Abstract
Introduction
Although clinical, functional, and biomarker data predict asthma exacerbations, newer approaches providing high accuracy of prognosis are needed for real-world decision-making in asthma. Machine learning (ML) leverages mathematical and statistical methods to detect patterns for future disease events across large datasets from electronic health records (EHR). This study conducted training and fine-tuning of ML algorithms for the real-world prediction of asthma exacerbations in patients with physician-diagnosed asthma.
Methods
Adults with ≥ 2 ICD9/10 asthma codes within 1 year and at least 30 days apart were identified from the Optum Panther EHR database between 2016 and 2023. An emergency department (ED), urgent care, or inpatient visit for asthma, while on systemic administration of corticosteroids, was considered an exacerbation. To predict factors associated with exacerbations in a 6-month study period, clinical information from patients was retrieved in the preceding 6-month baseline period. Clinical information included demographics, lab results, diagnoses, medications, immunizations, and allergies. Three models built using Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Transformers algorithms were trained and tested on independent datasets. Predictions were explained using the SHAP (SHapley Additive exPlanations) library.
Results
Of 1,331,934 patients with asthma, 16,279 (1.2%) experienced ≥ 1 exacerbation. XGBoost was the best predictive algorithm (area under the curve [AUC] = 0.964). Factors associated with exacerbations included a prior history of exacerbation, prednisone usage, high-dose albuterol usage, and elevated troponin I. Reduced probability of exacerbations was associated with receiving inhaled albuterol, vitamins, aspirin, statins, furosemide, and influenza vaccination.
Conclusion
This ML-based study on asthma in the real world confirmed previously known features associated with increased exacerbation risk for asthma, while uncovering not entirely understood features associated with reduced risk of asthma exacerbations. These findings are hypothesis-generating and should contribute to ongoing discussion of the strengths and limitations of ML and other supervised learning models in patient risk stratification.
Supplementary Information
The online version contains supplementary material available at 10.1007/s12325-024-03053-y.
Keywords: Asthma, Electronic health records, Exacerbations, Machine learning, Physician-diagnosed asthma, Real-world prediction, XGBoost
Key Summary Points
| Why carry out this study? |
| Machine learning (ML) and artificial intelligence have the potential to augment clinical decision-making and advance personalized medicine for patients with asthma. |
| In this study, we applied ML algorithms to a relatively large dataset of patients with asthma whose data were extracted from their electronic health records (EHR), with the aim of identifying features of patients with asthma that were associated with exacerbations. |
| What was learned from the study? |
| Factors associated with exacerbations included a prior history of exacerbation, prednisone usage, high-dose albuterol usage, and elevated troponin I. Reduced probability of exacerbations was associated with receiving inhaled albuterol, vitamins, aspirin, statins, furosemide, and influenza vaccination. |
| This study on asthma in the real world confirmed previously known features associated with increased exacerbation risk, while uncovering not entirely understood features associated with reduced risk of exacerbations. |
| These findings suggest that ML algorithms can be used on electronic health records to assess features of asthma associated with future risk of exacerbations. Patients should be informed if such methods are used to augment clinical decision-making and establish prognosis. |
| These results are hypothesis-generating and are not meant to guide clinical decision-making or treatment selection for prevention of asthma exacerbations. Future research will focus on uncovering the causal relationships among the most predictive features identified in our study. |
Introduction
Automation in medicine seems to be an unavoidable trend that, if well implemented, should benefit patients and the healthcare system overall [1]. The direct benefits for patients include augmentation of clinical decision-making by their treating physician which should lead to better prognosis [2].
Predicting asthma exacerbations continues to be a challenge with one exacerbation being treated every 3 s in the USA [3]. Having a history of exacerbations is one well-known feature associated with increased risk of having future exacerbations. More recently, biomarkers of type 2 inflammation such as elevated blood eosinophils and/or fractional exhaled nitric oxide (FeNO) can also identify patients who are at increased risk of future exacerbations [4]. These biomarkers, however, only denote type 2 inflammation that is predominant in many, but not all, patients with asthma [5].
In the past decade several groups have shared their attempts in predicting asthma exacerbation using different approaches and data types. For a more detailed description of those studies, we invite readers to peruse our recent review paper [3]. Herein, we will briefly mention those that are most relevant to this work. Tong et al. developed an XGBoost-based ML model to predict asthma exacerbations in the following 12 months [6]. The model delivered an AUC of 0.9 mainly by using features related to loss of asthma control and race. The dataset used for building and testing the model contained 82,888 data instances. Zein et al. analyzed data extracted from electronic health records (EHRs) of 60,302 patients with asthma and built a light gradient boosting machine model able to predict asthma exacerbation with AUC 0.88 [7]. Authors identified age, use of long-acting β-agonists, high-dose inhaled glucocorticoids and oral glucocorticoids as the most important predictors of future exacerbation. Goto et al. tested several supervised ML models on data from a small dataset 3206 adult patients with asthma [8]. The best models at predicting asthma and chronic obstructive pulmonary disease (COPD) exacerbations were tree-based.
Because most of the aforementioned data come from clinical studies in patients with severe asthma, we wished to assess what additional features might be important in predicting asthma exacerbations in the general population with asthma who suffer largely from mild forms of the condition [9]. To this end, we developed and evaluated ML algorithms that predict asthma attacks on a relatively large dataset of patients with asthma whose data were extracted from their EHR in the Optum Panther Database. Our main objective was to find out what features (inputs) were associated with asthma exacerbations (the output or label). If new, unknown, or poorly understood features were to be found, we would further our understanding and improve our ability to predict asthma exacerbations.
Methods
Data Preparation
Data from adult patients with asthma were extracted from the Optum Panther EHR database from January 1, 2016 to June 30, 2023. Patients with at least two ICD9/10 asthma codes (493.90, 49392, 49300, 49381, 49302, 49310, 493, 493.91, 493.82, 493.9, 493.12, 493.01, 493.11, 493.0, 493.1, 518.3, J45, J45.2, J45.20, J45.21, J45.22, J45.3, J45.30, J45.31, J45.32, J45.4, J45.40, J45.41, J45.42, J4.55, J45.50, J45.51, J45.52, J45.90, J45.901, J45.902, J45.909, J45.990, J45.991, J45.998, J82.83, J82, J67.8) within 1 year but at least 30 days apart were selected, and their information was collected. An exacerbation was defined as an ER or urgent care visit or inpatient visit for asthma which required systemic administration of corticosteroids [10]. If multiple exacerbation events occurred within 1 week, only the first exacerbation event was considered.
Anchor dates are dates in the patient clinical history and were used to establish baselines and prediction periods. For patients who experienced at least one exacerbation, the anchor date (exacerbation anchor date) was the date when the exacerbation occurred (one anchor date for each exacerbation). Meanwhile, for patients who never experienced an exacerbation, the anchor date (non-exacerbation anchor date) was a random date when a code for asthma was recorded. Once the anchor dates were established, we applied sliding non-overlapping windows to create baseline periods (Fig. 1). Each exacerbation anchor date could have up to four baseline windows: the two most recent ones were used to predict an exacerbation, while the two oldest ones were used to predict a non-exacerbation event. Non-exacerbation anchor dates could have up to two baseline windows, both used to predict a non-exacerbation event. Each baseline window spanned 6 months.
Fig. 1.
Sliding window algorithm schematic. An anchor date is a date when an exacerbation or a non-exacerbation event (random asthma code) occurred. When the anchor date corresponded to an exacerbation, up to four 6-month non-overlapping intervals were drawn. Of those four intervals, the data collected within the two most recent ones were used to predict an asthma exacerbation in the following 6 months, and the data collected within the two oldest ones were used to predict the lack of an exacerbation. Meanwhile, when the anchor date corresponded to a non-exacerbation, up to two 6-month non-overlapping intervals were drawn, and data collected in both were used to predict the lack of exacerbation. Gray intervals indicate intervals that were discarded according to the criteria described in the text
Baseline periods were excluded if the patient was ≤ 18 years of age at any point during that period, or if the patient first appeared in the Optum database after the period began (incomplete coverage), or if there were no valid diagnosis codes (patients never interacted with the healthcare system and thus no data were available for prediction). Furthermore, for each patient, non-exacerbation windows were removed if they overlapped with an exacerbation window.
Supplemental Fig. 1 depicts the flowchart for the exclusion/inclusion process, anchor date identification, and baseline period filtering.
Optum EHR is a commercial database of electronic health records (EHR). This study complied with all applicable laws regarding patient privacy, using Health Insurance Portability and Accountability Act (HIPAA)-compliant de-identified retrospective data sources, and permission was not required. No direct patient contact or primary collection of individual patient data occurred.
To predict factors associated with asthma exacerbations in a 6-month study period, clinical information from patients was retrieved in the preceding 6-month baseline period. Clinical information, including demographics, lab results, diagnoses, medications, observations, immunizations and allergies, was extracted from Optum EHR.
Demographics
Demographic patient information included gender, ethnicity, race, year of birth. Patients with missing birth year were excluded. Missing values for gender, ethnicity and race were filled with the “unknown” string.
Diagnosis
Diagnosis included diagnosis code (i.e., J45), diagnosis code type (i.e., ICD10), diagnosis date, and interaction type (i.e., ED, ambulatory visit, etc.). Only diagnoses with ICD9 and ICD10 codes were considered.
Laboratory Results
Laboratory tests were performed by different entities and used different measurement units and normal ranges. To facilitate our analysis, we utilized the “in_range” flag that indicated whether the lab result was within the normal range or not. The in_range flag was generated by the laboratory professionals. We further labelled immunoglobulin E (IgE) and eosinophil lab tests as normal or abnormal on the basis of custom thresholds. These thresholds were 148 kU/L for IgE, 150 and 300 cells/µL for absolute blood eosinophils, and 2% and 4% relative blood eosinophils [11]. While binarizing predictors can result in a significant loss of information, it was necessary to enable their processing alongside other lab results without introducing additional LSTM or Transformer blocks to the model.
Observations
Observations included clinical features such as body mass index (BMI), heart rate, blood pressure, and temperature. Each observation was labelled as Y and N depending on the result was within or outside the normal range (in range flag), respectively.
Medications
Optum provides three sets of medications data: prescriptions, medications administered to the patient during the episode of care, and medications the patient told the healthcare professional (HCP) they were taking at the time of check-in or hospitalization. To limit sparsity of the dataset, we aggregated the National Drug Code (NDC) at common drug name level using a mapping table we curated.
Immunizations
The vaccine table contained NDC of the immunization and immunization date. We converted NDC of each vaccine to the vaccine common name to reduce sparsity of the data.
Allergies
The allergies table contained allergy name and onset date. We utilized the common name of the allergen to represent patient allergies. Allergies were treated as a demographic feature: if an allergy developed before the end of the window, it was included as predictors of the event with which that window was associated.
Except for demographics, allergies, and immunizations, every other data type was represented as a time series, and for each datapoint we calculated the distance in days between the datapoint time stamp and the anchor date.
Models and Models Trained
We built and trained three different ML models based on Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Transformers algorithms. Additional details are provided in the Supplemental Methods.
Diagnosis, observations, medications (prescription, medication administered to the patient, medications the patient was taking), and labs were passed to the Transformers and LSTM model as a time series, tokenized, and embedded as vectors. The length of each series was trimmed at the 99th percentile. This step was needed to ignore and clip very long and abnormal sequences of datapoints associated with patients that were not representative of the population with asthma. Clipping sequences led to faster training and lower memory and computation usage without sacrificing model performance. Allergies and immunization data were tokenized and embedded as vectors. Patient demographics was one-hot encoded. Figure 2 shows a high-level overview of the architecture of LSTM and Transformer models.
Fig. 2.
Architecture of the LSTM (A) and Transformer models (B). In the proposed architectures, each data type is processed independently by either LSTM or Transformer blocks. The result of each block is then aggregated by a common dense layer before being passed to the final classification layer. DEM demographic data; DIAG CODES diagnostic codes; EMBED embedding neural network layer; ENC TYPE how the patient interacted with the healthcare system (e.g., emergency visit, phone call, etc.); IMM immunization; IN RANGE (Y/N) flag indicating whether lab results are within the normal range, Yes/No; LABS NAME-IN RANGE (Y/N) Lab tests and their IN RANGE flag are concatenated together; LSTM long short-term memory; MED ADMIN medications administered to patients during the episode of care; MED TAKEN medications the patients told their healthcare provider they were taking; NN neural network; OBS CODES clinical observations; OBS CODES-IN RANGE (Y/N) Observations and their IN-RANGE flag are concatenated together; PRESCR prescriptions given to patients
For the XGBoost algorithm, variable temporal information was discarded, and codes were de-duplicated and one-hot encoded. For labs and observations, the in-range value was concatenated to the name of the lab or observation, respectively; meanwhile diagnosis codes and their encounter types were concatenated together.
Results
Cohort Description
Optum EHR contains clinical data of more than 100 million patients in the USA. We identified 1,331,934 patients with asthma whose data were analyzed between January 2016 and June 2023. Of these, 16,279 (1.2%) patients had experienced ≥ 1 exacerbation with a mean number of exacerbations per patient of 1.65. Demographic information is summarized in Table 1. African Americans and women exacerbated more frequently than individuals of other races and men, respectively. There were no differences in the use of oral and inhaled corticosteroids between non-exacerbators and exacerbators.
Table 1.
Cohort demographics and baseline characteristics
| All patients (1,331,934) | Never exacerbated (1,315,655) | Exacerbated (16,279) | |
|---|---|---|---|
| Age (mean) | 49 years | 49 years | 42 years |
| Gender | |||
| Male | 401,924 | 397,000 | 4924 |
| Female | 929,355 | 918,003 | 11,352 |
| Unknown | 655 | 652 | 3 |
| Race | |||
| White | 982,116 | 973,131 | 8985 |
| African American | 231,517 | 225,164 | 6353 |
| Asian | 20,751 | 20,642 | 109 |
| Other/unknown | 97,550 | 96,718 | 832 |
| Ethnicity | |||
| Non-Hispanic | 1,130,898 | 1,116,633 | 14,265 |
| Hispanic | 88,597 | 86,896 | 1701 |
| Other/unknown | 112,439 | 112,126 | 313 |
| CS | |||
| Oral | 302,529 | 299,035 | 3494 |
| Inhaled | 431,132 | 426,195 | 4937 |
Asthma Exacerbation Prediction
Herein we trained three different supervised models for the prediction of asthma exacerbation in patients with asthma: XGBoost, LSTM, and Transformers algorithms, which were developed and trained and later tested on a test dataset.
The prediction performances on the test set of each model are described in Table 2. XGBoost showed the best receiver operating characteristics and best precision-recall (PR) curves. Both plots show the performance of a classification model at all classification thresholds, but the latter factors in the class imbalance (Fig. 3). As, expected, because of the imbalanced nature of the data, the ROC-AUC is much higher than the PR-AUC, for all three models. While all three algorithms were able to accurately classify most of the events (accuracy closer to 1), only XGBoost could detect more than 50% of the exacerbations.
Table 2.
Prediction performance of each model
| XGBoost | LSTM | Transformer | |
|---|---|---|---|
| AUC | 0.964 | 0.904 | 0.910 |
| PR-AUC | 0.647 | 0.430 | 0.419 |
| Precision | 0.729 | 0.521 | 0.524 |
| Recall | 0.529 | 0.424 | 0.380 |
| F1-score | 0.613 | 0.468 | 0.441 |
| Accuracy | 0.988 | 0.983 | 0.983 |
| NPV | 0.991 | 0.989 | 0.989 |
Fig. 3.
Classification performance graphs for XGBoost. This figure illustrates the performance evaluation of the XGBoost model through the receiver operating characteristic (ROC) curve (left) and the precision-recall (PR-ROC) curve (right). AUC area under the curve; LSTM Long Short-Term Memory; XgBoost Extreme Gradient Boosting
Model Prediction Explanation with SHAP
Predictions from the XGBoost model were explained using the SHAP library [12]. Features of asthma associated with exacerbations included having a medical history of exacerbation in the previous 6 months, having received systemic corticosteroids, high doses of inhaled albuterol (0.83 mg), and abnormal levels of troponin I (Figs. 4, 5, 6). Features associated with no exacerbations were increased age, chronic use of any of the following: aspirin, atorvastatin, multivitamins (multivitamin oral tablet), vitamin D (cholecalciferol 0.025 oral tablets), furosemide, or low dose of albuterol (0.09 mg) (Figs. 4, 5, 6). African American and Hispanic ethnicities were risk factors for experiencing asthma exacerbations (Fig. 4).
Fig. 4.
SHAP values for age, race, and ethnicity. A SHAP values decreased with age suggesting that older patients have lower risk of experiencing asthma exacerbation. B, C African American and Hispanic ethnicities were risk factors for experiencing asthma exacerbations denoted by the positive values in the Y axes. Hisp Hispanic; SHAP SHapley Additive exPlanations
Fig. 5.
SHAP values for troponin I lab test, atorvastatin, and high- and low-dose albuterol medications. A A positive test for troponin I is associated with higher risk of asthma exacerbation. B–D Patients taking atorvastatin or low-dose albuterol were less likely to report an asthmatic attack in the following 6 months. Patients taking high-dose albuterol were positively associated with future asthmatic events. Lab laboratory test result; Med patients were taking or not the medication during the check-in or hospitalization; SHAP SHapley Additive exPlanations
Fig. 6.
SHAP values for eosinophil lab results and furosemide, aspirin, multivitamin and cholecalciferol medications. A–D Patients who reported taking furosemide, aspirin, cholecalciferol, or multivitamin were less likely to experience asthmatic attacks in the following 6 months. E Elevated eosinophil blood counts in the blood were an indicator of future exacerbations (i.e., SHAP values are higher). Abs absolute count; Lab laboratory test result; Med patients were taking or not the medication during the check-in or hospitalization; SHAP SHapley Additive exPlanations
Discussion
In this study, we trained three ML algorithms capable of predicting adult asthma exacerbations significantly better than those reported in existing literature. While direct performance comparisons are not advisable because of differences in training and testing datasets, asthma definitions, exacerbation criteria, and prediction windows varying across studies, we speculate that the superior performance of our models may be attributed to the size of our dataset, which is approximately 22 times larger than the dataset used by Zein et al., for example.
We have identified several features of asthma that are already known risks factors for exacerbations. We have also identified features that are less well understood and may be either causal or confounding. In this regard, we have used SHAP analysis which is a strong way to explore the causality and contribution that a particular feature plays in a ML model. The goal of this study was to deepen our understanding of asthma exacerbations, and all our findings are associative rather than definitive. These findings are hypothesis-generating and should contribute to ongoing discussion of the strengths and limitations of ML and other supervised learning models in patient risk stratification.
The major limitation of this study is the interpretability of the findings, largely due to the study design and innovative yet still exploratory ML methodology. One limitation of the study design is that it focuses only on severe asthma exacerbations and treats patients with one exacerbation the same as patients who experience multiple exacerbations. In addition, the database used does not include the population of patients with asthma who may show other features of their disease but attend healthcare systems who do not use Optum. Unfortunately, our EHR dataset was not able to evaluate some well-known factors previously associated with asthma exacerbations, such as level of type 2 inflammation, blood eosinophils or FeNO, asthma severity, smoking status, and exposure to pollution or other environmental triggers, as these features were either not available or were excluded because of the amount of missing data. Because data pre-processing was tailored for the Optum-EHR database, our results might not be generalizable to other EHR/EMR data. Finally, all analyses did not include causal analyses, and results may be spurious or the result of multicausal associations.
In agreement with literature findings, African Americans exacerbated more frequently than individuals of other races. This observation may be partially related to differential socioeconomic status, healthcare access and quality, and the physical environment and is not limited to any single race, ethnicity, or country of residence.
The risk of asthma exacerbations is increased in patients with previous exacerbations and particularly in patients with high type 2 asthma [4, 5]. In our study, we found that the use of systemic corticosteroids was associated with exacerbations but also with reduction of exacerbations (Figs. 4, 5, 6). This is not surprising, as bursts of systemic corticosteroids are prescribed not only to control asthma symptoms to avert exacerbations but also during treatment of such events. Thus, bursts of corticosteroids can be associated with reductions of exacerbations if administered on the verge of an acute event and they provided effective treatment, or they could be associated with the occurrence of exacerbations if administered once the event has developed and required systemic corticosteroids. Similarly, the use of high doses of inhaled albuterol associated with increased number of exacerbations is a well-known risk factor particularly in the absence of steroids as β-2 agonists lack anti-inflammatory effects and can induce tolerance and overreliance while masking the severity of asthma [13]. However, high doses of inhaled β-2 agonists are indicated in controlled settings to treat acute attacks to reduce airflow deterioration in the ED. Thus, together with systemic corticosteroids, inhaled β-2 agonists can be temporally associated with worsening of exacerbations if used prior to seeking medical help, or with reductions of exacerbations if used to treat airflow limitation in the hospital setting. In addition, as previously reported in our broad study population of patients with asthma, high levels of peripheral blood eosinophils were a feature associated with exacerbations [12] (Fig. 6).
In this study, we also found that high levels of troponin I were associated with asthma exacerbations. High levels of troponin I have been previously described in association with COPD exacerbations, and reports exist of increased levels also in asthma attacks due to eosinophilic infiltration—particularly in EGPA (eosinophilic granulomatosis with polyangiitis) [14, 15]. High doses of inhaled β-2 agonists or hypoxemia during an asthma attack do increase myocardial damage also and may be responsible for the increase of troponin I in our study [16, 17]. Although it is possible that some patients included in this study may have had comorbid COPD, our methodology only included exacerbations in patients with a primary diagnosis of asthma when the exacerbation was recorded. The question that remains unanswered in our analysis is if troponin I levels are increased prior to exacerbations constituting a predictive feature or whether such high levels reflect the severity of the attack and may constitute a marker of poor long-term survival in patients with asthma, as reported in COPD [18].
The reduction of exacerbations associated with increasing age (Fig. 4) that we observed in our study may not be a uniform trend and the relatively low incidence we found may be caused by the chronic use of low-dose aspirin or atorvastatin. In this regard, although in 2–23% of adults with asthma both aspirin and non-steroidal anti-inflammatory drugs can provoke exacerbations, the chronic use of low-dose aspirin has been reported in association with lower incidence of asthma in previous studies [19–21]. Similarly, the use of statins has been reported in association with reduced incidence of asthma exacerbations [22, 23].
Another finding of our present study is the association with the prescription of furosemide and reduced incidence of asthma exacerbations. Indeed, use of furosemide has been reported to reduce the severity of exacerbations when added to usual acute care in the ED by providing additional relief of bronchospasm as measured by improvements of forced expiratory volume in the first second in several controlled studies, and furosemide is still in use in some ED protocols [24, 25]. Yet other studies have shown no benefits in the treatment of asthma exacerbations [26]. Whether chronic use of aspirin, statins, or furosemide reduces the incidence of exacerbations requires further studies.
The association with intake of multivitamins is difficult to interpret owing to the varied formulations available. Lastly, authors have observed that intake of vitamin D can be associated with reduced incidence of asthma diagnosis [27].
The goal of this study was to deepen our understanding of asthma exacerbations, and as mentioned earlier, all our findings are associative rather than definitive. These findings are hypothesis-generating and should contribute to ongoing discussion of the strengths and limitations of ML and other supervised learning models in patient risk stratification. The results are not meant to guide clinical decision-making or treatment selection for prevention of asthma exacerbations. Future research will focus on uncovering the causal relationships among the most predictive features identified in our study.
Conclusion
Our study using ML algorithms in a broad population of patients with asthma confirms the association of previously known features that increase risks of exacerbations or reduce the incidence of such events. Some of these features we describe can be causal or confounding, and further studies are required to dissect the strength and nature of such associations.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
Medical Writing/Editorial Assistance
Editorial review of this manuscript was provided by Kate Smigiel, PhD, of Amgen Inc.
Author Contributions
Nestor A. Molfino developed the concept. Gianluca Turcatel and Yi Xiao ran the analyses. Gilles Gnacadja, Julie Kim and Scott Caveney contributed to interpretation of results. Gianluca Turcatel and Nestor A. Molfino wrote the initial paper. All coauthors contributed to and approved the final manuscript.
Funding
This study was sponsored by Amgen Inc. The journal’s Rapid Service and Open Access Fees were funded by Amgen Inc.
Data Availability
Qualified researchers may request data from Amgen clinical studies. Complete details are available at the following: https://wwwext.amgen.com/science/clinical-trials/clinical-data-transparency-practices/
Declarations
Conflict of Interest
All authors (Gianluca Turcatel, Yi Xiao, Scott Caveney, Gilles Gnacadja, Julie Kim and Nestor A. Molfino) are employees and stockholders of Amgen Inc.
Ethical Approval
This study complied with all applicable laws regarding patient privacy, using Health Insurance Portability and Accountability Act (HIPAA)-compliant de-identified retrospective data sources. No direct patient contact or primary collection of individual patient data occurred.
Footnotes
Prior Presentation: Results presented in this manuscript have been presented in slide format at the 2023 Annual Meeting of the American College of Chest Physicians (CHEST), held October 8–11 in Honolulu, HI, USA.
References
- 1.Acosta JN, Falcone GJ, Rajpurkar P, Topol EJ. Multimodal biomedical AI. Nat Med. 2022;28(9):1773–84. [DOI] [PubMed] [Google Scholar]
- 2.Moor M, Banerjee O, Abad ZSH, et al. Foundation models for generalist medical artificial intelligence. Nature. 2023;616(7956):259–65. [DOI] [PubMed] [Google Scholar]
- 3.Molfino NA, Turcatel G, Riskin D. Machine learning approaches to predict asthma exacerbations: a narrative review. Adv Ther. 2024;41(2):534–52. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Couillard S, Do WIH, Beasley R, Hinks TSC, Pavord ID. Predicting the benefits of type-2 targeted anti-inflammatory treatment with the prototype Oxford asthma attack risk scale (ORACLE). ERJ Open Res. 2022;8(1):00570-2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Couillard S, Petousi N, Smigiel KS, Molfino NA. Toward a predict and prevent approach in obstructive airway diseases. J Allergy Clin Immunol Pract. 2023;11(3):704–12. [DOI] [PubMed] [Google Scholar]
- 6.Tong Y, Messinger AI, Wilcox AB, et al. Forecasting future asthma hospital encounters of patients with asthma in an academic health care system: predictive model development and secondary analysis study. J Med Internet Res. 2021;23(4):e22796. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Zein JG, Wu CP, Attaway AH, Zhang P, Nazha A. Novel machine learning can predict acute asthma exacerbation. Chest. 2021;159(5):1747–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Goto T, Camargo CA Jr, Faridi MK, Yun BJ, Hasegawa K. Machine learning approaches for predicting disposition of asthma and COPD exacerbations in the ED. Am J Emerg Med. 2018;36(9):1650–4. [DOI] [PubMed] [Google Scholar]
- 9.Wang Z, Li Y, Gao Y, et al. Global, regional, and national burden of asthma and its attributable risk factors from 1990 to 2019: a systematic analysis for the global burden of disease study 2019. Respir Res. 2023;24(1):169. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Fuhlbrigge A, Peden D, Apter AJ, et al. Asthma outcomes: exacerbations. J Allergy Clin Immunol. 2012;129(3):S34-48. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Agress A, Oprea Y, Roy S, et al. The association between malignancy, immunodeficiency, and atopy in IgE-deficient patients. J Allergy Clin Immunol Pract. 2024;12(1):185–94. [DOI] [PubMed] [Google Scholar]
- 12.Price D, Wilson AM, Chisholm A, et al. Predicting frequent asthma exacerbations using blood eosinophil count and other patient data routinely available in clinical practice. J Asthma Allergy. 2016;9:1–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Venkatesan P. 2023 GINA report for asthma. Lancet Respir Med. 2023;11(7):589. [DOI] [PubMed] [Google Scholar]
- 14.Harvey MG, Hancox RJ. Elevation of cardiac troponins in exacerbation of chronic obstructive pulmonary disease. Emerg Med Australas. 2004;16(3):212–5. [DOI] [PubMed] [Google Scholar]
- 15.Ikeda M, Ohshima N, Kawashima M, Shiina M, Kitani M, Suzukawa M. Severe asthma where eosinophilic granulomatosis with polyangiitis became apparent after the discontinuation of dupilumab. Intern Med. 2022;61(5):755–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Carroll CL, Coro M, Cowl A, Sala KA, Schramm CM. Transient occult cardiotoxicity in children receiving continuous beta-agonist therapy. World J Pediatr. 2014;10(4):324–9. [DOI] [PubMed] [Google Scholar]
- 17.Yalta K, Yalta T, Gurdogan M, Palabıyık O, Yetkın E. Cardiac biomarkers in the setting of asthma exacerbations: a review of clinical implications and practical considerations. Curr Allergy Asthma Rep. 2020;20(6):17. [DOI] [PubMed] [Google Scholar]
- 18.Pavasini R, d’Ascenzo F, Campo G, et al. Cardiac troponin elevation predicts all-cause mortality in patients with acute exacerbation of chronic obstructive pulmonary disease: Systematic review and meta-analysis. Int J Cardiol. 2015;191:187–93. [DOI] [PubMed] [Google Scholar]
- 19.Mullur J, Buchheit KM. Aspirin-exacerbated respiratory disease: updates in the era of biologics. Ann Allergy Asthma Immunol. 2023;131(3):317–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Kurth T, Barr RG, Gaziano JM, Buring JE. Randomised aspirin assignment and risk of adult-onset asthma in the Women’s Health Study. Thorax. 2008;63(6):514–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Barr RG, Kurth T, Stampfer MJ, Buring JE, Hennekens CH, Gaziano JM. Aspirin and decreased adult-onset asthma: randomized comparisons from the Physicians’ Health Study. Am J Respir Crit Care Med. 2007;175(2):120–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Park C, Jang JH, Kim C, et al. Real-world effectiveness of statin therapy in adult asthma. J Allergy Clin Immunol Pract. 2024;12(2):399-408.e396. [DOI] [PubMed] [Google Scholar]
- 23.Kim JH, Wee JH, Choi HG, et al. Association between statin medication and asthma/asthma exacerbation in a national health screening cohort. J Allergy Clin Immunol Pract. 2021;9(7):2783–91. [DOI] [PubMed] [Google Scholar]
- 24.Inokuchi R, Aoki A, Aoki Y, Yahagi N. Effectiveness of inhaled furosemide for acute asthma exacerbation: a meta-analysis. Crit Care. 2014;18(6):621. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Hodroge SS, Glenn M, Breyre A, et al. Adult patients with respiratory distress: current evidence-based recommendations for prehospital care. West J Emerg Med. 2020;21(4):849–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Rodríguez Vázquez JC, Pino Alfonso PP, Gassiot Nuño C, Páez PI. Usefulness of inhaled furosemide in a bronchial asthma attack. J Investig Allergol Clin Immunol. 1998;8(5):290–3. [PubMed] [Google Scholar]
- 27.Tibrewal C, Modi NS, Bajoria PS, et al. Therapeutic potential of vitamin D in management of asthma: a literature review. Cureus. 2023;15(7):e41956. [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
Data Availability Statement
Qualified researchers may request data from Amgen clinical studies. Complete details are available at the following: https://wwwext.amgen.com/science/clinical-trials/clinical-data-transparency-practices/