Objective: Infants hospitalized with bronchiolitis may experience persistent symptoms linked to future chronic lung diseases like bronchiectasis. Identifying phenotypes during hospitalization could guide targeted interventions. As traditional clustering requires large datasets, this study explores whether Unsupervised Feature Extraction Algorithms (UFEAs) and clustering can identify high-risk profiles in a small dataset of Indigenous infants.
Methods: We included 128 Indigenous infants hospitalized with bronchiolitis at the Royal Darwin Hospital, Northern Territory, Australia. Eight UEFAs were applied to reduce the dimensionality of 22 variables across 2-17 dimensions. A support vector machine classifier assessed the effectiveness of each UFEA in classifying bronchiectasis. Kernel Principal Component Analysis with nine dimensions performed best, and these dimensions were used for clustering.
Results: Six clinical profiles were identified. Profile C, the highest-risk group with the most infants with bronchiectasis (45%), preterm birth (95%), low birth weight (86%), weight-for-length z-score < -2 (62%), household smoke exposure (90%), and antibiotics prescribed before hospitalization (100%). Profile D, the second-highest risk, had bronchiectasis (30%), the highest wet/productive cough (45%), crackles/crepitations (36%), and wheeze (18%). Profile F infants included bronchiectasis (22%), oxygen supplementation (91%), and lobar collapse/consolidation on chest X-rays (65%). Profile A included bronchiectasis (5%) and household smoke exposure (30%), and Profile E showed bronchiectasis (9%) and household smoke exposure (36%). Profile B, the lowest-risk group, with no bronchiectasis (0%), preterm birth (15%), low birth weight (10%), and any bacteria (5%).
Conclusion: Using UFEAs and clustering, we reduced dataset dimensionality, effectively identifying six unique, clinically significant risk profiles in Indigenous infants.
扫码关注我们
求助内容:
应助结果提醒方式:
