Developing hybrid mechanistic and data-driven personalized prediction models for platelet dynamics

TL;DR

Chemotherapy-induced thrombocytopenia presents strong inter-patient variability, challenging personalized risk prediction. The authors compare hybrid mechanistic-UDE frameworks based on the Friberg hematopoiesis model with purely data-driven ARX-GRU/NARX approaches across diverse data regimes, using a NHL-B patient dataset. They find that data-driven ARX-GRU predictions excel when data are plentiful, especially for high-toxicity trajectories, while hybrid and mechanistic methods provide robust performance under data sparsity and help preserve physiological plausibility. The work offers a generalizable framework for individualized prediction of treatment-related toxicities and provides public code to extend the approach to other hematopoietic toxicities and interventions.

Abstract

Hematotoxicity, drug-induced damage to the blood-forming system, is a frequent side effect of cytotoxic chemotherapy and poses a significant challenge in clinical practice due to its high inter-patient variability and limited predictability. Current mechanistic models often struggle to accurately forecast outcomes for patients with irregular or atypical trajectories. In this study, we develop and compare hybrid mechanistic and data-driven approaches for individualized time series modeling of platelet counts during chemotherapy. We consider hybrid models that combine mechanistic models with neural networks, known as universal differential equations. As a purely data-driven alternative, we utilize a nonlinear autoregressive exogenous model using gated recurrent units as the underlying architecture. These models are evaluated across a range of real patient scenarios, varying in data availability and sparsity, to assess predictive performance. Our findings demonstrate that data-driven methods, when provided with sufficient data, significantly improve prediction accuracy, particularly for high-risk patients with irregular platelet dynamics. This highlights the potential of data-driven approaches in enhancing clinical decision-making. In contrast, hybrid and mechanistic models are superior in scenarios with limited or sparse data. The proposed modeling and comparison framework is generalizable and could be extended to predict other treatment-related toxicities, offering broad applicability in personalized medicine.

Figures (3)

• Figure 1: Combining mechanistic and data-driven models. Mechanistic models and data-driven approaches alone fail to perform well for individual sparse patient data (a). To deal with this problem, we employ transfer learning and hybrid modeling approaches, where both, individual patient dynamics and existing mechanistic knowledge is combined in model training (b). Results showed a strong advantage of the data-driven ARX-GRU approach, if sufficient data is provided for training (c).
• Figure 2: Average prediction performance per patient group and data scenario. We compare performances of the different modeling approaches (rows) using SMSE (see \ref{['eqn:SMSE']}) averaged over all patients of the respective groups (panels) per number of therapy cycles used for training (columns). The SMSE is calculated for each patient using data and corresponding model predictions from all cycles not used for model training. The average of SMSE across all patients is calculated and compared between groups. To emphasize differences in SMSE across modeling approaches, we use a color scale where lighter colors represent lower SMSE values, indicating better prediction performance. The best model per column is indicated by a blue box and SMSE values not significantly inferior compared to the best performance per column are marked by an asterisk.
• Figure 3: Individual predictions for four selected patients. We present predictions of the Friberg (blue), Henrich(orange, dashed), ARX-GRU (red, dash-dotted) and UDE-add (pink, dotted) model for NHL patients 4, 7, 114 and 1678. The split between training and prediction time points is indicated by a vertical black line. Data points are shown as black circles. For improved legibility, we use different scales per patient (rows). While patients 4, 114 and 1678 are from group De14, patient 7 belongs to group Sp21.