Ensemble Machine Learning Using Quantitative Chest CT and Clinical Biomarkers to Predict Asthma Severity and Outcomes
Year: 2021
Type: conference-abstract
Abstract: Background: There is expanding demand for methods to predict disease progression and outcomes in chronic lung diseases such as asthma.There have been several early efforts to exploit quantitative patterns in chest CT images to predict disease severity (e.g.GOLD stage in COPD) and to classify subtle textural differences critical to diagnosing phenotypes of ILD using deep learning with convolution neural networks (CNN's).However, such CCN's do not incorporate well developed patterns of disease derived from expert radiological assessments, and non-imaging clinical biomarkers.We hypothesize that "expert filters" in the form of decision trees (DTs) can be combined into an ensemble classifier with CNN's to exploit unseen quantitative patterns in chest CT images while incorporating expert knowledge to classify patterns of asthma progression.Methods: Chest CT scans of 46 subjects from phase 3 of the severe asthma research program (SARP 3) were acquired using a quantitative protocol at baseline and 3-year follow-up.Scans were reformatted to capture 4 slices per patient (sagittal right and left lungs, axial at level of carina, and coronal through the carina) in a 2.5-dimensional (2.5D) design.These image planes were used to train a 2.5D CNN to predict 2 outcomes: improved lung function vs. declinestable disease.Severe asthma biomarkers of instability and progression, including sinusitis, GERD, BMI, sex, maximum post-albuterol reversibility, log (blood eosinophil counts), and post-triamcinolone FEV1 %p were then used to develop an ensemble classifier of four DT's and six CNN's trained on 19 subjects using bootstrapping with replacement, validated on 11, and tested on 16.The resulting classifier used both majority (MV) and highest performing (HPV) voting across the DT and CNN networks to test performance.Results: Preliminary results for the network demonstrated an overall accuracy for MV vs HPV of 75%, and 81% respectively.In the best performing HPV classifier, 8 of 9 of asthmatics with progression were identified (sensitivity = 89%) and 5 of 7 improving asthmatics were identified (specificity = 71%).The classifier struggled to identify healthy subjects, incorrectly designating one as declining progressive asthma and the other as stable asthma.Conclusions: These results are promising but require inclusion of more scans and subjects for training to refine performance particularly from healthy subjects.Training of the classifier has recently expanded to 410 CT scans acquired in SARP 3, including 42 healthy subjects.
Authors: S.B. Fain, S. Jiang, K. Carey, L.C. Denlinger, N.N. Jarjour, M.L. Schiebler, R.L. Sorkness, A.D. Hahn, V.Z. Tejeda
Institution: University of Wisconsin–Madison
Language: English
Subfield: Physiology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
OpenAlex ID: W4246133136