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

Area of Science:

• Oncology
• Pharmacology
• Computational Biology

Background:

• Thrombocytopenia is a frequent dose-limiting side effect of cytotoxic chemotherapy.
• Accurate prediction of individual thrombocytopenia risk is clinically significant for optimizing chemotherapy dosages.

Purpose of the Study:

• To develop and evaluate non-linear auto-regressive networks with exogenous inputs (NARX) for predicting individual platelet dynamics during chemotherapy.
• To compare the performance of different NARX architectures and transfer learning (TL) approaches against a semi-mechanistic model.

Main Methods:

• Utilized NARX networks, including feed-forward networks (FNN) and gated recurrent units (GRU), to model platelet dynamics.
• Employed transfer learning (TL) with a semi-mechanistic hematotoxicity model to address sparse individual patient data.
• Trained and validated models on a large dataset of patients with high-grade non-Hodgkin's lymphoma.

Main Results:

• The NARX network with a GRU architecture demonstrated superior prediction performance.
• The NARX model significantly improved prediction accuracy, particularly for patients with irregular platelet dynamics and well-spaced measurements.
• Transfer learning enhanced individual prediction capabilities.

Conclusions:

• NARX networks offer a viable approach for predicting individual responses to chemotherapy-induced thrombocytotoxicity.
• Recommends at least three well-spaced measurements per cycle (baseline, nadir, recovery) for effective model learning.
• Future work aims to generalize the approach to other treatments and blood cell types.