Data sources and processing
Up until now, ICIs approved by FDA for antitumor treatment include pembrolizumab, nivolumab, atezolizumab, avelumab, durvalumab, cemiplimab and ipilimumab. Therefore, in this study, these seven ICIs were chosen as the study drugs. Spontaneous ADE reports were retrieved from July 1, 2014 (considering the FDA marketing approval of the first ICIs, pembrolizumab on September, 2014) to December 31, 2019 in the FAERS database (https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html), and EV (http://www.adrreports.eu/en/search.html) was accessed and queried on February 29, 2020. In this study, the processing of data downloaded from FAERS database followed the customized strategy described previously[12]. Given that the drug names in the EV and FAERS databases are not standardized, all drug names were standardized into active substances with relevant Anatomical Therapeutic Chemical (ATC) codes before data analysis. Besides, we detected and eliminated duplicates and multiple records (reports with at least overlaps in 3 on 4 of considered key-fields, including event date, age, gender, and reporter country). And the incomplete records with missing event dates, age, gender, or reporter country were removed for further study. Furthermore, all ADEs reported in these databases are coded by preferred terms (PTs) from the Medical Dictionary for Drug Regulatory Activities (MedDRA). Previous studies have demonstrated several fatal ICIs-associated side effects, including myocarditis, colitis, hepatitis, pneumonitis, nephritis, and so on[5, 6, 8]. Therefore, to identify the cases of fatal ICIs-associated ADRs, we searched these spontaneous reporting pharmacovigilance databases using the following MedDRA PTs: myocarditis, colitis, hepatitis, pneumonitis, and nephritis. And these fatal ICIs-associated ADRs were further analyzed in this study by signal-detection algorithms.
Only FAERS, can realize signal detection by using open database, but if you pay for it, other databases can also do it. Therefore, this study used the open database to obtain the number of fatal ICIs-associated ADRs in two major databases, the fatality rate caused by ICIs-associated ADRs in EV database, the age and gender distribution of ICIs-associated myocarditis in EV and FAERS databases, and the signal value of ICIs-associated myocarditis in FAERS database.
Data mining algorithm
The data mining methods used to detect the ADR signals in spontaneous reporting systems are mainly the disproportionality methods[13, 14], which are based on spontaneous reports submitted for a lot of drugs and ADRs[15]. All the reports included in the FAERS database from July 1, 2014 to December 31, 2019 were selected to determine the ADR signals in the present study.
To detect the ADRs signals, both Frequentist (non-Bayesian) methods and Bayesian methods were used to calculate disproportionality by using reporting odds ratio (ROR)[16], proportional reporting ratios (PRR)[17], and information component (IC) of Bayesian confidence propagation neural networks (BCNPP)[18], which are mainly based on a two-by-two contingency table (Supplementary material, Table S1).
PRR and ROR have the advantages of easy calculation and high sensitivity, and the results of PRR and ROR are highly consistent. Therefore, PRR and ROR methods are often used to estimate signals of ADRs. The calculation formulas of ROR and PRR are ROR=(a/c)/(b/d), PRR=a(c+d)/c(a+b), respectively. When an adverse event is new and rare and the second row of two-by-two contingency table involves all drugs (that is, a or b in two-by-two contingency table is a very small number, even zero), PRR and ROR are not applicable. In this scenario, BNCPP can be used to calculate the signal values of ADRs[19]. BCNPP method uses the Bayesian discrimination principle based on the fourfold table. The core of the BCNPP method is to calculate the value of IC. The calculation formula of the IC is IC = . Since our study involves some rare ICIs-related ADRs, we use all three data mining methods to determine the signal values at the same time.
For ROR, the threshold criteria of ADR signal are a≥2 and the lower bound of the 95% two-sided confidence interval (CI) is greater than one[12]. For PRR, the signal judgment criteria are a≥3, x2≥4, and PRR≥2[17]. For the IC, the conditions for signal generation are IC>0 and the lower bound of the 95% two-sided CI>0[20]. The higher the value, the stronger the signal appears to be[12, 19].
MATLAB R2019b software was used to detect the ADR signals in this study.