A computational phenotype for pediatric asthma exacerbations requiring hospitalization using electronic health record data Authors: Colin Rogerson, Danielle Severns, Jason Stemple, Vanessa Monroig, Amy E Hanson, Eneida Mendonca Published: 2026 ABSTRACT AbstractBackgroundComputational phenotypes can be used to improve patient selection for observational studies. We sought to derive and validate a computational phenotype for pediatric asthma exacerbation requiring hospitalization.MethodsRetrospective cohort study using electronic health record (EHR) data from a single quaternary children’s hospital. We used ICD 9 and 10 diagnostic codes and medication data to develop multiple iterations of a computational phenotype. Encounters identified by each phenotype were manually reviewed by expert chart reviewers, and positive predictive value (PPV) was calculated. Sensitivity was obtained by comparison with an established cohort in the Virtual Pediatric Systems (VPS) database.Measurements and main resultsOur cohort included 19 015 pediatric encounters from 2014 to 2022 admitted with an indication of asthma. Starting with 3 broad criteria (a diagnostic code for asthma, receipt of albuterol, and receipt of systemic steroids), we iteratively added phenotype criteria to improve the PPV. The initial phenotype using albuterol and systemic steroids had 65% PPV. Restricting the timeframe for receipt to the first 24 h of the encounter and increasing the amount of albuterol used improved the PPV to 85%. Encounters using only dexamethasone or not having a diagnostic code for asthma were tested and found to have low PPV. The final phenotype included the use of >1 nebulized albuterol treatment or continuous albuterol and the use of systemic steroids in the first 24 h of hospitalization, plus a diagnostic code for asthma, and excluded encounters that only used dexamethasone. This phenotype achieved a PPV of 93%. Using this as a reference metric, defining the cohort based only on diagnostic codes had a PPV of 34%. Applying these criteria to the VPS cohort admitted to the intensive care unit yielded a sensitivity of 97%.ConclusionsWe created a computational phenotype for pediatric asthma exacerbation requiring hospitalization using structured EHR data elements, which achieved a high positive predictive value and sensitivity. This phenotype can be leveraged for future observational studies in this population, but may require institution-specific adjustments. Introduction The widespread use of electronic health records (EHR) enables the collection of granular data that can facilitate clinical research through real-world evidence generation.1 While this has improved the data available for retrospective observational studies across clinical settings, observational research continues to struggle with inherent biases relative to prospective experimental studies, particularly in patient screening and selection.2,3 The use of standardized diagnostic codes, such as the International Classification of Diseases (ICD), has become common for selecting and defining patient cohorts for retrospective observational studies.4,5 However, more recent studies have shown that, when tested for validity, ICD codes alone are often inaccurate in screening of optimal patient populations for observational comparisons.6–9 Computational phenotypes involve a compilation of multiple distinct data elements that can improve the ability to select a valid patient population for retrospective studies.10 These typically involve the use of structured data elements routinely captured in EHR systems in the process of patient care, and have been shown to accurately define patient populations.11–14 The process of creating computational phenotypes and formally testing their validity in unique populations is a step toward using modern informatics tools to improve the quality of observational clinical research. Asthma is the most common chronic disease of childhood, and while most children can be successfully managed in the outpatient setting, children suffering from more severe asthma exacerbations requiring hospitalization are at risk of significant morbidity.15,16 There are no generalized, widely accepted computational phenotypes for pediatric asthma exacerbation requiring hospitalization, and many studies rely solely on diagnostic ICD-10 codes.17 The use of these codes alone may be inaccurate for this population because asthma is a chronic disease that may be added to an encounter but may not be the primary reason for the encounter, and the higher disease acuity can add complexities that are not adequately captured using only ICD-10 codes. The ability to accurately identify this population is imperative to facilitate future research leading to improved clinical outcomes. The objective of this study was to derive and validate a computational phenotype for pediatric asthma exacerbations requiring hospitalization using multiple structured data elements collected in the EHR during routine clinical care. Methods Study population and data elements 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 required or obtained. The study population included all hospital admission encounters between January 1, 2015, and December 31, 2022, for children aged 2-18 years who met at least 1 of the following inclusion criteria. Included encounters must have had either an International Classification of Disease 9th or tenth revision diagnostic code for asthma (ICD9: 493, ICD10: J45 including hierarchical descendants), the receipt of a single dose of nebulized albuterol, or the receipt of an oral or intravenous dose of systemic steroids (prednisone, prednisolone, dexamethasone, or methylprednisolone). Hospital admission time was defined as arrival at the emergency department (ED) for encounters that initially presented to our hospital’s ED, and as arrival to the hospital floor or pediatric intensive care unit (PICU) for those that were directly admitted. Asthma exacerbations are episodes of bronchospasm and dyspnea requiring systemic steroids and beta agonist treatments. Admission criteria for asthma at our hospital include the need for beta agonists more frequently than every 4 h. In our institution, asthma exacerbations requiring advanced respiratory support therapies (ie, high flow nasal cannula, bilevel positive pressure support, or invasive mechanical ventilation) and/or ancillary asthma therapies including aminophylline, terbutaline, or magnesium infusions require PICU admission and were termed critical asthma. Study data were obtained primarily from our institution’s Cerner Millennium Enterprise Data Warehouse. Relevant data tables were queried using Structured Query Language (SQL) scripts specific to the stated inclusion criteria. Data elements obtained included patient demographics (age), diagnostic codes (ICD 9/10), and medications (name, route, and date/time stamp). Timing of medications was measured as the difference between medication order time and hospital admission time in hours. A separate population of PICU encounters was obtained from the Virtual Pediatric Systems (VPS) database.18 These encounters included all PICU encounters over the same time period screened by trained clinical nurse specialists and assigned a primary diagnosis of critical asthma. We chose to use this cohort as a “gold standard” population to measure sensitivity (rather than the initial cohort) because of the larger sample size of clinician-diagnosed hospitalized asthma without need for greatly expanded manual chart reviews. These encounters had the same clinical data obtained from the Cerner Data Warehouse as above. Phenotype iterations and validation Multiple iterations of the computational phenotype were conducted with refinements to improve phenotype performance. We began with a broad approach to maximize sensitivity and iteratively added criteria to the phenotype to improve positive predictive value (PPV). The first phenotype iteration included the use of any nebulized albuterol, and increasingly strict criteria were added systematically. These additional criteria included the timing of the albuterol dose (within 24 h of hospitalization), amount of albuterol (at least 2 intermittent doses or continuous nebulization), the use of systemic steroids (prednisone, prednisolone, methylprednisolone, or dexamethasone) also within 24 h of hospitalization, and a diagnostic code for asthma. We also tested phenotypes that seemed to consistently identify patients that appeared computationally similar to our target population but were not accurate on in-depth chart review. These included patients who received only dexamethasone without other systemic steroids during their hospitalization, and patients who met all other criteria, but did not have a diagnostic code for asthma. Each phenotype iteration had manual chart reviews performed to validate performance, which ranged in volume from 10 to 90, depending on PPV. After each set of 10 charts, we assessed the phenotype’s PPV. If the PPV was >70%, we reviewed another 10 charts and reassessed. The next criterion was a PPV of 90%. If, after 20 chart reviews, the PPV was >90%, we continued to 40 charts, and then to 90 charts if the PPV remained >90%. Charts were selected at random from the list specified by each phenotype iteration. All chart reviews were performed by at least 2 separate board-certified pediatric clinicians with experience in pediatric asthma and included reviews of admission notes and daily progress notes to confirm the primary diagnosis and active problems. Each identified encounter was reviewed by a minimum of 2 reviewers and labeled (Yes/No) as having asthma as a primary reason for hospital admission. The determination of asthma as a primary reason for hospitalization was based on the principal treatments the patient received that required hospitalization, with bronchodilator therapy being the most indicative of asthma as the primary reason for hospitalization. Reviewers conducted reviews independently and asynchronously and were blinded to the results of the other reviewer. All discrepancies were adjudicated by a third clinician reviewer. Positive predictive value (PPV) was determined for each phenotype iteration by dividing the number of true positive cases determined by expert chart review by the number of phenotype-identified positive cases. Sensitivity was measured using the VPS identified cohort. Sensitivity was measured as the number of encounters meeting phenotype criteria in this cohort divided by the total number of total encounters in the cohort. All data management and statistics were done using R version 4.2.3.19 Results Initial iterations The initial cohort included 19 015 inpatient encounters for children aged 2-18 years with any of the asthma inclusion criteria (Table 1). The broadest phenotype criteria were a diagnostic code for asthma (n = 8286), receipt of any albuterol (n = 13 699), or receipt of any systemic steroids (n = 8703) (Figure 1). The first computational phenotype created and validated was encounters with any albuterol AND any systemic steroid use at ANY point during the hospital encounter, with no requirement for a diagnostic code (Table 2). This phenotype identified 6301 encounters. For this phenotype, 20 encounters were randomly selected and reviewed, and 13 were determined positive (PPV 65.0%) with 95% agreement between the 2 initial reviewers. The next phenotype iteration added a time criterion to the medication receipts (albuterol and systemic steroids) within the first 24 h of the hospital encounter and required receipt of either 2 or more doses of intermittent albuterol, or the receipt of continuous nebulized albuterol. Of the 4177 encounters identified by this phenotype, 40 encounters were randomly selected and reviewed, and 35 were determined positive (PPV 87.5%) (Figure 2) with 90% agreement between the initial 2 reviewers. Table 1Phenotype iterations.1. Any albuterol and systemic steroid during encounter2. Receipt of either 2 intermittent albuterol doses or continuous albuterol and a systemic steroid in the first 24 h of encounter3. Final phenotype Receipt of either 2 or more intermittent albuterol doses or continuous albuterol in the first 24 h of the encounterReceipt of a systemic steroid in the first 24 h of the encounterReceipt of a systemic steroid other than dexamethasone during the encounterDiagnostic code for asthma present Phenotype iterations. Receipt of either 2 or more intermittent albuterol doses or continuous albuterol in the first 24 h of the encounter Receipt of either 2 or more intermittent albuterol doses or continuous albuterol in the first 24 h of the encounter Receipt of a systemic steroid in the first 24 h of the encounter Receipt of a systemic steroid in the first 24 h of the encounter Receipt of a systemic steroid other than dexamethasone during the encounter Receipt of a systemic steroid other than dexamethasone during the encounter Diagnostic code for asthma present Diagnostic code for asthma present Venn diagram illustrating intersections between each of the initial 3 broad phenotype inclusion criteria: Any receipt of albuterol, any receipt of systemic corticosteroids, and a diagnostic code for asthma. Bar graph illustrating positive predictive value for identifying asthma exacerbations requiring hospitalization of each phenotype iteration. Table 2Initial Cohort Characteristics.CharacteristicCohort (n=19,015) Male10 467 (55.0)Age (years)9 (5-14)Race White13 155 (69.2) Black5 098 (26.8) Other762 (4.0)Hispanic2 008 (10.6)Inclusion Criteria Diagnostic Code8 286 (43.6) Received Albuterol13 699 (72.0) Received Steroid8 703 (45.8)Continuous variables measured as medians with interquartile range Categorical variables measured as frequency with percentage. Initial Cohort Characteristics. Continuous variables measured as medians with interquartile range Categorical variables measured as frequency with percentage. Negative phenotypes Several themes emerged from the first 60 chart reviews, and the feedback from reviewers was used to identify specific negative phenotypes that were decreasing the PPV of the phenotype. The first negative phenotype tested included encounters that met the criteria from the previous iteration (≥2 albuterol doses and systemic steroids in the first 24 h) in which the only systemic steroid received was dexamethasone. For this phenotype, we identified 156 encounters, randomly selected and reviewed 10 charts, and only 1 was determined positive (PPV 10.0%). The second negative phenotype tested included encounters that met the medication criteria used previously, but did not have a diagnostic code for asthma. With this phenotype, we identified 454 encounters, randomly selected and reviewed 10, and determined 3 to be positive (PPV 30.0%). Final phenotype The information gained from each of the previous phenotypes, positive and negative, was compiled to create the final computational phenotype. This phenotype included receipt of at least 2 doses of intermittent nebulized albuterol OR continuous albuterol in the first 24 h of the hospital encounter, the receipt of a dose of systemic steroids (dexamethasone [cannot be the only steroid used in the hospital encounter], prednisone, prednisolone, or methylprednisolone) in the first 24 h of the hospital encounter, and the presence of a diagnostic code for asthma. For this phenotype, we identified 2828 encounters, randomly selected and reviewed 90 charts, and determined 84 to be positive (PPV 93.3%) with 92% agreement between the 2 initial reviewers. Comparing the initial, broad criteria phenotypes: a diagnostic code for asthma alone resulted in a PPV of 34.1%, a single dose of albuterol at any time had a PPV of 20.6%, and a dose of steroids achieved a PPV of 32.5%. The encounters that met all criteria for the final phenotype but were determined not to be admitted to the hospital for asthma were primarily children with multiple complex chronic conditions who were treated for acute bronchospasm at some point during their encounter but were admitted for another reason. Assessment of sensitivity We identified 1575 PICU encounters in the VPS database with a primary admission diagnosis of critical asthma. The initial phenotype of ANY receipt of albuterol and systemic steroids identified 1546 encounters, yielding a sensitivity of 98.1% (Figure 3). The second phenotype of continuous albuterol or ≥2 doses of intermittent albuterol and systemic steroids in the first 24 h of the hospital encounter identified 1534 encounters, yielding a sensitivity of 97.4%. The final phenotype described above identified 1520 encounters, yielding a sensitivity of 96.5% for identifying children admitted to the PICU for asthma. Bar graph illustrating sensitivity for identifying asthma exacerbations requiring admission to the pediatric intensive care unit of each phenotype iteration. Discussion In this study, we created a computational phenotype for pediatric asthma exacerbation requiring hospitalization using structured EHR data elements obtained in the routine process of clinical care. We used an iterative approach that helped to refine and improve the phenotype criteria to improve performance. The final phenotype achieved a high PPV in identifying children admitted to the hospital for asthma and sensitivity for identifying children admitted to the PICU for critical asthma determined by a large number of manual expert chart reviews and a gold standard PICU population. The high PPV achieved by this phenotype makes it ideal for retrospective observational clinical research to more accurately determine study inclusion with a high level of confidence. Accurate patient selection is 1 of the most important features of a valid clinical study to avoid biased results. Many previous studies in pediatric asthma have relied on diagnostic codes alone to identify patients for study inclusion.20,21 Afshar et al derived and validated a computational phenotype for pediatric asthma which relied mostly on ICD codes, but also included medication data. While they achieved high performance metrics (PPV of 94.2% and Sensitivity of 98.7%), their phenotype required data from multiple clinical encounters over a large period. While useful for identifying asthma in a longitudinal cohort, many children receive their initial diagnosis after being hospitalized with no previous symptoms, and the need for data from multiple longitudinal encounters limits its utility in the inpatient setting.22 Our study found that, for hospitalized children, using a diagnostic code alone for study inclusion had a PPV of 34.1%. The incremental inclusion of more criteria that were clinically relevant to the target population resulted in consistent increases in phenotype performance, and the final phenotype achieved a strong PPV of 93.3%. The encounters misclassified by our final phenotype included mostly children with bronchospasm as a component of their complex chronic diseases who were hospitalized primarily for other diagnoses. This population is difficult to phenotype due to the co-occurrence of multiple disease states and presents common challenges for observational researchers. Further refinement of the phenotype could potentially improve the phenotype’s performance in this population, but it may come at the expense of excluding children with chronic diseases who are admitted with asthma. This study demonstrates that a more sophisticated approach significantly outperforms the broad criteria approaches; and this phenotype could be used to improve patient selection for future studies focused on pediatric asthma exacerbations requiring hospitalization. The use of a different cohort for the measurement of sensitivity is an important consideration in our study. The ideal scenario would include a large, validated, gold standard cohort of children admitted to the hospital for asthma that could be used to derive and validate a computational phenotype and measure both PPV and sensitivity. While it is more straightforward to evaluate a new phenotype’s PPV by testing the cohort identified by the phenotype via manual chart review, it is more difficult to identify which encounters are not being identified and to measure sensitivity. This often requires large-volume chart reviews by multiple investigators to identify encounters the phenotype is missing. Our study team was limited in this regard and thus chose to use the VPS identified cohort as a validated cohort to assess sensitivity. While this was an efficient approach, it does add bias as we cannot determine the exact percentage of patients admitted to the general floor that the phenotype is missing. Thus, while the PPV can be interpreted in the context of all asthma admissions, the measured sensitivity can only be interpreted in the context of children admitted to the PICU. An important feature of computational phenotypes is the availability of the specific data elements included in the phenotype. When developing our phenotypes, we were strategic in ensuring that incorporated data elements were elements already being routinely captured in a structured format during the process of clinical care. Using only these types of data elements, we have shown that this phenotype could be used in future studies with access limited only to medication and diagnostic code data, making it a more generalizable phenotype. Other data elements, such as unstructured clinical notes, could potentially provide additional data that may improve phenotype performance. However, the need for advanced natural language processing methods and access to clinical notes provide significant barriers to many retrospective studies. Computational phenotypes may require adjustment in varying clinical contexts or based on differing institutional practices. For our purposes, we created this phenotype to study children hospitalized at our institution for asthma. Our focus is on clinical responsiveness to interventions such as medications and respiratory support. Thus, our phenotype seeks to identify the most ill children with asthma. Other studies examining children with less severe asthma attacks that may be discharged from the emergency department, outpatient asthma, or children admitted for other reasons who experience an asthma exacerbation after hospitalization would likely need to adjust the phenotype to optimize its use for these purposes. Our phenotype is also somewhat specific to our institution. Our institutional practice is to treat initially with nebulized albuterol shortly after hospital presentation, as well as systemic steroids. While studies have shown that dexamethasone is equally effective in treating asthma exacerbations,23 our institution’s typical practice is to use prednisone, prednisolone, or methylprednisolone in children requiring hospital admission. Thus, these specific criteria may not perform as well in other institutions with different clinical practices and would need specification. One final consideration is that diagnostic codes are often assigned to encounters at discharge, thus limiting the timeframe of use for the phenotype. If a study aimed to identify patients soon after admission for prospective clinical phenotyping or an interventional study, the use of the diagnostic code in the phenotype would be prohibitive. This study has several strengths and limitations. We used a team of expert pediatric clinicians with extensive experience in asthma to validate each phenotype and used several iterations to refine the phenotype into a highly predictive product. Study limitations include the retrospective nature of using EHR data, which introduces missingness and historical bias. This phenotype was also developed and validated in a single institution, and both nuances in EHR data capture and clinical practices may differ in other sites which may bias the model and produce less robust precision and recall when applied at other centers. Further research is needed to validate this phenotype in data from other institutions. Clinical practices may have changed over the time period, which may also add bias to the results of the phenotype. Our sensitivity analysis was performed in a PICU population, which may introduce bias that impairs sensitivity of the phenotype in the population of asthma admitted to the general floor. Investigating the negative phenotypes showed that 30% of encounters that lacked a diagnostic code but met all other criteria were determined to be admitted primarily for asthma, which shows limitations in our phenotype’s sensitivity. There is some inherent subjectivity in the determination of the primary reason for hospital admission, which may vary between reviewers and add bias. Although our hospital system uses the same albuterol protocol for asthma both in the PICU and on the floor, differences in illness severity and steroid use may bias the sensitivity of the phenotype for floor patients. Lastly, our phenotype was created in a Cerner-based EHR system and may have some nuances limiting its deployment in other EHR systems. Conclusion We created a computational phenotype for pediatric asthma exacerbation requiring hospitalization using structured EHR data elements, achieving a high positive predictive value and sensitivity. This phenotype can be leveraged for future observational studies in this population but may require institution-specific adjustments.