Recommendation	Rationale
If prior knowledge suggests data from a particular organ system should be monitored, consider extreme value modelling on data arising from each trial for the compound of interest. For example, if preclinical data suggested a potential liver issue, prepare to model ALT; if another compound in the class showed kidney signals, prepare to model creatinine	Extreme-value modelling has been demonstrated, in various examples, to provide useful predictions of drug toxicities from early-phase data. If it is possible to pre-specify the modelling and prediction exercise, the results have greater credibility than if they are data driven, and resources can be allocated up front to ensure the work is done to appropriate deadlines [74, 76]
Some analyses will be data driven, suggested by observed extremes in the data. These could also be subjected to extreme value modelling and the statistical evidence thus acquired interpreted in context	Not all potential safety issues are known in advance, so some analyses are necessarily data driven. It is inappropriate to consider such analysis illegitimate or to yield unreliable results provided they are interpreted in context. Statistical inference is only one part of the larger process of scientific inference [74]
Extreme value modelling can commence as early as phase I; however, in most cases, phase II data need to be available for reliable inferences to be made	Experience suggests that phase I data may be sufficient for extreme value modelling to identify toxicity, but that sometimes the sample sizes are too small. Modelling and prediction have the most value to add when the volume of available data is low, so such exercises should be commenced as soon as possible [74]
Properly trained, Independent Data Monitoring Committees or Safety Review Boards are likely to benefit from extreme value modelling of unblinded data	When an IDMC exists, there are sometimes reasons for additional monitoring. It follows that applying proven methodology to emerging data will provide the best chance of identifying and characterising the safety issue as soon as possible [74]
When extreme value modelling does not find evidence of a safety signal in studies of short duration, extrapolation beyond observed durations of exposure is discouraged	It is reasonable to expect that some toxic effects of drugs will not manifest themselves until several weeks or months of exposure have occurred. If extreme values are not observed at relevant doses in short trials, proceed with caution, acknowledging that they could occur after longer durations of exposure
Multiplicity adjustment provides a useful tool to improve the positive predictive value in signal detection in clinical trial data. The use of multiplicity adjustment needs to be evaluated against the size of the available clinical trial database	The ability for ADR detection is highly influenced by ADR frequency in the source dataset. Thus, database size and event reporting frequency must be taken into consideration when the use of multiplicity adjustments for ADR candidate selection is considered [75]
The use of Bayesian Hierarchical Models can improve the efficiency of signal detection through borrowing of strength from other relevant events in the clinical trial dataset. This must be weighed against the more complex computational requirements of Bayesian modelling	Bayesian Hierarchical Models provided the best performance with regard to positive predictive value, specificity, sensitivity and negative predictive value, mainly owing to their ability to “borrow strength” across similar terms [75]
The use of more specific MedDRA® groupings can further improve signal detection in clinical trial data	The use of narrow-term groupings for analysis provided slightly better results for signal detection compared with the analysis based on MedDRA® PTs alone [75]