Prediction of Epilepsy Phenotype in Intra-amygdala Kainic Acid Mouse Model of Epilepsy

Animal models of drug-resistant epilepsy represent an important resource for discovering new drug targets and testing experimental medicines. A major limitation, however, is the loss of time and resources from generating mice with low or high rates of spontaneous seizures. Intra-amygdala microinjection of kainic acid in mice is one of the most widely regarded models of drug-resistant epilepsy. Mice develop acute status epilepticus, which abates after a few hours and then, within a few days, mice display spontaneous seizures (epilepsy). The frequency of spontaneous seizures varies between mice, with some developing low or high seizure rates. The ability to predict soon after status epilepticus, which mice will go on to develop a moderate frequency of seizures, would enable a significant reduction in resources and EEG reviewing time, as well as lead to humane early end-points. In this study, we developed a transfer learning-based method for predicting the emergent spontaneous seizure rates in the intra-amygdala kainic acid model based on the acute EEGs recorded in mice during status epilepticus. The method was trained on data from 28 mice and subsequently tested on data from 16 mice, achieving an accuracy of 75% on the test set in classifying emergent epilepsy as moderate or an outlier (low-frequency or high-frequency seizure rate). This approach holds great promise for researchers, aiding in the analysis of seizure rates within EEGs of the intra-amygdala kainic acid mouse model and preclinical drug development and compliance with the Reduction, Refinement and Replacement (3Rs).


INTRODUCTION
Epilepsy is a neurological disorder characterised by recurrent seizures [13].Seizures are caused by excessive and synchronous discharges of groups of neurons in the brain.According to the World Health Organisation (WHO), about 50 million people globally suffer from epilepsy, making it one of the most prevalent neurological disorders.Epilepsy Ireland reports a prevalence of about 40,000 people in Ireland, accounting for 0.79% of Ireland's population.
Electroencephalogram (EEG) is a diagnostic tool used in clinical settings to aid in the identification of epilepsy [15].The EEG records the weak electrical signals detectable over the surface of the brain.Seizures can be identified in EEG recordings based on changes to the frequency, amplitude and morphology of these waves.There is much interest in developing automated systems for the analysis of EEG to overcome the time-consuming need to identify and quantify seizures and background abnormalities.However, very few automatic approaches to detect seizures in mouse models of epilepsy have been developed.Casillas-Espinosa et al. [3] developed software based on advanced time-frequency analysis that detects seizures in rodent models of acquired and genetic epilepsy, while Wei et al. [16] proposed XGBoost-based detection of spontaneous seizures in multiple mouse models of epilepsy.Lashkari et al. [5] predicted the occurrence of absence epilepsy in seven genetic absence epilepsy Wag/Rij rats using phase-space geometry.Li et al. [6] predicted seizure episodes using the interictal recordings of five dogs as one of the independent datasets.
Animal models of epilepsy are critical tools in understanding the mechanisms of epilepsy and testing novel experimental therapeutics.Historically, drug screening has used acute evoked seizures, but there is increasing recognition that experimental drugs should also be screened in models in which animals exhibit spontaneous (epileptic) seizures [8,14].To this end, the model using intra-amygdala microinjection of kainic acid (IAKA) is increasingly being adopted for drug target identification and preclinical drug development.In the model, an acute status epilepticus (SE) is evoked, damaging select networks within limbic structures, including the hippocampus.Within a few days, spontaneous seizures emerge, which continue thereafter, typically at rates of 3 -10 seizures per day [9].The seizures are highly refractory to anti-seizure medications, and the molecular profile of the model correlates well with findings in human temporal lobe epilepsy [4,19,20].One of the key challenges, however, is inter-animal variability in the epilepsy phenotype.Some of the mice in this model develop a low-frequency seizure rate which is less useful for drug screening, or a high-frequency seizure rate, which leads to brain damage and eventual death of the mouse.Identifying those animals at the earliest possible stage would eliminate expensive and time-consuming EEG analysis, save resources and lead to earlier, human end-points.Hence, this study sought to predict, ahead of time, the epilepsy phenotype based on EEG recordings during the initial status epilepticus.

MATERIALS AND METHODS
To classify the IAKA mouse phenotype into moderate or outlier (i.e.low or high seizure frequency rate), this study follows two technical steps.First, transfer learning is adopted to classify the spectrogram of the SE periods.Second, the decision tree algorithm was used to post-process the output of the transfer learning to derive the final classification of mice with moderate or outlier frequency rates.Figure 2 shows the methodological framework.

EEG Data
This EEG dataset was collected from experiments performed at the Royal College of Surgeons in Ireland (RCSI).Ethical approval for the animal experiments was approved by the research ethics committee at RCSI with a license from the Health Products Regulatory Authority (AE19127/001), Dublin, Ireland.
Adult male C57BL/6 mice (weight: 28-30g; age 10 weeks) obtained from Harlan (UK) were employed, following the experimental protocol outlined in [9].Briefly, mice were anaesthetized using isoflurane, with an initial induction of 5% and maintenance at approximately 2%.Under surgical anaesthesia, mice were fitted with an implantable EEG telemetry device (DSI; Data Systems International) to facilitate continuous EEG monitoring.The transmitters (HD-X02 model) can record bilateral EEG signals from screws embedded in the skull.These transmitters were implanted in a subcutaneous pocket simultaneously during the placement of a cannula for microinjection of kainic acid.The cannula was positioned on the dura mater over the right hemisphere, with the following coordinates relative to bregma: IA: AP = -0.95mm, L = +2.85mm.To maintain the body temperature of the mice, a feedback-controlled heat blanket from Harvard Apparatus Ltd (UK) was employed [16].
After recovery, SE was induced in mice by intra-amygdala kainic acid injection.An injection cannula was inserted into the guide cannula into the basolateral amygdala, and kainic acid (0.3µg) was injected.The cannula was withdrawn, and then seizure activity was recorded.After 40 minutes, mice received intraperitoneal injections of lorazepam (8 mg/kg) to reduce morbidity and mortality.Figure 1 shows the overview of the IAKA mouse model.
Continuous EEG recordings were made for the next two weeks in a group of 29 mice.Spontaneous seizures were counted based on well-established criteria i.e. high amplitude, high-frequency polyspike discharge lasting ≥ 10 seconds.A number of spontaneous seizures were classified as 'moderate', 'low' or 'high' based on the spontaneous seizures recorded during the two-week period.This was performed systematically on the basis of using the mean spontaneous recurrent seizures (SRS) number/day ± 2 SEM from days 7-13 as upper and lower cutoffs for the SRS high (H) and SRS low (L) groups.For an individual mouse to be classified as SRS high (H) or low (L) it must exhibit an SRS number greater than Mean + 2 SEM or less than Mean -2 SEM on a minimum of four days from days 7-13 post-KA, respectively.If these cutoffs were not met, mice were classified as SRS moderate (M) by exclusion.Mice with SUDEP (sudden unexpected death in epilepsy) were also classified as SRS high (H).
The training was executed by employing 80% of the EEGs from the 28 mice (moderate frequency mice: 14; outlier frequency mice: 14).The remaining 20% were used for validation purposes.For testing, EEGs from 16 mice (6 moderate frequency and 10 outlier frequency) were used.Comprehensive information regarding the number and duration of EEG recordings utilized for training and testing is available in Table 1.

Data Pre-processing
To remove powerline interference, a notch filter (50Hz) was applied to each channel of the EEG recordings.The desired frequency band signals were extracted using a Butterworth filter (infinite impulse response) within the range of 0.1 to 64 Hz.One of the mice exhibited very high artefacts and was excluded from the study.Subsequently, spectrograms were generated utilizing the SE period of EEG recordings, with each resulting spectrogram being classified as either 'moderate' or 'outlier'.

Method Development
The approach taken in this study for predicting seizure rates involved utilizing a pre-trained transfer learning model known as ResNet50 [18].In this method, ResNet50 was utilized as a feature extractor, with the initial layers being frozen.The input to the model was an image with dimensions (256×256×3) Byte.The final output layer of ResNet50 was removed, and instead, two fully connected layers were added -one with 128 units and another with 64 units.Each of these fully connected layers was followed by a rectified linear unit (ReLU) activation function.To enhance the model's generalization ability and prevent overfitting, a regularization technique (L2 regularization with a strength of 0.05) was applied.For the seizure rate prediction, the model employed a softmax activation function in its output layer.The training process utilized the Adam optimization method, with a batch size of 64.The training was set to run for 100 epochs.Additionally, to prevent unnecessary training and to optimize model performance, the early stopping technique was implemented with a patience value of 20.

Post-processing
Studies indicate that the signal of every channel does not necessarily undergo substantial changes during a seizure [17].To address this, a decision tree was employed to combine the predictions of ResNet50 for each spectrogram originating from two channels (EEG1 and EEG2) within the training dataset.Subsequently, this decision tree model was used to process the outputs of ResNet50 for the two channels in the test dataset.By applying this decision tree model across multiple EEG channels, the ultimate classification of each spectrogram as either 'moderate' or 'outlier' was achieved.The method architecture is illustrated in Figure 2.

Performance Evaluation
For each model, we computed the accuracy, precision, recall and F1 score as the metric evaluation.
Where: TP: the number of mice having both predicted and actual labels as moderate FP: the number of mice predicted to be moderate when the actual label is outlier TN: the number of mice having both predicted and actual labels as outlier FN: the number of mice predicted to be outlier when the actual label is moderate

RESULTS
The present study used status epilepticus to trigger epilepsy in mice.The number of spontaneous seizures recorded over 2 weeks was used to classify mice as having developed a moderate course of epilepsy or an outlier (low/high) frequency seizure rate phenotype.The acute EEG recordings during status epilepticus were used to train a transfer learning model to classify the epilepsy phenotype, with a view to being able to predict moderate or outlier frequency rates of spontaneous seizures.Table 2 provides a comprehensive overview of the epilepsy phenotype model's performance on the training, validation and independent test sets, and the outcomes of the post-processing technique.The table illustrates that the model effectively discerns between moderate and outlier mice within the EEG data set.For the validation set, the model's capabilities are highlighted, achieving an accuracy of 82%, a precision of 87%, and an F1 score of 0.82.
For the independent test set, the model's performance achieved an accuracy of 68% and an F1 score of 0.65.These results emphasize the model's generalizability and ability to maintain a respectable level of accuracy on unseen data.The post-processing method was incorporated by considering multiple EEG recording channels, the accuracy of classification gives a notable 7% enhancement.This advancement is mirrored in the F1 score, which experiences a concurrent increase of 0.1.

DISCUSSION
The focus of this study was to forecast the rate of spontaneous seizure exhibited by mice in the animal model of drug-resistant epilepsy following the stimulation of status epilepticus with intraamygdala kainic acid.We developed a transfer learning-based method for predicting the emergent spontaneous seizure rates in the intraamygdala kainic acid mouse model based on the acute EEGs recorded in mice during status epilepticus.Although there are myriads of research in predicting the next seizure episode in humans and animal models of epilepsy, to the best of our knowledge, and at the time of this study, no study has attempted the prediction of spontaneous seizure rate from animal models of epilepsy.Consequently, we exploit the integration of transfer learning techniques and a decision tree algorithm method to discriminate moderate and outlier seizure rates in mice.
Some work [10-12, 18, 21, 22] utilised transfer learning-based approaches in predicting and detecting seizures from EEG. Transferring the knowledge learned from a previous task to a new domain overcomes the issue of overfitting a neural network/deep learning model trained with limited data and reduces training time and computation resources.
EEG channel selection has a potential value in seizure detection and prediction.One of the major purposes of EEG channel selection is the reduction of computational resources [1], which may possibly lead to the loss of crucial information.In the study of drug-resistant epilepsy in a mouse model, the mouse brain activity was measured with two channels of EEG.Two EEG channels were considered in this study.Our method overcomes the complexities associated with multi-channel EEG analysis.Being that different channels of EEG may have varying seizure information, the post-processing module improves on the misclassification of the transfer learning module and profers a substantial performance of 75% accuracy, precision, recall and F1 score.
With the growing adoption of the mouse intra-amygdala kainic acid model for drug screening, including by the National Institute of Neurological Disorders and Stroke epilepsy therapy screening program [20], the present findings have immediate practical applications in how the model is used in epilepsy research.Use of this classifier will enable researchers to quickly and early select out mice that will have too low or too high a baseline spontaneous seizure rate, leading to more homogenous groups of mice and compliance with 3Rs directives and humane early end-points [7], when using the model for drug screening.
A limitation of the current study is understanding the influence of dependent and independent variables.ResNet50 is a convolutional neural network, and neural networks are known culprits of inexplainability [2].Overcoming this bottleneck, future work will examine the application of traditional machine learning and compare the performance of the two methods.Furthermore, future work will expand to incorporate features such as mouse behaviour during status epilepticus or a more granular analysis of features such as high-frequency oscillations (HFOs).Moreover, we will assess the potential applicability of the presented approach to other prevalent models, such as pilocarpine, which are widely used, as well as genetic models (e.g., the Scn1a model of Dravet syndrome).

CONCLUSION
The inter-animal variability in the IAKA epilepsy phenotype has been a major hurdle for researchers.Overcoming this challenge, an integrated transfer learning and decision tree algorithm was devised to forecast the incidence of spontaneous seizures in the IAKA mouse model into moderate and outlier rates of spontaneous seizures.The performance of this method will reduce the loss of time and resources from generating mice with low or high rates of spontaneous seizure and enhance the humanely early endpoint of mice with low or high seizure rates.This method will significantly aid researchers in scrutinising the mice in their animal models of drug-resistant epilepsy, thus speeding up the discovery and testing of new therapeutic targets and candidate drugs for epilepsy.

Figure 1 :
Figure 1: Overview of the Intra-amygdala Kainic Acid mouse model.

Figure 2 :
Figure 2: Methodological framework for seizure rate prediction