Martin Johnson (1), Paul Metcalfe (2), Henning Schmidt (3), Karthick Vishwanathan (4), Nidal Al-Huniti (4), Helen Tomkinson (1) Giovanni Di Veroli (1)
(1) Quantitative Clinical Pharmacology, Early Clinical Development, IMED Biotech Unit, AstraZeneca, Cambridge, UK, (2) Data Science Solutions, GMD, AstraZeneca, Cambridge, UK, (3) IntiQuan GmbH, Basel, Switzerland, (4) Quantitative Clinical Pharmacology, Early Clinical Development, IMED Biotech Unit, AstraZeneca, Boston, USA
Objectives:
Artificial Neural Networks (NN) are an important component of modern Machine Learning. They have been used in a wealth of applications such as recognition algorithms, Engineering, Biology or Finance. However, the use of artificial NN in pharmacometrics is limited [1, 2]. Artificial NN are powerful learning algorithms which is capable of establishing input and output relationships for complex data where these relationships might not be obvious (artificial NN do not assume any specific structural model). We examined the applicability of artificial NN to analyse tumour progression and identify associated factors using data from a non-small cell lung cancer (NSCLC) phase III study.
Methods:
Osimertinib is an approved drug for the treatment of patients with T790M mutation positive advanced NSCLC who have progressed on or after EGFR tyrosine kinase inhibitor (TKI) therapy [3]. Data used in this analysis were collected from a phase III study (AURA3; NCT01801632; osimertinib vs chemotherapy) including patients with advanced NSCLC. Longitudinal sum of the longest diameter (SLD) collected from 257 patients (1695 SLD measurements) dosed with osimertinib were analysed. Baseline SLD, age, gender, race, metastatic status, WHO status, smoking status, information about previous therapy, lactose dehydrogenase (LDH) at baseline, albumin level (at baseline, week 6, week 12 and week 18), neutrophil to lymphocyte ratio (NLR) (at baseline, week 6, week 12 and week 18) were used as features (covariates) to predict the longitudinal changes in SLD. Time variable was included in the model as covariate. An artificial NN approach as implemented in ‘nnet’ R package was applied to predict the longitudinal SLD. In addition, functions from ‘caret’ R package were used to build the model. The available dataset (1695 observations) were randomly partitioned in 80/20 ratio as training and test dataset and artificial NN were trained using the training dataset and the model was validated using test dataset. Number of nodes (hidden units) and decay rate were tuned and root mean squared error (RMSE) used as a metric to evaluate the model. Predictor variables were then identified by order of importance. Artificial NN approaches are data driven and hence, 1000 bootstrapped datasets were used to evaluate the sensitivity of the model outcome. Artificial NN was applied on 1000 bootstrapped datasets and the top 5 predictors were identified for each dataset. Further, the predictors were selected based on the 5 most frequent predictors identified from all the bootstrapped datasets.
Results:
Following the training procedure, we obtained an artificial NN which was able to predict the SLD up to 78 weeks. The predictions from the training and test dataset were close to observation with root mean square error (RMSE) of 0.0450 and 0.0655, respectively. Based on the 1000 bootstrapped datasets, baseline SLD, NLR at week 6 and 18, LDH at baseline, and age at baseline were identified as the 5 most influential predictors for changes in longitudinal SLD. Albumin levels and systemic drug exposure were also identified as potential predictors, but to a lesser extent. Artificial NN identified predictors could be potentially tested as covariates in a pharmacometric model describing changes in longitudinal SLD over time. Because the model was built using SLD up to 78 weeks, the outcome beyond 78 weeks may be unreliable.
Conclusions:
Our artificial NN model is able to predict the trends in SLD up to 78 weeks. Baseline SLD, NLR at weeks 6 and 18, LDH at baseline, and age at baseline were identified as the most influential predictors for changes in longitudinal SLD. We plan to further investigate the generalisation of this result by analysing larger number of clinical data via artificial NN.
References:
[1] Mok TS, Wu Y-L, Ahn M-J, Garassino MC, Kim HR, Ramalingam SS, et al. Osimertinib or Platinum–Pemetrexed in EGFR T790M–Positive Lung Cancer. New England Journal of Medicine 2017 376 (7) : 629 – 40.
[2] Chow HH, Tolle KM, Roe DJ, Elsberry V, Chen H. Application of Neural Networks to Population Pharmacokinetic Data Analysis. Journal of Pharmaceutical Sciences 1997 86 (7) : 840 – 846.
[3] Gobburu JV and Chen EP.Artificial Neural Networks as a Novel Approach to Integrated Pharmacokinetic-Pharmacodynamic Analysis J Pharm Sci. 1996 May; 85 (5) : 505 – 510.
Reference: PAGE 27 (2018) Abstr 8609 [www.page-meeting.org/?abstract=8609]
Poster: Drug/Disease Modelling - Oncology