Table 2.
Summary of studies employing ML to predict asthma outcomes
| References | Algorithms implemented | Predictive features | Notes |
|---|---|---|---|
| Farion et al. [59] | Naïve Bayes | Not assessed | Predictions made by physicians were more accurate than those made with the Naïve Bayes model or based on PRAM score |
| Xu et al. [60] | Random forest | 160 SNPs | Highest predictive power was obtained by incorporating the 160 most significant SNPs |
| Finkelstein et al. [61] |
Bayesian classifier, Adaptive Bayesian network, Support vector machines |
Not assessed | Prediction accuracy drops by decreasing the prediction time windows (from day 7 to day 1) |
| Blakey et al. [18] | Logistic regression | History of asthma-related events (acute oral corticosteroid courses, emergency visits), frequency of healthcare utilization, lung function, smoking status, blood eosinophilia, rhinitis, nasal polyps, eczema, gastroesophageal reflux disease, obesity, age, and sex | |
| Patel et al. [62] |
Decision trees, Lasso logistic regression, Random forests, Gradient boosting machines |
Oxygen saturation, respiratory rate, triage acuity, pulse rate, weight-for-age z-score, age, socioeconomic status, and weather variables | Gradient boosting produced the highest predictive power |
| Goto et al. [63] |
Lasso regression, Random forest, XGBoost, Deep neural network |
Critical care outcome: arrhythmia, respiratory rate, congestive heart failure, temperature, oxygen saturation, arrival mode (ambulance vs walk-in), asthma status Hospitalization outcome: age, congestive heart failure, arrival mode, asthma status, COPD status, oxygen saturation, respiratory rate |
XGBoost and random forest were the best-performing algorithms for critical care and hospitalization prediction, respectively |
| Zein et al. [64] |
Logistic regression, Random forest, LightGBM |
Non-severe exacerbation outcome: history of sinusitis, treatment with combination iCS and LABA or with HDiCS, and leukotriene inhibitors, high BMI, eosinophilia, low blood albumin Emergency department visit outcome: age, Black/African American race, a history of non-severe exacerbations, history of severe asthma, eosinophilia, low blood albumin Hospitalization outcome: a history of non-severe exacerbations, low hemoglobin, high BMI |
LightGBM generated the best predictions for non-severe exacerbation, emergency department visit, and hospitalization |
| Noble et al. [19] | Logistic regression | Previous hospitalization, older age, being underweight, smoking, history of asthma attacks and blood eosinophilia | |
| Lugogo et al. [65] | Gradient boosting machines | Mean number of daily albuterol inhalations during the 4 days prior to the prediction, inhalation parameters in the 4 days prior to prediction (PIF, inhalation volume, and inhalation duration), and comparison to the baseline values for these inhalation parameters | |
| Zhang et al. [24] |
Logistic regression, Decision tree, Naïve Bayes, Perceptron algorithms |
Not assessed | Logistic regression was the best-performing model |
| Tong et al. [66] | XGBoost | Features related to prior emergency department visits and asthma medications, race | |
| Overgaard et al. [67] |
Logistic regression, Support vector machine, Random forest, Gaussian Naïve Bayes, Perceptron |
Not assessed | Three best-performing algorithms were logistic regression, random forest, and perceptron |
| Lan et al. [68] |
Conditional random forest, Conditional tree, Generalized linear model |
Top four predictive features were prednisone usage, race, daily particulate matter exposure, and obesity | |
| Haque et al. [69] | Deep neural network regression | Not assessed | Model was trained to predict ACT score |
| Halner et al. [70] | Random forest |
Top five predictors were breathlessness, sputum purulence, use of long-acting muscarinic antagonist, number of unscheduled primary care and emergency department visits in the previous 12 months, and ICS use |
Outcome was treatment failure, defined as the need for additional SCS and/or antibiotics, hospital readmission or death within 30 days of initial emergency department visit |
| de Hond et al. [71] |
XGBoost, Support vector machines, Logistic regression |
Not assessed | Logistic regression provided the best predictive performance |
FEV1 forced expiratory volume in 1 s, PRAM pediatric respiratory assessment measure, SNPs single nucleotide polymorphisms, iCS inhaled corticosteroids, LABA long-acting beta agonists, HDiCS high-dose inhaled corticosteroids, ACT Asthma Control Test, SCS systemic corticosteroids