Emma Vestesson (1), Silke Gastine (1), Iek Fan Cheng (2), Lamia Samrin (2), Pritesh Patel (2), Sujith Samarasinghe (2), Neil Sebire (1, 3), Joe Standing (1)
1. UCL Great Ormond Street Institute of Child Health, London, UK 2. Great Ormond Street Hospital, London, UK 3. NIHR GOSH BRC, London, UK
Objectives:
The primary aim of this project is to develop a proof-of-concept therapeutic drug monitoring tool that can be used in clinical practice and that can easily be reused for different PK models and drugs. The secondary aim is to understand how the precision of estimates of concentration estimates and the area under the curve (AUC) depend on the number of data points used in the prediction to help us understand how sensitive the developed our model is to data quality.
Methods:
An R shiny app (1) (2) was developed that allows clinicians to run maximum a posteriori estimations for drug dosing without the need to write any code. A two-compartment population pharmacokinetic model was applied for patients with osteosarcoma given high-dose Methotrexate as a case study (for full details see (3)) but can be easily applied elsewhere. The app prompts the user to enter patient characteristics eg age, sex and kidney function, and then estimates when the drug concentration is likely to be below 0.2 micromol/L. The app also returns estimated individual model parameters and the AUC. A number of features have been implemented to ensure patient safety and improve user experience. Two levels of restrictions are implemented on input values; if the value entered is outside of the range of values that the model was developed on then the user sees a warning but can still get an estimate and if the value is extreme eg an age of 200 years an error is displayed and no results are returned.
To test how the precision of predictions changes with more data points, time points were added sequentially, the concentration time curve was estimated, and compared to the measured concentrations at the same time points. The mean absolute percentage difference between observed and predicted were calculated for each prediction run. For each dosing occasion, the AUC for the prediction run with the lowest and highest mean absolute percentage error (MAPE) were compared.
Results:
The output from the app was compared to the original model with very similar prediction results. All validations added on input values were tested extensively to make sure the appropriate warning were displayed. The user interface of the app was shown to multiple clinical pharmacists and will be user tested more formally in the future.
The same data that was used to develop the model was used to test the sensitivity of the predictions and AUC. The data set had 46 patients (300 occasions, 943 blood levels), resulting in 943 MAP predictions. When looking at data from all occasions, there was no clear pattern of how the MAPE changed as the number of samples used in the prediction increased. For example, MAPE was 36, 18.4, and 19.9 when one, three and eight samples were used respectively. The relative difference in the AUC between the worst and the best prediction run (highest and lowest MAPE, respectively), was on average 1.2% higher for the best run than the worst run with an interquartile range of -2.5% lower to 3.1% higher.
Conclusions:
It is possible to build a therapeutic drug monitoring tool using only open source tools. No clear relationship between prediction precision and the number of levels used in the prediction was found. There was some variation in the the AUC between the best and the worst prediction run but the difference seems to be too small to be clinically relevant. Future work includes rerunning the predictions with new data that was not used in the development of the model.
References:
[1] Chang W, Cheng J, Allaire J, Xie Y, McPherson J. Shiny: Web application framework for r. 2019. Available from: https://CRAN.R-project.org/package=shiny
[2] R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2019. Available from: https://www.R-project.org/
[3] Johansson ÅM, Hill N, Perisoglou M, Whelan J, Karlsson MO, Standing JF. A population pharmacokinetic/pharmacodynamic model of methotrexate and mucositis scores in osteosarcoma. Therapeutic Drug Monitoring. 2011 Dec;33(6):711–8. doi: 10.1097/FTD.0b013e31823615e1
Reference: PAGE () Abstr 9516 [www.page-meeting.org/?abstract=9516]
Poster: Methodology - Other topics