Predicting epileptic seizures

Abstract

The present disclosure relates to a system for predicting epileptic seizures, the system comprising: an input interface configured to receive electroencephalogram signals, wherein the electroencephalogram signals represent electrical activity of a brain of a subject; a spike encoder operatively coupled to the input interface, the spike encoded configured to convert the electroencephalogram signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space;a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of an epileptic seizure. The disclosure further relates to a computer-implemented method for processing electroencephalogram signals representing electrical activity of a brain of a subject and to a computer- implemented method executed on a computing device for training a spiking neural network for processing physiological electroencephalogram signals.

Description

[0001] P7152PC00

[0002] 1

[0003] System and method for predicting epileptic seizures

[0004] The present disclosure relates to a system for predicting epileptic seizures using a spiking neural network. The disclosure further relates to a computer-implemented method for processing electroencephalogram signals representing electrical activity of a brain of a subject using the spiking neural network and to a computer-implemented method for training the spiking neural network.

[0005] Background

[0006] Epilepsy is a prevalent neurological disorder characterized by the occurrence of spontaneous seizures. The ability to predict these seizures can significantly improve the quality of life for individuals suffering from epilepsy, as it enables timely intervention and management of the condition. Conventionally, seizure prediction has relied heavily on the analysis of electroencephalogram (EEG) signals, which provide a direct measure of electrical activity in the brain. These signals are typically processed using various computational techniques to identify pre-ictal states indicative of an impending seizure.

[0007] Despite advancements in EEG technology and signal processing algorithms, current methods still face significant challenges. The primary limitation is the lack of specificity and sensitivity in the predictions. Many systems fail to consistently differentiate between pre-ictal and interictal states (normal brain activity outside of seizures), leading to high rates of false positives and false negatives. Additionally, existing prediction methods often require substantial computational resources and real-time processing capabilities, which can limit their applicability in portable or wearable devices that are ideal for continuous monitoring.

[0008] Furthermore, traditional seizure prediction systems do not adequately exploit the temporal dynamics of neural activity, as they often rely on static thresholds or simple feature extraction methods that fail to capture the complex, nonlinear patterns typical of EEG data. This limitation restricts their effectiveness, as the predictive algorithms cannot adequately adapt to the variability inherent in individual patient data, which is crucial for personalized medicine approaches. P7152PC00

[0009] 2

[0010] It is therefore an objective of the present disclosure to develop a system for predicting epileptic seizures that addresses these shortcomings. The proposed system aims to enhance the accuracy, reliability, and computational efficiency of seizure prediction. This system is intended to improve the specificity and sensitivity of seizure predictions while being suitable for implementation on portable devices, thereby making real-time monitoring feasible and effective for patients with epilepsy.

[0011] Summary

[0012] The present disclosure relates to a system for predicting epileptic seizures, the system comprising: an input interface configured to receive electroencephalogram signals, wherein the electroencephalogram signals represent electrical activity of a brain of a subject; a spike encoder operatively coupled to the input interface, the spike encoder configured to convert the electroencephalogram signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of an epileptic seizure.

[0013] This system leverages a spiking neural network to enhance the specificity and sensitivity of seizure predictions by utilizing the dynamics of spike data. The inclusion of a recurrent layer allows the system to handle high-dimensional data effectively, thus improving the system's ability to identify complex patterns indicative of impending seizures. By projecting spike-encoded EEG signals into a higher-dimensional space, the system enhances the separability of pre-ictal and non-pre-ictal states, which can lead to more accurate predictions.

[0014] The system may be configured such that the recurrent layer includes neurons that are randomly connected, supporting robust feature extraction from EEG signals by capitalizing on the diverse response patterns of the neurons. The readout layer, trained specifically to identify pre-ictal activities, uses a subset of neurons from the recurrent layer, optimizing the processing power and reducing computational load, making the P7152PC00

[0015] 3 system suitable for real-time applications on portable devices. The dual-channel spike encoding approach allows the system to handle varying signal amplitudes, thus providing a nuanced analysis that accommodates different intensities of brain activity, which can vary widely among patients and across different seizures.

[0016] The system features a combination of a recurrent layer and a readout layer. The recurrent layer can handle and enhance the complexity of the incoming EEG data. By projecting these signals into a higher-dimensional space, the recurrent layer effectively captures and retains the intricate temporal dynamics inherent in the brain's electrical activity, which can be used to identify the nuanced patterns that precede seizures.

[0017] This high-dimensional representation of spike data is then analyzed by the readout layer. The readout layer comprises a selected subset of neurons, optimized to focus on the most informative features of the pre-ictal state data extracted from the recurrent layer's output. This may reduce the computational demand on the system while enhancing the efficiency and accuracy of the seizure prediction. The training of these neurons is tailored to differentiate between various pre-ictal activities, enabling the system to recognize the onset of seizures with increased precision.

[0018] Overall, the combination of these two layers within the neural network not only increases the predictive capability of the system but also ensures that it can operate efficiently in real-time environments, such as wearable devices for continuous monitoring. This makes it an invaluable tool for epilepsy management, potentially reducing the incidence and impact of unexpected seizures on patients' daily lives.

[0019] Prediction of epileptic seizures represents a substantially more difficult technical challenge than the detection of seizures. Seizure detection, which aims to identify an ictal event while or after it is already occurring, can rely on strong and overt features in the electroencephalogram signals, such as large-amplitude oscillations or characteristic rhythmic patterns. By contrast, seizure prediction requires recognition of pre-ictal activity, which often consists of subtle, noisy, and patient-specific changes in brain dynamics that occur minutes or even seconds before a seizure. These pre-ictal signals are less pronounced and may be confounded by normal variations in brain activity. The presently disclosed system addresses this difficulty by using a spike encoder, a spiking neural network configured to process a series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer P7152PC00

[0020] 4 configured to project the series of spikes into spike data of a higher dimensional space; and a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space. Moreover, as described in further detail below, the system may comprise a dual-channel spike encoding, specialized competing readout neurons that separately learn to respond to pre-ictal versus normal interictal activity. Together, these features may enable the system to discriminate between safe and pre-ictal brain states with improved sensitivity and specificity, thus tackling the harder problem of prediction rather than mere detection.

[0021] The present disclosure further relates to computer-implemented methods for processing electroencephalogram signals representing electrical activity of a brain of a subject and for training a spiking neural network to perform such predictions. These methods ensure that the system is continually refined based on the latest available data, which includes labeled EEG signals representing known pre-ictal and non-pre- ictal states. This training enables the readout layer to adaptively improve its predictive accuracy over time, using real-world data to refine its algorithms.

[0022] Description of drawings

[0023] Various embodiments are described hereinafter with reference to the drawings. The drawings are examples of embodiments and are intended to illustrate some of the features of the presently disclosed system and method for processing electroencephalogram signals representing electrical activity of a brain of a subject. The drawings are not to be construed as limiting to the presently disclosed invention.

[0024] Fig. 1 shows an embodiment of the presently disclosed system for predicting epileptic seizures.

[0025] Fig. 2 shows an example of generation of spikes using a high channel and a low channel.

[0026] Fig. 3 shows an example of spike generation using a threshold.

[0027] Fig. 4 shows an example of an implementation of decision making using the reduced set of readout neurons.

[0028] Fig. 5 shows EEG signals and categorizing of the EEG signals into pre-ictal activity leading to seizure and activity not leading to seizure. P7152PC00

[0029] 5

[0030] Fig. 6 shows a flow chart of a method according to an embodiment of the presently disclosed method for processing electroencephalogram signals representing electrical activity of a brain of a subject.

[0031] Fig. 7 shows a flow chart of a method according to an embodiment of the presently disclosed computer-implemented method for training a spiking neural network for processing electroencephalogram signals.

[0032] Detailed description

[0033] The present disclosure relates to a system for predicting epileptic seizures, the system comprising: an input interface configured to receive electroencephalogram signals, wherein the electroencephalogram signals represent electrical activity of a brain of a subject; a spike encoder operatively coupled to the input interface, the spike encoded configured to convert the electroencephalogram signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of an epileptic seizure. The recurrent layer may be a reservoir.

[0034] Fig. 1 shows an example of the presently disclosed system 100 for predicting epileptic seizures. In the example there is an input interface 101 configured to receive electroencephalogram signals 102. A spike encoder 103 is operatively coupled to the input interface 101 . The spike encoder is configured to generate two series of spikes 106 from input channels 104. The system 100 further comprises a spiking neural network 105 comprising a recurrent layer 107 and a readout layer 108 comprising a set of readout neurons N0,i , No, 2 trained to distinguish between different pre-ictal activities. The example further comprises a decision unit 109 and an output interface 110.

[0035] The system for predicting epileptic seizures begins with an input interface designed to receive electroencephalogram (EEG) signals. EEG signals represent the electrical activity of a brain of a subject, capturing fluctuations in voltage resulting from ionic P7152PC00

[0036] 6 current flows within neurons. This activity is indicative of the brain's functional state at any given time and can be used for identifying potential anomalies that precede epileptic seizures.

[0037] As used herein, the term “subject” refers to a human or animal individual from whom electroencephalogram signals can be obtained. The subject may be a person affected by epilepsy or another neurological condition, or a healthy individual, for example in a research, monitoring, or training context. Similarly, the subject may be an animal, such as in veterinary applications or experimental models. Thus, the term “subject” is intended to be broad and not limited to patients under medical care.

[0038] The system may be further equipped with multiple electrodes that are strategically placed on the patient’s scalp according to standardized EEG placement protocols, such as the International 10-20 system. These electrodes are preferably non-invasive and designed to detect minute electrical changes on the scalp that arise from the brain's neural activity. The role of the input interface is not only to capture these signals accurately but also to ensure that they are recorded with high fidelity, minimizing noise and artifacts that could interfere with the analysis. This is achieved through the use of high-quality, sensitive electrodes and appropriate filtering technology.

[0039] The input interface may also include analog-to-digital converters (ADCs). These converters transform the analog EEG signals into digital form so that they can be processed by computational tools. The ADCs sample the continuous EEG waveform at a high frequency, ensuring that the digital representation maintains the integrity of the original signal’s temporal and amplitude resolutions. This digitization allows for the advanced processing techniques required in later stages of the seizure prediction system.

[0040] Once the EEG signals are digitized, they are transmitted from the input interface to the spike encoder. This transmission is carried out via secure and efficient data communication channels to ensure real-time processing capabilities. The input interface is configured to handle a wide range of EEG signal intensities and variations, making the system robust and adaptable to different patient conditions and environments. P7152PC00

[0041] 7

[0042] In one embodiment of the system for predicting epileptic seizures, the input interface comprises analog-to-digital converters (ADCs) configured to sample streams of analog electroencephalogram signals into series of digital electroencephalogram signals. These ADCs are configured to sample streams of analog electroencephalogram (EEG) signals and convert them into a series of digital signals. The inclusion of ADCs is useful for facilitating the transition from analog signals, which represent the natural electrical activity of the brain as detected by scalp electrodes, to digital form, which is necessary for subsequent digital processing and analysis.

[0043] The ADCs may operate at various sampling rates, which can be selected based on the specific requirements of the EEG analysis. Typically, EEG signals are sampled at a high frequency to ensure that the digital conversion retains the detailed information present in the original analog signal. This high sampling rate is useful for capturing the rapid fluctuations in brain activity that are indicative of potential epileptic events. Moreover, the system may be configured with ADCs that provide high-resolution digital outputs. This high resolution allows for a more precise representation of the EEG signal’s amplitude, which can be important for detecting subtle changes in brain activity that precede a seizure. The precise configuration of the ADCs - such as bit rate and conversion speed - can be tailored to optimize both the accuracy and the efficiency of the seizure prediction system.

[0044] Additionally, the use of ADCs in the input interface ensures that the EEG data, once digitized, is compatible with modern digital signal processing (DSP) techniques. These DSP techniques are applied downstream by other components of the system, such as the spike encoder and the spiking neural network, to analyze the EEG data and identify predictive indicators of epileptic seizures. This configuration ensures a seamless integration of the input interface with the rest of the system, facilitating a robust and responsive platform for seizure prediction.

[0045] In one embodiment of the system for predicting epileptic seizures, the spike encoder is configured to generate at least two series of spikes, utilizing a dual-channel approach that includes a high channel and a low channel. This configuration enables the spike encoder to process electroencephalogram (EEG) signals with enhanced specificity by categorizing the spikes based on their amplitude thresholds. P7152PC00

[0046] 8

[0047] The high channel may be tailored to detect spikes or variations in amplitudes that exceed a higher amplitude threshold, which typically represent more significant or abrupt changes in the brain's electrical activity. These high-amplitude-related spikes can be indicators of potential seizure activity, as they often precede or occur during epileptic events. Conversely, the low channel may be designed to capture variations exceeding a lower amplitude threshold, which includes subtler variations in the EEG signals.

[0048] The dual-channel configuration can enhance the sensitivity and accuracy of the seizure prediction system. By separately processing high and low amplitude spikes, the system can more accurately distinguish between different types of neural activities, which can be used, for example, to reduce false positives and enhance the reliability of seizure predictions. This approach allows for a nuanced analysis of the EEG data, where significant events are highlighted by the high channel, and ongoing neural states are continuously monitored by the low channel.

[0049] The dual-channel configuration provides a more detailed and dynamic understanding of the patient's brain activity, enabling the spiking neural network to make more informed predictions based on a rich dataset that includes both pronounced and subtle neural signals. This dual-channel approach not only improves the predictive capabilities of the system but also ensures that the monitoring is comprehensive, covering a broader range of neural activities that could be relevant for detecting and predicting seizures.

[0050] In one embodiment of the system for predicting epileptic seizures, the spike encoder is configured to generate a series of spikes for signals exceeding a comparison threshold through the high channel and through the low channel otherwise.

[0051] The presently disclosed dual-channel spike encoding approach may be particularly advantageous when combined with the pre-ictal detection functionality of the spiking neural network. By feeding both high-amplitude spikes, which emphasize pronounced changes, and low-amplitude spikes, which capture more subtle background activity, into the pre-ictal-sensitive and non-pre-ictal-sensitive readout neurons, the system achieves a more reliable separation between seizure-related and normal activity. The cooperation of dual-channel encoding with the discriminative readout neurons enables the system to improve both sensitivity and specificity in detecting pre-ictal activity. P7152PC00

[0052] 9

[0053] Fig. 2 shows an example of generation of spikes using a high channel and a low channel. It can be seen that SFEhigh spikes are generated every time the amplitude exceeds a first amplitude threshold (compared to the preceding value). Similarly, SFEiow spikes are generated every time the amplitude exceeds a second amplitude threshold (compared to the preceding value). As can be seen the encoder is not necessarily updated every time a new value is measured. The updates can occur at a slower rate.

[0054] Fig. 3 shows an example of spike generation using a threshold, wherein the base is updated to the sampled value (reset) every time the value exceeds the threshold.

[0055] The high channel may be set up to generate a first series of spikes from EEG signals that exceed a first, higher amplitude threshold. This function may be valuable for identifying significant neural events that often correlate with pre-seizure activity or actual seizures. By focusing on these higher amplitude spikes, the high channel ensures that the more critical and potentially alarming spikes are captured and prioritized for analysis.

[0056] Conversely, the low channel is designed to capture a second series of spikes from EEG signals that exceed a second amplitude threshold, which is set lower than the first. This allows for the inclusion of subtler brain activities in the analysis. These activities, though less pronounced, can provide important context about the patient's overall neural state and help in distinguishing between normal brain activity variations and those indicative of a developing seizure.

[0057] The spike encoder may utilize a dynamic approach to generate spikes, wherein, if a current amplitude of an electroencephalogram signal is a greater than a preceding amplitude of the electroencephalogram signal and a comparison threshold, a spike is generated through the high channel, else if a current amplitude of an electroencephalogram signal is a greater than a preceding amplitude of the electroencephalogram signal and lower than the comparison threshold, a spike is generated through the low channel.

[0058] For the high channel, a spike may be generated if the current amplitude of the EEG signal exceeds both the preceding amplitude and a predefined high threshold value. P7152PC00

[0059] 10

[0060] This criterion ensures that only significant increases in neural activity, which may indicate emerging seizure activity, trigger the high-priority alert system.

[0061] The system may be configured such that a first threshold defines how much increase in amplitude generates and spike, and a second threshold defines which channel, i.e. either the high channels or the low channel, receives the spike.

[0062] Conversely, for the low channel, a spike may be generated if the current amplitude exceeds the preceding amplitude but only needs to surpass a lower threshold value. This setup captures more gradual or less pronounced changes in brain activity, which are still important but may not necessarily indicate an immediate seizure risk. This method of generating spikes based on comparing current and preceding amplitudes to respective thresholds enhances the system’s ability to detect and respond to different levels of neural activity changes, providing a more nuanced and responsive approach to seizure prediction.

[0063] A recurrent layer within a spiking neural network is a component for processing and enhancing signal data. This layer comprises a network of neurons that are interconnected in a complex pattern, often initialized randomly. These connections may include both feedforward and recurrent pathways, allowing the network to not only process incoming data but also retain a memory of past inputs. This memory capability is useful for recognizing and analyzing temporal patterns in data, such as the electroencephalogram (EEG) signals used in seizure prediction systems.

[0064] The primary role of the recurrent layer is to transform incoming spike sequences into a higher-dimensional representation. This expanded data space enhances the network's ability to differentiate between various neural activity states, making it more effective at identifying precursors to epileptic seizures. The dynamic nature of this layer, with its ongoing internal feedback loops, enables it to adapt and respond to new information while considering historical data, providing a robust mechanism for real-time signal analysis.

[0065] In one embodiment of the system for predicting epileptic seizures, the spiking neural network includes a recurrent layer equipped with a set of randomly connected neurons. This configuration facilitates a diverse and dynamic representation of neural activities, P7152PC00

[0066] 11 enhancing the network’s ability to process complex patterns inherent in electroencephalogram (EEG) signals.

[0067] The recurrent layer's randomly connected neurons serve as a rich source of temporal and spatial variations in neural signaling, which is useful for capturing the nuanced changes that may indicate the onset of an epileptic seizure. From this extensive network, the readout layer selectively utilizes a subset of these neurons to focus on the most relevant features for seizure prediction. This selective approach allows the system to concentrate computational resources on the most informative signals, thereby increasing efficiency and improving the accuracy of seizure forecasts. The subset may be a reduced selection of reservoir neurons whose spiking activity is connected to the readout neurons. The subset may be randomly chosen, predefined by connectivity rules, or selected during training.

[0068] The readout layer's strategy involves analyzing the outputs from this subset of neurons to identify specific indicators that predict epileptic seizures. By doing so, it effectively filters and consolidates the high-dimensional data processed by the recurrent layer, isolating patterns that are most strongly associated with pre-ictal states.

[0069] Optionally, the system may include a number, for example 128, of Leaky Integrate-and- Fire (LIF) neurons within the recurrent layer. LIF neurons are particularly suitable for this application due to their ability to closely mimic biological neuronal behavior through the integration and firing mechanism based on threshold activation. The use of LIF neurons can enhance the system's biological plausibility and responsiveness, providing a more accurate simulation of neuronal dynamics during seizure onset. This setup ensures that the network is not only robust in handling variations in EEG data but also efficient in its operation, making real-time seizure prediction more feasible and reliable.

[0070] The recurrent layer may comprise a reservoir network comprising a set of neurons, such as a set of randomly connected neurons, wherein the readout layer is configured to use a subset of the set of neurons to identify the indicators predictive of an epileptic seizure. The subset of the recurrent layer neurons may be connected by a fully connected or sparsely connected layer, directly or indirectly, to a set of readout neurons. In the case of a fully connected layer, every neuron in the chosen subset from the recurrent layer is linked to each neuron in the readout layer. This connectivity ensures that all detailed interactions between the extracted features (neural activities) and the predictive mechanisms are meticulously considered. Such an arrangement is P7152PC00

[0071] 12 beneficial for capturing complex patterns and dependencies within the spike data, thereby facilitating a thorough analysis and enhancing the predictive accuracy of the system.

[0072] Alternatively, employing a sparsely connected layer involves selectively establishing connections between neurons, based on specific criteria such as their functional relevance to seizure prediction, proximity, or similarity in response patterns. Sparse connectivity focuses on the most significant neural pathways, reducing computational demands and potentially increasing the efficiency and responsiveness of the system. This approach helps in minimizing unnecessary computational complexity and can be particularly effective in reducing overfitting, thereby improving the model's generalizability across different seizure types and patient conditions.

[0073] The readout layer may be configured to consolidate the spike data of the higher dimensional space from the recurrent layer to identify the indicators predictive of an epileptic seizure.

[0074] One function of the readout layer is to sift through the high-dimensional spike data to identify specific patterns or indicators that are predictive of an epileptic seizure. These indicators might include particular sequences or clusters of spikes, unusual levels of neural activity, or abnormal synchronizations across different neuron groups, all of which could suggest the imminent onset of a seizure.

[0075] The readout layer may utilize advanced algorithms and neural processing techniques to effectively distill this enriched data into actionable insights. It may examine the enhanced spike data for correlations, trends, and other statistical features that have been previously associated with pre-ictal states. By focusing on these characteristics, the readout layer is able to isolate the most relevant information from the broad spectrum of neural signals received.

[0076] This consolidation process involves sophisticated data reduction and filtering techniques that preserve essential diagnostic information while discarding redundant or irrelevant data. The goal is to streamline the amount of data to be analyzed, thereby speeding up the decision-making process without compromising the accuracy of seizure predictions. This is particularly important in real-time monitoring systems, where speed and accuracy are crucial for timely intervention. P7152PC00

[0077] 13

[0078] The readout layer may be supported by machine learning models that have been trained on historical data to recognize the specific signatures of seizures. These models continuously refine their predictive accuracy based on feedback from ongoing monitoring and new data inputs. The integration of machine learning enhances the adaptability of the readout layer, allowing it to improve over time and potentially personalize its predictive capabilities to individual patients' patterns.

[0079] In one embodiment of the system for predicting epileptic seizures, the training of the readout layer involves a reward-based mechanism, which enhances the accuracy and specificity of the system in detecting seizures. This training approach strategically employs reward-modulated learning, where specific readout neurons are conditioned to respond to particular types of neural activities.

[0080] Fig. 5 shows EEG signals and categorizing of the EEG signals into pre-ictal activity leading to seizure and activity not leading to seizure. The upper diagram shows EEG amplitudes over time. The lower diagram shows corresponding labelling of the EEG amplitudes, wherein T means pre-ictal activity leading to seizure and ‘2’ means pre- ictal activity not leading to seizure. This can be used as training data for training the reduced set of readout neurons.

[0081] The readout layer may comprise at least a first readout neuron which has been rewarded when it spikes during interictal activity leading to a seizure during training. Rewarding the first readout neuron neurons specifically when it responds to seizure activity ensures that it becomes highly sensitive to the patterns and signals that typically precede or accompany a seizure. This conditioning helps the neuron to specialize in recognizing and reacting to seizure-indicative data.

[0082] The readout layer may comprise at least a second readout neuron which has been rewarded when it spikes during no-seizure interictal activity. Training this neuron to be active during normal brain activity, and rewarding it for correct identification of such states, helps the system to distinguish between normal and pre-ictal (before seizure) neural activities effectively. This differentiation may reduce false positives, where the system might incorrectly predict a seizure when none is imminent. P7152PC00

[0083] 14

[0084] In one embodiment, the readout layer comprises a combination of at least a first readout neuron and a second readout neuron, each trained to respond to different classes of electroencephalogram (EEG) activity. The first readout neuron is conditioned during training by rewarding spiking activity in response to interictal signals that precede or accompany epileptic seizures, while the second neuron is trained to suppress spiking activity under the same conditions. As a result of this reward-based training, just the first readout neuron becomes highly sensitive to seizure-indicative patterns, allowing it to specialize in identifying pre-ictal activity. The second readout neuron, by contrast, may be rewarded when it spikes during no-seizure interictal activity, while spiking by the first neuron is actively punished by the training algorithm. Through this reinforcement, the second readout neuron is trained to reliably respond to EEG patterns corresponding to normal brain states. The complementary training of the first and second readout neurons enables the system to establish a functional distinction between seizure-related and non-seizure-related activity. The cooperation of these specialized neurons reduces the incidence of false positives, since the second neuron provides a balancing signal against erroneous seizure predictions, while the first neuron provides sensitivity to pre-ictal signatures. The joint effect is an improvement in both sensitivity and specificity of seizure prediction.

[0085] The first and second readout neurons may be regarded as competing neurons, each specialized to represent opposing brain states. The first readout neuron is rewarded during training when it spikes in response to pre-ictal activity. Conversely, the second readout neuron is rewarded when it spikes during safe interictal states. During operation, the two neurons may act in competition: when the first neuron dominates in spiking activity relative to the second neuron, for example within a moving time window, this indicates a transition toward a pre-seizure state. When the second neuron dominates, the system interprets the activity as a safe brain state. This competitive interplay may enhance the robustness of the classification.

[0086] These training strategies employ principles of reinforcement learning, where neurons are essentially 'taught' through feedback - positive reinforcement in the form of rewards - to refine their responses based on the accuracy of their activity relative to actual brain states. Rewarding neurons for specific types of spikes ensures that each neuron's sensitivity is aligned with its target stimulus, enhancing the overall reliability of the system. P7152PC00

[0087] 15

[0088] The system may be further configured to estimate a likelihood of a seizure occurring by analyzing spikes of the first readout neuron and the second readout neuron within a moving time window. The analysis may be conducted within a moving time window that continuously updates to reflect the most recent data, allowing the system to evaluate transient neural patterns that may signify an impending seizure.

[0089] By continuously monitoring spikes within this moving time window, the system integrates outputs from both neuron types, assessing their activity patterns and levels. For example, an increase in activity from the first neuron type compared to the second may indicate a rising probability of a seizure. This method of real-time assessment leverages the distinct training of the neurons to enhance the system’s predictive accuracy.

[0090] The moving time window may extend over a predetermined interval, such as a few seconds up to several minutes, and is continuously updated as new electroencephalogram data is received and processed. By considering the spikes of the first readout neuron, which is trained to respond to pre-ictal activity, together with the spikes of the second readout neuron, which is trained to respond to normal interictal activity, the system derives a relative measure of pre-ictal versus non-pre-ictal state probabilities. Evaluating these spiking activities within the temporal context of the moving time window improves the robustness of seizure prediction by filtering out transient fluctuations or noise that may otherwise cause spurious detections. This temporal smoothing enables the system to provide a likelihood score of seizure imminence, rather than a binary output, thereby enhancing both sensitivity and specificity. For example, a gradual increase in the firing rate of the first readout neuron relative to the second readout neuron over successive time windows may indicate that a seizure is becoming more likely. Conversely, stable or dominant spiking of the second readout neuron within the time window may indicate that the brain activity remains in a normal state. Accordingly, the moving time window analysis may combine the discriminative specialization of the readout neurons with temporal integration, resulting in a more reliable estimate of pre-ictal activity and reducing the number of false positives compared to instantaneous classification.

[0091] Furthermore, the system may utilize adaptive thresholds that adjust based on learned data and ongoing monitoring, which refines the decision-making process regarding the issuance of seizure alerts. This adaptive approach ensures that the system remains P7152PC00

[0092] 16 sensitive to shifts in brain activity patterns while maintaining high specificity in its predictions. The overall functionality provided by this method allows for effective, realtime predictions that are critical for timely interventions, potentially enabling preemptive measures that can mitigate or even prevent the occurrence of a seizure.

[0093] The reduced set of readout neurons in the readout layer may be Leaky Integrate-and- Fire neurons. The use of LIF neurons in the readout layer may enhance the system’s ability to process and interpret the complex, high-dimensional spike data forwarded from the recurrent layer. By leveraging the intrinsic properties of LIF neurons, the system can more accurately track and analyze the temporal sequences of spikes, identifying patterns that are indicative of either normal brain activity or potential seizure activity.

[0094] Moreover, the system may be configured to compute or extract a likelihood of whether a seizure is ongoing or imminent based on the real-time states of these LIF neurons. This computational process involves analyzing the firing rates, synchronization, and other relevant neuronal behaviors captured by the LIF neurons. By continuously monitoring these states, the system can assess changes over time and determine the probability of a seizure occurring.

[0095] This functionality allows for an adaptive response mechanism where, if the analyzed neuronal states suggest a high likelihood of seizure, the system can trigger alerts or take other predefined actions. This predictive capability is useful for providing timely interventions and managing epilepsy more effectively. By implementing a targeted approach using LIF neurons in the readout layer, the system ensures that the seizure predictions are not only accurate but also timely, facilitating proactive management of the condition.

[0096] The system for predicting epileptic seizures may be further configured to compute or extract a likelihood that a seizure is ongoing or will occur based on states of the reduced set of readout neurons. By integrating these neuronal outputs, the system can assess whether the patterns align more closely with those typically observed during pre-ictal (pre-seizure), ictal (during seizure), or interictal (normal) states. This assessment is continuously updated, allowing the system to dynamically monitor the patient's neurological state P7152PC00

[0097] 17

[0098] The readout layer may use different types of algorithms or layers. In one embodiment, the readout layer is a fully connected Reward-modulated Spike-Timing-Dependent Plasticity layer. This configuration harnesses the power of neural plasticity to optimize seizure prediction based on the precise timing of spikes received from the preceding layers. R-STDP is a form of synaptic plasticity where synaptic strengths are adjusted not only based on the relative timing of pre- and post-synaptic spikes but also incorporating feedback in the form of rewards. This reward feedback enhances the ability of the system to strengthen connections that lead to successful seizure predictions and weaken those that do not, thereby refining the system's predictive accuracy over time.

[0099] Moreover, the weights in the readout layer may be dynamically adjusted based on the timing of the spikes from the spike data projected into a higher dimensional space by the recurrent layer. This ability to adjust weights based on spike timing may be useful for adapting the neural network’s response to varying patterns of neural activity that characterize different stages of seizure development. By continually refining these weights, the system becomes increasingly adept at recognizing the subtle nuances that distinguish between pre-ictal and normal brain activity, enhancing both sensitivity and specificity in seizure prediction.

[0100] Additionally, the system may incorporate a decision unit configured to estimate the likelihood of a seizure occurring. This unit integrates the outputs from the readout layer, which reflect the current state and dynamics of brain activity. The decision unit may analyze these outputs to determine the probability of a seizure event. This probability can be calculated by assessing various indicators such as the intensity, pattern, and synchronization of neuronal spikes, which have been identified as predictive of seizures.

[0101] In a further configuration of the system for predicting epileptic seizures, the decision unit is designed to enhance its predictive accuracy by estimating the likelihood of a seizure based on the states of a reduced set of readout neurons within a moving time window. This configuration allows the system to dynamically analyze changes in neuronal activity over specified intervals, continually updating its assessments to reflect the most current data. P7152PC00

[0102] 18

[0103] The moving time window provides a framework within which data is continuously collected and analyzed. This approach ensures that the decision unit has access to a constantly refreshed stream of information, allowing it to detect emerging patterns that may indicate the onset of a seizure. The time window shifts forward with each new data point, ensuring that the analysis remains relevant to the present moment and is not unduly influenced by older, potentially outdated data.

[0104] Within this window, the decision unit can monitor the states of the readout neurons, focusing particularly on those that have been identified as critical for recognizing pre- ictal activities. These neurons have been selected for their ability to capture specific neural dynamics associated with the lead-up to seizures. By analyzing the output of these neurons - such as their firing rates, pattern regularity, and synchronization - the decision unit can assess whether the observed activities align with typical pre-seizure patterns.

[0105] Fig. 4 shows an example of an implementation of decision making using the reduced set of readout neurons. In this example, the reduced set of readout neurons comprises two neurons to distinguish between different pre-ictal activities. The first neuron, Neuron 1 / Ni, is rewarded for pre-ictal activity leading to seizure. The second neuron, Neuron 2 / N2, is rewarded for pre-ictal activity not leading to seizure. In the example of the decision unit, the sums of spikes for the first neuron and the second neuron are calculated during a time window. If the total sum of the two neurons is greater than a first threshold and the sum of the first neuron is greater than a second threshold, the decision unit estimates that there is a seizure, otherwise the decision unit estimates that there is no seizure.

[0106] In a further configuration of the system for predicting epileptic seizures, the decision unit is designed with specific criteria to enhance its responsiveness and precision in detecting ongoing seizures. In one embodiment, the decision unit is configured to indicate that there is a seizure occurring if a total number of spikes of a first readout neuron and a second readout neuron is greater than a total threshold, and the number of spikes of the first readout neuron is greater than a seizure threshold.

[0107] In this embodiment the first readout neuron is specifically sensitive to neural patterns that typically occur during seizures, while the second readout neuron monitors general neural activity, including non-seizure states. The decision unit evaluates the output P7152PC00

[0108] 19 from these neurons under specific conditions: if the combined total number of spikes from both neurons exceeds a predetermined overall threshold, and if the number of spikes from the first readout neuron alone surpasses a separate, higher threshold indicative of seizure activity, then the system concludes that a seizure is likely occurring.

[0109] The effectiveness of the classification may arise from the combination of the spikecounting mechanism with the specialization of the readout neurons. The first readout neuron may be trained to respond preferentially to pre-ictal activity, while the second readout neuron may be trained to respond to normal interictal activity. Their combined activity therefore enables the decision unit to distinguish between different types of pre- ictal activity. For example, subtle pre-ictal states leading gradually to seizure onset can be distinguished from transient fluctuations that are not indicative of an impending seizure.

[0110] When these conditions are met, the decision unit signals that a seizure is occurring. This signal can then trigger a range of responses, from alerting medical personnel to activating on-device interventions designed to help manage the seizure.

[0111] The present disclosure further relates to a computer-implemented method for processing electroencephalogram signals representing electrical activity of a brain of a patient, the method comprising: receiving the electroencephalogram signals; converting the electroencephalogram signals into a series of spikes using a spike encoder; processing the series of spikes through a spiking neural network to identify indicators associated with pre-ictal activity, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; outputting the indicators for use in seizure prediction or a monitoring system.

[0112] The present disclosure further relates to a method for predicting epileptic seizures, the method comprising: receiving electroencephalogram signals representing electrical activity of a brain of a patient; converting the electroencephalogram signals into a series of spikes using a spike encoder; processing the series of spikes through a P7152PC00

[0113] 20 spiking neural network to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; outputting, directly or indirectly, the indicators predictive of an epileptic seizure.

[0114] A person skilled in the art will recognize that the presently disclosed method for processing electroencephalogram signals representing electrical activity of a brain of a patient may be performed using any embodiment of the presently disclosed system for predicting epileptic seizures, and vice versa.

[0115] Fig. 6 shows a flow chart of a method 200 according to an embodiment of the presently disclosed method for processing electroencephalogram signals representing electrical activity of a brain of a patient. The method may comprise the step of receiving electroencephalogram signals representing electrical activity of a brain of a patient (201). The electroencephalogram signals may then be converted into a series of spikes using a spike encoder (202). The method may further comprise the step of processing the series of spikes through a spiking neural network to identify indicators associated with pre-ictal activity (203). The method may further comprise the step of outputting the indicators for use in seizure prediction or a monitoring system (204).

[0116] The present disclosure further relates to a computer program having instructions which, when executed by a computing device or computing system, cause the computing device or computing system to carry out any embodiment of the presently disclosed method for predicting epileptic seizures. The computer program may be stored on any suitable type of storage media, such as non-transitory storage media

[0117] The system for predicting epileptic seizures may comprise processing circuitry to carry out any of the disclosed processing steps. As would be understood by a person skilled in the art, the processing circuitry may be any suitable processing circuitry, such as a single processor or a multi-core / multiprocessor system, provided as part of, for example, a portable system, or as a separate part. The system may further comprise peripheral components, such as one or more memories, which may be used for storing instructions that can be executed by any of the processors. The one or more memories P7152PC00

[0118] 21 may include random access memory (RAM) and / or a read-only memory (ROM), or any suitable type of memory. The system may further comprise any of: internal and external network interfaces, input and / or output ports, modules for sending and receiving data wirelessly, communication interface that allows software and / or data to be transferred between the system and external devices.

[0119] The present disclosure further relates to a computer-implemented method executed on a computing device for training a spiking neural network for processing physiological electroencephalogram signals, the method comprising: obtaining a spiking neural network operatively coupled to a spike encoder for converting physiological electroencephalogram signals into a series of spikes, wherein the spiking neural network is configured to process the series of spikes through a spiking neural network to identify indicators associated with pre-ictal activity, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons; training the reduced set of neurons in the readout layer to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space by providing labeled electroencephalogram signals representing known pre-ictal and non-pre-ictal states. The physiological electroencephalogram signals may be physical, time-varying voltage signals. The signals may transfer biological signals into a reduced and computationally efficient representation usable for processing in the presently disclosed systems and methods.

[0120] A person skilled in the art will recognize that the presently disclosed method for training a spiking neural network may be performed using any embodiment of the presently disclosed system for predicting epileptic seizures, and vice versa.

[0121] Fig. 7 shows a flow chart of a method 300 according to an embodiment of the presently disclosed method for training a spiking neural network for predicting epileptic seizures. The method may comprise the step of obtaining a spiking neural network operatively coupled to a spike encoder for converting electroencephalogram signals into a series of spikes, wherein the spiking neural network is configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a P7152PC00

[0122] 22 reduced set of readout neurons (301). The method may further comprise the step of training the reduced set of neurons in the readout layer to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space by providing labeled electroencephalogram signals representing known pre-ictal and non-pre-ictal states (302).

[0123] The system described in the present disclosure is not limited to prediction of epileptic seizures. In further embodiments the system may be used to predict, monitor or categorize neurological diseases. The present disclosure therefore further relates to a system for predicting a neurological event or state, the system comprising: an input interface configured to receive physiological signals; a spike encoder operatively coupled to the input interface, the spike encoded configured to convert the physiological signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of the neurological event or state, wherein the spiking neural network comprises: a recurrent-layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of the neurological event or state, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between the neurological event or state and other events or states based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of the neurological event or state.

[0124] The physiological signals may be, but are not limited to, neurophysiological signals, i.e. physiological signals related to the nervous system, and / or bioelectrical signals, including any electrical signal generated by biological processes, including, for example, electroencephalogram signals, electrocardiograms and electromyograms.

[0125] The neurological event or state can be any suitable neurological event or state. For example, the system can be used to identify stages of, and / or predict future stages of Parkinson’s Disease by analyzing neural patterns and brain activity changes. P7152PC00

[0126] 23

[0127] The system may also be used to monitor states and / or progress of, for example traumatic Brian Injury, mental health disorders, cognitive decline, dementia, ADHD etc.

[0128] The system for predicting a neurological event or state may include any of the above described features or embodiments present in the corresponding system for predicting epileptic seizures.

[0129] The system can be further expanded to include more general use, such as use within any biometric monitoring, industrial applications, such as verification, load or stress detection, in automotive applications, telecommunications, for example, in traffic monitoring, allocation of bandwidth, network monitoring and optimization. Other examples of use include supply chain and logistics, climate monitoring and forecasting, industrial automation and surveillance.

[0130] Accordingly, the present disclosure therefore further relates to embodiments of a system for predicting an event or state, the system comprising: an input interface configured to receive physical signals; a spike encoder operatively coupled to the input interface, the spike encoded configured to convert the physical signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of the event or state, wherein the spiking neural network comprises: a recurrent-layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of the event or state, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between the event or state and other events or states based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of the event or state.

[0131] Accordingly, the physical signals may be electrical representations of a physical activity, including, but not limited to, data or communications system in telecommunication system and / or signal representing parameters such as measurements in, for example, an industrial process or in climate or weather monitoring or forecasting system. P7152PC00

[0132] 24

[0133] The system for predicting an event or state may include any of the above described features or embodiments present in the corresponding system for predicting epileptic seizures.

[0134] Further details

[0135] 1 . A system for predicting epileptic seizures, the system comprising: an input interface configured to receive electroencephalogram signals, wherein the electroencephalogram signals represent electrical activity of a brain of a patient; a spike encoder operatively coupled to the input interface, the spike encoded configured to convert the electroencephalogram signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent-layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of an epileptic seizure.

[0136] 2. The system for predicting epileptic seizures according to item 1 , wherein the recurrent layer comprises a reservoir network comprises a set of neurons, such as a set of randomly connected neurons, wherein the readout layer is configured to use a subset of the set of neurons to identify the indicators predictive of an epileptic seizure.

[0137] 3. The system for predicting epileptic seizures according to item 2, wherein the subset of neurons is connected by a fully connected or sparsely connected layer, directly or indirectly, to a set of readout neurons. P7152PC00

[0138] 25

[0139] 4. The system for predicting epileptic seizures according to any one of the preceding items, wherein the readout layer comprises at least a first readout neuron which has been rewarded when it spikes during interictal activity leading to seizure during training.

[0140] 5. The system for predicting epileptic seizures according to any one of the preceding items, wherein the readout layer comprises at least a second readout neuron which has been rewarded when it spikes during no-seizure interictal activity.

[0141] 6. The system for predicting epileptic seizures according to items 4 and 5, further configured to estimate a likelihood of a seizure occurring by analyzing spikes of the first readout neuron and the second readout neuron within a moving time window.

[0142] 7. The system for predicting epileptic seizures according to any one of the preceding items, wherein the reduced set of readout neurons in the readout layer are Leaky Integrate-and-Fire neurons.

[0143] 8. The system for predicting epileptic seizures according to any one of the preceding items, wherein the system for predicting epileptic seizures is further configured to compute or extract a likelihood that a seizure is ongoing or will occur based on states of the reduced set of readout neurons.

[0144] 9. The system for predicting epileptic seizures according to any one of the preceding items, wherein the input interface comprises analog-to-digital converters configured to sample streams of analog electroencephalogram signals into series of digital electroencephalogram signals.

[0145] 10. The system for predicting epileptic seizures according to item 9, wherein the spike encoder is configured to generate two series of spikes to the spiking neural network, wherein the spike encoder comprises a high channel and a low channel.

[0146] 11. The system for predicting epileptic seizures according to item 10, wherein the spike encoder is configured to generate a series of spikes for signals exceeding P7152PC00

[0147] 26 a comparison threshold through the high channel and through the low channel otherwise.

[0148] 12. The system for predicting epileptic seizures according to any one of the preceding items, wherein the recurrent layer comprises a set of randomly connected neurons having randomly initialized weights.

[0149] 13. The system for predicting epileptic seizures according to any one of the preceding items, wherein the readout layer is configured to consolidate the spike data of the higher dimensional space from the recurrent layer to identify the indicators predictive of an epileptic seizure.

[0150] 14. The system for predicting epileptic seizures according to any one of the preceding items, wherein the readout layer is a fully connected Reward- modulated Spike-Timing-Dependent Plasticity layer.

[0151] 15. The system for predicting epileptic seizures according to any one of the preceding items, wherein the readout layer has weights that have been adjusted based on timing of spikes of the spike data of a higher dimensional space.

[0152] 16. The system for predicting epileptic seizures according to any one of the preceding items, further comprising a decision unit configured to estimate a likelihood of a seizure occurring based on the indicators predictive of an epileptic seizure.

[0153] 17. The system for predicting epileptic seizures according to item 17, wherein the decision unit is configured to estimate likelihood of a seizure occurring based on states of the reduced set of readout neurons within a moving time window.

[0154] 18. The system for predicting epileptic seizures according to item 15, wherein, if a total number of spikes of a first readout neuron and a second readout neuron is greater than a total threshold, and the number of spikes of the first readout neuron is greater than a seizure threshold, the decision unit is configured to indicate that there is a seizure occurring. P7152PC00

[0155] 27

[0156] 19. A method for predicting epileptic seizures, the method comprising: receiving electroencephalogram signals representing electrical activity of a brain of a patient; converting the electroencephalogram signals into a series of spikes using a spike encoder; processing the series of spikes through a spiking neural network to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; outputting, directly or indirectly, the indicators predictive of an epileptic seizure.

[0157] 20. A computer program having instructions which, when executed by a computing device or computing system, cause the computing device or computing system to carry out the method for predicting epileptic seizures according to item 19.

[0158] 21 . A method for training a spiking neural network for predicting epileptic seizures, the method comprising: obtaining a spiking neural network operatively coupled to a spike encoder for converting electroencephalogram signals into a series of spikes, wherein the spiking neural network is configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons; training the reduced set of neurons in the readout layer to distinguish between different pre-ictal activities based on the spike data of the higher P7152PC00

[0159] 28 dimensional space by providing labeled electroencephalogram signals representing known pre-ictal and non-pre-ictal states.

Claims

P7152PC0029Claims1 . A system for predicting epileptic seizures, the system comprising: an input interface configured to receive electroencephalogram signals, wherein the electroencephalogram signals represent electrical activity of a brain of a subject; a spike encoder operatively coupled to the input interface, the spike encoded configured to convert the electroencephalogram signals into a series of spikes; a spiking neural network operatively coupled to the spike encoder, the spiking neural network configured to process the series of spikes to identify indicators predictive of an epileptic seizure, wherein the spiking neural network comprises: a recurrent-layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; an output interface configured to, directly or indirectly, communicate the indicators predictive of an epileptic seizure.

2. The system for predicting epileptic seizures according to claim 1 , wherein the recurrent layer comprises a reservoir network comprises a set of neurons, such as a set of randomly connected neurons, wherein the readout layer is configured to use a subset of the set of neurons to identify the indicators predictive of an epileptic seizure, and wherein the subset of neurons is connected by a fully connected or sparsely connected layer, directly or indirectly, to a set of readout neurons.

3. The system for predicting epileptic seizures according to any one of the preceding claims, wherein the readout layer comprises at least a first readout neuron which has been rewarded when it spikes during interictal activity leading to seizure during training.P7152PC00304. The system for predicting epileptic seizures according to claim 3, wherein the readout layer further comprises at least a second readout neuron which has been rewarded when it spikes during no-seizure interictal activity.

5. The system for predicting epileptic seizures according to claims 3 and 4, further configured to estimate a likelihood of a seizure occurring by analyzing spikes of the first readout neuron and the second readout neuron within a moving time window.

6. The system for predicting epileptic seizures according to any one of the preceding claims, wherein the system for predicting epileptic seizures is further configured to compute or extract a likelihood that a seizure is ongoing or will occur based on states of the reduced set of readout neurons.

7. The system for predicting epileptic seizures according to any one of the preceding claims, wherein the input interface comprises analog-to-digital converters configured to sample streams of analog electroencephalogram signals into series of digital electroencephalogram signals, wherein the spike encoder is configured to generate two series of spikes to the spiking neural network, wherein the spike encoder comprises a high channel and a low channel, and wherein the spike encoder is configured to generate a series of spikes for signals exceeding a comparison threshold through the high channel and through the low channel otherwise.

8. The system for predicting epileptic seizures according to any one of the preceding claims, wherein the recurrent layer comprises a set of randomly connected neurons having randomly initialized weights.

9. The system for predicting epileptic seizures according to any one of the preceding claims, wherein the readout layer is configured to consolidate the spike data of the higher dimensional space from the recurrent layer to identify the indicators predictive of an epileptic seizure.

10. The system for predicting epileptic seizures according to any one of the preceding claims, wherein the readout layer has weights that have beenP7152PC0031 adjusted based on timing of spikes of the spike data of a higher dimensional space.

11. The system for predicting epileptic seizures according to any one of the preceding claims, further comprising a decision unit configured to estimate a likelihood of a seizure occurring based on the indicators predictive of an epileptic seizure, wherein the decision unit is configured to estimate likelihood of a seizure occurring based on states of the reduced set of readout neurons within a moving time window.

12. The system for predicting epileptic seizures according to claim 11 , wherein, if a total number of spikes of a first readout neuron and a second readout neuron is greater than a total threshold, and the number of spikes of the first readout neuron is greater than a seizure threshold, the decision unit is configured to indicate that there is a seizure occurring.

13. A computer-implemented method for processing electroencephalogram signals representing electrical activity of a brain of a subject, the method comprising: receiving the electroencephalogram signals; converting the electroencephalogram signals into a series of spikes using a spike encoder; processing the series of spikes through a spiking neural network to identify indicators associated with pre-ictal activity, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons trained to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space; outputtingthe indicators for use in seizure prediction or a monitoring system.

14. A computer program having instructions which, when executed by a computing device or computing system, cause the computing device or computing systemP7152PC0032 to carry out the method for processing electroencephalogram signals representing electrical activity of a brain of a subject according to claim 13.

15. A computer-implemented method executed on a computing device for training a spiking neural for processing physiological electroencephalogram signals, the method comprising: obtaining a spiking neural network operatively coupled to a spike encoder for converting physiological electroencephalogram signals into a series of spikes, wherein the spiking neural network is configured to process the series of spikes through a spiking neural network to identify indicators associated with pre-ictal activity, wherein the spiking neural network comprises: a recurrent layer configured to project the series of spikes into spike data of a higher dimensional space; a readout layer configured to identify the indicators predictive of an epileptic seizure, wherein the readout layer comprises a reduced set of readout neurons; training the reduced set of neurons in the readout layer to distinguish between different pre-ictal activities based on the spike data of the higher dimensional space by providing physiological labeled electroencephalogram signals representing known pre-ictal and non-pre-ictal states.

Images