Landscape and future directions of machine learning applications in closed-loop brain stimulation

Abstract

Brain stimulation (BStim) encompasses multiple modalities (e.g., deep brain stimulation, responsive neurostimulation) that utilize electrodes implanted in deep brain structures to treat neurological disorders. Currently, BStim is primarily used to treat movement disorders such as Parkinson’s, though indications are expanding to include neuropsychiatric disorders like depression and schizophrenia. Traditional BStim systems are “open-loop” and deliver constant electrical stimulation based on manually-determined parameters. Advancements in BStim have enabled development of “closed-loop” systems that analyze neural biomarkers (e.g., local field potentials in the sub-thalamic nucleus) and adjust electrical modulation in a dynamic, patient-specific, and energy efficient manner. These closed-loop systems enable real-time, context-specific stimulation adjustment to reduce symptom burden. Machine learning (ML) has emerged as a vital component in designing these closed-loop systems as ML models can predict / identify presence of disease symptoms based on neural activity and adaptively learn to modulate stimulation. We queried the US National Library of Medicine PubMed database to understand the role of ML in developing closed-loop BStim systems to treat epilepsy, movement disorders, and neuropsychiatric disorders. Both neural and non-neural network ML algorithms have successfully been leveraged to create closed-loop systems that perform comparably to open-loop systems. For disorders in which the underlying neural pathophysiology is relatively well understood (e.g., Parkinson’s, essential tremor), most work has involved refining ML models that can classify neural signals as aberrant or normal. The same is seen for epilepsy, where most current research has focused on identifying optimal ML model design and integrating closed-loop systems into existing devices. For neuropsychiatric disorders, where the underlying pathologic neural circuitry is still being investigated, research is focused on identifying biomarkers (e.g., local field potentials from brain nuclei) that ML models can use to identify onset of symptoms and stratify severity of disease.

Model-Based Closed-Loop Control of Thalamic Deep Brain Stimulation

“…Golshan et al (2018) [71] utilized the wavelet coefficients of STN-LFP beta frequency range as features and further developed a support vector machine (SVM) classifier for the behaviors of PD patients. Numerous prior studies have established high performance using SVM classifiers with input features such as phase amplitude coupling [25], Hjorth parameters [72], beta band power [25], and burst duration [70]. Using power densities within the beta [69] and gamma [73] bands as features, Hidden Markov Models [70][74][73], SVM [71], convolutional neural networks (CNN) [70][75][72], linear discriminant analysis (LDA) [70], and logistic regression [76] were performed.…”

“…In recent closed-loop DBS control systems, machine learning methods have been developed to map biomarker features (input) to patients' observed states (output) and could further deliver an appropriate DBS setting [25][69] [70]. Therefore, it is imperative to identify key biomarkers that need to be extracted from the LFP and EMG as they will serve as input features for a judiciously Hjorth parameters [72], beta band power [25], and burst duration [70]. Using power densities within the beta [69] and gamma [73] bands as features, Hidden Markov Models A key improvement to existing machine learning methods is to incorporate physiological characterizations of the input-output mapping.…”

“…In recent closed-loop DBS control systems, machine learning methods have been developed to map biomarker features (input) to patients' observed states (output) and could further deliver an appropriate DBS setting [25][69] [70]. Therefore, it is imperative to identify key biomarkers that need to be extracted from the LFP and EMG as they will serve as input features for a judiciously Hjorth parameters [72], beta band power [25], and burst duration [70].…”

“…ζ(𝑥(𝑡)) is the standardization of 𝑥(𝑡); 𝑥̅ (𝑡) and 𝑠𝑑[𝑥(𝑡)] are the mean and standard deviation of 𝑥(𝑡), respectively. The polynomial order is 25 and {𝜑 0 , 𝜑 showed that the consistent model parameters fitted based on concatenated DBS frequencies in a certain range -in this case, [10,200] Hz -can be consistently applied to unobserved frequencies (e.g., 25 Hz and 180 Hz) in the same range [31]. Thus, we implement the polynomial estimated EMG as the biomarker to control the DBS frequencies in the range [10,200] Hz.…”

“…Regardless of the type of the feedback biomarker, DBS setting is determined based on neural (e.g., LFP) or behavioral (e.g., EMG) signals in the most existing closed-loop DBS controllers [19][17][21][25][26][27]. However, these methods suffer from the lack of mechanistic understanding of the physiological mechanisms underlying the DBS and disease-related neuronal circuits.…”