Anisia Talianu(1,2), Martin Johnson(1)
(1) Clinical Pharmacology Quantitative Pharmacology, BioPharmaceuticals R&D, AstraZeneca R&D, Cambridge, UK, (2) Imperial College London, UK.
Introduction: Mechanistic pharmacokinetic-pharmacodynamic (PKPD) modelling of longitudinal data is essential to understand changes in the physiological system due to therapeutic interventions. The applications and effectiveness of PKPD models as tools for informed decision making in the drug development process is well documented. Artificial neural networks [1-3]and deep learning methods [4, 5]have been explored as promising and alternative tools to predict PK and PD changes with higher accuracy than PKPD models. We aim to assess whether deep learning methods, including recurrent neural networks, can predict disease progression in ovarian carcinoma patients based on time-series laboratory data.
Objectives: In this proof-of-concept study, we hypothesised that Long Short-Term Memory (LSTM) neural networks could accurately predict clinical response to treatment reflected by changes in the sum of longest tumour diameter (SLD), based on a series of longitudinal biomarker data. We used the relationship between longitudinal cancer antigen 125 (CA125) and SLD in ovarian cancer after chemotherapy treatment to evaluate this hypothesis.
Methods: The prototype LSTM models were trained to predict inter visit SLD (for next clinical visit) based on time-series CA125 levels. We used a published PKPD model to simulate the time series CA125 and SLD data. Simulated time-series CA125 levels were provided as inputs to the LSTM layer and further passed onto the Feed-Forward (FF) layer, then to the output layer, which predicted SLD. Baseline SLD and time elapsed since the start of treatment with chemotherapy were included as covariates. These covariates were fed as an input to the FF network along with the CA125 sequences processed by the LSTM layers. Data were split into training (70%), validation (15%) and test (15%) sets and used for model development, model evaluation and model performance (in unseen data) purposes, respectively. Root-mean-squared-error (RMSE) and mean-percentage-error (MPE) were used as metrics to assess the adequacy of model predictions. Keras and TensorFlow software libraries, as implemented in python programming language, were used to develop this algorithm. Model hyperparameters, including the learning rate, the number of epochs and the layer activation functions, were tuned to optimise the learning process. A 5-fold cross-validation[6] method was used to identify the optimal set of hyperparameters.
Results: The model architecture presenting the smallest error consisted of a three-layered LSTM model sequentially connected to a FF dense layer made up of 8 processing nodes and a single node output layer. The optimal activation functions obtained were hyperbolic tangent, rectified linear unit, and sigmoid for the LSTM, FF, and output layers, respectively.
For the base model, inter visit SLDs were forecasted with an 8.1 mm RMSE and 31.1% MPE while using training datasets and 7.9 mm RMSE and 31.3 % MPE while using the test set. The addition of covariates improved model performance, as prediction error decreased to 4.1 mm RMSE and 17.7 % MPE on the test set, compared to the model without covariates (7.9 and 31.1%, RMSE and MPE, respectively).
The current model structure is limited with predicting the inter visit SLD only a week ahead of the visit and it requires biomarker data until a week before the visit for a CT scan. It will be valuable and efficient to forecast inter visit SLDs at least 6 to 12 weeks ahead of CT scan visit. Our ongoing efforts include exploring implementation to account for 6 to 12 weeks of forecasting.
Conclusions: Our results support the use of deep learning approaches as promising forecasting tools in oncology. The LSTM model built using patient covariates achieved predictions with adequate accuracy. Our findings suggest that the integration of extensive and informative patient features would further improve the performance of the proposed model.
References:
[1] Chow, H.H., Tolle, K.M., Roe, D.J., Elsberry, V., and Chen, H.: ‘Application of neural networks to population pharmacokinetic data analysis’, J Pharm Sci, 1997, 86, (7), pp. 840-845
[2] Gobburu, J.V., and Chen, E.P.: ‘Artificial neural networks as a novel approach to integrated pharmacokinetic-pharmacodynamic analysis’, J Pharm Sci, 1996, 85, (5), pp. 505-510
[3] Johnson, M., Metcalfe, P., Schmidt, H., Vishwanathan , K., Al-Huniti, N., Tomkinson, H., and Di Veroli, G.: ‘Artificial neural networks can facilitate the analysis and prediction of longitudinal tumour size data: an example from a non-small cell lung cancer phase III study’, PAGE 27 (2018) Abstr 8609 [www.page-meeting.org/?abstract=8609]
[4] Lee, H.C., Ryu, H.G., Chung, E.J., and Jung, C.W.: ‘Prediction of Bispectral Index during Target-controlled Infusion of Propofol and Remifentanil: A Deep Learning Approach’, Anesthesiology, 2018, 128, (3), pp. 492-501
[5] Liu, X., Liu, C., Huang, R., Zhu, H., Liu, Q., Mitra, S., and Wang, Y.: ‘Long short-term memory recurrent neural network for pharmacokinetic-pharmacodynamic modeling’, Int J Clin Pharmacol Ther, 2021, 59, (2), pp. 138-146
[6] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., and Dubourg, V.J.t.J.o.m.L.r.: ‘Scikit-learn: Machine learning in Python’, 2011, 12, pp. 2825-2830
Reference: PAGE 29 (2021) Abstr 9815 [www.page-meeting.org/?abstract=9815]
Poster: Drug/Disease Modelling - Oncology