Machine learning differentiation of kawasaki disease

Abstract

Machine learning techniques for differentiating, in a febrile pediatric patient, Kawasaki disease from non-Kawasaki disease are presented. The techniques include: receiving clinical data for the patient, the clinical data including symptom data and laboratory data; inputting the clinical data for the patient to a machine learning system, where respective likeness scores for a plurality of phenogroups are output from the machine learning system, where the plurality of phenogroups include at least one Kawasaki disease phenogroup and at least one non-Kawasaki disease phenogroup, where the machine learning system has been trained with historical clinical data using dimensionality reduction and a clustering algorithm; determining a phenogroup classification for the patient based on the respective likeness scores for the plurality of phenogroups; and providing the phenogroup classification for the patient.

Description

MACHINE LEARNING DIFFERENTIATION OF KAWASAKI DISEASE Related Application

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 538,368, entitled “Machine Learning Differentiation of Multisystem Inflammatory Syndrome in Children and Kawasaki Disease in Covid-19 Patients,” and filed September 14, 2023. Government Funding

[0002] This invention was made with government support under grant HD105591 awarded by the National Institutes of Health. The government has certain rights in the invention. Field

[0003] This disclosure relates generally to machine learning algorithms and classifiers in the medical field. Background

[0004] Multisystem inflammatory syndrome in children (“MIS-C”) is multisystem inflammatory disease linked to SARS-CoV-2 infection. MIS-C has overlapping clinical presentation with that of Kawasaki disease (“KD”). The etiology of KD remains unknown, but strong epidemiologic and mechanistic data support KD being a post infectious vasculitis. The incidence of MIS-C is correlated with the incidence of SARS- CoV-2 infections (although this is affected by specific strains). Initial data shows that the physical manifestations of MIS-C are increasingly overlapping with those of KD as SARS-CoV-2 variants have become milder.Summary

[0005] According to various embodiments, a machine learning method of differentiating, in a febrile pediatric patient, Kawasaki disease from non-Kawasaki disease is presented. The method includes: receiving clinical data for the patient, the clinical data comprising symptom data and laboratory data; inputting the clinical data for the patient to a machine learning system, wherein respective likeness scores for a plurality of phenogroups are output from the machine learning system, wherein the plurality of phenogroups comprises at least one Kawasaki disease phenogroup and at least one non-Kawasaki disease phenogroup, wherein the machine learning system has been trained with historical clinical data using dimensionality reduction and a clustering algorithm; determining a phenogroup classification for the patient based on the respective likeness scores for the plurality of phenogroups; and providing the phenogroup classification for the patient.

[0006] Various optional features of the above method embodiments include the following. The method may include comprising treating the patient based on the phenogroup classification. The at least one non-Kawasaki disease phenogroup may include a COVID-19 only phenogroup. The at least one non-Kawasaki disease phenogroup may include a multisystem inflammatory syndrome in children (MIS-C) phenogroup. The at least one non-Kawasaki disease phenogroup may include a post- viral hyperinflammatory response phenogroup. The at least one Kawasaki disease phenogroup may include a more severe Kawasaki disease phenogroup and a less severe Kawasaki disease phenogroup. The historical clinical data may include historical symptom data, historical laboratory data, and historical diagnosis data, for a plurality of historical patients. The dimensionality reduction may include uniform manifold approximation and projection. The clustering algorithm may include aGaussian mixture model. The method may include providing a quantitative certainty for the phenotype classification, wherein the quantitative certainty for the phenotype classification is based on the respective likeness scores for the plurality of phenogroups, excluding a phenogroup corresponding to the phenogroup classification.

[0007] According to various embodiments, a machine learning system for differentiating, in a febrile pediatric patient, Kawasaki disease from non-Kawasaki disease is presented. The system includes: an electronic processor; and persistent memory comprising non-transitory computer readable instructions that, when executed by the electronic processor, configure the electronic processor to perform actions comprising: receiving clinical data for the patient, the clinical data comprising symptom data and lab data; providing the clinical data for the patient to a machine learning system, wherein respective likeness scores for a plurality of phenogroups are output from the machine learning system, wherein the plurality of phenogroups comprises at least one Kawasaki disease phenogroup and at least one non-Kawasaki disease phenogroup, wherein the machine learning system has been trained with historical clinical data using dimensionality reduction and a clustering algorithm; determining a phenogroup classification for the patient based on the respective likeness scores for the plurality of phenogroups; and outputting the phenogroup classification for the patient.

[0008] Various optional features of the above system embodiments include the following. The actions may further include providing a treatment recommendation based on the phenogroup classification. The at least one non-Kawasaki disease phenogroup may include a COVID-19 only phenogroup. The at least one non- Kawasaki disease phenogroup may include a multisystem inflammatory syndrome inchildren (MIS-C) phenogroup. The at least one non-Kawasaki disease phenogroup may include a post-viral hyperinflammatory response phenogroup. The at least one Kawasaki disease phenogroup may include a more severe Kawasaki disease phenogroup and a less severe Kawasaki disease phenogroup. The historical clinical data may include historical symptom data, historical lab data, and historical diagnosis data, for a plurality of historical patients. The dimensionality reduction may include uniform manifold approximation and projection. The clustering algorithm may include a Gaussian mixture model. The actions may further include providing a quantitative certainty for the phenotype classification, wherein the quantitative certainty for the phenotype classification is based on the respective likeness scores for the plurality of phenogroups, excluding a phenogroup corresponding to the phenogroup classification.

[0009] Combinations, (including multiple dependent combinations) of the above-described elements and those within the specification have been contemplated by the inventors and may be made, except where otherwise indicated or where contradictory. Brief Description of the Drawings

[0010] Various features of the examples can be more fully appreciated, as the same become better understood with reference to the following detailed description of the examples when considered in connection with the accompanying figures, in which:

[0011] Fig. 1 is a flow diagram for a method of training a machine learning system for differentiating, in a COVID-19 patient, a non-Kawasaki disease (e.g., multisystem inflammatory syndrome in children (MIS-C)) from Kawasaki disease;

[0012] Fig.2 is a flow diagram for a machine learning method of differentiating, in a COVID-19 patient, MIS-C or another non-Kawasaki disease from Kawasaki disease;

[0013] Fig.3 is a scatterplot representing risk factor contribution towards cluster formation by quantifying the sum of squared distances between and within clusters;

[0014] Figs.4A and 4B depict charts showing spatial organization of patients in a training cohort overlay with enrollment group / clinical diagnosis and with phenogroup; and

[0015] Fig. 5 is a schematic diagram of a computer system suitable for implementing a machine learning system for differentiating, in a COVID-19 patient, Kawasaki disease from non-Kawasaki disease. Description of the Examples

[0016] Reference will now be made in detail to example implementations, illustrated in the accompanying drawings. Wherever convenient, the same reference numbers will be used throughout the drawings to refer to the same or like parts. In the following description, reference is made to the accompanying drawings that form a part thereof, and in which is shown by way of illustration specific exemplary examples in which the invention may be practiced. These examples are described in sufficient detail to enable those skilled in the art to practice the invention and it is to be understood that other examples may be utilized and that changes may be made without departing from the scope of the invention. The following description is, therefore, merely exemplary.

[0017] The issue of properly recognizing children with MIS-C (or another non-Kawasaki disease) became, and remains to this date, important during the COVID-19 pandemic. Identifying patients with MIS-C, for example, is challenging forat least three reasons. First, neither KD nor MIS-C has a gold standard diagnostic test available; the diagnosis is based on symptomatology. Second, given that MIS-C is a new clinical entity, the CDC criteria for MIS-C is necessarily broad and non-specific. Third, the criteria for MIS-C overlaps with the diagnostic criteria for KD. This has resulted in diagnostic confusion, some degree of hypervigilance, and likely over- diagnosis. For example, a study by the inventors found that 50.4% of patients clinically labelled as having COVID-19 / no MIS-C would meet a strict interpretation of the CDC criteria for MIS-C, despite them not being diagnosed with MIS-C clinically. Conversely, of patients clinically labelled as having MIS-C, 33.3% of the patients also met the diagnosis criteria for incomplete KD, and a further 11.8% met the criteria for complete KD. This diagnostic confusion persists, with 32.6% of the MIS-C patients enrolled in International Kawasaki Disease Registry (“IKDR”) in the Omicron era (Nov / Dec 2021- Jun 2022) meeting the diagnosis criteria for incomplete KD, 19.6% meeting the criteria for complete KD, and still 41.4% of those having COVID-19 / no MIS-C patients meeting the strict criteria for MIS-C. Some embodiments directly address this problem by using unsupervised machine learning to group patients with similar profiles and use this classification as the basis for a diagnostic algorithm to differentiate MIS-C (or another non-KD) from KD in a new patient.

[0018] In general, some embodiments provide a clinical diagnostic classifier that identifies patients with a high probability of having a non-Kawasaki disease (e.g., MIS-C) or KD, versus COVID without the non-Kawasaki disease.

[0019] By way of non-limiting example, embodiments of the invention are described herein in reference to a reduction to practice. According to the reduction to practice, the diagnosis algorithm was to be designed using clinical presentation, history of exposure, and classic clinical biomarkers. Inflammatory and cardiacbiomarker profiles were also investigated relative to model performance. The reduction to practice started with 101 features available on the first day of hospital admission (excluding all diagnostic labels). The features were used to generate phenogroups using a Gaussian mixture model (GMM). Features that did not improve either of the sum of squares (between- or within-clusters) were removed from the analysis. The inventors selected both an appropriate number of clusters (N=7) and the covariance method (full) for the GMM using the Bayesian information criteria (BIC) and the Silhouette score. The UMAP distance metric (Canberra) that maximized the between-clusters sum of squares and minimized the within-clusters sum of squares was selected for clustering.

[0020] For the reduction to practice, seven phenogroups were generated by the GMM. Two clusters included patients with a KD profile, with one such representing the high end of the severity spectrum for KD and another such cluster representing a milder form of KD. Three clusters included patients with non-Kawasaki disease profiles, namely, and by way of non-limiting example, two MIS-C profiles, with one such cluster representing the moderate-to-severe version of MIS-C, one such cluster representing MIS-C in high-risk patients, and another such cluster representing a less severe profile that is still compatible with MIS-C. Finally, two clusters included patients with COVID who did not meet the clinical criteria for MIS-C, one with mild inflammation and one with severe respiratory complications. Once trained, the GMM algorithm generated a likeness score for each patient against the profile of each phenogroup. Using this likeness score allows for automatically and objectively classifying patients in the first two clusters as KD patients, classifying patients in the second three clusters as MIS-C patients (or, more generally, non-Kawasaki disease), and classifying patients in the sixth and seven cluster as neither KD nor MIS-C patients.

[0021] Phenogroup distribution and characteristics remained similar when the reduction to practice was applied to an external validation cohort of 2,938 patients enrolled in the IKDR, and 87.1% of those patients could be classified as KD or MIS-C with >80% certainty. It is usually not feasible to generate the standard performance metrics for classification problems with algorithms created using unsupervised machine learning given the lack of a gold standard. However, the inventors were able to generate such metrics by measuring performance in the subset of patients with no overlapping clinical labelling (e.g., MIS-C patients not meeting KD criteria and vice versa). Performance metrics for the KD and MIS-C patients for the reduction to practice are reported in Tables 1 and 2 below and show excellent performance for both diseases in both the training and validation subsets. Training #F #B AUC p value Se / Sp PP / NPV TP TN FP FN Uncertainty MIS-C 4717 0.82 <.001 77% / 84% 62% / 92% 11235869 33 2.0% KD 47 17 0.84 <.001 82% / 85% 75% / 90% 19938066 44 1.3% Table 1 Validation #F #B AUC p value Se / Sp PP / NPV TP TN FP FN Uncertainty MIS-C 4717 0.82 <.001 85% / 79% 78% / 86% 799869226144 2.6% KD 4717 0.80 <.001 82% / 78% 76% / 84% 632696201138 2.6% Table 2

[0022] In Tables 1 and 2, #F represents the number of features in the final model, #B represents the number of biomarkers with established clinical utility in the final model, FN represents false negatives, FP represents false positives, NPVrepresents negative predictive value, Uncertainty represents the calculated average percent uncertainty in cluster membership, PP-V represents positive predictive value, Se represents sensitivity, Sp represents specificity, TN represents true negative, and TP represents true positive.

[0023] Note that the performance metrics (e.g., AUC) may change over time, and embodiments may be re-trained occasionally (e.g., every few years).

[0024] The inventors also investigated whether adding inflammatory and cardiac biomarker profiles to the classifier of the reduction to practice further improved model performance. As part of the 101 features included, the inventors investigated 42 laboratory biomarkers; of those, 17 were found to improve classifier performance in diagnostic utility testing using a decision curves based on the within- and between- cluster sum of squares (a modification of the clinical utility test for unsupervised machine learning). Notable contributory biomarkers not usually part of standard clinical testing included: BNP, NT-ProBNP and Troponin I.

[0025] These and other features and advantages are shown and described presently in reference to Figs.1-5.

[0026] Fig.1 is a flow diagram for a method 100 of training a machine learning system for differentiating, in a COVID-19 patient, a non-Kawasaki disease (e.g., multisystem inflammatory syndrome in children (MIS-C)) from Kawasaki disease. The method 100 may be implemented using a computer system such as is shown and described herein in reference to Fig.5. By way of non-limiting example, the method 100 is described in reference to the reduction to practice referred to throughout this disclosure.

[0027] At 102, the method 100 obtains historical patient data. Data for the reduction to practice were obtained in collaboration with the IKDR. The IKDR enrollspediatric patients (<18 years old) from more than 40 clinical sites across the world. The IKDR enrolls three groups of patients: (1) patients with SARS-CoV-2 infection with MIS-C (adhering to Centers for Disease Control and Prevention criteria) regardless of whether they have overlapping signs of KD, (2) patients with SARS-CoV-2 infection investigated for but eventually not diagnosed with MIS-C, and (3) patients with KD but without SARS-CoV-2 infection (diagnosed between 2018 and 2022). Only hospitalized patients were eligible for inclusion in the IKDR. Local investigators collected all the related data and entered them into a secured and dedicated IKDR REDCap database.

[0028] A subset of 990 patients enrolled in IKDR were concurrently enrolled in a sub-study aimed at systematically collecting deep phenotypic data for the purpose of creating prediction models for the diagnosis, treatment and management of MIS-C and KD. Model training for the reduction to practice was performed with the data from these 990 patients with validation performed in the remainder of the available IKDR cohort (n=2938).

[0029] A group of 101 clinical factors including clinical presentation, medical history, history of COVID exposure, clinical signs and symptoms at admission, laboratory investigations, and comorbidities were identified as potential predictive features for the reduction to practice. All features were known at or close to the time of hospital admission (i.e., at the time of diagnosis). All diagnostic labels were removed from the potential pool of predictive features, this includes the clinical diagnostic label (i.e., the label that was given by the submitting institution) and whether the patient meet the clinical case definition for either MIS-C or KD.

[0030] At 104, the method 100 preprocesses the data obtained at 102. For the reduction to practice, predictive data imputation was implemented to address themissing instances in the collected data using Decision Tree algorithm iterative imputer (maximum number of iterations = 10) where the imputation was solved for each missing instance based on the known features of a given patient. Post imputation, the data set was investigated for any spurious outliers and limited them by winsorizing the top and lower 2.5% values for each feature in the dataset. All features were then standardized, which was helpful for clustering methods relying on several distance metrics computations quantifying patient similarity. The scaled dataset was then split into training and validation cohorts.

[0031] At 106, the method 100 reduces the dimensionality. This may be accomplished using spectral embedding, by way of non-limiting example.

[0032] In general, the reduction to practice applied unsupervised clustering machine learning methods to congregate patients with similar profiles and used this classification as the basis for a diagnostic algorithm to differentiate MIS-C (or other non-KD) from KD. This was achieved with a two-step protocol: (1) using Uniform Manifold Approximation and Projection (UMAP) for spectral embedding, and (2) generating phenogroups using a Gaussian mixture model (GMM). Phenogroup identity was then assigned based on investigating the risk factors associated within each cluster. This data driven clustering unfolded any hidden similar characteristics with respect to the phenogroup and could classify patients according to the disease severity. Further, phenogroup clusters can provide insights, allowing to select key subset of risk factors that can adequately distinguish patients for disease management.

[0033] UMAP is a non-linear dimensionality reduction technique and was employed in the reduction to practice to optimally summarize the scaled training cohort data with higher dimensions (101 risk factors) projected onto three lower dimensions.UMAP allowed for capturing the feature changes effect at different levels and retaining the most significant information present in the higher dimensions for diagnostic classification in the lower dimensions. After projecting the patients into lower dimensions, a GMM algorithm was implemented to assign cluster membership to each patient.

[0034] At 108, the method 100 includes clustering. For the reduction to practice, a GMM algorithm was used as described above.

[0035] For development of the reduction to practice, two clustering algorithms, K-means and GMM, were considered for cluster formation. GMM was chosen over K-means due to its flexible cluster shapes and probabilistic nature. However, the reduction to practice initialized the GMM clusters using K-means in it to achieve an optimal convergence, improving the cluster formation. An optimal number of clusters using GMM was determined (for the reduction to practice, and by way of non-limiting example, five clusters) using Bayesian information criterion (BIC) and silhouette coefficient score metrics with the elbow method.

[0036] UMAP based on the stochastic gradient algorithm computes a lower dimensional diagnostic landscape, spectrally embedding the patients with similar characteristics by quantifying their relationships to each other in the higher dimensions and preserving the same observed relationship in lower dimensions with a similarity score. Developing the reduction to practice involved investigating a nine-by-four grid with nine UMAP distance computation metrics quantifying similarity between patients in higher dimensions against four types of GMM covariance clustering algorithms. The UMAP distance metric (Canberra) with GMM full type covariance method maximized the between-clusters sum of squares and minimized the within-clusters sum of squares was selected for clustering. Other UMAP parameters including number ofneighbors (n_neighbors = 20), learning rate (learning rate = 0.2), spread (spread = 0.3), minimum distance (min_dist = 0) to address cluster density, were varied around their default values and were chosen preserving the global structure while improving the cluster separation.

[0037] At 110, the method 100 removes non-contributing risk factors. Note that during hyperparameter tuning, 108 may precede 110. For training the final model, 110 may precede 108.

[0038] For the reduction to practice, after identifying the optimal number of clusters, GMM covariance type methods and all UMAP parameters, the inventors proceeded to eliminate non-contributing risk factors from the model. Out of the 93 original features used for clustering, risk factors that did not improve either of the sum of squares (between- or within-clusters) from the cluster analysis were removed. This was accomplished by sequentially removing each feature from the dataset, implementing the cluster analysis, and recording its effect on the sum of squares metrics. These metrics were then compared with the cluster model metrics with all the features included (full model) and then removed if they did not improve either of the sum of squares. After optimization, UMAP spectral embedding together with the GMM clustering has provided distinct patient clusters and allowed for pursuing deep cluster phenotyping in classifying the patients.

[0039] At 112, the method 100 phenomaps patients with cluster assignments.

[0040] For the reduction to practice, phenomapping of patients with GMM cluster assignments was implemented to characterize homogeneous diagnostic patient clusters that significantly improved classification and prediction of the patient diagnosis with overlapping presentation of the disease signs and symptoms. The individual associations between risk factors and the cluster assignments wereconsidered. Kruskal-Wallis rank sum, Pearson’s Chi-squared and Fisher’s exact test were implemented and p-value was assessed for statistical significance, comparing the cluster features grouped by the GMM cluster labels. The results were then analyzed to identify the set of features that characterize each of the GMM clusters and assign appropriate phenotype label to the GMM clusters. All patients from the five GMM clusters were classified into three disease phenogroups, with MIS-C, KD and COVID labels.

[0041] The GMM algorithm generates a likeness score for each patient against the profile of each phenogroup, and using this likeness score allows for classifying patients between the phenogroups, e.g., according to highest likeness score. A patient with high likeness score of association to a phenogroup will have low likeness scores in the remaining phenogroups and vice versa.

[0042] Further, the likeness score can be used to calculate a certainty equivalent around the high likeness score for each patient, by calculating the composite mean of the remaining phenogroup’s likeness scores, excluding their phenogroup association likeness score.

[0043] With the likeliness score for each patient belonging to MIS-C, KD and COVID phenogroups, the reduction to practice assigned the predicted class labels to each patient. These predicted labels for the MIS-C and KD patients were then compared with the nonoverlapping cohort, whose true labels were adhered with respect to CDC criteria for MIS-C and the AHA criteria for KD. Standard classification metrics including area under the curve, sensitivity, specificity, observed to expected ratio representing classifier calibration, positive and negative prediction values, counts of the true, false negatives and positives were summarized for the training andvalidation cohorts. These results are presented herein, e.g., in reference to Tables 1 and 2.

[0044] Fig. 2 is a flow diagram for a machine learning method 200 of differentiating, in a COVID-19 patient, MIS-C (or another non-Kawasaki disease) from Kawasaki disease. The method 200 may be implemented using a computer system such as is shown and described herein in reference to Fig.5. The method 200 may be practiced using a machine learning system trained according to the method 100, as shown and described herein in reference to Fig.1.

[0045] At 202, the method 200 receives clinical data for a patient. The clinical data may be as described herein in reference to Fig.1, for example. By way of non- limiting example, the clinical data may include any combination of the 47 data types used in the reduction to practice. For example, the clinical data may include any, or any combination, of BNP, NT-ProBNP and / or Troponin I. The clinical data may be received by retrieval from electronic memory, over a network connection, of according to a different technique according to various embodiments.

[0046] At 204, the method 200 provides clinical data to a trained machine learning system. The clinical data provided to the trained machine learning system may first undergo pre-processing, e.g., as shown and described herein in reference to Fig.1. The trained machine learning system may be trained according to the method 100 as shown and described in reference to Fig.1.

[0047] At 206, the method 200 obtains likeness scores for each of plurality of phenogroups. According to some embodiments, the plurality of phenogroups may include any two or more of: MIS-C, KD, and / or no MIS-C or KD. In general, the plurality of phenogroups may include one or more Kawasaki disease phenogroups and one ormore non-Kawasaki disease phenogroups. The likeness scores may be obtained from the trained machine learning system as described herein in reference to Fig.1.

[0048] At 208, the method 200 determined a phenogroup classification for the patient. The phenogroup classification may be determined according to the likeness scores, as shown and described herein in reference to Fig.1, for example.

[0049] At 210, the method 200 provides the phenogroup classification for the patient. The phenogroup classification may be provided by display on a computer monitor or other device, provided over a network connection, or provided to a clinical data system, by way of non-limiting examples.

[0050] According to various embodiments, the phenogroup classification may be provided to a different diagnostic or prognostic system, such as a machine learning diagnostic or prognostic system. For example, according to various embodiments, the phenogroup classification may be provided to a machine learning system trained to predict an expected response to therapy, and / or determining an optimal therapy, for patients with Kawasaki disease and / or a non-Kawasaki disease. As another example, according to various embodiments, the phenogroup classification may be provided to a machine learning system trained to predict a risk of complications in patients with Kawasaki disease and / or a non-Kawasaki disease. A machine learning system according to one of the aforementioned examples may be trained using a training corpus that includes a plurality of labeled phenogroup examples. The labels may be set according to the particular machine learning system, e.g., the labels may include therapy responses or complications.

[0051] According to various embodiments, 212 and 214 are optional and may be implemented in conjunction with, e.g., in parallel or at the same time as, 208 and 210, for example.

[0052] At 212, the method 200 determines a classification certainty. The classification certainty may be determined as the certainty equivalent described herein in reference to Fig.1, for example.

[0053] At 214, the method 200 provides the classification certainty. The classification certainty may be provided by display on a computer monitor or other device, provided over a network connection, or provided to a clinical data system, by way of non-limiting examples.

[0054] Further details regarding the development and accuracy of the reduction to practice are presented in detail below.

[0055] The training set for the reduction to practice included 990 patients, of whom 410 (41.4%) had a clinical diagnosis of MIS-C, 343 (34.6%) of KD, and 236 (23.8%) were COVID patients with neither MIS-C or KD. The validation dataset included 2938 patient of whom 1736 (59.1%) with MIS-C, 1146 (39.0%) with KD, and 56 (2.0%) with COVID and neither MIS-C or KD. The training set was enriched for patients in the latter group to facilitate model training, however the validation group is more representative of the patient distribution clinically. In the training and validation cohorts respectively, 135 (33.3%) and 309 (17.8%) of patients clinically labelled as having MIS-C met the criteria for incomplete KD, while 48 (11.8%) and 250 (14.4%) met the criteria for complete KD. Of patients clinically labelled as KD, 153 (44.6%) met the CDC criteria for MIS-C in the training cohort, while 695 (60.6%) did so in the validation cohort.

[0056] The reduction to practice used UMAP spectral embedding to group similar patients into homogeneous clusters. First, that the pre-selected predictive features previously identified clustered similar patients was confirmed. Based on an empirical search in the hyperparameter space, an optimal number of clusters wasdetermined to be five, and a full covariance method was used based on the BIC and the Silhouette score. The UMAP distance metric (Canberra) that maximized the between-clusters sum of squares and minimized the within-clusters sum of squares was selected for clustering.

[0057] Second, feature reduction was applied, and it was determined that 47 out of 93 features did not improve either of the sum of squares (between- or within- clusters). Such features were removed them from the analysis.

[0058] Fig.3 is a scatterplot 300 representing risk factor contribution towards cluster formation by quantifying the sum of squared distances between and within clusters. In particular, the scatterplot 300 the removal of individual variables. The scatterplot 300 is divided into four quadrants, considering the full model (cluster model with all features included) as the origin point. The pertinent negative features present in quadrant A were not contributing for clustering, and removing them from the model increased the distance between the clusters and decreased the within cluster distance relative to the full model. Risk factors present in quadrants B, C, D were considered towards developing a parsimonious clinical classifier. Those features in quadrant A were removed and the remaining 46 features were used to generate the final clusters.

[0059] The reduction to practice utilized clinical phenomapping as described presently in reference to Figs.4A and 4B.

[0060] Figs.4A and 4B depict charts showing spatial organization of patients in a training cohort overlay with enrollment group / clinical diagnosis 402 and with phenogroup 404. In a reduction to practice, a clinical classifier was trained using a training cohort with 990 patients projected onto UMAP spatial dimensions and clustered them by calculating a similarity score. As shown in Figs.4A and 4B, patient clusters are represented with clinical diagnosis 402 labels on the UMAP axesdemonstrate poorly defined boundaries between the clusters. Further, patient clusters are represented with seven phenogroup 404 cluster labels resulting from GMM full covariance method. GMM clusters with the distinct boundaries were then investigated with patient clinical characteristics to map an appropriate phenotype label to each GMM cluster.

[0061] As shown in Fig.4A (402) there is substantial overlap in patient profile between enrollment groups in the training cohort (N=990) with a poorly defined boundary between patients with MIS-C and COVID / no MIS-C and a large number of MIS-C patients showing a KD profile and vice versa. Fig.4B (404) demonstrates the seven phenogroups that were generated by the GMM, clusters B (N=146) and F (N=126) include patients with a KD profile, with cluster B representing the high end of the severity spectrum for KD and cluster F a milder form of KD. The profile of patients in cluster C (N=134) is typical of the moderate-to-severe version of MIS-C, while patients in cluster D (N=233) have a less severe profile that is still compatible with MIS-C and those in cluster E having high-risk features for MIS-C complications. Finally, patients in cluster A (N=148) represent the patients with COVID and mild inflammation who do not meet the clinical criteria for MIS-C and those in cluster G (n=87) represent those with severe COVID and respiratory compromise.

[0062] Clinical phenomapping of the clusters was performed for the reduction to practice with the GMM cluster labels and descriptive statistics quantifying the individual associations between risk factors and the cluster assignments were implemented, as shown in Table 3. Statistically significant associations were observed to characterize the five clusters into three phenogroups with KD, MIS-C and neither KD or MIS-C labels. The GMM algorithm generates a likeness score for each patient against the profile of each phenogroup and using this likeness score allows forclassifying patients in (404) in clusters B and F as KD patients, and patients in clusters C, D and E as MIS-C patients while patients in cluster A and G have neither KD or MIS-C.In Ea EuA(BAEt SeA C dig s r ra la fre te op bic i hcxgani(e ha arnorn eua c / k)n cit mt acsApsn M / B / A yal pr teeia / Widlafr e)esrio nh dlckicpiaenstiltee eEntaca A tioo(Cstmnf Nau ere(yocra nths / riei Wcaar / ae nSno )st / Cs)au Athsiraibn bean84 864 8 8 8 9646464689N 43 8 1 8A 08 3 3 4. (1 ,(0(3.(3(6 (1(58(5 N% 2 1 .5 0 5 %=(1 (2 9 8(33(8(19(.9 (2 ,(4 N.0% .9% 3% .7 1 68 4. %=) ) )% %))%)1))123302 4(0(1 3 2 22 85 1 E0 (.(3(1.(1(6.8 1 ,4N% 7)% 7)% 7% 9))% 3)%(3 %=) 1).8)13425 1 8 F( 91 (8 1 3( 6 5 3.6 (1 ,N.8% .3(42)% 7 .(1(6)% 8% 5))% 8)%(3 3. %=)0))112501 9 1 4 12 (G (00%( 90%(2( 0813( 515(5.26 8(.8,N%=) )% %))%)% 5).7)) 187<< p0 04 0 -. .20 00. . 0 va2003 .0l1 5 01 ue2OSC o So B LaA OCVOTeD i Cth u u la tm herththck inerIV H DIDi CstONst aVeg edgn areaas A ) Aictsma / ex + or I atCos ctAiap t yD+iv neo ee Oisers r Paiicciss coV ut / nIan istan an ficr Ce O si D d i / HVst+C ciIeaOspDnttaV+w Ianic idthm Dci s / CtIe< a6 D nw (eneokns -987 8 9 4 84 84 831072N6 6 64693 12 13 2A (, 2 1 3 2(6(983 (.2 (91.1N( ( 7 545%= 1.60.8(30(201.% 0% 89 .3 %=.0 % 4 5))% %) 1%)) 116) )%)%)84 4( 9 1444(26 3D 5(((12 , 5 2 34N9 56%(2.(1.(1(2% 1)% 63)% 5%=) 24% 0% 9% 7%))133) ) ) )71 7 12 1 E 5 3 1 3( (5 6 (0 (1 ,N( 14 (2 ( 555 .% 2 (% 94 .7 4% %=.)13% .6 9% .(430))%))134) )% %))11 6( 4 18 3 F7 ( , 6 3 14 1714(5(1(313 N(5 (2(1(1% 4% 5% 1 %%=) 1.5% .8% 3 6)) ) )12) )% %5) )53 2 8G2 2 2(6(52 5.3(0 (9 (8.,N( 580% 8%= 3.(03.( 007.(3% %))%)))18 % % 5% 07) ) )%)<0 < p.0 0 -0.0 val0 0 <0 0101 ue. .265.37.00041 6O rEalx Rmtra Ceshe C ir ov nmHSmigyPi ju u hmr Pe rA ACp -e e- le le sr th om mC CoO Oumcan n esttoxi xi gia rbV V cio tlcs iea sly tiovsupte msms ti sn tiensid IiDI+D+chchmana pitnhispre p pr gn gctie / k / kg ad s es r io oin ars no nooatu rto-cdi w wge eess / pneo nrea ar ad cne neepl a (deredicm ia conxp xpnth gogy Cris)descc diosuosudionon ti r rd on e eb ( i s ?ef do a tior ye s n) s98 98 98 98 98 92 97 98 98 96 9)) ) ) ))) ) ) ) ) )13 2 4 2 5 5 1 5 2 1 6 5( 1.( 0( 9( 2 (1 3 .3 4 (3 1 4 9 797 1 3 2(3.%6% 0% 2% 9%0 91. ( (1 .(1(1(5(4) 11 4.4 0 7% 6 0 1 3) %)% % % %)) ) ) ) ) ) ) ) )11 1 18 03 1 47 11 ( 5. 9 4 8 2 11 7(9 (9(( 1 0 39 0 (7 (3 (7 (1 (9(548 9 3 (8.6 .2(.2 .2 .0 .7 . .8% 2% 5% 8% 9 9) 4 2.9 % % % % 2% %)) ) )%))) ) ) ) ) )3 2 2 1 1 1 5 3( 7 3 23 (2 ( ( ( ( 61 3 .2 9 9 4 6 3 2.4% .38%.03%.42 . 8%.3%12.6(4(3( 32( 21( 61( 86( 137)) ) ) )) 1 .0) %)%)%)%)%)%)<0 <.0 0 <0 <0 <0 <0 <0 <0 <0 0 <00.0 .0 .0 .0 .0 .0 .0 .0 .0 .0101 01 01 01 01 01 01 0 41 8 01IrrKPt o CH DC JaM A V D AR So CK iaawlo e y h u y rt o ia b h r oM Mawbilaymnf au d sis a p inllsnd al hg rmii rrhdoino e utghe e atyak oi rsio ch eaiceia tistin e mainrrh hroet e si t akd pi hn e , s gaea at nc coisdeoupllepa om mpisasn ileser)aso npl et ashmet ee A e(en gAH oatlH A tyypoA rKh KD ica eDclpo ac riftormi ttereegr ii aaaly)) ) ) ) ) ) ) ) ) ) ) ) ) )10 7 1 8 2 2 2 3 4 9 6 1 2 4 3 3 3(7(.5 4( 4 4 0 (1 5 (3 5 7 16 2 5 7 6 (25.% 2 1% 1(63(18(15 .5(26 .0(71(50 (8(16(34(28(27 .3))%)%)%)% %))% %))%)% 7)% % % % % %)) ) ) ) )68 1 1 1 4 3 1 3 1 4( 7 3 3 5 5 5 9 0 2 957(1(42.5(1(14.(24.(24.(27.2( 4 3 7 8 8 4 7 48.(3(2(1(3(1(3(2(7% % % % % % % % 0 5% 6% 4% 0% 4% 5% 2 5)) ) ) ) ) ) )% % %)) ) ) ) ) ) ) )5( 1 15 (1 1 2(2 7(9 1(2 3 2((3 1 3 2((3 8 2(9 1(8 3(0 1(6 6 2(1 (2 0.7 .1 1 2 9 1 .4 2 .4 3 3 2 3 1(7 .(0% % 3% 5% 1% 4% % 4 % 2 3 1 4 8 0 3% %)) ) ) ) ) )%))%)%)%)%)%)%)) )<0 <0 0 <0 <0 <0 <0 < < < < < < < < <.0.00.00.00.00.0 00.0 00.0 0.0 0.0 0.0 0.0 0.0 0.0 0.011 1 1 1 1 1 01 01 01 01 01 01 01 01Ch Bl l B B BA A A A C ASo o e ic a s la lk lb h n yOLrio t n ti eA n dd an ar so pp ar ni alumar oscotisthaoreeur a be toarin hilta n it e n ae imnct iape rgy xianuat setrtra ph erirteer to i ac n ns sa f o ie s sp th icge pn e m rap s attiidnae as see e75 7 1 5 5 6 7 6 7 9 9 992224831282103649N75757 937 947414 2 2 0 1A3 2 4 34.7 .8 2) ) 8) .2 .) 03 2 4 3 ))11% % % % %) ) ) ) 34) ) ) ) )101 4 1. .9 8 22 0 0. .6 03 57 69 2 F1 3 (1 ,N0 0 6 3 3( 0 62(4(3.0 .( ( ( ( 5 63(0(05.(2(35 33(3. 0.0 42 68(83(5 %=)) 1%) ) 4) 0) 4) ) ) ) 125)% 0)% 5)% 0)%)103 5 1. 1 23 0.0 1 (G1 2 1 2 19 .1(4( 13( .9 3 64 45 13 36 8.8,N2(1(2 (1 3(82 (2.9 .8 22(3(0.(051(48(6(7 %)=4 .% 3% .1% 6% 1%) ) 3) .1) 5) ) ) 4))187) ) ) ) )<0 < < < < < < p- < < <.0 0 00.0 .00 0 02.0 .00 00.0 00.0 00.0 00.0 v0a0 0l.0 .0 03.2 05.0 0.011 5 1 9 1 1 1 1 ue 021 4 01 01Po PlN MLayL INH H H H H GFi Fe E ED D t sa e amacRS S S emeGb r ry os ir -dste uiule tr co ro p tmt ch at Tr Tr Cm-o T ri rg at nitotihr in eo ctimophpihao e ocyde p oop roea l oc obcr g neoncy phb eiili run ls gt es teshy nd in nintiin it te ls rurvoT I e se big p d ne ro imna teese inntationrate82 78 61 57 58 33 29 9 27 79 88 53 2 4 5 4 4 3 555 5 2 8 2 123 9 6 9335657786739124.4 33 07 4 0.5 .9 37 .3 5 17 37 .18 29 44 2 11 34 31 .7 36 2 0.1 4 1. .1( ( ( ( ( ( ( ( ( 8( 7 .4 7 7( 1( 5( 6( 4 00. 12 3. 0. 2. 33 0.1 41 87 52 (1(4 (8 1. 52 24 0.(28.(135 29 .0 .80 .80 87 .25 10 1 9 1 4 6 .3 1 5 .4 .1 .4( ( ( ( ( ( ( ( 7( 1( 8 .2 5( 5( 7( 3 0 ( 30.6 10 3 5.2 0.4 1.7 20 0.1 13 24 65(12(3. 76 2.4 15(25(0.131(1.) 5) ) 7) 8) 2) 7) ) ) ) ) 1) ) 0) 0) ) 4 .) 5) 99)4.1 24 0 6 0. .5 1.4 73 1.1 1 52 5 12 35 6 4.6 72 4 0.0 3 1. .2( 8( 3( 5( 6( 2( 4( 0 ( 7 4 .2 0 1 4 4 6 0 40.6 14 0 45 .9 0) .5 13 .1 37 8 00 .(117 11)6(06) 4()21(6(4) .4 2(1) .(582 1(36 1(0) .(13.(21.7) ) ) ) ) ) ) ) ) 5) ) 7)<0 < < < < < < < < < <.0 00.0 00.0 00.0 00.0 00.0 0 00.0 .00 02.0 00.0 00.0 0 <0.0 0 <.0 0 <.0 0 <.0 0 <.0 0 0 <.0 .0 00.011 1 1 1 1 1 4 1 1 1 01 01 01 01 01 01 2 01WT T So C R Pa P PNh rit i o re e r ro ro Te gl ty al diua dtib tia thca -Pbl coebimor l ninlool t rh omlcri oBdid ir ec e us bdicroe mbtiolnn NinPe nl lbl scotpimco oula eun nstt t intime85 17 68 8 7 8 3 3 1 26 4 4352761807740399 1.0 .7 1( 6 11 3 47 . 04. 3 23 7 7 34 (2 13 .6 (7, 2,5.(3 1(1 ((1(0 (8.26.3) .50 4 3 8 .5 ) ) 6(18.3 22 75) ) .0) ) 9) ) )15 2. .5 0 10 7 3 35 27 .82 3 ( 01 1 .8 (8 4(5(. 1(7.7 (1(01( .7 3.4(1,33,99 .00 ) (2.7 3) .3 6) 5) 3 .4 6 8) 2) ) ) 8) )10 2. .0 22 1 15 3 43 5 .0 3 ( 1 6. ( 7(5(. 14 .( .7 1( 1( 5 2.3 5 6 8. ( ,1 ,7) 1 15 2 (3 25 0. (5 7) 3) 31 8. 43 26) ) .6) ) 1 0) ) )11 1. .4 7 12 9 3 45 . 2( . 4 05( 5 0 8 32 (2. 14 .5 (7, 3,2. 04 .(88) (3(2(0(.4 .4 50)1.)23(31.9 39 89) 6) 3) ) 7) ) )11 1 1 4. .9 86 13 35 61 .2 99 3 ((6( 1 1 .0 (10 7. 0(1.5 (3(05(.9 5.3(1 ,7 ,426 .94 2 () 3. 0) .5 6) 4) 6 2.4 30 2) ) 5) 6) ) )13 1. .3 59 1 12 3 45 3 .3 3 ( 3. (( . 0 7 2 2 1 2 2 15( ( 5 ( ( .2 4 ( ,2 ,4. 06 .) 7 17 5 (3 13 0.(6 3.69 5. 80 12) ) .0) ) 39 ) ) 2) ) )8 1.9 .6 1( 9 9 3 48 . 26. 6 44 1 2 33 (1. 13 .5 (1,4 8.( 71 0) .(787 ) (3(35(0(787.6(4 9 7) .3) ) .77 ) ) 6 .7 8) ) )<0 < < < <. 000.1 03.0 00.0 00.0 0 0 <0.0 .0 0 <0.0 0 <.0 0.01 1 1 1 01 2 01 01 01C17602_P17602-02 0184.0251-PCT

[0063] The tailored clinical classifier was then evaluated predicting the phenogroup of the patients in validation cohort and their associated likeness score. Patients in the validation cohort were spectrally embedded onto UMAP spatial dimensions and grouping of the similar patients was implemented. Standard classifier performance metrics were then derived from the subset of patients with no overlapping diagnosis present in the training and validation cohorts (see Tables 1 and 2). Phenogroup distribution and characteristics remained similar in the validation cohort and 87.1% of those patients could be classified as KD or MIS-C with >80% certainty.

[0064] A discussion regarding results according to the reduction to practice and alternate embodiments follows.

[0065] The increasing prevalence of MIS-C during the SARS-CoV-2 pandemic with clinical features overlapping KD makes distinguishing these entities critical to refining prognostic and therapeutic understanding. Unstructured machine learning provides clinicians an additional tool for discrimination between these clinically similar diseases. The three distinct input clusters (MIS-C, KD, SARS-CoV-2) of the non- limiting example presented herein resulted in, again by way of non-limiting example, seven distinct output clusters, with the majority of patients being classified similarly both clinically and via the machine generated clusters. The machine generated clusters created more distinction between patients with initially overlapping KD and MIS-C phenotypes, refining them into clusters D and B appearing to correlate with low severity KD and moderate / high severity MIS-C respectively. The closer proximity of these phenotypically less and more severe entities of KD and MIS-C may highlight distinct pathophysiologic disease processes between KD and MIS-C.

[0066] Many of the 47 clinical features entered into the reduction to practice that did not help in creating distinct clusters are frequently used in both diagnosing and monitoring KD and MIS-C, including, but not limited to, the inflammatory markers CRP and fibrin degradation production D-dimer. The seven distinct output clusters created from the remaining 47 features resulted in transformation of the input groups of MIS-C and SARS-CoV-2 correlated in five more distinct clusters (A, C, D, E, and G), which appear to represent three groups of MIS-C patients (C, D and E) and two clusters of SARS-CoV-2 infection without MIS-C or KD (A, G). Those patients initially classified as KD generating the two distinct output clusters (B and D) appeared to correlate clinically with more (B) and less (D) severe KD, but were not correlative of complete vs incomplete KD.

[0067] The use of unstructured machine learning resulted in a degree of correlation between most clinically diagnosed KD, MIS-C, SARS-CoV-2 and the five machine generated clusters. However, other embodiments may utilize any number of clusters, generally at least three. The machine generated clusters identified a number of patients clinically diagnosed with MIS-C that clustered as SARS-CoV-2 and vice versa. The machine clusters also appeared to provide better discrimination between patients with features overlapping MIS-C and KD. These are particularly important findings as the therapeutic and prognostic pathways for these diseases, and the more novel MIS-C in particular, are refined.

[0068] Further, other embodiments may be used to differentiate between Kawasaki disease and any non-Kawasaki disease. For example, embodiments may be used to differentiate between Kawasaki disease and a post-viral hyperinflammatory reaction.

[0069] Note that the disclosed technique may be updated with each new SARS- CoV-2 variants; future versions may be trained and / or evaluated on a different patient population and / or with a different combination of underlying features. For example, alternate embodiments may use other febrile controls in addition to COVID-19 / No MIS- C patients. In such embodiments, the COVID phenogroup may then be replaced by a “neither” group.

[0070] Fig. 5 illustrates a client-server system 500 according to various embodiments. The system 500 includes a client computer 510 and a server computer 520. The client computer 510 is communicatively coupled with the server computer 520 over a network 530, such as the internet.

[0071] In general, some embodiments may be deployed using a cloud-hosted multi-tenant application, e.g., hosted by the server computer 520 and accessible remotely by the client computer 510 using a healthcare provider portal, for example. A staging area / encrypted bridge may be deployed within each site’s internal firewall for the packaging, encryption and transfer of data between the sites (e.g., at the client computer 510) and the application (e.g., at the server computer 520).

[0072] The client computer 510 includes one or more processors 512, persistent storage 514, such as a hard drive, and a display 516. Processor(s) 512 may include a processor, a microprocessor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or another type of processor that interprets and executes instructions. Persistent storage 514 may include a magnetic storage medium, such as a hard disk drive, or a removable memory, such as a flash memory. Persistent storage 514 may store data and / or program instructions. Client computer 510 also includes an input device 518, which may include a component that permits a user to input information to client computer 510, such as a control button, akeyboard, a keypad, or another type of input device. Display 516 may include a component that outputs information to the operator, such as a light emitting diode (LED) display, a CRT display, or another type of visual output device.

[0073] Client computer 510 may also include a transient memory, which may include a random-access memory (RAM) or another type of dynamic storage device that stores information or instructions for execution by processor(s) 512. Client computer 510 may include a communication interface, which may include any transceiver-like component that allows client computer 510 to communicate with other devices, such as server computer 520, or networks. In some examples, the communication interface may include a wireless interface, a wired interface, or a combination of a wireless interface and a wired interface. In some examples, the communication interface may receive data over a network, such as network 530, e.g., from server computer 530, and may forward the data for storage in a computer readable storage medium (e.g., persistent storage 514).

[0074] The server computer 520 includes one or more processors 522 and persistent storage 524. The client computer 510 and server computer 520 may include additional components other than those explicitly depicted in Fig.5. Server computer 520 includes one or more processors 522, and persistent storage 524, such as a hard drive. Processor(s) 522 may include a processor, a microprocessor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or another type of processor that interprets and executes instructions. Persistent storage 524 may include a magnetic storage medium, such as a hard disk drive, or a removable memory, such as a flash memory. Persistent storage 524 may store data and / or program instructions. Server computer 520 may also include an input device, which may include a component that permits a user to input information to server computer520, such as a control button, a keyboard, a keypad, or another type of input device. Server computer 520 may also include a display, which may include a component that outputs information to the operator, such as a light emitting diode (LED) display, a CRT display, or another type of visual output device.

[0075] Server computer 520 may also include a transient memory, which may include a random-access memory (RAM) or another type of dynamic storage device that stores information or instructions for execution by processor(s) 522. Server computer 520 may include a communication interface, which may include any transceiver-like component that allows server computer 520 to communicate with other devices, such as client computer 510, or networks. In some examples, the communication interface may include a wireless interface, a wired interface, or a combination of a wireless interface and a wired interface. In some examples, the communication interface may send data over a network, such as network 530, e.g., to client computer 510.

[0076] Certain examples can be performed using a computer program or set of programs. The computer programs can exist in a variety of forms both active and inactive. For example, the computer programs can exist as software program(s) comprised of program instructions in source code, object code, executable code or other formats; firmware program(s), or hardware description language (HDL) files. Any of the above can be embodied on a transitory or non-transitory computer readable medium, which include storage devices and signals, in compressed or uncompressed form. Exemplary computer readable storage devices include conventional computer system RAM (random-access memory), ROM (read-only memory), EPROM (erasable, programmable ROM), EEPROM (electrically erasable, programmable ROM), and magnetic or optical disks or tapes.

[0077] 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 using computer readable program instructions that are executed by a processor, e.g., as shown in Fig.5.

[0078] These computer readable program instructions may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus, e.g., as shown in Fig. 5, 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. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0079] In embodiments, the computer readable program instructions may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state- setting data, configuration data for integrated circuitry, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++, or the like, andprocedural programming languages, such as the C programming language or similar programming languages. The computer readable program instructions may execute entirely on a 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.

[0080] As used herein, the terms “A or B” and “A and / or B” are intended to encompass A, B, or {A and B}. Further, the terms “A, B, or C” and “A, B, and / or C” are intended to encompass single items, pairs of items, or all items, that is, all of: A, B, C, {A and B}, {A and C}, {B and C}, and {A and B and C}. The term “or” as used herein means “and / or.”

[0081] As used herein, language such as “at least one of X, Y, and Z,” “at least one of X, Y, or Z,” “at least one or more of X, Y, and Z,” “at least one or more of X, Y, or Z,” “at least one or more of X, Y, and / or Z,” or “at least one of X, Y, and / or Z,” is intended to be inclusive of both a single item (e.g., just X, or just Y, or just Z) and multiple items (e.g., {X and Y}, {X and Z}, {Y and Z}, or {X, Y, and Z}). The phrase “at least one of” and similar phrases are not intended to convey a requirement that each possible item must be present, although each possible item may be present.

[0082] The techniques presented and claimed herein are referenced and applied to material objects and concrete examples of a practical nature that demonstrably improve the present technical field and, as such, are not abstract, intangible or purely theoretical. Further, if any claims appended to the end of this specification contain one or more elements designated as “means for [perform]ing [a function]…” or “step for [perform]ing [a function]…”, it is intended that such elements are to be interpreted under 35 U.S.C. § 112(f). However, for any claims containingelements designated in any other manner, it is intended that such elements are not to be interpreted under 35 U.S.C. § 112(f).

[0083] While the invention has been described with reference to the exemplary examples thereof, those skilled in the art will be able to make various modifications to the described examples without departing from the true spirit and scope. The terms and descriptions used herein are set forth by way of illustration only and are not meant as limitations. In particular, although the method has been described by examples, the steps of the method can be performed in a different order than illustrated or simultaneously. Those skilled in the art will recognize that these and other variations are possible within the spirit and scope as defined in the following claims and their equivalents.

Claims

What is claimed is:

1. A machine learning method of differentiating, in a febrile pediatric patient, Kawasaki disease from non-Kawasaki disease, the method comprising: receiving clinical data for the patient, the clinical data comprising symptom data and laboratory data; inputting the clinical data for the patient to a machine learning system, wherein respective likeness scores for a plurality of phenogroups are output from the machine learning system, wherein the plurality of phenogroups comprises at least one Kawasaki disease phenogroup and at least one non-Kawasaki disease phenogroup, wherein the machine learning system has been trained with historical clinical data using dimensionality reduction and a clustering algorithm; determining a phenogroup classification for the patient based on the respective likeness scores for the plurality of phenogroups; and providing the phenogroup classification for the patient.

2. The method of claim 1, further comprising treating the patient based on the phenogroup classification.

3. The method of claim 1, wherein the at least one non-Kawasaki disease phenogroup comprises a COVID-19 only phenogroup.

4. The method of claim 1, wherein the at least one non-Kawasaki disease phenogroup comprises a multisystem inflammatory syndrome in children (MIS-C) phenogroup.

5. The method of claim 1, wherein the at least one non-Kawasaki disease phenogroup comprises a post-viral hyperinflammatory response phenogroup.

6. The method of claim 1, wherein the at least one Kawasaki disease phenogroup comprises a more severe Kawasaki disease phenogroup and a less severe Kawasaki disease phenogroup.

7. The method of claim 1, wherein the historical clinical data comprises historical symptom data, historical laboratory data, and historical diagnosis data, for a plurality of historical patients.

8. The method of claim 1, wherein the dimensionality reduction comprises uniform manifold approximation and projection.

9. The method of claim 1, wherein the clustering algorithm comprises a Gaussian mixture model.

10. The method of claim 1, further comprising providing a quantitative certainty for the phenotype classification, wherein the quantitative certainty for the phenotype classification is based on the respective likeness scores for the plurality of phenogroups, excluding a phenogroup corresponding to the phenogroup classification.

11. A machine learning system for differentiating, in a febrile pediatric patient, Kawasaki disease from non-Kawasaki disease, the system comprising: an electronic processor; and persistent memory comprising non-transitory computer readable instructions that, when executed by the electronic processor, configure the electronic processor to perform actions comprising: receiving clinical data for the patient, the clinical data comprising symptom data and lab data; providing the clinical data for the patient to a machine learning system, wherein respective likeness scores for a plurality of phenogroups are output from the machine learning system, wherein the plurality of phenogroups comprises at least one Kawasaki disease phenogroup and at least one non-Kawasaki disease phenogroup, wherein the machine learning system has been trained with historical clinical data using dimensionality reduction and a clustering algorithm; determining a phenogroup classification for the patient based on the respective likeness scores for the plurality of phenogroups; and outputting the phenogroup classification for the patient.

12. The system of claim 11, wherein the actions further comprise providing a treatment recommendation based on the phenogroup classification.

13. The system of claim 11, wherein the at least one non-Kawasaki disease phenogroup comprises a COVID-19 only phenogroup.

14. The system of claim 11, wherein the at least one non-Kawasaki disease phenogroup comprises a multisystem inflammatory syndrome in children (MIS-C) phenogroup.

15. The system of claim 11, wherein the at least one non-Kawasaki disease phenogroup comprises a post-viral hyperinflammatory response phenogroup.

16. The system of claim 11, wherein the at least one Kawasaki disease phenogroup comprises a more severe Kawasaki disease phenogroup and a less severe Kawasaki disease phenogroup.

17. The system of claim 11, wherein the historical clinical data comprises historical symptom data, historical lab data, and historical diagnosis data, for a plurality of historical patients.

18. The system of claim 11, wherein the dimensionality reduction comprises uniform manifold approximation and projection.

19. The system of claim 11, wherein the clustering algorithm comprises a Gaussian mixture model.

20. The system of claim 11, wherein the actions further comprise providing a quantitative certainty for the phenotype classification, wherein the quantitative certainty for the phenotype classification is based on the respective likeness scoresfor the plurality of phenogroups, excluding a phenogroup corresponding to the phenogroup classification.

Images