Data-Driven Characterization of Neuronal Markers During Deep Brain Stimulation for Patients With Parkinson's Disease

Deep brain stimulation of the subthalamic nucleus (STN DBS) has developed into a standard therapy for treating refractory stages of Parkinson's disease (PD). The large number of DBS systems nowadays routinely implanted represent open loop technology. These so-called continuous DBS (cDBS) systems are relatively simple from a technical perspective, as they deliver uninterrupted high-frequency stimulation pulse trains typically 24 hours a day. The stimulation is applied to the target area, like the STN, without taking into account the current level of PD symptoms or the motor state of the patient. Changes to the stimulation parameters -like pulse width, amplitude or frequency- can be applied only by a trained expert during a so-called adjustment session, which usually takes place in the clinic. This limits the number of adjustment sessions to at most a few per year. This may be sufficient to adapt the system to long-term changes of a patient's state as induced by PD progress, which take place over months and years, but certainly is not sufficient to react upon varying daily conditions or changes on even smaller temporal scales. Despite being a widely accepted approach, cDBS is known to cause several side effects such as speech impairment or tolerance to treatment due to chronic continuous stimulation, and has disadvantages with regard to energy efficiency and battery life of the implanted stimulation device.

In contrast to the available cDBS systems, it would be desirable to have adaptive DBS (aDBS) systems, that provide stimulation on demand only and, for example, reduce or stop stimulation delivery during periods of inactivity or when the motor performance of the patient is sufficiently high. Even though a few aDBS prototypes have been reported in literature, they are investigated in research contexts only and have not yet been included into clinical routines.

To realize the closed loop control of a patient's motor symptoms by an aDBS approach, at least one information source describing the motor state of the patient is required. On the one hand, this information may be accessible via external sensors or wearables, which record e.g. muscle tone, tremor, kinematic information etc. in every-day situations or during the execution of specific motor tasks. Alternatively, the information may also be expressed by specific brain signals, so-called neural markers, which correlate with the motor state and can act as its surrogate.

Informative neural markers can be extracted from several brain areas and with different recording technologies. Activity in the subthalamic nucleus (STN) and other basal ganglia can be measured both during and after the implantation of the DBS electrodes in the form of local field potentials (LFP) or microelectrode recordings (MER). Signals recorded either during stimulation, from small time windows between stimulation sequences, or with stimulation absent can provide information about the clinically relevant motor state of PD patients. Additionally, it has been shown that neural signal recordings via magneto- or electroencephalogram (MEG/EEG) and electrocorticogram (ECoG) may provide valuable complementary information compared to the signals obtained from basal ganglia.

On a clinical level, the motor state of the patients can be assessed using part III of the Unified Parkinson's Disease Rating Scale (UPDRS-III) test battery. Its assessment, however, is rather time consuming and requires the involvement of a clinician (neurologist) and consequently the full UPDRS-III score cannot be used for a aDBS implementation. Unfortunately, with the current state of research, the information about the motor behavior cannot simply be replaced by information collected via brain signals. The reasons is, that the relation between relevant neural markers of the LFP and MER recordings, and the individual motor symptoms (e.g. as described by the UPDRS-III) is far from complete and requires further investigation.

To characterize candidates of neural markers, which can be utilized as surrogates for the motor state, it is important to investigate two questions: (1) (How) does the marker change upon applying DBS? (2) Is this change related to the clinical effects of DBS observed e.g. a change in the UPDRS-III score? In this context, selected oscillatory components have been described. The power of LFP oscillatory components in the beta range (12-30 Hz) has been reported to drop upon DBS and, despite unclear causal relation and action mechanisms, it has also been correlated to motor parkinsonian symptoms as bradykinesia and rigor. Furthermore, the interaction of band power of other frequency components with specific PD motor symptoms has been described. An example is the relation between the delta and gamma band power recorded from the STN with dyskinetic symptoms and the correlation of high gamma band power with UPDRS-III scores, and the modulation of high gamma through DBS or L-Dopa. Additionally, DBS stimulation has also been observed to influence cross-frequency coupling between cortical-cortical, cortical-subcortical and subcortical-subcortical structures.

Most studies on the effect of DBS on the motor system and on informative neural markers report on global effects observed in group studies. However, grand average findings may not provide sufficient information to control aDBS systems for an individual patient. This is underlined by many recent studies from the field of brain-computer interfaces (BCI), where informative neural signatures have been found to be subject-specific, and where subject-specific methods for extracting informative neural markers have been applied successfully. Hence we propose to refine the level of data analysis beyond the level of group statistics.

Apart from neural markers being subject-specific, the implicit dynamics of both, the neural markers and the DBS effects, should be considered:

• Dynamics of the neural markers Even within an individual user and a single day, the adaptation of DBS parameters may be required in order to compensate non-stationary characteristics displayed by neural markers on several temporal scales : (a) On the scale of hours to minutes, due to, e.g., changes in wakefulness/tiredness or circadian cycle. (b) On the scale of minutes to seconds, variations e.g. in the attention level, workload. (c) On even smaller time scales due to the current status of the motor system (task preparation vs. task onset vs. sustained ongoing tasks, high force vs. precision tasks, isometric vs. movement tasks etc.). It must be expected, that the individually informative neural markers, which can be exploited to realize the closed-loop aDBS system, are subject to change their informative content in the above-mentioned time scales and scenarios.
• Dynamics of the DBS effects Depending on the DBS parameters (e.g. intensity, frequency, duration, pulse shape) of the stimulation pattern applied in the immediate past, the effects onto (1) the motor system and onto (2) the informative neural markers are known to persist from several seconds to minutes even after stimulation has been turned off [Bronte-Stewart et al. 2009]. Due to this washout effect of DBS, the stimulation strategy of an aDBS system will probably benefit from taking the (short term) stimulation history into account. The duration and temporal dynamics of this so-called washout period depends on the kind of motor symptom studied. It has been reported to be longer for akinesia (minutes - hours) as opposed to rigidity (minutes). Thus it can be hypothesized, that the dynamics of the washout effects for the motor symptoms and for the neural markers are not the same.

The applicants of this proposal want to make a substantial step forward into the direction of a fully closed-loop aDBS system. To reach this goal, it is necessary to develop data analysis methods for brain signals, which are capable of identifying the aforementioned informative neural markers, and to utilize them as input to decode the current motor state. For both tasks, machine learning methods have been successfully investigated and utilized in the context of closed loop BCI systems. Methods developed in this field allow for single-trial decoding of non-invasive EEG signals and invasive signals like ECoG and LPF. The machine learning methods enable the detection of movement intentions in single-trial and the decoding imagined or executed movements. Furthermore, latest research of the applicants has shown, that BCI approaches allow to even predict the task performance of an upcoming motor task, which may be valuable information for brain state dependent closed-loop applications.