Methods and systems for processing cell count data from multiple tests

Abstract

The present disclosure provides a method of processing cell count data from multiple tests. The method of processing cell count data from multiple tests comprises the steps of: determining an overlap parameter between the multiple tests using a flow cytometer; pre-processing the overlap parameter; and developing a classifier by supervised machine learning.

Description

Technical Field

[0001] This disclosure relates to a method for processing cell count data, and more particularly to a method for processing cell count data from multiple tests. Background Technology

[0002] Flow cytometry (FC) is essential for the diagnosis and monitoring of blood disorders. To operate a flow cytometer, a blood, bone marrow, or tissue cell sample is placed in a suspension and injected into the analyzer. The cells are arranged in a line and then passed in front of a laser beam, scattered light, and fluorescent lamps. The cells are then counted and classified. The data is stored in a computer and reported as histograms or scatter plots for further analysis.

[0003] Current analytical methods and software primarily rely on skilled analysts who manually examine and interpret a series of two-dimensional graphs through complex, sequential gated processes. This labor-intensive approach faces challenges due to a shortage of skilled professionals and the growing demand for rapid, accurate, and reproducible functional group (FC) analysis results. Furthermore, the lack of consistency between manual and subjective analysis of clinical FC data often leads to discrepancies in interpretation. Moreover, because each medical device manufacturer develops specific kits for particular applications, data processing, analysis, and results reporting must be customized for each kit and each application.

[0004] Much effort has been put into reconciling antibody groups, sample processing, analysis, and results reporting for various disease diagnoses. However, the main challenge lies in ensuring that laboratories fully comply with these guidelines.

[0005] The concept of standardization exists within a scope that can evolve over time and across different laboratories. Harmonization, as a practical initial step, may ultimately lead to the development of a more comprehensive set of standardized methods.

[0006] In the field of flow cytometry, although there is a consensus that standardized methods can improve the reproducibility of flow cytometry measurements and studies between laboratories, significant differences still exist between different centers. Much effort has been devoted to reconciling antibody groups, sample processing, analysis, and results reporting for various disease diagnoses.

[0007] To minimize heterogeneity and improve the accuracy of clinical flow cytometry tests, the medical community has developed and published numerous guidelines, as cited below.

[0008] Borowitz MJ, Craig FE, Digiuseppe JA, Illingworth AJ, Rosse W,Sutherland DR, Wittwer CT, Richards SJ; Clinical Cytometry Society.Guidelines for the diagnosis and monitoring of paroxysmal nocturnalhemoglobinuria and related disorders by flow cytometry. Cytometry B ClinCytom. 2010 Jul;78(4):211-30. doi: 10.1002 / cyto.b.20525. PMID: 20533382.。

[0009] Davis BH, Holden JT, Bene MC, Borowitz MJ, Braylan RC, Cornfield D,Gorczyca W, Lee R, Maiese R, Orfao A, Wells D, Wood BL, Stetler-Stevenson M.2006 Bethesda International Consensus recommendations on the flow cytometricimmunophenotypic analysis of hematolymphoid neoplasia: medical indications.Cytometry B Clin Cytom. 2007;72 Suppl 1:S5-13. doi: 10.1002 / cyto.b.20365.PMID: 17803188.。

[0010] Borowitz MJ, Wood BL, Keeney M, Hedley BD. Measurable ResidualDisease Detection in B-Acute Lymphoblastic Leukemia: The Children's Oncology Group (COG) Method. Curr Protoc. 2022 Mar;2(3):e383. doi: 10.1002 / cpz1.383.PMID: 35263042.

[0011] Van Dongen JJ, Lhermitte L, Böttcher S, Almeida J, van der Velden VH, Flores-Montero J, Rawstron A, Asnafi V, Lécrevisse Q, Lucio P, Mejstrikova E, Szczepański T, Kalina T, de Tute R, Brüggemann M, Sedek L, Cullen M, LangerakAW, Mendonça A, Macintyre E, Martin-Ayuso M, Hrusak O, Vidriales MB, Orfao A; EuroFlow Consortium (EU-FP6, LSHB-CT-2006-018708). EuroFlow antibody panels for standardized n-dimensional flow cytometric immunophenotyping of normal, reactive and malignant leukocytes. Leukemia. 2012 Sep;26(9):1908-75. doi:10.1038 / leu.2012.120. Epub 2012 May 3. PMID: 22552007; PMCID: PMC3437410.).

[0012] These guidelines aim to establish standardized practices for antibody groups, sample processing, analysis, and results reporting. Summary of the Invention

[0013] However, the main challenge is ensuring that laboratories fully comply with these guidelines.

[0014] While the goal of standardization is to achieve reproducibility, reconciliation can, in certain cases, provide consistent interpretations, counts, and even patterns. These concepts exist within a scope that can evolve over time and across different laboratories. Reconciliation serves as a practical initial step and may ultimately lead to the development of more comprehensive standardization methods.

[0015] In our research, we have demonstrated that supervised machine learning (ML) methods can effectively detect immune phenotypic abnormalities and identify different diseases at the specimen level using flow cytometry data obtained from the same panel measurements on the same or two different instrument models from the same manufacturer.

[0016] Therefore, to improve the reproducibility of data analysis across laboratories with different panels but harmonic parameter subsets, this disclosure covers exemplary embodiments of methods and associated systems for automatically classifying flow cytometry counting data. Data is measured through various reagent panels and obtained through different flow instrument models, all for the same purpose, while utilizing common markers measured across reagent panels.

[0017] The purpose of this disclosure is to provide a method for processing cell count data from multiple tests, including: (a) using flow cytometry to determine overlap parameters between the multiple tests; (b) preprocessing the overlap parameters; and (c) developing a classifier through supervised machine learning.

[0018] In some embodiments, the determination of overlap parameters is performed by identifying overlap parameters in different tests that describe the same biomarker.

[0019] In some embodiments, preprocessing the overlap parameters includes rearranging the overlap parameters and normalizing them.

[0020] In some embodiments, rearranging overlapping parameters includes performing parameter alignment.

[0021] In some embodiments, normalization includes z-score normalization, min-maximum normalization, quantile normalization, or rank normalization, and can be applied per locus or per sample.

[0022] In some embodiments, preprocessing the overlap parameters further includes extracting the distribution and embedding phenotypic features into the phenotypic representation.

[0023] In some embodiments, the distribution is a high-dimensional cell distribution.

[0024] In some embodiments, the high-dimensional cell distribution is performed using a Gaussian Mixture Model (GMM), Principal Component Analysis (PCA), or deep neural network methods such as Convolutional Neural Networks (CNN) and autoencoders.

[0025] In some embodiments, the representation is a high-dimensional phenotypic representation.

[0026] In some embodiments, embedding phenotypic features into a distribution is achieved through Fisher vectorization, which embeds phenotypic features into a high-dimensional phenotypic representation vector at the sample level.

[0027] In some embodiments, the preprocessing of the overlap parameters further includes inputting the vector along with the corresponding diagnostics into the support vector machine.

[0028] In some embodiments, developing a classifier includes data training and sample classification.

[0029] In some embodiments, the method is used to count cells, sort cells, determine cell function, determine cell characteristics, or detect microorganisms such as bacteria, fungi, or yeast, or to search for biomarkers, or to assist in the diagnosis and potential treatment of blood and bone marrow cancers.

[0030] In some embodiments, visualization is also included after developing the classifier.

[0031] In some embodiments, the overlap parameter is selected from the group that includes light scattering properties and fluorescent labels.

[0032] Another object of this disclosure is to provide a system for processing cell count data from multiple tests, which is signal-connected to one or more cell count data providing devices to receive cell count data from the cell count data providing devices. The system includes: a storage module; and an automatic classification module configured to be signal-connected to the storage module; wherein multiple program codes are stored in the storage module, and wherein after the automatic classification module executes the multiple program codes stored in the storage module, the automatic classification module performs the steps of a method for processing cell count data from multiple tests.

[0033] This disclosure claims priority to U.S. Provisional Patent Application No. 63 / 607,305, filed December 7, 2023, the entire contents of which are incorporated herein by reference. Attached Figure Description

[0034] Figure 1A schematic diagram of a system for processing cell count data from multiple tests according to an embodiment of this disclosure is shown.

[0035] Figure 2 A flowchart illustrating a method for processing cell count data from multiple tests according to an embodiment of this disclosure.

[0036] Figure 3 The figure shows the experimental results of single-parameter characteristic selection provided for the embodiments disclosed herein.

[0037] Figure 4 This diagram illustrates an example of developing and visualizing panel-independent sample classification using cell count data according to an embodiment of this disclosure.

[0038] Figure 5 This disclosure provides a detailed example architecture for developing and visualizing panel-independent sample classification using cell count data, according to embodiments of the present disclosure.

[0039] Figure 6A and 6B This displays a visualization of exemplary cross-panel AML samples versus samples in a non-tumor classification, according to embodiments of this disclosure.

[0040] Figure 7 This shows exemplary results of developing panel-independent sample classification using cell count data according to embodiments of this disclosure, requiring the minimum number of identification parameters.

[0041] Figure 8 This illustrates an example architecture for developing and visualizing panel-independent sample classification using cell count data according to embodiments of this disclosure.

[0042] Figure 9A and 9B This displays an example visualization of samples across panel AML and non-tumor classifications according to embodiments of this disclosure.

[0043] Figure 10 This illustrates an example architecture for developing and visualizing panel-independent sample classification using cell count data according to embodiments of this disclosure.

[0044] Figure 11 This displays a visualization of samples in an example cross-panel lymphatic abnormality versus no-abnormality classification according to an embodiment of this disclosure.

[0045] Figure 12 Example block diagrams are provided to illustrate artificial intelligence (AI) systems that can implement at least some of the operations described herein, according to embodiments of this disclosure.

[0046] Figure 13 This diagram illustrates an example of z-score normalization according to an embodiment of this disclosure.

[0047] Figure 14 This diagram shows an example of density contour lines according to an embodiment of the present disclosure.

[0048] Figure 15 This diagram shows an example of a classifier according to an embodiment of the present disclosure.

[0049] Figure 16 Example schematic diagrams showing evaluation metrics according to embodiments of this disclosure.

[0050] Figure 17 Example schematic diagrams showing evaluation metrics according to embodiments of this disclosure.

[0051] Figure 18 This diagram illustrates an example of GMM and Fisher vectorization according to an embodiment of this disclosure.

[0052] Figure 19 This diagram illustrates an example of cross-validation according to an embodiment of this disclosure.

[0053] Figure 20 This diagram illustrates an example of compensation according to an embodiment of the present disclosure. Detailed Implementation

[0054] The present disclosure will be described in more detail below with reference to the embodiments and accompanying drawings, so that the purpose, features and effects of the present disclosure can be more clearly understood.

[0055] It should be noted that the steps described herein may be performed sequentially, in reverse order, or the order may be appropriately changed or skipped during the control process. It should be clarified that "the first step may be performed after the second step" in this disclosure can mean "the first step immediately follows the second step" and / or "the second step is followed by other steps (such as the third step), and then the first step."

[0056] Furthermore, it should be noted that in this disclosure, terms such as "first," "second," and "third" are used to distinguish differences between elements, rather than to limit the elements themselves or represent a specific order of elements. It should also be noted that in the following description, the same components or steps may be represented by the same reference numerals.

[0057] Furthermore, the term "coupling" as used in this disclosure can mean "direct connection" and / or "indirect connection." Specifically, "the first component is configured to be coupled to the second component" can mean "the first component is configured to be directly connected to the second component" and / or "the first component is configured to be indirectly connected to the second component." Also, the term "signal connection" as used in this disclosure can mean transmitting a signal in any manner.

[0058] This disclosure provides a method for creating automated classification models across different groups using parameters that are common measurements designed for the same diagnostic or monitoring purpose. This disclosure can significantly improve the reproducibility of data analysis between laboratories with different panels but harmonic subsets of parameters.

[0059] The cell count data used in this disclosure are from flow cytometry counts performed as part of a clinical evaluation of bone marrow samples from patients with assessed potential hematologic disorders from various institutions.

[0060] Cross-test data or data from multiple tests refer to data from two or more tests. Specifically, cross-test data or data from multiple tests refer to data processed by two or more experiments using different models of flow cytometers, different panels, different kits, different treatment methods, different pretreatment methods, or with differences.

[0061] Figure 1 This diagram illustrates a system for processing cell count data from multiple tests according to an embodiment of the present disclosure.

[0062] In some embodiments, the system 100, which processes cell count data from multiple tests, is adapted to be signal-connected to one or more cell count data providing devices to receive cell count data from the cell count data providing devices. In some embodiments, the system 100 may be an electronic device, such as a computer, server computer, client computer, personal computer, desktop computer, laptop computer, tablet computer, workstation, server, cloud server, smartphone, personal digital assistant, and / or computing device. The system 100 includes an automatic classification module 110, a storage module 120, an I / O interface 130, and a communication interface 140.

[0063] In some embodiments, the automatic classification module 110 is configured to be signal-connected to the storage module 120, and in some embodiments, also signal-connected to the I / O interface 130 and the communication interface 140. In some embodiments, the automatic classification module 110 may be a processor, CPU, GPU, or other computing unit capable of executing a program or executing program code stored in the storage module 120. The automatic classification module 110 is configured to execute a method for processing streaming cell counting data from multiple tests after executing multiple program codes stored in the storage module 120.

[0064] In some embodiments, storage module 120 is configured to store program code, programs, processes, and / or operations of the disclosed method that can be performed by automatic classification module 110. In some embodiments, storage module 120 may include one or more non-volatile memories and one or more volatile memories. In some embodiments, volatile memories may be finished products well known to those skilled in the art to which this disclosure pertains, such as, but not limited to, various types of dynamic random access memory or static random access memory. In some embodiments, non-volatile memories may be finished products well known to those skilled in the art to which this disclosure pertains, such as, but not limited to, various types of read-only memory or flash memory. In some embodiments, cell count data from multiple tests may also be stored in storage module 120.

[0065] In some embodiments, I / O interface 130 is configured to allow user operating system 100 to perform the methods or processes, programs, or operations disclosed herein. In some embodiments, I / O interface 130 may provide a visual and / or auditory user interface, such as a display unit, touch screen, projector, speaker, telephone voice, keyboard, mouse, motion detection, and voice recognition, as a medium for operating system 100.

[0066] In some embodiments, the communication interface 140 is configured to communicate with data external to the system 100. In some embodiments, the communication interface 140 is signal-connected to one or more flow cytometers, different panels, or other cell counting data providing devices external to the system 100, receiving cell counting data from multiple tests via physical signal lines and / or virtual signal lines. In some embodiments, the communication interface 140 is signal-connected to a database, server, cloud server, or other storage device via physical signal lines and / or virtual signal lines to receive cell counting data from multiple tests. In some embodiments, the physical signal line may be, for example, a network signal cable conforming to Internet Protocol, a USB cable, an RJ45 cable, an RS232 cable, etc., but is not limited thereto. In some embodiments, the virtual signal line may be, for example, a wireless communication method conforming to wireless communication protocols such as Wi-Fi, 4G / 5G / 6G, Bluetooth, short-range communication, etc., but is not limited thereto.

[0067] Figure 2 A flowchart illustrating a method for processing cell count data from multiple tests according to an embodiment of this disclosure. Figure 2 The method shown for processing flow cytometry cell count data from multiple tests can be performed by an automatic classification module. Figure 2 The method shown for processing flow cytometry counting data from multiple tests includes the following steps: (S100) determining overlap parameters between multiple tests using a flow cytometer; (S200) preprocessing the overlap parameters; and (S300) developing a classifier through supervised machine learning.

[0068] In step S100, flow cytometry is used to determine overlap parameters between multiple tests. In some embodiments, the method utilizes common markers extracted from various sources as comparable features.

[0069] In some embodiments, data are sourced from different manufacturers and instruments, using different panels, wherein reagents are included to measure multiple markers / parameters for flow cytometry analysis. In some embodiments, several different datasets and different panels are integrated, for example, five different datasets and four different panels are integrated, which together contain ten common parameters to provide cell count data from multiple tests.

[0070] In some embodiments, the data formats of the relevant files are managed and processed according to specific protocols to ensure seamless integration into existing workflows. In some embodiments, the method provides broad compatibility, supporting various cell counting data file formats, such as Flow Cytometry Standard (FCS) formats 2.0, 3.0, 3.1 and LMD list mode files, ensuring diverse applications across various scientific disciplines.

[0071] In some embodiments, parameters from multiple tests can be collected from a single flow cytometer instrument or panel. In some embodiments, parameters from multiple tests can be collected from multiple flow cytometer instruments or panels.

[0072] In some embodiments, the naming of markers associated with cell count data may vary, for example, when markers have different aliases or the same marker is labeled with different fluorophores. In some embodiments, additional alignment work may be required to ensure the consistency and accuracy of data comparisons, such as automatically searching for basic marker information from a database to check the information and relevance of each marker, determining spectral ranges to identify markers, etc.

[0073] In step S200, the overlap parameters are preprocessed. Specifically, this involves correcting for biases between different sources to ensure that the method for processing flow cytometry cell counting data is consistent and reliable across various datasets.

[0074] In some embodiments, the method employs one or more preprocessing steps, including noise reduction and data normalization techniques (e.g., sample z-score normalization, site z-score normalization, and rank normalization). In some embodiments, calibration techniques utilizing reference materials may also be employed.

[0075] Therefore, this method can achieve high accuracy by using all overlapping parameters and enables robust comparisons and analyses across different data sources. Furthermore, through feature selection, this method has been shown to produce relatively fair results using one or more parameters. This flexibility allows for effective performance tuning to meet specific analytical needs.

[0076] In some embodiments, the results of two feature selection experiments across five-point AML classification are shown below. Data were collected from AML and non-tumor samples from five laboratories. These laboratories used four panels and three instrument models for measurement. A total of ten common parameters were found across the different panels, including FSC-A, FSC-H, SSC-A, CD7, CD19, CD34, CD45, CD56, CD117, and HLA-DR.

[0077] For single-parameter feature selection, Table 1 presents the experimental results of the single-parameter feature selection performance provided in this embodiment. Figure 3 The figure shows the experimental results of the single-parameter characteristic selection performance provided in this embodiment.

[0078] CD117 97.08% 94.41% SSC-A 95.16% 89.78% CD34 93.23% 87.42% HLA-DR 92.04% 86.50% FSC-H 83.96% 77.67% CD7 75.31% 71.13% CD56 73.32% 67.91% FSC-A 72.53% 66.50% CD19 67.85% 63.26% CD45 67.47% 60.97%

[0079] Table 1

[0080] Refer to Table 1 and Figure 3 This experiment aims to evaluate the importance of each parameter by training and validating our method using only one common parameter. (See table below.) Figure 3 As shown, our method can still achieve high and comparable performance using only one parameter, just like CD117 and SSC-A.

[0081] Table 2 shows the experimental results of the performance of sequential feature selection according to embodiments of the present disclosure.

[0082] 10 FSC-A, FSC-H, SSC-A, CD7, CD19, CD34, CD45, CD56, CD117, HLA-DR - 99.94% 9 FSC-A, SSC-A, CD7, CD19, CD34, CD45, CD56, CD117, HLA-DR FSC-H 99.90% 8 FSC-A, SSC-A, CD19, CD34, CD45, CD56, CD117, HLA-DR CD7 99.90% 7 FSC-A, SSC-A, CD19, CD45, CD56, CD117, HLA-DR CD34 99.79% 6 FSC-A, SSC-A, CD19, CD45, CD117, HLA-DR CD56 99.92% 5 SSC-A, CD19, CD45, CD117, HLA-DR FSC-A 99.95% 4 SSC-A, CD19, CD117, HLA-DR CD45 99.95% 3 SSC-A, CD19, CD117 HLA-DR 99.40% 2 SSC-A, CD117 CD19 97.93% 1 CD117 SSC-A 97.60%

[0083] Table 2

[0084] Referring to Table 2, this experiment removed parameters one by one to evaluate the minimum number of parameters required for the method. The parameters to be discarded for each parameter number were determined by comparing which discarded parameters produced the best results. The table below lists the optimal parameter set for each parameter number. Performance remained unchanged or slightly decreased due to parameter discarding. The method achieved the best AUC with only five and four parameters. SSC-A and CD117 were the last two parameters retained, demonstrating their importance to our method for classifying AML and non-tumor diseases. This result is similar to that of previous single-parameter feature selection experiments.

[0085] In step S300, a classifier is developed through supervised machine learning. The details of developing a classifier through supervised machine learning are described below.

[0086] In some embodiments, overlapping parameters are determined by identifying overlapping parameters in different tests that describe the same biomarker. For example, by identifying the information labeled on the overlapping parameters, information such as name, type, application target, spectral or biochemical characteristics, and overlapping parameters describing the same biomarker are identified.

[0087] In some embodiments, preprocessing of the overlap parameters includes rearranging and normalizing the overlap parameters. In some embodiments, rearranging the overlap parameters includes parameter alignment. In some embodiments, because some common markers may not be placed in the same tube or measured in the same fluorescence channel, the alignment process involves rearranging the data frames so that each data frame has parameters arranged in the same order. For example, before alignment, the channels used to measure CD45 in each panel are located in different row positions. After alignment, the row order of these two parameters is the same so that the data from the two panels can be combined for subsequent model training.

[0088] In some embodiments, normalization includes z-score normalization. In some embodiments, z-score normalization is defined as... .

[0089] The mean and standard deviation for each sample are calculated based on the fluorescence intensity of all cells and channels. σ is a small value added to the denominator to maintain numerical stability.

[0090] In some embodiments, the preprocessing of the overlap parameters further includes extracting the distribution and embedding phenotypic features into the phenotypic representation.

[0091] In some embodiments, the distribution is a high-dimensional cellular distribution.

[0092] In some embodiments, the high-dimensional cell distribution is determined using a Gaussian mixture model (GMM). In some embodiments, a deep neural network approach is employed to generate the sample representation.

[0093] In some embodiments, it exhibits a high-dimensional phenotypic appearance.

[0094] In some embodiments, embedding phenotypic features into the distribution is performed through Fisher vectorization, which embeds phenotypic features into high-dimensional phenotypic representation vectors at the sample level.

[0095] In some embodiments, the preprocessing of the overlap parameters further includes inputting the vector along with the corresponding diagnostics into the support vector machine.

[0096] In some embodiments, the development of a classifier includes data training and sample classification.

[0097] In some embodiments, the method is used for counting cells, sorting cells, determining cell function, identifying cell characteristics, detecting microorganisms (e.g., bacteria, fungi, or yeast), searching for biomarkers, or assisting in the diagnosis and potential treatment of blood and bone marrow cancers. In some embodiments, the method is also used for assessing treatment response, monitoring residual disease, identifying biomarkers, cell quality control, etc.

[0098] In some embodiments, the method further includes visualizing the classifier after its development.

[0099] In some embodiments, the overlap parameter is selected from the group of light scattering properties and fluorescent labels.

[0100] Figure 4 This illustration shows an example diagram of developing and visualizing panel-independent sample classification using cell count data according to an embodiment of this disclosure. The cell count data used in this disclosure comes from flow cytometry counts as part of a clinical evaluation of bone marrow samples from patients assessed for potential blood disorders from various institutions.

[0101] Crossover data or data from multiple experiments refer to data from two or more experiments. Specifically, crossover data or data from multiple experiments refer to data obtained after treatment using different types of flow cytometers, different panels, different reagent kits, different treatment methods, different pretreatment methods, or other methods that differ from each other.

[0102] Example 1: Cross-panel AML vs. non-tumor classification across two panels.

[0103] In some implementations, 103 flow cytometry standards (FCS) records from bone marrow samples were used, from two different sources shown in Table 3.

[0104]

[0105] Table 3

[0106] Data from Roswell Park Comprehensive Cancer Center (RPCCC) were obtained using a Beckman Coulter Navios EX cell analyzer and measured using a ClearLLab10C panel, as shown in Table 4a; and data from the University of Pittsburgh Medical Center (UPMC) were obtained using a BD FACSCantoII and measured using a diagnostic panel developed by UPMC, as shown in Table 4b.

[0107] Panel composition

[0108] a) ClearLLab10C

[0109] B Kappa Lamda CD10 CD5 CD200 CD34 CD38 CD20 CD19 CD45 T TCRγδ CD4 CD2 CD56 CD5 CD34 CD7 CD8 CD3 CD45 M1 CD16 CD7 CD10 CD13 CD64 CD34 CD14 HLA-DR CD11b CD45 M2 CD15 CD123 CD117 CD13 CD33 CD34 CD38 HLA-DR CD19 CD45

[0110] b) Diagnostic panel developed by UPMC

[0111] 1 CD36 CD123 CD64 CD33 CD34 CD14 HLA-DR CD45 2 CD15 CD56 CD117 CD13 CD11b CD16 HLA-DR CD45 3 Kappa Lambda CD5 CD19 CD10 CD38 CD20 CD45 4 CD16 & 57 CD7 CD4 CD3 CD56 CD8 CD2 CD45 5 CD7 CD13&33 CD19 - CD56 - - -

[0112] Table 4

[0113] Figure 5 This disclosure provides a detailed example architecture for developing and visualizing panel-independent sample classification using cell count data, according to embodiments of the present disclosure. In some embodiments, the classification model development process of the present disclosure is as follows: Figure 5 As shown. In some embodiments, we utilize flow cytometry list pattern data, treating each light scattering characteristic and fluorescent label as a different parameter. In some embodiments, data preprocessing involves parameter alignment and z-score normalization. In some embodiments, a Gaussian mixture model (GMM) is used to extract the high-dimensional cell distribution, followed by Fisher vectorization to embed phenotypic features into high-dimensional phenotypic representation vectors at the specimen level. In some embodiments, these vectors, along with their corresponding diagnoses, are input into a support vector machine (SVM) for AML and non-tumor classification.

[0114] In some embodiments, we identify 25 fluorescence parameters that are commonly measured in the RPCCC and UPMC panels for further data processing, including parameter alignment and z-score normalization, to develop a cross-panel classification model (Model D) and use three different scenarios as controls: Model A - trained using all ClearLLab10C panel parameters on the RPCCC group; Model B - trained using these 25 overlapping parameters between the ClearLLab10C and UPMC panels on the RPCCC group; Model C - trained using these 25 overlapping parameters between the ClearLLab10C and UPMC panels on the UPMC group; and Model D - trained using a combination of these 25 overlapping parameters between the ClearLLab10C and UPMC panels on the RPCCC and UPMC groups.

[0115] The classification performance of these models is shown in Table 5. Model A, trained using all ClearLLab 10C panel parameters on the RPCCC group, achieves 100% AUC. However, when only the overlapping parameters between the ClearLLab 10C and UPMC panels are used, the model's AUC drops slightly to 99.86% (Model B). The cross-panel classifier (Model D), trained on the combined RPCCC and UPMC datasets using 25 overlapping parameters, achieves 99.06% AUC.

[0116] Sample-level classification model performance

[0117] parameter All markers measured in the ClearLLab10C panel 25 biomarkers commonly measured in the diagnostic panel developed by ClearLLab10C and UPMC Dataset RPCCC Only RPCCC Only UPMC only RPCCC+UPMC Sample size 53 53 50 103 threshold 0.542 0.27 0.27 0.27 ACC 100.00% 98.10% 96.00% 96.10% AUC 100.00% 99.86% 99.52% 99.06% Sensitivity 100.00% 100.00% 100.00% 100.00% Specificity 100.00% 92.31% 84.00% 92.16% PPV 100.00% 93.10% 86.20% 92.86% NPV 100.00% 100.00% 100.00% 100.00% RPR 0.00% 7.69% 16.00% 7.84% FNR 0.00% 0.00% 0.00% 0.00%

[0118] Abbreviations: RPCCC: Roswell Park Comprehensive Cancer Center; AML: Acute Myeloid Leukemia; ACC: Accuracy, defined as the agreement between model predictions and sample diagnoses; AUC: Area under the ROC curve.

[0119] The 25 markers measured together in the ClearLLab10C and UPMC panels are: CD2, CD3, CD4, CD5, CD7, CD8, CD10, CD11b, CD13, CD14, CD15, CD16, CD19, CD20, CD33, CD34, CD38, CD45, CD56, CD64, CD117, CD123, HLA-DR, Kappa, and Lamda.

[0120] *Note: In the context of this study, “normal” (non-tumor) samples from RPCCC include individuals with lymphoma but without bone marrow involvement, bone marrow donors, and residual specimens from hip replacement surgery. Similarly, “normal” samples from UPMC refer to patients with cytopenia.

[0121] Table 5

[0122] Figure 6A and 6B This displays a visualization of exemplary cross-panel AML and non-tumor classification samples according to embodiments of this disclosure. Figure 7 This shows exemplary results of using cell count data to identify the minimum parameters required for developing panel-independent sample classification according to embodiments of this disclosure.

[0123] In some embodiments, dimensionality reduction is applied to the phenotypic vector at each sample level to visualize the samples from the training model. Each sample is represented as a point on a 3D graph, while color represents the actual diagnosis. Figure 6A and 6B As shown in the figure, the visualization of the model D samples shows that AML and non-tumor samples in the RPCCC and UPMC datasets are well separated.

[0124] In some embodiments, to evaluate the importance of each parameter and its interaction with other parameters, and to determine whether the number of parameters can be reduced while maintaining performance, we conduct feature selection experiments. In these experiments, we pair the parameter with the highest accuracy with each of the remaining parameters to determine the combination that produces the highest accuracy. Subsequently, we individually add the remaining parameters to determine the next combination with the highest accuracy. Figure 7 As shown, the model developed using CD117, CD123, CD15, CD34 and combinations of optical parameters achieves equivalent performance with an accuracy of 97.01%, while the accuracy of model D is 96.01% (Table 5).

[0125] Example 2: Cross-panel AML versus non-tumor classification across three panels.

[0126] In some embodiments, this disclosure is discussed below in conjunction with examples of cross-panel AML across three panels and non-tumor classification.

[0127] In this example, there are 139 flow cytometry standard (FCS) records among peripheral blood and bone marrow samples obtained from three different sources as shown in Table 6.

[0128] Basic features

[0129] FC record total 53 50 36 139 Flow panel ClearLLab10C UPMC Diagnostic Euroflow AML / MDS instrument FASCanto 0 50 0 50 Navios EX 53 0 0 53 FACSLyric 0 0 36 36 sample PB 5 0 0 5 BM 48 50 36 134 diagnosis AML 27 25 17 69 Non-tumor 26 25 19 70

[0130] Table 6

[0131] Table 7 shows the composition of the data panel at the sample level; Table 8 shows the composition of the data panel at the cell level; Table 9 shows the training set for cell-level classification.

[0132] In some embodiments, Roswell Park Comprehensive Cancer Center (RPCCC) data were obtained using a Beckman Coulter Navios EX cell analyzer (Navios EX) and measured using a ClearLLab10C panel, as described in the first column of Table 7; data from the University of Pittsburgh Medical Center (UPMC) were obtained using a BD FACSCantoII and measured using a diagnostic panel developed by UPMC, as shown in the second column of Table 7; and data from the National Taiwan University Cancer Center (TUCC) were obtained using a Beckman Coulter FACSLyric flow cytometer and measured using a Euroflow AML / MDS panel, as described in the third column of Table 7.

[0133]

[0134] Table 7

[0135]

[0136] Table 8

[0137]

[0138] Table 9

[0139] In some embodiments, to prevent bias towards the majority class, we balance the training set size for each class across the three datasets.

[0140] In some embodiments, we identify seven fluorescence parameters commonly measured across all RPCCC, UPMC, and TUCC panels for further data processing, including parameter alignment and z-score normalization, to develop a cross-panel classification model (Model D). In some embodiments, because common markers may not be placed in the same tube or measured using the same fluorescence channel, the alignment process involves rearranging data frames so that each frame has parameters arranged in the same order. For example, before alignment, CD34 measured in the ClearLLab10C panel, CD34 measured in Tube 1 of the UPMC panel, and CD34 measured in the Euroflow AML / MDS panel might be located in different row positions. After alignment, the row order of these three parameters is ensured to be identical so that data from the three panels can be combined for model training.

[0141] In some embodiments, z-score normalization is defined as .

[0142] In some embodiments, the mean and standard deviation for each sample are calculated based on the fluorescence intensity of all cells and channels. 𝜖 is a small value added to the denominator to maintain numerical stability.

[0143] In some embodiments, the model (Model D) is trained by aligning the optical parameters (FSC-A, FSC-H, SSC-A) with the parameters measured in all three panels: CD34, CD45, CD117, HLA-DR, to ensure that each data frame has the same parameter order, and these aligned data undergo normalization, sample performance, and classifier training steps.

[0144] In some embodiments, a Gaussian mixture model (GMM) is used to capture the complex cell distribution. Then, in some embodiments, the Fisher gradient vectorization method is applied to embed phenotypic features into the resulting sample-level high-dimensional phenotypic representation based on the learned probability distribution. Each vector is combined with a classification label such as AML or non-tumor. In some embodiments, a supervised machine learning algorithm (e.g., SVM) is then used to train the AML and non-tumor classifiers.

[0145] In some embodiments, three different scenarios are used as controls: Model A - trained using all ClearLab10C panel parameters on the RPCCC group; Model B - trained using all UPMC panel parameters on the UPMC group; Model C - trained using all Euroflow AML / MDS panel parameters on the TUCC group; Model D - trained using 7 overlapping parameters between ClearLab10C, UPMC, and Euroflow AML / MDS panels on the RPCCC group; Model E - trained using 7 overlapping parameters on the UPMC group; Model F - trained using 7 overlapping parameter pairs on the TUCC group; and Model G - trained using 7 overlapping parameters on a combined RPCCC, UPMC, and TUCC group.

[0146] In some embodiments, the classification performance of these models is shown in Table 10. Model A, trained using all ClearLLab10C panel parameters on the RPCCC group, achieves 100% AUC. However, when only the overlapping parameters between the three panels are used, the model's AUC drops slightly to 92.88% (Model D). The cross-panel classifier (Model G), trained on a combined RPCCC, UPMC, and TUCC dataset using 7 overlapping parameters, achieves 99.68% AUC.

[0147]

[0148] Table 10

[0149] Example 3: Cross-panel AML versus non-tumor classification across five panels.

[0150] In some embodiments, this disclosure is discussed below in conjunction with examples of cross-panel AML across five panels and non-tumor classification.

[0151] In this example, there are 220 flow cytometry standard (FCS) records from bone marrow samples obtained from five different sources as shown in Table 11.

[0152]

[0153] Table 11

[0154] Data from each source was measured and obtained using different panels and instrument sets, as listed in Table 11. The detailed composition of the panels is shown in Table 12.

[0155]

[0156]

[0157] Table 12

[0158] Figure 8 This illustrates an example architecture for developing and visualizing panel-independent sample classification using cell count data according to embodiments of this disclosure. The classification model development process of this example is as follows: Figure 8 As shown. In some embodiments, flow cytometry list pattern data is used to process each light scattering characteristic and fluorescent label as a different parameter. In some embodiments, data preprocessing involves parameter alignment, compensation, downsampling, and z-score normalization. In some embodiments, a Gaussian mixture model (GMM) is used to extract the high-dimensional cell distribution, and then Fisher vectorization is used to embed the phenotypic features into the high-dimensional phenotype.

[0159] In some embodiments, Figure 8 The architecture pipeline shown is as follows: (1) Select and rearrange the general parameters between different panels; (2) Preprocess the data, including compensation, downsampling (scale = 0.2) and z-score normalization for each sample; (3) Transform the data of each sample into a one-dimensional representation vector using a Gaussian mixture model (GMM) and Fisher vectors; (4) Apply triple cross-validation to train and validate the support vector machine (SVM) for binary classification.

[0160] In some embodiments, 10 common optical and fluorescence parameters measured across all panels are identified for further data processing to develop a cross-panel classification model (Model K). In some embodiments, because common markers may not be placed in the same tube or measured within the same fluorescence channel, the alignment process involves rearranging data frames so that each data frame has parameters arranged in the same order. For example, before alignment, the channels used to measure CD34 in the ClearLLab10C and UPMC panels are located in different row positions. After alignment, it is ensured that the row order of these two parameters is the same so that data from both panels can be combined for model training.

[0161] In some embodiments, compensation is a standard procedure in flow cytometry to address the effects of spillover. Spillover is a common occurrence in conventional cell analyzers, where one channel can simultaneously collect light emitted by paired markers and markers from other channels, leading to inaccurate light intensity measurements. To address this issue, a compensation matrix is ​​used, obtained from the measurement standard beads, and defined as (x) ̂ = x[c]^(-1), where c is the compensation matrix, x is the intensity matrix of the channel with spillover, and x is the intensity matrix of the channel without spillover.

[0162] In some embodiments, during the downsampling step, the number of cells is reduced to a target proportion of the original number. This is accomplished by randomly sampling cells from the data.

[0163] In some embodiments, z-score normalization is defined as .

[0164] The mean and standard deviation for each sample are calculated based on the fluorescence intensity of all cells and channels. σ is a small value added to the denominator to maintain numerical stability.

[0165] This model (Model K) is trained by adjusting the optical and fluorescence parameters measured in all panels, including FSC-A, FSC-H, SSC-A, CD7, CD19, CD34, CD45, CD56, CD117, and HLA-DR. This process is performed for each sample to ensure that each data block has the same parameter order, and these aligned data undergo compensation, downsampling, normalization, sample representation, and classifier training steps.

[0166] In some embodiments, a Gaussian mixture model (GMM) is used to capture the complex cell distribution. Then, in some embodiments, the Fisher gradient vectorization method is applied to embed phenotypic features into the resulting sample-level high-dimensional phenotypic representation based on the learned probability distribution. Each vector is combined with a classification label such as AML or non-tumor. Supervised machine learning algorithms such as SVM are then used to train the AML and non-tumor classifiers.

[0167] In some embodiments, ten different scenarios are used as controls: Models A to E are trained using separate datasets and site-specific parameters. Models F to J are also trained using separate datasets, but only using 10 common parameters measured across the panel. Model K is trained using all data from five datasets and 10 common parameters.

[0168] In some embodiments, the classification performance of these models is shown in Table 3. By comparing the results of models trained on the same dataset using different parameter sets, it was found that the performance of all models decreased or even lost points. For example, models D and I were both trained on the UPMC dataset, but model D slightly outperformed model I on all indicators; models E and J were also trained on the same individual dataset TUH, but both models achieved 100% in all aspects. Model K was trained on all datasets and on commonly used measurement parameters in four different panels. It achieved excellent performance, outperforming most other models, with only two samples misclassified in 220 cases.

[0169]

[0170] Table 13

[0171] Figure 9A and 9B This displays an example visualization of samples across panel AML and non-tumor classifications according to embodiments of this disclosure. Dimensionality reduction is performed on the phenotypic vector at each sample level to visualize the samples from the training model. Each sample is represented as a point on a 3D graph, with color representing the actual diagnosis and shape representing the source. Figure 9A and 9B This displays a visualization of sample results across panels in AML and non-tumor classifications. (Example) Figure 9A and 9B As shown in the model D sample visualization, AML and non-tumor samples are well separated in all datasets.

[0172] Example 4: Classification of transpanel lymphatic abnormalities and no abnormalities in two panels.

[0173] In some embodiments, this disclosure is discussed below in conjunction with examples of cross-panel lymphatic abnormality and no abnormality classification across two panels.

[0174] In this example, there are 397 flow cytometry standard (FCS) records from bone marrow samples obtained from two different sources as shown in Table 14.

[0175]

[0176] Table 14

[0177] Data from each source was measured and obtained using different panels and instrument sets, as listed in Table 14. The detailed composition of the panels is shown in Table 15.

[0178]

[0179] Table 15

[0180] Figure 10 This illustrates an example architecture for developing and visualizing panel-independent sample classification using cell count data according to embodiments of this disclosure. The classification model development process of this example is as follows: Figure 10 As shown. In some embodiments, flow cytometry list pattern data is used to process each light scattering characteristic and fluorescent label as a different parameter. In some embodiments, data preprocessing involves parameter alignment, compensation, downsampling, and z-score normalization. In some embodiments, a Gaussian mixture model (GMM) is used to extract the high-dimensional cell distribution, and then Fisher vectorization is used to embed the phenotypic features into the high-dimensional phenotype.

[0181] In some embodiments, five optical and fluorescence parameters that are commonly measured in both panels are identified for further data processing to develop a cross-panel classification model. In some embodiments, because common markers may not be placed in the same tube or measured in the same fluorescence channel, the alignment process involves rearranging the data frames so that each data frame has parameters arranged in the same order. For example, before alignment, the channels used to measure CD45 in each panel are located in different row positions. After alignment, it is ensured that the row order of these two parameters is the same so that the data from both panels can be combined for model training.

[0182] In some embodiments, compensation is a standard procedure in flow cytometry to address the effects of spillover. Spillover is a common occurrence in conventional cell analyzers, where one channel can simultaneously collect light emitted by paired markers and markers from other channels, leading to inaccurate light intensity measurements. To address this issue, a compensation matrix is ​​used, obtained through measurement standard beads, defined as x̂ = xc^(-1), where c is the compensation matrix, x is the intensity matrix of the channel with spillover, and x̂ is the intensity matrix of the channel without spillover.

[0183] In some embodiments, during the downsampling step, the number of cells is reduced to a target proportion of the original number. This is accomplished by randomly sampling cells from the data.

[0184] In some embodiments, z-score normalization is defined as .

[0185] The mean and standard deviation for each sample are calculated based on the fluorescence intensity of all cells and channels. σ is a small value added to the denominator to maintain numerical stability.

[0186] In some embodiments, the model is trained by aligning optical and fluorescence parameters measured in two panels, including FSC-H, SSC-H, CD3, CD5, CD38, and CD45. This operation is performed for each sample to ensure that each data block has the same parameter order, and these aligned data undergo compensation, downsampling, normalization, sample representation, and classifier training steps.

[0187] In some embodiments, a Gaussian mixture model (GMM) is used to capture the complex cell distribution. Then, in some embodiments, the Fisher gradient vectorization method is applied to embed phenotypic features into the resulting sample-level high-dimensional phenotypic representation based on the learned probability distribution. Each vector is accompanied by a classification label, such as lymph node abnormality or absence of abnormality. Supervised machine learning algorithms such as SVM are then used to train a lymph node abnormality and absence of abnormality classifiers.

[0188] In some embodiments, the classification performance of these models is shown in Table 16. This model achieves good performance using only five parameters, which is much less than the original panel design.

[0189]

[0190] Table 16

[0191] Figure 11 This displays a visualization of samples in an example cross-panel lymphatic abnormality and no-abnormality classification according to an embodiment of this disclosure.

[0192] In some embodiments, dimensionality reduction is applied to the phenotypic vector at each sample level to visualize the samples from the training model. Each sample is represented as a point on a 3D graph, with color representing the actual diagnosis and shape representing the source. Figure 11 This displays a sample visualization of examples categorizing lymphatic abnormalities and normalities across panels. For example... Figure 11 As shown in the sample visualization, the abnormal lymphocyte samples and the normal samples in all datasets can be well separated.

[0193] Figure 12A block diagram illustrating an example of an artificial intelligence (AI) system 500 according to an embodiment of this disclosure is provided, wherein at least some of the operations described herein can be implemented. As shown, the AI ​​system 500 may include a set of layers that conceptually organize elements within an example network topology of the AI ​​system architecture to implement a specific AI model 530. Generally, the AI ​​model 530 is a computer-executable program implemented by the AI ​​system 500 for analyzing data to make predictions. Information can pass through each layer of the AI ​​system 500 to produce output for the AI ​​model 530. These layers may include a data layer 502, a structure layer 504, a model layer 506, and an application layer 508. The algorithm 516 of the structure layer 504 and the model structure 520 and model parameters 522 of the model layer 506 together constitute the example AI model 530. In some embodiments, an optimizer 526, a loss function engine 524, and a regularization engine 528 are used to refine and optimize the AI ​​model 530, and the data layer 502 provides resources and support for the application layer 508 to apply the AI ​​model 530.

[0194] Data layer 502 prepares data for AI model 530 as the foundation for AI system 500. As shown in the figure, data layer 502 may include two sub-layers: hardware platform 510 and one or more software libraries 512. Hardware platform 510 may be designed to execute the operations of AI model 530 and includes storage, memory, logic, and network computing resources, such as... Figure 10 Related resources. Hardware platform 510 can use one or more servers to process large amounts of data. Servers can perform backend operations such as matrix computation, parallel computing, machine learning (ML) training, etc. Examples of servers used by hardware platform 510 include central processing units (CPUs) and graphics processing units (GPUs). CPUs are electronic circuits designed to execute computer program instructions, such as arithmetic, logic, control, and input / output (I / O) operations, which can be implemented on integrated circuit (IC) microprocessors. GPUs were originally designed for graphics operations and output, but due to their large computing and memory resources, they can be used for AI applications. GPUs use parallel architectures, which generally make them more efficient than CPUs. In some cases, hardware platform 510 may include Infrastructure as a Service (IaaS) resources, i.e., computing resources (such as servers, memory, etc.) provided by cloud service providers. Hardware platform 510 may also include computer memory for storing data about AI model 530, applications of AI model 530, and training data for AI model 530. Computer memory can be a form of random access memory (RAM), such as dynamic RAM, static RAM, and non-volatile RAM.

[0195] Software library 512 can be considered a set of data and program code, including executable files, used to control the computing resources of hardware platform 510. The program code may include low-level raw languages ​​that form the basis of one or more low-level programming languages, enabling the server of hardware platform 510 to perform specific operations using these low-level raw languages. If low-level programming languages ​​are present, they do not need to be abstracted too much from the instruction set architecture of the computing resources, allowing them to run quickly with a small memory footprint. Examples of software libraries 512 that may be included in AI system 500 include Intel Math Kernel Library, Nvidia cuDNN, Eigen, and Open BLAS.

[0196] The structural layer 504 may include an ML framework 514 and algorithms 516. The ML framework 514 can be considered an interface, library, or tool that allows users to build and deploy AI models 530. The ML framework 514 may include open-source libraries, application programming interfaces (APIs), gradient boosting libraries, integration methods, and / or deep learning toolkits that work with the layers of the AI ​​system to facilitate the development of the AI ​​model 530. For example, the ML framework 514 may be distributed across multiple resources in the hardware platform 510 for the application or training of the AI ​​model 530. The ML framework 514 may also include pre-built components that implement and train the AI ​​model 530, allowing users to build and train the AI ​​model 530 using pre-built functions and categories. Therefore, the ML framework 514 can be used to facilitate the data engineering, development, hyperparameter tuning, testing, and training of the AI ​​model 530. Examples of ML frameworks 514 that can be used in the AI ​​system 500 include TensorFlow, PyTorch, Scikit-Learn, Keras, Cafffe, LightGBM, Random Forest, and Amazon Web Services.

[0197] Algorithm 516 can be a set of organized computer-operable algorithms for producing output data from a set of input data, and can be described using pseudocode. Algorithm 516 may include complex program code that allows computational resources to learn from new input data and create new / modified outputs based on what has been learned. In some implementations, algorithm 516 can be used to build AI model 530 by training on the computational resources of running hardware platform 510. This training allows algorithm 516 to make predictions or decisions without explicit programming. Once training is complete, algorithm 516 can be run on computational resources as part of AI model 530 to make predictions or decisions, improve computational resource efficiency, or perform tasks. Algorithm 516 can be trained using supervised learning, unsupervised learning, semi-supervised learning, and / or reinforcement learning.

[0198] Using supervised learning, algorithm 516 can be trained to learn patterns based on labeled training data (e.g., mapping input data to output data). The training data can be labeled by an external user or operator. For example, a user can collect a set of training data, such as by capturing data from sensors, images from a camera, or outputs from a model. In the example implementation, the training data may include the parameters described above. Users can label the training data based on one or more categories and train the AI ​​model 530 by inputting the training data into algorithm 516. The algorithm determines how to label new data based on the labeled training data. Users can conveniently collect, label, and / or input the data through the ML architecture 514. In some cases, users can convert the training data into a set of feature vectors to input into algorithm 516. Once trained, users can test algorithm 516 on new data to determine if it can accurately predict labels for the new data. For example, users can use cross-validation to test the accuracy of algorithm 516, and if the cross-validation result is below an accuracy threshold, algorithm 516 is retrained on new training data.

[0199] Supervised learning can involve classification and / or regression. Classification techniques involve teaching algorithm 516 to identify a new class of observations based on training data, and are used when the input data to algorithm 516 is discrete. In other words, when learning through classification techniques, algorithm 516 receives training data labeled with categories (e.g., classes) and determines the relationship between features observed in the training data and the categories. Once trained, algorithm 516 can classify new data by analyzing the features in the new data that map to the categories. Examples of classification techniques include reinforcement learning, decision tree learning, genetic programming, learned vector quantization, k-nearest neighbor (k-NN) algorithms, and statistical classification.

[0200] Regression techniques involve estimating the relationship between independent and dependent variables and are used when the input data for Algorithm 516 is continuous. Regression techniques can be used to train Algorithm 516 to predict or forecast relationships between variables. To train Algorithm 516 using regression techniques, the user can choose a regression method to estimate the model's parameters. The user collects and labels the training data input to Algorithm 516 so that Algorithm 516 can learn about data characteristics and the relationship between the dependent variables. Once training is complete, Algorithm 516 can predict missing historical data or future outcomes based on the input data. Examples of regression methods include linear regression, multiple linear regression, logistic regression, regression tree analysis, least squares method, and gradient descent. In example implementations, regression techniques can be used, for example, to estimate and impute missing data for machine learning-based preprocessing operations.

[0201] In unsupervised learning, Algorithm 516 learns patterns from unlabeled training data. Specifically, Algorithm 516, after training, can learn hidden patterns and insights from the input data, which can be used for data exploration or to generate new data. Here, Algorithm 516 has no predefined output, unlike the labels output when Algorithm 516 is trained using supervised learning. In other words, unsupervised learning is used to train Algorithm 516 to find the underlying structure of a set of data, group the data according to similarity, and represent the data in a compressed format.

[0202] Supervised learning can utilize several techniques: clustering, anomaly detection, and techniques for learning latent variable models. Clustering techniques involve grouping data into different clusters containing similar data, while other clusters contain different data. For example, during clustering, potentially similar data may be retained in groups with less or no similarity to another group. Examples of clustering techniques include density-based methods, hierarchical methods, partitioning methods, and grid-based methods. In one example, Algorithm 516 could be trained as a k-means clustering algorithm, which divides n observations into k clusters such that each observation belongs to the cluster whose nearest mean is used as the cluster prototype. Anomaly detection techniques are used to detect rare, previously unseen objects or events in data without prior knowledge of these objects or events. Anomalies can include data that is rare in a set, deviations from other observations, outliers inconsistent with the rest of the data, patterns that do not conform to well-defined normal behavior, and so on. When using anomaly detection techniques, Algorithm 516 can be trained as an Isolation Forest, a Local Anomaly Factor (LOF) algorithm, or a K-Nearest Neighbors (k-NN) algorithm. Latent variable techniques involve associating observable variables with a set of latent variables. These techniques assume that the observable variables are a result of an individual's stance on the latent variables, and that, after controlling for latent variables, the observable variables have nothing in common. Examples of latent variable techniques that Algorithm 516 can use include factor analysis, item response theory, latent profile analysis, and latent category analysis.

[0203] Model structure 520 describes the architecture of AI model 530 of AI system 500. Examples of structures that can be used as model structure 520 include decision trees, support vector machines, regression analysis, Bayesian networks, Gaussian processes, genetic algorithms, and artificial neural networks (or simply neural networks). Model structure 520 may include several structural layers, several nodes (or neurons) in each structural layer, and activation functions for each node. The activation function of each node defines how the node converts received data into data output. A structural layer may contain an input layer for nodes that receive input data and an output layer for nodes that produce output data. Model structure 520 may include one or more hidden node layers between the input and output layers. Model structure 520 may be an artificial neural network (or simply a neural network) that connects the nodes in the structural layers, allowing the nodes to interact with each other. Examples of neural networks include feedforward neural networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), autoencoders, and generative adversarial networks (GANs).

[0204] Model parameters 522 represent the relationships learned during training and can be used to make predictions and decisions based on input data. Model parameters 522 can assign weights and biases to the nodes and connections of model structure 520. For example, when model structure 520 is a neural network, model parameters 522 can weight and bias the nodes in each layer of the neural network, such that the weights determine the strength of the nodes, and the biases determine the threshold of the activation function of each node. Model parameters 522, combined with the activation functions of the nodes, determine how the input data is transformed into the desired output. Model parameters 522 can be determined and / or changed during the training of algorithm 516.

[0205] The loss function engine 524 determines the loss function, which is a metric used to evaluate the performance of the AI ​​model 530 during training. For example, the loss function engine 524 can measure the difference between the predicted output and the actual output of the AI ​​model 530 and use it to guide the optimization of the AI ​​model 530 during training to minimize the loss function. The loss function can be presented through the ML architecture 514, allowing the user to determine whether to retrain or otherwise modify the algorithm 516 if the loss function exceeds a threshold. In some cases, if the loss function exceeds a threshold, the algorithm 516 can be automatically retrained. Examples of loss functions include the binary cross-entropy function, hinged loss function, regression loss functions (e.g., mean squared error, quadratic loss, etc.), mean absolute error function, smoothed mean absolute error function, log-hyperbolic orthogonal loss function, and quantile loss function.

[0206] Optimizer 526 tunes model parameters 522 to minimize the loss function during training of algorithm 516. In other words, optimizer 526 uses the loss function generated by loss function engine 524 as guidance to determine which model parameters will produce the most accurate AI model 530. Examples of optimizers include gradient descent (GD), adaptive gradient algorithm (AdaGrad), adaptive moment estimation (Adam), root mean square propagation (RMSprop), radial basis function (RBF), and finite memory BFGS (L-BFGS). The type of optimizer 526 used can be determined based on the type of model structure 520 and the size of the data available in data layer 502 and computational resources.

[0207] The regularization engine 528 performs regularization operations. Regularization is a technique to prevent the AI ​​model 530 from overfitting or underfitting. Overfitting occurs when the algorithm 516 is too complex and overfits the training data, leading to poor performance of the AI ​​model 530. Underfitting occurs when the algorithm 516 fails to identify even the most basic patterns from the training data, resulting in poor performance on either the training or validation data. The regularization engine 528 can apply one or more regularization techniques to correctly fit the algorithm 516 to the training data, which helps constrain the resulting AI model 530 and improve its general applicability. Examples of regularization techniques include lasso (L1) regularization, ridge (L2) regularization, and elasticity (L1 and L2 regularization).

[0208] Figure 13 This diagram illustrates an example of z-fraction normalization according to an embodiment of this disclosure.

[0209] In some embodiments, z-score normalization is a strategy for normalizing data. This technique scales the values ​​of features to a mean of 0 and a standard deviation of 1.

[0210] z = (x - μ) / σ

[0211] Where x is the original value, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0212] refer to Figure 13 , Figure 13 Show an example where x=60, μ=50, σ=10. After z-fraction normalization, z=(60-50) / 10=1, then μ=0, σ=1.

[0213] In some embodiments, for sample z-score normalization, z-score normalization is performed separately for each channel of the sample. In some embodiments, for location z-score normalization, z-score normalization is performed separately for each channel of the dataset.

[0214] Figure 14This diagram shows an example of density contour lines according to an embodiment of the present disclosure.

[0215] In some embodiments, the probability density function (PDF) is similar to a density map, where a higher value indicates a higher probability that data falls into that region. Density contour lines represent the probability density function in two-dimensional space.

[0216] Figure 15 This diagram shows an example of a classifier according to an embodiment of the present disclosure.

[0217] In some embodiments, for a Support Vector Machine (SVM) classifier, the SVM operates by finding an optimal decision boundary (called a hyperplane) that best separates data points from different classes. This hyperplane is positioned to maximize the distance between it and its nearest data points. In some embodiments, for an Extreme Gradient Boosting (XGBoost) classifier, the XGBoost classifier is an ensemble learning algorithm that uses multiple decision trees for prediction. It employs a strategy called gradient boosting. Each tree focuses on correcting the errors of the previous tree, resulting in a more accurate ensemble model. The final prediction is typically obtained by summing the predictions from each decision tree.

[0218] Figure 16 and Figure 17 Example schematic diagrams showing evaluation metrics according to embodiments of this disclosure.

[0219] In some embodiments, the formula for ACC (accuracy) is as follows:

[0220]

[0221] The model's overall ability to correctly predict both actual positive and actual negative values.

[0222] In some embodiments, the formula for sensitivity = recall rate (true positive rate) is as follows:

[0223]

[0224] The model's ability to correctly predict actual positive results.

[0225] In some embodiments, the formula for specificity (true negative rate) is as follows:

[0226]

[0227] The model's ability to correctly predict actual negative outcomes.

[0228] In some embodiments, the formula is AUC (Area Under the ROC Curve), as shown in the figure below:

[0229] For the true positive rate (TPR):

[0230] Regarding the false positive rate (FPR):

[0231] In some embodiments, the formula for UAR (unweighted average recall rate) is as follows:

[0232]

[0233] The sum of category recall rates divided by the number of categories.

[0234] In some embodiments, the formula for weighted F1 is as follows:

[0235]

[0236] The F1 score combines accuracy and recall using the harmonic mean. The weighted F1 score is calculated by weighting the F1 score of each class by the number of real instances in each class and then taking the average.

[0237] In some embodiments, in cell-level classification, we use macro-averaging to calculate the average performance of ACC and AUC across multiple classes in the classification task. This assigns equal weight to each class, regardless of class imbalance.

[0238] Figure 18 This diagram illustrates an example of GMM and Fisher vectorization according to an embodiment of this disclosure.

[0239] In some embodiments, a Gaussian Mixture Model (GMM) is a probabilistic model that assumes the data is generated by a mixture of several Gaussian distributions, where each Gaussian distribution represents a cluster within the data. GMMs allow for more diverse cluster shapes and provide a probability for each data point to belong to each cluster, thus offering greater flexibility.

[0240] In some embodiments, Fisher vectors utilize Gaussian Mixture Models (GMMs) to extract the statistical properties of features. They calculate the gradient of each original feature based on the mean and variance of each Gaussian distribution. This process helps us understand the impact of features on Fisher vectors. Features with larger gradients have a more significant impact on Fisher vectors, demonstrating their potential importance in characterizing data distributions.

[0241] Figure 19 This diagram illustrates an example of cross-validation according to an embodiment of this disclosure.

[0242] In some embodiments, the cross-validation steps are as follows:

[0243] 1. Divide the data:

[0244] The dataset is divided into k subsets of equal size. These are often referred to as "folds". (We set k=5 in our experiments.)

[0245] 2. Training and Testing:

[0246] This model is trained k times, with each iteration using a k-1x fold for training and the remaining folds for testing. Therefore, in each iteration, different folds are used for testing, while other k-1 folds are used for training.

[0247] 3. Summary and Evaluation:

[0248] After training and testing the model k times, the test results of each iteration are combined to evaluate the model's performance.

[0249] Figure 20 This diagram illustrates an example of compensation according to an embodiment of this disclosure.

[0250] In some embodiments, fluorophores emit light at specific wavelengths, but their emission spectra often overlap, such as PE and FITC. This overlap means that the fluorescence of one fluorophore may be detected in the channel of another fluorophore, resulting in false positives. This is particularly problematic in multicolor experiments. To address this issue, compensation is needed to ensure that the detected fluorescence accurately reflects the fluorophore being measured.

[0251] In some embodiments, the compensation steps are as follows:

[0252] 1. Use flowutils to obtain the overflow matrix from the FCS archive.

[0253] 2. Then, using the overflow matrix as the coefficient matrix, matrix operations are performed on the FCS data to compensate for the spectral overlap effect.

[0254] In some embodiments, the steps of the methods for processing cell count data from multiple tests described herein may be further combined, substituted, repeated and / or modified to produce new embodiments within the scope of this disclosure.

[0255] In some embodiments, each step of the method for processing cell count data from multiple tests described herein can be stored in a computer-readable recording medium, which can be a non-transitory computer-readable recording medium, such as a hard disk, optical disk, magnetic disk, USB flash drive, or a database accessible via a network, but is not limited thereto. When the computer-readable recording medium loads a computer program product stored in it via a computing device and executes the computer program product, any of the methods for processing cell count data from multiple tests described herein can be implemented.

[0256] In some embodiments, the computer program product of this disclosure for processing cell count data from multiple tests may include steps of the method of processing cell count data from multiple tests as disclosed herein, such that a computing device, after loading and running the computer program product, can implement any of the methods of processing cell count data from multiple tests as disclosed herein.

[0257] Although this disclosure contains numerous details, these details should not be construed as limiting the scope of any invention or the scope that may be claimed, but rather as descriptions of specific features of specific embodiments of a particular invention. Certain features described in this disclosure in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments. Furthermore, although features above may be described as functioning in a certain combination, or even initially claimed in this way, in some cases one or more features may be removed from the claimed combination, and the claimed combination may be for sub-combinations or variations thereof.

[0258] Similarly, although operations are depicted in a specific order in the figures, this should not be construed as requiring that these operations must be performed in the specific or sequential order shown, or that all of the shown operations must be performed to achieve the desired result. Furthermore, the separation of the various system components in the embodiments described in this disclosure should not be construed as requiring such separation in all embodiments.

[0259] This disclosure only describes some implementations and examples, and other implementations, enhancements and changes can be made based on the content described and explained in this disclosure.

Claims

1. A method for processing cell count data from multiple tests, comprising: Use flow cytometry to determine the overlapping parameters between multiple tests; Preprocess the overlap parameters; and Develop classifiers through supervised machine learning.

2. The method as described in claim 1, wherein, The overlapping parameters are determined by identifying overlapping parameters in different tests that describe the same biomarker.

3. The method as described in claim 1, wherein, Preprocessing the overlapping parameters includes rearranging the overlapping parameters and normalizing them.

4. The method of claim 3, wherein, Rearranging the overlapping parameters includes performing parameter alignment.

5. The method of claim 3, wherein, The normalization includes z-score normalization.

6. The method of claim 3, wherein, Preprocessing the overlapping parameters also includes extracting the distribution and embedding phenotypic features into the phenotypic representation.

7. The method of claim 6, wherein, The distribution system is a high-dimensional cell distribution.

8. The method of claim 7, wherein, The high-dimensional cell distribution system was determined using a Gaussian mixture model (GMM).

9. The method of claim 6, wherein, The aforementioned performance is a high-dimensional phenotypic performance.

10. The method of claim 6, wherein, Embedding the phenotypic features into the distribution involves using Fisher vectorization to embed the phenotypic features into a high-dimensional phenotypic representation vector at the sample level.

11. The method of claim 6, wherein, The preprocessing of the overlapping parameters also includes inputting the vector along with the corresponding diagnostics into the support vector machine.

12. The method of claim 1, wherein, Developing the classifier involves data training and sample classification.

13. The method of claim 1, wherein, The method is used to count cells, sort cells, determine cell function, determine cell characteristics, or detect microorganisms such as bacteria, fungi or yeast, or to find biomarkers, or to assist in the diagnosis and potential treatment of blood and bone marrow cancers.

14. The method of claim 1, wherein, The development of the classifier also includes visualization.

15. The method of claim 1, wherein, The overlap parameters are selected from a group that includes light scattering properties and fluorescent labels.

16. A system for processing cell count data from multiple tests, wherein an applicable signal is connected to one or more cell count data providing devices to receive the cell count data from the cell count data providing devices, the system comprising: Storage module; and An automatic sorting module is configured to connect to the storage module via a signal. in, Multiple program codes are stored in the storage module, and Wherein, after the automatic classification module executes a plurality of program codes stored in the storage module, the automatic classification module performs the steps of the method for processing cell count data from a plurality of tests as described in any one of claims 1 to 15.

Images