Producing NONMEM dataset using a standard SAS® program

PDF of poster

Objectives: The overall objective is to implement a system which will ensure results from NONMEM analysis to be ready in time for decisions made about drug development. One main problem is a bottleneck in the data building part of the process especially with the large datasets in phase 2 and 3. The aim of this project is to greatly reduce the time taken to prepare a dataset for NONMEM from the raw data in the database. 

Method: Standard specification for the required structure of NONMEM dataset was developed to minimise changes in what is required between studies. Key data components were identified. Date/time and value of PK and/or PD observations (plasma and urine) and the dosing history and route (non steady state and/or steady state; intravenous and/or extravascular administration) formed the main body of the dataset. The covariates are from demographic, laboratory, co-medications and adverse events data. These covariates were grouped as follows: covariates that only exist once per patient, covariates that change once per visit, and covariates like co-medications which can start and stop at various times throughout the study. Standard rules for replacing missing covariates were also defined. A SAS programme was developed to reflect the standard requirements of a NONMEM dataset.

Result: Six standard input datasets are created from the various source datasets within the study database. These are then used in the rest of the program which will remain mainly unchanged. The standard datasets contain dosing data, observation (PK and PD) data, time-independent covariates (e.g. sex or race), time-dependent covariates (e.g. weight), laboratory data and co-medication data.  

The dosing and observation data is set together to create the main body of the NONMEM dataset. The demographic covariates are prepared by replacing missing values, and then merged together by patient number. The time-dependent covariates are then merged by patient and visit number, and the missing values replaced by carry forward and carry backward method. Laboratory covariates are then merged and missing values are once again replaced. Finally the co-medications are merged and flags are assigned if the patient was taking a particular medication at each time when there was a dosing or observation measurement.

Conclusion: The time taken to produce the NONMEM dataset was reduced to a quarter of what it took in the past. This has ensured the NONMEM analysis can be performed very soon after the database is locked, and in time for decisions made about future drug development.