Asthma exacerbations can be predicted effectively using machine-learning (ML)-based models, according to systematic review and meta-analysis findings published in BMC Pulmonary Medicine.

Investigators sought to estimate the predictive performance and practical use of ML-based prediction models for asthma exacerbations.

They conducted a systematic review and meta-analysis searching the Web of Science, EMBASE, Cochrane Library, and PubMed databases through mid-December 2022 for studies published in English that developed prediction models for asthma exacerbations by applying ML methods. Patients with asthma were at least 5 years of age.

The meta-analysis included 11 studies (identifying 23 prediction models); 6 of those were published in the past 3 years (suggesting a popular trend in applying ML algorithms in asthma); 10 were retrospective studies; 1 was prospective. Participants in 9 of the studies were included regardless of asthma severity, control levels, or treatment. Prediction windows varied from 0 days to 4 years, with 7 of the studies keeping the prediction window within 1 year. Individual study sample sizes ranged from 94 patients to more than 750,000 participants.

Most frequently used ML methods included logistic regression, random forest, and boosting. Logistic regression models incorporated a relatively minor number of predictors, whereas random forest and boosting models applied significantly more predictors. Investigators found the most common important predictors were systemic steroid use, short-acting beta2-agonists, and emergency department visits. Additional common predictors included age, asthma diagnosis, and exacerbation history.

“

This study showed that ML could achieve great performance in predicting asthma exacerbations. Future studies should focus on improving models’ generalizability and practicability, thus driving the application of these models in clinical practice.

For the 21 prediction models in the meta-analysis reporting area under the curve of the receiver operating characteristics (AUROC), the pooled AUROC was 0.80; 95% CI, 0.77-0.83, suggesting a good discrimination ability. Notably, model performance is directly related to sample size and requires large numbers (5 studies had <10,000 participants; 6 studies had >10,000 participants). Subgroup pooled analysis of studies with less than 10,000 participants vs more than 10,000 participants showed AUROC of 0.68 vs 0.82, suggesting a large dataset is necessary for the development of ML prediction models.

The boosting method achieved the best performance between different ML models (AUROC, 0.84; 95% CI, 0.81-0.87) in subgroup analysis. Investigators noted the ML prediction models could predict which patients were at high risk for exacerbation for periods ranging from several days to several years.

The researchers’ assessment of the overall quality of studies found a high risk for bias for all 11 studies, and high risk for applicability in 10 of the studies. Overall pooled sensitivity and specificity showed extremely high heterogeneity.

Systematic review and meta-analysis limitations include heterogeneity within studies.

“This study showed that ML could achieve great performance in predicting asthma exacerbations,” the researchers concluded. “Future studies should focus on improving models’ generalizability and practicability, thus driving the application of these models in clinical practice.”

Source link