Global Deep Forecasting with Patient-Specific Pharmacokinetics

TL;DR

The paper tackles the challenge of forecasting blood glucose by incorporating patient-specific pharmacokinetics into deep learning forecasts. It introduces a PK encoder within a hybrid global-local architecture, enabling explicit modeling of time-varying drug concentration effects from sparse dosing data and multiple insulin doses under linear PK assumptions. Across simulated and real-world OhioT1DM data, PK-augmented models (notably TFT-PK, NBEATSx-PK, and NHITS-PK) consistently outperform baselines, with hybrid global-local variants delivering the strongest gains and clear evidence of learning treatment effects. The work offers a practically impactful approach for early warnings of hypo/hyperglycemia and personalized dosing insights, while outlining limitations and directions for extending to nonlinear PK and broader patient populations.

Abstract

Forecasting healthcare time series data is vital for early detection of adverse outcomes and patient monitoring. However, it can be challenging in practice due to variable medication administration and unique pharmacokinetic (PK) properties of each patient. To address these challenges, we propose a novel hybrid global-local architecture and a PK encoder that informs deep learning models of patient-specific treatment effects. We showcase the efficacy of our approach in achieving significant accuracy gains in a blood glucose forecasting task using both realistically simulated and real-world data. Our PK encoder surpasses baselines by up to 16.4% on simulated data and 4.9% on real-world data for individual patients during critical events of severely high and low glucose levels. Furthermore, our proposed hybrid global-local architecture outperforms patient-specific PK models by 15.8%, on average.

Figures (15)

• Figure 1: Insulin dose is recorded, but plasma insulin concentration is not directly measurable and typically requires invasive procedures and laboratory tests.
• Figure 2: Pharmacokinetic models of plasma insulin concentration over time for various insulin types.
• Figure 3: Medication administrations are represented as sparse variables in time series data. We propose a pharmacokinetic (PK) encoder to effectively capture time-dependent plasma drug concentration (Conc.).
• Figure 4: Our hybrid global-local architecture combines global model parameters shared across patients with patient-specific pharmacokinetic (PK) parameters. Sparse treatment time series features, such as medication doses, are input into the PK encoder to generate patient-specific treatment concentration curves $\check{\textbf{x}}$. PK encoder output $\check{\textbf{x}}$ and any other exogenous features $\textbf{x}$ are used as deep learning model inputs. During each training step, the model learns relevant hidden states and output parameters $\hat{\theta}$ across patients, while $\textbf{k}^{(i)}$ values are updated for each patient $i$ using gradient descent optimization.
• Figure 5: The hybrid global-local NHITS-PK model trained on all OhioT1DM individuals (orange, solid median) has a lower MAE than NHITS-PK local models (blue, dashed median).

Theorems & Definitions (2)

• Proposition 1
• Proposition 2