Data-Driven Personalization of Automated Insulin Delivery

Technical Summary: Data-Driven Personalization of Automated Insulin Delivery

Problem Statement
Automated Insulin Delivery (AID) systems currently rely on controllers tuned at the population level, often using exhaustive searches over virtual cohorts (e.g., the FDA-accepted UVA/Padova simulator). While this ensures safety across a broad population, it fails to account for significant inter- and intrapatient variability in insulin needs caused by meal patterns, physical activity, circadian rhythms, and fluctuations in insulin sensitivity. Existing adaptive approaches often depend on re-identifying physiological model parameters (e.g., insulin sensitivity), which requires high model fidelity and identifiability. Alternatively, direct adaptation methods like Extremum-Seeking Control (ESC) struggle with the noise inherent in Continuous Glucose Monitoring (CGM) and metabolic variability, often requiring aggressive filtering that slows convergence or introduces oscillations.

Methodology
The authors propose a real-time, data-driven personalization framework that adapts controller parameters directly from an individual's daily glycemic data without re-identifying the underlying physiological model.

1. Control Architecture: The system utilizes a Model Predictive Control (MPC) framework where the controller output $u$ is a function of state $x$ and tunable parameters $\theta = (\alpha, \beta)$.

   • $\beta$ scales the operating point to target a specific fasting glucose.
   • $\alpha$ adjusts the conservativeness regarding insulin stacking.
   • Unlike global scaling factors (e.g., $\gamma$ for Total Daily Insulin), these parameters allow the controller to leverage specific pharmacokinetic/pharmacodynamic (PK/PD) profiles, such as those of ultra-rapid insulin formulations.

2. Adaptation Law: The core of the approach is a Projected Gradient Descent algorithm operating on a daily risk metric $J_k$. The update rule is:
   $$\theta_{k+1} = \Pi_T (\theta_k - \eta \hat{\nabla} J_k + \xi_k)$$
   where $\Pi_T$ is the Euclidean projection onto the admissible set, $\eta$ is the learning rate, and $\xi_k$ is a bounded dither signal for persistent exploration.

3. Gradient Estimation: To handle CGM noise and metabolic variability, the authors introduce a Regularized Weighted Least Squares (RWLS) gradient estimator.

   • Spatial Proximity Weighting: Samples closer to the current operating point $\theta_k$ are weighted more heavily to approximate the local gradient.
   • Adaptive Regularization: The regularization coefficient $\rho$ is dynamically adjusted based on the weighted spatial variance of the data batch. This acts as a "dynamically scaled trust region," preventing large gradient steps when data is distant from the current estimate, thereby balancing noise attenuation against curvature bias.

4. Theoretical Validation:

   • Contraction Theory: The authors assume the closed-loop system is contractive (a property supported by the structure of T1D physiology, where insulin and glucose dynamics form a cascaded, dissipative system). This ensures that under periodic disturbances (daily routines), the system converges to a unique periodic orbit, defining a well-posed steady-state risk function $\bar{J}(\theta)$.
   • Stability: Using contraction theory, the authors prove that the closed-loop system with the adaptation law is Input-to-State Stable (ISS). They derive bounds on the learning rate $\eta$ that ensure exponential convergence to the optimal parameters under periodic disturbances and convergence to a neighborhood of the optimum under noisy, non-periodic conditions.

Key Contributions

1. RWLS Gradient Estimator: Development of a novel estimator with spatial proximity weighting and adaptive regularization specifically designed for AID personalization. The paper provides a theoretical error bound showing that the estimation error is controlled under curvature regularity and persistent excitation conditions.
2. Contraction-Based Analysis: Demonstration that under a nominal daily routine, the closed-loop glucose dynamics reduce to a periodic orbit. This establishes a theoretical foundation for analyzing the adaptation law as a projected gradient descent on a steady-state risk metric.
3. ISS Validation: Proof that the system with the performance-based adaptation scheme (PAS) is Input-to-State Stable, validated through in silico trials under realistic meal, behavioral, and metabolic variability.

Results
The method was evaluated using the 100-adult cohort of the FDA-accepted UVA/Padova T1D simulator under two scenarios: constant conditions and conditions with meal/metabolic variability.

• Risk Reduction: The proposed method reduced glycemic risk compared to non-adaptive baselines and existing $\gamma$-scaling methods.
• Time-in-Range (TIR): In the variability scenario, the method increased TIR (70–180 mg/dL) by:
  • 2% after 4 weeks.
  • 3% after 8 weeks.
  • 4% after 17 weeks.
• Comparison: Adapting the specific parameters $(\alpha, \beta)$ yielded lower risk and higher TIR compared to adapting only the global scaling factor $\gamma$. A hybrid approach (initially adapting $\gamma$, then switching to $(\alpha, \beta)$) showed convergence to the lowest overall risk.
• Robustness: The system maintained stability and improved performance despite variations in meal timing ($\pm 20$ minutes), meal size, and insulin sensitivity.

Significance and Claims
The paper claims that this approach offers a robust, model-free alternative to physiological parameter identification for personalizing AID systems. By leveraging contraction theory, the authors ensure that the adaptation process does not destabilize the closed-loop system, even in the presence of significant metabolic noise. The method is particularly significant for integrating ultra-rapid insulin formulations, where simple global scaling of Total Daily Insulin (TDI) is insufficient to capture altered PK/PD temporal profiles. The authors emphasize that their framework allows controllers to safely reduce conservativeness regarding insulin stacking, thereby curbing postprandial peaks more aggressively while mitigating late-onset hypoglycemia, all while maintaining input-to-state stability.