Keywords: Real Word Data, Artificial Intelligence
Room Madison A
In the medical device effectiveness and safety analysis, while randomized controlled clinical trials (RCT) are well known gold standard with regard to scientific rigor, the analysis model of RCTs is usually simple. Even for observation studies using propensity score (PS) matching for comparison, the PS model is relatively not complicated. Rarely a PS model include more than 10 covariates. When Real World Data becomes more and more widely used in FDA’s regulatory practice, high dimension data bring new opportunity and challenge to the analysis community. For example, a safety analysis study monitoring breast implant rupture rate collected more than 180 different symptoms. How to identify the key variables and further uncover the causal relationships is essential to FDA’s evaluation whether the implant is safe. For modeling on such a high-dimension data, challenge usually includes, but not limited to: a) model selection becomes difficult; b) linear model results could be misleading due to collinearity or confounding even linear assumption holds; c) hidden patterns are not linear. This presentation discusses the approach of dimension reduction and pattern detection in FDA’s medical device regulatory decision making practice by means of data mining and machine learning on data with complicated structure.