WO2008109508A2 - Automatic parameter selection and therapy timing for increasing efficiency in responsive neurodevice therapies - Google Patents

Abstract

An automatic fully responsive neurotherapy system and method are provided that recognizes and automatically selects efficacious therapy parameters, to finely manipulate the on-off timing of therapy delivery aimed at improving efficacy over the entire lifespan of the device. The system incorporates auto tuning of the character, dose, and timing of neurostimulation across patients, and on-the-fly, on a patient-specific, event-by-event basis. 'Pre-' and 'post-therapy effectivity' features are derived from biological signals and seizure detection subsystems, and a self-organizing 'therapy rulebase' is developed from the storage/archiving subsystems to make sense of these 'effectivity' features, and subsequently command their trajectories in directions of clinical improvement.

Description

AUTOMATIC PARAMETER SELECTION AND THERAPY TIMING FOR INCREASING EFFICACY IN THE RESPONSIVE NEURODEVICE THERAPIES OF

BRAIN DISEASE AND INJURY

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims priority to U.S. Provisional Patent Application No. 60/892,649, filed March 2, 2007, the entirety of which is incorporated herein by reference.

BACKGROUND

The present invention is in the field of medical devices to treat neurological disorders of the brain. More specifically, the invention aims at a method to automatically select the character, dosage parameters and the timing of therapy delivery in an on- demand device with the goal of increasing efficacy in the treatment of epilepsy as well as other neurological and psychiatric disorders, where parameter autotuning, (automatically adapting to individual patients, performance criteria, and patterns), and therapy timing may be beneficial.

While a majority of patients diagnosed with brain disorders can be treated with drugs, surgery, diets, alternative remedies, and combinations of these and other therapies remain options with medically intractable (e.g. medicine resistant) conditions. For example, some 60 million people worldwide are estimated to have medication resistant epilepsy alone. These unrelenting ailments are not only devastating to patients, but impact heavily on personal and family relationships, and the patient's community as well. A number of technologically advanced neuroprosthetic, external, brain-computer interface, and hybrid devices have been envisioned that offer renewed hope for patients with neurological disorders that are not optimally treated with standard therapies. A flagship triumph of this approach is the cochlear implant for the profoundly deaf, considered to be the first restoration of a human sense. The system involves sound signal transduction, processing, and electrical stimulation of inner-ear nerves for the brain to interpret as sound.

Unlike profound sensorineural hearing loss, which can remain stable over time, epilepsy, migraine, recurrent depression, pain, bipolar-disorder, anxiety-, sleep-, abnormal movements-, and other central and peripheral nervous system disorders often display paroxysmal onset and offset of disruptive events. In such cases, it is advantageous for a neuro-device to utilize event detection in order to deliver therapy only at times when it is needed, just before or during the disruptive events, thereby conserving battery life (if battery powered) and reducing unnecessary potential side effects, such as effects of continuous intervention on brain tissue.

Most prior neurostimulation devices for the treatment of medically refractory epilepsy, including vagal nerve stimulation (VNS), deep brain stimulation (DBS), and transcranial magnetic stimulation (TMS), do not embody automatic event detection because they lack signal transduction and processing means. VNS is the only device cleared for sale in the U.S.; all others are experimental in epilepsy, as of this writing. These sensorless devices are programmed to deliver electrical or magnetic stimuli, inducing currents in direct brain or peripheral nerves periodically, every certain number of minutes for a preset duration, around the clock, or at the time instant when an experiment happens to occur, irrespective of brain activity. The responsive neurostimulation system (RNS) disclosed in U.S. Patent No. 6,016,449 to Fischell et al. is an improvement in which intracranial EEG is sensed and optionally stored, allowing one to program an automatic detector for epileptiform activity and use it to trigger stimulation at the approximate times when such therapy should be beneficial. Methods for detecting neurological events from brain signals are well known. Methods for fine-tuning the parameters of such event detectors are also known in the field and art.

Methods for tuning the parameters and the timing of therapeutic neurostimulation are less readily available and have met with inconsistent efficacy in clinical practice. U.S. Patent No. 6,161,045 to Fischell et al. discloses the use of electrically evoked epileptiform activity during presurgical EEG monitoring and brain mapping in order to obtain stimulation parameters for use in a subsequently implanted responsive neurostimulator. This method is not applicable to tuning stimulation parameters as is most frequently needed, during post-implantation and long-term use. U.S. Patent No. 6,480,743 to Kirkpatrick et al. improves upon prior RNS systems by synchronizing the timing of a therapy waveform to a characteristic of the EEG waveform shortly after event detection, for the purpose of injecting therapy variability in a controlled manner to avoid gradual loss of efficacy (or worse, potentially kindling) as a result of learning, accommodation, or habituation in the brain. While this system can maintain clinical efficacy constant over time, it lacks a method to specifically improve such efficacy over time, and to do so automatically. Furthermore, it relies on characteristics of the EEG waveform itself, such as a halfwave duration, rather than on characteristics of the epileptic network, such as a phase distribution. Post-implant adjustments to stimulation parameters have been heretofore relegated to human manual intervention. This remains a slow, costly, guessing exercise with uncertain results. On the one hand, basic research and clinical experience have yet to mature, but on the other hand, this slow approach is doomed by the mathematical intractability of the combinatorially explosive parameter space inherent with pulsatile stimuli. The tuning of therapy parameters needs to occur much more quickly than currently practiced in relation to how seizures wax and wane. Moreover, the selection of the therapy parameters should occur automatically without human intervention, and should be guided by efficacy metrics in a grand feedback scheme. Substantially higher endpoint thresholds must be surpassed for implanted neuro- devices to outweigh the mortality and morbidity risks of surgery, and associated costs.

SUMMARY

Briefly, an automatic fully responsive neurotherapy system is provided that recognizes and automatically select efficacious therapy parameters, and to finely manipulate the on-off timing of therapy delivery aimed at improving efficacy over the entire lifespan of the system. The system incorporates autotuning of the character, dose, and timing of neurotherapy across patients, and on-the-fly, on a patient- specific, event- by-event basis. Significant advantages over prior systems can be achieved by leveraging neurotechnology through "pre-" and "post-stimulation effectivity" features derived from the biological signal sensing and seizure detection subsystems. A self-organizing "therapy rulebase" is developed to make sense of the "effectivity" features, and subsequently command their trajectories in directions of clinical improvement. The system and method work in effect as a multistart stochastic optimization algorithm aimed at maximizing neurotherapy efficacy and clinical benefit. BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is an example of a block diagram of neuro therapy device or system that performs automatic therapy parameter selection.

FIG. 2 is a diagram illustrating the relationship between the flow charts of FIGs. 3, 4 and 7 that together depict a method for automatic therapy parameter selection.

FIG. 3 is a flowchart of a method for learning general seizure (SZ) onset and stereotypical seizure onset.

FIG. 4 is a flowchart that depicts a method of creating an initial therapy rulebase. FIG. 5 illustrates an example of a stylized plot of a brain signal showing a pre- stimulation period (a), a stimulation period (b), a stimulation artifact period (c)

(consisting of stimulation period (b) plus after-effect), and a post-stimulation period (d), from which pre- and post-effectivity features are extracted across channels during the good-signal periods indicated in the diagram, according to an embodiment of the present invention. FIG. 6 shows an example of an EEG signal, Hilbert-based instantaneous phase, and the dependence of optimal therapy timing on the critical phase range.

FIG. 7 is a flowchart that depicts a method of learning efficacious therapy parameters online and applying them on-the-fly.

FIG. 8 is an example of a more detailed block diagram that also shows interaction of the major functional components of the system shown in FIG. 1 during normal therapy mode.

DETAILED DESCRIPTION Overview An automatic fully responsive neurotherapy system and method are provided to recognize and automatically select efficacious therapy parameters. In one form, the method involves sensing at least one biological condition indicative of neurological or other biological activity in a patient and generating a sensor signal representative thereof; analyzing the sensor signal to detect occurrence of a neurological disorder event and to identify at least one stereotypical onset pattern associated with the neurological disorder event; storing therapy rulebase data that maps real-time conditions of the patient to at least one parameter for at least one therapy to be applied to the patient; and automatically selecting parameters for a therapy for the patient based on the sensor signal and therapy rulebase data.

In one form, the system comprises a biological sensor subsystem (comprising at least one biological sensor) that is configured to sense at least one biological condition indicative of neurological activity in a patient and to generate a sensor signal representative thereof; a therapy application subsystem (comprising at least one therapy applicator) that is configured to deliver at least one therapy to the patient in order to treat a neurological disorder event in the detected in the patient; and a controller coupled to the at least one biological sensor and to the at least one therapy applicator. The controller is configured to: analyze the sensor signal to detect occurrence of a neurological disorder event in the patient and to identify at least one stereotypical onset pattern associated with the neurological disorder event; store therapy rulebase data that maps real-time conditions of patient to at least one parameter for a least one therapy to be applied to the patient by the therapy application subsystem; and automatically select parameters for a therapy for the patient based on the sensor signal and the therapy rulebase data.

Referring first to FIG. 1 , a neurotherapy device or system is shown at 1000. The system comprises a neurotherapy controller subsystem 1100, a biological sensor subsystem 1200 and a therapy applicator subsystem 1300. The controller subsystem 1100 is coupled to the sensor subsystem 1200 and the therapy applicator subsystem 1300. In one form, at least a portion (if not all) of the subsystems 1100, 1200 and 1300 are configured to be contained in appropriate housings so that they can be implanted within and/or attached to the body of a patient. To this end, there is a communication interface 1400 that is connected to the controller subsystem 1100 to allow for two-way (wired or wireless) communication between the system 1000 and a computing device shown at 2000. The controller subsystem 1100 may be one or any combination of a processing device, such as a microprocessor or microcontroller that executes one or more software or computer programs stored in computer readable medium, a digital signal processor, an application specific integrated circuit (ASIC), or any other combination of programmable or fixed logic devices. It is understood that there may be analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) in the system 1000 to facilitate signal analysis in the digital domain, and supply of therapy control signals to the analog domain, if necessary. The computing device 2000 has a communication interface 2100 that transmits signals to, and receives signals from, the communication interface 1400 in the system 1000. A medical profession or other trained person may interact with the system 100 via the computing device 2000 using a display and user interface equipment shown at 2200. The computing device 2100 may be an off-the-shelf computer (hand-held, laptop, desktop, etc.) or a customized computing device configured to interact with the system 1000.

The biological sensor subsystem 1200 comprises one or more sensors that are configured to sense any biological signal that represents or indicates neurological activity of a patient. One example of a sensor is an electroencephalogram (EEG) sensor. Other sensors and conditions that may be monitored include, without limitation: cardiac signals, neural activity from other brain structures, nuclei or functional regions associated with other neurological disorders such as depression, mania, pain, schizophrenia; systems that control bodily movement, either of the limbs, trunk, head and neck, or internal movement such as peristalsis. Other conditions that could be monitored are those that generate neural or other signals, chemical, growth or change in composition or physical characteristics. Examples include stomach secretion of acid, release of neurotransmitters, rate of growth of tissue such as peripheral nerves after injury, tumor growth or reduction, etc. The therapy applications subsystem 1300 comprises one or more therapy application devices that can treat a neurological disorder event in a patient. One example of such a device is an implantable pulse generator (IPG) that supplies electrical or magnetic stimulation to one or more sites within a patient's nervous system. Other types of therapy applicators include, without limitation, drug delivery devices such as those that release antiepileptic or neuromodulatory substances in the brain to treat abnormal neural activity, mood, behavior, awareness, etc., devices that release chemicals, drugs and other substances or electrical stimulation in the heart, limbs, viscera, peripheral nerves, blood vessels or other structures. The system might also accommodate devices that record and/or deliver light to modulate light-activated drugs, substances or other forms of therapy.

A medicine-refractory patient typically has electrodes implanted on or near epileptic brain foci as part of a presurgical evaluation in the hope of capturing and localizing activity delineating the epileptic network. During a brain mapping procedure, in appropriate patients, an external neurostimulation device is used to determine if regions of seizure onset overlap regions of eloquent cortex whose surgical removal could cause significant functional (cognitive, motor, sensory, emotional, psychiatric etc.) impairment. In contrast to epilepsy surgery, which is irreversible, an implanted neurostimulator is a viable therapeutic option, even in those patients who are not surgery candidates, because interventions can be turned on, off, and modified at will, via software. Brain mapping generates some experimental data regarding current intensity and frequency levels for which epileptiform after-discharges are evoked or unpleasant sensations are reported by the patient in various locations. These thresholds may vary from location to location, depending upon characteristics of the brain underlying stimulation electrodes, and the electrode-tissue interface.

The term "neurological disorder event" is used herein and it meant to include, without limitation, the aforementioned neurological diseases described above, as well as abnormal or undesired movements, behavior, pain, confusion, cognitive dysfunction, loss of awareness, coma, syncope, or similar events caused by dysfunction in other organs, such as heart, lungs, kidneys, liver etc.

FIG. 2 shows that the method for automatic therapy parameter selection is depicted by the flow charts (connected to each other by appropriate connector labels) of FIGs. 3, 4 and 7. FIGs. 3 and 4 illustrate aspects of the method that pertain to learning modes in order to develop a stereotypical onset discriminator, data for a stereotypical therapy rulebase, and improvement-seeking therapy profiles. Once the processing shown in FIGs. 3 and 4 is complete, the system can engage in "normal" therapy mode, where the processing shown in FIG. 7 occurs every time there is a therapy, or on demand by a physician upon interrogation of the system. hi the flow charts of FIGs. 3, 4 and 7, the trapezoidal boxes indicate that some degree of manual input/human intervention may be performed in that step.

Referring now to FIG. 3, a first phase of the method is shown in which learning of characteristics associated with general seizure onset and stereotypical seizure onset for a patient is made. At 10, therapy parameters of the therapy application subsystem 1300 are initialized for a newly implanted system 1000 in the patient. For example, the therapy application subsystem 1300 may comprise an IPG. Numerous techniques are known for initializing the stimulation parameters, examples of which are disclosed in aforementioned U.S. Patent No. 6,161,045 to Fischell et al. These parameters are saved, but at 20 the therapy application subsystem 1300 itself is initially disabled, since therapy delivery is to be contingent upon the detection of epileptiform events, yet to be determined.

During patient recovery after implantation of the system 1000, electrode impedances may fluctuate due to hemorrhage, edema, or other response to surgical implantation. As soon as stable EEG signals can be registered, a monitoring period is devoted to tuning seizure detection or prediction parameters that are appropriate for the particular patient. In one embodiment, at 30, initial parameters are set on the sensitive rather than on the specific end of the spectrum, using a universal seizure detector such as linelength, or a combination of useful quantitative features, in order to avoid missing seizures. At 40, three or more EEGs are automatically captured by the device, with the onset of epileptiform EEG activity occurring one or more minutes into each recording. Whenever practical, these raw EEG signal epochs are telemetered from the implanted device to a permanent, cumulative EEG archive 50 such as a database located remotely. This may require transmission to an interim receiver, either mobile with the patient or stationary, and later transmission to a central processing center. Archiving EEG data allows offline analyses and ensures minimal loss of data as the device begins overwriting previous recordings due to limited storage capacity. Tuning of seizure detection or prediction parameters preferably involves evaluation of a performance metric 70, and iterations of the steps 40 through 90 in FIG. 3, until a desired blend of sensitivity and specificity or receiver operating characteristic point are achieved. Other biological signals related to neurological activity that are detected by the biological sensor subsystem 1200 may also be processed in a manner similar to that shown in FIG. 3.

At 60, a stereotypical onset pattern discriminator function is provided. It is well known that seizure onset patterns within an individual are often stereotyped, though a single patient may have several different stereotyped onsets, depending upon the nature of his/ her disorder. Meta-analyses of published studies suggest an average 60% of all within-patient seizures may be similar, possibly reflecting a stable unitary pathophysiological substrate for a patient's most frequent kind of seizure. There are several advantages associated with the stereotypical onset pattern discriminator 60. First, it is expected that the learning of efficacious therapy parameters for the stereotypical onset will converge since the "consistent" problem source can afford a consistent therapeutic solution (while the tuning of parameters for the non-stereotypical class may need to continue indefinitely). It is important to notice that these stereotypic onsets may sometimes require different responses due to changes in patient environment or circumstances, such as missing seizure medications, exposure to seizure-inducing substances (e.g. alcohol, other medications) or conditions (sleep deprivation, stress etc.). Second, a measure of therapeutic success becomes more reliable than previously possible since the stereotypical onset discriminator predicts which events would have evolved into full-blown seizures had no interventive stimulation been applied. Some events without a clearly stereotyped onset may go on to become seizures, but others in this non- stereotypical class could be self-limiting and cease without any therapeutic response, thus creating ambiguity regarding stimulation "success" in prior art systems.

In one embodiment, the stereotypical onset discriminator function is a post- detector for the universal seizure detector developed in steps 40 through 90 shown in FIG. 3. For only those EEG segments that have been flagged as "seizure" by the detector, input features such as linelength are further examined in order to separate stereotypical from non-stereotypical onset classes. Alternatively, a separate dedicated classifier can be built to be run in parallel with the seizure detector. In both cases, the captured electrograms may be expertly labeled as stereotypical vs. non-stereotypical in order to form a training set upon which tuning of the discriminator parameters is based during an offline session. In any event, after a stereotypical onset pattern is determined at 60, an evaluation is made at 70 as to the performance of the detection of a stereotypical onset. Then, at 80, it is determined whether the detection performance was acceptable (within some predetermined performance ranges of thresholds). If the performance was determined not to be acceptable, then at 90, the seizure detection parameters as well as the stereotypical onset pattern discriminator are tuned accordingly.

Turning now to FIG. 4, a method is shown for creating an initial therapy rulebase from data gathered during operation system 1000 in a patient. At 100, the therapy application subsystem 130 in the patient is enabled for responsive therapy operation via a programming wand and/or other means such that at 110, any subsequent epileptiform event that is automatically captured by the device will include one or more minutes of pre-therapy plus one or more minutes of post- therapy EEG for each event. Step 120 defines extraction of "pre-therapy and post-therapy effectivity" features, also referred to as effectivity features hereinafter. These features are EEG signatures extracted peri-ictally, just prior to and just after therapy, such that the acute effects of the intervention can be quantified. More generally, an effectivity feature is any observable parameter, measure or quantifiable phenomenon, which may be associated with device efficacy. In automatic mode of the IPG device, each therapy is triggered by a seizure detection that occurred less than 400 ms earlier. Thus, the seizure detection time also forms a natural fiduciary point for comparisons with non-stimulated (but nevertheless detected) events. Several alternative hypotheses can be tested in terms of these data classes. For example, (I) pre- therapy features (augmented with then-in-effect therapy parameters) are predictive of seizure termination, (II) post-therapy features in non-terminated seizures show nevertheless reduced seizure severity compared to post-detection features in non- stimulated seizures, (III) paired post-therapy to pre-therapy features (ratios) show a "normalizing" effect on EEG compared to paired post-detection to pre-detection features (ratios) in non-treated events, etc. Therapy artifacts can interfere with EEG recording, so it is customary to start observation of post-event features about 2 seconds after event onset time as shown in FIG. 5. Our research analyses of some 5,000 stimulated and 6,000 non-stimulated events recorded in humans have lent support to all three exemplary alternative hypotheses listed above. The underlying idea is that the acute, microtemporal efficacy of responsive therapy can be easily quantified and automated using the disclosed effectivity features, and furthermore serves as a shortcut to the macrotemporal improvement in quality of life and other clinical endpoints.

In one embodiment, the character of the pre-effectivity features is the same as that of the post-effectivity features. For example, a pre-therapy linelength compared to a post- therapy linelength is still a linelength, the only difference lying on temporal location with respect to the fiduciary point. However, different post-event feature characters may be useful, for example, when measuring seizure spread across channels, which is undefined pre-ictally (although this is derivable as channel count of off-limit energy or cross- channel sum of energies — a common pre-effectivity feature). Accordingly, this scheme can be used to control changes in epileptic networks that reliably precede seizures, providing a means to arrest seizure generation prior to actually ictal onset events. These seizure "precursor" events can be measured multi-scale, and may come from a variety of sensors measuring noise, unit and multi-unit activity (at the cellular level) and extend down to wider-spread field potentials in a variety of bandwidths, down to DC range activity. Examples of effectivity features include brain signal energy, mean frequency, rhythmicity, complexity, high-frequency oscillation, terminal phase, in-phase synchronization, phase locking, network phase distribution, and cross-channel global correlation. Further examples of effectivity features include "clock" variables such as time of day, day of menstrual period, etc., associated with ultra-, infra-, and circadian rhythms and their relation to device efficacy for certain nocturnal/diurnal/awakening and catanienial epilepsies. Still further examples of effectivity features include "external trigger" variables such as light, noise, music, text, sleep, oxygen, food, serum glucose, fluctuations in hormone or other measurable concentrations in serum, other bodily fluids or by non-invasive means etc. Some measure parameters may involve measuring stimuli such as light/ strobe exposure, touch, sound, cognitive tasks and other stimuli associated with device efficacy in certain reflex epilepsies. Pathological brain states are associated with extreme brain signal energies (e.g., higher ictally, lower postictally), extreme frequencies (e.g., slowed background, chirping overshoot ictally), higher rhythmicity/periodicity, extreme complexities (e.g., higher linelength at onset, lower fractal dimension ictally), certain high-frequency oscillation patterns, certain ranges of instantaneous phase values at the time of therapy (e.g., lower vulnerability of synchronized state, contributing to ineffective therapy), higher in-phase synchronizations between two channels, higher phase difference locks between two channels, higher peakedness in the phase distributions of network oscillators, higher global correlations across spatially distinct channels, certain hours of the day or night, certain days of the month, certain photic frequencies, higher levels of noise, certain spectral content of music, certain words being read, sleep deprivation, hyperventilation, and consumption of certain foods. Effectivity features make it apparent when the brain enters in and out of pathological states (as in well-known seizure detection methods), but furthermore reveal relationships between these states and the parameters and timing of efficacious therapy.

In one particular example, when the biological sensor signal represents brain function of the patient, the signal may analyzed for linelength features to derive at least one of instantaneous amplitude, frequency and phase for the electrical stimulation. Linelength has been used in the lie detection, image processing, and seizure detection arts as a simple, generic "motion detector." Not generally known, as disclosed herein, is the fact that linelength can also be used in implantable devices to monitor instantaneous frequency and phase of a biological signal. Consider a sinusoid x(t) = Acos[2πft], with frequency of oscillation/slow enough for the digitally sampled waveform to still resemble the continuous sinusoid (up to ~F_s/8, or 4x oversampling, where F_s is the sampling frequency). The linelength L of this sinusoid over an observation window of length TV is approximately the L of one half- wave (sum of vertical segments from one extremum to the next = 2A) times the number of half- waves that fit in the length-TV window. Equivalently, that is IA times twice the number of full-waves that fit in the length-TV window (/^"such waves fit in F_s samples), so L ~ (2A)(2)(fN/F_s) = (4N/F_s)Af. Thus, the L of an oversampled sinusoid is proportional to its amplitude and frequency as claimed (and obeys homogeneity but not additivity). From this, the estimate of frequency in Hz is

f ^F'^L

Note direct dependence on A, so for this estimate to work on a chirp, the amplitude has to vary very slowly compared to the oscillation. Relative changes in phase of the signal can be tracked with the integral of the instantaneous frequency estimate. Having interrogated the implanted device, the captured treated events are uploaded to the EEG archive 50) and the effectivity features are calculated initially offline at 120. All of this information may be made available for human review on a computer or hand-held device. At 130, neurologists or other expertly trained personnel can then score the treated events in terms of the acute efficacy of the therapy, as seen on EEG and optionally aided by post-effectivity quantities. In a simple scheme, the scoring consists of annotating each treated event as "successful" vs. "unsuccessful," where success is defined as seizure termination within, e.g., 5 seconds of the onset of therapy. For the convergence and reliability reasons described before, at 140, the expert scorer further stratifies or classifies each stimulated event into either stereotypical or non- stereotypical seizure onset. These further samples of onset types can be added to the discriminator 60 training set in order to refine its parameters for its upcoming unsupervised, online application. In addition, at 150, events that are classified as non- stereotypical events are stored in an archive. From this point forward, the focus of the system and method described herein is on learning efficacious therapy parameters for seizures of the stereotypical class. This is not meant to imply that non-stereotypical events are to be ignored or left untreated, but the distinction is made so that any parallel effort on the non-stereotypical class is understood as "experimental" and likely requiring indefinite parameter "hunt" without convergence of the algorithm. It should also be understood that "convergence" in this context does not imply arriving at a single fixed set of therapy parameters, but rather at a single therapy rulebase (which will contain several sets of therapy parameters to be loaded on-the-fiy conditioned on pre-effectivity features, and is derived from the table described next). It is important to note that non-stereotyped seizures may have properties in common with stereotyped events, making them amenable to therapy derived from stereotyped events. Similarly, non-stereotyped events are likely to fall into a separate, likely reproducible class, allowing a similar, or identical type treatment strategy, developed in parallel to the higher frequency of occurrence stereotyped onset patterns.

The collection of pre- and post-therapy effectivity features, augmented with then- in-effect therapy parameters, along with the initially manually scored therapy outcomes, is organized in a table or data structure called a stereotypical therapy archive 160. This table represents the "cumulative experience" of the device with the particular patient in regards to therapy parameters, conditions under which therapy was applied as reflected by the pre-effectivity features, and the immediate outcome as reflected by post-effectivity features and success score for each stimulated event. The z^th row represents a particular therapy trial, while the/^h column represents a particular augmented feature coordinate. For example, the stereotypical therapy archive can look like Table 1 below: TABLE 1

The therapy profile [ X_k ] is a row vector containing the therapy and channel configuration parameters. For example, in the case of an IPG therapy device, such parameters may include current amplitude, pulse frequency, pulse width per phase, pulse train (burst) duration, limit on number of bursts per episode, electrode surface areas, stimulating electrode subset, bipolar electrode pairings and polarities, etc. Selecting a subset of electrodes from a relatively large number of channels via software partially addresses the electrode location question; physically varying the actual topographic locations post-surgery is clearly impracticable. During the initial stages of populating the table (FIG. 4), the therapy profile is held constant as X_\ for a number of trials, for example, as many as 800 or more therapy events can be collected in a few days, then it is switched to a different profile J_-2 and held for some time and so on. The various X_k profiles are preferably quite distinct from each other, as if representing a random sampling of the large therapy parameter space. These initial profiles can be reasoned or alternatively obtained via pseudorandom number generators within known safety limits. A degree of randomness is important since the whole system will work in effect as a multistat! stochastic optimization algorithm, thus discouraging entrapment into local extrema of a "clinical utility function" as it learns therapy parameters that work best for the patient and the real-time conditions at hand. Examples of parameters that might be used with other forms of therapy might include: (1) amount, concentration, rate, frequency (e.g. how often) and spatial zone of delivery, as well as the specific location of delivery of a drug, substance or medication delivered in response to changes in brain activity; (2) the intensity, frequency, spatial and dispersion of light application to a particular area; (3) the degree of temperature, magnetic or electric field strength delivered to a particular area, as well as its time course, frequency of delivery and spatial extent of delivery. Example of parameters to be measured in such instances include parameters that measure cellular as well as organ function (e.g. electrical), metabolic activity, perfusion/ blood flow, biochemical products of activity (e.g. break-down products of neurotransmitters, neurotransmitter concentration(s), amino acids, lactate and similar metabolites, PH, oxygen, CO2 and other substance concentrations (e.g. glucose), The pre- and post-effectivity features are stored as row vectors x^w and y⁽⁽' respectively on the z^th therapy trial. In general, the calculation of each feature vector coordinate at a given instant of time requires observation of a "window" of the raw (neurologically-related) brain signals, opened to a certain extent spatially (multiple channels) and temporally (interval from a point in the past to the present time). For certain features computed recursively, this window is semi-infinite in extent but exponentially forgetful of the past. This window "slides" across the time axis when a feature time series is desired, however, in the stereotypical therapy archive, the features x^ and y^ are associated with a single "snapshot" around a therapy event. For example, x^(;) can be extracted from a 2-second window right-aligned to the therapy onset time, while y^w can be extracted from a 2-second window left-aligned to 2 seconds after the therapy onset time, as shown in FIG. 5. The latter allows enough time for therapy artifact to subside, yet is early enough to take a valid reading of any acute post-therapy effect. Other methods of dealing with therapy artifact are known in the art. The success score determined at 130 (FIG. 4) may be a scalar S, which in the simple scheme is either 0 (unsuccessful) or 1 (successful seizure termination), but the skilled practitioner will recognize many other schemes are possible, including multilevel and continuous scoring, without departing from the spirit of the invention. The observation period used in order to arrive at a success score need not strictly coincide with the post-effectivity feature observation window. For example, if success is defined as seizure termination within 5 seconds after the beginning of therapy, then the scorer (human or machine) will rely on the 3 -second window that is left-aligned with and completely overlapping the 2-second post-effectivity window ((d) plus 1 more second in FIG. 5).

The stereotypical therapy archive can be interpreted as a statistical sample of possibly dependent realizations drawn from a joint probability density function (PDF) p(X,x,y,S), where [X,x,y,S] is a row vector of random variables (all column names in the table). Integrations of this PDF, JJ- • • f p(X,x,y,S)dXdxdydS , can inform how frequent or how rare it is to observe the variables X, x, y, and S occurring together around the specific values

(a row in the stereotypical therapy archive table). In the steady-state condition of many brain anomalies, this function tends to be stationary or cyclostationary, i.e., the input variables vary with time, but the PDF transformation itself does not vary, or does so only periodically with the circular clock variables (in which case augmenting the effectivity feature vector with such variables converts the cyclostationary PDF into a stationary one). For example, if an individual reports having seizures often on awakening (this pattern tends to persist over the course of years), then a circular histogram of the marginalized distribution of the linelength feature would show peaks at certain hours of the day, regardless of when the patient is examined, as long any changes in medication, etc. have been given time to settle.

The stereotypical therapy archive is preferably populated with events that have been captured without long device-off periods between them, and with device-on periods representing a relatively uniform sampling though time. This motivates an ergodicity assumption to be made, so that the statistical properties of [X,x,y,S] can be inferred from their temporal behavior chronicled in the stereotypical therapy archive table. That is, the stereotypical therapy archive table (normalized as a multidimensional histogram) is itself a sampling of the actual p(X,x,y,S) function with some noise. It is known that direct joint density estimation can over fit the noise when the variables exceed approximately 6 dimensions. Therefore, additional assumptions such as naive Bayes conditional independence of the variables may be useful. An important aspect of the invention is that several clinically relevant questions can be answered based on statistics of the stereotypical therapy archive. For example, the acute efficacy over the entire device-on lifetime for a given patient is the prior probability of a 1 in the success score column of the table, P(S=I), whose unbiased estimator is the fraction of Is to the total number of stimulated events. Since the choice of therapy profile X can be experimentally manipulated during the initial stages (FIG. 4), the table is better interpreted as a family of samplings of the conditional distributions p(x,y,S\X=X_k) for k = 1 ,2, ,K (subscripts are dropped to simplify notation, but it should be understood from the context that these pQ's are different functions), rather than as a single function. This deemphasizes any role for how frequently the therapy profile X_k was used, P(X=X_k), which is initially manipulated, and instead emphasizes the probabilistic role of the remaining variables: effectivity features and therapy outcome, which are initially simply observed. Of clinical relevance is the calculation of probabilities conditioned on therapy profile. For example, the probability of success given a specific therapy, P(S=WX=X_k) for some k, is the fraction of Is in the score column for only the ^-labeled block of rows to the total length of that block. Whereas the experiment wise success rate may be only P(S=I) = 10% for a patient, restriction to the single best performing set of therapy parameters X from among the [X_k), may show efficacy of P(S=WX=X^*) = 30%. The use of a manually chosen fixed set of therapy parameters X that impressionistically worked best among several tried, is effectively the method employed with current devices. Thus, archive data may be stored that represents the effectivity features accumulated over time for a patient such that a probability of therapy efficacy can be computed based on the archive data.

It is well known that neurons possess non-trivial phase resetting curves, wherein pulsatile electrical stimulation at certain phases of the natural spiking/bursting period of the neuron will advance the next spike/burst, whereas at certain other phases it will delay or even abolish the next spike/burst. A macroscopic counterpart of this behavior has been observed experimentally and in computational models of stochastic phase resetting, wherein a network of coupled oscillators engaged in a limit cycle has certain relatively rare vulnerability periods during which pulsatile stimulation of the kind used in IPG devices can desynchronize the macro-oscillation. More generally and advantageously over previously known systems, the method described herein prescribes a probability of therapy success as a function of internal state and external input conditions present at the time of therapy delivery (as measured by the pre-effectivity features). Thus, whereas probability of success may be 3-10% blindly, and 20-40% with a fixed best known therapy profile, that chance may typically increase to P(S=I |xe x ,K=X^**) = 60-90% with a variable best known therapy profile X^** = X^*(x) chosen on-the-fly on the /^th trial based on real-time conditions x. The notation xe x means that pre-effectivity vector x is currently visiting a particular region of the feature space (e.g., a crisp or fuzzy cell around x) denoted x , such as one with certain values of instantaneous phase, certain hours of the day, etc.

Some of the effectivity features, such as instantaneous phase, turn over their full range of values several times within a few seconds. Such fast exploratory features can be used to optimize the timing of therapy delivery by withholding stimulation until x enters a favorable region x known to increase the conditional success rate. For example, as shown in FIG. 6, the circled region 190 represents a "critical phase range" discovered by the device in which stimulation has been more effective in the past. After automatic seizure detection at line 192, the first time the monitored phase falls within this desired range, the stimulator is triggered as shown at line 194). Without loss of generality, therapy timing can be treated as yet another parameter to be tuned based on the effectivity features according to the automatic parameter selection techniques disclosed herein.

In an alternative embodiment, the probability function of interest is P(ye y Ix_^X). This can be used to seek therapy profiles conditioned on x that specifically drive post- effectivity features y into a certain region y , such as one with decreased seizure spread, decreased energy, decreased complexity, decreased synchronization, etc., regardless of whether such outcome should be called "successful."

Reference is made back to FIG. 3. Due to size and power constraints, it is possible that some forms of the implantable system may not have the capability to store the stereotypical therapy archive and/or calculate its statistics in real time using processing and storage technologies currently available, though that may change in the future. In any event, due to current technological limitations, an initial stereotypical therapy rulebase 170 is extracted from the stereotypical therapy archive 160 and periodically downloaded into the device. The rulebase 170 is a much reduced look-up table that directly maps or translates each coarse-grained real-time condition x into its pre-optimized therapy profile X . The therapy to choose at any given time is the one that maximizes probability of acute therapy success given the real-time conditions and the cumulative experience thus far:

X^** = argmax { P(S=I | xe x ,X=X*) }.

[** } In other words, this rule says "if the pre-effectivity feature vector x is currently in region x , select the therapy profile from among the K known so far that was most frequently associated with success whenever x was visiting that same region in the past." The stereotypical therapy rulebase is updated at 290 as described hereinafter. Automatic selection of therapy parameters can be made in real-time based on the effectivity features. The stereotypical therapy rulebase table 170 requires only two columns, one each for the premise xe x and the consequent X=X^**. The/^h row represents a "rule" of the form "IF xe x ^w THEN X=X^**U) " The number of rows is determined by the desired granularity and/or the number of partition regions discernible from the stereotypical therapy archive. An efficient addressing scheme can be used in order to identify which of the coarse-grained regions x ^w an incoming feature vector x belongs to without the need for distance calculations. For example, the coordinates of x can be converted to binary and the most significant bits used as the address into the look-up table. Such analog-to- digital converters can be implemented at minimal power cost in implantable devices using two spiking electronic neurons configured as a hybrid state machine See for example R. Sarpeshkar et al., "An ultra-low-power programmable analog bionic ear processor," /EEE Transactions on Biomedical Engineering, pp. 711-727, April 2005. The firing of individual or multiple simultaneous rules, and the inference mechanism used to declare a set of therapy parameters in the stereotypical therapy rulebase can be practiced according to any of the well known methods used with crisp or fuzzy rulebases. Thus far the stage for increased clinical efficacy has been set by defining effectivity observations [X,x,y,S] and on-the-fly selection of therapy profile X from among a finite set of reference profiles [X_k)- The system also includes a module or function 180 for tuning the reference profiles themselves in a direction of therapeutic benefit. In a first-principled model of the system described herein, optimizing therapy parameters would be a matter of running millions of simulations on a computer and choosing the best inputs predicted by the responses of the model. In clinical trials there is no such model. Accordingly, each of the therapy trials is a datum that may be used to progressively build an empirical model of a continuous "efficacy response surface" or "clinical utility function" U(X). Changes to any of the reference therapy profiles X_k can then be guided by the gradient vector VU(X) in directions of utility ascent (or cost descent). If clinical utility is defined as the acute success rate, then the x-dependent therapy profile sought by the continuous optimization algorithm is

X^**=X^*(x) = argmax { POS=I | xe x JC) },

X where X is allowed to vary within a therapy parameter continuum, not just within the discrete reference set [X_k)- The gradient aspect, in conjunction with the use of multiple randomly seeded therapy profiles in the reference set, makes the overall technique a multistat! stochastic optimization algorithm aimed at maximizing neurostimulator efficacy.

For the gradient to provide direction-of-improvement information, especially when Table 1 contains only a few reference profiles, the system needs the ability to evaluate the utility function over a "continuous" space of therapy parameters. A smooth interpolating function can be synthesized using multivariate kernel smoothing, splines, neural networks, or any other method of continuous function synthesis from a discrete set of data points. Then gradient calculations can be performed numerically or analytically. Given an old profile X_0Id which represents, for example, the current worst performer, or the current best performer, or the centroid of the reference set, etc., a new profile can be calculated as X_new =X_oi_d + ct V [/(J-Ow), where α is a step size preferably scheduled to shrink with time. This new profile is added to an ever-increasing reference set, or it can be used to replace the worst performer in a fixed-size reference set. In sum, the therapy rulebase data may be adjusted based on changes associated with a multivariate function that represents effectivity observations from among a finite set of therapy parameter reference profiles. As the implanted device comes to "know the patient" over the first few months, it becomes ready to be switched to a largely unsupervised, autonomous mode. Notably, the determination of seizure termination success vs. failure changes from qualitative review by neurologist to quantitative determination by post-effectivity features. Reference is now made to FIG. 7 that shows an implementation of this online autonomous mode/process. At 200, pre-effectivity features x are continuously monitored using hybrid digital-analog circuitry and rolling buffers in a similar manner as practiced in currently available seizure detection devices. At 210, when a stereotypical seizure onset is detected, function 220 looks up the location of input x in the x column of the on-board stereotypical therapy rulebase, reads the corresponding set of pre-optimized therapy parameters in the X^** column, and loads these parameters into the system for timed delivery. Because there is only memory transfer but no calculation involved, the function 220 occurs almost instantaneously. Any further delay between this time and actual therapy delivery depends on whether the fired rule contains microtiming features such as instantaneous phase. At 230, post-effectivity features are followed up for a few seconds after therapy delivery to determine if seizure activity has been abolished. The prediction implied by use of the stereotypical seizure onset discriminator is that full-blown seizure activity would be present at this time. At 240, it is determined whether the therapy was a success. That is, if post-effectivity features are within thresholds indicative of successful termination, the whole event automatically receives success score S=I at 250. On the other hand, if therapy failed to stop the seizure, the success score S is set to 0 at 260 and at 270 a flag is set indicating request for a new, improved therapy profile to be deployed the next time feature x visits region x . In both cases, the [X,x,y,S] vector associated to this event is used to update the stereotypical therapy rulebase at 290 stored in the on-board memory for future asynchronous uploading into the stereotypical therapy archive at 280. It should be noted that in this autonomous mode, storage of the raw electrograms is no longer necessary. Finally, the stereotypical therapy rulebase may be updated on the next device interrogation, with reinforcement for the known successful profiles, or with a new therapy profile computed from gradients if such request was made by the device. Thus, the therapy rulebase data stored within the system 1000 that is attached to or implanted in the patient may be derived on-demand or periodically from a more comprehensive collection of stereotypical therapy data stored separately from the system 1000 in a device or system that is not implanted or attached in the patient is derived.

Referring now to FIG. 8 in conjunction with FIG. 7, operation of the therapy mode of the system is described. FIG. 8 illustrates an example of the system 1000 in which the biological sensor subsystem 1100 comprises signal transducers 300 that sense EEG signals and the therapy application subsystem 1300 comprises an IPG 400. That FIG. 8 deals with EEG signals only and a stimulation therapy is meant only by way of example. Other biological signals may be processed in a similar manner, and other therapy regimens may be applied as well. Starting from the brain, the signal transducers 300 are provided that include electrodes and electronics to obtain digital EEG data. A storage buffer 370, e.g., a rolling buffer, continually stores EEG data for a most recent duration of time, e.g., the last 2 minutes, analogous to a stripchart. When commanded at time tθ, a pre-event EEG contained (from tθ-2min. to tθ) in the rolling buffer 370 is copied into electrogram storage unit (e.g., memory) 380. When commanded a second time 2 minutes later, the same copy operation occurs, but the rolling buffer now contains post-event EEG. Thus, a complete 4-minute electrogram from tθ-2min. to tθ+2min with respect to an EEG event is obtained. Given that the electrogram storage may have limited storage capacity, only a finite number of such electrograms that can be stored on-board the patient implanted or carried equipment. Therefore, the storage buffer 370 can be periodically interrogated for transfer of EEG data into an external EEG archive (50 shown in FIGs. 3 and 4) through a telemetry and programming interface device 500 as is known in the art.

At block 310, stereotypical seizure onset features are extracted from the EEG data. At 320, pre-effectivity features are extracted from the real-time EEG data. A seizure detector 330 receives as input the seizure features triggers the above electrogram storage operation by of a timer 360. This is needed for all seizures rather than only a stereotypical subset in order to document clinical endpoints.

The stereotypical onset discriminator (SOD) 335 operates as a refinement immediately after seizure detection and triggers the firing of a pre-optimized set of therapy parameters that are loaded into the neurostimulator 400 "on-the-fly" and that are determined to be appropriate for the existing brain condition x(t), based on the continuously extracted pre-effectivity features. To this end, an inference engine 340 matches the input x to the "closest" (not necessarily in the Euclidean sense) known condition in the stored stereotypical therapy rulebase 350 in order to formulate the "best" therapy parameters. The SOD 330 asserted condition also triggers the neurostimulator 400 through a small delay shown at 390 to ensure the appropriate parameters have finished loading. When the therapy parameters include optimal timing characteristics, such as those in a phase-based therapy procedure, then the necessary additional delay will be introduced in a controlled manner. The signal processing blocks shown in FIG. 8, such as blocks 310, 320, 330, 335, 340, 350, 360, 370 and 390 may be implemented by one or more software or computer programs stored in computer readable medium and executed by a computer or microprocessor.

Signal sensing and therapy delivery to the brain may be made by any electrode wire structures heretofore known or hereinafter developed. For example, the wires/electrodes may be mesh-networked sensors and microstimulators that are operated wirelessly. It should be understood that the system and method described herein may be deployed in various platforms. For example, and not by way of limitation, the system and method described herein may be fully automated and software-driven, and embodied in an external-wearable device (e.g., clipped to the belt) and worn by a patient during the learning/training phases. The patient wears the device after the appropriate IPG components, drug delivery components and related electrodes are implanted and the body-wearable device downloads data collected from the implanted device (e.g., IPG and related electrodes) processes and trains on the data, (or off-loads the data remotely for training on a separate computing apparatus), then uploads the trained software from the separating computing apparatus. After training/learning is complete, the patient need not wear the computing device any longer.

Further still, the system and method of the present invention may be part of a closed-loop continuous control system. That is, the on-the-fly parameter loading operation to the IPG can be made so fast that the unit implanted in the patient may provide essentially continuous feedback control. The implementation of a controller using a therapy rulebase as described herein is similar to fuzzy logic controllers and gain- scheduled controllers (where if the state x(t) of the system is in region A then use parameters A' to control; if in region B then use B', etc.). Other Uses and Applications of the Effectivity Features and Therapy Rulebase Data

There are numerous other uses and applications of the effectivity features described herein. One application is to analyze the effectivity features in order to identify sites in the patient where a therapy would be most effective based on at least one of: sites where earliest neurological disorder activity is detected, sites where high frequency activity occurs early in a neurological disorder event, synchrony with respect to all possible pairs of signals common to two or more adjacent or functionally connected regions, as measured by the appropriate sensors. This concept also applies to detectable and reliable patterns of statistical correlation or neuronal activity, such as fast ripple, ripples, high frequency oscillations, single cell or multicellular activity, and abstract parameters that might measure this activity.

Another use of the effectivity features is to adjust a sensitivity of a neurological disorder event detection process (e.g., seizure detector). Although an ideal neurological event detector would have maximum specificity for a given false negatives tolerance, such performance may be suboptimal in the closed loop of responsive therapy (the objective is to optimize therapy efficacy; this may or may not imply the optimal seizure detector). The effectivity features disclosed herein provide an objective metric that can be used to guide the neurological event detector sensitivity setting. For example, clinical trial experience to date suggests certain patients benefit from hyper-sensitive seizure detectors. These settings trigger therapy more often, even if at the expense of numerous false positives. As another example, if a parameter of interest is known to occur in cellular networks generating epileptic seizures, such as fast ripples (e.g. characteristic 500 Hz oscillations), the system could automatically alter its sensing parameters to enhance detection of these waveforms (by changing sliding windows for feature calculation, gain in the sensed signal, or altering the among particular features may be weighted in a classified-detection and classification system. These are just examples of a variety of options for changing the system's ability to pick up signals of interest for a particular task, as the system "learns" to recognize and map signals that may be important to modulating the target tissue.

Still another use is to store a therapy archive comprising data representing the therapy rulebase data accumulated over time, and deriving from the archive markers to predict efficacy of a therapy in other patients. For example, the aggregate therapy rules learned numerous patients (e.g., 100 or more) during near completion of clinical trials may serve as a machine "registry" or normative database against which predictions can be made for newly enrolled patients. The registry can be data-mined for pre-effectivity features that more frequently resulted in successful therapy outcomes. This provides the bio-signal equivalent of "surrogate markers" for predicting therapy responders, and for matching optimal therapy profiles to new patients. In this way if a particular strategy or chain of automated adjustments in a system lead to the desired clinical result, the eventual target values, parameter/ feature space and optimization trajectory for these features can be used as a potential template to be attempted in other patients. Again, this would not necessarily be used to program the system intelligence to follow the same exact method to optimize therapy in another patient. Rather, it would be included as a possible or potentially preferred strategy for arriving at the appropriate therapy, to be reinforced or abandoned, depending upon whether or not it gave successful results, or led to satisfying optimization or descent-direction criteria.

Another application of the therapy rulebase data is to generate parameters for a preventative treatment designed to interfere with generation of neurological disorder event in a patient, where the parameters for the preventative treatment may comprise at least one of a micro, macro, spatial or frequency characteristic. In this case, successful therapeutic parameters can be employed as a responsive target after detection of conditions or changes in parameters that the device learns to associate with oncoming undesired events. In such a case there can either be movement of intervention parameters in the direction of previously successful therapy in other patients, or perhaps rapid transition to these parameters, should deterioration of conditions be so rapid as to indicate that an undesired target event is imminent. Once the system acquires experience with its patient's events and responses to interventions, this same approach will govern its own stereotypical response, though still with frequent or continuous assessment, to insure that the desired response occurs, given continuous plasticity and changes in circuits and cellular behavior. The system and methods described herein may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The foregoing embodiments are therefore to be considered in all respects illustrative and not meant to be limiting.

Claims

What is claimed is:

1. A method comprising: sensing at least one biological condition indicative of neurological activity in a patient and generating a sensor signal representative thereof; analyzing the sensor signal to detect occurrence of a neurological disorder event and to identify at least one stereotypical onset pattern associated with the neurological disorder event; storing therapy rulebase data that maps real-time conditions of the patient to at least one parameter for at least one therapy to be applied to the patient; and automatically selecting parameters for a therapy for the patient based on the sensor signal and therapy rulebase data.

2. The method of claim 1 , and further comprising analyzing the sensor signal after a therapy has been applied to at least one treated neurological disorder event, and generating effectivity features that represent observable parameters, measures or quantifiable phenomenon just prior to and just after the therapy is applied to treat the neurological disorder event.

3. The method of claim 2, and further comprising generating a measure of success for the therapy based on the effectivity features, and classifying the neurological disorder event as stereotypical or non-stereotypical.

4. The method of claim 3, wherein generating effectivity features is performed with respect to detected and treated neurological disorder events in the patient over time, and further comprising updating the therapy rulebase data based on the effectivity features and the measure of success associated with detected and treated neurological disorder events.

5. The method of claim 4, wherein storing comprises storing the therapy rulebase data within a device that is attached to or implanted in the patient, and further comprising storing in a device or system not implanted or attached in the patient a more comprehensive collection of stereotypical therapy data from which the therapy rulebase data stored in the device is derived.

6. The method of claim 4, and further comprising analyzing the effectivity features to identify sites in the patient where a therapy would be most effective based on at least one of: sites where earliest neurological disorder activity is detected, sites where high frequency activity occurs early in a neurological disorder event, synchrony with respect to all possible pairs of sensor signals common to two or more adjacent or functionally connected regions.

7. The method of claim 4, and wherein automatically selecting comprises selecting parameters for a therapy in real-time based on the effectivity features.

8. The method of claim 4, wherein automatically selecting comprises selecting timing of a therapy based on the effectivity features.

9. The method of claim 4, and further comprising storing archive data representing the effectivity features accumulated over time for the patient, and further comprising computing a probability of therapy efficacy based on the archive data.

10. The method of claim 4, and further comprising adjusting a sensitivity to detection of a neurological disorder event based on the effectivity features.

11. The method of claim 4, and further comprising storing therapy archive data representing the therapy rulebase data accumulated over time, and deriving markers to predict efficacy of a therapy in other patients.

12. The method of claim 1, and further comprising adjusting the therapy rulebase data based on changes associated with a multivariate function that represents effectivity observations from among a finite set of therapy parameter reference profiles.

13. The method of claim 1, wherein the therapy rulebase data maps real-time conditions of the patient into at least one parameter for a therapy that comprises at least one of: applying an electrical stimulation to the patient, applying a magnetic stimulation to the patient, applying a therapeutic drug to the patient, and applying light to modulate light-activated drugs or substances.

14. The method of claim 12, wherein the sensor signal represents brain function of the patient, and further comprising analyzing a linelength feature from the sensor signal and deriving at least one of instantaneous amplitude, frequency and phase for the electrical stimulation.

15. The method of claim 1, and further comprising generating from the therapy rulebase data parameters for a preventative treatment to interfere with generation of a neurological disorder event in the patient, wherein the parameters for the preventative treatment comprise at least one of a micro, macro, spatial or frequency characteristic.

16. The method of claim 1, wherein the neurological disorder event comprises abnormal or undesired movements, behavior, pain, confusion, cognitive dysfunction, loss of awareness, coma, syncope, or similar events caused by dysfunction in heart, lungs, kidneys and liver.

17. A system comprising: a biological sensor subsystem that is configured to sense at least one biological condition indicative of neurological activity in a patient and generates a sensor signal representative thereof; a therapy application subsystem that is configured to deliver at least one therapy to the patient in order to treat a neurological disorder event in the detected in the patient; a control subsystem coupled to the biological sensor subsystem and the therapy application subsystem, wherein the control subsystem is configured to analyze the sensor signal to detect occurrence of a neurological disorder event in the patient and to identify at least one stereotypical onset pattern associated with the neurological disorder event, store therapy rulebase data that maps real-time conditions of patient to at least one parameter for a least one therapy to be applied to the patient by the therapy application subsystem, and automatically select parameters for a therapy for the patient based on the sensor signal and the therapy rulebase data.

18. The system of claim 17, wherein the control subsystem is configured to analyze the sensor signal after a therapy has been applied to at least one treated neurological disorder event, and generate effectivity features that represent observable parameters, measures or quantifiable phenomenon just prior to and just after the therapy is applied to treat the neurological disorder event.

19. The system of claim 18, wherein the control subsystem is configured to generate a measure of success for the therapy based on the effectivity features, and classify the neurological disorder event as stereotypical or non-stereotypical.

20. The system of claim 19, wherein the control subsystem is configured to generate effectivity features respect to detected and treated neurological disorder events in the patient over time, and update the therapy rulebase data based on the effectivity features and the measure of success associated with detected and treated neurological disorder events.

21. The system of claim 17, wherein the therapy application subsystem is configured to apply at least one of an electrical stimulation to the patient, a magnetic stimulation to the patient, and a therapeutic drug to the patient.

22. The system of claim 17, wherein at least a portion of the biological sensor subsystem, therapy application subsystem and control subsystem is configured to be implanted within or attached to a body of the patient.

23. A device comprising: at least one biological sensor that is configured to sense at least one biological condition indicative of neurological activity in a patient and to generate a sensor signal representative thereof; at least one therapy applicator that is configured to deliver at least one therapy to the patient in order to treat a neurological disorder event in the detected in the patient; a controller coupled to the at least one biological sensor and to the at least one therapy applicator, wherein the controller is configured to: analyze the sensor signal to detect occurrence of a neurological disorder event in the patient and to identify at least one stereotypical onset pattern associated with the neurological disorder event; store therapy rulebase data that maps real-time conditions of patient to at least one parameter for a least one therapy to be applied to the patient by the therapy application subsystem; and automatically select parameters for a therapy for the patient based on the sensor signal and the therapy rulebase data.

24. The device of claim 23, wherein the controller is configured to analyze the sensor signal after a therapy has been applied to at least one treated neurological disorder event, and generate effectivity features that represent observable parameters, measures or quantifiable phenomenon just prior to and just after the therapy is applied to treat the neurological disorder event.

25. The system of claim 24, wherein the controller is configured to generate a measure of success for the therapy based on the effectivity features, and classify the neurological disorder event as stereotypical or non-stereotypical.

26. The system of claim 25, wherein the controller is configured to generate effectivity features respect to detected and treated neurological disorder events in the patient over time, and update the therapy rulebase data based on the effectivity features and the measure of success associated with detected and treated neurological disorder events.