines indicating different alleles is formed in the detection area of ​​the lateral flow test strip; Step S4: During the preset color development acquisition period starting from the sampling time, continuously acquire the color development image of the detection line at preset time intervals to obtain a dynamic color development image sequence; Step S5: Input the dynamic color development image sequence into the pre-constructed deep learning interpretation model. The deep learning interpretation model extracts the spatiotemporal features of the color development dynamics of each detection line and outputs the ABO blood type classification result of the sample to be tested and the confidence level of the classification result based on the features.

2. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 1, characterized in that, In step S3, the colorimetric image of the detection line is generated by a detection antibody labeled with upconversion nanoparticles, which emit 540nm and 660nm dual-color visible light under 980nm near-infrared excitation light. In step S4, the continuous acquisition is multispectral image acquisition, and ambient light sensor data is recorded synchronously during the acquisition process.

3. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 2, characterized in that, The deep learning interpretation model includes a background compensation module; The background compensation module includes a U-Net-based background semantic segmentation sub-network, used to segment the matrix region and detection line region of the dynamic color image sequence; The background compensation module also combines a multispectral physical model and uses the segmentation results and ambient light sensor data to remove sample matrix absorption attenuation and ambient light reflection interference. The background area of ​​the lateral flow test strip is pre-printed with a rare earth fluorescent reference strip. The background compensation module performs radiometric calibration using the rare earth fluorescent reference strip and outputs the absolute fluorescence intensity ratio of each detection line. The spatiotemporal characteristics of the color development kinetics are constructed based on the absolute fluorescence intensity ratio.

4. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 3, characterized in that, The sample to be tested is whole blood spot, semen spot, or saliva spot; The deep learning interpretation model eliminates sample matrix effects through the background compensation module, and the dynamic colorimetric image sequence is constructed based on the absolute fluorescence intensity ratio, so that the ABO blood typing results and confidence levels are consistent with the results obtained using purified DNA samples.

5. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 1, characterized in that, In step S5, the deep learning interpretation model includes a multi-component signal decoupling module based on a multi-head attention mechanism, which is used to perform deconvolution integral decomposition on the spatiotemporal features of the colorimetric dynamics of each detection line, output at least one primary blood type dynamic curve and one secondary blood type dynamic curve, and output the primary blood type classification result and the secondary blood type classification result, as well as the estimated ratio of primary and secondary blood type components, based on the primary blood type dynamic curve and the secondary blood type dynamic curve.

6. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 5, characterized in that, The input to the multi-component signal decoupling module is the gray value-time series of each detection line, which is then input to the multi-head attention layer after position encoding. The multi-head attention layer uses each detection line channel as an independent key-value pair and learns the contribution weight of different blood type components to the color development rate of each detection line through query vectors, thereby realizing the deconvolution integral solution.

7. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 5, characterized in that, The primary blood group kinetic curve and the secondary blood group kinetic curve correspond to the kinetic characteristics of blood group antigen-antibody reactions from different individuals. The estimated ratio is calculated based on the ratio of the area under the curve of the primary blood group kinetic curve to that of the secondary blood group kinetic curve.

8. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 5, characterized in that, When the deconvolution integral solution can only separate one effective dynamic curve, the deep learning interpretation model determines that the sample to be tested is a single human source sample, outputs only a unique blood type classification result, and marks the mixing ratio as a single source.

9. The method for rapid on-site ABO blood typing based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in claim 5, characterized in that, In step S5, before performing the deconvolution integral solution, the multi-component signal decoupling module also performs sample environmental factor decoupling preprocessing on the spatiotemporal characteristics of the color development dynamics of each detection line. The sample environmental factor decoupling preprocessing includes: inputting the chromogenic kinetic spatiotemporal features of each detection line into a pre-trained variational autoencoder to extract sample environmental latent variables independent of blood type components, wherein the sample environmental latent variables characterize the degradation degree and / or inhibitor residue level of the sample to be tested; and separating the sample environmental latent variables from the chromogenic kinetic spatiotemporal features to obtain environmentally corrected chromogenic kinetic spatiotemporal features of each detection line, which are used as inputs to the deconvolution integral solution.

10. A rapid on-site ABO blood typing kit based on RPA-LFD technology and artificial intelligence-assisted interpretation, characterized in that, ABO blood typing is performed based on the rapid on-site ABO blood typing method based on RPA-LFD technology and artificial intelligence-assisted interpretation as described in any one of claims 1-9, including: The recombinase polymerase isothermal amplification system contains a specific primer set for amplifying characteristic single nucleotide polymorphism sites of the ABO blood group gene; The CRISPR / Cas12a trans-cleavage system comprises crRNA designed for the characteristic single nucleotide polymorphism site, Cas12a protein, and a reporter probe labeled with a detectable hapten. The CRISPR / Cas12a trans-cleavage system is used to activate the Cas12a trans-cleavage activity after the crRNA specifically binds to the amplification product, cleaving the reporter probe and releasing a detectable reporter signal. A lateral flow test strip, wherein a sample pad, a conjugate pad, a detection area and an absorption pad are sequentially arranged along the chromatography direction on the lateral flow test strip, and multiple detection lines and a control line are fixed in the detection area, the multiple detection lines being used to capture the detectable report signals corresponding to different alleles; An image acquisition device, comprising a multispectral CMOS camera and an ambient light sensor, is used to continuously acquire a dynamic color development image sequence of the detection area at preset time intervals within a preset color development acquisition period calculated from the time of sample addition. The image interpretation device has a pre-trained deep learning interpretation model built in. The deep learning interpretation model is used to extract the spatiotemporal features of the color development dynamics of each detection line in the dynamic color image sequence, and outputs the ABO blood type classification result of the test sample and the confidence level of the classification result based on the features.

Images