Distinction of an Assortment of Deep Brain Stimulation Parameter Configurations for Treating Parkinson’s Disease Using Machine Learning with Quantification of Tremor Response through a Conformal Wearable

Deep brain stimulation offers an advanced means of treating Parkinson’s disease in a patient specific context. However, a considerable challenge is the process of ascertaining an optimal parameter configuration. Imperative for the deep brain stimulation parameter optimization process is the quantification of response feedback. As a significant improvement to traditional ordinal scale techniques is the advent of wearable and wireless systems. Recently conformal wearable and wireless systems with a profile on the order of a bandage have been developed. Previous research endeavors have successfully differentiated between deep brain stimulation “On” and “Off” status through quantification using wearable and wireless inertial sensor systems. However, the opportunity exists to further evolve to an objectively quantified response to an assortment of parameter configurations, such as the variation of amplitude, for the deep brain stimulation system. Multiple deep brain stimulation amplitude settings are considered inclusive of “Off” status as a baseline, 1.0 mA, 2.5 mA, and 4.0 mA. The quantified response of this assortment of amplitude settings is acquired through a conformal wearable and wireless inertial sensor system and consolidated using Python software automation to a feature set amenable for machine learning. Five machine learning algorithms are evaluated: J48 decision tree, K-nearest neighbors, support vector machine, How to cite this paper: LeMoyne, R., Mastroianni, T., Whiting, D. and Tomycz, N. (2020) Distinction of an Assortment of Deep Brain Stimulation Parameter Configurations for Treating Parkinson’s Disease Using Machine Learning with Quantification of Tremor Response through a Conformal Wearable and Wireless Inertial Sensor. Advances in Parkinson’s Disease, 9, 21-39. https://doi.org/10.4236/apd.2020.93003 Received: July 7, 2020 Accepted: August 18, 2020 Published: August 21, 2020 Copyright © 2020 by author(s) and Scientific Research Publishing Inc. This work is licensed under the Creative Commons Attribution International License (CC BY 4.0). http://creativecommons.org/licenses/by/4.0/


Introduction
Deep brain stimulation offers a significant advance for the treatment of people with Parkinson's disease. With the considerable assortment of parameter configurations, such as amplitude, frequency, pulse width, and polarity, a patient specific treatment strategy is feasible [1] [2] [3]. The essence of efficacious intervention through the deep brain stimulation system is contingent upon converging the parameter configuration to an optimized setting, which can present a laborious process [2]- [7].
The deep brain stimulation system parameter optimization process relies upon the establishment of quantified feedback, such as through ordinal scales.
However, the ordinal scale approach is inherently subjective, for which the reliability is controversial [3] [8] [9] [10] [11] [12]. A recommended alternative is the incorporation of a wearable and wireless inertial sensor system to provide feedback for intervention efficacy in a quantified and objective manner [3] [9] [13]- [28]. Recent technological evolutions have produced conformal wearable and wireless inertial sensor systems that impart minimal encumbrance relative to previous applications with a profile on the order of a bandage, such as the BioStamp nPoint [18] [29] [30].
The opportunities of the wearable and wireless system have been further evolved with the application of machine learning to ascertain considerable classification accuracy, such as for distinguishing between deep brain stimulation set to "On" and "Off". The machine learning classification endeavors have relied upon the operation of the Waikato Environment for Knowledge Analysis (WEKA) [3] [22] [23] [24] [25] [28] [29] [30]. WEKA presents an assortment of machine learning classification algorithms [31] [32] [33]. These machine learning algorithms offer contextually unique performance capabilities, and the most appropriate machine learning algorithm is integrally correlated with the intended application [24] [25] [28]. Beyond the scope of differentiating the "On" R. LeMoyne et al.
DOI: 10.4236/apd.2020. 93003 23 Advances in Parkinson's Disease and "Off" status settings for deep brain stimulation exists the opportunity to distinguish between actual parameter configurations for deep brain stimulation through machine learning, such as with the variation of the amplitude parameter in conjunction with an assortment of machine learning algorithms. The objective of the research endeavor is to evaluate the efficacy of machine learning algorithms with respect to the parametric variation of deep brain stimulation for the treatment of Parkinson's disease with the BioStamp nPoint providing quantified feedback. Five machine learning algorithms are considered: J48 decision tree, K-nearest neighbors, support vector machine, logistic regression, and random forest. The amplitude for deep brain stimulation is the selected parameter for variation respective of the following settings: "Off" status as a baseline, 1.0 mA, 2.5 mA, and 4.0 mA. Two machine learning performance parameters are considered for determining the most suitable algorithm: classification accuracy and time to develop the machine learning model.

Parkinson's Disease and Traditional Therapy
With respect to the United States of America, approximately one million people have been diagnosed with Parkinson's disease, which is neurodegenerative and proportional to age [34] [35] [36]. The neurological basis for Parkinson's disease is associated with degeneration of the substantia nigra, which leads to diminished dopamine available for the caudate and putamen [34] [37]. Parkinson's disease characteristically involves the movement disorder of resting tremor, which has an approximate frequency of four to five per second [34] [38].
Traditional medication therapy for Parkinson's disease involves the prescription of levodopa [34] [37] [39]. Occasionally the medication therapy diminishes in efficacy, and an alternative intervention is sought, such as the thalamotomy and pallidotomy. These neurological techniques permanently disrupt pathways pertaining to the thalamus and globus pallidus internal segment [20] [44]. The primary subsystem of the deep brain stimulation system is the implantable pulse generator that is battery powered. The electric signal originating for the implantable pulse generator transmits through electrode leads that terminate at a prescribed deep brain neurological structure [4].  [45]. Upon completion of the surgical procedure for the implementation of the deep brain stimulation system, the optimization of the parameter configuration is ascertained through the variation of the four available parameter settings: am-

Ordinal Scale Approach for Quantifying Parkinson's Disease Status
The ordinal scale technique is a standard means for the determination of the

Preliminary Wearable and Wireless Systems, such as the Smartphone, for Parkinson's Disease Quantification
Functionally wearable inertial sensor systems equipped with accelerometers have been successfully applied for ascertaining therapy efficacy for movement disorders inclusive of Parkinson's disease [48]- [53]. The evolution to wearable and wireless inertial sensor systems has rendered tethering and manual techniques for uploading the signal data obsolete [3]  LeMoyne and Mastroianni during 2010 succeeded in the preliminary demonstration of the smartphone as a wearable and wireless inertial sensor system for quantifying Parkinson's disease hand tremor through the smartphone's accelerometer. The acquired accelerometer signal data was conveyed wirelessly as an email attachment through the Internet for pending post-processing [19]. An immediate observed utility was that the experimental location and post-processing resources could be situated anywhere in the world with Internet access [3] [13]- [19].

Integration of Wearable and Wireless Systems for Deep Brain Stimulation Treatment of Parkinson's Disease with Machine Learning Classification
Wearable and wireless systems, such as the smartphone, have been successfully

Machine Learning for the Distinction of Deep Brain Stimulation Scenarios
The combination of machine learning with wearable and wireless inertial sensor systems to provide quantified feedback for establishing a feature set differentiating deep brain stimulation system tuning scenarios has been advocated and suc- Additionally, the BioStamp nPoint is certified as a 510(k) medical device by the FDA. This certification permits the acquisition of medical grade data [66].

Material and Methods
One female subject with an age in the mid-60's diagnosed with Parkinson's disease in 2011 was selected for this preliminary demonstration from the perspective of engineering proof of concept. The subject has been receiving deep brain  [30]. Figure 1 presents the representative mounting technique, and    [67].
The experiment involved the recording of accelerometer signal data from the BioStamp nPoint representing a conformal wearable and wireless system with parametric variation of the amplitude parameter for deep brain stimulation to the following settings: "Off" status as a baseline, 1.0 mA, 2.5 mA, and 4.0 mA.
The BioStamp nPoint incorporated a sampling rate of 250 Hz. The duration of the sampling for each deep brain stimulation amplitude setting was sufficient to acquire five trials lasting for two seconds. Contact with the table to the measured hand was prevented by extending the wrist of the subject beyond an elevated support.
The experimental protocol described below was applied for the four prescribed deep brain stimulation amplitude settings: 1) Mount the BioStamp nPoint by adhesive medium to the dorsum of the hand with a longitudinal and symmetric orientation respective of the third metacarpal.
2) Situate the respective forearm of the subject on an elevated support so that the Parkinson's disease hand tremor does not collide with the table.
3) Initiate the BioStamp nPoint recording with a duration that is capable of acquiring five trials for each prescribed deep brain stimulation amplitude parametric setting ("Off" status as a baseline, 1.0 mA, 2.5 mA, and 4.0 mA) for a two second duration.

4)
With the completion of recording the trial data, wirelessly transmit the acquired inertial sensor data to the secure Cloud computing environment.

Results
The progressive increase of the amplitude parameter for deep brain stimulation induces the attenuation of Parkinson's disease hand tremor. The attenuating trend of Parkinson's disease hand tremor is quantified through the BioStamp nPoint, which constitutes a conformal wearable and wireless inertial sensor system, and the three dimensional orthogonal accelerometer signal is post-processed to the respective acceleration magnitude using Python. The deep brain stimulation system set to "Off" status represents a baseline for Parkinson's disease hand tremor as demonstrated by     All five machine learning algorithms achieved considerable classification accuracy. Figure 7 presents their attained classification accuracy for differentiating between deep brain stimulation set to "Off" status as a baseline, amplitude set to 1.0 mA, amplitude set to 2.5 mA, and amplitude set to 4.0 mA for a subject with Parkinson's disease tremor. The Parkinson's disease tremor is quantified by a conformal wearable and wireless inertial sensor system. Figure 8 represents the time to develop the machine learning models that achieve their associated classification accuracy.
The J48 decision tree achieved 90% classification accuracy. Two instances were misclassified. One instance involving deep brain stimulation set to "Off" status as a baseline was misclassified as being deep brain stimulation set to an Logistic regression and random forest both achieve 95% classification accuracy. They both misclassified an instance of deep brain stimulation set to "Off" status as a baseline as an amplitude setting of 1.0 mA. The logistic regression required 0.07 seconds to develop, and the random forest machine learning model was developed in 0.09 seconds.

Discussion
The selection of the best machine learning algorithm is determined based on two  . Machine learning classification accuracy achieved for the J48 decision tree, K-nearest neighbors, support vector machine, logistic regression, and random forest with respect to deep brain stimulation set to "Off" status as a baseline, amplitude set to 1.0 mA, amplitude set to 2.5 mA, and amplitude set to 4.0 mA. The subject's Parkinson's disease hand tremor was quantified by a conformal wearable and wireless inertial sensor system. Figure 8. Time to develop the J48 decision tree, K-nearest neighbors, support vector machine, logistic regression, and random forest machine learning classification models. Note that the J48 decision tree and K-nearest neighbors machine learning algorithms require less than 0.01 seconds to develop their machine learning models.
However, the time to develop the machine learning model may become more relevant in scenarios, for which the computational processing time is more protracted. This observation would be pertinent for a design requirement of deriving the classification accuracy through an associated wearable system rather than a Cloud computing environment. With respect to this design requirement scenario K-nearest neighbors would be most preferable, since the 95% classification accuracy is associated with a time to develop the machine learning classification model of less than 0.01 seconds.
These progressive evolutions further realize the broad objective of achieving real-time optimization of the parameter configuration for a deep brain stimulation system providing therapy for a person with a movement disorder. Additionally, the evaluation of more subjects is warranted in light of the successful preliminary research. These achievements further develop the presence of Network Centric Therapy, which synergizes the amalgamated capabilities of conformal wearable and wireless inertial sensor systems with data access to Cloud computing resources and acuity of machine learning to distinguish the quantified response to various therapy strategies. Network Centric Therapy has global healthcare implications as patients can be treated with internationally renowned medical talent from anywhere in the world [

Conclusions
The efficacy of five machine learning algorithms (J48 decision tree, K-nearest neighbors, support vector machine, logistic regression, and random forest) has been successfully evaluated with respect to differentiating an assortment of deep brain stimulation parameter configurations, such as amplitude set to of "Off" status as a baseline, 1.0 mA, 2.5 mA, and 4.0 mA, for the treatment of Parkinson's disease. The composition of a feature set suitable for machine learning using WEKA is derived from the quantification of Parkinson's disease hand tremor based on a conformal wearable and wireless inertial sensor system with connectivity to a secure Cloud computing environment, which has a profile on the order of bandage and can be readily mounted about the dorsum of the hand by an adhesive medium. Post-processing of the recorded acceleration signal was facilitated through software automation enabled through Python.
In order to ascertain the best machine learning algorithm, two performance parameters were assigned. The primary performance parameter was classification accuracy, and secondary performance parameter was the time to develop the machine learning model. The support vector machine achieved 100% classification accuracy, but this machine learning algorithm required 0.19 seconds to construct the machine learning model, which is the greatest for the five machine learning algorithms under consideration. In the event that the time to construct the machine learning model becomes relatively more significant K-nearest neighbors attains 95% classification accuracy with less than 0.01 seconds to develop the machine learning model.