Sepsis mortality prediction model

Abstract

The present disclosure describes techniques for generating a sepsis mortality prediction model. The sepsis mortality prediction model can predict tire likelihood of sepsis mortality for a subject. The sepsis mortality prediction model can be generated by using feature selection to select a model subset of genetic marker(s) from training data. Candidate sepsis mortality prediction models can be trained using the model subset. The sepsis mortality prediction model can be selected from the candidate prediction models based on determining a performance metric associated with each candidate prediction using a test subset of the training data.

Description

SEPSIS MORTALITY PREDICTION MODELSTATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0001] This invention was made with government support under N626451920001 awarded by the Naval Medical Logistics Command. The government has certain rights in tire invention.CROSS REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Patent Application 63 / 578,492, filed on August 24, 2023, which is hereby incorporated by reference in its entirety.FIELD OF THE DISCLOSURE

[0003] Described herein are methods, systems, and computational environments for stratifying individuals with sepsis or at risk of developing sepsis, and for predicting severe disease in individuals with sepsis or at risk of developing sepsis. Also described are systems and methods for generating topological networks and clusters identifying disease-response phenotypes, systems and methods for selecting prognostic or diagnostic features and host biomarkers, and systems and methods for predicting clinical outcomes. Also described are methods of detecting panels of host biomarkers, methods of assessing risk factors in an individual with sepsis or at risk of developing sepsis, and methods of treating a patient determined to have an elevated risk of severe disease from sepsis.BACKGROUND

[0004] Expeditious and accurate information for clinical decision-making is critical for improving outcomes for infectious disease patients, particularly if a dysregulated host response to the infection leads to the potentially life-threatening organ dysfunction known as sepsis. Early recognition and characterization of an infection and the ensuing host response are essential components for preventing the development and / or mitigating the severity of sepsis. However, current diagnostic and prognostic assays are either insensitive or not expediently useful, if available at all. Tire use of specific host response biomarkers can improve our ability to quickly and accurately phenotype infectious disease states and predict their clinical course. This will be highly informative not just in traditional clinical settings, but also in low resource environments, military' operations, and for at-home monitoring.SUMMARY OF THE DISCLOSURE

[0005] This Summary is provided to introduce a selection of concepts in a simplified form that arc further described below in the Detailed Description. This Summary' is not intended to identify all key features or essential features of the claimed subject matter, nor is it intended to be used alone as an aid in detennining the scope of the claimed subject matter.

[0006] Techniques for determining a prediction associated with sepsis mortality of a subject and / or a treatment recommendation using a sepsis mortality prediction model and for generating the sepsis mortality prediction model are discussed herein.

[0007] In embodiments, there are provided a method of generating a sepsis mortality prediction model for a subject comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of the plurality' of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality' of trained candidate prediction models as tire sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis.

[0008] In embodiments, there are provided a method of predicting sepsis mortality of a subject comprising: receiving value of a patient parameter associated with the subject; executing a sepsis mortality prediction model using the value to generate a prediction associated with sepsis mortality, wherein the sepsis mortality prediction model is generated by performing operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy' metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset; detennining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trainedcandidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis; and outputting, by the sepsis mortality prediction model based on value, the prediction associated with the onset of sepsis mortality within the threshold time after presentation of suspected sepsis in the subject.

[0009] In embodiments, there are provided a system for generating a sepsis mortality prediction model for a subject comprising: one or more processors; and one or more non-transitory computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perform operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, tire model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset; detennining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis.

[0010] In embodiments, there are provided a system for predicting sepsis mortality of a subject comprising: one or more processors; an output component; and one or more computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perform operations comprising: receiving value of a patient parameter associated with the subject; executing a sepsis mortality prediction model using the value to generate a prediction associated with sepsis mortality, wherein the sepsis mortality prediction model is generated by perfonning operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset;determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis; and outputting, by the sepsis mortality prediction model based on value, the prediction associated with the onset of sepsis mortality within the threshold time after presentation of suspected sepsis in the subject.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Tire detailed description is described with reference to the accompanying figures. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. The use of the same reference numbers in different figures indicates similar or identical components or features. The figures are merely exemplary to illustrate certain features that can be used singularly or in combination with other features, and the present disclosure should not be limited to the embodiments shown.

[0012] FIG. 1 illustrates a block diagram of an example system for generating a sepsis mortality prediction model.

[0013] FIG. 2 illustrates a flow diagram of an example process of generating the sepsis mortality prediction model.

[0014] FIG. 3 illustrates a flow diagram of an example process for performing the feature selection step while generating the sepsis mortality prediction model.

[0015] FIG. 4 illustrates a flow diagram of an example process of selecting the sepsis mortality prediction model from candidate models.

[0016] FIG. 5 illustrates a block diagram of an example system for using tire sepsis mortality prediction model.

[0017] FIG. 6 illustrates a flow diagram of an example process of using the sepsis mortality prediction model.

[0018] FIG. 7 illustrates a flow diagram of an example process of selecting a genetic marker subset to be used as an input for tire sepsis mortality prediction model.

[0019] FIGs. 8A-8I illustrate experimental design, demographics, and initial prognostic modeling associated with predicting sepsis mortality.

[0020] FIGs. 9A-9B illustrate topological data analysis (TDA) of a combined global sepsis cohort dataset.

[0021] FIGs. 10A-10C illustrate TDA group-specific genes and a clinical tool.

[0022] FIGs. 11A-1 IB illustrate gene-set enrichment comparison of TDA groups and 28-day sepsis mortality.

[0023] FIG. 12 illustrates a sepsis endotype model.DETAILED DESCRIPTION

[0024] Tire following detailed description is presented to enable any person skilled in the art to make and use the subject of the application. For purposes of explanation, specific nomenclature is set forth to provide a thorough understanding of the present disclosure. However, it will be apparent to one skilled in the art that these specific details are not required to practice the subject of the application. Descriptions of specific applications are provided only as representative examples. The present application is not intended to be limited to the embodiments shown, but is to be accorded the widest possible scope consistent with the principles and features disclosed herein.

[0025] Technical and scientific terms used herein have the meanings commonly understood by one of ordinary skill in the art to which the present disclosure pertains, unless otherwise defined.

[0026] As used herein, the singular forms “a,” “an,” and “the” designate both the singular and the plural, unless expressly stated to designate the singular only.

[0027] The present disclosure provides systems and methods for predicting mortality within a subject, individual, or patient within a threshold time of presentation with suspected sepsis. In embodiments, the threshold time can be 28 days of presentation with suspected sepsis.

[0028] As used herein, the term “presentation with suspected sepsis” can refer to detectable symptoms such as before there are perceivable, noticeable, or measurable signs of sepsis or severe disease resulting from sepsis in the individual(s).

[0029] As used herein, the term “host gene” refers to a gene belonging to an individual who is infected by a pathogen.

[0030] As used herein, the tenns “administer,” “administration,” or “administering” refer to (1) providing, giving, dosing and / or prescribing, such as by either a health professional or his or her authorized agent or under their direction, and (2) putting into, taking or consuming, such as by a health professional or the individual, and is not limited to any specific dosage forms or routes of administration, unless otherwise stated.

[0031] As used herein, the terms “ameliorating” or “preventing” progression of sepsis include alleviating or preventing tire development of one or more symptoms thereof, or impeding orpreventing an underlying mechanism of severe disease, and achieving any therapeutic and / or prophylactic benefit.

[0032] As used herein, the term “sepsis” refers to the potentially life-threatening physical reaction of tire host to an infection. Tire way sepsis is defined clinically continues to evolve, but recent definitions include the 2001 SCCM / ESICM / ACCP / ATS / SIS “Sepsis-2”, and tire 2016 SCCM / ESICM “Sepsis-3”. Both definitions, and any future updates to the clinical definitions or international standards for defining sepsis, apply here.

[0033] As used herein, tire term “severe disease” is defined as sepsis with any degree of end organ damage (e.g., kidney, respiratory, or liver failure). Sepsis patients who go on to develop severe disease will require significant medical intervention (e.g., admission to a hospital or intensive care unit, ventilation, renal replacement therapy) in order to avert permanent physical damage, long-term sequelae, and / or death.

[0034] As used herein the terms “marker(s)” and “biomarker(s)” are used interchangeably to refer to a measurable substance from a biological sample such as one or more nucleic acid markers (also referred to as a genetic marker or genetic biomarker). Tire tenn “host biomarker” further indicates that the measurable substance is derived from the infected individual, rather than the infecting pathogen.

[0035] In embodiments, the nucleic -acid biomarker(s) are at least one or more of: adrenoceptor B2 (ADRB2), CD 177 molecule (CD 177), carboxypeptidase vitellogenic like (CPVL), C-X3-C motif chemokine receptor 1 (CX3CR1), defensin a3 (DEFA3), Fc receptor like 5 (FCRL5), G protein subunit y2 (GNG2), interleukin 10 receptor subunit a (IL10RA), kinesin light chain 3 (KLC3), oleoyl -ACP hydrolase (OLAH), pyruvate kinase Ml / 2 (PKM), radical S-adenosyl methionine domain containing 2 (RSAD2), STE20 related adaptor B (STRADB), tyrosylprotein sulfotransferase 1 (TPST1), tetraspanin 5 (TSPAN5), tetratricopeptide repeat domain 9C (TTC9C), or zinc finger with KRAB and SCAN domains 1 (ZKSCAN1).

[0036] As used herein, the term “stratification” refers to the division of a group of individuals into subgroups, based on one or more shared characteristics, such as derived from the observable or measured biological parameters. For example, the division can be based on a characteristic already known relevant to tire outcome, such as age, sex, or having a pre-existing condition, or it can be based on clusters identified in observable or measured biological parameters using any of a variety of data cluster analysis techniques.

[0037] In embodiments, data can be stratified prior to feature selection. This data stratification can be achieved by using unsupervised or supervised machine learning models, including but not limited to topological data analysis, k-means clustering, hierarchical clustering, nearest neighborclustering, non-linear clustering (e.g., t-distributed stochastic neighbor embedding), consensus clustering, or spectral clustering.

[0038] As used herein, the term “clustering” refers to the grouping of individuals or samples based on one or more shared characteristics, such as derived from the observable or measured biological parameters. For example, these can comprise one or more host biomarkers, one or more clinical outcomes data, one or more administrative health data, or a combination thereof. Clustering is performed using dedicated mathematical algorithms, here primarily via topological data analysis or cluster analysis methods.

[0039] As used herein, the term “data quality control” refers to analytic approaches including visual and mathematical approaches to cleaning data, reformatting data, applying missing data algorithms, normalizing data, standardizing data, and / or reducing the dimensionality of data based on specific criteria.

[0040] In embodiments, data quality control involves at least one of differential expression algorithms, principal component analysis, k-nearest neighbor imputation algorithms, three-sigma rule algorithms, or empirical Bayes method algorithms. Differential expression algorithms determine the fold-change from a reference sample and the p-value for the statistical difference between the sample and the reference value, which are used as decision metrics for inclusion or exclusion. Principal component analysis identifies the key variables in a multidimensional data set that explain the differences in the observations (variance) and can be used to determine if groups separate according to a priori knowledge about the samples. Nearest neighbor imputation utilizes k-nearest neighbor algorithms to predict discrete and continues values for a potential missing value . Using three-sigma rule algorithms, biomarker data generated from multiplex assays can be subsetted by a variance metric, wherein a threshold for variance is set as an inclusion or exclusion criteria (e.g., only markers with a variance greater than three standard deviations will be included). Empirical Bayes method algorithms utilize tire estimated distributions from the data to establish prior distributions, and are used to approximate values in a data set and subset data based on the parameters of the estimated distribution.

[0041] As used herein, the term “topological data analysis” or “TDA” refers to the analysis of datasets using techniques from topology, a study of the properties of a geometric space which allows defining continuous deformation of subspaces. Extraction of information from datasets that are high-dimensional, incomplete, and noisy is generally challenging. In practice, TDA methods such as the “Mapper” algorithm, enable dimensionality reduction, visualization and clustering of complex data sets.

[0042] As used herein, the terms ‘'individual”, “subject”, “patient”, or “test individual” indicates a mammal, in particular a human or non-human primate. Tire test individual can or can not be in need of an assessment of sepsis and / or severe disease. In embodiments, the test individual is assessed prior to the detection of symptoms of sepsis. In embodiments, the test individual is assessed prior to the onset of any detectable symptoms of sepsis. In embodiments, the test individual does not have detectable symptoms of any type of sickness or condition. In embodiments, the test individual has an exposure, injury, wound, or condition that puts them at risk of developing sepsis, such as: having a viral or bacterial infection, such as but not limited to: urinary' tract infection, meningitis, endocarditis, or septic arthritis; undergoing a medical surgical or dental procedure; having an open wound or trauma, such as but not limited to: a blast injury, a crush injury, an extremity wound, a gunshot wound, or a wound received in combat; suffering a nosocomial infection; having undergone medical interventions such as central line placement or intubation; having diabetes; being HIV positive; undergoing hemodialysis; and / or undergoing an organ transplant procedure (donor or receiver). In embodiments, the individual does not have a condition that puts them at risk of severe disease from sepsis, prior to application of the methods described herein. In embodiments, the individual has a condition that puts them at risk of severe disease from sepsis.

[0043] As used herein, the term “clinical outcome” indicates a measurable status or change in the health, function or quality of life of an individual with sepsis or at risk of developing sepsis. Examples include, but arc not limited to: severity or duration of symptoms, need for organ support (e.g., ventilation, renal replacement therapy, or vasoactive medications), response to treatment, admission to a hospital or intensive care unit, length of stay in a hospital or intensive care unit, mortality, long-term morbidity (e.g., time to returning to activities or quality of life), incidence of long-term sequelae of infectious diseases (e.g., chronic kidney disease, cardiovascular disease, or chronic pulmonary' disease), and re-hospitalization. Clinical outcomes can be recorded as categorical data (e.g., “yes / no”, “prcscncc / abscncc”, an ordinal scale), continuous data (e.g., blood pressure), temporal data (e.g., duration of symptoms, days hospitalized), or time-to-event data (e.g., days to death, time to return to normal daily activities).

[0044] Examples of clinical outcome data include, but are not limited to any one or more of the following: severity or duration of symptoms, time to symptom onset or abatement, need for organ support, duration of organ support, response to treatment, admission to a hospital or intensive care unit, length of stay in a hospital or intensive care unit, mortality, time to death, duration of morbidity (e.g., time to returning to normal daily activities or quality of life), incidence of longterm sequelae of infectious diseases, and re-hospitalization.

[0045] As used herein, the term "mortality" indicates death of the individual due to sepsis or due to severe disease caused by sepsis.

[0046] As used herein, the term ‘‘training data” can include one or more biological effectors and / or one or more non-biological effectors.

[0047] As used herein, the term “biological effector” is used to mean a molecule, such as, but not limited to a protein, a peptide, a carbohydrate, a complex lipid, a fatty acid, an amino acid, a biogenic amine, a nucleic acid, a glycoprotein, or a proteoglycan, that can be assayed. Specific examples of biological effectors can include cytokines, growth factors, antibodies, hormones, cell surface receptors, cell surface proteins, lipid mediators, or carbohydrates. More specific examples of biological effectors include, but are not limited to the genes described herein. Not all “biological effector(s)” and / or “non-biological effector(s)” can be applicable to predicting mortality due to sepsis or due to severe disease due to sepsis. The relevant “biological effectors” and / or “non- biological effector(s)” applicable to predicting mortality due to sepsis or due to severe disease with sepsis are identified during feature selection step(s) as described throughout this disclosure.

[0048] In embodiments, the biological effectors are soluble. In embodiments, the biological effectors are membrane-bound, such as a cell surface receptor. In embodiments, the biological effectors are intracellular. In embodiments the biological effectors are nucleic acids (e.g., messenger RNA, transfer RNA, micro RNA, long-noncoding RNA, silencing RNA, short hairpin RNA, or DNA). In embodiments, the biological effectors are detectable in a fluid sample of an individual, such as serum, and / or plasma. In embodiments, the biological effectors are measurable in a biological sample of an individual, such as blood plasma, wound effluent, or sputum.

[0049] As used herein, the term non-biological effector is a clinical parameter that is generally considered not to be a specific molecule. Although not a specific molecule, a non-biological effector can nonetheless still be quantifiable, cither through routine measurements or through measurements that stratify the data being assessed. For example, heart rate, change in heart rate over time, respiratory rate, body temperature, blood pressure, body mass index, and other parameters would be a non-biological effector component of the risk profile. All these components are measurable or quantifiable using routine methods and equipment. Other non-biological components include data that can not be readily or routinely quantifiable or that can require a practitioner’s judgment or opinion. For example, peripheral vascular disease, pulmonary’ hypertension, heart failure can be a quantifiable aspect of the risk profile. While there can be published guidance on classifying and diagnosing these aspects of the risk profile, assigning a numerical value to the severity, still involves observation and, to a certain extent, judgment or opinion. In some instances, the quantity or measurement assigned to anon-biological effector couldbe binary, e.g., “0” if absent or “1” if present. In other instances, the non-biological effector aspect of the risk profile can involve qualitative components that cannot or should not be quantified.

[0050] At least some of the training can be assayed, detected, measured, and / or determined from a sample taken or isolated from an individual. “Biological sample,” “sample” and “test sample” are used interchangeably herein. In embodiments, the sample can be body fluid taken from the individual.

[0051] Examples of the test sample or sources of training data include, but are not limited to: biological fluids and / or tissues isolated from an individual or patient, which can be tested by the methods of the present application described herein, and include but are not limited to: whole blood, peripheral blood, capillary blood, serum, plasma, cerebrospinal fluid, wound effluent, urine, amniotic fluid, peritoneal fluid, pleural fluid, lymph fluids, various external secretions of the respiratory, intestinal, and genitourinary tracts, various components of exhaled breath, tears, sweat, saliva, white blood cells, tissue biopsies, and combinations thereof.

[0052] As described herein, the term “treatment” refers to providing medical care to alleviate sepsis symptoms or cure sepsis or suspected sepsis. Examples of treatment can include initiation or broadening of antibiotic therapy, balancing fluids and electrolytes, renal replacement therapy, adjustment of mechanical ventilation, targeted or empiric anti-inflammatory or immunomodulatory drugs, hemodynamic adjustments, calcium channel blocker medications, or surgical intervention. Benefits of such early treatment can include: reduced severity or duration of sepsis, reduced need for organ support (e.g. , ventilation, renal replacement therapy, or vasoactive medications), reduced length of stay in a hospital or intensive care unit, reduced risk of mortality, reduced long-term morbidity (e.g., time to returning to activities or quality of life), decreased incidence of long-term sequelae of infectious diseases (e.g., chronic kidney disease, cardiovascular disease, or chronic pulmonary disease), decreased re-hospitalization rates, and / or reduced medical costs. In embodiments, adjusting current treatment comprises changing dose of current antibiotic, changing to a different antibiotic, changing dose of non-steroidal anti- inflammatory drugs, or initiating or adjusting insulin therapy.

[0053] As described herein, the term “endotype(s)” refer to biologically defined subclasses of clinical sy ptoms that differentiate a heterogeneous group of subjects based on differing molecular expression profiles. Once differentiated, the underlying pathobiology can be more directly targeted and treated. Examples of different endotypes of sepsis include immunosuppressed, immunometabolic, acute inflammation, and immunocompetent.

[0054] As described herein, the term “immunosuppressed” refers to a reduction of the activation of efficacy of the immune system of an individual or subject.

[0055] As described herein, the term ‘'immunometabolic” refers to being related to a metabolic pathway that regulates the functions of an immune system of an individual or subject.

[0056] As described herein, the temi “acute inflammation” refers to an immediate, rapid, adaptive, and / or temporary response by an immune system of an individual or subject to injury, illness, or infection.

[0057] As described herein, the term “immunocompetent” refers to a normally operating immune system of an individual or subject.

[0058] FIG. 1 illustrates a block diagram of an example system 100 for generating a sepsis mortality prediction model. Hie example system 100 includes biological sample collection device(s) 102, a biomarker detection device 104, first computing device(s) 106, and second computing device(s) 122. The first computing device(s) 106 includes an input component 108, first processor(s) 110, first memory 112, a training data component 114, a clinical outcome component 116, a prediction model component 118, and an output component 120. The second computing device(s) 122 includes second processor(s) 124, second memory 126, and a prediction model component 128. In embodiments, a sepsis mortality prediction model can be generated at the first computing device(s) 106 and transferred to the second computing device(s) 122. Alternatively, the sepsis mortality prediction model can be generated at the second computing device 122.

[0059] The biological sample collection device(s) 102 can be one or more devices that are positioned at or proximate to a plurality of subjects to collect biological sample(s) from the subjects, the subjects being different than an individual or a subject that the generated sepsis mortality prediction model is used for. In embodiments, the biological sample(s) can include one or more of whole blood, peripheral blood, capillary blood, serum, plasma, cerebrospinal fluid, wound effluent, urine, amniotic fluid, peritoneal fluid, pleural fluid, lymph fluids, various external secretions of the respiratory, intestinal, and genitourinary tracts, various components of exhaled breath, tears, sweat, saliva, white blood cells, and / or tissue biopsies. In embodiments, each body fluid collection device of the device(s) 102 can include a needle and a fluid storage container physically coupled to the needle.

[0060] In embodiments, the biological sample(s) collected by the biological sample collection device(s) 102 can be transferred to biomarker detection device 104. The biomarker detection device 104 can detect, identify, and quantify one or more biomarker(s) from the biological sample. Hie biomarker(s) can be detected, identified, and quantified (in other words, determined) by performing an assay such as an immunoassay or a gene expression assay on the collected biological sample.

[0061] In embodiments, the detecting, identifying, and quantifying the biomarker(s) can include detecting a presence of one or more biomarkers from the biological sample, identifying theone or more biomarkers, and / or determining a value associated with and / or a concentration of each detected and identified biomarker.

[0062] In embodiments, as each biomarkcr is detected, identified, and quantified by the biomarker detection device 104, biomarker data associated with the detected, identified, and quantified biomarker (e.g., biomarker name, presence, and / or concentration) can be transmitted or transferred to the first computing device(s) 106 and stored at the training data component 114. In embodiments, the biomarker detection device 104 can include a communication component that can transmit tire biomarker data to the first computing device(s) 106 using a wired connection or a wireless connection. The first computing dcvicc(s) 106 can receive the biomarkcr data using its communication component. Alternatively, the biomarker data can be transmitted or transferred to the first computing device(s) 106 after all biomarker parameter(s) are determined. In embodiments, the biomarker data received by the first computing device(s) 106 from the biomarker detection device 104 can be stored in the training data component 114 as a feature to use to generate the sepsis mortality prediction model. Alternatively, the biomarker detection device can instead transmit raw data to the first computing device(s) 106, where tire processor! s) 110 uses the raw data to determine the biomarker data.

[0063] In embodiments, the first computing device(s) 106 is where the sepsis mortality prediction model is being generated. As illustrated, the first computing device(s) 106 generates the sepsis mortality prediction model, and transmits or transfers the generated sepsis mortality prediction model to the second computing dcvicc(s) 122 where the sepsis mortality prediction model is used to generate a prediction metric associated with sepsis mortality. Alternatively, the first computing device(s) 106 can generate and utilize the sepsis mortality prediction model. The sepsis mortality prediction model can be generated by a single first computing device 106 or multiple first computing devices 106 (e.g., by a server cluster).

[0064] In addition to the biomarker data, the first computing device(s) 106 can receive additional data that can be stored at the training data component 114 of tire memory 112. In embodiments, the additional data can include administrative health data. Examples of administrative health data include, but are not limited to any one or more of: baseline demographics (e.g., age, sex, ethnicity), physiological parameters (e.g., body mass index, heart rate, respiratory rate, body temperature), comorbid conditions including but not limited to immunocompromising conditions (e.g., history' of chronic kidney disease, history of hepatic disease, pulmonary' hypertension, dementia, having diabetes, being HIV positive, tobacco use, alcohol use, drug use, or pregnancy), past surgical history (e.g.. central line placement, organ transplant donor orrecipient), and environmental or social exposures (e.g., living situation, travel history, contact with livestock.

[0065] In embodiments, the clinical outcome component is configured to store clinical outcome(s) associated with the training data. In embodiments, the stored clinical outcome(s) can indicate whether one or more features (e.g., the biomarker(s), the administrative health data, and / or the like), leads to sepsis mortality. Examples of clinical outcomes include, but are not limited to any one or more of the following: severity or duration of symptoms, time to symptom onset or abatement, need for organ support, duration of organ support, response to treatment, admission to a hospital or intensive care unit, length of stay in a hospital or intensive care unit, mortality, time to death, duration of morbidity (e.g., time to returning to normal daily activities or quality of life), incidence of long-term sequelae of infectious diseases, and re-hospitalization. In embodiments, the clinical outcome(s) can further include a likelihood (e.g., a percentage) associated with sepsis mortality in association with one or more features of the training data. The likelihood associated with sepsis mortality can be determined from, for example, the mortality, the time to death, the duration of morbidity, and / or the like.

[0066] In embodiments, the input component 108 can include a mouse, a keyboard, a touch screen, and / or the like. In embodiments, a user can utilize the input component to input clinical outcome(s) and / or the administrative health data.

[0067] In embodiments, the training data can be stored at the training data component 114 in a first data structure (e.g., a database) and the clinical outcome(s) and the clustering can be stored at the clinical outcome component 116 in a second data structure. Alternatively, the training data and the clustering can be stored in the memory 112 in a single data structure. In embodiments, the processor(s) 110 can be used to generate the sepsis mortality prediction model at the prediction model generating component using the training data. Generating the sepsis mortality prediction model is described in additional detail at FIG. 2-4, as well as throughout this disclosure. In embodiments, the generated sepsis mortality prediction model can be stored at the prediction model component 118 or at another location within the memory 112. The sepsis mortality prediction model being stored within the memory 112 can be transferred or transmitted to the prediction model component 128 within the memory 126 of the second computing device(s) 122. The processor(s) 124 can use the sepsis mortality prediction model to determine a sepsis mortality prediction metric of a subject or individual. Using the prediction model to determine the sepsis mortality prediction metric is discussed in additional detail in FIGS. 5 and 6, as well as throughout this disclosure.

[0068] FIG. 2 illustrates a flow diagram of an example process 200 of generating a sepsis mortality prediction model. In embodiments, some or all of process 200 can be performed by oneor more components described in association with FIG. 1, as described herein. Additionally, some portions of process 200 can be omitted, replaced, and / or reordered. For example, the example process 200 can be performed by the processor(s) 110 of the first computing device(s) 106, and the generated sepsis mortality prediction model can correspond to the generated wound prediction model detailed in association with FIG. 1.

[0069] At operation 202, the process can include receiving training data and clinical outcome(s). In embodiments, the training data can correspond to data stored within the training data component 114 and data stored within the clinical outcome component 116 described in association with FIG. 1, as well as throughout this disclosure. In embodiments, the training data can include the biomarker data, the administrative health data 208, and the clinical outcome(s) 206. In embodiments, the biomarker data can include nucleic acid marker(s) 204. The administrative health data 208 can correspond with the administrative health data described in association with FIG. 1, as well as throughout this disclosure. The clinical outcome(s) 206 can correspond with the clinical outcome(s) described in association with FIG. 1, as well as throughout this disclosure.

[0070] In embodiments, the nucleic acid marker(s) 204 can include, but are not limited to one or more of: adhesion G protein-coupled receptor El (ADGRE1), adrenoceptor (32 (ADRB2), angiotensin II receptor associated protein (AGTRAP). AKT serine / threonine kinase 1 (AKT1), 5'- aminolevulinate synthase 2 (ALAS2), alkaline phosphatase, biomineralization associated (ALPL), ankyrin repeat domain 22 (ANKRD22), annexin A3 (ANXA3), arginase 1 (ARG1), BCL2 like 1 (BCL2L1), BMX non-rcccptor tyrosine kinase (BMX), chromosome 6 open reading frame 62 (C6orf62), carbonic anhydrase 2 (CA2), C-C motif chemokine ligand 5 (CCL5), C-C motif chemokine receptor 3 (CCR3), CD4 molecule (CD4), CD24 molecule (CD24), CD 177 molecule (CD 177), CD274 molecule (CD274), cell division cycle 34, ubiquitin conjugating enzyme (CDC34), complement factor D (CFD), chitinase 3 like 1 (CHI3L1), carbohydrate sulfotransferase 2 (CHST2), C-type lectin domain family 4 member E (CLEC4E), cytidinc / uridinc monophosphate kinase 2 (CMPK2), cytochrome C oxidase assembly factor 1 homolog (COA1), carnitine palmitoyltransferase 1A (CPT1A), carboxypeptidase vitellogenic like (CPVL), chondroitin sulfate N-acetylgalactosaminyltransferase 1 (CSGALNACT1). cystatin C (CST3), C-X3-C motif chemokine receptor 1 (CX3CR1), DNA damage inducible transcript 4 (DDIT4), defensin a3 (DEFA3), defensin a4 (DEFA4), DNA J heat shock protein family (Hsp40) member C 1 (DNAJC 1), DNA damage regulated autophagy modulator 1 (DRAM1), deoxyuridine triphosphatase (DUT), dual specificity tyrosine phosphorylation regulated kinase 3 (DYRK3), erythrocyte membrane protein band 4.2 (EPB42), family with sequence similarity 174 member C (FAM174C), F-box and WD repeat domain containing 2 (FBXW2). Fc receptor like 5 (FCRL5). ferrochelatase (FECFI),fibroblast growth factor binding protein 2 (FGFBP2), Fms related receptor tyrosine kinase 3 (FLT3), formyl peptide receptor 1 (FPR1), GATA binding protein 1 (GATA1), GTPase, IMAP family member 4 (GIMAP4), GTPase. IMAP family member 7 (GIMAP7), GTPase, IMAP family member 8 (GIMAP8), G protein subunit y2 (GNG2), granulysin (GNLY), G protein-coupled receptor 65 (GPR65), growth factor receptor bound protein 10 (GRB10), glutathione S-transferase K! (GSTK1), H3 histone pseudogene 6 (H3F3AP4). hemoglobin subunit a2 (HBA2), hemogen (HEMGN), HECT and RLD domain containing E3 ubiquitin protein ligase family member 6 (HERC6), H3.2 histone [putative] (HIST2H3PS2), major histocompatibility complex, class I, B (HLA-B), major histocompatibility complex, class II, DQ Bl (HLA-DQB1), high mobility group box 2 (HMGB2), 15 -hydroxyprostaglandin dehydrogenase (HPGD), hydrogen voltage gated channel 1 (HVCN1), isoamyl acetate hydrolyzing esterase 1 [putative] (IAH1), intercellular adhesion molecule 1 (ICAM1), immediate early response 5 (IER5). interferon a inducible protein 6 (IFI6), interferon a inducible protein 27 (IFI27), interferon induced protein 44 (IFI44), interferon induced protein with tetratricopeptide repeats 1 (IFIT1), interferon induced protein with tetratricopeptide repeats 2 (IFIT2), interleukin ip (IL1B), interleukin 1 receptor type 1 (IL1RA), interleukin 1 receptor type 2 (IL1R2), interleukin 10 receptor subunit a (IL 1 ORA), interaction protein for cytohesin exchange factors 1 (IPCEF1), interferon regulatory factor 2 binding protein 2 (IRF2BP2), ISG15 ubiquitin like modifier (ISG15), JUN proto-oncogene, AP-1 transcription factor subunit (JUN), potassium voltage-gated channel subfamily E regulatory subunit 1 (KCNE1), kinesin light chain 3 (KLC3), kelch like family member 24 (KLHL24), kringle containing transmembrane protein 1 (KREMEN1), long intergenic non-protein coding RNA 861 (LINC00861), lymphocyte antigen 6 family member E (LY6E), MAPK associated protein 1 (MAPKAP1), mediator complex subunit 28 (MED28), MicroRNA 6724-4 (MIR6724-4), matrix metallopeptidase 8 (MMP8), multimerin 1 (MMRN1). myeloperoxidase (MPO). mannose receptor C type 2 (MRC2), mitochondrially encoded 12S rRNA (MT-RNR1), MX dynamin like GTPase 2 (MX2), nuclear factor, erythroid 2 like 3 (NFE2L3), 2'-5'-oligoadenylate synthetase 3 (OAS3), oleoyl-ACP hydrolase (OLAH), olfactomedin 4 (OLFM4), peptidase inhibitor 3 (PI3), phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit (PIK3CB), PITH domain containing 1 (PITHD1). pyruvate kinase Ml / 2 (PKM). perilipin 2 (PLIN2), DNA polymerase 5 interacting protein 3 (POLDIP3), RAL GTPase activating protein catalytic subunit a2 (RALGAPA2), RAN binding protein 9 (RANBP9), REST corepressor 1 (RCOR1), Rh associated glycoprotein (RHAG), RNA, U1 small nuclear 2 (RNU1-2), RNA, U1 small nuclear 4 (RNU1-4), ribosomal protein L37a (RPL37A), ribosomal protein L38 (RPL38), ribosomal protein Si l (RPS11), ribosomal protein S18 (RPS18), radical S-adenosyl methionine domain containing 2(RSAD2), SI 00 calcium binding protein A 8 (S100A8), SI 00 calcium binding protein A9 (S100A9), SI 00 calcium binding protein A 12 (S100A12), SAM domain, SH3 domain and nuclear localization signals 1 (SAMSN1), Sin3A associated protein 30 (SAP30), strawberry notch homolog 1 (SBNO1), selenium binding protein 1 (SELENBP1), sialic acid binding Ig like lectin 10 (SIGLEC10), solute carrier family 25 member 6 (SLC25A6), solute carrier family 25 member 39 (SLC25A39), solute carrier family 39 member 8 (SLC39A8), solute carrier family 4 member 1 [Diego Blood Group] (SLC4A1), synuclein a (SNCA), small nucleolar RNA, H / ACA box 44 (SNORA44), superoxide dismutase 2 (SOD2), spectrin a, erythrocytic 1 (SPTA1), STE20 related adaptor p (STRADB), syntaxin 6 (STX6), switching B cell complex subunit SWAP70 (SWAP70), spectrin repeat containing nuclear envelope protein 2 (SYNE2), T-box transcription factor 21 (TBX21), TRAF interacting protein with forkhead associated domain (TIFA), toll like receptor 7 (TLR7). transmembrane and coiled-coil domain family 2 (TMCC2), transmembrane protein 35B (TMEM35B), transmembrane protein 273 (TMEM273), thymosin 10 (TMSB10), TNF a induced protein 6 (TNFAIP6), tyrosylprotein sulfotransferase 1 (TPST1), tripartite motif containing 4 (TRIM4), tetraspanin 5 (TSPAN5), tetratricopeptide repeat domain 9C (TTC9C), ubiquitin protein ligase E3 component N-recognin 5 (UBR5), UNC-93 homolog Bl, TLR signaling regulator (UNC93B1), WASH complex subunit 2C (WASHC2C), XIAP associated factor 1 (XAF1), tyrosine 3-monooxygenase / tryptophan 5 -monooxygenase activation protein e (YWHAH), and zinc finger With KRAB and SCAN domains 1 (ZKSCAN1).

[0071] In embodiments, the administrative health data 208 can further include, but are not limited to one or more of: baseline demographics, physiologic parameters, comorbid conditions including but not limited to immunocompromising conditions, past surgical history, and environmental or social exposures.

[0072] At operation 210, the process can include subsetting the training data. In embodiments, the operation 214 can include performing data quality control such as data subsetting algorithms 212A-212N (referred to generically as data subsetting 212). In embodiments, the data subsetting 212 can vary depending on the characteristics of each feature. For example, some of the biomarker data (e.g., those that involve host biomarkers) can be determining (e.g., measured) using multiplex assays that generate data on thousands of markers. Subsetting such data can be performed by conducting a differential expression analysis wherein the fold-change from a reference sample and the p-value for the statistical difference between the sample and the reference value are used as decision metrics for inclusion or exclusion. In other examples, other biomarker data (e g., those not related to host biomarkers) generated from multiplex assays can be subsetted by a variance metric, wherein a threshold for variance is set as an inclusion or exclusion criteria (e.g., only markers witha variance greater than three standard deviations will be included). While these methods of data quality control are discussed, many more are contemplated.

[0073] In embodiments, the operation 210 can further include determining a training subset and a test subset from the training data. In embodiments, the training subset can be used for generating the sepsis mortality prediction model. For example, the operation 210 can include selecting a portion (e.g., 80%) of the training data as the training subset, and leaving the rest (e.g., the 20%) as the test subset.

[0074] At operation 214, the process can include performing topological data analysis (also referred to as TDA) on the training subset. In embodiments, the topological data analysis uses unsupervised approaches (such as by using TDA algorithms 216A-216N) to represent such data in a structured, two-dimensional network that retains the geometric ‘shape" (topology) of the data correlations. Individuals or samples with a high degree of similarity, for example of host gene, form groups of highly interconnected nodes which represent distinct subgroups / populations within the dataset. In embodiments, perfonning tire TDA includes grouping individuals with comparable gene expression profiles in an unbiased manner and identifying trends and phenotypes within the sepsis cohort. Groups of individuals with similar gene expression profiles are defined within tire TDA structure based on node density and connectivity.

[0075] In embodiments, additionally or alternatively, the operation 214 can include performing cluster analysis on the training subset. Performing the cluster analysis can include using supervised or unsupervised approaches to discretize highly complex data based on similarities in the test subset. Both the cluster analysis and the topological data analysis result in the assignment of individuals or samples to a discrete set of groups / clusters based on multiple shared characteristics.

[0076] In embodiments, the operation 214 can also be performed on the test subset in addition to the training subset.

[0077] At operation 220, the process can include performing feature selection on the training subset that has been processed through TDA to determine a model subset. Additional details with respect to the operation 220 are described in association with FIG. 3, as well as throughout this disclosure.

[0078] At operation 222, the process can include training candidate prediction models. In embodiments, the operation 222 can include inputting the model subset into each candidate prediction model to train each candidate prediction model. Additional details with respect to the operation 222 are described in association with FIG. 4, as well as throughout this disclosure.

[0079] At operation 224, the process can include selecting prediction model from the trained candidate prediction models. In embodiments, the operation 226 can include using the test subset to determine a performance metric for each candidate prediction model and using the performance metrics to determine one of the candidate prediction models as the prediction model. Additional details with respect to the operation 226 are described in association with FIG. 4, as well as throughout this disclosure.

[0080] FIG. 3 illustrates a flow diagram of an example process 300 of performing feature selection of the training subset described in association with FIG. 2. In embodiments, some or all of process 300 can be perfonned by one or more components described in association with FIG. 1, as described herein. Additionally, some portions of process 300 can be omitted, replaced, and / or reordered. For example, the example process 300 can be performed by the processor(s) 110 of the first computing device(s) 106, and the generated sepsis mortality prediction model can correspond to the generated wound prediction model detailed in association with FIG. 1. In embodiments, the example process 300 can correspond to the operation 222 of FIG. 2.

[0081] At operation 302, the process can include receiving training data. In embodiments, the training data can include some or all of the data stored in the training data component 114. The training data can include features such as the biomarker data (e.g., the nucleic acid marker(s)) and the administrative health data described in association with FIG. 2.

[0082] At operation 304, the process can include perfonning a data quality check on the training subset. In embodiments, the operation 304 can include determining a completeness of the training data. Determining the completeness can include determining whether the training subset contains missing value(s) for one or more features, and if so, the operation 304 can include filling in the missing value(s) to ensure completeness of the training subset.

[0083] At operation 306, the process can include determining a relevancy metric for each feature in the training subset. In embodiments, the relevancy metric can be a relevancy score. In embodiments, the relevancy score can be indicative of how closely a feature from the training subset correlate with predicting sepsis mortality within 28 days of presentation with suspected sepsis. In embodiments, the operation 306 can further include using the clinical outcome(s) to determine an association between the feature and sepsis mortality within 28 days of presentation with suspected sepsis. In embodiments, the feature that is determined to be closest in correlation with predicting sepsis mortality within 28 days of presentation with suspected sepsis can have the highest relevancy score. If the operation 306 is being perfonned for the first time, then the process skips to performing operation 312.

[0084] At operation 308, the process can include determining a redundancy metric for each feature in the training subset. In embodiments, the redundancy metric can be a redundancy score. In embodiments, the operation 308 can be performed after the operation 306 has been performed for at least a second time (e.g., a second time, a third time, a fourth time, and so on). In embodiments, the operation 308 can include determining the relevancy score for a feature in the training subset based on determining a distance between the feature and each of the selected features that were selected in operation 314. In embodiments, the feature with the highest redundancy score is indicative of having the strongest correlation with one or more selected features.

[0085] At operation 310, the process can include determining a difference between the relevancy metric and the redundancy metric of each feature in the training subset. In embodiments, the operation 310 can include detennining a difference between the relevancy score and the redundancy score of each feature. Alternatively, the operation 310 can include determining a quotient between the relevancy score and the redundancy score.

[0086] At operation 312. the process can include ranking features based on the differences determined at the operation 310. In embodiments, if the operation 312 is being performed for the first time, then the features can be ranked based on their relevancy scores.

[0087] At operation 314, the process can include selecting the feature with the highest rank. In embodiments, the operation 314 can further include removing tire selected feature from the training subset before proceeding to operation 316.

[0088] At operation 316, the process can include determining whether a target threshold of features has been selected. In embodiments, the target threshold can be 8 features, 10, features, or 13 features. If no, then the process can return to the operation 310 where the operation 310 can be performed on the training subset that now excludes the selected fcaturc(s). If yes, then the process can proceed to operation 318 where a model subset is generated that includes only the selected features. In embodiments, the selected features of the model subset can include ADRB2, CPVL, CX3CR1, DEFA3, FCRL5, IL10RA, STRADB, TPST1, CD177, and RSAD2.

[0089] FIG. 4 illustrates a flow diagram of an example process 400 of generating a sepsis mortality prediction model. In embodiments, some or all of process 400 can be performed by one or more components described in association with FIG. 1, as described herein. Additionally, some portions of process 400 can be omitted, replaced, and / or reordered. For example, the example process 400 can be performed by the processor(s) 110 of the first computing device(s) 106, and the generated sepsis mortality prediction model can correspond to the generated sepsis mortalityprediction model detailed in association with FIG. 1. In embodiments, the example process 400 can correspond to the operations 224 and 226 of FIG. 2.

[0090] At operation 402, the process can include receiving a model subset. In embodiments, the model subset can correspond to the model subset detennined at the operation 318 of FIG. 3.

[0091] At operation 404. the process can include training first candidate prediction model using the model subset. In embodiments, the first candidate prediction model can be configured to predict a likelihood of sepsis mortality using a random forest algorithm. In embodiments, the operation 404 can include inputting tire model subset into the first candidate prediction model to generate a first prediction, determining a first difference between the first prediction and an expected result, the expected result being based on one or more clinical outcomes, and adjusting, based on the first difference, one or more parameters of the first candidate prediction model to generate updated first candidate prediction model such that subsequent updated first prediction(s) approaches the expected result. In embodiments, training the first candidate prediction model can include iteratively updating the first candidate model using the model subset until the updated first prediction approaches a threshold similarity with the expected result or until a threshold iteration has been reached.

[0092] At operation 406, the process can include a training second candidate prediction model using the model subset. In embodiments, the second candidate prediction model can be configured to predict a likelihood of sepsis mortality using a logical regression algorithm. In embodiments, the operation 406 can include inputting the model subset into the second candidate prediction model to generate a second prediction, determining a second difference between the second prediction and the expected result, and adjusting, based on the second difference, one or more parameters of the second candidate prediction model to generate updated second candidate prediction model such that subsequent updated second prediction(s) approaches the expected result. In embodiments, training the second candidate prediction model can include iteratively updating the second candidate model using the model subset until the updated second prediction approaches a threshold similarity with the expected result or until a threshold iteration has been reached.

[0093] At operation 408, the process can include training a third candidate prediction model using the model subset. In embodiments, the third candidate prediction model can be configured to predict a likelihood of sepsis mortality using a neural network machine algorithm. In embodiments, the operation 404 can include inputting the model subset into the third candidate prediction model to generate a third prediction, determining a third difference between the third prediction and the expected result, and adjusting, based on the third difference, one or more parameters of the third candidate prediction model to generate updated third candidate prediction model such thatsubsequent updated third prediction(s) approaches the expected result. In embodiments, training the third candidate prediction model can include iteratively updating the third candidate model using the model subset until the updated third prediction approaches a threshold similarity with the expected result or until a threshold iteration has been reached.

[0094] At operation 410, the process can include training n-th candidate prediction model using the model subset. In embodiments, training the n-th candidate prediction model can include training a fourth candidate prediction model, a fifth candidate prediction model, a sixth candidate prediction model, up to an n-th candidate prediction model is trained. For example, the n-th candidate prediction model can be the sixth candidate prediction model, and tire operation 410 can include training the fourth candidate prediction model, the fifth candidate prediction model, and the sixth candidate prediction model. In embodiments, the fourth candidate prediction model can be configured to predict a likelihood of sepsis mortality using a Bayesian optimally classifier, the fifth candidate prediction model can be configured to predict a likelihood of sepsis mortality using classification and regression tree, and the sixth candidate prediction model can be configured to predict a likelihood of sepsis mortality using support vector machines. Candidate prediction models can be configured to predict a likelihood of sepsis mortality using one or more of cluster analysis, unsupervised machine learning algorithm, supervised machine learning algorithm, minimum redundancy maximum relevance, Student’s t-test, Mann-Whitney U test, bootstrap aggregating, boosting, Bayesian model combination, a bucket of models, or stacking.

[0095] In embodiments, the operation 410 can include inputting the model subset into the third candidate prediction model and up to the n-th candidate prediction model to generate a third prediction and up to an n-th prediction, determining a difference between each of the prediction(s) up to n-th prediction and the expected result, and adjusting one or more parameters of each of the prediction model(s) up to the n-th candidate prediction model to generate updated candidate prediction model(s) such that subsequent updated prediction(s) approaches the expected result. In embodiments, training each of the prediction modcl(s) up to n-th candidate prediction model can include iteratively updating the candidate model(s) using tire model subset until the updated prediction(s) approaches a threshold similarity with the expected result or until a threshold iteration has been reached.

[0096] In the example where the n-th candidate prediction model is the sixth candidate prediction model, the operation 410 can include determining a fourth prediction, a fifth prediction, and a sixth prediction, determining a fourth difference between the fourth prediction and the expected result, determining a fifth difference between the fifth prediction and the expected result, determining a six difference between the sixth prediction and the expected result, adjusting, basedon the fourth difference, one or more parameters of the fourth candidate prediction model to generate updated fourth candidate prediction model, such that subsequent updated fourth updated prediction(s) approaches the expected result, adjusting, based on the fifth difference, one or more parameters of the fifth candidate prediction model to generate updated fifth candidate prediction model, such that subsequent updated fifth updated prediction(s) approaches the expected result, and adjusting, based on the sixth difference, one or more parameters of the sixth candidate prediction model to generate updated sixth candidate prediction model, such that subsequent updated sixth updated prediction(s) approaches the expected result.

[0097] At operation 412, the process can include determining first performance mctric(s) associated with the first candidate prediction model. In embodiments, the operation 412 can include perfonning internal cross-validation of the first candidate prediction model with the test subset and / or external cross-validation of the first candidate prediction model with an independent dataset unrelated to the training data.

[0098] In embodiments, perfonning the internal cross-validation of the first candidate prediction model can include determining, using the test subset, a first area under the receiver operating characteristic curve (AUROC) associated with the first candidate prediction model, In embodiments, performing the external cross-validation of the first candidate prediction model can include determining, using the independent dataset, a second AUROC associated with the first candidate prediction model to include also as the first performance metrics.

[0099] At operation 414, the process can include detennining second perfonnance metric(s) associated with the second candidate prediction model. In embodiments, the operation 414 can include perfonning internal cross-validation of the second candidate prediction model with the test subset and external cross-validation of the second candidate prediction model with the independent dataset.

[0100] In embodiments, performing the internal cross-validation of the second candidate prediction model can include determining, using the test subset, a first AUROC associated with the second candidate prediction model as second performance metrics. In embodiments, performing the external cross-validation of the second candidate prediction model can include determining, using the independent dataset, second AUROC associated with the second candidate prediction model to include also as the second performance metrics.

[0101] At operation 416, the process can include determining a third perfonnance metric(s) associated with the third candidate prediction model. In embodiments, tire operation 414 can include performing internal cross-validation of the third candidate prediction model with the testsubset and external cross-validation of the third candidate prediction model with the independent dataset.

[0102] In embodiments, perfonning the internal cross-validation of the third candidate prediction model can include detennining, using the test subset, a first AUROC associated with the third candidate prediction model as the third performance metrics. In embodiments, perfonning the external cross-validation of the third candidate prediction model can include determining, using the independent dataset, a second AUROC associated with the third candidate prediction model to include also as the third performance metrics.

[0103] At operation 418, the process can include detennining n-th second perfonnance metric(s) associated with the n-th candidate prediction model. In embodiments, the operation 414 can include performing internal cross-validation of the n-th candidate prediction model with the test subset and external cross-validation of the n-th candidate prediction model with the independent dataset.

[0104] In embodiments, perfonning the internal cross-validation of the n-th candidate prediction model can include determining, using the test subset, a first AUROC associated with the n-th candidate prediction model as n-th performance metrics. In embodiments, performing the external cross-validation of the n-th candidate prediction model can include determining, using the independent dataset, a second AUROC associated with the n-th candidate prediction model to include also as the n-th performance metrics.

[0105] In tire example where the n-th candidate prediction model is the sixth candidate prediction model, the operation 418 can include performing internal cross-validation and external cross-validation of the fourth candidate prediction model to determine fourth performance metrics, performing internal and external cross-validation of the fifth candidate prediction model to determine fifth performance metrics, and performing internal and external cross-validation of the sixth candidate prediction model to determine sixth performance metrics. The fourth, fifth, and sixth perfonnance metrics can also include AUROCs associated with the fourth candidate prediction model, AUROCs associated with the fifth candidate prediction model, and AUROCs associated with the sixth candidate prediction model.

[0106] At operation 420, the process can include selecting one of the candidate prediction modcl(s) as the sepsis mortality prediction model. In embodiments, the operation 422 can include selecting the candidate prediction model of the candidate prediction models with the highest AUROC value. In embodiments, the operation 420 can select, based on the second perfonnance metric being the highest, the second candidate prediction model that uses the logical regression algorithm as the sepsis mortality prediction model.

[0107] FIG. 5 illustrates a block diagram of an example system 500 for using a sepsis mortality prediction model to determine tire likelihood of sepsis mortality in a subject. In embodiments, the sepsis mortality prediction model can correspond to the sepsis mortality prediction model generated as described in association with FIGS. 1-4. The example system 500 includes a patient sample collection device 502, and a biomarker detection device 504, and a computing device 506, In embodiments, the patient sample collection device 502 can correspond with the biological sample collection device(s) 102, and the biomarker detection device 504 can correspond with the biomarker detection device 104, and tire computing device 506 can correspond with the computing device 122 of FIG. 1. The computing device 506 includes an input component 508, proccssor(s), memory 512, which further includes a patient parameter component 514, a prediction model component 516. and an output component 518.

[0108] In embodiments, the patient sample collection device 502 can collect sample(s), such as blood, from a patient. The patient sample collection device 502 can be positioned at or proximate to the subject to collect the sample(s) from the patient. In embodiments, the patient sample collection device 502 can include a needle and a fluid storage container physically coupled to the needle.

[0109] In embodiments, the sample(s) collected by the patient sample collection device 502 can be transferred to the biomarker detection device 504. The biomarker detection device 504 can detect, identify, and quantify one or more biomarker(s) from the body fluid. The biomarker(s) can be detected, identified, and quantified (in other words, determined) by performing an immunoassay(s) and / or performing mass spectrometry (e.g., liquid-chromatography mass spectrometry) on the collected body fluid sample.

[0110] In embodiments, the detecting, identifying, and quantifying the biomarker(s) can include detecting a presence of one or more biomarkers from the sample(s), identifying the one or more biomarkers, and / or determining a concentration of each detected and identified biomarker.

[0111] In embodiments, as each biomarker is detected, identified, and quantified by the biomarker detection device 504, biomarker data associated with the detected, identified, and quantified biomarker (e.g., biomarker name and concentration) can be transmitted or transferred to the computing device 506 and stored at the patient parameter component 514. In embodiments, the biomarkcr detection device 504 can include a communication component that can transmit the biomarker data to the computing device 506 using a wired connection or a wireless connection. The computing device 506 can receive the biomarker data using its communication component. Alternatively, the biomarker data can be transmitted or transferred to the computing device 506 after all biomarker parameter(s) are determined. In embodiments, the biomarker detection device504 can further determine that the biomarker data includes only biomarkers that correlate with the model subset described in association with FIGS. 1-4 before transmitting the biomarker data to the computing device 506. Alternatively, the processor(s) 510 can remove biomarkers that are do not correlate with the model subset from the biomarker data prior to storing the biomarker data in the patient parameter component 514.

[0112] In embodiments, the input component 508 can include a mouse, a keyboard, a touch screen, and / or the like. In embodiments, a user can utilize the input component to supplement data within the model parameter component or input information about the subject such as identifying infomiation about the subject (c.g., name, age, gender, etc.).

[0113] In embodiments, the processor(s) 510 can execute the sepsis mortality prediction model stored at the prediction model component 516 based on tire data from the patient parameter component 514 to predict the likelihood of sepsis mortality within 28 days of presentation with suspected sepsis. Predicting the likelihood of sepsis mortality is described in association with FIG. 6, as well as throughout this disclosure.

[0114] In embodiments, the output component 518 can be a display used to output a report associated with a prediction score being above a threshold score. The report can include a warning that, without treatment, sepsis mortality is likely to occur within 28 days of presentation with suspected sepsis and, as the likelihood of sepsis mortality, a sepsis mortality score. The report can also include a treatment recommendation to lower the likelihood of sepsis mortality. The recommendation can include, based on detennining a presence and / or dosage of one or more target antibiotics in the patient sample, initiating treatment using the target antibiotics or increasing the dose of the antibiotics, based on determining a level of fluids and a level of electrolytes in the patient sample, balancing the level of the fluids and the level of the electrolytes, based on determining the health of one or more kidney of the patient, renal replacement therapy, based on determining a presence and / or a dosage of empiric anti-inflammatory or immunomodulatory' drug(s) in the patient sample, initiating and / or increasing the dosage of the empiric antiinflammatory or immunomodulatory drug(s), determining, based on a presence and / or dosage of calcium blockers in the patient sample, initiating and / or increasing the dosage of calcium blockers, adjustment of mechanical ventilation, and / or surgical intervention. The report can further be transmitted to a user device of tire clinician and / or the surgeon (e.g., to a mobile phone, a tablet, a personal computer, and / or the like) using the communication component of the computing device 506.

[0115] FIG. 6 illustrates a flow diagram of an example process 600 of using a sepsis mortality prediction model to predict the likelihood of sepsis mortality within 28 days of presentation withsuspected sepsis. In embodiments, the sepsis mortality prediction model can be the sepsis mortality prediction model described in association with FIGS. 1-5. In addition to what is discussed below in association with FIG. 6, using the sepsis mortality prediction model is further discussed in association with the Examples section, as well as throughout this disclosure.

[0116] At operation 602. the process can include receiving model parameter(s). In embodiments, the model parameter(s) can be biomarker data from a patient sample that correlates with the features within the model subset described in association with FIG. 3. In embodiments, the model parameter(s) can be received by the computing device 506 at the patient parameter component 514. Alternatively, the model paramctcr(s) can be determined by the proccssor(s) 510 from biomarker data received from the biomarker detection device 504 prior to being stored in the patient parameter component 514. Additional details with respect to the operation 602 are described in association with FIG. 5, as well as throughout this disclosure.

[0117] At operation 604, the process can include determining a sepsis mortality prediction score using the sepsis mortality prediction model based on the model parameter(s). In embodiments the score can be a weighted prediction. In embodiments, expression levels of CD 177 and RSAD2 in the patient sample are compared to fixed reference values. Based on distances of the CD 177 and RSAD2 to centroids, a set of four probability scores are determined, one for each endotype. In embodiments, the probability scores add up to 100%. In embodiments, the prediction model uses logistic regression to predict sepsis mortality for each endotype: ADRB2 and DEFA3 for endotype 1; CX3CR1 and FCRL5 for endotype 2; CPVL, TPST1, and STRADB for endotype 3; and IL10RA for endotype 4. In embodiments, the probability scores are then used to weight each of the endotype-specific predictions. In embodiments, the prediction model adds the four weighted predictions to generate a prediction score. Additional details with respect to the operation 608 are described in association with FIG. 5, as well as throughout this disclosure.

[0118] At operation 606, the process can include determining whether the likelihood is greater than or equal to a threshold likelihood. In embodiments, the operation 606 can include determining whether the prediction score exceeds a threshold score. If yes, then at the operation 608, the process can include outputting a report indicating that sepsis mortality is likely to occur within 28 days of presentation with suspected sepsis. The report can indicate a warning that sepsis mortality is likely to occur within 28 days of presentation with suspected sepsis, a probability associated with the sepsis mortality, and a treatment recommendation. Additional details with respect to the operation 606 are described in association with FIG. 5, as well as throughout this disclosure. If no, then the process can return to the operation 602, where the process continues to gather additional patient sample(s). determine updated model parameter(s) from the patient sample(s), and use updatedmodel parameter(s) to predict whether sepsis mortality is likely to occur within 28 days of presentation with suspected sepsis.

[0119] FIG. 7 illustrates a flow diagram of an example process 700 of determining a genetic marker subset that is predictive of sepsis mortality. In embodiments, some or all of process 700 can be performed by one or more components described in association with FIG. 1. as described herein. Additionally, some portions of process 700 can be omitted, replaced, and / or reordered. For example, the example process 700 can be performed by the processor(s) 110 of the first computing device(s) 106, and tire generated sepsis mortality prediction model can correspond to the generated sepsis mortality prediction model detailed in association with FIG. 1. In embodiments, the example process 400 can be a portion of the operations 224 and 226 of FIG. 2 along with the processes described in association with FIGS. 3 and 4.

[0120] At operation 702, the process can include receiving genetic marker data. In embodiments, the genetic marker data can include a plurality of genes. In embodiments, the genetic marker data can include a plurality of protein-coding, non-sex-linked genes. In embodiments, the genetic marker data can alternatively include differentially expressed genes. In embodiments, the contents of the genetic marker data can be genes collected from individuals with sepsis. In embodiments, the genetic marker data can include the mortality of each individual at the threshold time (e g., whether the individual survived after 28 days of sepsis). In embodiments, samples can be collected from each individual using sample collection devices that correspond to the biological sample collection dcvicc(s) 102. In embodiments, the samples can be assayed using a biomarkcr detection device corresponding to the biomarker detection device 104. In embodiments, the gene marker data can be determined from tire assayed sample by the biomarker detection device or the processor(s) 110.

[0121] At operation 704, the process can include executing features selection on the genetic marker data. In embodiments, the feature selection can correspond to the feature selection described in association with FIG. 3 or another maximum relevancy minimum redundancy algorithm. For each genetic marker subsets 706-712, different thresholds can be set for executing tire feature selection. For example, for operation 706 to determine a first genetic marker subset, the feature selection can be used to select 8 features, and for operation 708 to determine a second genetic marker subset, the feature selection can be used to select 13 features. In embodiments, each feature is associated with a different gene in the genetic marker data.

[0122] In embodiments, the genetic marker data can first be stratified using TDA prior to the feature selection. In embodiments, the TDA can be performed on a subset of genes from the genetic marker data that is less than the total of the genetic marker data. For example, the genetic markerdata can include 3061 genes and out of the 3061 genes, 1000 genes can be the most variable genes. In this example, TDA would be performed on the 1000 genes. In embodiments, the TDA can cluster the subset of genes into a plurality of gene expression endotypes. For example, the TDA can cluster the subset of genes into four clusters that represent varying levels of 28-day sepsis mortality risk between 11% and 22%. In embodiments, the feature selection can be performed within each of the four clusters.

[0123] At operation 706, the process can include determining the first genetic marker subset. In embodiments, the operation 706 can include executing the feature selection on the genetic marker data as is without stratifying the genetic marker data using TDA or another clustering or cluster analysis algorithm. In embodiments, the operation 706 can further include executing the feature selection to select a first plurality of features from the genetic marker data as tire first genetic marker subset. For example, the first genetic marker subset can be GNG2, CST3, CX3CR1, FCRL5, FLT3, CPT1A, CCR3, and PLIN2.

[0124] At operation 708, the process can include determining tire second genetic marker subset. In embodiments, tire operation 708 can include stratifying the genetic marker data using TDA or another clustering or cluster analysis algorithm to cluster the genetic marker data into the plurality of gene expression endotypes described in association with the operation 704. In embodiments, the feature selection can be executed forthe genes at each gene expression endotype. For example, the TDA can be performed to cluster the 1000 most variable genes into four gene expression cndotypcs that represent varying levels of 28-day sepsis mortality risk between 11% and 22%. In embodiments, the feature selection can select a second plurality of features as the second genetic marker subset with a different number of features than the first genetic marker subset. For example, the second genetic marker subset can be ADRB2, CPVL, CX3CR1, DEFA3, FCRL5, GNG2, IL10RA, KLC3, PKM, STRADB, TPST1, TTC9C and ZKSCAN1.

[0125] At operation 710, the process can include determining a third genetic marker subset. In embodiments, the operation 710 can include selecting a third plurality of features from the second genetic marker subset such that the third genetic marker subset can be used in a weighted sum prediction model that determines a prediction score for sepsis mortality based on a weighted sum of the plurality of endotypes, such that the weights is based on comparing a plurality of genes that are not in the third genetic marker subset to reference values associated with those plurality of gene. For example, the feature selection can result in selecting ADRB2 and DEFA3 for endotype 1; CX3CR1 and FCRL5 for endotype 2; CPVL, TPST1 and STRADB for endotype 3; and IL10RA for endotype 4 where the CD 177 and RSAD2 is used to detennine the likelihood of the patient belonging to each of the gene expression endotype. Expression levels of CD 177 and RSAD2 in thepatient sample are compared to fixed reference values for these two genes, and based on their distances to the centroids, a set of four probability scores is determined (one for each endotype, adding up to 100%). Logistic regression is used to predict 28-day mortality for each of the four endotypes. The probability scores are used to weight each of tire endotype -specific predictions, and the four weighted predictions are summed to determine the prediction score.

[0126] At operation 712, the process can include determining a fourth up to an N-th genetic marker subset. For example, the operation 712 can include determining the fourth genetic marker subset. In such an example, the process can include determining a fourth plurality of features from the genetic marker data that is different the first, second and third genetic marker subsets. For example, the fourth genetic marker subset can be CPVL. CX3CR1, DEFA3, FCRL5, GNG2, ILIORA, PKM. RSAD2, STRADB, TPST1. TSPAN5, TTC9C, ZKSCAN1, RAB21, SCYL2, XRCC5. The fourth genetic marker subset can also be used to determine a prediction score using a weighted sum associated with CD 177 and RSAD2 as described in association with the operation 710.

[0127] At operations 714-720. the process can include determining a first, second, third, fourth, and up to an n-th performance metric to validate a performance for each of the genetic marker subsets. In embodiments, the performance metrics can be determined using k-fold cross- validation using the prediction model selected as described in association with FIG. 4. For example, the k-fold cross-validation can be executed using 10 folds and 10 repeats, and each performance metric can be an average AUROC over all of the runs of the k-fold cross-validation.

[0128] At operation 722. tire process can include selecting the genetic marker subset to be used for predicting sepsis mortality. In embodiments, the operation 722 can include ranking each of the performance metrics and selecting the genetic marker subset with the highest performance metric. For example, if the fourth performance metric is the highest, then the fourth genetic marker subset would be selected. In embodiments, if rankings include genetic marker subsets that uses the weighted sum prediction model at tire top (e.g., the third or the fourth being ranked tire two highest), then further validation can be performed (e.g. using k-fold cross-validation) for those genetic marker subsets using the weighted sum prediction model to determine which genetic marker subset has the highest performance metric and be selected as the genetic marker subset. In embodiments, the machine learning algorithm that is used for the weighted sum model can be the same machine learning algorithm used in tire selected prediction model from FIG. 4. In embodiments, the selected genetic marker subset and / or the weighted prediction model can be used on a subject or individual with or at risk for sepsis in steps that correspond with those described in association with FIG. 6. In embodiments, the prediction score determined by the weighted prediction model can becompared to a threshold score such that when the prediction score is above the threshold score, then the subject or individual is predicted to not survive past a threshold time (e.g., 28 days).

[0129] While one or more examples of the techniques described herein have been described, various alterations, additions, permutations, and equivalents thereof are included within the scope of the techniques described herein.

[0130] In the description of examples, reference is made to the accompanying drawings that form a part hereof, which show by way of illustration specific examples of the claimed subject matter. It is to be understood that other examples can be used and that changes or alterations, such as structural changes, can be made. Such examples, changes or alterations are not necessarily departures from the scope with respect to the intended claimed subject matter. While the steps herein can be presented in a certain order, in some cases the ordering can be changed so that certain inputs are provided at different times or in a different order without changing the function of the systems and methods described. The disclosed procedures could also be executed in different orders. Additionally, various computations that are herein need not be performed in the order disclosed, and other examples using alternative orderings of the computations could be readily implemented. In addition to being reordered, the computations could also be decomposed into subcomputations with the same results.

[0131] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as example fonns of implementing the claims.

[0132] The components described herein represent instructions that can be stored in any type of computer-readable medium (also referred to as a computer-readable storage medium or computer-readable storage medium) and can be implemented in software and / or hardware. All of the methods and processes described above can be embodied in, and fully automated via, software code modules and / or computer-executable instructions executed by one or more computers or processors, hardware, or some combination thereof. Some or all of the methods can alternatively be embodied in specialized computer hardware. A computer-readable medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer-readable storage medium can be any tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. In embodiments, the memories 112, 126, and 512 of FIGS. 1 and 5 respectively can be computer-readable mediums.

[0133] Program code embodied on a computer-readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing. Computer program code for carrying out operations for aspects of the present disclosure can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, Smalltalk, C++, or the like, and conventional procedural programming languages, such as the C programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer, and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0134] Conditional language such as, among others, “can,” “could,” “can” or “might,” unless specifically stated otherwise, arc understood within the context to present that certain examples include, while other examples do not include, certain features, elements and / or steps. Thus, such conditional language is not generally intended to imply that certain features, elements, and / or steps are in any way required for one or more examples or that one or more examples necessarily include logic for deciding, with or without user input or prompting, whether certain features, elements and / or steps are included or are to be performed in any particular example.

[0135] Conjunctive language such as the phrase “at least one of X, Y or Z,” unless specifically stated otherwise, is to be understood to present that an item. term. etc. can be either X, Y. or Z, or any combination thereof, including multiples of each element. Unless explicitly described as singular, “a” means singular and plural.

[0136] Any routine descriptions, elements, or blocks in the flow diagrams described herein and / or depicted in tire attached figures should be understood as potentially representing modules, segments, or portions of code that include one or more computer-executable instructions for implementing specific logical functions or elements in the routine. Alternate implementations are included within the scope of the examples described herein in which elements or functions can bedeleted, or executed out of order from that shown or discussed, including substantially synchronously, in reverse order, with additional operations, or omitting operations, depending on the functionality involved as would be understood by those skilled in the art.

[0137] Many variations and modifications can be made to tire above-described examples, the elements of which are to be understood as being among other acceptable examples. All such modifications and variations are intended to be included herein within the scope of this disclosure and protected by the following claims.

[0138] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0139] These computer program instructions can also be stored in a computer-readable medium that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the computer-readable medium produce an article of manufacture including instructions which implement the function / act specified in the flowchart and / or block diagram block or blocks. The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus, or other devices to produce a computer-implemented process such that tire instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0140] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical fimction(s). It should also be noted that, in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession can.in fact, be executed substantially concurrently, or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustration, and combinations of blocks in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware-based systems that perfonn the specified functions or acts, or combinations of special purpose hardware and computer instructions.

[0141] Although the figures show a specific order of method steps, the order of the steps can differ from what is depicted. Also, two or more steps can be performed concurrently or with partial concurrence. Such variation will depend on the software and hardware systems chosen and on the designer's choice. All such variations are within the scope of the disclosure. Likewise, software implementations could be accomplished with standard programming techniques with rule-based logic and other logic to accomplish the various connection steps, processing steps, comparison steps, and decision steps.

[0142] As will be understood by one of ordinary skill in tire art, each embodiment disclosed herein can comprise, consist essentially of, or consist of its particular stated element, step, ingredient, or component. Thus, the terms “include” or “including” should be interpreted to recite: “comprise, consist of. or consist essentially of.” The transition term “comprise” or “comprises” means includes, but is not limited to, and allows for the inclusion of unspecified elements, steps, ingredients, or components, even in major amounts. The transitional phrase “consisting of’ excludes any element, step, ingredient, or component not specified. Tire transition phrase “consisting essentially of’ limits the scope of the embodiment to the specified elements, steps, ingredients, or components and to those that do not materially affect the embodiment.EXEMPLARY EMBODIMENTS

[0143] 1: A method of generating a sepsis mortality prediction model for a subject comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; executing a feature selection algorithm on the training data, wherein executing the feature selection algorithm comprises: determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; and generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of tire training data less than a total amount of tire training data; training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of theplurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis.

[0144] 2: The method of embodiment 1, further comprising: determining whether one or more values associated with the training data is missing; and filling in, based on determining one or more missing values associated with the training data, the one or more missing values.

[0145] 3: The method of embodiments 1 or 2, wherein the biomarker data comprises one or more nucleic acid markers.

[0146] 4: The method of any one of embodiments 1-3, wherein the model subset comprises a level of: adrenoceptor [32 (ADRB2); carboxypeptidase vitellogenic like (CPVL); C- X3-C motif chemokine receptor 1 (CX3CR1); defensin a3 (DEFA3); Fc receptor like 5 (FCRL5); interleukin 10 receptor subunit a (IL10RA); STE20 related adaptor [3 (STRADB); tyrosylprotein sulfotransferase 1 (TPST1); CD177 molecule; and / or radical S-adenosyl methionine domain containing 2 (RSAD2).

[0147] 5: The method of any one of embodiments 1-4, further comprising: generating a second data structure comprising genetic marker data; determining, based on executing the feature selection algorithm on the genetic marker data, a first genetic marker subset; determining, based on executing a topological data analysis algorithm on the genetic marker data, a plurality’ of clusters associated with sepsis mortality; determining, based on executing the feature selection algorithm on each of the plurality of clusters, a second genetic marker subset, the second genetic marker subset being different from the first genetic marker subset; determining, based on executing the sepsis mortality prediction model on the first genetic marker subset, a first genetic marker performance metric; determining, based on executing the sepsis mortality prediction model on the second genetic marker subset, a second genetic marker performance metric; and selecting, based on the higher of the first genetic marker performance metric and second genetic marker perfonnance metric, the first genetic marker subset or the second genetic marker subset as an input for the sepsis mortality prediction model.

[0148] 6: The method of any one of embodiments 1-5, wherein determining the relevancy metric comprises: detennining, based on a correlation between a biomarkcr feature of the plurality of biomarker features and a likelihood of sepsis mortality, a relevancy score associated with the biomarker feature.

[0149] 7: The method of embodiment 6, wherein generating the model subset comprises: determining, for each feature and based on a distance between a selected feature from tire trainingdata and each biomarker feature, a redundancy metric associated with each biomarker feature; ranking, based on a difference between the redundancy metric and relevancy metric, the biomarker features associated with a training subset; selecting a biomarker feature associated with a largest difference; generating updated training subset by eliminating the biomarker feature associated with the largest difference from a training subset; and generating the model subset, the model subset comprising the biomarker feature associated with the largest difference.

[0150] 8: The method of any one of embodiments 1-7, wherein determining the performance metric comprises determining, by inputting a test subset into the trained candidate prediction model, a magnitude of an area under a receiver operator curve (AUROC) associated with the trained candidate prediction model.

[0151] 9: The method of embodiment 8, wherein selecting the sepsis mortality prediction model comprises selecting a candidate prediction model of the plurality of candidate prediction models with a highest magnitude of AUROC.

[0152] 10: Tire method of any one of embodiments 1-9, wherein training the plurality of candidate prediction models comprises: detennining, for each candidate prediction model of the plurality of candidate prediction models and based on inputting the model subset into the candidate model, a prediction associated with the candidate prediction model; determining a difference between the prediction and the clinical outcomes; and adjusting, based on the difference, a parameter of the candidate prediction model.

[0153] 11: Hie method of any one of embodiments 1-10, wherein the biomarker data being associated with biological samples from a plurality of individuals different from the subject.

[0154] 12: A method of predicting sepsis mortality of a subject comprising: receiving value of a patient parameter associated with the subject; executing a sepsis mortality prediction model using the value to generate a prediction associated with sepsis mortality, wherein the sepsis mortality prediction model is generated by performing operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; executing a feature selection algorithm on the training data, wherein executing the feature selection algorithm comprises: determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarkcr feature in the biomarkcr data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of tire training data; training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of the plurality of trained candidate predictionmodels: and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis; and outputting, by the sepsis mortality prediction model based on value, the prediction associated with an onset of sepsis mortality within the threshold time after presentation of suspected sepsis in the subject.

[0155] 13: Tire method of embodiment 12, wherein the prediction associated with sepsis mortality comprises a predictions score associated with sepsis mortality within a threshold time after presentation of suspected sepsis.

[0156] 14: The method of embodiments 12 or 13, wherein the patient parameter comprises biomarker data and the value comprises a level of a biomarker feature in the biomarker data.

[0157] 15: The method of embodiment 14, wherein the value comprises a level of: adrenoceptor (32 (ADRB2); carboxypeptidase vitellogenic like (CPVL); C-X3-C motif chemokine receptor 1 (CX3CR1): defensin a3 (DEFA3); Fc receptor like 5 (FCRL5); interleukin 10 receptor subunit a (IL10RA); STE20 related adaptor [3 (STRADB); tyrosylprotein sulfotransferase 1 (TPST1); CD 177 molecule; and / or radical S-adenosyl methionine domain containing 2 (RSAD2).

[0158] 16: Tire method of any one of embodiments 12-15, further comprising: receiving a second value associated with biomarker data, the biomarker data being determined from a sample associated with the subject; and wherein the prediction associated with the sepsis mortality is further detennined based on the second value.

[0159] 17: The method of any one of embodiments 12-16, wherein generating the sepsis mortality prediction model further comprises: generating a second data structure comprising genetic marker data; determining, based on executing the feature selection algorithm on the genetic marker data, a first genetic marker subset; determining, based on executing a topological data analysis algorithm on the genetic marker data, a plurality of clusters associated with sepsis mortality; determining, based on executing the feature selection algorithm on each of the plurality of clusters, a second genetic marker subset, the second genetic marker subset being different from the first genetic marker subset; determining, based on executing the sepsis mortality prediction model on the first genetic marker subset, a first genetic marker performance metric; determining, based on executing the sepsis mortality prediction model on the second genetic marker subset, a second genetic marker performance metric; and selecting, based on the higher of the first genetic marker performance metric and second genetic marker performance metric, the first genetic marker subset or the second genetic marker subset as an input for the sepsis mortality prediction model,wherein the value of the patient parameter is a value associated with the selected marker genetic subset.

[0160] 18: The method of embodiment 14, wherein outputting the prediction associated with sepsis mortality comprises: detennining, based on the biomarker data, a presence or a dosage of an antibiotic; determining, based on tire presence or the dosage of the antibiotic, a dosage of the antibiotic to administer to the subject; and outputting a recommendation associated with administering the dosage of the antibiotic.

[0161] 19: A system for generating a sepsis mortality prediction model for a subject comprising: one or more processors; and one or more non-transitory computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perfonn operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data: training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from tire plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model detennines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis.

[0162] 20: A system for predicting sepsis mortality of a subject comprising: one or more processors; an output component; and one or more computer-readable media storing computerexecutable instructions that, when executed, cause the one or more processors to perfonn operations comprising: receiving value of a patient parameter associated with the subject; executing a sepsis mortality prediction model using the value to generate a prediction associated with sepsis mortality, wherein the sepsis mortality prediction model is generated by performing operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data: training aplurality of candidate prediction models using the model subset: determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model detennines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis; and outputting, by the sepsis mortality prediction model based on value, the prediction associated with an onset of sepsis mortality within the threshold time after presentation of suspected sepsis in the subject.

[0163] While the cxamplary embodiments described above arc described with respect to one particular implementation, it should be understood that, in the context of this document, the content of the example clauses can also be implemented via a method, device, system, computer-readable medium, and / or another implementation. Additionally, any of the exemplary embodiments 1-20 above and / or the examples 1-11 below can be implemented alone or in combination with any other one or more of the exemplary embodiments 1-20 above and / or the examples 1-11.EXAMPLES

[0164] Example 1: Materials and Methods.

[0165] Tire prognostic gene sets / algorithms identified as part of the ACESO observational sepsis study present a significant improvement in the prediction of severe outcomes in patients with sepsis (specifically, mortality within 28 days of presentation with suspected sepsis) over the current gold-standard clinical tool (qSOFA score). Application of these gene sets / algorithms will enable rapid stratification of high / low-risk patients with sepsis when they are first encountered and diagnosed in-clinic. Thus, it will enable better triage decisions, and potentially impact treatment decisions. We have already demonstrated the feasibility of implementing one of these gene sets using a fieldable point-of-care device which uses existing technology, as detailed below. Further validation and optimization of this tool and its associated prognostic algorithm can be done if desired.

[0166] In embodiments, the disclosure provides three sets of host genes, the expression of which can be used to determine mortality within 28 days of presentation with suspected sepsis. It should be stressed that each individual gene set can be used for this purpose; they are not intended to be used in conjunction.

[0167] The first gene set is agnostic of potential sepsis endotypes and is intended to be used ‘'as is” without any further information required to obtain the prediction. The second gene set requires the assignment of a patient to one of four separately defined gene expression endotypes.One method for generating these assignments is described for the third gene set, and requires the measurement of two additional genes, but other assignment methods are possible. The third gene set represents a fieldable sequencing assay that works in conjunction with a weighted, predictive algorithm for sepsis, which is based on measuring two specific genes for sepsis endotype assignment. Uris final gene set and its associated algorithm have been translated to a point-of-care device with automated analysis software.

[0168] All three gene sets were developed using blood gene expression data (RNAseq) obtained from the ACESO observational sepsis study (n=494), which was performed in Cambodia, Ghana, and the USA. Briefly, 3061 protein-coding, non-scx-linkcd genes were subject to widely used data QC and normalization methods. Differentially expressed genes between 28-day survivors and non-survivors were then used for feature selection, either with or without stratification. Feature selection was performed using mRMR (minimum redundancy maximum relevance), followed by logistic regression.

[0169] For the first gene set, no stratification of the subjects was performed. Feature selection yielded a set of 8 genes, the expression of which can predict 28-day mortality with an AUROC of 0.81. The following genes are included: GNG2, CST3, CX3CR1, FCRL5, FLT3, CPT1A, CCR3 and PLIN2.

[0170] To identify the second and third gene sets, the 1000 most variable genes from the 3061 initial preprocessed gene list were first stratified using topological data analysis (TDA). This ‘‘clustering” method yielded four gene expression “sepsis endotypes'’ with vary ing levels of 28-day mortality risk (ranging from 11% to 22%). Feature selection was then performed within each of these groups, using the methods already outlined. This yielded a set of 13 genes which included: ADRB2, CPVL, CX3CR1, DEFA3, FCRL5, GNG2, IL10RA, KLC3, PKM, STRADB, TPST1, TTC9C and ZKSCAN1. These 13 genes are only intended to be used in conjunction with assignment of a patient to one of four specific host-gene expression endotypes, with a subset of the gene list associated with mortality prediction for each endotype.

[0171] The third gene set and its associated algorithm were developed for implementation on a fieldable point-of-care assay. It represents a modified version of the second gene set, which was reduced to 8 genes: ADRB2, CPVL, CX3CR1, DEFA3, FCRL5, IL10RA, STRADB, and TPST1. To these were added 2 further genes that are used to determine a patient’s likelihood of belonging to one of the four sepsis gene expression endotypes: CD177 and RSAD2. The combined set of 10 genes is again intended to be used “as is” without any further information required to obtain the prediction. However, it does this using the Predict d28 algorithm, which is tied to this specific gene list. The Predict_d28 algorithm first uses CD 177 and RSAD2 to determine the likelihood of thepatient belonging to each of the four sepsis gene expression endotypes. Expression levels of CD 177 and RSAD2 in the patient sample are compared to fixed reference values for these two genes, and based on their distances to the centroids, a set of four probability scores is determined (one for each endotype, adding up to 100%). In the second step, tire algorithm uses logistic regression to predict 28-day mortality for each of the four endotypes, based on the specific sets of predictive genes: ADRB2 and DEFA3 for endotype 1: CX3CR1 and FCRL5 for endotype 2: CPVL. TPST1 and STRADB for endotype 3; and IL 1 ORA for endotype 4. The probability scores from step 1 are then used to ‘ weight” each of the endotype-specific predictions. In the final step, the algorithm adds up the four weighted predictions to give a final prediction score - if this score exceeds a set threshold, then the algorithm predicts the patient will die within 28 days of taking the measurement.

[0172] The third gene set and its associated algorithm were translated onto a fieldable, point- of-care device based on existing technology (Oxford Nanopore MinlON device). The assay was then tested on a set of 60 new samples obtained from the ACESO sepsis observational study in Cambodia, Ghana, and the United States. Further validation of the sepsis cndotypcs and refinement of tire models is possible.

[0173] The ability to predict outcomes in sepsis can be a major asset for clinical decision making, in particular in resource-constrained settings. The current gold standard is the qSOFA score,50which was developed to '‘identify high-risk patients for in-hospital mortality with suspected infection outside the ICU”. We determined the performance of the qSOFA score in the ACESO sepsis observational cohort, in order to compare it with the performance of the first and second gene lists outlined above. In all cases, the gene lists that are described herein significantly outperformed the qSOFA score on the same set of patients (see Example below).

[0174] Example 2: Study sites and subjects.

[0175] Five hundred and six patients across three sites had specimens available for this study (Table 1). Study protocols were approved by the Naval Medical Research Center (NMRC) Institutional Review Board (IRB) (Cambodia sepsis study # NMRC.2013.0019; Ghana sepsis study # NMRC.2016.0004-GHA; Duke sepsis study (Duke # PR000054849) in compliance with all applicable Federal regulations governing the protection of human subjects as well as host country IRBs. Tire study protocol in Cambodia was approved by the Cambodian National Ethics Committee for Health Research (NECHR). Tire protocol in Ghana was approved by the Committee on Human Research, Publication, and Ethics (CHRPE) at Kwame Nkrumah University of Science & Technology. All procedures were in accordance with the ethical standards of the Helsinki Declaration of the World Medical Association. All patients, or their legally authorized representatives, provided written informed consent.Table 1N = Count or number of patientsSD = Standard DeviationIQR = Inner Quartile Range (25% - 75% of data)28-day mortality = Number of patients that died within 28 days of presentation with suspected sepsis* Wilcoxon rank sum test results. All other results are from Welch Two Sample t-test

[0176] Example 3: Demographic comparisons.

[0177] To compare gender, size, age, and mortality in the study, we considered all the patients with available 28-day mortality information. For continuous variables (age) were compared using Welch's two-samplc t-test, while discrete variables (mortality) used the Wilcoxon rank sum test. To show the distribution of individual data, medians, and interquartile ranges, the age was plotted and colored in relation to gender, mortality, and site using functionalities of ggplot, and ggridges packages in R statistical environment.

[0178] Example 4: RNA-seq library preparation from patient peripheral blood.

[0179] RNA for sequencing was prepared as previously described. Briefly, the peripheral blood RNA was collected in Pagane RNA tubes (PreAnalytiX), and total RNA was purified using PAX gene Blood miRNA KIT (Qiagen) and depleted of human rRNA and globin using Globin- Zero Gold rRNA Removal Kit (Illumina). The vendor (Aetna) prepared the paired-end RNA sequencing according to standard procedures and sequenced generating ~ 50 million, 150-bp long paired-end reads per sample.

[0180] Example 5: Genome alignment and sequencing data pre-processing.

[0181] Sequencing data were aligned to the human genome (GRCh38). Briefly, all samples underwent quality control using Fats. Passing samples were aligned to the genome using Hista2, and transcripts were assembled using Strangite. Low-expression features (counts per million < 10), sex-linked features (located on chromosomes X, and Y), and features not mapping to known genes were removed to decrease noise and avoid gender bias in feature selection and modeling. To normalize the data between different study sites, the raw read counts were normalized using Median Ratio Normalization (Moar) method. Normalized data were then transformed using Variance Stabilizing Transformation (VST). The processing yielded 3061 genes for 506 patients from across the three study sites.

[0182] Example 6: Differential gene analysis ‘within’ and ‘between’ TDA groups

[0183] The normalized expression table of 3061 genes for 505 patients was fdtered to only contain 494 patients with known mortality outcomes. A total of fourteen (14) comparisons wereperformed "between ’ TDA groups to account for even' variation. Any patients that could be attributed to the TDA group under scrutiny were removed leaving only patients exclusive to individual TDA groups. " Within ’ TDA groups, patients were compared by 28-day mortality, contrasting those that died within 28 days of presentation with suspected sepsis and those that survived beyond this point. The significance of the mean difference was evaluated using a Welch two-sample t-test and the final value was adjusted for multiple comparisons using Benjamini & Hochberg41 model. Data were sorted by adjusted or p-value along with log2 fold change (L2FC) to highlight the most significantly changed genes.

[0184] Example 7: Gene set enrichment analyses (GSEA) ‘within’ and ‘between’ TDA groups.

[0185] Gene set enrichment analysis was performed using the results of differential gene expression analyses "within’ and "between’ TDA groups. For each contrast, the log2 fold changes (L2FC) were calculated using group means and ranked in decreasing order. For the "between’ TDA comparison, only genes with significant p.adjust value (<0.05) were considered for the analysis. All genes were considered for the "within’ TDA comparison due to the low count of statistically significant values. For GSEA analysis of the entire cohort comparing day 28 mortality, we used genes with significant p.value (<0.05). The selected genes were compared against Molecular Signature Database (MSigDB) gene set representing 'Hallmark’ pathways accessed using the msigdbr package and analyzed using the clusterProfiler package. For the frill cohort, we set " pvalueCutoff parameter to 0.05 to detect significant pathways only, whereas for other analyses, we set this parameter to 1 and marked the statistically significantly enriched pathways in the figures with boxes for a clearer comparison.

[0186] Example 8: Gene expression predicts mortality in global sepsis cohorts.

[0187] Total RNA from peripheral blood collected at enrollment (July 2016 - October 2017) was sequenced from 494 subjects in observational studies of sepsis in Ghana, Cambodia, and Durham, North Carolina (Fig. 8A). Male and female subjects were similar in age and 28-day mortality in the combined cohort (Fig. 8B). The overall 28-day mortality was 17%, with individual cohorts having a mortality of 8% (Duke), 13% (Cambodia), and 32% (Ghana) (Fig. 8C and D). Inspection of the data following preprocessing revealed no site bias (Fig. 8E). Differentially expressed genes (DEGs) between 28-day survivors and non-survivors include transcripts previously linked to sepsis outcomes such as IL1R2, OLAH, and CX3CR1 (Fig. 8F). Gene set enrichment analysis (GSEA) was used to identify biological pathways enriched in non-survivors. Hypoxia was the top enriched term while interferon response was significantly reduced. Feature selection using the Minimum Redundancy Maximum Relevance (MRMR) algorithm wasperformed using 391 differentially expressed genes to develop a prognostic classifier for 28-day mortality. The use of MRMR resulted in a list of the top 50 ranked genes in decreasing relation to the risk of sepsis-related mortality. Average performance versus the number of features used in logistic regression was determined and resulted in eight transcripts (Fig. 8H). These top eight were used as inputs to logistic regression. Performance with qSOFA scores that combine mental status, respiratory rate, and blood pressure was also evaluated in these same subjects. The eight-transcript model had improved performance (AUROC=.81) versus the model using qSOFA alone (AUROC 0.72) in the 441 subjects where matched data were available (Fig. 81). These results show that a molecular prognostic for sepsis mortality can be derived from diverse global cohorts.

[0188] With respect to FIG. 8, (8A) data for this study was generated from subjects enrolled in three different sites located in Cambodia. Duke (USA), and Ghana. (8B) The mean and median age were nearly identical (~50 years of age) for each gender. The median and quartiles are indicated by boxplot and the mean as red dotted lines. (8C, D) The 28-day mortality was also very similarly distributed between genders and averaged a ~17% death ratio for the entire study cohort. (8E) Principal component analysis of the 3061 selected genes did not reveal any site-specific or mortality-specific patient clustering. (8F) Volcano plot of differential gene expression analysis comparing subjects that died by day 28 to those that survived. Labels show for top 10 most significantly changed genes (p. adjust < 0.05). The bar insert shows a total number of significantly changed genes (p.adjust < 0.05). (8G) Results of gene set enrichment analysis (GSEA) using Molecular Signature Database (MSigDB) and Hallmark gene data sets show significantly different pathways (p.adjust < 0.05). Numbers in grey show the percentage of pathway coverage. (8H) The 28-day mortality-based log2 fold changes of eight (8) genes were used for prognostic mortality modeling. (81) Receiver operating characteristic (ROC) curve showing performance of eight-(8)- gene model (AUC = 0.812 ± 0.070) in predicting 28-day mortality relative to the predictive power of quick sepsis-related organ failure assessment (qSOFA, AUC = 0.716 ± 0.075) score. All modeled curves were generated with repeated stratified k-fold cross-validation described in the methods.

[0189] Example 9: Data dimension reduction identifies patient subgroups that improve prognostic performance.

[0190] Genome-wide expression profiling has been used successfully to identify septic shock subclasses that describe biological heterogeneity19,20. We hypothesized that heterogeneity within, and across global cohorts could compromise the performance of the molecular sepsis prognostic. To identify relevant sub-groups for improved prognostic models, we used topological data analysis (TDA) to cluster patients in an unsupervised, data-driven manner21. In TDA a 2-dimensional topological network is created which is based on the similarity between data points as well as theoverall distribution of the data in n-dimensional space (Fig. 9A). This provides an intuitive means of stratifying subjects with similar gene expression profiles into groups where relative position within the network can reveal shared biology or endotypes. TDA was performed resulting in the identification of 5 subgroups with a major left-right axis described by 28-day mortality (Fig. 9B). Groups tl, t2, and t3 all exhibit similar rates of elevated mortality (22%, 22%, and 20%, respectively) versus the low mortality groups t4 and t5 (10% and 11%, respectively) (Fig. 9B). Feature selection was repeated within the TDA subgroups to ask if 28-day mortality prognostic performance could be improved using a stratified approach. Groups t4 and t5 contain too few non survivors individually for this analysis but share a subset of subjects so we combined them into one low mortality group t4-5 (28-day mortality=l 1%). A total of 13 features across these four final groups were identified by step-up performance analysis of the previously identified 50 MRMR transcripts (Fig. 10A). Models consisting of features for specific TDA groups were fitted and evaluated using only the subjects belonging to the respective TDA group. Performance increased in each of the 3 high-mortality subgroups and was decreased in the low-mortality t4-5 group versus what w as achieved for the combined global cohort (Figs. 81 and 10C). All these subgroup-specific features and models proved to be more effective than qSOFA (Fig. 10C).[001911 With respect to FIG. 9, (9A) the normalized expression of the top 1000 genes from 506 subjects was used to perform Topological Data Analysis (TDA). This decomposition method groups similar patients into nodes and produces a relationship netw ork with exclusive and shared patient membership between nodes. (9B) The sepsis TDA network was divided into four groups (left) based on 28-day mortality (right) colored by mortality ranging from green=0% and red=40%. (9C) Patients are stratified into three high-mortality groups (tl, t2, t3) and one low -mortality group (t4).

[0192] With respect to FIG. 10, (10A) the TDA group-specific feature selection identified a total of thirteen genes (13) for stratified 28-day mortality prognostic. Heatmap of hierarchically clustered genes in 28-day mortality comparison across entire study cohort showing TDA group membership and direction of expression change. (10B) TDA overlay showing expression of t o of the biomarker genes across the entire study cohort. TDA subgroups are indicated with dashed lines and labels. (10C) Receiver operating characteristic (ROC) curves showing the performance of prognostic classifiers including the 8-gene model (red) based on the entire cohort versus quick sepsis-related organ failure assessment (qSOFA) score (purple), and TDA-stratified models (blue). All modeled curves were generated with repeated stratified k-fold cross-validation described in the methods.

[0193] Example 10: Molecular definition of sepsis endotypes.

[0194] GSEA using molecular features was used to explore the TDA stratified groups. We first compared GSEA results across the major left-right mortality axis (FIG. 11A and B). Type I and II interferon responses and allograft rejection were elevated in tire low mortality t4-5 group, as were pathways linked to cell growth and proliferation including MYC and E2F targets. Key interferon-stimulated genes including IF127 and ISG15 were significantly elevated in this group. These signatures are consistent with an active adaptive immune response in the low-mortality subgroup. In contrast, the high -mortality groups were enriched for terms including hypoxia and coagulation which are characteristic of severe sepsis with subgroup tl being the most elevated. The individual gcnc-lcvcl inspection identified genes such as CYP1B1, S100A12, and IL1R2 that were most elevated in this high-mortality group tl. Notably, these genes characterize a CD14+immunosuppressive monocytic myeloid-derived suppressor cell (MDSC) population MSI recently described in bacterial sepsis and COVID-19 as negatively associated with survival. Subgroup tl showed the least interferon response signatures by GSEA versus all other groups. Taken together these data are consistent with an immunosuppressed phenotype in group tl. Pathways linked to inflammation including TNFa and IL6-JAK-STAT3 signaling were also elevated in the high mortality groups compared to the low mortality group. Comparison across the three high mortality groups revealed that group t3 is most enriched for acute inflammatory signatures. We assessed differences between groups t2 and t3 and noted a significant difference in heme metabolism consistent with the clinical hemoglobin\hematocrit laboratory' results. We also asked if we could identify pathways linked to mortality within the potential endotypes. A subset of signatures including interferon response and allograft rejection that is a gene-set describing adaptive immune response were positively linked to survival in all groups (FIG. 1 IB). However, we did note pathways with subgroup-specific differences by outcome including tire “inflammatory response” term that was significantly enriched in 28-day survivors in the immunocompetent low-mortality group t4-5.

[0195] Example 11: Classification of sepsis cndotypcs.

[0196] TDA decomposition of host-gene expression from our combined global sepsis cohort suggests at least 4 endotypes (Fig. 12). The low mortality t4-5 group with less severe clinical correlates is enriched for immunocompetent molecular and hematological signatures most notably for the adaptive immune response. This is in direct contrast to high mortality group tl which has reduced lymphocytes and interferon gene expression signatures and has molecular markers consistent with immunosuppressive cells. High-mortality group t3 shares key clinical features of severe illness with group tl, but GSEA shows that there is a robust innate and adaptive immune response in this group. Group t2 is notable for heme metabolism signatures and reducedmitochondrial gene expression linked to oxidative phosphorylation. Overall, results from our combined global cohort are similar to recent reports of biological endotypes using unsupervised analysis of clinical and multi-omics data from sepsis studies. Sweeney and colleagues identified three stable patient subgroups in a combined analysis of bacterial sepsis cohorts designated Inflammopathic, Adaptive, and Coagulopathic23. Another study found three stable sepsis subclasses coined immune-innate (IN), immune-coagulant (IC), and immune-adaptive (IA)7. Finally, Scicluna and colleagues identified four endotypes in European cohorts designated Molecular Diagnosis and Risk Stratification of Sepsis (Mars) 1-413. The Marsl group was characterized by reduced expression of T-cell and adaptive immune genes versus a Mars3 group with robust innate immune activation. Consistent across all these studies is the association of reduced mortality and less severe clinical correlates with endotypes most strongly defined by signatures of a functional adaptive immune response similar to our “immunocompetent” group t4-5. In contrast, subclasses in these studies that have reduced adaptive immune signatures, coagulopathies, elevated mortality, and poor clinical scores align with our “immunosuppressed” group tl . The t3 group described here, with innate pathway signatures linked to “acute inflammation” and high mortality shares these features with tire “inflammopathic” group reported by Sweeney and colleagues23Finally, group t2 characterized by heme biosynthesis is consistent with described “immunometabolic” endotypes and free heme has been shown to contribute to sepsis pathogenesis.

Claims

CLAIMSWHAT IS CLAIMED IS:

1. A method of generating a sepsis mortality prediction model for a subject comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; executing a feature selection algorithm on the training data, wherein executing the feature selection algorithm comprises: determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; and generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data: training a plurality of candidate prediction models using the model subset: determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis.

2. The method of claim 1, further comprising: determining whether one or more values associated with the training data is missing; and filling in, based on determining one or more missing values associated with the training data, the one or more missing values.

3. The method of claim 1 , wherein the biomarker data comprises one or more nucleic acid markers.

4. The method of claim 1, wherein the model subset comprises a level of: adrenoceptor |32 (ADRB2); carboxypeptidase vitellogenic like (CPVL);C-X3-C motif chemokine receptor 1 (CX3CR1); defensin a3 (DEFA3);Fc receptor like 5 (FCRL5): interleukin 10 receptor subunit a (IL 1 ORA):STE20 related adaptor (STRADB); tyrosylprotein sulfotransferase 1 (TPST1);CD 177 molecule; and / or radical S-adenosyl methionine domain containing 2 (RSAD2).

5. Tire method of claim 1. further comprising: generating a second data structure comprising genetic marker data; determining, based on executing tire feature selection algorithm on the genetic marker data, a first genetic marker subset; determining, based on executing a topological data analysis algorithm on the genetic marker data, a plurality of clusters associated with sepsis mortality; determining, based on executing the feature selection algorithm on each of the plurality of clusters, a second genetic marker subset, the second genetic marker subset being different from the first genetic marker subset; determining, based on executing the sepsis mortality prediction model on tire first genetic marker subset, a first genetic marker performance metric; determining, based on executing the sepsis mortality prediction model on the second genetic marker subset, a second genetic marker performance metric; and selecting, based on the higher of the first genetic marker performance metric and second genetic marker performance metric, the first genetic marker subset or the second genetic marker subset as an input for the sepsis mortality’ prediction model.

6. The method of claim 1, wherein determining the relevancy metric comprises: determining, based on a correlation between a biomarker feature of the plurality of biomarker features and a likelihood of sepsis mortality, a relevancy score associated with the biomarker feature.

7. Tire method of claim 6, wherein generating the model subset comprises:determining, for each feature and based on a distance between a selected feature from the training data and each biomarker feature, a redundancy metric associated with each biomarker feature; ranking, based on a difference between the redundancy metric and relevancy metric, the biomarker features associated with a training subset; selecting a biomarker feature associated with a largest difference; generating updated training subset by eliminating the biomarker feature associated with the largest difference from a training subset; and generating the model subset, tire model subset comprising the biomarker feature associated with the largest difference.

8. Tire method of claim 1, wherein determining the performance metric comprises determining, by inputting a test subset into the trained candidate prediction model, a magnitude of an area under a receiver operator curve (AUROC) associated with the trained candidate prediction model.

9. Tire method of claim 8, wherein selecting the sepsis mortality prediction model comprises selecting a candidate prediction model of the plurality of candidate prediction models with a highest magnitude of AUROC.

10. Tire method of claim 1, wherein training the plurality of candidate prediction models comprises: determining, for each candidate prediction model of the plurality of candidate prediction models and based on inputting the model subset into the candidate model, a prediction associated with the candidate prediction model; determining a difference between the prediction and the clinical outcomes; and adjusting, based on the difference, a parameter of the candidate prediction model.

11. Tire method of claim 1, wherein the biomarker data being associated with biological samples from a plurality of individuals different from the subject.

12. A method of predicting sepsis mortality of a subject comprising: receiving value of a patient parameter associated with the subject;executing a sepsis mortality prediction model using the value to generate a prediction associated with sepsis mortality, wherein the sepsis mortality prediction model is generated by performing operations comprising: generating a data structure storing training data, tire training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features: executing a feature selection algorithm on the training data, wherein executing the feature selection algorithm comprises: determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models: and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis; and outputting, by the sepsis mortality prediction model based on value, tire prediction associated with an onset of sepsis mortality within the threshold time after presentation of suspected sepsis in the subject.

13. The method of claim 12, wherein the prediction associated with sepsis mortality comprises a predictions score associated with sepsis mortality within a threshold time after presentation of suspected sepsis.

14. The method of claim 12, wherein the patient parameter comprises biomarker data and the value comprises a level of a biomarker feature in the biomarker data.

15. Tire method of claim 14, wherein the value comprises a level of: adrenoceptor (32 (ADRB2);carboxypeptidase vitellogenic like (CPVL);C-X3-C motif chemokine receptor 1 (CX3CR1); defensin a3 (DEFA3);Fc receptor like 5 (FCRL5); interleukin 10 receptor subunit a (IL 1 ORA):STE20 related adaptor 0 (STRADB); tyrosylprotein sulfotransferase 1 (TPST1);CD 177 molecule; and / or radical S-adenosyl methionine domain containing 2 (RSAD2).

16. Tire method of claim 12, further comprising: receiving a second value associated with biomarker data, the biomarker data being determined from a sample associated with the subject; and wherein the prediction associated with the sepsis mortality is further determined based on the second value.

17. Tire method of claim 12, wherein generating the sepsis mortality prediction model further comprises: generating a second data structure comprising genetic marker data; determining, based on executing tire feature selection algorithm on the genetic marker data, a first genetic marker subset; determining, based on executing a topological data analysis algorithm on the genetic marker data, a plurality of clusters associated with sepsis mortality; determining, based on executing the feature selection algorithm on each of the plurality of clusters, a second genetic marker subset, the second genetic marker subset being different from the first genetic marker subset; determining, based on executing the sepsis mortality prediction model on tire first genetic marker subset, a first genetic marker performance metric; determining, based on executing the sepsis mortality prediction model on the second genetic marker subset, a second genetic marker performance metric; and selecting, based on the higher of the first genetic marker performance metric and second genetic marker performance metric, the first genetic marker subset or the second genetic marker subset as an input for the sepsis mortality prediction model, wherein the value of the patient parameter is a value associated with the selected marker genetic subset.

18. The method of claim 14, wherein outputting the prediction associated with sepsis mortality comprises: determining, based on tire biomarker data, a presence or a dosage of an antibiotic; determining, based on the presence or the dosage of the antibiotic, a dosage of tire antibiotic to administer to the subject; and outputting a recommendation associated with administering the dosage of the antibiotic.

19. A system for generating a sepsis mortality prediction model for a subject comprising: one or more processors; and one or more non-transitory computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perform operations comprising: generating a data structure storing training data, tire training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis.

20. A system for predicting sepsis mortality of a subject comprising: one or more processors; an output component; andone or more computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perform operations comprising: receiving value of a patient parameter associated with the subject; executing a sepsis mortality prediction model using the value to generate a prediction associated with sepsis mortality, wherein the sepsis mortality prediction model is generated by perfonning operations comprising: generating a data structure storing training data, the training data comprising biomarker data and clinical outcomes associated with the biomarker data and the biomarker data comprising a plurality of biomarker features; determining, based on an association between the biomarker data and the clinical outcomes, a relevancy metric for each biomarker feature in the biomarker data; generating, based on the relevancy metric of each biomarker feature, a model subset, the model subset comprising a subset of the training data less than a total amount of the training data; training a plurality of candidate prediction models using the model subset; determining a performance metric for each trained candidate prediction model of the plurality of trained candidate prediction models; and selecting, based on the performance metric, a candidate model from the plurality of trained candidate prediction models as the sepsis mortality prediction model, wherein the sepsis mortality prediction model determines a likelihood of sepsis mortality within a threshold time after presentation of suspected sepsis; and outputting, by the sepsis mortality prediction model based on value, the prediction associated with an onset of sepsis mortality within the threshold time after presentation of suspected sepsis in the subject.

Images