Fractional-Order Model Predictive Control for Neurophysiological Cyber-Physical Systems: A Case Study using Transcranial Magnetic Stimulation

Fractional-order dynamical systems are used to describe processes that exhibit temporal long-term memory and power-law dependence of trajectories. There has been evidence that complex neurophysiological signals like electroencephalogram (EEG) can be modeled by fractional-order systems. In this work, we propose a model-based approach for closed-loop Transcranial Magnetic Stimulation (TMS) to regulate brain activity through EEG data. More precisely, we propose a model predictive control (MPC) approach with an underlying fractional-order system (FOS) predictive model. Furthermore, MPC offers, by design, an additional layer of robustness to compensate for system-model mismatch, which the more traditional strategies lack. To establish the potential of our framework, we focus on epileptic seizure mitigation by computational simulation of our proposed strategy upon seizure-like events. We conclude by empirically analyzing the effectiveness of our method, and compare it with event-triggered open-loop strategies.

💡 Research Summary

The paper presents a novel closed‑loop transcranial magnetic stimulation (TMS) framework for mitigating epileptic seizures by leveraging fractional‑order system (FOS) models within a model predictive control (MPC) scheme. Recognizing that electroencephalogram (EEG) signals exhibit long‑range temporal dependencies and power‑law behavior, the authors adopt a discrete‑time fractional‑order state‑space representation Δ^α xₖ₊₁ = A xₖ + wₖ, where the vector α encodes the memory depth of each channel and A captures spatial coupling among electrodes. This modeling choice reduces the number of parameters compared with traditional multivariate autoregressive (MV‑AR) models, while still faithfully representing the underlying long‑memory dynamics.

System identification is performed on a 10‑second ictal segment (sampled at 160 Hz) from subject 11 of the CHB‑MIT scalp EEG database. The identified matrices A and α (four‑channel case) reveal fractional exponents ranging from 0.66 to 1.07, indicating heterogeneous memory characteristics across channels. Process noise is modeled as additive white Gaussian noise with variance σ²_w = 0.2.

To embed the FOS into an MPC framework, the authors convert the fractional model into a p‑step augmented linear time‑invariant (LTI) system (˜A, ˜B). This enables the use of standard quadratic programming solvers for the finite‑horizon optimal control problem. The cost function penalizes state energy (Q = I) and stimulation effort (R = εI, ε > 0) to simultaneously drive the brain activity toward a low‑energy regime and limit the intensity of the magnetic pulses for safety. Constraints on the control input enforce clinically acceptable voltage bounds.

The MPC operates with a prediction horizon P and a shorter control horizon M; at each sampling instant the optimizer computes a sequence of future inputs, but only the first M inputs are applied before the horizon is shifted and the problem is re‑solved. This receding‑horizon strategy introduces robustness against model‑plant mismatches and external disturbances.

Three stimulation strategies are compared in simulation: (1) a pure open‑loop protocol that delivers a pre‑programmed 16 Hz sinusoid (amplitude 0.5) irrespective of the brain state; (2) an event‑triggered open‑loop approach that activates the same waveform upon detection of seizure onset; and (3) the proposed FOS‑MPC closed‑loop controller. The open‑loop methods fail to suppress the seizure dynamics and can even exacerbate them, largely because they ignore real‑time EEG feedback. The event‑triggered scheme improves performance modestly but suffers from detection latency and limited adaptability. In contrast, the FOS‑MPC continuously adapts the stimulation waveform based on the estimated state, achieving rapid reduction of the state energy while respecting the input constraints. The controller also demonstrates resilience to stochastic disturbances modeled as Poisson‑distributed wavelet bursts of varying amplitude.

The authors discuss extensions to time‑varying fractional‑order models (Δ^α xₖ₊₁ = Aₖ xₖ + wₖ) to capture non‑stationary brain dynamics, and argue that the receding‑horizon MPC inherently compensates for such variations. They highlight that the combination of fractional‑order modeling (capturing long‑memory effects) and MPC (providing predictive, constrained, and robust control) overcomes the limitations of conventional linear models and static stimulation protocols.

In conclusion, the study provides compelling evidence that a fractional‑order model predictive control architecture can effectively and safely mitigate epileptic seizures using closed‑loop TMS. The approach outperforms traditional open‑loop and event‑triggered strategies, offers built‑in robustness to model uncertainties and external disturbances, and is readily extensible to other neuro‑stimulation applications such as Parkinson’s disease, Alzheimer’s disease, depression, and anxiety.