Victoria Chan (1), Itziar Irurzun Arana (2), and Martin Johnson (2).
NULL
Objectives:
Neutropenia is a common dose-limiting toxicity in oncology patients following chemotherapy or targeted therapies. The Friberg PKPD model [1] has been established as the gold-standard model to describe drug induced myelosuppression. However, large feedback parameter (γ) estimates of this model can cause stability issues which result in non-physiological predictions due to the magnitude of oscillations and/or time to return to equilibrium [2]. In addition, considering a single chain of compartments to describe haematological toxicity may not be always sufficient and more mechanistic QSP models are being developed to describe such complex behaviours.
When modelling aims to predict and not to have a mechanistic understanding of the system, deep learning methods [3,4] have demonstrated to be a promising and alternative tool to PKPD models, with the ability to have higher accuracy in the predictions while eliminating the burden of structural model development.
This proof-of-concept study hypothesised that Long Short-Term Memory (LSTM) neural networks could accurately predict neutrophil time-course undergoing drug treatment based on a series of exposure timepoints and baseline characteristics.
Methods:
For this proof-of-concept study, we generated time course of plasma concentrations and neutrophil count data for a hypothetical small molecule using the Friberg model. We simulated four different doses in 21-day cycles for a total of 10 cycles with 1000 patients per dose. No residual variability was added in this first step of this project. These simulated data were split into training (64%), validation (16%) and test (20%) sets and used for model development, model evaluation and model performance purposes, respectively.
Simulated time-series of Cmax, as well as patient baseline covariates (e.g. gender, ecog status, prior-chemotherapy, neutrophil count at baseline), were provided as inputs to the LSTM layer and further passed onto the Feed-Forward (FF) layer, then to the output layer, which predicted neutrophil counts. 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.
Goodness of fit plots and mean-absolute-percentage-error (MAPE) were used to assess the adequacy of model predictions.
Results:
The model architecture consisted of a three-layered LSTM model sequentially connected to a FF dense layer and a single node output layer.
For the base model, which only included time-series of Cmax values as input, neutrophil counts over time were forecasted with a 36.5% MAPE for the training dataset. The addition of covariates improved model performance: goodness of fit plots revealed an adequate predictive performance of the neutrophil count data for training and test datasets, and MAPE values of 6% for the training dataset and 5.23% for the test dataset showed good accuracy.
Evaluation of model performance in additional doses not included in the train/validation/test datasets also showed very good accuracy with <10% MAPE. However, testing the algorithm on simulations made with 42-day cycles did not show adequate results, with MAPE values increasing to 30%. Therefore, for scenarios which were not used to train the model, the current LSTM-based model predictions might not be accurate.
Conclusions:
Our results support deep learning approaches as promising forecasting tools for adverse events in oncology. Future work will involve adding residual error variability to the hypothetical dataset generated for this work to have a more realistic representation of real patient data as well as training the algorithm with varying cycle length data.
References:
[1] Friberg et al, Journal of Clinical Oncology. 2002, 20(24):4713-21.
[2] Fornari et al. CPT Pharmacometrics Syst Pharmacol. 2020, 9(9):498-508.
[3] Lee HC et al. Anesthesiology. 2018 128, (3), 492-501.
[4] Liu X et al. Int J Clin Pharmacol Ther. 2021, 59, (2), 138-146.
Reference: PAGE 30 (2022) Abstr 10086 [www.page-meeting.org/?abstract=10086]
Poster: Methodology – AI/Machine Learning