Matrix Decomposition in Meta-Analysis for Extraction of Adverse Event Pattern and Patient-Level Safety Profile

The purpose of assessing adverse events (AEs) in clinical studies is to evaluate what AE patterns are likely to occur during treatment. In contrast, it is difficult to specify which of these patterns occurs in each patient. To tackle this challenging issue, we construct a new statistical model including nonnegative matrix factorization (NMF) by incorporating prior knowledge of AE-specific structures such as severity, combination therapy, and patient information. In our model, the large patient-AE occurrence matrix is decomposed into two small rank matrices. A matrix decomposition in which the elements of these matrices are restricted to positive values is called NMF. Because insufficient information is derived from a single clinical study, we have to extend our model to analyze data from multiple clinical studies simultaneously, as in a meta-analysis. To this end, we assume that a treatment-dependent term can be separated from a study-dependent term. The extracted typical treatment-specific AE patterns coincided with background knowledge. We demonstrated the extraction of AE patterns and patient-level safety profiles using the data in the Project Data Sphere (submitted).

Next, using the data sets in Vivli, we would like to incorporate temporal order of AEs, patient information including age, weight and blood test values into our statistical model. We believe that our research will allow us to extract treatment-specific AE patterns and understand what patient information makes a certain pattern more likely to occur, and then will enable the proposal of drugs that are likely to reduce AEs for each patient.