Computer-implemented method and system for determining diseases affecting blood cell morphological characteristics and cytoplasmic complexity - Patent Application 20070122997

Abstract

1. A computer-implemented method for determining diseases affecting morphological characteristics and cytoplasmic complexity of blood cells, the method comprising: a) obtaining blood parameters from at least one patient, the blood parameters comprising at least two measured characteristics of single cells of multiple blood pictures, the single cells comprising white blood cells, the measured characteristics of the white blood cells comprising size, cytoplasmic granularity, and nuclear structure; b) generating at least one scatter plot, each axis of the scatter plot comprising a different measured characteristic of the single cells; c) determining at least one cluster in the at least one scatter plot, the cluster comprising subpopulations of white blood cells, the subpopulations comprising monocytes, lymphocytes, basophils, neutrophils, and eosinophils; d) integrating the characteristic elements of the determined clusters to generate a one-dimensional global vector, wherein the arrangement of the characteristic elements in the global vector includes an arrangement according to the associated cluster; e) performing dimensionality reduction of the global vector; f) diagnosing at least one disease by an ensemble, wherein the ensemble includes at least one machine learning model and at least one deep learning model, wherein the at least one machine learning model receives at least one reduced global vector as an input variable and the at least one deep learning model receives at least one scatter plot image as an input variable; and g) automatically generating a report including at least one result regarding the diagnosing of the at least one disease.

Description

[Technical Field]

[0001] The present invention relates to clinical laboratory diagnostics and describes a computer-implemented method for determining diseases that affect the morphological characteristics and cytoplasmic complexity of blood cells.

[0002] Technical background Clinical blood analysis is one of the most accessible diagnostic methods in clinical medicine, as it reflects the systemic pathological processes in the human body based on the quantitative evaluation of cellular composition and blood morphology. Circulating blood cells constantly infiltrate almost all tissues in the human body at a high flow rate, and their states of maturation, activation, proliferation, and senescence reflect the current pathophysiological state.

[0003] Routine measurements of whole blood counts, including the measurement of single-cell characteristics among tens of thousands of blood cells, provide a comprehensive overview of these pathophysiological conditions. Every routine blood count provides high-dimensional single-cell information on white blood cells. Various morphological characteristics, such as size, complexity, lobulation, and granularity, are measured by hematology analyzers, allowing differentiation of five subpopulations of mature white blood cells: neutrophils, lymphocytes, monocytes, eosinophils, and basophils (Takubo T, Tatsumi N. Further evolution and leukocyte differential using an automated blood cell counter. Rinsho Byori. 1995 Sep;43(9):925-30. Japanese. PMID: 7474456.).

[0004] Currently, clinical decisions are based solely on a few derived statistics. Because of the overwhelming amount of information available from a patient's blood profile, physicians often are unable to fully interpret it in the context of the clinical situation. (Chaudhury A, Noiret L, Higgins JM. White blood cell population dynamics for risk stratification of acute coronary syndrome. Proc Natl Acad Sci US A. 2017 Nov 14;114(46):12344-12349. doi: 10.1073 / pnas.1709228114. Epub 2017 Oct 27. PMID: 29087321; PMCID: PMC5699055.)

[0005] However, early attempts to detect infections early by determining immature granulocytes and to determine the prognosis of some malignancies by counting white blood cells with atypical characteristics have recognized the great potential of the complete blood picture (Statland BE, Winkel P, Harris SC, Burdsall MJ, Saunders AM. Evaluation of biologic sources of variation of leukocyte counts and other hematologic quantities using very precise automated analyzers. Am J Clin Pathol. 1978 Jan;69(1):48-54. doi: 10.1093 / ajcp / 69.1.48. PMID: 563672.).

[0006] While these efforts have met with limited success, they suggest potential for improved clinical decision support. (Gijsberts CM, den Ruijter HM, de Kleijn DPV, Huisman A, Ten Berg MJ, van Wijk RHA, Asselbergs FW, Voskuil M, Pasterkamp G, van Solinge WW, Hoefer IE. Hematological Parameters Improve Prediction of Mortality and Secondary Adverse Events in Coronary Angiography Patients: A Longitudinal Cohort Study. Medicine(Baltimore). 2015 Nov;94(45):e1992. doi: 10.1097 / MD.0000000000001992. PMID: 26559287; PMCID: PMC4912281.)

[0007] The following paper examines the diagnosis of acute coronary syndrome (ACS) using white blood cell analysis with a statistical and mathematical approach (Chaudhury A., Noiret L., Higgins J. White blood cell population dynamics for risk stratification of acute coronary syndrome. Proceedings of the National Academy of Sciences, 2017;114(46), S.12344-12349).

[0008] As the amount of digital data provided by analytical systems increases, there is a growing possibility of applying machine learning methods from the field of artificial intelligence (AI) to improve the validity of the diagnostic information available for the benefit of patients and healthcare providers (Pieszko K., Hiczkiewicz J., Budzianowski P., Rzezniczak J., Budzianowski J., Blaszczynski J., Slowinski R., Burchardt, P. Machine-learned models using hematological inflammation markers in the prediction of short-term acute coronary syndrome outcomes. Journal of Translational Medicine, 2018;16(1):334).

[0009] overview In order to improve the quality of medical care for patients, it can be considered a challenge to propose alternative methods for determining diseases, especially those that affect the morphological characteristics and cytoplasmic complexity of blood cells.

[0010] This problem is solved by a computer-implemented method having the features of independent claim 1. Advantageous embodiments and developments can be taken from the dependent claims and the following description.

[0011] 1. A computer-implemented method for determining diseases affecting the morphological characteristics and cytoplasmic complexity of blood cells, the method comprising: a) obtaining blood parameters from at least one patient, the blood parameters comprising at least two measured characteristics of single cells from multiple blood counts, the single cells comprising white blood cells, the measured characteristics of the white blood cells comprising size, cytoplasmic granularity, and nuclear structure; b) generating at least one scatter plot, each axis of the scatter plot comprising a different measured characteristic of the single cell; c) determining at least one cluster in the at least one scatter plot, the cluster comprising a subpopulation of white blood cells, the subpopulation comprising monocytes, lymphocytes, basophils, neutrophils, and eosinophils; d) determining at least one cluster in the at least one scatter plot, the cluster comprising a subpopulation of white blood cells, the subpopulation comprising monocytes, lymphocytes, basophils, neutrophils, and eosinophils; A method is proposed that includes the steps of: integrating the characteristic elements of the determined clusters to generate a one-dimensional global vector, wherein the arrangement of the characteristic elements in the global vector includes an arrangement according to the associated cluster; e) performing dimensionality reduction of the global vector; f) determining at least one disease by an ensemble, wherein the ensemble includes at least one machine learning model and at least one deep learning model, wherein the at least one machine learning model receives at least one reduced global vector as an input variable and the at least one deep learning model receives at least one scatter plot image as an input variable; and g) automatically generating a report including at least one result regarding the determination of the at least one disease.

[0012] The implementation of the method according to the invention is further described below.

[0013] A patient is admitted to the emergency room of a medical institution. An acute coronary syndrome (ACS) is suspected, which affects the morphological characteristics and cytoplasmic complexity of blood cells. First, venous whole blood is collected from the cubital vein using a 4 ml blood collection vacuum system, for example, into a Vacutest tube (KIMA, Italy), and for example, 7.2 mg of K3EDTA is applied to the inner surface of the tube wall. This sample is then used to determine the disease using the method according to the present invention. Here, sampling is not part of the method.

[0014] After blood collection, the tube is inverted and rotated horizontally and vertically for 30 seconds to mix. Clinical blood testing is then performed in open mode on an automated hematology analyzer, such as the CELL-DYN Sapphire (Abbott Laboratories, USA). Here, single cells in the whole blood picture are measured in high-dimensional detail, and the measurements include the characteristics of single cells, as shown in Figure 1.

[0015] The measurements are copied from the analytical instrument, for example as an FCS file or in another format, and transferred to an accessible PC, mobile computer, or cloud for processing. These measurements include blood parameters such as characteristics of white blood cells, including neutrophils, eosinophils, basophils, lymphocytes, and monocytes, and characteristics such as size, granularity, lobulation, and complexity (Figure 2).

[0016] The process involves automatically discriminating at least three subpopulations (neutrophils, lymphocytes, and monocytes) by overlaying the measured characteristics on at least one scatter plot of size versus complexity and / or size versus lobulation, as shown, for example, in Figure 3. In Abbott Cell-DYN Sapphire, cell size is determined by measuring the axial light loss (ALL) of each cell, cytoplasmic and nuclear complexity is determined by measuring the inter-angular scatter intensity (IAS), and lobulation is determined by measuring the polarized side scatter intensity (PSS).

[0017] The determination of subpopulations differentiated according to their characteristics is carried out automatically by hierarchical cluster analysis, which is understood as a method for discovering the similarity structure of a data set, where clusters are composed of measurement variables that are closer to each other than to the measurement variables of other clusters (Figure 3). In addition to hierarchical cluster analysis, the following cluster algorithms can also be used: K-Means, K-Medians, Affinity Propagation, Mean-shift, Spectral Clustering, Ward Hierarchical Clustering, DBSCAN, OPTICS, Gaussian Mixtures and Birch.

[0018] Then, elements of at least three subpopulations are sorted and merged to generate a global vector,

number

number

number

[0019] The global vector with each characteristic element can then be standardized by rescaling the values ​​to the range [0,1] or [-1,1], or by rescaling the data to have a mean of 0 and a standard deviation of 1 (unit variance).

number

number

number

[0020] After standardization, principal component analysis (PCA) is applied. PCA can also be applied before or instead of standardization. The goal of PCA is to reduce the dimensionality of the standardized global vector features while preserving as much feature variability (information) as possible (Dunteman GH, Principal Component Analysis, 1989). Large vector sizes can cause problems in machine learning because predictive models based on this data risk being retrained (Kabari LG, Nwame BB, Principal Component Analysis (PCA) ~ An Effective Tool in Machine Learning, International Journals of Advanced Research in Computer Science and Software Engineering, ISSN: 2277-128X (Volume-9, Issue-5)). Furthermore, many representations may be redundant or highly correlated, which can also lead to reduced diagnostic accuracy. Furthermore, PCA significantly improves computational speed. In addition to PCA, sparse principal component analysis, kernel principal component analysis, truncated singular value decomposition, independent component analysis, nonnegative matrix factorization, and latent Dirichlet allocation are also available.

[0021] PCA reduces the number of elements in the global vector from n to the number of principal components p.

number

[0022] Machine learning is a branch of artificial intelligence and a general term for the "artificial" generation of knowledge from experience. Artificial systems learn from examples and are able to generalize them after the learning phase is complete. To do this, machine learning algorithms build statistical models based on training data. This means that the examples provided during the learning phase are not simply memorized, but patterns and regularities in the training data are recognized. In this way, the system is also able to evaluate unknown data (learning transfer).

[0023] When at least two (machine) learning models are used for prediction, this is called an ensemble. Ensemble is a machine learning paradigm in which multiple models are trained to solve the same problem and combined to achieve better results (Langley P., Elements of Machine Learning, 1996), and the ensemble can include artificial neural networks, k-nearest neighbor models, random forest models, decision trees, Adaboost models, gradient boosting machines (GBM), bootstrap aggregation models, stacking generalization models, logistic regression models, Bayesian models, and / or support vector machines (SVM).

[0024] A global vector, e.g., as shown in FIG. 4, is used as an input variable of at least one machine learning model, while at least one global scatter plot, e.g., as shown in FIG. 5, is used as an input image in the form of an input matrix of at least one deep learning model, where the global scatter plot consists of the at least one scatter plot, and the deep learning model can include a convolutional neural network (CNN).

[0025] In the field of deep learning, CNNs are a type of artificial neural network most commonly used for analyzing visual images (Valueva, MV; Nagornov, NN; Lyakhov, PA; Valuev, GV; Chervyakov, NI (2020). “Application of the residue number system to reduce hardware costs of the convolutional neural network implementation”. Mathematics and Computers in Simulation. Elsevier BV. 177: 232-243. doi:10.1016 / j.matcom.2020.04.031. ISSN 0378-4754.). CNNs essentially consist of alternating filter (convolutional) and aggregation (pooling) layers, followed by one or more layers of “regular” fully connected neurons (dense / fully connected layers). CNNs use their filters to recognize structures in images regardless of their location. Here, the type of filter is not predetermined; rather, the neural network learns it. With each increasing level of filtering, the network's level of abstraction increases, and which level of abstraction ultimately activates the back layer is determined by the characteristics of the predetermined class that must be recognized.

[0026] A scatter plot is created by plotting the measured properties of single white blood cells against each other, such as size vs. complexity, size vs. granularity, size vs. lobulation, complexity vs. granularity, complexity vs. lobulation, and granularity vs. lobulation, as shown, for example, in FIG. 6. Some analytical instruments may measure fewer properties, such as only three properties: cell size, cell complexity, and cell contents. It can be assumed that virtually all analytical instruments measure at least two properties, namely, cell size and complexity, and that the simplest scatter plot plots at least these two properties against each other. Individual scatter plots from FIG. 6 can be added to a global scatter plot, and the order of the individual scatter plots can be reversed in the global scatter plot.

[0027] The disease determination in this method is characterized by determining the sum of votes from multiple trained machine learning models and deep learning models (hard voting), and the disease diagnosis with the most votes is selected, as shown in Figure 7, for example.

[0028] Alternatively, the method is characterized in that the sum of the probabilities of multiple trained machine learning models and deep learning models is calculated (soft voting), and the disease diagnosis is output from the weighted average of the maximum sum probability of each ensemble, as shown in Figure 8, for example.

[0029] The machine learning model and the deep learning model are trained based on a pre-created database, as shown in FIG. 9 for example, where the database consists of measured blood parameters. The database is made up of two datasets X 訓練 and X 試験 Divided into X 訓練 is used to train the model, and X 試験 is used to evaluate the quality of the model.

[0030] For training and evaluation of models that require global vectors as input variables, these vectors are derived from the dataset according to the approach described above,

number

number

[0031] To train deep learning models that require global scatterplots as input variables, these global scatterplots are derived from the dataset by the methods described above,

number

[0032] A true disease (positive) or true healthy control (negative) case corresponds to each global vector and global scatterplot in the dataset. The true vector is a response vector or one-hot encoded response matrix Y 訓練 In one-hot coding, categorical data is converted into numerical data for use in machine learning. Specifically, categorical features are converted into binary features that are coded "one-hot," meaning 1 if a feature is represented in this column and 0 otherwise.

[0033] Models are trained on a training dataset using supervised and / or semi-supervised approaches. Supervised learning is a branch of machine learning, where learning refers to the ability of an artificial intelligence to reproduce regularities. The result is a response matrix Y 訓練The learning algorithm seeks to find hypotheses that will produce the most reliable predictions possible. Here, a hypothesis should be understood as a mapping that assigns a possible output value to each input value (Rostamizadeh, Afshin., Talwalkar, Ameet.: Foundations of machine learning. MIT Press, Cambridge, MA 2012, ISBN 978-0-262-01825-8.). Therefore, the method is based on a predetermined output to be learned, the result of which is a response matrix Y 訓練 The results of the learning process can be compared to known correct results, i.e., "monitored" (Guido, Sarah, Rother, Kristian: Einfuehrung in Machine Learning mit Python Praxiswissen Data Science. Heidelberg, ISBN 978-3-96009-049-6.).

[0034] Semi-supervised learning means that during training, we 訓練 This is a machine learning approach that combines a small amount of data with known (labeled) outcomes with a large amount of data with unknown (unlabeled) outcomes. In the first step, a first model is initially trained using labeled data from a pre-created database. Unlabeled data is then initially classified using the trained model and labeled accordingly (pseudo-labeling). The database can be augmented with unlabeled data and corresponding pseudo-labels to be used to train the next model. Here, the pseudo-labels are considered to be true diagnoses.

[0035] After training the models and aggregating them into an ensemble, we use the test dataset X 試験 Check the quality of the ensemble for One way to estimate the quality of the model is to group the results by a confusion matrix.

[0036] [Table 1]

[0037] where TN represents true negatives, FN represents false negatives, FP represents false positives, and TP represents true positive predictions. The confusion matrix shows the number of diseased or healthy patients correctly or incorrectly predicted by the model. Here, sensitivity: Sensitivity=TP / (TP+FN) indicates the probability that a positive diagnosis is correctly classified as positive. For example, sensitivity in a medical diagnosis corresponds to the proportion of patients who are actually recognized as having the disease. Here, in addition to sensitivity, specificity: Specificity=TN / (TN+FP) indicates the probability that a disease will be correctly ruled out as negative. For example, specificity in a medical diagnosis corresponds to the proportion of healthy people who are actually confirmed to be free of the disease. The specificity of a test indicates the probability that an uninfected person will actually be recognized.

[0038] As part of an experimental study, in collaboration with General Hospital No. 2 in St. Petersburg, a database containing 221 cases was retrospectively created. In 110 patients, a diagnosis of acute coronary syndrome (55 cases of STEMI and 55 cases of NSTEMI) was classified by a cardiologist. In 111 patients, the disease was excluded and they therefore served as healthy controls. In all patients, the first blood sample was taken in the emergency room and measured on a CELL-DYN Sapphire hematology analyzer (Abbott Laboratories, USA). The database was compiled from a training dataset X containing 154 cases. 訓練 and a test dataset X containing 67 examples. 試験 The training dataset X was randomly divided into 訓練 was used to train the machine learning and deep learning models. Test dataset X 試験 The accuracy and performance of the model were evaluated on the test dataset X 試験The dataset included 17 STE-ACS, 18 NON-STE-ACS, and 32 negative control cases. After training the models and aggregating them into an ensemble, we used the test dataset X 試験 The quality of the ensemble was checked for the following: 30 of 32 healthy patients were then correctly classified as healthy (TN), and 34 of 35 diseased patients were correctly classified as diseased (TP). This resulted in a sensitivity of 97.14% and a specificity of 93.75%, indicating very high ensemble quality. For example, as shown in Figure 11, comparing the accuracy of the described method with that of Abbott Technologies' high-sensitivity Troponin-I test shows that the described method can provide better sensitivity and specificity immediately after the first blood test.

[0039] A trained model that is trained on a training dataset and shows high accuracy on a test dataset can be used as part of a method in a laboratory information system for predicting diseases that affect the morphological features and cytoplasmic complexity of blood cells. Since this database can be expanded at any time with new labeled and / or unlabeled patient examples, the model can be retrained in a supervised and / or semi-supervised approach to obtain better models and, if necessary, to replace older models.

[0040] In the exemplary illustration of FIG. 10, all method steps of a computer-implemented method for the diagnosis of diseases affecting the morphological characteristics and cytoplasmic complexity of blood cells are summarized.

[0041] The novelty of the presented solution is achieved by using, for the first time, global vectors and global scatter plots, automatically obtained by cluster analysis from measured characteristics of white blood cells, as input variables for an ensemble of machine learning and deep learning models, compared with existing analogues, with the aim of determining diseases affecting the morphological features and cytoplasmic complexity of blood cells. In the case of the diagnosis of NSTE-ACS, as part of an experimental study, we achieved more accurate and faster results than contemporary conventional standard solutions immediately after the first blood sample, as shown by way of example in Figure 11.

[0042] In an advantageous embodiment, the step of obtaining the blood parameters of the multiple blood images includes obtaining them from a hematology analyzer.

[0043] In an advantageous embodiment, the step of performing dimensionality reduction of the global vector comprises principal component analysis, and the global vector is standardized before and / or after dimensionality reduction.

[0044] In an advantageous embodiment, at least one scatterplot image is standardized to at least one deep learning model, which may preferably be done before and / or during transfer to the associated deep learning model.

[0045] In an advantageous embodiment, the step of determining the clusters comprises cluster analysis, and the cluster analysis comprises a hierarchical clustering method.

[0046] In an advantageous embodiment, the step of determining at least one disease comprises soft voting and / or hard voting.

[0047] In an advantageous embodiment, the machine learning model that receives the reduced global vector as an input variable in the ensemble comprises an artificial neural network, a k-nearest neighbor, a random forest, an Adaboost, a gradient boosting machine (GBM), and / or a support vector machine (SVM).

[0048] In an advantageous embodiment, the deep learning model that receives at least one scatterplot image as an input variable includes a convolutional neural network.

[0049] In an advantageous embodiment, all models of the ensemble are trained on a pre-created database containing measured properties and / or scatter plots of single cells, and the database is extendable with new measured properties and / or scatter plots of single cells, so that the database can be constantly expanded with new data for training the models to improve their accuracy.

[0050] In an advantageous embodiment, at least one result report is computer generated, the result report comprising graphics and / or informative text and / or probabilities and / or scores relating to at least one disease, and the display of the result report comprises a display on a computer device and / or a mobile device and / or a testing device.

[0051] In an advantageous embodiment, the method further comprises the steps of: acquiring blood parameters from at least one patient, the blood parameters comprising measured properties of single cells of a number of blood images, the single cells comprising red blood cells and platelets; b) creating at least one histogram in addition to or instead of at least one scatter plot, wherein one of the axes of the histogram comprises the number of single cells; f) using at least one histogram image in addition to or instead of at least one scatter plot image as an input variable of at least one deep learning model of the ensemble for determining at least one disease, and complementing the database of claim 8 with the histogram.

[0052] The present invention also provides a system for determining diseases affecting the morphological characteristics and cytoplasmic complexity of blood cells, the system comprising a computer having a calculation unit, a memory unit connected thereto, and an input unit, the system comprising: a) acquiring blood parameters acquired from at least one patient by an analytical instrument through the input unit, the blood parameters including at least two measured characteristics of single cells of many blood pictures, the single cells including white blood cells, the measured characteristics of the white blood cells including size, cytoplasmic granularity, and nuclear structure; b) generating at least one scatter plot by the calculation unit, each axis of the scatter plot including various measured characteristics of the single cells; c) determining at least one cluster in at least one of the at least one scatter plot by the calculation unit, the cluster including subpopulations of white blood cells, the subpopulations including monocytes, , lymphocytes, basophils, neutrophils, and eosinophils; d) integrating the characteristic elements of the determined clusters by a calculation unit to generate a one-dimensional global vector, wherein the arrangement of the characteristic elements in the global vector includes an arrangement according to the associated cluster; e) reducing the dimension of the global vector by the calculation unit; f) determining at least one disease by an ensemble using the calculation unit, the ensemble including at least one machine learning model and at least one deep learning model, wherein the machine learning model receives the at least one reduced global vector as an input variable and the at least one deep learning model receives the at least one scatter plot image as an input variable; and g) automatically generating a report by the calculation unit including at least one result regarding the determination of the at least one disease.

[0053] The system may further be designed to perform one or more of the method steps described above. It should be noted that the memory unit connected to the computing unit is designed to store and supply the corresponding algorithm to the computing unit for execution, and to store and supply data to the computing unit when executing the algorithm or the method. The input unit may be any device capable of transferring data from the outside to the system. It may be a data or signal interface, an input device, a data carrier with a corresponding interface, or any other suitable device. [Brief explanation of the drawings]

[0054] [Figure 1] 1 shows the characteristics of white blood cells measured using a laser beam on a hematology analyzer (Abbott Cell Dyn Sapphire): size, lobulation, granularity, and complexity. This is an example for an Abbott Cell Dyn Sapphire-type analyzer. Other analyzers, such as Sysmex-type analyzers, can measure cell size, cell complexity, and cell contents. As mentioned above, suitable analyzers typically measure at least two characteristics: cell size and internal complexity. [Figure 2] Figure 1 shows an example of a series of subpopulations of white blood cells sorted according to properties measured by cluster analysis, including basophils, neutrophils, lymphocytes, monocytes, and eosinophils. All subpopulations are described by their properties. [Figure 3] FIG. 1 shows an example of the determination of at least three subpopulations (neutrophils, lymphocytes, monocytes) according to at least three properties (size, complexity, lobulation) by plotting the measured properties in at least two scatter plots superimposed on Cartesian coordinates: size vs. complexity and size vs. lobulation. [Figure 4] FIG. 1 illustrates an example of standardization and reduction of global vectors used as input variables for machine learning models. [Figure 5]FIG. 1 illustrates an example of a global scatter plot used as an input image in matrix form for a deep learning model. [Figure 6] FIG. 10 shows an example of creating a global scatter plot by plotting the measured properties of single white blood cells against each other in individual scatter plots: size vs. complexity, size vs. granularity, size vs. lobulation, complexity vs. granularity, complexity vs. lobulation, and granularity vs. lobulation, where the order of the individual scatter plots can be swapped in the global scatter plot. [Figure 7] FIG. 1 illustrates an example of the hard voting method described above for disease determination. [Figure 8] FIG. 10 shows an example of the soft voting method also described above for disease determination. [Figure 9] Figure 1 shows an example of using a database for training and evaluating a model. The database contains measured blood parameters and corresponding true diagnoses (labeled database). The database is divided into two parts, X-training and X-testing, where X-training is used for training and X-testing is used for evaluating the model. For models requiring global vectors as input variables, the global vectors are obtained from X-training and X-testing using the approach described above. For models requiring global scatter plots as input variables, the global scatter plots are obtained using the approach described above. [Figure 10]1 is a diagram illustrating an exemplary diagram summarizing all steps of a computer-implemented method. A patient's blood sample, collected prior to the method, is subjected to testing in the method according to the present invention, referred to herein as step (1). In step (2), the characteristics of single cells in the blood sample are measured by an analyzer. In step (3), the measurement data is transferred as raw data to a computer. In step (4), a scatter plot is created. In step (5), subpopulations are discriminated and determined according to their characteristics by cluster analysis, and then integrated to generate a global vector. In step (6), the global vector is standardized and reduced. In step (7), the reduced vector is used as an input variable for at least one trained machine learning model, while the individual scatter plots obtained in step (4) are integrated to generate a global scatter plot, which is used as an input image for at least one trained deep learning model. In step (8), a diagnosis is made using a hard or soft voting approach. In step (9), a diagnosis is output on a graphical surface or in the form of an automatically generated report. In step (11), the predicted diagnoses can be stored in a pre-created database together with the corresponding raw data obtained in step (3) to extend the database with new examples and train a new model using a semi-supervised approach. The pre-created database can also be extended in step (11) with raw measurement data and corresponding true diagnoses. The purpose of the database is model training and evaluation. Here, the database is divided into a training dataset and a test dataset in steps (12) and (13). In step (18), at least one machine learning model is trained and evaluated, using the global vectors obtained in steps (14) and (17) as input variables. In step (19), at least one deep learning model is trained and evaluated, using the scatter plots obtained in steps (15) and (16) as input variables. The model in step (7) can be replaced at any time by a newly trained model in steps (18) and (19) as long as the newly trained model in steps (18) and (19) achieves better results on the test dataset. [Figure 11] FIG. 1 shows a comparison of the accuracy of the method of the present invention with Abbott Technologies' high-sensitivity Troponin-I test.

[0055] The described method can be implemented based on the results of clinical blood analysis obtained during the first few minutes of a patient's emergency room visit, before they undergo any other laboratory or instrumental testing processes, thereby significantly improving the quality of medical care.

[0056] An example of application of this method to a patient who was confirmed to have a "positive" diagnosis as a result of applying this method The following example further illustrates the method according to the present invention. A 36-year-old patient entered the care department with a provisional diagnosis of "arteriosclerotic heart disease, acute coronary syndrome without ST-segment elevation, acute heart failure, Killip class I" two hours after a typical pain syndrome. An electrocardiogram was performed in the reception section, and the electrocardiogram also showed no ST-segment elevation. A venous blood sample was then taken for laboratory testing. The blood sample taken outside the present method was subjected to testing (step 1 in Figure 10), which also included a high-sensitivity cardiac troponin I method and clinical blood analysis.Laboratory studies showed the following results: urea 4.2 mmol / L (3.0-9.2); ALT 16 units / L (0-55); AST 12 units / L (5-34); total protein 70 g / L (64-83); creatinine 74 μmol / L (64-111); total bilirubin 6.2 μmol / L (3.4-20.5); glucose 7.5 mmol / L (3.9-5.5); potassium 3.7 mm ol / L (3.5-5.1); sodium 137 mmol / L (135-145); ionized calcium 1.23 mmol / L (1.13-1.32); APTV 78.7 seconds (25.1-36.5); MNO 0.97 (0.2-0.5); 90-1.20); prothrombin 118.0% (70.0-140.0); prothrombin time 11.0 seconds (9.4-12.5); white blood cells 12.4 10E9 / L (4.0-9.0); (NEUT) neutrophils 10.0×109 / L (2.0-5.5); (NEUT%) neutrophils 80.0% (48.0-78.0); (LYM) lymphocytes 1.79×109 / L (1.20-3.00); (LYM%) lymphocytes 14.3% (19.0-37.0); (MON) monocytes 0.57 109 / L (0.09-0.60); (MON%) Monocytes 4.6% (3.0-11.0); (EOS) Eosinophils 0.07×109 / L (0.00-0.30); (EOS%) Eosinophils 0.52% (1.00-5.00); (BAS) Basophils 0.07×109 / L (0.00-0.06); (BAS%) Basophils 0.52% (0.00-1.00); (HGB) Hemoglobin 134g / L (130-160); (HCT) Hematocrit 40.7% (40.0-48.0); (RBC) Red blood cells 4.45 10 × 10⁻² / L (4.00-5.60); (MCH) mean hemoglobin concentration in red blood cells 30.1 pg (24.0-34.0); (MCHC) mean hemoglobin concentration in red blood cells 32.9 g / dL (30.0-38.0); (MCV) mean corpuscular volume 91.4 fL (75.0-95.0); (RDW-CV) red blood cell distribution approximately 11.0% (11.5-16.0%).5); (PLT) platelets 262 × 10⁻² / L (180-400); (MPV) mean platelet volume 11.5 fL (7.4-10.4); cardiac troponin I (high-sensitivity method) 37.9 ng / mL (upper limit of normal 26.2 ng / mL).

[0057] After blood measurement (step 2 in Figure 10), the raw measurement data was transferred to a PC (step 3 in Figure 10). Two scatter plots were then generated for the trained deep learning model (step 4 in Figure 10), which were then merged to generate a global scatter plot. In parallel, a global vector with 4216 elements was derived (step 5 in Figure 10).

number

number

[0058] After standardization, PCA was applied to reduce the dimensionality of the features from 4216 elements to 4 elements, called principal components, while preserving as much variability (information) as possible in the features (step 5 in Figure 10). After applying principal component analysis, all 4216 elements of the standardized global vector were reduced to vectors in a 4-dimensional subspace:

number

[0059] At first glance, it is difficult to extract information for patient diagnosis from the values ​​of this reduced vector. For this purpose, an ensemble of trained machine learning models is used. This ensemble consists of an ensemble of artificial neural networks, an ensemble of k-nearest neighbor models, an ensemble of random forest models, an ensemble of Adaboost models, an ensemble of gradient tree boosting models, and an ensemble of support vector machine models, each trained with a pre-created database. The standardized and reduced global vector is used as the input vector for all ensembles, while the global scatterplot is used for the deep learning model. In the above case, the votes for ACS from the individual ensembles are counted (hard voting): [Table 2]

[0060] The final result of the ACS test by hard voting was positive. The decision was made to perform percutaneous coronary intervention. The patient underwent coronary angiography followed by transluminal dilatation and stenting of the infarct-dependent coronary artery.

[0061] Coronary angiography No. 7175 on June 7, 2018: Left-sided blood circulation. Left coronary artery: main trunk - no stenosis. Anterior interventricular branch - 5-50% stenosis at the ostium, subocclusion in the middle third. 90% stenosis in the middle to proximal third. Diagonal branch - no stenosis. Circumflex branch (BB) - no stenosis. Obtuse marginal branch - no stenosis. Right coronary artery: hypoplastic. Acute branch: no stenosis. Posterolateral branch (ROB) - no stenosis. Posterior interventricular branch - no stenosis.

[0062] Coronary angioplasty and PMV stent placement No. 7176 on June 7, 2018: The mean third of the PMV stenosis was BC 2.0 x 20.0 mm, p=18 atm. A 2.75 x 33.0 mm drug-coated stent was placed in the middle third of the BC 2.0 x 20.0 mm, p=16 atm. Control: TIMI III grade blood flow. On two projections on June 8, 2018, no infiltrates were detected on chest X-rays. The roots were structural and not enlarged, with the left one partially obstructed. The pulmonary pattern was unchanged. The diaphragm was contoured. The cardiac shadow was unremarkable. The sinuses were not obstructed.

[0063] The following treatments were administered: beta-blockers, anticoagulants, dual antiplatelet therapy, statins (Crestor dose reduced from 20 to 10 mg / day due to elevated transaminase levels), and gastroprotective drugs. The patient refused rehabilitation treatment at a nursing home.

[0064] During the postoperative period, the peak cardiac troponin I concentration during dynamic observation reached 7522.5 ng / ml. The hospitalization lasted 12 days. The final diagnosis was "atherosclerotic heart disease." On June 7, 2018, the patient presented with an acute myocardial infarction of the left ventricular anterior parietal high lateral segment without ST-segment elevation. On June 7, 2018, coronary angioplasty and stent placement were performed. The patient was discharged on June 19, 2018, and further outpatient follow-up was performed at his residence.

Claims

1. 1. A computer-implemented method for determining diseases affecting morphological characteristics and cytoplasmic complexity of blood cells, the method comprising: a) obtaining blood parameters from at least one patient, the blood parameters comprising at least two measured characteristics of single cells of a number of blood pictures, the single cells comprising white blood cells, the measured characteristics of the white blood cells comprising size, cytoplasmic granularity, and nuclear structure; b) generating at least one scatter plot, each axis of said scatter plot containing different measured properties of said single cells; c) determining at least one cluster in at least one of said at least one scatter plot, said cluster comprising a subpopulation of leukocytes, said subpopulation comprising monocytes, lymphocytes, basophils, neutrophils and eosinophils; d) generating a one-dimensional global vector by integrating the characteristic elements of the determined clusters, wherein the arrangement of the characteristic elements in the global vector includes an arrangement according to the associated cluster; e) performing a dimensionality reduction of the global vector; f) determining at least one disease, said determination being performed using an ensemble of at least one machine learning model and an ensemble of at least one deep learning model, said machine learning model receiving at least one reduced global vector as an input variable, and said at least one deep learning model receiving at least one scatter plot image as an input variable; and g) automatically generating a report including at least one result regarding the determination of at least one disease; Including, the step of acquiring the blood parameters includes acquiring them from a hematology analyzer; The step of determining the clusters comprises cluster analysis, and the cluster analysis comprises a hierarchical clustering method. method.

2. The computer-implemented method of claim 1 , wherein the step of reducing the dimension of the global vector comprises principal component analysis, and wherein the global vector is standardized after the dimension reduction.

3. The computer-implemented method of claim 1 , further comprising normalizing the at least one scatterplot image to the at least one deep learning model.

4. 2. The computer-implemented method of claim 1, wherein the step of diagnosing at least one disease includes a soft voting method and / or a hard voting method for diagnosing the disease by voting using the results of a plurality of disease diagnosis models based on a global vector and the results of a plurality of disease diagnosis models based on a global scatter plot.

5. 2. The computer-implemented method of claim 1, wherein the machine learning model that receives the reduced global vector as an input variable in an ensemble comprises an artificial neural network, a k-nearest neighbor, a random forest, an Adaboost, a gradient boosting machine (GBM), and / or a support vector machine (SVM).

6. 10. The computer-implemented method of claim 1, wherein the deep learning model that receives at least one scatterplot image as an input variable comprises a convolutional neural network.

7. 2. The computer-implemented method of claim 1, wherein all models of the ensemble are trained based on a pre-created database, the database including measured properties and / or scatter plots of the single cells, and the database is extendable with new measured properties and / or scatter plots of single cells.

8. 2. The computer-implemented method of claim 1, further comprising generating at least one result report on a computer, the result report including graphics and / or informative text and / or probabilities and / or scores relating to at least one disease, and displaying the result report including displaying it on a computer device and / or a mobile device and / or a testing device.

9. The computer-implemented method of claim 7, further comprising the steps of: acquiring blood parameters from the at least one patient, the blood parameters comprising measured characteristics of single cells of many blood images, the single cells comprising red blood cells and platelets; creating at least one scatter plot; creating at least one histogram in addition to or instead of the at least one scatter plot, wherein one of the axes of the histogram comprises the number of the single cells; determining at least one disease; using at least one histogram image in addition to or instead of the at least one scatter plot image as an input variable of at least one deep learning model of an ensemble for determining at least one disease; and supplementing the database of claim 7 with the histogram.

10. 1. A system for determining diseases affecting morphological characteristics and cytoplasmic complexity of blood cells, the system comprising a computer having a calculation unit, a memory unit connected thereto, and an input unit, the system comprising: a) obtaining blood parameters acquired by an analytical device from at least one patient through the input unit, the blood parameters including at least two measured characteristics of single cells of multiple blood images, the single cells including white blood cells, the measured characteristics of the white blood cells including size, cytoplasmic granularity, and nuclear structure; b) generating at least one scatter plot by said computing unit, each axis of said scatter plot comprising different measured properties of said single cells; c) determining, by the computing unit, at least one cluster in at least one of the at least one scatter plots, the cluster comprising a subpopulation of leukocytes, the subpopulation comprising monocytes, lymphocytes, basophils, neutrophils, and eosinophils; d) combining the characteristic elements of the determined clusters by the arithmetic unit to generate a one-dimensional global vector, wherein the arrangement of the characteristic elements in the global vector includes an arrangement according to the associated cluster; e) performing a dimensionality reduction of the global vector by the computation unit; f) determining at least one disease by an ensemble using the computing unit, the ensemble including at least one machine learning model and at least one deep learning model, wherein the machine learning model receives at least one reduced global vector as an input variable, and the at least one deep learning model receives at least one scatter plot image as an input variable; and g) designed to automatically generate by said computing unit a report including at least one result relating to the determination of at least one disease; obtaining the blood parameters includes obtaining them from a hematology analyzer; determining the clusters includes cluster analysis, and the cluster analysis includes a hierarchical clustering method; system.

Images