On the need for adaptive learning in on-demand Deep Brain Stimulation for Movement Disorders.

The results presented in this paper indicate that future on-demand Deep Brain Stimulation (DBS) systems for chronic use in patients with movement disorders should continuously and adaptively "learn" in order to maintain high symptom control efficacy. In this work, two machine learning algorithms-Decision Tree and LArge Memory STorage And Retrieval (LAMSTAR) neural network, both with surface Electromyography and accelerometry as control signals-are used to predict onset of tremor after DBS has been switched off in two patients, one suffering from Parkinson's disease and the other from essential tremor. The novelty of this work is that training and testing are done by using different data recorded during sessions at least one week apart. The question is whether the applied algorithms are robust to long-term operation (as patient's control signal may change over time due to disease progression, displacement of the wearable sensor, etc.). Various metrics are used to compare the performance of the proposed approach to those available in the literature, where training and testing are done on data from the same recording session. It is shown that a 100% sensitivity is achieved for training and testing over the same session; however, the sensitivity reduces when tested over a different session. The ratio of predicted stimulation-off time to observed stimulation-off time value is also found to be lower when training and testing on data from separate sessions. These results point to the need of adaptive learning in on-demand DBS systems.

Who May Benefit From On-Demand Control of Deep Brain Stimulation? Noninvasive Evaluation of Parkinson Patients.

BACKGROUND: In closed-loop on-demand control (ODC) of deep brain stimulation (DBS), stimulation is applied only when symptoms appear. Following stimulation of a fixed duration, DBS is switched off until the symptoms reappear. By repeating these demand-driven cycles, the amount of stimulation delivered can be decreased, thereby reducing DBS side-effects and improving battery-life of the pulse-generator. This article introduces Ro metric for quantification of degree of benefit of ODC and explores candidate selection in tremor-dominant Parkinson's disease (PD). METHOD: The study was performed on nine PD patients previously implanted with Medtronic DBS systems. Accelerometer sensor was placed on the tremor-dominant hand to detect onset of tremor. Fixed duration of stimulation (DS) of 20-80 sec was applied. Once the tremor was observed, stimulation was switched on. These trials were repeated during resting, postural, and kinetic conditions. Ro metric was calculated as the ratio of stimulation-off tremor-free period to the DS. Ro calculated at different DS were compared for each patient. RESULTS: We found that for each patient, Ro varied with DS and an optimal DS* gave a higher percentage of stimulation-off time. Average Ro at DS* varied from 0.554 to 4.24 for eight patients giving 35%-80% stimulation-off time. CONCLUSIONS: Ro values can be used for selection of optimal DS* in ODC. Three of nine patients were found to be tremor-free without stimulation for >50% of total time with even up to 80% in one patient. Patients with low Ro may not benefit from ODC in DBS, where the trade-off between having side-effects and using ODC system will need to be assessed.

Real-time prediction of disordered breathing events in people with obstructive sleep apnea.

PURPOSE: Conventional therapies for obstructive sleep apnea (OSA) are effective but suffer from poor patient adherence and may not fully alleviate major OSA-associated cardiovascular risk factors or improve certain aspects of quality of life. Predicting the onset of disordered breathing events in OSA patients may lead to improved strategies for treating OSA and inform our understanding of underlying disease mechanisms. In this work, we describe a deployable system capable of performing real-time predictions of sleep disordered breathing events in patients diagnosed with OSA, providing a novel approach for gaining insight into OSA pathophysiology, discovering population subgroups, and improving therapies. METHODS: LArge Memory STorage and Retrieval artificial neural networks with 864 different configurations were applied to polysomnogram records from 64 patients. Wavelet transforms, measures of entropy, and other statistics were applied to six physiological signals to provide network inputs. Approximate statistical tests were used to determine the best performing network for each patient. The most important predictors of disordered breathing events in OSA patients were determined by analyzing internal network parameters. RESULTS: The average optimized individual prediction sensitivity and specificity were 0.81 and 0.77, respectively. Predictions were better than random guessing for all OSA patients. Analysis of internal network parameters revealed a high degree of heterogeneity among disordered breathing event predictors and may reveal patient subgroups. CONCLUSIONS: We report the first practical system to predict individual disordered breathing events in a heterogeneous group of patients diagnosed with OSA. The pattern of disordered breathing predictors suggests variable underlying pathophysiological mechanisms and highlights the need for an individualized approach to OSA diagnosis, therapy, and management.

Towards fully automated closed-loop Deep Brain Stimulation in Parkinson's disease patients: A LAMSTAR-based tremor predictor.

This paper describes the application of the LAMSTAR (LArge Memory STorage and Retrieval) neural network for prediction of onset of tremor in Parkinson's disease (PD) patients to allow for on-off adaptive control of Deep Brain Stimulation (DBS). Currently, the therapeutic treatment of PD by DBS is an open-loop system where continuous stimulation is applied to a target area in the brain. This work demonstrates a fully automated closed-loop DBS system so that stimulation can be applied on-demand only when needed to treat PD symptoms. The proposed LAMSTAR network uses spectral, entropy and recurrence rate parameters for prediction of the advent of tremor after the DBS stimulation is switched off. These parameters are extracted from non-invasively collected surface electromyography and accelerometry signals. The LAMSTAR network has useful characteristics, such as fast retrieval of patterns and ability to handle large amount of data of different types, which make it attractive for medical applications. Out of 21 trials blue from one subject, the average ratio of delay in prediction of tremor to the actual delay in observed tremor from the time stimulation was switched off achieved by the proposed LAMSTAR network is 0.77. Moreover, sensitivity of 100% and overall performance better than previously proposed Back Propagation neural networks is obtained.

On modeling the neuronal activity in movement disorder patients by using the Ornstein Uhlenbeck process.

Mathematical models of the neuronal activity in the affected brain regions of Essential Tremor (ET) and Parkinson's Disease (PD) patients could shed light into the underlying pathophysiology of these diseases, which in turn could help develop personalized treatments including adaptive Deep Brain Stimulation (DBS). In this paper, we use an Ornstein Uhlenbeck Process (OUP) to model the neuronal spiking activity recorded from the brain of ET and PD patients during DBS stereotactic surgery. The parameters of the OUP are estimated based on Inter Spike Interval (ISI) measurements, i.e., the time interval between two consecutive neuronal firings, by means of the Fortet Integral Equation (FIE). The OUP model parameters identified with the FIE method (OUP-FIE) are then used to simulate the ISI distribution resulting from the OUP. Other widely used neuronal activity models, such as the Poisson Process (PP), the Brownian Motion (BM), and the OUP whose parameters are extracted by matching the first two moments of the ISI (OUP-MOM), are also considered. To quantify how close the simulated ISI distribution is to the measured ISI distribution, the Integral Square Error (ISE) criterion is adopted. Amongst all considered stochastic processes, the ISI distribution generated by the OUP-FIE method is shown to produce the least ISE. Finally, a directional Wilcoxon signed rank test is used to show statistically significant reduction in the ISE value obtained from the OUP-FIE compared to the other stochastic processes.

Pathological tremor prediction using surface electromyogram and acceleration: potential use in 'ON-OFF' demand driven deep brain stimulator design.

OBJECTIVE: We present a proof of concept for a novel method of predicting the onset of pathological tremor using non-invasively measured surface electromyogram (sEMG) and acceleration from tremor-affected extremities of patients with Parkinson's disease (PD) and essential tremor (ET). APPROACH: The tremor prediction algorithm uses a set of spectral (Fourier and wavelet) and nonlinear time series (entropy and recurrence rate) parameters extracted from the non-invasively recorded sEMG and acceleration signals. MAIN RESULTS: The resulting algorithm is shown to successfully predict tremor onset for all 91 trials recorded in 4 PD patients and for all 91 trials recorded in 4 ET patients. The predictor achieves a 100% sensitivity for all trials considered, along with an overall accuracy of 85.7% for all ET trials and 80.2% for all PD trials. By using a Pearson's chi-square test, the prediction results are shown to significantly differ from a random prediction outcome. SIGNIFICANCE: The tremor prediction algorithm can be potentially used for designing the next generation of non-invasive closed-loop predictive ON-OFF controllers for deep brain stimulation (DBS), used for suppressing pathological tremor in such patients. Such a system is based on alternating ON and OFF DBS periods, an incoming tremor being predicted during the time intervals when DBS is OFF, so as to turn DBS back ON. The prediction should be a few seconds before tremor re-appears so that the patient is tremor-free for the entire DBS ON-OFF cycle and the tremor-free DBS OFF interval should be maximized in order to minimize the current injected in the brain and battery usage.

A neural network-based design of an on-off adaptive control for Deep Brain Stimulation in movement disorders.

The current Food and Drug Administration approved system for the treatment of tremor disorders through Deep Brain Stimulation (DBS) of the area of the brain that controls movement, operates open-loop. It does not automatically adapt to the instantaneous patient's needs or to the progression of the disease. This paper demonstrates an adaptive closed-loop controlled DBS that, after switching off stimulation, tracks few physiological signals to predict the reappearance of tremor before the patient experiences discomfort, at which point it instructs the DBS controller to switch on stimulation again. The core of the proposed approach is a Neural Network (NN) which effectively extracts tremor predictive information from non-invasively recorded surface-electromyogram(sEMG) and accelerometer signals measured at the symptomatic extremities. A simple feed-forward back-propagation NN architecture is shown to successfully predict tremor in 31 out of 33 trials in two Parkinson's Disease patients with an overall accuracy of 75.8% and sensitivity of 92.3%. This work therefore shows that closed-loop DBS control is feasible in the near future and that it can be achieved without modifications of the electrodes implanted in the brain, i.e., is backward compatible with approved DBS systems.

Adaptive control of deep brain stimulator for essential tremor: entropy-based tremor prediction using surface-EMG.

Entropy, as a measure of randomness in time-varying signals, is widely used in areas such as thermodynamics, statistical mechanics and information theory. This paper investigates the use of two commonly employed entropy measures, namely Wavelet Entropy and Approximate Entropy, as a predictor of tremor reappearance in Essential Tremor patients; the predictor input is a raw surface-electromyographic (sEMG) signal measured from tremor affected muscles of patients implanted with a Deep Brain Stimulator (DBS). A combination of both types of entropy measure is shown to successfully predict the occurrence of tremor few seconds before its visual manifestation. This result can potentially lead to a novel sEMG-based adaptive on-off DBS controller that can be added on to existing open-loop DBS systems with minimal changes; an adaptive DBS system provides stimulation only when needed thereby reducing the risk of brain over stimulation, delaying DBS intolerance and prolonging DBS battery life.

Stochastic modeling of the neuronal activity in the thalamus of Essential Tremor patient.

Several stochastic models, with various degrees of complexity, have been proposed to model the neuronal activity from different parts of the human brain. In this paper, we use an Ornstein-Uhlenbeck Process (OUP) to model the spike activity recorded from the thalamus of a patient suffering from Essential Tremor at the time of implantation of the electrodes for Deep Brain Stimulation. From the recorded data, which contains information about the spike times of a single neuron, we identify the model parameters of the OUP.We then use these parameters to numerically simulate the inter-spike interval distribution. We show that the OUP provides excellent fits to the data recorded both without any external stimulation as well as with stimulation. We finally compare the fits with other stochastic models commonly used and we show the superiority of the OUP model in general.

Adaptively controlling deep brain stimulation in essential tremor patient via surface electromyography.

OBJECTIVES: We present patient test outcomes to show that on-off control of deep brain stimulation sequences in essential tremor patients is achievable in a self-adaptive manner via non-invasive surface-electromyography, to prevent tremors in these patients. METHOD: In our study, an essential tremor patient, who underwent bilateral deep brain stimulation implantation 8 years earlier, was subjected to deep brain stimulation at 130 pulses/second, with a 90-microsecond pulse-width, in packets of durations from 20 to 73 seconds and was monitored with surface-electromyography. RESULTS: At the end of these stimulation packets, tremor-free intervals followed, averaging over 20 seconds, before tremor reappeared. Wavelet analysis of the eletromyographic signals allowed predicting onset of tremors at the end of the tremor-free intervals and was successful in all test cycles. Furthermore, once stimulation was restarted, the tremors disappeared within 0.5 seconds on average. When restarting stimulation approximately 2 seconds ahead of the end of tremor-free post-simulation intervals as predicted by visual inspection of unprocessed electromyograms, no tremors occurred during three successive cycles of stimulation-on and stimulation-off. Maximal ratio of tremor-free duration to stimulation duration was computed, to determine a best DBS (deep brain stimulation) duration range (20-35 seconds). CONCLUSIONS: We show existence of a tremor-free interval averaging over 20 seconds that follows applying stimulation packets of 20-35 seconds and that surface electomyogram allows predicting onset of tremor to facilitate activation of a next stimulation packet before tremor reappears. This establishes the feasibility of electromyographic-based predictive on-off control of deep brain stimulation in certain essential tremor patients. Best tremor-free duration to stimulation duration ratio may differ over the progression of the disorder and from patient to patient.

Stochastic modeling of the neuronal activity in the subthalamic nucleus and model parameter identification from Parkinson patient data.

Several stochastic models, with various degrees of complexity, have been proposed to model the neuronal activity from different parts of the human brain. In this article, we use a simple Ornstein-Uhlenbeck process (OUP) to model the spike activity recorded from the subthalamic nucleus of patients suffering from Parkinson's disease at the time of implantation of the electrodes for deep brain stimulation. From the recorded data, which contains information about the spike times of a single neuron, we identify and extract the model parameters of the OUP. We then use these parameters to numerically simulate the inter-spike intervals and the voltage across the neuron membrane. We finally assess how well the proposed mathematical model fits to the measured data and compare it with other commonly adopted stochastic models. We show an excellent agreement between the computer-generated data according to the OUP model and the measured one, as well as the superiority of the OUP model when compared to the Poisson process model and the random walk model; thus, establishing the validity of the OUP as a simple yet biologically plausible model of the neuronal activity recorded from the subthalamic nucleus of Parkinson's disease patients.

Automated prediction of apnea and hypopnea, using a LAMSTAR artificial neural network.

RATIONALE: The prediction of individual episodes of apnea and hypopnea in people with obstructive sleep apnea syndrome has not been thoroughly investigated. Accurate prediction of these events could improve clinical management of this prevalent disease. OBJECTIVES: To evaluate the performance of a system developed to predict episodes of obstructive apnea and hypopnea in individuals with obstructive sleep apnea; to determine the most important signals for making accurate and reliable predictions. METHODS: We employed LArge Memory STorage And Retrieval (LAMSTAR) artificial neural networks to predict apnea and hypopnea. Wavelet transform-based preprocessing was applied to six physiological signals obtained from a set of polysomnography studies and used to train and test the networks. MEASUREMENTS AND MAIN RESULTS: We tested prediction performance during non-REM and REM sleep as a function of data segment duration and prediction lead time. Measurements included average sensitivities, specificities, positive predictive values, and negative predictive values. Prediction performed best during non-REM sleep, using 30-second segments to predict events up to 30 seconds into the future. Most events were correctly predicted up to 60 seconds in the future. Apnea prediction achieved a sensitivity and specificity up to 80.6 +/- 5.6 and 72.8 +/- 6.6%, respectively. Hypopnea prediction achieved a sensitivity and specificity up to 74.4 +/- 5.9 and 68.8 +/- 7.0%., respectively. CONCLUSIONS: We report, to our knowledge, the first system to predict individual episodes of apnea and hypopnea. The most important signal for apnea prediction was submental electromyography. The most important signals for hypopnea prediction were submental electromyography and heart rate variability. This prediction system may facilitate improved therapies for obstructive sleep apnea.

A modified LAMSTAR neural network and its applications.

This paper describes a modification of the LArge Memory STorage And Retrieval (LAMSTAR) neural network. The purpose of the modification is to allow rare events a larger role in decision-making when they are strongly biased towards a particular decision. As a by-product, the modification also permits the introduction of a confidence measure. This measure allows comparison across different network inputs so that the user may choose the "best" solution. The authors have applied the modified LAMSTAR network to a financial forecasting problem.

Walking performance, medical outcomes and patient training in FES of innervated muscles for ambulation by thoracic-level complete paraplegics.

OBJECTIVE: To discuss functional electric stimulation (FES) gait training of upper motoneuron spinal cord injured complete paraplegics considering ambulation performance, physiologic and metabolic responses as well as psychologic outcome, while providing myologic insight into ambulation via FES when training starts many years post-injury. METHODS: Transcutaneous FES using the Parastep stimulation system, gait training methods with and without major emphasis on muscle reinforcement, cardiovascular and respiratory conditioning. Examination of myofiber tissues and correlation of normal muscles histology versus innervated muscles of upper motor neuron and of denervated muscles of lower motor neuron paraplegics. RESULTS: Published works in literature reviewed in this paper report average walking distance of 440 m/walk when major muscle reinforcement and preconditioning cardiovascular and respiratory systems precedes gait training, versus average 115 m/walk when undergoing direct gait training. Medical, metabolic and psychologic outcomes, as reported in several works, point to benefits of FES walking, including 60% increase in blood flow to lower extremities. Myofiber tissues of patients with upper motor neuron paralysis compare well with those of normal tissue even many years post-injury, while adipose tissue substitute muscle fibers in patients with lower motor neuron lesions. DISCUSSION: Transcutaneous FES allows considerably longer walking distances and speed at the end of training when training involves an extensive pre-conditioning program than with direct gait training. Medical and psychologic benefits are observed, especially concerning blood flow to the lower extremities. Myofiber examinations provide myologic understanding of effectiveness of FES many years post-injury.

Brain tumor demarcation by applying a LAMSTAR neural network to spectroscopy data.

The application of optical spectroscopy for intra-operatively delineating brain tumors has been studied in this paper. The classification of tissue as normal, tumor and boundary is done using the LAMSTAR neural network (NN). The objective is to combine both fluorescence and reflectance as attributes to be used for the demarcation, thus giving the identification greater specificity and sensitivity. The input word has seven sub-words, five with autofluorescence parameters and two with reflectance values. The mean and standard deviation for the fluorescence parameters that were used for setting the weights of the NN were obtained from previous work. The reflectance value was used with the fluorescence parameters through a two-step discrimination algorithm. The neural network was trained with 10 sets of each tumor, normal and boundary type of tissue parameters. The network was then tested with 15 complete input sets and 10 incomplete sets for the identification. A 100% success rate was obtained for the complete testing sets and 80% for the incomplete ones. The most significant self-organizing map layers of the network were also identified for each decision. A sensitivity of 97.1% and specificity of 94.73% were achieved, which is much higher than earlier published results of 89 and 76%, respectively.

A neural-network-based detection of epilepsy.

Diagnosis of epilepsy is primarily based on scalp-recorded electroencephalograms (EEG). Unfortunately the long-term recordings obtained from 'ambulatory recording systems' contain EEG data of up to one week duration, which has introduced new problems for clinical analysis. Traditional methods, where the entire EEG is reviewed by a trained professional, are very time-consuming when applied to recordings of this length. Therefore, several automated diagnostic aid approaches were proposed in recent years, in order to reduce expert effort in analyzing lengthy recordings. The most promising approaches to automated diagnosis are based on neural networks. This paper describes a method for automated detection of epileptic seizures from EEG signals using a multistage nonlinear pre-processing filter in combination with a diagnostic (LAMSTAR) Artificial Neural Network (ANN). Pre-processing via multistage nonlinear filtering, LAMSTAR input preparation, ANN training and system performance (1.6% miss rate, 97.2% overall accuracy when considering both false-alarms and 'misses') are discussed and are shown to compare favorably with earlier approaches presented in recent literature.

An overview of the state of the art of noninvasive FES for independent ambulation by thoracic level paraplegics.

This paper is an overview of the status of transcutaneous noninvasive (unbraced) functional electrical stimulation (FES) for independent standing and for independent ambulation by traumatic spinal-cord injured (SCI) paraplegics with complete spinal cord lesions at the thoracic level. The paper discusses aspects of patient selection, patient training, system performance, ambulation range, medical benefits and psychological benefits. It also considers problems relating to system adoption and long term system use. Furthermore, the paper discusses the various aspects of transcutaneous noninvasive FES as compared with implanted FES systems for ambulation by thoracic level SCI patients.