Abstract
RATIONAL: Primary ciliary dyskinesia (PCD) is a genetic condition that results in dysmotile cilia and abnormal mucociliary clearance. Despite advances in understanding the pathogenesis of PCD, diagnosis continues to be challenging. Here we used feature-based machine learning and image-based deep learning to objectively quantify the directed particle transport of motile cilia and detect PCD-related cilia dysfunction. METHODS: Fluorescent microspheres were captured on cultured multiciliated cells using high-speed video microscopy as a proxy for motile cilia function. An interactive Jython script was designed to automatically detect, track and extract raw track metrics from videos. Data was subsequently analyzed to approximate a quantifiable and visual signature of ciliary transport through a custom-built Python Package, CiliaTracks. RESULTS: Airway epithelial cells were obtained from 14 individuals with genetically confirmed PCD, 10 healthy donors, and 2 patients with cystic fibrosis. A total of 602 videos (301 PCD and 301 non-PCD) were captured. Quantitative and visual analyses of fluorescent microsphere trajectories, including kinematic metrics and trajectory plots, revealed distinct motility profiles between PCD and non-PCD samples. Classical machine learning models and a convolutional neural network were employed to classify PCD using both modalities, demonstrating excellent accuracy of 95-97%, and the capacity to differentiate PCD from normal cells or cystic fibrosis. CONCLUSION: Cilia-propelled microsphere transport exhibits unique trajectory patterns in PCD, enabling differentiation from non-PCD samples. Machine learning provides an objective and accurate framework for characterizing ciliary dysfunction, offering potential as a diagnostic tool for PCD.