Scholarlink

Purpose <p>Detecting asthma exacerbations presents challenges for traditional clinical methods. While machine learning has shown promise in predicting exacerbations, prior systematic reviews have rarely differentiated between shallow and deep learning approaches. This paper aims to identify studies that implemented either approach, compare their methodologies and characteristics, assess the consistency of their results and examine how model interpretability was addressed.</p> Method <p>We conducted a systematic review using the three-stage approach by Tranfield et al., guided by the PRISMA protocol. An extensive search across five databases, including PubMed, Scopus, and Web of Science, yielded 218 studies. Of these, 96 were excluded based on predefined criteria, leaving 23 for analysis.</p> Results <p>Among the included studies, 15 (65.21%) used only shallow algorithms, with Random Forest being the most common. Four studies (17.39%) applied deep learning exclusively, while another four combined both approaches. Performance metrics varied across studies: In one study, a deep learning model achieved the highest AUC (69.4%) compared to shallow models like Logistic Regression (67.4%), while in another, a shallow Decision Tree slightly outperformed a deep model in accuracy (89.8% vs. 89.5%), possibly due to dataset size limitations. Secondary data was most frequently used, featuring in 16 studies, while only 7 (30.43%) used primary data. Model explainability was addressed in just 8 studies (34.78%).</p> Conclusion <p>Although asthma exacerbation prediction models have shown promise, explainability remains crucial, especially for deep learning methods that are often seen as “black boxes”. Ensuring interpretability will foster trust and support effective clinical integration of both shallow and deep learning models. Additionally, larger datasets, particularly for deep learning, are essential to improve prediction accuracy and advance healthcare applications. This systematic review is registered with PROSPERO (CRD42024604818).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Asthma exacerbation prediction using shallow and deep learning approaches: A systematic review

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Stanley Ebhohimhen Abhadiomhen,
  • Royransom Chiemela Nzeh,
  • Modesta Ero Ezema,
  • Abel Onolunosen Abhadionmhen,
  • Blessing Chimezie Uzo,
  • Assumpta Obianuju Ezugwu

摘要

Purpose

Detecting asthma exacerbations presents challenges for traditional clinical methods. While machine learning has shown promise in predicting exacerbations, prior systematic reviews have rarely differentiated between shallow and deep learning approaches. This paper aims to identify studies that implemented either approach, compare their methodologies and characteristics, assess the consistency of their results and examine how model interpretability was addressed.

Method

We conducted a systematic review using the three-stage approach by Tranfield et al., guided by the PRISMA protocol. An extensive search across five databases, including PubMed, Scopus, and Web of Science, yielded 218 studies. Of these, 96 were excluded based on predefined criteria, leaving 23 for analysis.

Results

Among the included studies, 15 (65.21%) used only shallow algorithms, with Random Forest being the most common. Four studies (17.39%) applied deep learning exclusively, while another four combined both approaches. Performance metrics varied across studies: In one study, a deep learning model achieved the highest AUC (69.4%) compared to shallow models like Logistic Regression (67.4%), while in another, a shallow Decision Tree slightly outperformed a deep model in accuracy (89.8% vs. 89.5%), possibly due to dataset size limitations. Secondary data was most frequently used, featuring in 16 studies, while only 7 (30.43%) used primary data. Model explainability was addressed in just 8 studies (34.78%).

Conclusion

Although asthma exacerbation prediction models have shown promise, explainability remains crucial, especially for deep learning methods that are often seen as “black boxes”. Ensuring interpretability will foster trust and support effective clinical integration of both shallow and deep learning models. Additionally, larger datasets, particularly for deep learning, are essential to improve prediction accuracy and advance healthcare applications. This systematic review is registered with PROSPERO (CRD42024604818).