Artificial Intelligence Techniques for High-Precision Invasive and Non-invasive Neural Stimulation

The ability to systematically design stimuli that can invoke desired neural responses is one of the main challenges of neural engineering. The process of altering neural activity through targeted delivery of stimuli (e.g., electrical, magnetic, chemical, ultrasound) is referred to as neuromodulation. Neuromodulation therapies have shown promise in treating several neural disorders, such as alleviating Parkinsonian symptoms [1], facilitating stroke rehabilitation [2], regulating depression [3], and more [4]. The field of neuromodulation is one of the fastest-growing areas of medicine, impacting millions of patients [5]. Electrical currents are the most commonly used stimuli among the various stimuli modalities for controlling neural activity [6]. Therefore, it is essential to develop systematic approaches for designing electrical stimuli to enhance the efficacy of neuromodulation therapies in treating neural disorders.

Traditionally, electrical stimuli have been designed using domain-specific heuristics and traditional computational neuron models (e.g., Hodgkin-Huxley neuron models [7]). The accuracy of the underlying neural models limits the performance of these model-based approaches. For example, model-based approaches can be inflexible, as model-based approaches utilizing a particular type of neuron model (e.g., FitzHugh-Nagumo model [8]) cannot be easily extended to other neuron model types (e.g., Hodgkin-Huxley models). Furthermore, models can be inaccurate for the task of designing neural stimuli, and it is not a priori apparent that the model being utilized is inaccurate for the purpose of designing stimuli (as we demonstrate in Chapter 2). In certain instances, neural models may not be available altogether due to a paucity of experimental data (see Chapter 8). As for hand-designing electrical stimuli using domain-specific heuristics, it works well when the number of stimuli parameters is small (e.g., 1 or 2 parameters) but quickly becomes unfeasible as the number of parameters increases.

To address the above limitations of traditional stimuli design, this thesis explores the complementary approach of designing stimuli directly from data (without explicitly relying on neural models). Inspired by the recent success of artificial intelligence (AI) and machine learning (ML) tools in numerous fields (such as, image processing, language modeling, robotics, and protein folding), this thesis develops AI and ML techniques for designing electrical stimuli that evoke desired neural activity in the brain. Designing electrical stimuli can be conceptually divided into two parts: (i) designing where the current is flowing (the spatial aspect) and (ii) designing how the current changes temporally across the duration of stimulation (the temporal aspect).

In this thesis, we propose PATHFINDER and its extensions in Chapters 3 and 4 to design the temporal aspect of electrical stimuli directly from data. PATHFINDER utilizes a novel optimization framework to estimate a pseudoinverse of the forward mapping relating the stimuli parameters to the neural responses directly from data. The stimuli parameters are designed using the estimated pseudoinverse without explicitly relying on traditional neural models. Furthermore, through computational experiments, we demonstrate that PATHFINDER is more data efficient than existing pseudoinverse estimation techniques, which is crucial in the relatively data-scarce field of neural engineering. Using the PATHFINDER optimization framework, we also propose novel adaptive sampling techniques and dimensionality reduction techniques in Chapter 4, which can further increase the data-efficiency of PATHFINDER in designing electrical stimuli.

For designing the spatial aspect of electrical stimuli, we propose the HingePlace algorithm and its extensions in Chapters 6 and 8. A common way to inject currents into the brain is to place electrodes in the brain through neurosurgery invasively. Neurosurgery carries significant risks, and it is desirable to stimulate the brain through non-invasive means. A common alternative is transcranial electrical stimulation (tES), which injects current through electrodes placed on the scalp. A particular drawback of tES is the dispersion of the current due to the current traveling through the layers of the head, resulting in stimulation of off-target brain regions. The HingePlace algorithm and its extensions utilize convex optimization tools to design appropriate electrode montages that minimize this off-target neural activation. Through extensive computational (Chapter 6) and experimental studies in rodents and monkeys (Chapter 7), we demonstrate that HingePlace elicits focused neural stimulation. We further demonstrate that the HingePlace-designed electrode montages elicit lower scalp pain in human studies with appropriate regularization (Chapter 8).

Finally, we discuss how AI algorithms can help discover novel biophysical phenomena in Chapters 5 and 9. Chapter 5 shows how the AI algorithms revealed a surprising observation regarding charge/energy efficient waveforms for evoking muscle activity, namely, the temporal shape of charge/energy efficient waveforms depends on the spatial shape of the stimulating field. Similarly, Chapter 9 discusses how AI tools helped discover electrically skull-transparent electrode arrangements whose shape of the induced electric field is not affected by changing conductivities and thickness of different head layers. These results demonstrate that AI tools have benefits beyond their intended use and can lead to a deeper understanding of the mechanisms underlying neuromodulation mechanisms by discovering novel neural responses.