Classification method for classifying systemic inflammatory states

Abstract

A computer-implemented training method (110) of training at least one trainable model is disclosed for classifying a patient's health condition into a systemic inflammatory state (168, 170, 172) selected from a predetermined group of at least three systemic inflammatory states (168, 170, 172). The method comprises: a. providing the trainable model; b. retrieving labeled training patient data, the training patient data comprising gene expression data of patients having known systemic inflammatory states (168, 170, 172); and c. training the trainable model on the labeled training patient data. Further disclosed is a computer-implemented classification method (112) of classifying a patient's health condition into a systemic inflammatory state (168, 170, 172) selected from a predetermined group of at least three systemic inflammatory states (168, 170, 172) and systems, computer programs and computer-readable storage media for performing the methods.

Description

[0001] Classification method for classifying systemic inflammatory states

[0002] Technical Field

[0003] The invention generally refers to the topic of classifying a patient’ s health condition with respect to the patient’s inflammatory state, specifically the patient’s systemic inflammatory state and more specifically the patient’s inflammatory response state. Thus, the invention refers to a computer-implemented training method of training at least one trainable model for classifying a patient’s health condition into a systemic inflammatory state selected from a predetermined group of at least three systemic inflammatory states. Further, the invention relates to a computer-implemented classification method of classifying a patient’s health condition into a systemic inflammatory state selected from a predetermined group of at least three systemic inflammatory states. Moreover, the invention relates to corresponding training and classification systems, as well as to computer programs and computer-readable storage media for implementation of the training and classification methods. The methods and devices according to the present invention specifically may be used in the field of professional care, in order to reliably classify patients, specifically patients suffering from inflammatory symptoms, thereby being capable of deriving optimized treatment methods for the specific patients. Other applications, however, are also feasible.

[0004] Background art

[0005] Sepsis is a life-threatening medical emergency that specifically may occur when the body's response to an infection causes damage on its own tissues. Thus, early diagnosis of sepsis and septic shock is crucial, in order to enable the healthcare personnel to respond to the respective situation rapidly. Several stages of sepsis are known, which, in the following, are also referred to as septic states, inflammatory states or systemic inflammatory states. Thus, the first stage or state is also referred to as “systemic inflammatory response syndrome (SIRS)”. In this stage, typically, one or more of the following symptoms may be noted: abnormal body temperature, abnormal heart rate, abnormal respiratory rate, abnormal blood gas, abnormal white blood cell count. In the second stage or stage, often referred to as “sepsis”, SIRS occurs in response to an infectious process. The third stage or state typically is referred to as “severe sepsis”. This third stage is typically defined as sepsis with sepsis- induced organ dysfunction or tissue hypoperfusion. In this third stage, typically, one or more of the following additional symptoms may occur: hypotension, elevated lactate, decreased urine output. Severe sepsis is an infectious disease state associated with multiple organ dysfunction syndrome (MODS). Finally, the fourth state or stage is referred to as “septic shock”. This fourth stage denotes a severe sepsis in conjunction with persistently low blood pressure.

[0006] Depending on the results of the diagnosis of the septic or systemic inflammatory state, treatment in the form of number of actions or “bundles” is given to the patient. Thus, administration of antibiotics, monitoring of blood cultures, monitoring of lactate, hemoglobin determination, urine output monitoring, administration of high-flow oxygen and administration of intravenous fluids are typical treatment measures. Further, depending on the stage, other measures are required, such as surgical drainage, support for organ dysfunction such as hemodialysis, mechanical lung ventilation, transfusions, drug and fluid therapy or combinations thereof. Besides choosing appropriate measures, timing of these measures, however, is a crucial factor for successful treatment.

[0007] Disclosed are method and apparatus for identifying biomarkers and in particular for identifying biomarkers for use in making clinical assessments, such as early diagnostic, diagnostic, disease stage, disease severity, disease subtype, response to therapy or prognostic assessments. In one particular example, the techniques are applied to allow assessments of patients suffering from, suspected of suffering from, or with clinical signs of SIRS (Systemic Inflammatory Response Syndrome) being either infection-negative SIRS or infection-positive SIRS.

[0008] Besides traditional detection of symptoms, such as one or more of the above-mentioned symptoms, indicating a systemic inflammatory state, the evaluation of the level of specific biomarkers gained from patient samples is known for detecting systemic inflammatory states. Thus, as an example, EP 3011059 Bl and WO 2014 / 201516 A2 discloses methods and apparatus for identifying biomarkers and in particular for identifying biomarkers for use in making clinical assessments, such as early diagnostic, diagnostic, disease stage, disease severity, disease subtype, response to therapy or prognostic assessments. In one particular example, the techniques are applied to allow assessments of patients suffering from, suspected of suffering from, or with clinical signs of SIRS (Systemic Inflammatory Response Syndrome) being either infection-negative SIRS or infection-positive SIRS.

[0009] Further, means and methods of machine learning are known for diagnostic purposes. Thus, as an example, WO 2020 / 198068 Al discloses systems and methods for subject clinical condition evaluation using a plurality of modules. Modules comprise features whose corresponding feature values associate with an absence, presence or stage of phenotypes associated with the clinical condition. A first dataset is obtained having feature values, acquired through a first technical background from respective subjects in transcriptomic, proteomic, or metabolomic form, for at least a first of the plurality of modules. A second training dataset is obtained having feature values, acquired through a technical background other than the first technical background, from training subjects of the second dataset, in the same form as for the first dataset, of at least the first module. Inter-dataset batch effects are removed by co- normalizing feature values across the training datasets, thereby calculating co-normal- ized feature values used to train a classifier for clinical condition evaluation of the test subject.

[0010] In Mayhew, M.B., Buturovic, L., Luethy, R. et al.: “A generalizable 29-mRNA neural-net- work classifier for acute bacterial and viral infections.”, Nat Commun 11, 1177 (2020), identification of bacterial and viral infections by using a generalizable host-gene-expressionbased classifier is disclosed.

[0011] In Taroni JN, Greene CS: “Cross-platform normalization enables machine learning model training on microarray and RNA-Seq data simultaneously”, bioRxiv 118349; doi: https: / / doi.Org / 10.l 101 / 118349, the use of quantile and Training Distribution Matching normalization for supervised and unsupervised model training on microarray and RNA-seq data simultaneously is disclosed.

[0012] Further, for evaluation of biomarker levels, reference may also be made to Thompson JA, Tan J, Greene CS: “Cross-platform normalization of microarray and RNA-seq data for machine learning applications”, PeerJ. 2016 Jan 21;4:el621. doi: 10.7717 / peerj. l621. PMID: 26844019; PMCID: PMC4736986, and to Sweeney TE, Shidham A, Wong HR, Khatri P.: “A comprehensive time-course-based multicohort analysis of sepsis and sterile inflammation reveals a robust diagnostic gene set”, Sci Transl Med. 2015 May 13;7(287):287ra71. doi: 10.1126 / scitranslmed.aaa5993. PMID: 25972003; PMCID: PMC4734362. Despite the advantages implied by the above-mentioned means and methods, several challenges and room for improvement remain. Thus, many studies use only single-platform gene expression data to develop models for sepsis prediction or biomarker detection with only few data points. Consequently, the accuracy of prediction of known methods typically is rather poor. Further, known methods and devices for diagnosis of sepsis are rather nonspecific and enable the healthcare personnel to simply distinguish two different classes of systemic inflammatory states, such as sepsis from septic shock. As noted above, however, precise diagnosis of the patient’s systemic inflammatory state may be crucial, in order to take appropriate and state-specific therapeutic measures, thereby increasing the efficiency of the treatment.

[0013] Problem to be solved

[0014] It is therefore desirable to provide methods and devices which address the above-mentioned shortcomings of known methods and devices of diagnosis, specifically in the field of diagnosis of systemic inflammatory states. Specifically, methods and systems shall be proposed which increase the accuracy of diagnosis of systemic inflammatory states, and which also allow for distinguishing different inflammatory states, in order to enable the healthcare personnel to administer appropriate treatment.

[0015] Summary

[0016] This problem is addressed by a computer-implemented training method of training at least one trainable model for classifying a patient’s health condition into a systemic inflammatory state, by a computer-implemented classification method of classifying a patient’s health condition into a systemic inflammatory state, by corresponding training and classification systems, as well as by computer programs and computer-readable storage media for implementation of the training and classification methods, with the features of the independent claims. Advantageous embodiments which might be realized in an isolated fashion or in any arbitrary combinations are listed in the dependent claims as well as throughout the specification.

[0017] As used in the following, the terms “have”, “comprise” or “include” or any arbitrary grammatical variations thereof are used in a non-exclusive way. Thus, these terms may both refer to a situation in which, besides the feature introduced by these terms, no further features are present in the entity described in this context and to a situation in which one or more further features are present. As an example, the expressions “A has B”, “A comprises B” and “A includes B” may both refer to a situation in which, besides B, no other element is present in A (i.e. a situation in which A solely and exclusively consists of B) and to a situation in which, besides B, one or more further elements are present in entity A, such as element C, elements C and D or even further elements.

[0018] Further, it shall be noted that the terms “at least one”, “one or more” or similar expressions indicating that a feature or element may be present once or more than once typically will be used only once when introducing the respective feature or element. In the following, in most cases, when referring to the respective feature or element, the expressions “at least one” or “one or more” will not be repeated, non-withstanding the fact that the respective feature or element may be present once or more than once.

[0019] Further, as used in the following, the terms "preferably", "more preferably", "particularly", "more particularly", "specifically", "more specifically" or similar terms are used in conjunction with optional features, without restricting alternative possibilities. Thus, features introduced by these terms are optional features and are not intended to restrict the scope of the claims in any way. The invention may, as the skilled person will recognize, be performed by using alternative features. Similarly, features introduced by "in an embodiment of the invention" or similar expressions are intended to be optional features, without any restriction regarding alternative embodiments of the invention, without any restrictions regarding the scope of the invention and without any restriction regarding the possibility of combining the features introduced in such way with other optional or non-optional features of the invention.

[0020] In a first aspect of the invention, a computer-implemented training method of training at least one trainable model is disclosed. The trainable model is configured for classifying a patient’s health condition into a systemic inflammatory state, specifically an inflammatory response state, the systemic inflammatory state being selected from a predetermined group of at least three systemic inflammatory states. The training method comprises the following method steps, which, as an example, may be performed in the given order. However, a different order is also feasible. Further, it is possible to perform two or more of the method steps simultaneously or in a fashion overlapping in time. Further, it is also possible to perform one, more than one or even all of the method steps repeatedly.

[0021] The training method comprises the following steps: a. providing the trainable model; b. retrieving labeled training patient data, the training patient data comprising gene expression data of patients having known systemic inflammatory states; and c. training the trainable model on the labeled training patient data. The term “training” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to a process of determining parameters of at least one machine learning and / or deep learning model, specifically of the algorithm of the machine learning and / or deep learning model, specifically on at least one training data set or set of training data. The training specifically may comprise at least one optimization or tuning process, wherein a best parameter combination, e.g. according to at least one optimization procedure, is determined.

[0022] The term “trainable model” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The trainable model may be or may comprise at least one mathematical model configured for transforming one or more input values into one or more output values by using one or more parameters which may be adjusted in order to enable the model to be trained. The trainable model specifically may be a trainable mathematical model which is trainable on at least one training data set using one or more of machine learning, deep learning, neural networks, or other form of artificial intelligence. The term specifically may refer, without limitation, to the fact that the trainable model can be further trained, optimized or updated based on additional training data. Specifically, the trainable model is trained on a training dataset. The trainable model may be trained by using machine learning. The train- able model may be at least partially data-driven by being trained on data from historical training data.

[0023] The trainable model specifically may comprise at least one trainable model selected from the group consisting of: a decision tree model, specifically at least one of an XGBoost model and a Random Forest model; a Support Vector Machine (SVM) model, specifically, a linear kernel SVM; a nearest neighbors model, specifically a KNeighbors Classifier; a Bayes model, specifically a GaussianNB() model; a regression model, specifically at least one of a LogisticRegression; a linear and a nonlinear regression model, e.g. a regression model comprising transformed features, such as log-transformed or polynomial; an Artificial Neural Network (ANN), specifically a non-linear Artificial Neural Network (ANN), in particular a deep learning architecture such as Convolutional NN, Recurrent NN, Long Short Term Memory NN, and the like; a kernel-based method; a tree regression model; a distributed gradient-boosting framework model, specifically a light gradient-boosting machine model (LightGBM classifier). Consequently, the term “trained model” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to a trainable model which has gone through at least one training process as defined above, by applying, at least once, a set of training data to the trainable model. Specifically, one or more parameters of the trainable model might have been adapted, on the basis of the training data, in order to transform the trainable model into a trained model. Specifically, the trained model may be a trainable model which was trained on at least one training dataset, also denoted training data. The trained model specifically may be or may comprise a classifier, configured for classifying an object, on the basis of one or more input variables, describing the object. The trained model specifically may comprise at least one trained model selected from the group consisting of a decision tree model, specifically at least one of an XGBoost model and a Random Forest model; a Support Vector Machine (SVM) model, specifically, a linear kernel SVM; a nearest neighbors model, specifically a KNeighbors Classifier; a Bayes model, specifically a GaussianNB() model; a regression model, specifically at least one of a LogisticRegression; a linear and a nonlinear regression model, e.g. a regression model comprising transformed features, such as log-transformed or polynomial; an Artificial Neural Network (ANN), specifically a non-linear Artificial Neural Network (ANN), in particular a deep learning architecture such as Convolutional NN, Recurrent NN, Long Short Term Memory NN, and the like; a kernel-based method; a tree regression model; a distributed gradient-boosting framework model, specifically a light gradient-boosting machine model (LightGBM classifier).

[0024] The trainable model and / or the trained model specifically may be a model taken from a software library, providing a plurality of trainable models. As an example, the Scikit-learn open- source software library may be mentioned, for trainable models and / or trained models being programmed in the Python programming language. Additionally or alternatively, also, as an example, reference may be made to the extreme Gradient Boosting (XGBoost) open-source software library, specifically providing gradient boosting trainable models, being programmed in one or more of the programming languages C++, Java, Python, R, Julia, Perl and Scala. Other libraries, however, are also feasible, as well as customized trainable models not retrieved from software libraries.

[0025] The term “classify” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to the process of assigning an object to one or more predetermined or determinable classes. The classifying specifically may comprise a prediction or assignment providing an indication to which class out of a plurality of classes an object belongs, e.g. in accordance with one or more predetermined properties of the object.

[0026] The term “patient’s health condition” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to at least one item of information indicating at least one physical state or health state of the patient, e.g. with respect to one or more or even all body functions of the patient.

[0027] The term “systemic inflammatory state” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. Systemic inflammation refers to inflammation which affects the entire organism rather than bring a locally restricted inflammation. Systemic inflammation is, typically, characterized by systemic symptoms of inflammation. Such symptoms of systemic inflammation are described in the prior art and are known to the clinician. Typically, said symptoms for systemic inflammation are selected from the group consisting of hyperthermia or hypothermia, tachycardia, tachypnoea, leucocyte abnormalities, and organ failure. The term “systemic inflammatory state” specifically may refer, without limitation, to at least one item of information indicating that the patient’s health condition with respect to the various stages of sepsis. Thus, specifically, the systemic inflammatory state may be or may comprise a specific stage of the various known stages of sepsis. As will be outlined in further detail below, the predetermined group of at least three systemic inflammatory states may, thus, comprise the following systemic inflammatory states: the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock. Additionally, the following state may also be comprised: the patient being in an apparently healthy state.

[0028] Therein, the term “sepsis” generally is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to a body’s extreme response to an infection, indicated, as an example, by a plurality of physical findings, such as fever, low blood pressure, increased heart rate and difficulty of breathing. Other symptoms of the various stages of sepsis, herein, also referred to as states of sepsis, are listed above. As outlined above, the trainable model is configured for classifying a patient’s health condition into a systemic inflammatory state. Thus, as discussed above, the trainable model specifically may assign the patient’s health condition to one inflammatory state selected from the predetermined group of the at least three systemic inflammatory states.

[0029] The term “retrieve” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to the process of obtaining data from a data source. The data source may, however, very, in accordance with the specific application. Thus, in the context of the present invention, the retrieving of the labeled training patient data may take place by at least one of the following: downloading the labeled training patient data from at least one data source, such as from at least one data storage device, from a web- or cloud-based data storage device; obtaining the labeled training patient data via at least one computer network, such as the Internet; obtaining the labeled training patient data via at least one wire-based and / or wireless interface. The retrieving of the training data may fully or partially take place automatically, such as by automatic download of the labeled training patient data from publications of data on clinical studies, and / or may fully or partially take manually. Semi-automatic retrieving processes are also possible. Thus, as an example, data of clinical studies may be downloaded, e.g. automatically, and the data may be processed in order to generate the labelled training patient data, e.g. by manual labeling of the training patient data downloaded from a web-based data source and / or by other data cleaning or data processing steps. Additionally or alternatively, the retrieving of the labeled training patient data may also comprise the actual measurement of the data, e.g., by evaluating patient samples, such as by determining the gene expression data by appropriate measurements.

[0030] The term “labeled” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to the property of training data having, besides the actual training data, additional information indicating a classification of the data. Thus, as an example, the labeled training patient data may contain, specifically for each patient or each data set of a patient, besides the gene expression data, an indication of the known systemic inflammatory state of the patient. Thus, as an example, for each patient or each data set of the patient, besides the gene expression data off their respective patient, information may be comprised, indicating the systemic inflammatory state of the patient, e.g. one of the following systemic inflammatory states: the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock. Additionally, the following state may also be comprised: the patient being in an apparently healthy state. The systemic inflammatory state of the respective patients may be obtained by other diagnostic means, such as by detecting one or more of the above-mentioned symptoms of the respective patients contributing to the training patient data.

[0031] Consequently, the labeled training patient data may be obtained by using data from a plurality of patients, such as by using data from one or more clinical studies. As an example, data from at least 100 patients, specifically data from at least 500 patients or more, may be used for generating the labeled training patient data. For, each patient, firstly, gene expression data may be generated. Further, for each patient, information on the systemic inflammatory state of the respective patient may be obtained, such as the patient suffering from SIRS, the patient suffering from sepsis, the patient suffering from a septic shock. Additionally, the following state may also be comprised: the patient being in an apparently healthy state. As outlined above, diagnostic indicators may be used other than gene expression data for obtaining the systemic inflammatory state of the respective patient, such as by detecting the presence or absence of the characteristic symptoms for each of the systemic inflammatory states. Then, for each patient, the gene expression data are labeled at least with the respective information on the systemic inflammatory state of the respective patient, in order to obtain a labeled training data set for the respective patient. In addition, and optionally, additional information on the respective patient may be comprised in the labeling, such as information on the age of the patient, information on the gender of the patient, medical information on the patient and the like. The aggregate of the labeled training data set of the patients may then form part of the labeled training patient data.

[0032] Additionally or alternatively, as will be outlined in further detail below, the labeled training data may also be obtained by using data from one or more medical studies. Thus, as will be shown, various medical studies have been performed, indicating the gene expression data of the patients and also indicating the respective known systemic inflammatory states of the patients. Therein, training data from more than one medical study may be combined.

[0033] Specifically, one or more of the following medical studies, the results of which are available, may be used for providing the labeled training patient data:

[0034] GEO accession codes / study codes and download links:

[0035] GSE8121 : https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE8121

[0036] GSE9692: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE9692 GSE11755: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 1755 GSE13904: https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE13904 GSE57065: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE57065 GSE66099: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE66099 GSE26378: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE26378 GSE26440: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE26440 GSE27131 : https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE27131 GSE22098: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE22098

[0037] GSE63042: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE63042 GSE95233: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE95233 GSE9960: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE9960

[0038] GSE28750: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE28750 GSE131411 : https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE131411 GSE110487: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 10487

[0039] E-MTAB-1548: https: / / www.ebi.ac.uk / biostudies / arrayexpress / studies / E-MTAB-1548 GSE74224: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE74224 E-MTAB-4421 : https: / / www.ebi.ac.uk / biostudies / arrayexpress / studies / E-MTAB- 4421 ?query=E-MT AB-4421

[0040] SCP548: https: / / singlecell.broadinstitute.org / single_cell / study / SCP548 / an-immune-cell- signature-of-bacterial-sepsis-patient-pbmcs

[0041] GSE25504: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE25504

[0042] GSE54514: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE54514 GSE69528: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE69528 GSE119217: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 19217 GSE157103: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 57103 GSE154998: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 54998 GSE149689: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE149689

[0043] As will be outlined in further detail below, for the studies leading to the present invention, all of the above-mentioned clinical studies were used.

[0044] The term “gene expression data” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to all data pertaining to expressed genes in a certain cell or tissue. Typically, gene expression data may be nucleic acid sequences of the transcripts of expressed genes. The said nucleic acid sequences may be obtained by suitable nucleic acid sequencing techniques such as RNA-seq or other deep sequencing technologies. The transcripts may be transcribed RNA, typically mRNA, or all kinds of processed RNA. Based on the nucleic acid sequences of the transcripts of the expressed genes, the said genes may be identified, e.g., by comparing the nucleic acid sequences to those nucleic acid sequences of known genes contained in databases. The expressed gene can be identified based on a sequence match or essentially match in a sequence comparison. The name or designation of the identified expressed gene allocated to the determined nucleic acid sequence may also be part of the gene expression data as referred to herein. Moreover, the gene expression data may contain further information as to the abundance of the transcripts of the expressed gene, i.e. information as to the presence or absence or amount of transcripts found in a cell or tissue. The gene expression data specifically may be or may comprise gene expression data from peripheral blood mononuclear cells (PBMC).

[0045] One or more than one trainable model may be used. Even though the method generally relates to trainable models suited for classifying the patient’s health condition, as will be shown in further detail below, the trainable model specifically may be or may comprise at least one trainable SVM model and / or at least one trainable XGBoost model. These models, for the present application, out of the available trainable model is in standard software libraries, have shown to provide excellent classification results for classifying a patient’s health condition according to systemic inflammatory states.

[0046] As outlined above, the label of the training data specifically may contain information on the respective systemic inflammatory state. Thus, generally, the labelling of the training patient data to obtain the labeled training patient data may contain information on the respective systemic inflammatory state. The labeling may be part of the method or may be part of a preceding step of generating the training data, e.g. during one or more clinical studies, as outlined above and as will be outlined in further detail below.

[0047] The group of predetermined systemic inflammatory states may comprise or may consist of the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock. Thus, the trainable model, when being in a trained state, may be configured for classifying patients suspected to suffer from a systemic inflammatory state. Besides the three categories listed above, i.e. SIRS, sepsis, and septic shock, one or more additional states may be comprised by the predetermined group of the at least three systemic inflammatory states. Thus, as an example and as outlined above, the predetermined group of the at least three systemic inflammatory states, besides the systemic inflammatory states “the patient suffering from SIRS”, “the patient suffering from sepsis”, and “the patient suffering from a septic shock”, may, additionally, contain the state “the patient being in an apparently healthy state”, or, equivalently, “the patient apparently not suffering from any of SIRS, sepsis and septic shock”, notwithstanding the fact that the patient may suffer from other diseases.

[0048] The gene expression data of the labeled training patient data in step b., specifically may be derived from peripheral blood mononuclear cells (PBMC) and are, preferably, obtained by mRNA analysis.

[0049] The training method further specifically may comprise at least one hyperparameter tuning step for tuning hyperparameters of the trainable model. The term “hyperparameter tuning step” as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to the process of adjusting, specifically optimizing, one or more hyperparameters of the trainable model. Thus, trainable models, e.g. trainable models from standard software libraries such as Scikit-learn, typically require one or more hyperparameters to be predetermined before starting the training process. The hyperparameter tuning step may comprise the pre-adjusting of these hyperparameters, e.g. in order to provide a suitable convergence for the training process. As an example, the labeled training data or at least a part thereof may be used for an initial test for training the trainable model, trying different sets of hyperparameters, wherein the performance of the training process is evaluated on the basis of measurable training performance values such as the speed of convergence, and, wherein, the set of hyperparameters providing a superior performance may finally be chosen for the trainable model. The trainable model may then be trained with this specific set of hyperparameters.

[0050] The hyperparameter tuning step, specifically, may be performed at least once before performing step c. It is, however, also possible to implement a plurality of tuning steps, intermitting with one or more partial training steps, e.g. in order to improve or even optimize the tuning of the hyperparameters during the process.

[0051] As outlined above, the trainable model specifically may comprise or specifically may consist of a trainable SVM model. As also outlined in further detail below, the trainable SVM model specifically may be trained using the following hyperparameters:

[0052] Regularization parameter: C: 1.0;

[0053] Kernel: linear;

[0054] Penalty function: LI (Lasso);

[0055] Loss Function: squared hinge. As will be outlined in further detail below, this set of hyperparameters is suited to provide training results leading to a trained SVM model on the basis of the labeled training data derived from the above-mentioned studies, the trained SVM model being capable of classifying the patient’s health condition with high reliability. For further details on these hyperparameters, as an example, reference may be made to the Scikit-learn library.

[0056] As also outlined above, the trainable model, specifically in addition or as an alternative to the trainable SVM model, specifically may comprise or may consist of a trainable XGBoost model, i.e. a trainable model implementing a gradient boosting method. As also outlined in further detail below, the trainable XGBoost model specifically may be trained using the following hyperparameters: booster- gbtree', colsample_bytree= 1.0, gamma= 1, max_depth= 3, min_child_weight= 5, subsample= 0.6.

[0057] As will also be outlined in further detail below, this set of hyperparameters is suited to provide training results leading to a trained XGBoost model on the basis of the labeled training data derived from the above-mentioned studies, the trained XGBoost model being capable of classifying the patient’s health condition with high reliability. For further details on these hyperparameters, as an example, reference may be made to the description of gradient boosting models in the XGBoost library, available e.g. under https: / / xgboost.readthedocs.io / en / stable / index.html. Alternatively, however, other libraries may be used, with comparable hyperparameters, as the skilled person will now, such as the Scikit-learn library.

[0058] The training patient data, as discussed above, specifically may be obtained from one or more patient studies. Specifically, the labeled training patient data may be patient data obtained from a plurality of different studies. More specifically, the labeled training patient data may contain patient data from differing age groups. The label of the training patient data, as outlined above, may contain information on the study. Additionally or alternatively, the label of the training patient data may contain information on the patient, such as on the group the respective patient belongs to. More specifically, the label may contain age information, such as information on the respective age group of the patients. Thus, the labelling of the training patient data to obtain the labeled training data may additionally contain information on the respective age groups of the patients. The training method specifically may comprise splitting the labeled training patient data into a training data set and a test data set. Thus, more specifically, the splitting into a training data set and a test data set may comprise x%-y% splitting of the labeled training patient data, with x% being in the range 60 to 90, specifically in the range 65 to 75, and more specifically x%=70, and with y%=100-x. Therein, x% denotes the training data set and y% denotes the test data set.

[0059] The training method may further compromise a cross-validation in the training data set. Thus, the data of the training data set may be further split into k subsets or folds of equal size, with & being at least 2 or more folds, specifically in the range 3 to 10, more specifically k=5. The training and validation may be repeated k times, and each time the data of a different subset or fold may be used for validation. Additionally or alternatively, the training method may further comprise repeating the splitting of the labeled training data into the training data set and the test data set one or more times, specifically for a number of times in the range of 10 to 40 times, more specifically 20 times. Thus, a validation of the model may be performed with an xi%-yi% splitting of the labeled training patient data, and at least one additional X2%-y2% splitting of the labeled training patient data, with xi and X2 describing different partitions of the labeled training patient data, wherein, specifically, xi%=X2%. As an example, a degree of discrepancy between the results obtained for the multiple iterations may be used for qualifying and / or quantifying the training result.

[0060] Additionally or alternatively, the training method may further comprise repeatedly performing step c. with multiple splittings. The multiple splittings may comprise an xi%-yi% splitting of the labeled training patient data and at least one additional X2%-y2% splitting of the labeled training patient data, with xi and X2 describing different partitions of the labeled training patient data, wherein, specifically, xi%=X2%. Alternatively or additionally, the multiple splittings may comprise differing splittings of the labeled training data, specifically with xi%^X2%. Again, optionally, the training results, e.g. at least one numeric value qualifying and / or quantifying the training results, e.g. qualifying and / or quantifying the accuracy of the training, may be compared.

[0061] As outlined above, the training method specifically may comprise splitting the labeled training patient data into a training data set and a test data set. The training method may, then, further comprises the following step: d. at least one validation step, the validation step comprising testing the accuracy of the training by using the test data set. The testing of the accuracy specifically may comprise using at least one model accuracy metrics. Various model accuracy metrics are generally known in the field of machine learning. Specifically, one or more or even all of the following model accuracy metrics may be used: overall accuracy, balanced accuracy, weighted fl score. Other options, however, are also feasible.

[0062] The training method may further comprise the following step: e. determining at least one target set of relevant genes, by using at least one feature extraction method.

[0063] Again, various feature extraction methods are generally known to the skilled person in the field of machine learning. As an example, the feature extraction method may comprise considering a permutation feature importance. Additionally or alternatively, the importance of the features may be quantified, such as by quantifying the influence a variation of the value of the respective feature has on the accuracy of the model. The feature extraction method may then, as an example, also comprise considering a feature importance hierarchy. Thus, as an example, the feature extraction method may comprise selecting all features having a feature importance of at least one threshold value. Additionally or alternatively, the features may be ordered according to their feature importance, followed by selecting the most important features such that these selected features, in combination, result in a model accuracy metrics of at least one accuracy threshold value. Additionally or alternatively, the feature extraction method may comprise selecting a predetermined number of features highest in ranking with respect to their feature importance. Thus, as an example, the feature extraction method may also comprise, additionally or alternatively, selecting a group of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes, specifically for each age group of patients separately, more specifically at least one of: selecting a predetermined number of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes; selecting the genes of the gene expression data of the patients having a feature importance of at least one threshold value to be the target set of relevant genes; ordering the genes according to their feature importance, followed by selecting the most important genes which, in combination, result in a model accuracy according to a model accuracy metrics of the trained model of at least one accuracy threshold value to be target set of relevant genes; ordering the genes according to their feature importance, followed by selecting the most important genes to be target set of relevant genes, until adding a further gene results in a gain in a model accuracy of the trained model according to a model accuracy metrics which is below a gain threshold. For example, the feature extraction method may comprise ordering the genes according to their feature importance, specifically according to their feature importance with respect to a model accuracy according to a model accuracy metrics of the trained model. Genes having a high feature importance may be ordered above genes having a low feature importance. The group of the most important genes may then be selected such that the group of the most important genes may result in a model accuracy above at least one accuracy threshold value to be achieved. The group of the most important genes may be selected to be the target set of relevant genes. Alternatively or additionally, the ordered genes may be selected in an order beginning with a most important feature followed by selecting the second most important feature and so on. The most important genes may be selected to be the target set of relevant genes such that all of the most important features are selected until adding a further gene results in a gain in a model accuracy of the trained model according to a model accuracy metrics which is below a gain threshold. The further gene may not form part of the selected most important genes.

[0064] Again, additionally or alternatively, the feature extraction method may also comprise selecting a predetermined number of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes. Other means and methods of feature extraction, from the training results, may also be used in addition or as an alternative.

[0065] The feature extraction method may be performed such that a reasonable number of genes of the gene expression data of the patients are selected to be the target set of relevant genes. As an example, for practical applications, selecting a predetermined number of 3 to 20 genes is generally suitable for diagnostic handling, specifically a predetermined number of 5 to 15 genes, more specifically 10 genes.

[0066] In a further aspect of the present invention, a computer-implemented classification method of classifying a patient’s health condition into a systemic inflammatory state, specifically and inflammatory response state, is disclosed. The systemic inflammatory state is selected from a predetermined group of at least three systemic inflammatory states. The classification method comprises the following method steps, which, as an example, may be performed in the given order. However, a different order is also feasible. Further, it is possible to perform two or more of the method steps simultaneously or in a fashion overlapping in time. Further, it is also possible to perform one, more than one or even all of the method steps repeatedly

[0067] The classification method comprises the following steps: i. retrieving at least one trained model; ii. retrieving patient data of the patient, the patient data comprising gene expression data of the patient, the gene expression data comprising at least one predetermined target set of relevant genes; and iii. applying the trained model to the target set of relevant genes of the patient data, thereby assigning a systemic inflammatory state to the patient.

[0068] The trained model specifically may be trained by using the training method according to the present invention, such as according to any one of the embodiments described above and / or according to any one of the embodiments described in further detail below. Thus, for most of the terms, definitions and options of the classification method, reference may be made to the training method as described above or as described in further detail below.

[0069] As outlined in further detail above, the classification method specifically may assign the patient or the patient’s health condition to one of the systemic inflammatory states of the predetermined group of at least three systemic inflammatory states. As an example, the classification method may, by applying the trained model to the patient data, specifically to the relevant genes of the target set of relevant genes of the patient data, assign a specific systemic inflammatory state of the predetermined group of at least three systemic inflammatory states to the specific patient.

[0070] The trained model, as outlined above, specifically may be selected from the group consisting of a trained SVM model and a trained XGBoost model.

[0071] The patient specifically may be a patient suspected to suffer from an inflammatory disease. Specifically, the patient may be a patient having systemic inflammatory symptoms, such as one or more of the symptoms of the inflammatory stages listed above.

[0072] The group of predetermined systemic inflammatory states may be the same as discussed above in the context of the training method. Thus, as also outlined above, the group of predetermined systemic inflammatory states specifically may comprise or even may consist of: the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock. Besides the three categories listed above, i.e. SIRS, sepsis, and septic shock, one or more additional states may be comprised by the predetermined group of the at least three systemic inflammatory states. Thus, as an example and as outlined above, the predetermined group of the at least three systemic inflammatory states, besides the systemic inflammatory states “the patient suffering from SIRS”, “the patient suffering from sepsis”, and “the patient suffering from a septic shock”, may, additionally, contain the systemic inflammatory state “the patient being in an apparently healthy state”, or, equivalently, “the patient apparently not suffering from any of SIRS, sepsis and septic shock”, notwithstanding the fact that the patient may suffer from other diseases. Thus, generally, the group of predetermined systemic inflammatory states may further comprise: the patient being in an apparently healthy state.

[0073] The predetermined target set of relevant genes specifically may be or may comprise the target set of relevant genes determined during the training of the trained model. Thus, specifically, as discussed above, the trained model may be trained by using the training method of the present invention, according to any one of the embodiments described herein. As discussed above in the context of step e. of the training method, a target set of relevant genes may be determined during the training. This target set of relevant genes, in the classification method according to the present invention, may be used as the predetermined target set of relevant genes.

[0074] As will be discussed in further detail below in the context of the experimental data, the target set of relevant genes specifically may be selected from the group consisting of: a first group of relevant genes, comprising, preferably consisting of POMK, S100A6, LARGE2, ZNF492, SPINT1, NACA2, ZNF425, KCNA5, NEK6, and CFAP57; a second group of relevant genes, comprising, preferably consisting of SRPK3, S100A6, NMUR1, SGMS1, TRAF3, NEK6, ZDHHC4, TNFSF8, HIST1H4D, and DHFR; a third group of relevant genes, comprising, preferably consisting of RPL3L, AC0T4, ZNF425, LY86, APBB3, C9orfl35, DAPP1, ZPBP2, UPK1B, and SPDYA.

[0075] The term “relevant gene” as used in accordance with the present invention relates to a gene being suitable as a biomarker itself, e.g., due to its nucleic acid sequence, and / or a gene encoding a protein suitable as a biomarker for an inflammatory state as discussed above. In the latter case, the abundance of the transcripts or protein may serve as a biomarker for an inflammatory state as discussed above.

[0076] The relevant genes of the first, second and third group as referred to herein are listed in the following Tables (first group of relevant genes, second group of relevant genes and third group of relevant genes, respectively). The Tables indicate the abbreviations of the relevant genes, their full designation and an UniProt accession number for each relevant genes which, preferably, shows the human amino acid sequence of the protein encoded by the relevant genes.

[0077] Table: First group of relevant genes (all)

[0078] Table: Second group of relevant genes (adult)

[0079] Table: Third group of relevant genes (juvenile) It will be understood that relevant genes as referred to herein, in addition to those genes encoding the proteins deposited under the UniProt accession numbers also encompass variants of said genes. Such variants, typically, encode proteins that have at least the same essential biological and immunological properties as the aforementioned proteins deposited under the respective UniProt accession numbers. In particular, the said proteins share the same essential biological and immunological properties if they are detectable by the same specific assays referred to in this specification. Moreover, it is to be understood that a variant as referred to in accordance with the present invention shall encode a protein having an amino acid sequence which differs due to at least one amino acid substitution, deletion and / or addition wherein the amino acid sequence of the variant is still, preferably, at least 50%, 60%, 70%, 80%, 85%, 90%, 92%, 95%, 97%, 98%, or 99% identical with the specific amino sequence of the protein as deposited under the indicated UniProt accession number, preferably over the entire length of the said protein, respectively.

[0080] The degree of identity between two amino acid sequences in accordance with the present invention can be determined by algorithms well known in the art. Preferably, the degree of identity is to be determined by comparing two optimally aligned sequences over a comparison window, where the fragment of amino acid sequence in the comparison window may comprise additions or deletions (e.g., gaps or overhangs) as compared to the reference sequence (which does not comprise additions or deletions) for optimal alignment. The percentage is calculated by determining the number of positions at which the identical amino acid residue occurs in both sequences to yield the number of matched positions, dividing the number of matched positions by the total number of positions in the window of comparison and multiplying the result by 100 to yield the percentage of sequence identity. Optimal alignment of sequences for comparison may be conducted by the local homology algorithm disclosed by Smith, by the homology alignment algorithm of Needleman, by the search for similarity method of Pearson, by computerized implementations of these algorithms (GAP, BESTFIT, BLAST, FAST, PASTA, and TFASTA in the Wisconsin Genetics Software Package, Genetics Computer Group (GCG), 575 Science Dr., Madison, WI) or by visual inspection. Given that two sequences have been identified for comparison, GAP and BEST- FIT are preferably employed to determine their optimal alignment and, thus, the degree of identity. Preferably, the default values of 5.00 for gap weight and 0.30 for gap weight length are used. Variants referred to above may be allelic variants or any other species specific homologs, paralogs, or orthologs. Variants referred to above may be allelic variants or any other species specific homologs, paralogs, or orthologs.

[0081] The first group specifically may be used when the age of the patient is disregarded during classification, e.g. in case the age of the patient is unknown. The second group specifically may be used for adult patients. The third group specifically may be used for pediatric patients. Thus, as an example, the classification method as proposed herein may further comprise the step of retrieving the patient’s age and selecting the target set of relevant genes according to the patient’s age. Thus, as an example, the second group may be selected when the retrieved age of the patient indicates that the patient is an adult. Additionally or alternatively, the third group may be selected when the retrieved age of the patient indicates that the patient is a pediatric patient. Thus, generally, the second group may be used when the patient is an adult. The third group may be used when the patient is a child or an adolescent. The patient may be classified as an adult if the age of patient is at least 18 years. The patient may be classified as a child or adolescent if the age of the patient is 17 years or less.

[0082] As outlined above, the gene expression data of the patient data in step ii. specifically may be derived from peripheral blood mononuclear cells (PBMC). Preferably, the gene expression data of the patient data may be obtained by mRNA analysis.

[0083] The classification method may further comprise the following step: iv. outputting the assigned systemic inflammatory state obtained in step iii.

[0084] The output, as an example, may take place in one or more of the following fashions: by electronic output, by data transfer, by data storage, by displaying, by acoustic output, by haptic output, by printing a report or by a combination of one or more or even all of the before-mentioned options. Thus, as an example, the assigned systemic inflammatory state may be displayed on a computer display.

[0085] The classification method may further comprise the following step: v. providing at least one proposal for a therapeutic measure in accordance with the assigned systemic inflammatory state obtained in step iii.

[0086] The providing of the at least one proposal specifically may take place in one or more of the following fashions: by electronic output, by data transfer, by data storage, by displaying, by acoustic output, by haptic output, by printing a report or by a combination of one or more or even all of the before-mentioned options. Thus, as an example, the proposal may be displayed on a computer display. The method may imply using at least one transformation and / or algorithm for generating the at least one proposal depending on the assigned systemic inflammatory state. Thus, as an example, the method may imply using a lookup table having listed therein, one or more therapeutic measures for each systemic inflammatory state, such as one or more recommended therapeutic measures from textbooks.

[0087] In a further aspect of the present invention, a training system is proposed. The training system comprises at least one processor, the processor being configured, e.g. by software programming, for performing the training method according to the present invention, such as according to any one of the embodiments of the training method described above and / or according to any one of the embodiments of the training method described in further detail below.

[0088] In a further aspect of the present invention, a classification system is proposed. The classification system comprises at least one processor, the processor being configured, e.g. by software programming, for performing the classification method according to the present invention, such as according to any one of the embodiments of the classification method described above and / or according to any one of the embodiments of the classification method described in further detail below.

[0089] The term “processor” as generally used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to an arbitrary logic circuitry configured for performing basic operations of a computer or system, and / or, generally, to a device which is configured for performing calculations or logic operations. In particular, the processing unit may be configured for processing basic instructions that drive the computer or system. As an example, the processor may comprise at least one arithmetic logic unit (ALU), at least one floating-point unit (FPU), such as a math co-pro- cessor or a numeric coprocessor, a plurality of registers, specifically registers configured for supplying operands to the ALU and storing results of operations, and a memory, such as an LI and L2 cache memory. In particular, the processor may be a multi-core processor. Specifically, the processing unit may be or may comprise a central processing unit (CPU). Additionally or alternatively, the processor may be or may comprise a microprocessor, thus specifically the processing unit’s elements may be contained in one single integrated circuitry (IC) chip. Additionally or alternatively, the processing unit may be or may comprise one or more application-specific integrated circuits (ASICs) and / or one or more field-programmable gate arrays (FPGAs) or the like. The processing unit specifically may be configured, such as by software programming, for performing one or more evaluation operations.

[0090] The term "system" as used herein is a broad term and is to be given its ordinary and customary meaning to a person of ordinary skill in the art and is not to be limited to a special or customized meaning. The term specifically may refer, without limitation, to an arbitrary set of interacting or interdependent components parts forming a whole. Specifically, the components may interact with each other in order to fulfill at least one common function. The at least two components may be handled independently or may be coupled or connect-able. Consequently, the term “training system” generally refers to a system, as defined above, configured for performing at least one training method. Similarly, the term “classification system” generally relates to a system, as defined above, configured for performing at least one classification method.

[0091] In a further aspect of the present invention, a computer program is proposed, the computer program comprising instructions which, when the program is executed by at least one processor of a computer or a computer network cause the processor to perform the training method according to the present invention, such as according to any one of the embodiments of the training method described above and / or according to any one of the embodiments of the training method described in further detail below.

[0092] In a further aspect of the present invention, a computer program is proposed, the computer program comprising instructions which, when the program is executed by at least one processor of a computer or a computer network cause the processor to perform the classification method according to the present invention, such as according to any one of the embodiments of the classification method described above and / or according to any one of the embodiments of the classification method described in further detail below.

[0093] In a further aspect of the present invention, a computer-readable storage medium is proposed, specifically a non-transient computer-readable storage medium, the computer readable storage medium comprising instructions which, when the instructions are executed by at least one processor of a computer or a computer network cause the processor to perform the training method according to the present invention, such as such as according to any one of the embodiments of the training method described above and / or according to any one of the embodiments of the training method described in further detail below. In a further aspect of the present invention, a computer-readable storage medium is proposed, specifically a non-transient computer-readable storage medium, comprising instructions which, when the instructions are executed by at least one processor of a computer or a computer network cause the processor to perform the classification method according to the present invention, such as according to any one of the embodiments of the classification method described above and / or according to any one of the embodiments of the classification method described in further detail below.

[0094] As outlined above, the methods as proposed herein are computer-implemented. As used herein, the terms “computer-readable data carrier” and “computer-readable storage medium” specifically may refer to non-transitory data storage means, such as a hardware storage medium having stored thereon computer-executable instructions. The computer-readable data carrier or storage medium specifically may be or may comprise a storage medium such as a random-access memory (RAM) and / or a read-only memory (ROM).

[0095] Thus, specifically, the method steps of the methods as indicated above may be performed by using a computer or a computer network, preferably by using a computer program.

[0096] Further disclosed and proposed herein is a computer program product having program code means, in order to perform any one of the methods according to the present invention in one or more of the embodiments enclosed herein when the program is executed on a computer or computer network. Specifically, the program code means may be stored on a computer- readable data carrier and / or on a computer-readable storage medium.

[0097] The methods and devices as proposed herein provide a large number of advantages over methods and devices of similar kind. Specifically, the above-mentioned technical challenges may be addressed, wherein, specifically, the accuracy of diagnosis of systemic inflammatory states may be increased. Further, as opposed at least to most of the known methods, the present invention also allows for distinguishing different inflammatory states, in order to enable the healthcare personnel to administer appropriate treatment. The present invention, thus, specifically may allow for an early diagnosis of sepsis and septic shock, which is crucial to respond rapidly, by appropriate medical treatment.

[0098] Further, as opposed to diagnosis based on single-platform gene expression data, the present invention also allows for combining different platforms from various studies. Thereby, larger and more diverse training data sets can be achieved which generally may lead to more appropriate and more precise machine learning models. Further, the present invention allows for extending traditional 2-class classification models (e.g. sepsis vs. septic shock or sepsis vs. SIRS) to 3- or higher class classification models, such as SIRS vs. sepsis vs. septic shock or similar.

[0099] Summarizing and without excluding further possible embodiments, the following embodiments may be envisaged:

[0100] Embodiment 1 A computer-implemented training method of training at least one trainable model for classifying a patient’s health condition into a systemic inflammatory state, specifically an inflammatory response state, the systemic inflammatory response state being selected from a predetermined group of at least three systemic inflammatory states, the method comprising: a. providing the trainable model; b. retrieving labeled training patient data, the training patient data comprising gene expression data of patients having known systemic inflammatory states; and c. training the trainable model on the labeled training patient data.

[0101] Embodiment 2 The training method according to the preceding embodiment, wherein the trainable model is selected from the group consisting of a trainable SVM model and a trainable XGBoost model.

[0102] Embodiment 3 The training method according to any one of the preceding embodiments, wherein the labelling of the training patient data to obtain the labeled training patient data contains information on the respective systemic inflammatory state.

[0103] Embodiment 4 The training method according to any one of the preceding embodiments, wherein the group of predetermined systemic inflammatory states comprises, preferably consists of: the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock.

[0104] Embodiment 5 The training method according to the preceding embodiment, wherein the group of predetermined systemic inflammatory states further comprises: the patient being in an apparently healthy state.

[0105] Embodiment 6 The training method according to any one of the preceding embodiments, wherein the gene expression data of the labeled training patient data in step b. are derived from peripheral blood mononuclear cells (PBMC) and are, preferably, obtained by mRNA analysis.

[0106] Embodiment 7 The training method according to any one of the preceding embodiments, wherein the training method comprises at least one hyperparameter tuning step for tuning hyperparameters of the trainable model.

[0107] Embodiment 8 The training method according to any one of the preceding embodiments, wherein the trainable model comprises a trainable SVM model, wherein the trainable SVM model is trained using the following hyperparameters:

[0108] Regularization parameter: C: 1.0;

[0109] Kernel: linear;

[0110] Penalty function: LI (Lasso);

[0111] Loss Function: squared hinge.

[0112] Embodiment 9 The training method according to any one of the preceding embodiments, wherein the trainable model comprises a trainable XGBoost model, wherein the train- able XGBoost model is trained using the following hyperparameters: booster- gbtree', colsample_bytree= 1.0, gamma= 1, max_depth= 3, min_child_weight= 5, subsample= 0.6.

[0113] Embodiment 10 The training method according to any one of the preceding embodiments, wherein the labeled training patient data are patient data obtained from a plurality of different studies.

[0114] Embodiment 11 The training method according to any one of the preceding embodiments, wherein the labeled training patient data contain patient data from differing age groups.

[0115] Embodiment 12 The training method according to any one of the preceding embodiments, wherein the labelling of the training patient data to obtain the labeled training data additionally contains information on the respective age groups of the patients. Embodiment 13 The training method according to any one of the preceding embodiments, wherein the training comprises splitting the labeled training patient data into a training data set and a test data set.

[0116] Embodiment 14 The training method according to the preceding embodiment, wherein the splitting into a training data set and a test data set comprises a x%-y% splitting of the labeled training patient data, with x% being in the range 60 to 90, specifically in the range 65 to 75, and more specifically x%= 70, and with y%=100-x.

[0117] Embodiment 15 The training method according to any one of the two preceding embodiments, wherein the method comprises a cross-validation in the training data set.

[0118] Embodiment 16 The training method according to the preceding embodiment, wherein the method comprises repeatedly performing step c. with multiple splittings, and optionally comparing the training results.

[0119] Embodiment 17 The training method according to any one of the four preceding embodiments, wherein the method further comprises: d. at least one validation step, the validation step comprising testing the accuracy of the training by using the test data set.

[0120] Embodiment 18 The training method according to the preceding embodiment, wherein testing of the accuracy comprises using at least one model accuracy metrics, specifically at least one model accuracy metrics selected from the group consisting of overall accuracy, balanced accuracy, and weighted fl score.

[0121] Embodiment 19 The training method according to any one of the preceding embodiments, wherein the method further comprises: e. determining at least one target set of relevant genes, by using at least one feature extraction method.

[0122] Embodiment 20 The training method according to the preceding embodiment, wherein the feature extraction method comprises considering a permutation feature importance. Embodiment 21 The training method according to any one of the two preceding embodiments, wherein the feature extraction method comprises considering a feature importance hierarchy.

[0123] Embodiment 22 The training method according to the preceding embodiment, wherein the feature extraction method comprises selecting a group of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes, specifically for each age group of patients separately, more specifically at least one of selecting a predetermined number of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes; selecting the genes of the gene expression data of the patients having a feature importance of at least one threshold value to be the target set of relevant genes; ordering the genes according to their feature importance, followed by selecting the most important genes which, in combination, result in a model accuracy according to a model accuracy metrics of the trained model of at least one accuracy threshold value to be target set of relevant genes; ordering the genes according to their feature importance, followed by selecting the most important genes to be target set of relevant genes, until adding a further gene results in a gain in a model accuracy of the trained model according to a model accuracy metrics which is below a gain threshold.

[0124] Embodiment 23 The training method according to the preceding embodiment, wherein the feature extraction method comprises selecting a predetermined number of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes.

[0125] Embodiment 24 The training method according to the preceding embodiment, wherein the predetermined number of genes is 3 to 20, specifically 5 to 15, more specifically 10.

[0126] Embodiment 25 A computer-implemented classification method of classifying a patient’s health condition into a systemic inflammatory state, specifically an inflammatory response state, the systemic inflammatory response state being selected from a predetermined group of at least three systemic inflammatory states, the method comprising: i. retrieving at least one trained model; ii. retrieving patient data of the patient, the patient data comprising gene expression data of the patient, the gene expression data comprising at least one predetermined target set of relevant genes; and iii. applying the trained model to the target set of relevant genes of the patient data, thereby assigning a systemic inflammatory state to the patient.

[0127] Embodiment 26 The classification method according to the preceding embodiment, wherein the trained model is selected from the group consisting of a trained SVM model and a trained XGBoost model.

[0128] Embodiment 27 The classification method according to any one of the preceding embodiments referring to a classification method, wherein the patient is a patient having systemic inflammatory symptoms.

[0129] Embodiment 28 The classification method according to any one of the preceding embodiments referring to a classification method, the trained model being trained by using the training method according to any one of the preceding embodiments referring to a training method.

[0130] Embodiment 29 The classification method according to any one of the preceding embodiments referring to a classification method, wherein the group of predetermined systemic inflammatory states comprises, preferably consists of the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock.

[0131] Embodiment 30 The classification method according to the preceding embodiment, wherein the group of predetermined systemic inflammatory states further comprises: the patient being in an apparently healthy state.

[0132] Embodiment 31 The classification method according to any one of the preceding embodiments referring to a classification method, wherein the predetermined target set of relevant genes is selected from the group of target sets consisting of: a first group, comprising, preferably consisting of: POMK, S100A6, LARGE2, ZNF492, SPINT1, NACA2, ZNF425, KCNA5, NEK6, and CFAP57; a second group, comprising, preferably consisting of: SRPK3, S100A6, NMUR1, SGMS1, TRAF3, NEK6, ZDHHC4, TNFSF8, HIST1H4D, and DHFR; a third group, comprising, preferably consisting of: RPL3L, ACOT4, ZNF425, LY86, APBB3, C9orfl35, DAPP1, ZPBP2, UPK1B, and SPDYA. Embodiment 32 The classification method according to the preceding embodiment, wherein the second group is used when the patient is an adult.

[0133] Embodiment 33 The classification method according to any one of the two preceding embodiments, wherein the third group is used when the patient is a child or an adolescent.

[0134] Embodiment 34 The classification method according to any one of the preceding embodiments referring to a classification method, wherein the gene expression data of the patient data in step ii. are derived from peripheral blood mononuclear cells (PBMC) and are, preferably, obtained by mRNA analysis.

[0135] Embodiment 35 The classification method according to any one of the preceding embodiments referring to a classification method, further comprising: iv. outputting the assigned systemic inflammatory state obtained in step iii..

[0136] Embodiment 36 The classification method according to any one of the preceding embodiments referring to a classification method, further comprising: v. providing at least one proposal for a therapeutic measure in accordance with the assigned systemic inflammatory state obtained in step iii.

[0137] Embodiment 37 A training system comprising at least one processor, the processor being configured for performing the training method according to any one of the preceding embodiments referring to a training method.

[0138] Embodiment 38 A classification system comprising at least one processor, the processor being configured for performing the classification method according to any one of the preceding embodiments referring to a classification method.

[0139] Embodiment 39 A computer program comprising instructions which, when the program is executed by at least one processor of a computer or a computer network cause the processor to perform the training method according to any one of the preceding embodiments referring to a training method.

[0140] Embodiment 40 A computer program comprising instructions which, when the program is executed by at least one processor of a computer or a computer network cause the processor to perform the classification method according to any one of the preceding embodiments referring to a classification method.

[0141] Embodiment 41 A computer-readable storage medium, specifically a non-transient computer-readable storage medium, comprising instructions which, when the instructions are executed by at least one processor of a computer or a computer network cause the processor to perform the training method according to any one of the preceding embodiments referring to a training method.

[0142] Embodiment 42 A computer-readable storage medium, specifically a non-transient computer-readable storage medium, comprising instructions which, when the instructions are executed by at least one processor of a computer or a computer network cause the processor to perform the classification method according to any one of the preceding embodiments referring to a classification method.

[0143] Short description of the Figures

[0144] Further optional features and embodiments will be disclosed in more detail in the subsequent description of embodiments, preferably in conjunction with the dependent claims. Therein, the respective optional features may be realized in an isolated fashion as well as in any arbitrary feasible combination, as the skilled person will realize. The scope of the invention is not restricted by the preferred embodiments. The embodiments are schematically depicted in the Figures. Therein, identical reference numbers in these Figures refer to identical or functionally comparable elements.

[0145] In the Figures:

[0146] Figure 1 shows a flow chart of an embodiment of a computer-implemented training method and of an embodiment of a computer-implemented classification method;

[0147] Figure 2 shows a flow chart of an embodiment of a computer-implemented training method;

[0148] Figure 3 shows a diagram of a performance of different trainable models;

[0149] Figure 4 shows a flow chart of machine learning experiments; Figure 5 shows different accuracy metrics for different machine learning experiments with a trainable SVM model;

[0150] Figures 6A to 6C show the confusion matrices for the machine learning experiments; and

[0151] Figures 7A and 7B show different accuracy metrics for different experiments with a trainable XGBoost model (Figure 7A) and a trainable SVM model (Figure 7B).

[0152] Detailed description of the embodiments

[0153] Figure 1 shows a flow chart of an exemplary embodiment of a computer-implemented training method (denoted by reference number 110) and of an exemplary embodiment of a computer-implemented classification method (denoted by reference number 112). The computer- implemented training method 110 is a computer-implemented method of training at least one trainable model. The trainable model is configured for classifying a patient’s health condition into a systemic inflammatory state, specifically an inflammatory response state, the systemic inflammatory state being selected from a predetermined group of at least three systemic inflammatory states. The group of predetermined systemic inflammatory states may comprise or may consist of: the patient suffering from SIRS; the patient suffering from sepsis; the patient suffering from a septic shock. Additionally, one or more additional states may be comprised by the predetermined group of the at least three systemic inflammatory states. Thus, as an example and as outlined above, the predetermined group of the at least three systemic inflammatory states, besides the systemic inflammatory states “the patient suffering from SIRS”, “the patient suffering from sepsis”, and “the patient suffering from a septic shock”, may, additionally, contain the state “the patient being in an apparently healthy state”, or, equivalently, “the patient apparently not suffering from any of SIRS, sepsis and septic shock”, notwithstanding the fact that the patient may suffer from other diseases.

[0154] The training method 110 and / or the classification method 112 comprises the following method steps, which, as an example, may be performed in the given order. However, a different order is also feasible. Further, it is possible to perform two or more of the method steps simultaneously or in a fashion overlapping in time. Further, it is also possible to perform one, more than one or even all of the method steps repeatedly. The training method 110 comprises the following steps: a. (denoted by reference number 114) providing the trainable model; b. (denoted by reference number 116) retrieving labeled training patient data, the training patient data comprising gene expression data of patients having known systemic inflammatory states; and c. (denoted by reference number 118) training the trainable model on the labeled training patient data.

[0155] One or more than one trainable model may be used. Even though the method generally relates to trainable models suited for classifying the patient’s health condition, e.g. trainable models listed above, the trainable model specifically may be or may comprise at least one trainable SVM model and / or at least one trainable XGBoost model. These models, for the present application, out of the available trainable model is in standard software libraries, have shown to provide excellent classification results for classifying a patient’s health condition according to systemic inflammatory states. Examples of a trainable SVM model and a trainable XGBoost model are given in the following Figures.

[0156] The labeled training patient data may be obtained by using data from a plurality of patients, such as by using data from one or more clinical studies. As an example, data from at least 100 patients, specifically data from at least 500 patients or more, may be used for generating the labeled training patient data. For, each patient, firstly, gene expression data may be generated. Further, for each patient, information on the systemic inflammatory state of the respective patient may be obtained, such as the patient suffering from SIRS, the patient suffering from sepsis, the patient suffering from a septic shock. Additionally, the following state may also be comprised: the patient being in an apparently healthy state. As outlined above, diagnostic indicators may be used other than gene expression data for obtaining the systemic inflammatory state of the respective patient, such as by detecting the presence or absence of the characteristic symptoms for each of the systemic inflammatory states. Then, for each patient, the gene expression data are labeled at least with the respective information on the systemic inflammatory state of the respective patient, in order to obtain a labeled training data set for the respective patient. In addition, and optionally, additional information on the respective patient may be comprised in the labeling, such as information on the age of the patient, information on the gender of the patient, medical information on the patient and the like. The aggregate of the labeled training data set of the patients may then form part of the labeled training patient data. The gene expression data of the labeled training patient data in step b., specifically may be derived from peripheral blood mononuclear cells (PBMC) and are, preferably, obtained by mRNA analysis.

[0157] The training method 110 specifically may comprise splitting the labeled training patient data into a training data set and a test data set. Thus, more specifically, the splitting into a training data set and a test data set may comprise a x%-y% splitting of the labeled training patient data, with x% being in the range 60 to 90, specifically in the range 65 to 75, and more specifically x%= 70, and with y%=100-x. Therein, x% denotes the training data set and y% denotes the test data set.

[0158] The training method 110 may further compromise a cross-validation in the training data set. Thus, the data of the training data set may be further split into k subsets or folds of equal size, with & being at least 2 or more folds, specifically in the range 3 to 10, more specifically k=5. The training and validation may be repeated k times, and each time the data of a different subset or fold may be used for validation. Additionally or alternatively, the training method may further comprise repeating the splitting of the labeled training data into training data set and test data set one or more times, specifically for a number of times in the range 10 to 40 times, more specifically 20 times. Thus, a validation of the model may be performed with an xi%-yi% splitting of the labeled training patient data, and at least one additional X2%-y2% splitting of the labeled training patient data, with xi and x2describing different partitions of the labeled training patient data, wherein, specifically, xi%=x2%. As an example, a degree of discrepancy between the results obtained for the multiple iterations may be used for qualifying and / or quantifying the training result.

[0159] As shown in Figure 1, the training method 110 specifically may comprise splitting the labeled training patient data into a training data set and a test data set. The training method may, then, further comprises the following step: d. (denoted by reference number 120) at least one validation step, the validation step comprising testing the accuracy of the training by using the test data set.

[0160] The testing of the accuracy specifically may comprise using at least one model accuracy metrics. Various model accuracy metrics are generally known in the field of machine learning. Specifically, one or more or even all of the following model accuracy metrics may be used: overall accuracy, balanced accuracy, weighted fl score. Other options, however, are also feasible. Further, in the exemplary embodiment of Figure 1, the training method 110 may comprise the following step: e. (denoted by reference number 122) determining at least one target set of relevant genes, by using at least one feature extraction method.

[0161] Various feature extraction methods are generally known to the skilled person in the field of machine learning. As an example, the feature extraction method may comprise considering a permutation feature importance. Additionally or alternatively, the importance of the features may be quantified, such as by quantifying the influence a variation of the value of the respective feature has on the accuracy of the model. The feature extraction method may then, as an example, also comprise considering a feature importance hierarchy. Thus, as an example, the feature extraction method may comprise selecting all features having a feature importance of at least one threshold value. Additionally or alternatively, the features may be ordered according to their feature importance, followed by selecting the most important features such that these selected features, in combination, result in a model accuracy metrics of at least one accuracy threshold value. Additionally or alternatively, the feature extraction method may comprise selecting a predetermined number of features highest in ranking with respect to their feature importance. Thus, as an example, the feature extraction method may also comprise, additionally or alternatively, selecting a group of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes, specifically for each age group of patients separately, more specifically at least one of selecting a predetermined number of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes; selecting the genes of the gene expression data of the patients having a feature importance of at least one threshold value to be the target set of relevant genes; ordering the genes according to their feature importance, followed by selecting the most important genes which, in combination, result in a model accuracy according to a model accuracy metrics of the trained model of at least one accuracy threshold value to be target set of relevant genes; ordering the genes according to their feature importance, followed by selecting the most important genes to be target set of relevant genes, until adding a further gene results in a gain in a model accuracy of the trained model according to a model accuracy metrics which is below a gain threshold. Again, additionally or alternatively, the feature extraction method may also comprise selecting a predetermined number of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes. For example, the feature extraction method may comprise ordering the genes according to their feature importance, specifically accord- ing to their feature importance with respect to a model accuracy according to a model accuracy metrics of the trained model. Genes having a high feature importance may be ordered above genes having a low feature importance. The group of the most important genes may then be selected such that the group of the most important genes may result in a model accuracy above at least one accuracy threshold value to be achieved. The group of the most important genes may be selected to be the target set of relevant genes. Alternatively or additionally, the ordered genes may be selected in an order beginning with a most important feature followed by selecting the second most important feature and so on. The most important genes may be selected to be the target set of relevant genes such that all of the most important features are selected until adding a further gene results in a gain in a model accuracy of the trained model according to a model accuracy metrics which is below a gain threshold. The further gene may not form part of the selected most important genes. Other means and methods of feature extraction, from the training results, may also be used in addition or as an alternative.

[0162] The feature extraction method may be performed such that a reasonable number of genes of the gene expression data of the patients are selected to be the target set of relevant genes. As an example, for practical applications, selecting a predetermined number of 3 to 20 genes is generally suitable for diagnostic handling, specifically a predetermined number of 5 to 15 genes, more specifically 10 genes.

[0163] Further, as outlined above, Figure 1 shows an exemplary embodiment of the computer-implemented classification method 112 of classifying a patient’s health condition into a systemic inflammatory state, specifically an inflammatory response state. The systemic inflammatory state is selected from a predetermined group of at least three systemic inflammatory states, specifically, in this example, comprising the systemic inflammatory states as described above.

[0164] The classification method 112 comprises the following steps: i. (denoted by reference number 124) retrieving at least one trained model; ii. (denoted by reference number 126) retrieving patient data of the patient, the patient data comprising gene expression data of the patient, the gene expression data comprising at least one predetermined target set of relevant genes; and iii. (denoted by reference number 128) applying the trained model to the target set of relevant genes of the patient data, thereby assigning a systemic inflammatory state to the patient. Specifically, in the example of Figure 1, the trained model being retrieved in step i. of the classification method 112 may be the trainable model being trained by using the training method 110 as described above. However, in principle, in other embodiments of the classification method 112, the trained model may be retrieved independent from the training method 110.

[0165] Further, as shown in Figure 1, the classification method 112 may comprise the following step: iv. (denoted by reference number 130) outputting the assigned systemic inflammatory state obtained in step iii.

[0166] The output, as an example, may take place in one or more of the following fashions: by electronic output, by data transfer, by data storage, by displaying, by acoustic output, by haptic output, by printing a report or by a combination of one or more or even all of the before-mentioned options. Thus, as an example, the assigned systemic inflammatory state may be displayed on a computer display.

[0167] Additionally, as can be seen in Figure 1, the classification method 112 may further comprise the following step: v. (denoted by reference number 132) providing at least one proposal for a therapeutic measure in accordance with the assigned systemic inflammatory state obtained in step iii.

[0168] The providing of the at least one proposal specifically may take place in one or more of the following fashions: by electronic output, by data transfer, by data storage, by displaying, by acoustic output, by haptic output, by printing a report or by a combination of one or more or even all of the before-mentioned options. Thus, as an example, the proposal may be displayed on a computer display. The method may imply using at least one transformation and / or algorithm for generating the at least one proposal depending on the assigned systemic inflammatory state. Thus, as an example, the method may imply using a lookup table having listed therein, one or more therapeutic measures for each systemic inflammatory state, such as one or more recommended therapeutic measures from textbooks.

[0169] Further details and possible embodiments of the training method 110 and / or the classification method 112 will be discussed with regard to the following Figures. Thus, for details and / or possible embodiments of the training method 110 and / or the classification method 112, reference is made to the description of the following Figures. In Figures 2 to 6C, a further example of the present invention is shown. Therein, Figure 2 shows a flow chart of a further embodiment of the computer-implemented training method 110. In this example, the training method 110 may provide a machine learning algorithm based on gene expression data of patients from different public studies, such as from different microarrays, to classify systemic inflammatory states, e.g. sepsis conditions, specifically the systemic inflammatory states described above, e.g. “the patient suffering from SIRS”, “the patient suffering from sepsis”, and “the patient suffering from a septic shock”, and, in particular, to find relevant biomarkers. Further, different training methods 110 were performed for different age groups, such as pediatric, adults, as will be outlined in further details below.

[0170] The training method 110, in this example, comprises providing the trainable model (step a., denoted by reference number 114) and retrieving labeled training patient data (step b., denoted by reference number 116). Further, the training method 110 may comprise, after performing steps a. and b., at least one hyperparameter tuning step for tuning hyperparameters of the trainable model (denoted by reference number 133). The hyperparameter tuning step 133, as shown in Figure 2, may be performed at least once before performing step c. It is, however, also possible to implement a plurality of tuning steps, intermitting with one or more partial training steps, e.g. in order to improve or even optimize the tuning of the hyperparameters during the process.

[0171] The training method 110 may then proceed with training the trainable model in step c. (denoted by reference number 118), which is denoted in the following also as “machine learning experiment”, the at least one validation step (denoted by reference number 120), and the determining of the at least one target set of relevant genes, by using the at least one feature extraction method (denoted by reference number 122). In this example, the feature extraction method may specifically comprise considering a permutation feature importance, as will be outlined in further detail below.

[0172] In the example shown in Figures 2 to 6C, labeled patient data of 1024 patients from 27 individual studies were used. The gene expression data of the labeled training patient data in step b., specifically may be derived from peripheral blood mononuclear cells (PBMC) and are, preferably, obtained by mRNA analysis. The gene expression data in this example comprised 17808 different genes derived from PBMC. The patient data were obtained via the following GEO accession codes / study codes: GEO accession codes / study codes and download links:

[0173] GSE8121 : https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE8121

[0174] GSE9692: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE9692

[0175] GSE11755: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 1755

[0176] GSE13904: https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE13904

[0177] GSE57065: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE57065

[0178] GSE66099: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE66099

[0179] GSE26378: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE26378

[0180] GSE26440: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE26440

[0181] GSE27131 : https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE27131

[0182] GSE22098: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE22098

[0183] GSE63042: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE63042

[0184] GSE95233: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE95233

[0185] GSE9960: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE9960

[0186] GSE28750: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE28750

[0187] GSE131411 : https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE131411 GSE110487: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 10487

[0188] E-MTAB-1548: https: / / www.ebi.ac.uk / biostudies / arrayexpress / studies / E-MTAB-1548 GSE74224: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE74224

[0189] E-MTAB-4421 : https: / / www.ebi.ac.uk / biostudies / arrayexpress / studies / E-MTAB- 4421 ?query=E-MT AB-4421

[0190] SCP548: https: / / singlecell.broadinstitute.org / single_cell / study / SCP548 / an-immune-cell- signature-of-bacterial-sepsis-patient-pbmcs

[0191] GSE25504: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE25504

[0192] GSE54514: https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE54514

[0193] GSE69528: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE69528 GSE119217: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 19217 GSE157103: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 57103 GSE154998: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSEl 54998 GSE149689: https: / / www.ncbi. nlm.nih.gov / geo / query / acc.cgi?acc=GSE149689

[0194] Further, in the example shown in Figures 2 to 6C, a pre-study was performed comparing different trainable models, such as a decision tree model (denoted by reference number 134), a Random Forest model (denoted by reference number 136), a light GBM classifier (denoted by reference number 138), a KNeighbors Classifier (denoted by reference number 140), a Support Vector Machine (SVM) model (denoted by reference number 142), a GaussianNBQ model (denoted by reference number 144) and a LogisticRegression model (denoted by reference number 146). The results of this pre-study are shown in Figure 3. Specifically, Figure 3 shows a diagram with the performance different trainable models. The performance of the trainable models were evaluated using the balanced accuracy (denoted by reference number 145) and the overall accuracy (denoted by reference number 147). For the pre-study, the labeled patient data were split into 70% training data and 30% testing data. The training data were used to find the best model parameters using a 5-fold cross validation. The training data set and the testing data set were normalized. The splitting into 70% training data and 30% testing data was repeated 20 times, specifically with each split having the 5-fold cross- validation in each set of training data. The pre-study, as can be seen in Figure 3, showed that the SVM model performs better than the other tested trainable models. Thus, for the machine learning experiments of Figures 4 to 6C, the SVM model was selected for the trainable model. The trainable SVM model was trained using the following hyperparameters:

[0195] Regularization parameter: C: 1.0;

[0196] Kernel: linear;

[0197] Penalty function: LI (Lasso);

[0198] Loss Function: squared hinge.

[0199] In this example, several machine learning experiments were defined according to Table 1 :

[0200] Table 1: Overview on the machine learning experiment and studies.

[0201] A flow chart of the machine learning experiments is shown in Figure 4. For the machine learning experiments, a naive approach was used. The machine learning experiments, i.e. the training in step c. of the training method 110, may provide the best trained SVM model for each use case, e.g. for all data or stratified by age group, based on the normalized patient data. Prior to the machine learning experiment, as shown in Figure 2, the hyperparameter tuning step 133, e.g. a grid search, may be performed to identify the best model parameters for the linear SVM using cross-validation.

[0202] Further, the machine learning experiment of the example shown in Figures 2 to 6C comprised splitting the labeled training patient data into the training data set and the test data set (denoted by reference number 148). In this example, the splitting comprised splitting the labeled training patient data into 70% of training data and 30% of test data. Further, a hyperparameter tuning step was used (denoted by reference number 150).

[0203] As shown in Figure 4, the machine learning experiment further comprised training and testing the trainable SVM model (denoted by reference number 152). In a training and testing step, generally, several model accuracy metrics may be computed. For example, one or more of the overall accuracy, balanced accuracy, and weighted fl score may be computed. Finally, after the training and testing step 152, the machine learning experiment of the example shown in Figures 2 to 6C comprised determining of the at least one target set of relevant genes, by using the at least one feature extraction method (denoted by reference number 122), which, in the example shown in Figures 2 to 6C, specifically comprised considering a permutation feature importance, as will be outlined in further detail below. Each of these steps may be performed repeatedly, specifically, in this example, for a number of 20 times (denoted by reference number 153).

[0204] The results of the machine learning experiments are shown in Figures 5 to 6C. Specifically, Figure 5 shows different accuracy metrics for different experiments with the trainable SVM model. Figure 5 shows the overall accuracy 154, the balanced accuracy 156, and the weighted fl score 158 for the different machine learning experiments performed in this example. Specifically, Figure 5 shows these model accuracy metrics for the experiments “PEDIATRIC” (denoted by reference number 160), “ALL STUDIES” (denoted by reference number 164) and “ADULT” (denoted by reference number 166). As can be seen in Figure 5, the trained SVM model achieves accuracy scores of 0.8 or higher, except for the balanced accuracy of the PEDIATRIC experiment.

[0205] Figures 6A to 6C show the confusion matrices for the machine learning experiments of Figure 4. Specifically, , Figure 6A shows the confusion matrix for the machine learning experiment “ALL STUDIES”, Figure 6B shows the confusion matrix for the machine learning experiment “ADULT” and Figure 6C shows the confusion matrix for the machine learning experiment “PEDIATRIC”. In the Figures 6 A to 6C, the rows of the confusion matrices represent the true systemic inflammatory states “the patient suffering from sepsis” (denoted by reference number 168), “the patient suffering from SIRS” (denoted by reference number 170), and “the patient suffering from a septic shock” (denoted by reference number 172), wherein the columns represent the predicted systemic inflammatory states. As can be seen in Figures 6A to 6C, the confusion matrices are normalized confusion matrices, thus, representing the respective relative frequency. Further, in the example shown in Figures 2 to 6C, as outlined above, the determining of the at least one target set of relevant genes, by using the at least one feature extraction method comprised considering a permutation feature importance. The permutation feature importance provided the most relevant genes. As outlined above, the splitting was repeated 20 times and, thus, weights of each gene were summed and ranked accordingly. In order to identify the most relevant genes, permutation feature importance was performed for each machine learning experiment. The permutation feature importance generally may comprise a decrease in a model score when a single gene value is randomly shuffled. This shuffle may break the relationship between the gene and the target and, thus, the decrease in the model score may be indicative of how much the model depends on the gene. The permutation feature importance may benefit from being model agnostic and can be calculated many times with different permutations of the gene. For determining the permutation feature importance, all genes (PBMC) which are in common between the studies were considered as input features (n=17808).

[0206] For identifying the relevant genes, the following approach was used: The feature importance was determined based on permutation feature importance from each of the splittings. Specifically, the 70% training data and 30% testing data split was repeated for a number of 20 times, and the sum of weights for each splitting was computed for the full PBMC dataset, and the top ten highest genes in ranking were identified for each machine learning experiment.

[0207] The relevant genes, which then may be used as a target set of relevant genes in the classification method 112, for the example shown in Figures 2 to 6C, is shown in Table 2.

[0208] Table 2: The relevant genes determined by using the example shown in Figures 2 to 6C.

[0209] The training method 110, in this example, was performed using different trainable models as a starting point. Although Artificial Neural Networks were not tested, yet, in principle, they are also feasible for being used as the trainable model for the training method 110 and / or the classification method 112 according to the present invention. Further, the training method 110 in this example used a 5-fold cross validation approach by not considering a certain percentage of the labeled patient data for training in order to test the trainable model, wherein the splitting of the labeled patient data was repeated for a number of 20 times. Alternatively or additionally, a study wise validation may also be performed for testing the trainable model.

[0210] Figures 7A and 7B show further experiments comparing training results for a trainable XGBoost model with training results for a trainable SVM model. The experiments were performed on the basis of the training data used in Figures 2 to 6C above, so reference regarding the studies and the training data may be made to the description of these Figures. Therein, accuracy metrics for different experiments with training a trainable XGBoost model (Figure 7A) and training a trainable SVM model (Figure 7B) are shown. Specifically, Figures 7A and 7B show the overall accuracy 154, the balanced accuracy 156, and the weighted fl score 158 for the machine learning experiments “PEDIATRIC” (denoted by reference number 160), “ALL STUDIES” (denoted by reference number 164) and “ADULT” (denoted by reference number 166). As outlined above, the labeled patient data were the same used for the example of Figures 2 to 6C. However, in this second example, a trainable XGBoost model was used as the trainable model. The trainable XGBoost model was trained using the following hyperparameters: booster- gbtree', colsample_bytree= 1.0, gamma= 1, max_depth= 3, min_child_weight= 5, subsample= 0.6.

[0211] The results of Figures 7A and 7B are show in comparison in Table 3:

[0212] Table 3: Accuracy results for a training of a trainable SVM model and a trainable XGBoost model.

[0213] These results show that, after training, the trained XGBoost model is basically as accurate for use as trained models in a classification method of classifying a patient’s health condition into a systemic inflammatory state as the SVM model, so both models may generally be used.

[0214] List of reference numbers training method classification method providing the trainable model retrieving labeled training patient data training the trainable model validation step determining a target set of relevant genes retrieving a trained model retrieving patient data of the patient applying the trained model to the target set of relevant genes outputting the assigned systemic inflammatory state providing at least one proposal hyperparameter tuning step decision tree model

[0215] Random Forest model light GBM classifier

[0216] KNeighbors Classifier

[0217] SVM model

[0218] GaussianNB() model balanced accuracy

[0219] LogisticRegression model overall accuracy splitting the labeled training patient data intermitting hyperparamter tuning step training and testing step repetitions overall accuracy balanced accuracy weighted fl score

[0220] PEDIATRIC

[0221] ALL STUDIES

[0222] ADULT the patient suffering from sepsis the patient suffering from SIRS the patient suffering from a septic shock

Claims

F. Hoffmann-La Roche AG September 13, 2024Roche Diagnostics GmbH RD38192PC ST / DBRoche Diagnostics Operations, Inc.Claims1. A computer-implemented training method (110) of training at least one trainable model for classifying a patient’s health condition into a systemic inflammatory state (168, 170, 172) selected from a predetermined group of at least three systemic inflammatory states (168, 170, 172), the method comprising: a. providing the trainable model; b. retrieving labeled training patient data, the training patient data comprising gene expression data of patients having known systemic inflammatory states (168, 170, 172); and c. training the trainable model on the labeled training patient data.

2. The training method (110) according to the preceding claim, wherein the trainable model is selected from the group consisting of a trainable SVM model and a trainable XGBoost model.

3. The training method (110) according to any one of the preceding claims, wherein the group of predetermined systemic inflammatory states (168, 170, 172) comprises: the patient suffering from SIRS (170); the patient suffering from sepsis (168); the patient suffering from a septic shock (172).

4. The training method (110) according to any one of the preceding claims, wherein the gene expression data of the labeled training patient data in step b. are derived from peripheral blood mononuclear cells (PBMC) and are obtained by mRNA analysis.

5. The training method (110) according to any one of the preceding claims, wherein the labeled training patient data are patient data obtained from a plurality of different studies.

6. The training method (110) according to any one of the preceding claims, wherein the training comprises splitting the labeled training patient data into a training data set and a test data set, wherein the method further comprises:d. at least one validation step, the validation step comprising testing the accuracy of the training by using the test data set.

7. The training method (110) according to any one of the preceding claims, wherein the method further comprises: e. determining at least one target set of relevant genes, by using at least one feature extraction method.

8. The training method (110) according to the preceding claim, wherein the feature extraction method comprises considering a feature importance hierarchy, wherein the feature extraction method comprises selecting a group of genes of the gene expression data of the patients with the highest ranking in the feature importance hierarchy to be the target set of relevant genes.

9. A computer-implemented classification method (112) of classifying a patient’s health condition into a systemic inflammatory state (168, 170, 172) selected from a predetermined group of at least three systemic inflammatory states (168, 170, 172), the method comprising: i. retrieving at least one trained model; ii. retrieving patient data of the patient, the patient data comprising gene expression data of the patient, the gene expression data comprising at least one predetermined target set of relevant genes; and iii. applying the trained model to the target set of relevant genes of the patient data, thereby assigning a systemic inflammatory state (168, 170, 172) to the patient.

10. The classification method (112) according to the preceding claim, wherein the trained model is selected from the group consisting of a trained SVM model and a trained XGBoost model.

11. The classification method (112) according to any one of the preceding claims referring to a classification method (112), the trained model being trained by using the training method according to any one of the preceding claims referring to a training method.

12. The classification method (112) according to any one of the preceding claims referring to a classification method (112), wherein the predetermined target set of relevant genes is selected from the group of target sets consisting of:- a first group, comprising POMK, S100A6, LARGE2, ZNF492, SPINT1, NACA2, ZNF425, KCNA5, NEK6, and CFAP57;- a second group, comprising SRPK3, S100A6, NMUR1, SGMS1, TRAF3, NEK6, ZDHHC4, TNFSF8, HIST1H4D, and DHFR;- a third group, comprising RPL3L, ACOT4, ZNF425, LY86, APBB3, C9orfl35, DAPP1, ZPBP2, UPK1B, and SPDYA.

13. The classification method (112) according to any one of the preceding claims referring to a classification method, further comprising: iv. outputting the assigned systemic inflammatory state (168, 170, 172) obtained in step iii.

14. The classification method (112) according to any one of the preceding claims referring to a classification method (112), further comprising: v. providing at least one proposal for a therapeutic measure in accordance with the assigned systemic inflammatory state (168, 170, 172) obtained in step iii.

15. A training system comprising at least one processor, the processor being configured for performing the training method (110) according to any one of the preceding claims referring to a training method (110).

16. A classification system comprising at least one processor, the processor being configured for performing the classification method (112) according to any one of the preceding claims referring to a classification method (112).

17. A computer program comprising instructions which, when the program is executed by at least one processor of a computer or a computer network cause the processor to perform the training method (110) according to any one of the preceding claims referring to a training method (110).

18. A computer program comprising instructions which, when the program is executed by at least one processor of a computer or a computer network cause the processor to perform the classification method (112) according to any one of the preceding claims referring to a classification method (112).

19. A computer-readable storage medium comprising instructions which, when the instructions are executed by at least one processor of a computer or a computer network cause the processor to perform the training method (110) according to any one of the preceding claims referring to a training method (110).

20. A computer-readable storage medium comprising instructions which, when the instructions are executed by at least one processor of a computer or a computer network cause the processor to perform the classification method (112) according to any one of the preceding claims referring to a classification method (112).

Images