This is an intriguing proposal using matrix decomposition to predict likelihood of future adverse events based on prior adverse reactions. However, I am concerned that the investigators are biasing their results by excluding patients who have very few or no adverse events. They should consider the impact of this exclusion on the generalizability of their findings to a real-world patient population. In addition, I was surprised by the limited number of covariates being used in the modeling (no information on patient co-morbidity, concomitant medication use, etc.). Finally, will the model results obtained from analyses on one trial be validated in another trials’ patient population? Or will the 4 trials samples be combined for matrix decomposition analyses?