AI/ML for Predictive Closed-Loop Neuromodulation in HW-Constrained Architectures 

Neuromodulation and BCI have demonstrated clinical efficacy across epilepsy, depression, chronic pain, and movement disorders, yet therapeutic outcomes vary considerably between individuals and over time. Fixed stimulation protocols do not fully account for individual neurophysiology or changes in neural state that may influence treatment response. A central challenge is therefore to identify when stimulation is likely to produce the intended effect and use this information to personalise its timing and parameters.

Advances in implantable neural sensing interfaces, including iEEG, SEEG, DBS, and imec Neuropixel technology, now enable continuous, high-resolution access to neural circuit dynamics. Coupled with emerging bioelectronic medicine platforms for stimulation, these technologies create a foundation for truly adaptive closed-loop systems. Realising this potential requires models that translate ongoing recordings into predictions of stimulation response and actionable control decisions. This requires research into AI/ML methods that combine predictive capabilities within the hardware requirements of implantable systems, including memory, computational, energy, and latency constraints.

This PhD research will design AI/ML frameworks that extract neurophysiological biomarkers from brain and peripheral nerve recordings, model neural excitability and network dynamics, and optimize stimulation parameters in real time. The work will focus on identifying latent neural states predictive of therapeutic response and designing adaptive control strategies that personalize therapy at the individual level. The suitability of deploying the identified methods in imec-defined closed-loop neuromodulation or BCI implantable systems will also be assessed, looking at the algorithmic complexity and resource and computation time requirements.

The research will leverage imec's expertise in implantable neurotechnology and high-precision bi-directional neural interfaces, as well as available recording and stimulation hardware, bridging fundamental neuroscience, machine learning, and translational medicine through both preclinical and clinical datasets.

Concrete research questions

1. Neurophysiological Biomarker Discovery and Response Prediction

Design machine-learning models to identify predictive neural and physiological biomarkers and predict individual responses to neurostimulation, using invasive and non-invasive recordings, which are suitable to deploy in specific HW-constrained imec system architectures. Outcomes will include immediate evoked responses and stimulation-induced (neuro)physiological changes, with longer-term treatment responses investigated in suitable clinical datasets.

Determine how ongoing neural and physiological states influence stimulation responsiveness within individuals, and identify whether particular states indicate periods when stimulation is more likely to produce the intended effect. 

Design explainable models to identify neurophysiological characteristics associated with response variability and generate hypotheses for experimental investigation.

Investigate whether combining neural recordings with peripheral physiological signals and clinical information improves stimulation-response prediction, and determine which sources provide additional predictive value. 

2. Hardware-Aware AI for Personalised Closed-Loop Neuromodulation

Investigate data-efficient personalisation methods that adapt population models using limited individual recordings and stimulation responses, such as: subject-specific calibration, transfer learning, few-shot learning, and adaptation of pretrained foundation models.

Evaluate implementation complexity and timing requirements of deploying these models in specific HW-constrained architectures, identifying cost-benefit tradeoffs and best-suited methods to integrate.

Design predictive controllers that use individual response models and measured feedback to select stimulation timing and parameters, using computational models of neural responses or exploratory preclinical studies. 

Investigate predictive neuromodulation strategies that anticipate pathological network states before symptom onset.

3. Experimental Evaluation and Clinical Relevance

From one of the use cases listed below, establish predictive validity using existing preclinical and clinical recordings, with evaluation on unseen individuals and future sessions, and distinguishing prediction of immediate stimulation responses from prediction of longer-term stimulation treatment benefit. 

Evaluate the feasibility of deploying some of the closed-loop models in existing imec HW platforms, using synthetic input. 

Optional Extension: Conditional Generative Modelling of Stimulation Responses

Investigate conditional diffusion models that generate neural or physiological responses given preceding activity and stimulation parameters, evaluating whether they capture response variability beyond predictions of the average response.

Determine whether synthetic training data improve response prediction and personalisation when individual recordings are limited, comparing generative augmentation with conventional augmentation and models trained exclusively on real data.

Use Cases

Primary application: Preclinical neuromodulation for epilepsy

Evaluate adaptive stimulation strategies for suppressing epileptic activity, using vagus nerve stimulation, transcranial magnetic stimulation, or optogenetic stimulation.

Potential other applications include:

Neuromodulation for depression and other affective disorders.

Peripheral nerve stimulation for other therapeutic indications.

Closed-loop brain–computer interfaces incorporating stimulation.

Imec roadmap

This research will contribute to the development of intelligent bioelectronic medicine and next-generation closed-loop neuromodulation or BCI systems. By combining advanced AI/ML techniques with emerging neural interface technologies, and addressing limitations existing and envisioned HW systems are imposing on these techniques, the project aims to move beyond conventional open-loop stimulation paradigms towards predictive, personalized, and explainable therapies that can be part of future intelligent neurotech clinical solutions.

The work directly supports imec's roadmap in neurotechnology, implantable devices, brain-computer interfacing, neuropixel, and peripheral nerve stimulation. In the long term, it will enable the development of autonomous neuromodulation systems capable of sensing neural states, predicting pathological events, and delivering optimized interventions in real time, ultimately improving patient outcomes while advancing our understanding of brain and nervous system function.