Online Program

Return to main conference page
Tuesday, September 24
Tue, Sep 24, 4:15 PM - 5:30 PM
Recent Applications and Developments in Machine Learning

Application of Machine-Learned Quantitative Imaging Biomarkers in Clinical Trials in Subjects with Interstitial Lung Disease (300996)

Matthew Brown, University of California at Los Angeles 
Jonathan Goldin, University of California at Los Angeles 
*Grace Hyun J Kim, University of California at Los Angeles 

Keywords: Machine Learning, Minimal clinically importance difference, lung fibrosis

Prognostic, predictive classifiers, and surrogate endpoints are one of important characteristics of biomarkers. These characteristics can independently and prospectively be evaluated and validated in a setting of clinical trial as an outcome. Machine learning (ML) is one of artificial intelligence (AI) technique driven by an algorithm that led into a development of computer aided diagnosis and has been applied to many health-related studies. There are two parts of validation for testing a developed algorithm: (1) analytic validation: the robustness of algorithm, which can be tested using repeatability and reproducibility study; (2) clinical validation: Minimal clinically importance difference (MCID) is another key important factor to be utilized as a secondary or exploratory endpoint in a clinical trial. We developed a quantitative lung fibrosis (QLF) score using a supervised machine-learning technique to estimate an extent of fibrotic reticulation on high-resolution computed tomography, medical images (AUC=0.96 with gold standard from three expert readers). The score is generated from a trained algorithm using a support vector machine with texture features. After analytic research was followed to understand the robustness, repeatability, and reproducibility of the measurement, QLF score has been used in 3 NIH clinical trials, and 7 industrial clinical trials. Reproducibility of systematic sampling is estimated as 0.20% in whole lung and 1% in the most severe lobe. MCID is estimated as 0.5%-1% in whole lung and 2% in the most severe lobe. Significant associations were found between the primary outcome and QLF score (rho: -0.40 to -0.60), as well as in the skin biopsy processed with gene expression and pathway enrichment analyses. We will present the recent applications where the scores from ML are used and tested as a secondary or exploratory outcome in the phase 2 and 3 clinical trials. Furthermore, we will present new ML and deep learning applications.