Predicting chemotherapy-induced thrombotoxicity by NARX neural networks and transfer learning

Thrombocytopenia is a common side effect of cytotoxic chemotherapies, which is often dose-limiting. Predicting an individual’s risk is of high clinical importance, as otherwise, a small subgroup of patients limits dosages for the overall population for safety reasons.

We aim to predict individual platelet dynamics using non-linear auto-regressive networks with exogenous inputs (NARX). We consider different architectures of the NARX networks, namely feed-forward networks (FNN) and gated recurrent units (GRU). To cope with the relative sparsity of individual patient data, we employ transfer learning (TL) approaches based on a semi-mechanistic model of hematotoxicity. We use a large data set of patients with high-grade non-Hodgkin’s lymphoma to learn the respective models on an individual scale and to compare prediction performances with that of the semi-mechanistic model.

Of the examined network models, the NARX with GRU architecture performs best. In comparison to the semi-mechanistic model, the network model can result in a substantial improvement of prediction accuracy for patients with irregular dynamics, given well-spaced measurements. TL improves individual prediction performances.

NARX networks can be utilized to predict an individual’s thrombotoxic response to cytotoxic chemotherapy treatment. For reasonable model learning, we recommend at least three well-spaced measurements per cycle: at baseline, during the nadir phase and during the recovery phase. We aim at generalizing our approach to other treatment scenarios and blood lineages in the future.