Research
NEUROPHYSICS
The current clinical process for evaluating neurological or neurodevelopment disorders from an electroencephalography (EEG) recording, relies on a clinician’s manual inspection. The electrodynamic signal of the brain is exceptionally complex, and EEG recordings can be minutes, hours, even days long. In contrast, pathologic activity may occur in just a fraction of a second.
Quantifying epileptogenicity, the tendency of the brain to have a seizure, is a particularly interesting starting point for EEG analysis, as it has a physical manifestation unlike many other neurological conditions. Seizures occur upon the spontaneous synchronization of neurons in the brain. Thus the electrodynamic state of a highly epileptogenetic brain will differ from the state of “normative” brain which has a low chance of neuronal synchrony. Deciphering these electrodynamic states into actionable digital biomarkers of disease requires creative and novel methodologies.
Our approach considers the state of the brain as a trajectory in a nonlinear dynamical system.
A dynamical system changes over time based on a series of variables and constraints. In the brain, the variables would be the neurons and the constraints are the neural connectivity.
The brain has over 10 to the 11 neurons or dimensions, so the above phase space diagrams are just a 3 dimensional visual rendering.
In a non-epileptic patient, the brain electrodynamics follow a particular trajectory, which leads to no seizures. Even though the trajectory doesn’t overlap with one that does has seizures, some significant external input could push the brain to seize. However in an epileptic brain, the barrier to seize is much lower, and the brain dynamics can switch between the seizure trajectory and the non-seizure trajectory. The above diagram illustrates this with the separate versus overlapping trajectories.
We compute multiple nonlinear invariants that quantify the trajectory and represent the dynamical system, such as determinism, sample entropy, and the Lyapunov exponents. We compute these features, for all sensors on the 10-20 system, at multiple frequency bands.
Determinism
Sample Entropy
Lyapunov Exponents
This creates a 3D tensor describing the electrodynamics of the brain. Identifying the vectors within this tensor that align with a particular pathology would be a digital biomarker.
DIGITAL BIOMARKERS
The choice and dosage of antiseizure medication is currently guided by trial and error. A neuro-physiological biomarker for the effect of medication on epileptic seizure activity would be useful for clinical decision making. As an early step in this direction, we evaluate changes in brain dynamics of patients on varying amounts of medication.
Deep learning models have achieved state-of-the-art results in a wide variety of tasks in computer vision, natural language processing and speech recognition. The fact that deep learning models automatically learn abstract feature representations from raw features, while also optimizing on the target prediction tasks, makes them an attractive solution for analyzing wearable device data.
Sleep-related epilepsies are particularly difficult to diagnose during routine EEGs because the patient is awake, and the typical epileptic signatures are not visible. Our approach identifies a fundamental brain electrodynamic difference in patients with a sleep-related epilepsy, even when traditional neurophysiology approaches are unable.
Epilepsy is a spectrum disorder consisting of various heterogenous neurological conditions. We apply our nonlinear approach to a cohort of patients with heterogenous epilepsy conditions to see if the brain dynamics of patients with epilepsy differs fundamentally differ from patients with "normal" brain activity.
REAL-TIME MEDICATION EFFECTS
Currently there is a large variation in the anti-seizure drug choice and dosage, across clinicians as well as institutions. While those decisions are guided by clinical factors, it can often resort to months of trial and error. Comparing the brain electrodynamics of a patient with epilepsy when on versus off anti-seizure medication, can provide insights regarding medication efficacy and also presents a paradigm of different inferred epileptogenicity levels. Computing nonlinear measures, such as sample entropy, determinism and the Lyapunov exponents, on a group of patients with epilepsy, we were able to distinguish between the different brain states (high versus low medication) using only the comparison of two thirty-second EEGs. We were also able to retrospectively localize the seizure onset or irritative region by measuring its relative reaction to medication compared with the rest of the brain.
RESULTS
The results show a difference between the dynamics of the irritative zone of the brain for ASM-on periods versus off. The below figures illustrate the change in sample entropy for the 4 paradigms: ASM- off and non-irritative, ASM-off and irritative, ASM-on and non-irritative, ASM-on and irritative. The change in dynamics of the irritative region from off to on ASM, was at least 95% greater than the change in corresponding dynamics on the non-irritative region, for sample entropy, laminarity,
determinism, and the Lyapunov exponents. Also, for the irritative region, the dynamical state of ASM-on periods converges 34% closer to the state of the non-irritative region, as compared to the state when off ASMs.
The sensors are grouped into 2 zones: the clinically determined seizure onset zone, and the rest of the brain. In the below graphs, the nonlinear invariants are plotted over multiple frequencies across each group. The effect of medication on the brain electrodynamics can be observed on the seizure onset region of the brain by seeing the difference between the solid blue lines, and the solid red lines. It can also be observed that the corresponding effect on the rest of the brain, the dotted line, is much smaller across all nonlinear measures. Moreover, the brain dynamics of the seizure onset region of the brain changes to be more similar to the rest of the brain when on on medication, indicating efficacy.
To better illustrate the results, below are pictographs on the 10-20 EEG system for each nonlinear invariant.The colouring corresponds to the area under the multi scale curve bounded by the Nyquist limit. The location of the seizure onset zone differs for each patient, this selection is illustrative. There is an observable difference between the brain when on versus off medication, particularly for the seizure onset zone. Moreover, medication changes the electrodynamics in the seizure onset zone to be similar to the rest of the brain, indicating efficacy.
For more details, please refer to the following for more info:
Sathyanarayana, Aarti, Rima El Atrache, Michele Jackson, Aliza S. Alter, Kenneth D. Mandl, Tobias Loddenkemper, and William Bosl. "Measuring drug effects on brain dynamics through electroencephalography". American Epilepsy Society (December 2019) (abstract) (poster)
"SEE WHAT CAN'T BE SEEN"
Benign childhood epilepsy with centro-temporal spikes (BECTS), also known as Rolandic Epilepsy, is one of the most common forms of focal childhood epilepsy. Despite its prevalence, it is often misdiagnosed or missed entirely. While most research has focused on early warning signs of seizure onset, a more difficult challenge is to monitor brain electrodynamics in order to detect functional characteristics - biomarkers - that indicate the brain has entered a dynamical state in which seizures are likely to occur. This has the potential to predict clinical outcomes.
Seizures associated with BECTS occur most frequently during sleep. They present with unilateral tonic and clonic features, and at times generalized tonic clonic seizures. Because these seizures occur during sleep, BECTS often goes undiagnosed. Moreover, BECTS usually remits before the age of 18. Due to the high rates of remission during adolescence as well as the long-term effects of anti-epileptic medication, children diagnosed with BECTS often go untreated. Despite the name, this benign epilepsy has been repeatedly associated with developmental delays4. The decision of whether to treat BECTS with anti-epileptic drugs or not remains unclear.
Statistically significant group differences between BECTS patients and controls were found for several nonlinear features, particularly determinism, (p < 10e-4 for all frequency bands). Determinism is a measure of dynamical stability and non-randomness, which has been hypothesized to be related to seizure risk.
RESULTS
The below heatmap illustrates the area under the curve (AUC) of a multiscale curve for determinism, on the awake brain of a BECTS patient versus a control. The irritative zone indicates the region of the brain that BECTS patients have identifiable epileptiform activity at night. The BECTS patients have no identifiable epileptiform activities when awake.
BECTS
Controls
Sample Entropy
Lmax
Lmean
TT
Frequency
Value
Frequency
Lmean
Lmax
Value
Sample Entropy
TT
Frequency
Lmean
Lmax
Value
Sample Entropy
TT
Frequency
Lmean
Lmax
Value
Sample Entropy
TT
For more details, please refer to the following publication for more info:
Sathyanarayana, Aarti, Rima El Atrache, Michele Jackson, Aliza S. Alter, Kenneth D. Mandl, Tobias Loddenkemper, and William Bosl. "A Digital Biomarker for Benign Childhood Epilepsy with Centrotemporal Spikes (BECTS)." Preprint available at SSRN 3487819 (2019).(link) (poster)
HETEROGENEOUS EPILEPSIES
Does this methodology still work with a more heterogeneous epilepsy dataset? We decided to find by testing on a ohort of patients that included different epileptic encephlopathies. We were still able to find a clean distinction when compared with controls.
RESULTS
We used a random forest to identify which EEG sensors were of most use to the prediction of case versus control. The below figure shows these sensors and the corresponding box plots color coded by frequency band, for determinism.
To ensure that the nonlinear measures were not being influences by epileptiform activity, we compared our sensor selection with the frequency of spiking across all patients for each sensor. As can be seen below, the sensors were not the same.