Development of artificial neural networks for the prediction of outlying and influential individuals from pharmacokinetic and pharmacodynamic models

Detta är en Master-uppsats från Uppsala universitet/Institutionen för farmaci

Sammanfattning: Nonlinear Mixed effect models are often used to describe population pharmacokinetics (PK) and Pharmacodynamics (PD) and play an important part of drug development both from regulatory and industry point of view. However, they can be time consuming and computationally expensive to develop. This thesis is a part of a larger collaboration between Uppsala University and two pharmaceutical companies, with the aim to develop a suite of software that can automate the model building process with more efficiency. One aspect that is important during the model building process is to detect how much the population parameter estimates are influenced by particular individuals. The results of this might lead to reconsideration of the model structure, as well as exclusion of these individuals from the dataset. The current tools available to detect this use case deletion diagnostics (CDD) to run the model multiple times with each subject removed from the dataset to examine whether the population estimates alter when that individual is removed. Another important aspect is whether an individual is an outlier from the population parameter predictions, which is obtained from simulating the model and evaluating the residuals (simeval). Both of these tools are computationally expensive and can take a lot of time, in particular CDD. Therefore, we developed a tool using machine learning (ML) algorithms that can predict these individuals based on other criteria, which will decrease the runtime in an automated model building procedure, whilst maintaining the robustness of the current methods described above.  To create a training database for the ML models, predictors were extracted from 27 previously published models and the CDD and simeval diagnostic tools were run on these models to obtain that true values we want the ML model to predict. The database was then used to train two artificial neural networks (ANN) which is an efficient and powerful method in ML. To enable ‘on-the-fly’ predictions, the developed ANN models were deployed using tflite into pharmpy. The resulted ANNs were able to predict outlying individuals with 79% sensitivity, 83% precision, and 99.1% specificity. While the influential individuals ANN was able to predict with 58% sensitivity, 63% precision and 99.6% specificity. Both ANNs offered a rapid assessment of influential individuals and outlying individuals and were able to make predictions in a matter of sub-seconds compared to hours using traditional methods. 

  HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)