AI Peer Review

In this work, we model the risk of hospitalization for a large number of drug-disease combinations. Our methodology is akin to genome-wide association studies (GWAS), in which a simple model is used to estimate the effect of a large number of loci in a hypothesis-free manner. As in GWAS, this screen risks spurious relationships and requires further analysis, but complements target-driven repurposing.

We fitted a separate model for each drug-disease pair. We considered all hospitalizations resulting in an ICD10 code contained in the CCSR category of choice as an event... Using only the information from the assessment center, we could not consider how long subjects had been on treatment, neither how long they would stay on it, nor the medication dosage. We also could not measure if subjects changed treatment over time.

We found that the bias towards unfavorable effects did not vanish when reducing the pair Huang distance cut-off, implying that this discrepancy is due to non-observable variables. Based on the previous results, we chose the combination of covariates, time, and Huang distance cut-offs that would result in the most balanced number of total associations. We used this configuration for all further analysis.

Using individual ICD10 codes for our analysis would be challenging due to the small number of events for each code... Grouping drugs and diseases by functionality allowed us to gather enough events to estimate potential effects while keeping fine-grained categories. We estimated the risk of hospitalization, making our method inherently preventive although some results could generalize to already hospitalized patients.