Abstract
Rationale
More targeted management of severe acute pediatric asthma could improve clinical outcomes.
Objectives
To identify distinct clinical phenotypes of severe acute pediatric asthma using variables obtained in the first 12 h of hospitalization.
Methods
We conducted a retrospective cohort study in a quaternary care children's hospital from 2014 to 2022. Encounters for children ages 2–18 years admitted to the hospital for asthma were included. We used consensus k means clustering with patient demographics, vital signs, diagnostics, and laboratory data obtained in the first 12 h of hospitalization.
Measurements and Main Results
The study population included 683 encounters divided into derivation (80%) and validation (20%) sets, and two distinct clusters were identified. Compared to Cluster 1 in the derivation set, Cluster 2 encounters (177 [32%]) were older (11 years [8; 14] vs. 5 years [3; 8]; p < .01) and more commonly males (63% vs. 53%; p = .03) of Black race (51% vs. 40%; p = .03) with non‐Hispanic ethnicity (96% vs. 84%; p < .01). Cluster 2 encounters had smaller improvements in vital signs at 12‐h including percent change in heart rate (−1.7 [−11.7; 12.7] vs. −7.8 [−18.5; 1.7]; p < .01), and respiratory rate (0.0 [−20.0; 22.2] vs. −11.4 [−27.3; 9.0]; p < .01). Encounters in Cluster 2 had lower percentages of neutrophils (70.0 [55.0; 83.0] vs. 85.0 [77.0; 90.0]; p < .01) and higher percentages of lymphocytes (17.0 [8.0; 32.0] vs. 9.0 [5.3; 14.0]; p < .01). Cluster 2 encounters had higher rates of invasive mechanical ventilation (23% vs. 5%; p < .01), longer hospital length of stay (4.5 [2.6; 8.8] vs. 2.9 [2.0; 4.3]; p < .01), and a higher mortality rate (7.3% vs. 0.0%; p < .01). The predicted cluster assignments in the validation set shared the same ratio (~2:1), and many of the same characteristics.
Conclusions
We identified two clinical phenotypes of severe acute pediatric asthma which exhibited distinct clinical features and outcomes.
Keywords: asthma, informatics, machine learning, pediatrics
1. INTRODUCTION
Pediatric asthma is the most common chronic disease of childhood and is a frequent cause of hospitalization. Between 2013 and 2021, over 24,000 children were hospitalized for asthma each year in the United States, and over 3000 of these required admission to the intensive care unit. 1 While most children with asthma can be managed in the outpatient setting, many are unresponsive to initial treatment with inhaled bronchodilator therapy and require inpatient treatment for severe acute asthma. 2 , 3 New therapeutic regimens are being developed for outpatient asthma treatment focused on prevention of severe acute asthma, 4 however, there is limited guidance for inpatient management of severe acute asthma, and providers have wide variability in management practices. 1 , 5 , 6 Further research evaluating the effectiveness of therapies for severe acute asthma in children is needed to identify optimal treatment approaches and improve clinical outcomes.
Previous studies have identified unique sub‐types of asthma. Research done by the Severe Asthma Research Program used unsupervised clustering algorithms to identify four distinct clinical phenotypes of asthma with unique disease profiles and clinical outcomes. 7 Translational studies have identified separate biologic asthma endotypes, including the type‐2 high, allergic, and neutrophil‐predominant endotypes. 8 , 9 , 10 , 11 Children admitted to the hospital for severe acute asthma likely have separate mechanistic disease pathways at work that appear similar to the bedside clinicians, yet, due to the underlying pathophysiologic mechanism, may be clinically distinct from one another and may respond differently to therapies. The objective of this study was to identify distinct clinical phenotypes in severe acute pediatric asthma using unsupervised machine learning methods. We sought to do this using data readily available in the electronic health record within the first 12 h of hospital admission.
2. METHODS
This was a retrospective cohort study at a single, quaternary care, academic pediatric institution. This study was reviewed by the Indiana University Institutional Review Board (IRB #17481) and determined to be Human Subjects Research Exempt. No informed consent was obtained or required since the study met criteria for Human Subjects Research Exemption. The study population included all patients admitted to the hospital between January 1, 2014 and December 31, 2022 for pediatric asthma. The population age range was 2–18 years. The determination of asthma diagnosis was made using a deterministic computational phenotype including an International Classification of Disease 9th or 10th revision (ICD 9/10) diagnostic code for asthma, the receipt of at least two doses of nebulized albuterol or continuous nebulized albuterol within the first 24 h of admission, and the receipt of an oral or intravenous dose of systemic steroids (prednisone, prednisolone, dexamethasone, or methylprednisolone) in the first 24 h of hospitalization. This phenotype had a positive predictive value of 96% when evaluated with 100 randomly sampled manual chart reviews. We also included only patients who had laboratory tests performed within the first 12 h of presentation to the hospital or emergency department. Patients with a current tracheostomy at the time of admission or who were admitted to the hospital for less than 12 h were excluded. Findings from this study were reported according to the STROBE and TRIPOD guidelines. 12 , 13
Data for this study was obtained from the Indiana University Health Cerner Enterprise Data Warehouse. The population was determined using the above inclusion criteria. Data obtained for the study included patient demographics, diagnostic data, vital signs obtained longitudinally at hourly intervals, respiratory data including device and fraction of inspired oxygen, the Pediatric Asthma Severity Score 14 (PASS), laboratory data, medication data (Supporting Information S1: Table S1).
Encounter level data was used for the unsupervised clustering algorithm. Data pre‐processing included transformation of categorical data into binary format using “One‐hot encoding”. 15 While patient race may have some biologic influence in pediatric asthma, it is a social construct that also reflects other factors including social and environmental drivers of health. 16 In our study, we chose to include patient race as a binary variable (Black or non‐Black) as Black race has been shown in prior studies to influence clinical outcomes in asthma. 17 , 18 This decision was made based on the possibility that, due to several possible genetic, social, and environmental factors, patients identifying as Black race may have a different phenotype of pediatric severe acute asthma with differential clinical outcomes that can be identified through machine learning methods. Normal values were imputed for missing laboratory data and temperature was transformed to fever or no fever using 38.0℃ as the cut‐off. Pneumonia diagnosis was measured using ICD diagnostic codes present in the clinical record during each encounter. Data missingness for specific variables is detailed in Supporting Information S1: Table S1. We used consensus k‐means clustering as the unsupervised learning method to identify the phenotypes. The full cohort was split into an 80% derivation cohort and a 20% validation cohort using total hospital length of stay (LOS) as a sample partition. The k‐means clustering algorithm was applied to the derivation cohort using Euclidean distance as the distance metric and completing 25 iterations. The selection of the optimal number of clusters was done using both the elbow method 19 and in‐depth evaluation of the clinical features seen in varying numbers of clusters. The clustering was performed using encounter‐level demographics, diagnostic data, vital signs obtained on admission and at 12‐h after admission, the initial PASS value and the 12‐h PASS value, and laboratory data obtained in the first 12 h.
After derivation of the clusters, a logistic regression model was created to predict cluster assignment based on the same input variables used in the clustering algorithm. This model was trained on the derivation cohort, then applied to the validation cohort. The predicted cluster characteristics were compared between each other, and to the cluster characteristics found in the original clusters, including both the characteristics included in the logistic regression model (those available in the first 12 h of hospitalization), and outcome variables. Categorical variables are presented as counts with percentages; continuous variables are primarily presented as medians with interquartile range. Changes in vital signs and the PASS are presented as percentage change from the initial values obtained on hospital admission. Percent change in the PASS is described using mean with standard deviation to improve interpretability as the PASS has a small range of discrete values. Categorical variables were compared using Chi‐square tests or Fisher's exact tests for categories with sizes of five or less, and continuous variables were compared using Mann‐Whitney U tests. All data processing and statistics was done using R version 4.2.3. 20
3. RESULTS
The total study population of hospitalized children with a diagnosis of asthma contained 2828 encounters. Of these, 683 individual encounters had laboratory tests obtained in the first 12 h of hospitalization and were included in the clustering analysis. The derivation cohort contained 547 encounters (80%), and the validation cohort contained 136 encounters (20%) (Table 1). The derivation cohort median age was 7 years (IQR: 4–11), and encounter characteristics included 43.7% Black race, 55.9% Male, 61.4% included PICU admission, and the median hospital length of stay was 3.3 days (IQR: 2.1–5.4). There were no significant differences in any of the variables between the derivation and validation cohorts.
Table 1.
Population characteristics.
| Variable | Derivation | Validation | p value |
|---|---|---|---|
| N = 547 | N = 136 | ||
| Age | 7 (4; 11) | 6 (4; 9) | .17 |
| Sex (% Male) | 55.9 | 61.0 | .33 |
| Race (% Black) | 43.7 | 39.0 | .37 |
| Ethnicity (% Hispanic or Latino) | 12.1 | 13.2 | .82 |
| Pneumonia (%) | 41.5 | 43.4 | .76 |
| Complex Chronic Condition (%) | 10.6 | 12.5 | .46 |
| Fever (%) | 16.3 | 17.6 | .80 |
| HR Initial | 135 (117; 151) | 139 (119; 152) | .43 |
| HR 12‐h change (%) | −6.3 (−16.4; 5.1) | −8.2 (−17.5; 3.6) | .26 |
| RR Initial | 30 (24; 39) | 30 (24; 40) | .95 |
| RR 12‐h change (%) | −6.7 (−25.0; 13.6) | −5.4 (−24.1; 17.5) | .51 |
| PASS Initial | 8 (8; 8) | 8 (8; 8) | .36 |
| PASS 12‐h change (%) (Mean) (Std) | −2.9 (14.6) | 0.2 (16.6) | .11 |
| White Blood Cell Count | 11.6 (9.0; 15.3) | 12.0 (8.2; 15.7) | .93 |
| Neutrophil % | 81.0 (70.0; 89.0) | 80.0 (70.0; 89.0) | .39 |
| Lymphocyte % | 10.0 (6.0; 18.0) | 11.0 (6.0; 20.0) | .26 |
| Eosinophil % | 0.0 (0.0; 2.0) | 0.0 (0.0; 1.0) | .29 |
| Magnesium Use (%) | 49.2 | 52.9 | .49 |
| Continuous Albuterol (%) | 53.7 | 50.0 | .49 |
| Ipratropium (%) | 80.8 | 80.1 | .96 |
| HFNC (%) | 45.2 | 50.7 | .28 |
| BiPAP (%) | 12.6 | 13.2 | .96 |
| IMV (%) | 10.4 | 15.4 | .13 |
| PICU Admission (%) | 61.4 | 66.2 | .36 |
| PICU LOS (days) | 1.8 (1.1; 3.3) | 1.8 (1.2; 3.3) | .33 |
| Hospital LOS (days) | 3.3 (2.1; 5.4) | 3.3 (2.1; 5.5) | .71 |
| Mortality (%) | 2.4 | 1.5 | .75 |
Note: Continuous variables presented as medians with interquartile ranges unless otherwise specified.
Abbreviations: BiPAP, bi‐level positive airway pressure; HFNC, high flow nasal cannula; HR, heart rate; IMV, invasive mechanical ventilation; LOS, length of stay; PASS, Pediatric Asthma Severity Score; PICU, pediatric intensive care unit; RR, respiratory rate.
The clustering algorithm identified two distinct clusters of patients in the derivation cohort (n = 547) (Figure 1). Cluster 1 encompassed 68% of the encounters (n = 370), and Cluster 2 encompassed 32% (n = 177). Patients in Cluster 2 were older and more commonly Males of Black Race with non‐Hispanic ethnicity (Figure 2) (Table 2). Cluster 2 encounters had significantly smaller improvements in bedside clinical measures at 12‐h postadmission including heart rate, respiratory rate, and the PASS. Comparing the laboratory values, encounters in Cluster 2 had lower percentages of neutrophils, higher percentages of lymphocytes, higher percentages of eosinophils, and no difference in total white blood cell count. Cluster 2 encounters had lower rates of magnesium use, continuous albuterol use, ipratropium use, and higher rates of aminophylline use. Examining respiratory support use, Cluster 2 encounters had lower rates of HFNC use, equal rates of BiPAP use, and higher rates of IMV use. Outcomes for encounters in Cluster 2 included longer hospital LOS, longer pediatric intensive care unit (PICU) LOS for those admitted to the PICU, and a higher mortality rate.
Figure 1.
Principle component plot illustrating the consensus k means clustering results on the derivation set.
Figure 2.
Six boxplots comparing clinical features of the k‐means identified cluster 1 encounters and cluster 2 encounters.
Table 2.
Cluster characteristics.
| Variable | Cluster 1 | Cluster 2 | p value |
|---|---|---|---|
| N = 370 | N = 177 | ||
| Age | 5 (3; 8) | 11 (8; 14) | <.01* |
| Sex (% Male) | 52.7 | 62.7 | .03* |
| Race (% Black) | 40.3 | 50.8 | .03* |
| Ethnicity (% Hispanic or Latino) | 15.9 | 4.0 | <.01* |
| Pneumonia (%) | 44.3 | 35.6 | .06 |
| Complex chronic condition (%) | 16.5 | 30.0 | .06 |
| Fever (%) | 18.1 | 12.4 | .12 |
| HR Initial | 142 (129; 155) | 116 (102; 131) | <.01* |
| HR 12‐h change (%) | −7.8 (−18.5; 1.7) | −1.7 (−11.7; 12.7) | <.01* |
| RR Initial | 34 (28; 42) | 24 (20; 28) | <.01* |
| RR 12‐h change (%) | −11.4 (−27.3; 9.0) | 0.0 (−20.0; 22.2) | <.01* |
| PASS initial | 8 (8; 9) | 8 (8; 8) | <.01* |
| PASS 12‐h change (%) (Mean) (Std) | −4.4 (15.9) | 0.2 (10.9) | <.01* |
| White blood cell count | 11.6 (9.1; 15.0) | 11.2 (8.6; 15.8) | .99 |
| Neutrophil % | 85.0 (77.0; 90.0) | 70.0 (55.0; 83.0) | <.01* |
| Lymphocyte % | 9.0 (5.3; 14.0) | 17.0 (8.0; 32.0) | <.01* |
| Eosinophil % | 0.0 (0.0; 1.0) | 1.0 (0.0; 4.0) | <.01* |
| Magnesium use (%) | 56.5 | 33.9 | <.01* |
| Continuous albuterol (%) | 58.1 | 44.6 | <.01* |
| Ipratropium (%) | 88.6 | 64.4 | <.01* |
| HFNC (%) | 53.0 | 28.8 | <.01* |
| BiPAP (%) | 13.0 | 11.9 | .82 |
| IMV (%) | 4.6 | 22.6 | <.01* |
| PICU Admission (%) | 64.3 | 55.4 | .05 |
| PICU LOS (days) | 1.6 (1.0; 2.8) | 2.5 (1.3; 4.8) | <.01* |
| Hospital LOS (days) | 2.9 (2.0; 4.3) | 4.5 (2.6; 8.8) | <.01* |
| Mortality (%) | 0.0 | 7.3 | NA |
Note: Continuous variables presented as medians with interquartile ranges unless otherwise specified.
Abbreviations: BiPAP, bi‐level positive airway pressure; HFNC, high flow nasal cannula; HR, heart rate; IMV, invasive mechanical ventilation; LOS, length of stay; PASS, Pediatric Asthma Severity Score; PICU, pediatric intensive care unit; RR, respiratory rate.
Statistically significant at an alpha level of 0.05.
The validation cohort consisted of 136 encounters, and upon applying the logistic regression model to this cohort, 91 (67%) of encounters were identified as Cluster 1 and 45 (33%) were identified as Cluster 2 (Table 3). Several of the same characteristics were identified by the model as significantly different in the Clusters including patient age, clinical improvement in the initial 12 h of hospitalization (heart rate and respiratory rate), and laboratory values (neutrophils, lymphocytes, and eosinophils) (Table 4). Other variables such as race, medication use, respiratory support use, and LOS were not significantly different.
Table 3.
Logistic regression coefficients.
| Variable | OR of cluster 2 assignment |
|---|---|
| Age | 2.43 |
| Black race | 3.22 |
| Hispanic or Latino ethnicity | 0.01 |
| Male sex | 15.3 |
| Pneumonia | 0.61 |
| Complex chronic condition | 3.00 |
| Bronchopulmonary dysplasia | 3.77 |
| Fever | 0.80 |
| HR initial | 0.87 |
| HR 12‐h | 1.07 |
| RR initial | 0.81 |
| RR 12‐h | 1.19 |
| SpO2 initial | 1.44 |
| SpO2 12‐h | 0.72 |
| FiO2 initial | 1.63 |
| FiO2 12‐h | 1.30 |
| PASS initial | 0.33 |
| PASS 12‐h | 2.24 |
| SBP initial | 1.31 |
| SBP 12‐h | 0.94 |
| DBP initial | 1.26 |
| DBP 12‐h | 0.92 |
| White blood cell count | 1.00 |
| Neutrophil % | 0.84 |
| Lymphocyte % | 1.30 |
| Eosinophil % | 3.17 |
| Hgb | 1.34 |
Abbreviations: DBP, diastolic blood pressure; FiO2, fraction of inspired oxygen; Hgb, hemoglobin; HR, heart rate; OR, odds ratios; PASS, Pediatric Asthma Severity Score; RR, respiratory rate; SBP, systolic blood pressure; SpO2, oxygen saturation.
Table 4.
Predicted cluster characteristics.
| Variable | Cluster 1 | Cluster 2 | p value |
|---|---|---|---|
| N = 91 | N = 45 | ||
| Age | 5 (3; 8) | 8 (5; 14) | <.01* |
| Sex (% Male) | 58.2 | 66.7 | .45 |
| Race (% Black) | 37.4 | 42.2 | .07 |
| Ethnicity (% Hispanic or Latino) | 17.6 | 4.4 | .06 |
| Pneumonia (%) | 47.3 | 35.6 | .27 |
| Complex chronic condition (%) | 23.1 | 40 | .06 |
| Fever (%) | 20.9 | 11.1 | .24 |
| HR initial | 141 (124; 157) | 128 (106; 141) | <.01* |
| HR 12‐h change (%) | −9.1 (−17.7; −1.5) | −5.0 (−16.0; 10.5) | .05* |
| RR initial | 34 (28; 42) | 24 (20; 30) | <.01* |
| RR 12‐h change (%) | −11.1 (−27.4; 12.5) | 0.0 (−14.3; 33.3) | <.01* |
| PASS initial | 8 (8; 8) | 8 (8; 8) | .10 |
| PASS 12‐h change (%) (Mean) (Std) | −1.8 (15.8) | 3.0 (17.8) | .23 |
| White blood cell count | 12.4 (8.3; 16.5) | 10.1 (7.9; 14.5) | .21 |
| Neutrophil % | 84.0 (74.5; 89.0) | 73.0 (61.0; 81.0) | <.01* |
| Lymphocyte % | 10.0 (6.0; 15.0) | 18.0 (10.0; 26.0) | <.01* |
| Eosinophil % | 0.0 (0.0; 1.0) | 1.0 (0.0; 4.0) | <.01* |
| Magnesium use (%) | 54.9 | 48.9 | .63 |
| Continuous albuterol (%) | 49.5 | 51.1 | 1.0 |
| Ipratropium (%) | 80.2 | 80.0 | 1.0 |
| HFNC (%) | 53.8 | 44.4 | .40 |
| BiPAP (%) | 12.1 | 15.6 | .77 |
| IMV (%) | 12.1 | 22.2 | .20 |
| PICU admission (%) | 68.1 | 62.2 | .62 |
| PICU LOS (days) | 1.8 (1.2; 2.8) | 1.8 (1.4; 5.3) | .18 |
| Hospital LOS (days) | 3.3 (2.2; 5.0) | 3.4 (2.1; 8.8) | .35 |
| Mortality (%) | 0.0 | 4.4 | .11 |
Note: Continuous variables presented as medians with interquartile ranges unless otherwise specified.
Abbreviations: BiPAP, bi‐level positive airway pressure; HFNC, high flow nasal cannula; HR, heart rate; IMV, invasive mechanical ventilation; LOS, length of stay; PASS, Pediatric Asthma Severity Score; PICU, pediatric intensive care unit; RR, respiratory rate.
Statistically significant at an alpha level of 0.05.
4. DISCUSSION
In this study we sought to identify distinct clinical phenotypes of severe acute pediatric asthma using machine learning methods. We were able to identify two clinically distinct phenotypes of severe acute pediatric asthma patients based on patient‐centric variables obtained in the first 12 h of hospitalization. We were then able to predict cluster assignment for patients in a separate validation cohort, and these patients exhibited similar clinical characteristics and outcomes. The identification of these clinical phenotypes could be relevant to clinical practice in the future for both prognostic and predictive enrichment.
The initial management of severe acute pediatric asthma includes inhaled bronchodilator therapy and systemic steroids, and while this approach is effective for most children, there are many that do not respond quickly. 21 Recent studies have focused on the identification and understanding of unique subtypes of asthma created through different pathophysiologic processes. 22 , 23 , 24 In this study, we used an unsupervised machine learning method to identify distinct clinical phenotypes in children admitted to the hospital for severe acute asthma. Two phenotypes of severe acute pediatric asthma were identified using this method, and each exhibits a distinct set of clinical characteristics, laboratory markers, and clinical outcomes. Cluster 2 assignment is associated with older age, a higher percentage of lymphocytes and eosinophils, a lower percentage of neutrophils, and a lower magnitude of improvement in vital signs 12‐h from hospital admission. Notably, patients didn't appear to have worse symptoms at presentation, as measured by admission vital signs and initial therapies given in the first 12‐h and received fewer adjunctive therapies (i.e., Magnesium and ipratropium). However, when examining later hospital outcomes and interventions, these patients received an equal amount of noninvasive positive pressure ventilation, and a higher rate of IMV. There are several possible explanations for this, the most clinically interesting would be if these patients potentially exhibit similar or even fewer symptoms of respiratory distress initially, but due to lack of response to standard therapies, their clinical status worsens while the majority of those in Cluster 1 improve. This worsening then leads to increased need for intensive care level interventions, such as IMV, and overall, increased hospital length of stay. It is also interesting that, of the 13 mortalities recorded in this study, all of them were in Cluster 2 patients with no mortalities in the Cluster 1 patients. The lack of responsiveness to clinical therapies, in theory, be linked to distinct pathophysiologic processes in Cluster 2 patients, however, that is outside the scope of this study. It is also possible that the decreased frequency of adjunctive therapies in Cluster 2 contributes to the delay in clinical improvement, however, it is unclear in this case why these patients would receive fewer therapies when treated in the same institution as Cluster 1 patients.
Continued research in the field of pediatric asthma is revealing distinct endotypes based on immunopathology. The ADEPT study 9 , 25 identified the Type‐2 high endotype with overactivation of T‐helper 2 cells leading to elevated IL‐4, IL‐5, IL‐9, and IL‐13, 10 the allergic endotype, 11 the neutrophil‐predominant endotype, 22 and the severe therapy‐resistant endotype. 24 These endotypes have also been identified in patients admitted to the pediatric intensive care unit. 23 The concept of differential response to treatment in pediatric asthma is based on the premise that different biologic processes are occurring in different patients, and addressing the unique mechanisms underlying these processes may improve clinical treatment and outcomes. While the incorporation of laboratory tests in our study limited our sample size, it is vital to include these biologic variables in addition to the clinical variables to attempt to connect the identified phenotypes to underlying biologic endotypes. While it is beyond the scope of the current study to establish direct linkages between the distinct clinical phenotypes identified in this study and specific biologic endotypes, that is the rationale for heterogeneous therapeutic effects. The implication for future research is that it may be possible to identify biologic endotypes of pediatric asthma in the first 12 h of hospitalization using routine clinical data and laboratory tests that would allow bedside clinicians to select treatments that are more effective for specific asthma endotypes based on established mechanistic pathways.
The distinct phenotypes identified using the unsupervised machine learning method were able to be reproduced in a separate validation cohort of severe acute pediatric asthma encounters. The logistic regression model, trained and applied only on patient‐centric data obtained in the first 12 h of hospitalization, was able to classify patients in the validation cohort to the two phenotypes. The overall encounter ratio of approximately 2:1 between the clusters was maintained in the validation cohort, as well as several of the clinical characteristics. These included clinical factors such as age, vital sign responsiveness, and laboratory markers. The length of stay difference and the increased use of IMV were not seen in the validation cohort. The most likely reason these factors were not determined to be statistically significant is likely due to the small sample size in the validation cohort, and potentially would be found statistically significant in a cohort that was equal in size to the derivation cohort. This finding implies that the phenotypes may be generalizable to other cohorts and potentially other institutions, and similar phenotypes could be identified using simple logistic regression with the same independent variables included.
The goal of phenotype identification is to further the practice of precision medicine, to provide the right treatment to the right patient at the right time. 26 Phenotyping seeks to accomplish this through both prognostic enrichment and predictive enrichment. 27 The identification of distinct clinical phenotypes of severe acute pediatric asthma could potentially enable both types of enrichment. For prognostic enrichment, if bedside clinicians could identify Cluster 2 patients 12 h after hospital admission, this could increase the awareness of potential delays in clinical improvement, and prompt clinicians to treat more aggressively with increased adjunctive therapies or earlier increased respiratory support. Predictive enrichment would come later, but it is possible that children in Cluster 2 may respond differently to specific adjunctive asthma therapies, and this knowledge could translate into more specific treatment pathways. This study represents an early step in this direction, but the results will need to be validated in additional data from other pediatric centers.
This study has several limitations. It is a retrospective observational study and thus prone to unmeasured confounding factors. It incorporates data only from a single institution and thus may have institutional biases in the data that would limit generalizability. In an effort to capture relevant biologic variables that could be physiologically important in identifying distinct asthma phenotypes, the study population included only patients admitted to the hospital with laboratory tests obtained in the first 12 h of admission. However, the majority (75% in our study) of severe acute pediatric asthma encounters do not have basic laboratory tests obtained in our institution, which makes our population unique from the general population of severe acute pediatric asthma. Our population likely represents more clinically severe pediatric asthma patients, which may limit generalizability to the entire population. The sample size was small relative to other, similar studies using these types of methods, and thus many potential differences, in the validation cohort especially, may be underpowered to find statistical significance. There was also a large amount of data missingness in the PASS variable, which limits its influence in the study findings.
5. CONCLUSION
We identified two clinical phenotypes of severe acute pediatric asthma using clinical data in the first 12 h of hospitalization, which exhibited distinct clinical features and outcomes. This work has potential for prognostic and predictive enrichment but requires validation in data from other institutions.
AUTHOR CONTRIBUTIONS
Colin Rogerson: Conceptualization; investigation; writing–original draft; methodology; validation; visualization; writing–review and editing; software; formal analysis; project administration; data curation. L. Nelson Sanchez‐Pinto: Conceptualization; investigation; methodology; validation; writing–review and editing; supervision. Benjamin Gaston: Investigation; writing–review and editing; methodology; supervision. Sarah Wiehe: Investigation; writing–review and editing; supervision. Titus Schleyer: Investigation; writing–review and editing; methodology; validation; supervision. Wanzhu Tu: Conceptualization; investigation; writing–review and editing; methodology; validation; project administration; supervision. Eneida Mendonca: Conceptualization; investigation; methodology; validation; visualization; writing–review and editing; formal analysis; project administration; supervision; data curation.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflicts of interest.
Supporting information
Supporting information.
Rogerson C, Nelson Sanchez‐Pinto L, Gaston B, et al. Identification of severe acute pediatric asthma phenotypes using unsupervised machine learning. Pediatr Pulmonol. 2024;59:3313‐3321. 10.1002/ppul.27197
DATA AVAILABILITY STATEMENT
Due to protected health information being contained in the data set used to conduct this study, the data is not freely available for public use. A deidentified version of the data is available upon request from the corresponding author.
REFERENCES
- 1. Rogerson CM, Hogan AH, Waldo B, White BR, Carroll CL, Shein SL. Wide institutional variability in the treatment of pediatric critical asthma: a multicenter retrospective study. Pediatr Crit Care Med. 2024;25(1):37‐46. 10.1097/pcc.0000000000003347 [DOI] [PubMed] [Google Scholar]
- 2. Shah R, Saltoun CA. Chapter 14: acute severe asthma (status asthmaticus). Allergy Asthma Proc. 2012;33(suppl 1):47‐50. 10.2500/aap.2012.33.3547 [DOI] [PubMed] [Google Scholar]
- 3. Carroll CL, Sala KA. Pediatric status asthmaticus. Crit Care Clin. 2013;29(2):153‐166. 10.1016/j.ccc.2012.12.001 [DOI] [PubMed] [Google Scholar]
- 4. Avery C, Perrin EM, Lang JE. Updates to the pediatrics asthma management guidelines. JAMA Pediatr. 2021;175(9):966‐967. 10.1001/jamapediatrics.2021.1494 [DOI] [PubMed] [Google Scholar]
- 5. Rogerson CM, White BR, Smith M, et al Institutional variability in respiratory support use for pediatric critical asthma: a multicenter retrospective study. Ann Am Thorac Soc. 2024;21(4):612‐619. 10.1513/AnnalsATS.202309-807OC [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Russi BW, Roberts AR, Nievas IF, Rogerson CM, Morrison JM, Sochet AA. Noninvasive respiratory support for pediatric critical asthma: a multicenter cohort study. Respir Care. 2024;69(5):534‐540. 10.4187/respcare.11502 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Moore WC, Meyers DA, Wenzel SE, et al. Identification of asthma phenotypes using cluster analysis in the severe asthma research program. Am J Respir Crit Care Med. 2010;181(4):315‐323. 10.1164/rccm.200906-0896OC [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Silkoff PE, Moore WC, Sterk PJ. Three major efforts to phenotype asthma: severe asthma research program, asthma disease endotyping for personalized therapeutics, and unbiased biomarkers for the prediction of respiratory disease outcome. Clin Chest Med. 2019;40(1):13‐28. 10.1016/j.ccm.2018.10.016 [DOI] [PubMed] [Google Scholar]
- 9. Gans MD, Gavrilova T. Understanding the immunology of asthma: pathophysiology, biomarkers, and treatments for asthma endotypes. Paediatr Respir Rev. 2020;36:118‐127. 10.1016/j.prrv.2019.08.002 [DOI] [PubMed] [Google Scholar]
- 10. Kuruvilla ME, Lee FEH, Lee GB. Understanding asthma phenotypes, endotypes, and mechanisms of disease. Clin Rev Allergy Immunol. 2019;56(2):219‐233. 10.1007/s12016-018-8712-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Akar‐Ghibril N, Casale T, Custovic A, Phipatanakul W. Allergic endotypes and phenotypes of asthma. J Allergy Clin Immunol Prac. 2020;8(2):429‐440. 10.1016/j.jaip.2019.11.008 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP. The strengthening the reporting of observational studies in epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370(9596):1453‐1457. 10.1016/s0140-6736(07)61602-x [DOI] [PubMed] [Google Scholar]
- 13. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594. 10.1136/bmj.g7594 [DOI] [PubMed] [Google Scholar]
- 14. Maue DK, Krupp N, Rowan CM. Pediatric asthma severity score is associated with critical care interventions. World J Clin Pediatr. 2017;6(1):34‐39. 10.5409/wjcp.v6.i1.34 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Okada S, Ohzeki M, Taguchi S. Efficient partition of integer optimization problems with one‐hot encoding. Sci Rep. 2019;9(1):13036. 10.1038/s41598-019-49539-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Zurca AD, Suttle ML, October TW. An antiracism approach to conducting, reporting, and evaluating pediatric critical care research. Pediatr Crit Care Med. 2022;23(2):129‐132. 10.1097/pcc.0000000000002869 [DOI] [PubMed] [Google Scholar]
- 17. Urquhart A, Clarke P. US racial/ethnic disparities in childhood asthma emergent health care use: national health interview survey, 2013‐2015. J Asthma. 2020;57(5):510‐520. 10.1080/02770903.2019.1590588 [DOI] [PubMed] [Google Scholar]
- 18. Hughes HK, Matsui EC, Tschudy MM, Pollack CE, Keet CA. Pediatric asthma health disparities: race, hardship, housing, and asthma in a national survey. Acad Pediatr. 2017;17(2):127‐134. 10.1016/j.acap.2016.11.011 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Sammouda R, El‐Zaart A. An optimized approach for prostate image segmentation using K‐Means clustering algorithm with elbow method. Comput Intell Neurosci. 2021;2021:4553832. 10.1155/2021/4553832 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. R: A Language and Environment for Statistical Computing. Version 4.0.3. R Foundation for Statistical Computing; 2022. https://www.R-project.org
- 21. Expert panel report 3 (EPR‐3) : guidelines for the diagnosis and management of asthma‐summary report 2007. J Allergy Clin Immunol. 2007;120(5 suppl):S94‐S138. 10.1016/j.jaci.2007.09.043 [DOI] [PubMed] [Google Scholar]
- 22. Grunwell JR, Stephenson ST, Tirouvanziam R, Brown LAS, Brown MR, Fitzpatrick AM. Children with neutrophil‐predominant severe asthma have proinflammatory neutrophils with enhanced survival and impaired clearance. J Allergy Clin Immunol Prac. Feb 2019;7(2):516‐525.e6. 10.1016/j.jaip.2018.08.024 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Cottrill KA, Rad MG, Ripple MJ, et al. Cluster analysis of plasma cytokines identifies two unique endotypes of children with asthma in the pediatric intensive care unit. Sci Rep. 2023;13(1):3521. 10.1038/s41598-023-30679-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Fainardi V, Esposito S, Chetta A, Pisi G. Asthma phenotypes and endotypes in childhood. Minerva Med. 2022;113(1):94‐105. 10.23736/s0026-4806.21.07332-8 [DOI] [PubMed] [Google Scholar]
- 25. Silkoff PE, Strambu I, Laviolette M, et al. Asthma characteristics and biomarkers from the airways disease endotyping for personalized therapeutics (ADEPT) longitudinal profiling study. Respir Res. 2015;16:142. 10.1186/s12931-015-0299-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. König IR, Fuchs O, Hansen G, von Mutius E, Kopp MV. What is precision medicine? Eur Respir J. 2017;50(4):1700391. 10.1183/13993003.00391-2017 [DOI] [PubMed] [Google Scholar]
- 27. Stanski NL, Wong HR. Prognostic and predictive enrichment in sepsis. Nat Rev Nephrol. 2020;16(1):20‐31. 10.1038/s41581-019-0199-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Supporting information.
Data Availability Statement
Due to protected health information being contained in the data set used to conduct this study, the data is not freely available for public use. A deidentified version of the data is available upon request from the corresponding author.