Patient specific seizure prediction for time intervals using TQWT and deep learning

Patient specific seizure prediction for time intervals using TQWT and deep learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Patient specific seizure prediction for time intervals using TQWT and deep learning YongUn Jo, Do Chang Oh This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5714799/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Seizures are sudden activations of part or all of the brain, which are the primary symptoms of epilepsy. Epileptic seizures are characterized by their sudden and unpredictable nature and pose significant risks to patients’ daily lives. In many patients with epilepsy, specific symptoms are observed for a short period before a seizure occurs. Research is ongoing to develop technology that can predict seizures by detecting these symptomatic signals. In this paper, we used the CHB-MIT database, which contains more than 150 seizures in approximately 1,000 hours of EEG data collected from 23 children with intractable seizures. Tunable Q-factor Wavelet Transform (TQWT), a specific type of wavelet transform, was applied to the data to predict seizures by identifying inter-ictal and pre-ictal states using a relatively simple deep learning classifier. Each classification was performed individually for each patient, and seizure prediction during a specific period was achieved using k-fold cross-validation, a technique commonly employed in deep learning. By combining the results, the period with the highest probability of seizure occurrence for each patient was designated the specific warning period. Ultimately, it was possible to predict seizures 15 minutes in advance, achieving an average sensitivity of 0.97, an F1 score of 0.90, and a false discovery rate (FDR) of 0.13 for all patients. Additionally, a specific warning period was established for each patient, ranging from a minimum of 2.5 minutes to a maximum of 15 minutes. . inter-ictal pre-ictal seizure prediction patient specific specific warning period Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Introduction According to data released by the World Health Organization (WHO) [ 1 ], approximately 50 million people worldwide suffer from epilepsy, which is known to be the most common neurological disease. In particular, patients with epilepsy face a risk of premature death that is up to three times higher than that of the general population, with a significant number of fatalities attributed to seizures related to the condition. Epilepsy can be caused by various factors, including congenital brain abnormalities, genetic disorders, acquired head injuries, and brain tumors. A diagnosis of epilepsy is made when two or more spontaneous seizures occur. Epileptic seizures are characterized by irregular and nonperiodic activation of part or all of the brain, resulting in various symptoms that vary from patient to patient. These symptoms include temporary loss of consciousness, cognitive dysfunction, and paralysis of specific body parts or the entire body. Patients with epilepsy experience a higher incidence of seizure-related injuries than the general population. These injuries can range from minor injuries, such as fractures and bruises, to more severe physical complications, including traffic accidents. Furthermore, individuals with epilepsy may encounter psychological challenges, such as depression and anxiety. For this reason, extensive research has been conducted to detect seizures through various biological signals to prevent injuries caused by seizures. Additionally, further studies have focused on predicting and preparing for seizures [ 2 ]. Public databases commonly used in seizure research include the Hospital iEEG, CHB-MIT sEEG, the American Epilepsy Society Seizure Prediction Challenge iEEG, and the European Epilepsy Database. The CHB-MIT database comprises approximately 1,000 hours of recordings and includes 198 seizures from long-term monitoring of 23 refractory patients at Boston Children's Hospital. Seizures can be broadly categorized into three stages: inter-ictal (the period between seizures), pre-ictal (the period before a seizure), and ictal (the period during a seizure). Among these, the pre-ictal stage is when precursor symptoms of a seizure manifest, allowing for predicting seizures through their recognition. However, a standard or optimal point for dividing the three stages of seizures has not been clearly defined. In the American Epilepsy Society’s Seizure Prediction Challenge, the pre-ictal stage was established as one hour before the onset of the seizure. In some studies, researchers autonomously choose a very short period before a seizure, such as 5 minutes or less. In other cases, they may choose a relatively longer period of 20 to 30 minutes or even an extended period of 90 minutes. The primary reason for selecting the stage before the seizure is to predict the patient’s seizure in advance and to ensure an adequate response time. Most existing studies on seizure prediction using electroencephalogram(EEG) have been conducted under patient-specific conditions. According to Jana and Mukherjee, The EEG signal patterns are unique for each epilepsy patient [ 2 ]. One of the challenges is the variations across seizures even for the same patient [ 3 ]. Consequently, research on seizure prediction using EEG is challenging to generalize across a large patient population, so it is primarily considered patient specific. In particular, patient-specific characteristics include not only the EEG pattern but also the duration and timing of the onset of pre-ictal symptoms. Bandarabadi et al. selected various pre-ictal periods for several patients, ultimately predicting seizures within a range of 5 to 173 minutes, and established 44.3 minutes as the optimal interval for seizure prediction [ 4 ]. Moghim and Corne set the pre-ictal period as ranging from 0 to 20 minutes and developed an algorithm that selects the optimal pre-ictal period for each patient [ 5 ]. The potential between electrodes changes as a result of effects such as eye movement or muscular activity, causing an artifact [ 5 ]. Therefore, in research using EEG, preprocessing to eliminate various types of noise plays a significant role, and the results can vary considerably depending on the selected preprocessing method [ 6 ]. For instance, Shen et al. performed frequency analysis using the Digital Wavelet Transform (DWT) in conjunction with an algorithm that extracts eight types of eigenvalues [ 7 ]. In contrast, Zhang and Parhi preprocessed the signal by extracting 44 feature values, including spectral power analysis [ 8 ]. Tsiouris et al. extracted and integrated multiple features using various techniques, including cross-correlation estimation, analysis in both the time and frequency domains, and graph theory [ 9 ]. The preprocessing methods utilized in the previous study can be classified into two main types: feature value extraction in the time domain, such as zero crossing and cross-correlation estimation, and analysis in the frequency domain, which includes Fourier transform and wavelet transform. The feature value extraction method in the time domain is a traditional technique that consistently demonstrates strong performance in biosignal analysis. In contrast, frequency analysis is particularly well-suited to electroencephalography (EEG), as it enables the examination of standardized frequency bands, including delta, theta, alpha, beta, and gamma waves. In this paper, TQWT was applied to data from 21 individuals in the CHB-MIT database, excluding certain data that have significantly different shapes, and seizures are predicted by identifying pre-ictal EEG patterns using a straightforward Convolutional Neural Network (CNN) model. Additionally, by employing the k-fold cross-validation technique in deep learning, we were able to evaluate predictive performance at specific time points and identify the most appropriate seizure response time interval for each patient. Seizure prediction for time intervals with TQWT and deep learning • CHB-MIT database In most cases, the CHB-MIT database includes the 18 EEG channels from the International 10–20 system. Additionally, in some cases, ECG or Vagus Nerve Stimulation (VNS) signals or dummy channels may be used to meet the required number of channels. This channel configuration can result in numerous changes, even within a single case, making it difficult to select batches of data. Therefore, even among studies that use the same CHB-MIT database, some cases may be excluded. Truong et al. argued that pre-ictal ranges may be unreliable if channel configurations differ significantly, or seizures are repeated multiple times within 30 minutes [ 10 ]. Excluding these data, only 13 cases were ultimately used. Jana and Mukherjee used 23 reliable cases that included patient information, but only 22 EEG channels commonly used in many cases were selected, and the rest were excluded. Tsiouris et al. used 24 cases, excluding some data that excluded the 18 EEG channels commonly used in most cases. As a special case, Liu et al. used all 24 cases, but some signals such as electrocardiogram (ECG) were removed and only EEG signals were used [ 11 ]. A total of 21 cases were used in this paper, excluding cases chb12 and 15 with changes in electrode configuration (for example, change from FP1-F7 to FP1-ref) and chb24 with a very short data size. Most of the data contained in the CHB-MIT database are commonly used EEG channels, but some data have additional EEG channels or contain other signals, such as ECG. In particular, the configuration of channels changes depending on the measurement time, even for the same patient, which makes it difficult to apply deep learning classifiers that require constant data. Therefore, only 18 EEG channel data commonly used in most cases were selected, and cases with fewer than 18 channels were excluded. Pre-ictal data for classification learning and testing were extracted up to 15 minutes before seizure onset, and inter-ictal data were extracted from a period without seizures for 1 hour, and the amount of the two data types was balanced as much as possible. • TQWT for pre-processing To predict seizures more accurately, accurate feature extraction and noise removal via preprocessing are important. The most commonly used pre-processing transforms are the continuous wavelet transform (CWT) [ 12 ], short-time Fourier transform (STFT) [ 10 ], and discrete wavelet transform (DWT) [ 13 ]. Some researchers have applied deep learning to raw EEG data for seizure prediction [ 14 – 16 ]. However, methods that are too complex or extract too many feature values are not suitable in terms of efficiency, and therefore DWT, which requires less calculation, is used more than CWT. Tunable Q Factor Wavelet Transform (TQWT) is an improved version of the basic Discrete Wavelet Transform (DWT), as illustrated in Fig. 1 , and was selected as the preprocessing technique. Similar to the existing DWT, TQWT performs multiple conversion steps by combining high-frequency and low-frequency filters, but is a technology designed to intensively analyze the desired frequency band by introducing a Q-factor that can be selected by the user during the conversion process. Due to these characteristics, TQWT was judged to be suitable for the analysis of EEG, where specific frequency bands have special meaning. $$\:{y}_{high}\left[n\right]=\left({x}^{*}{h}_{1}\right)\left[n\right]=\:\sum\:_{k=-\infty\:}^{\infty\:}x\left[k\right]{h}_{1}\left[n-k\right]$$ 1 $$\:{y}_{low}\left[n\right]=\left({x}^{*}{h}_{0}\right)\left[n\right]=\:\sum\:_{k=-\infty\:}^{\infty\:}x\left[k\right]{h}_{0}[n-k]$$ 2 $$\:{H}_{0}^{\left(j\right)}\left(\omega\:\right)=\left\{\begin{array}{cc}\:\:\prod\:_{m=0}^{j-1}{H}_{0}\left(\frac{\omega\:}{{\alpha\:}^{m}}\right),&\:if\left|\omega\:\right|\le\:{\alpha\:}^{j}\pi\:\\\:0,&\:\:\:\:\:\:if{\alpha\:}^{j}\pi\:\:\le\:\left|\omega\:\right|\le\:\pi\:\end{array}\right.\:$$ 3 $$\:{H}_{1}^{\left(j\right)}\left(\omega\:\right)=\left\{\:\:\begin{array}{cc}{H}_{1}\left(\frac{\omega\:}{{\alpha\:}^{j-1}}\right)\prod\:_{m=0}^{j-2}{H}_{0}\left(\frac{\omega\:}{{\alpha\:}^{m}}\right),\:\:&\:\:\:if\:\left(1-\beta\:\right){\alpha\:}^{j-1}\pi\:\:\le\:\left|\omega\:\right|\le\:\:{\alpha\:}^{j-1}\pi\:\\\:0,&\:\begin{array}{c}\:\:\:\:\:\:\:\:\:\:\\\:for\:other\:\omega\:\in\:\left[-\pi\:,\:\pi\:\right]\end{array}\end{array}\right.\:\:$$ 4 $$\:\beta\:=\frac{2}{Q+1},\:\:\:\:\alpha\:=1-\frac{\beta\:}{r}$$ 5 DWT uses the results of passing the input signal \(\:x\left(n\right)\) through a high-frequency filter \(\:{h}_{1}\left(n\right)\) and a low-frequency filter \(\:{h}_{0}\left(\text{n}\right)\) , respectively, as shown in equations ( 1 ) and ( 2 ). This is advanced to TQWT, and scaling of low and high frequencies is applied at the j-th stage, respectively, as shown in equations ( 3 ), ( 4 ), and (5). When the sampling frequency of x(n) is \(\:{f}_{s}\) , α, and β are scaled to \(\:{\alpha\:}{\text{f}}_{\text{s}}\) ( \(\:0<{\alpha\:}<1\) ) and \(\:{\beta\:}{\text{f}}_{\text{s}}\) ( \(\:0<{\beta\:}<1\) ), respectively, and the condition for lowering the sampling frequency of the filter based on the Q factor is presented in Eq. ( 5 ), where r means oversampling. Selesnick recommended that the r value be between 2 and 4 so that the sampling frequency of the low- and high-frequency filters can be changed by adjusting the Q-factor. • CNN model The CNN model used in this study, as shown in Fig. 2 , a relatively straightforward architecture, comprising two convolutional layers that integrate convolution and max pooling, followed by two fully connected layers. Due to the nature of the CHB-MIT database, a substantial amount of data is available for classification between pre-ictal and inter-ictal states; therefore, the dropout technique was applied to the fully connected layer. Because the goal was to predict seizures through real-time monitoring, which requires fast processing speed and efficient power consumption in wearable hardware capable of independent operation, the model was designed to be as light as possible. The input signal was processed by applying a 30-second window to the original signal, preprocessing was then performed using TQWP, and the prediction results were updated through deep learning for new input values ​​every 30 seconds. • Prediction and validation by time interval Epileptic seizures are characterized by their sudden and unpredictable nature, posing significant risks to patients’ daily lives. Accurate and reliable seizure prediction systems can provide alerts before a seizure occurs, as well as give the patient and caregiver provider enough time to take appropriate measures [ 17 ]. The pre-ictal stage is not too close to the actual seizure point, and the period sufficiently secured to respond to subsequent seizures is called the Seizure Prediction Horizon (SPH). The main goal of epilepsy prediction is to detect seizures within the range of Seizure Prediction Horizon (SPH). An appropriate SPH should include an appropriate time range for taking adequate intervention or preventive measures before actual seizures. A long prediction range can cause patient anxiety and pose a challenge to using neural network prediction models, while a short prediction range may result in insufficient preparation time for both patients and healthcare providers, ultimately failing to achieve the goal of effective epilepsy prediction [ 17 ]. Seizure symptoms can vary significantly from one patient to another due to various factors, including differences in the body parts where the seizure occurs. Therefore, various drugs and treatments are available, and the response time required for these treatments varies. Instead of the existing method for setting SPH commonly for all patients, an appropriate seizure response time should be established for each patient. However, due to the irregularity of seizure symptoms and the limitations of prediction models, it is not possible to specify a 100% accurate timing for seizures. Therefore, as shown in Fig. 3 , the seizure response period SPH was commonly set, and within it, the specific interval with the highest probability of seizure for each patient was designated as the specific warning period. In the literature on seizure prediction, the time for seizure prediction, SPH, has been variably set within a range of 5, 15, 30 minutes, or longer [ 18 – 21 ]. In general, there is no definition of how to select SPH, but if the SPH is too long, the patient may remain tense throughout the duration. Conversely, if the SPH is too short, there may not be sufficient time to respond effectively to the seizure, ultimately deviating from the study's objectives. In the case of the CHB-MIT database, many of the data containing seizures have a short pre-seizure period, so if the seizure prediction range is too long, it is difficult to secure data within an appropriate range that can be used. Therefore, it is essential to secure data by setting the seizure prediction horizon (SPH) within an appropriate range that is neither excessively long nor too short in response to seizures, and in this study, the SPH was set to 15 minutes. In existing studies, the pre-ictal period indicated the possibility of seizure occurrence in a wide SPH range, so the probability of seizure occurrence at a specific point in time cannot be determined. However, because the symptoms of seizures vary among patients, and different anticonvulsant medications must be administered or treatment methods executed accurately and promptly, it is essential to establish a timeframe during which seizures are likely to occur for each individual patient. Therefore, we used the k-fold cross-validation structure, an ensemble validation technique of deep learning, and divided it into 6 sections of 2.5 minutes within the range of 0 to 15 minutes (SPH) before the seizure occurred, as shown in Fig. 4 . Seizure prediction was performed for each section by alternating learning and testing for these sections six times, and the section with the highest probability of seizure occurrence was set individually for each patient. Experimental results • Performance index F1 score, sensitivity, and false discovery rate (FDR) were selected as performance indicators to evaluate the effectiveness of seizure prediction. The F1 score is an indicator similar to the commonly used accuracy, specifically designed to prevent distortion of accuracy that can occur when there is a significant imbalance in the number of samples across different classes within a dataset. The F1 score was calculated using sensitivity and precision, as shown in Fig. 5 and Eq. ( 6 ), making it less sensitive to the total volume of data. $$\:\mathbf{F}1\:\mathbf{s}\mathbf{c}\mathbf{o}\mathbf{r}\mathbf{e}\:=\:2\varvec{*}\frac{\left(\varvec{S}\varvec{e}\varvec{n}\varvec{s}\varvec{i}\varvec{t}\varvec{i}\varvec{v}\varvec{i}\varvec{t}\varvec{y}\varvec{*}\varvec{P}\varvec{r}\varvec{e}\varvec{c}\varvec{i}\varvec{s}\varvec{i}\varvec{o}\varvec{n}\right)}{(\varvec{S}\varvec{e}\varvec{n}\varvec{s}\varvec{i}\varvec{t}\varvec{i}\varvec{v}\varvec{i}\varvec{t}\varvec{y}+\varvec{P}\varvec{r}\varvec{e}\varvec{c}\varvec{i}\varvec{s}\varvec{i}\varvec{o}\varvec{n})}$$ 6 Due to the nature of ictal EEG, including the CHB-MIT database used in this study, there is relatively more data in the normal state without seizures than data with seizures, so the F1 score is considered to be a more appropriate performance indicator than accuracy. The calculation of sensitivity and FDR from the confusion matrix is shown in Fig. 6 . Sensitivity is the ratio of actual seizures that the classifier correctly identifies as seizures. The probability of predicting an actual seizure in advance is an important performance indicator because it is directly related to patient safety. FDR is the rate at which a seizure is incorrectly identified as occurring when it is not expected. It represents the probability of failing to predict an actual seizure in advance. FDR is an important indicator in the medical field directly related to patient safety and is often used along with sensitivity. • Section prediction results Table 1. Results of seizure prediction for time intervals The shaded time section in Table 1, which shows a certain level of prediction performance for each patient, was selected as the section where seizures can be predicted with high probability for that patient. Excluding chb12,15 and chb24 due to differences in sensor configuration and data size, the prediction performance was achieved with an average sensitivity of up to 0.98, an average F1 score of up to 0.93, and an average FDR of 0.13 for the data from 21 patients (Fig. 7 ). In addition, we observed patterns of high-risk seizure timing for each patient, with intervals ranging from a maximum of 12.5 to 15 minutes to a minimum of 0 to 2.5 minutes before the seizure occurred. Since seizure auras vary significantly among patients, they can manifest in various forms, including the possibility of not appearing at all or occurring either early or late in the seizure process. Additionally, substantial individual differences in brain wave patterns complicate the identification of common features. Therefore, research on predicting seizures relies on a personalized system, and we also generated patient-specific predictions using the CHB-MIT database. However, we categorized the response patterns for each patient into five distinct types, as shown in Fig. 8 . We aim to respond more effectively to each patient by utilizing these patterns as a reference. Figure 8 shows five types of probability patterns observed during the time intervals when pre-ictal signals are detected. (a) and (b) illustrate situations in which symptomatic signals are likely to begin either very close to or far from the seizure point, (c) is a situation in which they occur consistently with a high probability, (d) and (e) are situations in which they become weaker or stronger at a certain point in time. For example, patient chb22 belongs to (a), chb09 to (b), chb01 to (c), chb17 to (d), and chb23 to (e), respectively. In the experimental results, the seizure prediction at each time interval was represented as a probabilistic value, indicating the likelihood of a seizure occurring at any moment within 15 minutes. Therefore, all patients must prepare for seizure treatment within 0 to 15 minutes from when a seizure is predicted and an alarm is activated. Establishing a response plan for each patient is essential based on their pre-ictal patterns within that timeframe. First, as in previous studies, SPH was established within a 15-minute range. Subsequently, the period with the highest probability of seizures for each patient was designated as a specific warning period to facilitate a more detailed analysis of seizure responses. Therefore, we prevented the patient's state of tension from prolonging, allowing for a more efficient response to seizures through a short-term, intensive approach. This method is similar to weather forecasting based on specific dates and time intervals. Conclusion and discussion As a result, in this paper, we applied TQWT using the CHB-MIT EEG database to predict seizures with a simple CNN. We achieved an average sensitivity of 0.98, an average F1 score of 0.93, and an average FDR of 0.13 across 21 patients. To provide sufficient time for patients to respond to seizures, a target SPH range of 15 minutes was set for each patient. In addition, by making predictions at 2.5-minute intervals for each patient, we were able to respond to seizures more effectively by presenting a specific warning period with the highest possibility of seizures and a probability pattern for seizure prediction. Therefore, in the future, it is essential to analyze the seizure patterns of each patient and integrate them into a common framework, as shown in Fig. 8 , and develop a more accurate seizure prediction model using an algorithm that autonomously identifies these patterns. The CHB-MIT database used in this paper consists of 24 data groups collected from 23 patients and encompasses nearly 1,000 hours of EEG recordings, which include more than 150 seizures. However, not all patients had the same EEG channel configuration, and the electrode connection structure of the chb12 and chb15 groups is significantly different from that of the other groups. Therefore, even when the same CHB-MIT database is used, the dataset is frequently configured and utilized in different ways based on the researcher's preferences. It is difficult to compare predictive performances, such as sensitivity, across multiple studies. We anticipate that various approaches can be compared in the future based on the data utilization method outlined in this paper, which employs groups with 18 common channels and a specified data length. Through the time section predictions, the results for each patient were displayed. However, this indicates that the probability of a seizure and its associated signs can occur at any moment, potentially leading to an actual seizure. Although this method does not guarantee a perfect prediction of 100% during the special caution period and 0% during other periods, it offers the advantage of helping patients understand their seizure probability patterns and prepare accordingly. In cases where signs of a seizure appear long before the actual event, the patient experiences a prolonged period of tension. In the future, we will also examine cases in which more than 15 minutes are required to respond effectively to a seizure. Declarations Data availability CHB-MIT Scalp EEG database that support the findings of this study have been deposited in the PhysioNet(https://doi.org/10.13026/C2K01R). Data is provided within the manuscript. Acknowledgements This study was funded by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education(NRF-2021R1I1A3043911) References The American Epilepsy Society’s seizure prediction challenge, https://www.kaggle.com/c/seizure-prediction, last access 2024/12/10. Jana, R., Mukherjee, I. Deep learning based efficient epileptic seizure prediction with EEG channel optimization. Biomed. Signal Process. Control 68, 102767, DOI: https://doi.org/10.1016/j.bspc.2021.102767 (2021). Elgohary, S., Eldawlatly, S. & Khalil, M. I. Epileptic seizure prediction using zero-crossings analysis of EEG wavelet detail coefficients. Proc. IEEE Conf. Comput. Intell. Bioinf. Comput. Biol. (CIBCB) , 1–8, DOI: https://doi.org/10.1109/CIBCB.2016.7758115 (2016). Bandarabadi, M., Teixeira, C. A., Rasekhi, J. & Dourado, A. Epileptic seizure prediction using relative spectral power features. Clin. Neurophysiol. 126, 237–248, DOI: https://doi.org/10.1016/j.clinph.2014.04.029 (2015). Moghim, N. & Corne, D. W. Predicting epileptic seizures in advance. PloS One 9, e99334, DOI: https://doi.org/10.1371/journal.pone.0099334 (2014). Teixeira, C. A. et al. Epileptic seizure predictors based on computational intelligence techniques: A comparative study with 278 patients. Comput. Methods Programs Biomed. 114, 324–336, DOI: https://doi.org/10.1016/j.cmpb.2014.01.004 (2014). Shen, M., Wen, P., Song, B. & Li, Y. An EEG-based real-time epilepsy seizure detection approach using discrete wavelet transform and machine learning methods. Biomed. Signal Process. Control 77, 103820, DOI: https://doi.org/10.1016/j.bspc.2022.103820 (2022). Zhang, Z. & Parhi, K. K. Low-complexity seizure prediction from iEEG/sEEG using spectral power and ratios of spectral power. IEEE Trans. Biomed. Circuits Syst. 10, 693–706, DOI: https://doi.org/10.1016/10.1109/TBCAS.2015.2477264 (2015). Tsiouris, K. M. et al. A long short-term memory deep learning network for the prediction of epileptic seizures using EEG signals. Comput. Biol. Med. 99, 24–37, DOI: https://doi.org/10.1016/j.compbiomed.2018.05.019 (2018). Truong, N. D. et al. Convolutional neural networks for seizure prediction using intracranial and scalp electroencephalogram. Neural Netw. 105, 104–111, DOI: https://doi.org/10.1016/j.neunet.2018.04.018 (2018). Liu, X. et al. Epileptic seizure prediction based on EEG using pseudo-three-dimensional CNN. Front. Neuroinform. 18, 1354436, DOI: https://doi.org/10.3389/fnins.2024.1354436 (2024). Khan, H., Marcuse, L., Fields, M., Swann, K. & Yener, B. Focal onset seizure prediction using convolutional networks. IEEE Trans. Biomed. Eng. 65, 2109–2118, DOI: https://doi.org/10.1109/TBME.2017.2785401 (2017). Kitano, L. A. S. et al. Epileptic seizure prediction from EEG signals using unsupervised learning and a polling-based decision process. Proc. Int. Conf. Artif. Neural Netw. (ICANN) , 27, 731–742, DOI: https://doi.org/10.1007/978-3-030-01424-773 (2018). Jana, R., Bhattacharyya, S. & Das, S. Patient-specific seizure prediction using convolutional neural networks. DoSIER Intell. Res. 51–60, DOI: https://doi.org/10.1007/978-981-13-2542-06 (2019). Jana, R., Bhattacharyya, S. & Das, S. Epileptic seizure prediction from EEG signals using DenseNet. Proc. IEEE Symp. Series Comput. Intell. (SSCI) , 1–6, DOI: https://doi.org/10.1109/SSCI44817.2019.9003059 (2019). Daoud, H. & Bayoumi, M. A. Efficient epileptic seizure prediction based on deep learning. IEEE Trans. Biomed. Circuits Syst. 13, 804–813, DOI: https://doi.org/10.1109/TBCAS.2019.2900723 (2019). Schelter, B. et al. Seizure prediction: The impact of long prediction horizons. Epilepsy Res. 73, 213–217, DOI: https://doi.org/10.1016/j.eplepsyres.2006.12.003 (2007). Niederhauser, J. J. et al. Detection of seizure precursors from depth-EEG using a sign periodogram transform. IEEE Trans. Biomed. Eng. 50, 449–458, DOI: https://doi.org/10.1109/TBME.2003.809112 (2003). Van Quyen, M. L. et al. Anticipating epileptic seizures in real time by a non-linear analysis of similarity between EEG recordings. Neuroreport 10, 2149–2155, DOI: https://doi.org/10.1097/00001756-199907050-00015 (1999). Park, Y., Luo, L., Parhi, K. K. & Netoff, T. Seizure prediction with spectral power of EEG using cost-sensitive support vector machines. Epilepsia 52, 1761–1770, DOI: https://doi.org/10.1111/j.1528-1167.2011.03176.x (2011). Howbert, J. J. et al. Forecasting seizures in dogs with naturally occurring epilepsy. PloS One 9, e81920, DOI: https://doi.org/10.1371/journal.pone.0081920 (2014). Table 1 Table 1 is available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files table1.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5714799","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":398841425,"identity":"3a8160b7-686d-4391-b6f6-ad3b2027020c","order_by":0,"name":"YongUn Jo","email":"","orcid":"","institution":"Konyang University","correspondingAuthor":false,"prefix":"","firstName":"YongUn","middleName":"","lastName":"Jo","suffix":""},{"id":398841426,"identity":"e4c12300-0f47-443c-9c07-7cc4d7ad3b63","order_by":1,"name":"Do Chang Oh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIiWNgGAWjYDCCAwwMzAwVCUgiPERpOUOyFsY2UrTwHW9++LhwXpqcuUTys4dfGOzkGXjOPsCrRfLMMWPjmdtyjC1npJkbyzAkGzbwthvg1WJwI4dNmndbReKGGwlm0hIMzAkM/Gz4HQbRMgekJf0bUEs9sVoacoBacswkPzAcTmDgbcOvBewXnmNpxgZn3pRJMxgcN2zjOYZfCzjEeGqS5QyOp2+T/FFRLc/Pk4ZfCwIIJDAw8wDDioBPkAH/AQbGH8QrHwWjYBSMghEEAPhFP+PrajWxAAAAAElFTkSuQmCC","orcid":"","institution":"Konyang University","correspondingAuthor":true,"prefix":"","firstName":"Do","middleName":"Chang","lastName":"Oh","suffix":""}],"badges":[],"createdAt":"2024-12-26 08:38:37","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5714799/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5714799/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":73316508,"identity":"d9788764-ad9c-4e05-834a-19d6f8b022dc","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":74112,"visible":true,"origin":"","legend":"\u003cp\u003eSimplified TQWT\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/e6a3500561f2a4b42707e6f0.png"},{"id":73317950,"identity":"c9897938-958f-467d-9039-b46ddc08b069","added_by":"auto","created_at":"2025-01-08 20:23:23","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":339566,"visible":true,"origin":"","legend":"\u003cp\u003eCNN model structure\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/71e44b6ee5bd16afea030700.png"},{"id":73316511,"identity":"65d3ae3d-318e-4825-a078-93895da7972b","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":28155,"visible":true,"origin":"","legend":"\u003cp\u003eExample of SPH and specific warning period\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/0942975f013338a356682588.png"},{"id":73316517,"identity":"3bd6c19e-4885-4c51-b1ed-e06cf7e96a1a","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":106759,"visible":true,"origin":"","legend":"\u003cp\u003eApplication of k-fold cross-validation\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/9fdb426da6645a7587bc5209.png"},{"id":73317418,"identity":"b7d0c2ec-ef47-4e16-bf05-4557b9be439a","added_by":"auto","created_at":"2025-01-08 20:15:23","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":47703,"visible":true,"origin":"","legend":"\u003cp\u003eExpressions between sensitivity, precision, and f1 scores\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/405ef73198b806a30aa4f2c4.png"},{"id":73316515,"identity":"0c59c49a-6ccc-42e0-9241-d9a192ac2d63","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":31928,"visible":true,"origin":"","legend":"\u003cp\u003eSensitivity and FDR in confusion Matrix\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/f5f1c02bee699d04f1df2359.png"},{"id":73316528,"identity":"b3bb4608-2f79-4644-9bff-18dee584c061","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":250451,"visible":true,"origin":"","legend":"\u003cp\u003eMaximum, minimum, and mean value for each patient’s sensitivity and F1 score.\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/063e2347d67ff083e32e276a.png"},{"id":73316520,"identity":"15249c2e-7108-4e04-b428-25992ee00a8b","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":201650,"visible":true,"origin":"","legend":"\u003cp\u003eFive categories of pre-ictal EEG patterns\u003c/p\u003e","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/3e8d7ad521953b477d99c558.png"},{"id":76269592,"identity":"b6096dec-3073-46bb-9fb9-a8724334a19b","added_by":"auto","created_at":"2025-02-14 08:24:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1509240,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/56656fa8-59e9-4a61-b5d1-fa0f0fda4f28.pdf"},{"id":73316509,"identity":"4945ceb6-4663-4514-a315-d308bc65859f","added_by":"auto","created_at":"2025-01-08 20:07:23","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":324784,"visible":true,"origin":"","legend":"","description":"","filename":"table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-5714799/v1/ef1cc72b87525fe8a447582e.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Patient specific seizure prediction for time intervals using TQWT and deep learning","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAccording to data released by the World Health Organization (WHO) [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], approximately 50\u0026nbsp;million people worldwide suffer from epilepsy, which is known to be the most common neurological disease. In particular, patients with epilepsy face a risk of premature death that is up to three times higher than that of the general population, with a significant number of fatalities attributed to seizures related to the condition. Epilepsy can be caused by various factors, including congenital brain abnormalities, genetic disorders, acquired head injuries, and brain tumors. A diagnosis of epilepsy is made when two or more spontaneous seizures occur. Epileptic seizures are characterized by irregular and nonperiodic activation of part or all of the brain, resulting in various symptoms that vary from patient to patient. These symptoms include temporary loss of consciousness, cognitive dysfunction, and paralysis of specific body parts or the entire body. Patients with epilepsy experience a higher incidence of seizure-related injuries than the general population. These injuries can range from minor injuries, such as fractures and bruises, to more severe physical complications, including traffic accidents. Furthermore, individuals with epilepsy may encounter psychological challenges, such as depression and anxiety. For this reason, extensive research has been conducted to detect seizures through various biological signals to prevent injuries caused by seizures. Additionally, further studies have focused on predicting and preparing for seizures [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003ePublic databases commonly used in seizure research include the Hospital iEEG, CHB-MIT sEEG, the American Epilepsy Society Seizure Prediction Challenge iEEG, and the European Epilepsy Database. The CHB-MIT database comprises approximately 1,000 hours of recordings and includes 198 seizures from long-term monitoring of 23 refractory patients at Boston Children's Hospital. Seizures can be broadly categorized into three stages: inter-ictal (the period between seizures), pre-ictal (the period before a seizure), and ictal (the period during a seizure). Among these, the pre-ictal stage is when precursor symptoms of a seizure manifest, allowing for predicting seizures through their recognition.\u003c/p\u003e \u003cp\u003eHowever, a standard or optimal point for dividing the three stages of seizures has not been clearly defined. In the American Epilepsy Society\u0026rsquo;s Seizure Prediction Challenge, the pre-ictal stage was established as one hour before the onset of the seizure. In some studies, researchers autonomously choose a very short period before a seizure, such as 5 minutes or less. In other cases, they may choose a relatively longer period of 20 to 30 minutes or even an extended period of 90 minutes. The primary reason for selecting the stage before the seizure is to predict the patient\u0026rsquo;s seizure in advance and to ensure an adequate response time.\u003c/p\u003e \u003cp\u003eMost existing studies on seizure prediction using electroencephalogram(EEG) have been conducted under patient-specific conditions. According to Jana and Mukherjee, The EEG signal patterns are unique for each epilepsy patient [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. One of the challenges is the variations across seizures even for the same patient [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Consequently, research on seizure prediction using EEG is challenging to generalize across a large patient population, so it is primarily considered patient specific. In particular, patient-specific characteristics include not only the EEG pattern but also the duration and timing of the onset of pre-ictal symptoms. Bandarabadi et al. selected various pre-ictal periods for several patients, ultimately predicting seizures within a range of 5 to 173 minutes, and established 44.3 minutes as the optimal interval for seizure prediction [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Moghim and Corne set the pre-ictal period as ranging from 0 to 20 minutes and developed an algorithm that selects the optimal pre-ictal period for each patient [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The potential between electrodes changes as a result of effects such as eye movement or muscular activity, causing an artifact [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Therefore, in research using EEG, preprocessing to eliminate various types of noise plays a significant role, and the results can vary considerably depending on the selected preprocessing method [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. For instance, Shen et al. performed frequency analysis using the Digital Wavelet Transform (DWT) in conjunction with an algorithm that extracts eight types of eigenvalues [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. In contrast, Zhang and Parhi preprocessed the signal by extracting 44 feature values, including spectral power analysis [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Tsiouris et al. extracted and integrated multiple features using various techniques, including cross-correlation estimation, analysis in both the time and frequency domains, and graph theory [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The preprocessing methods utilized in the previous study can be classified into two main types: feature value extraction in the time domain, such as zero crossing and cross-correlation estimation, and analysis in the frequency domain, which includes Fourier transform and wavelet transform. The feature value extraction method in the time domain is a traditional technique that consistently demonstrates strong performance in biosignal analysis. In contrast, frequency analysis is particularly well-suited to electroencephalography (EEG), as it enables the examination of standardized frequency bands, including delta, theta, alpha, beta, and gamma waves.\u003c/p\u003e \u003cp\u003eIn this paper, TQWT was applied to data from 21 individuals in the CHB-MIT database, excluding certain data that have significantly different shapes, and seizures are predicted by identifying pre-ictal EEG patterns using a straightforward Convolutional Neural Network (CNN) model. Additionally, by employing the k-fold cross-validation technique in deep learning, we were able to evaluate predictive performance at specific time points and identify the most appropriate seizure response time interval for each patient.\u003c/p\u003e\n\u003ch3\u003eSeizure prediction for time intervals with TQWT and deep learning\u003c/h3\u003e\n\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e\u0026bull; CHB-MIT database\u003c/h2\u003e \u003cp\u003eIn most cases, the CHB-MIT database includes the 18 EEG channels from the International 10\u0026ndash;20 system. Additionally, in some cases, ECG or Vagus Nerve Stimulation (VNS) signals or dummy channels may be used to meet the required number of channels. This channel configuration can result in numerous changes, even within a single case, making it difficult to select batches of data. Therefore, even among studies that use the same CHB-MIT database, some cases may be excluded. Truong et al. argued that pre-ictal ranges may be unreliable if channel configurations differ significantly, or seizures are repeated multiple times within 30 minutes [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Excluding these data, only 13 cases were ultimately used. Jana and Mukherjee used 23 reliable cases that included patient information, but only 22 EEG channels commonly used in many cases were selected, and the rest were excluded.\u003c/p\u003e \u003cp\u003eTsiouris et al. used 24 cases, excluding some data that excluded the 18 EEG channels commonly used in most cases. As a special case, Liu et al. used all 24 cases, but some signals such as electrocardiogram (ECG) were removed and only EEG signals were used [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eA total of 21 cases were used in this paper, excluding cases chb12 and 15 with changes in electrode configuration (for example, change from FP1-F7 to FP1-ref) and chb24 with a very short data size. Most of the data contained in the CHB-MIT database are commonly used EEG channels, but some data have additional EEG channels or contain other signals, such as ECG. In particular, the configuration of channels changes depending on the measurement time, even for the same patient, which makes it difficult to apply deep learning classifiers that require constant data. Therefore, only 18 EEG channel data commonly used in most cases were selected, and cases with fewer than 18 channels were excluded. Pre-ictal data for classification learning and testing were extracted up to 15 minutes before seizure onset, and inter-ictal data were extracted from a period without seizures for 1 hour, and the amount of the two data types was balanced as much as possible.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e• TQWT for pre-processing\u003c/h3\u003e\n\u003cp\u003eTo predict seizures more accurately, accurate feature extraction and noise removal via preprocessing are important. The most commonly used pre-processing transforms are the continuous wavelet transform (CWT) [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], short-time Fourier transform (STFT) [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], and discrete wavelet transform (DWT) [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Some researchers have applied deep learning to raw EEG data for seizure prediction [\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. However, methods that are too complex or extract too many feature values are not suitable in terms of efficiency, and therefore DWT, which requires less calculation, is used more than CWT.\u003c/p\u003e \u003cp\u003eTunable Q Factor Wavelet Transform (TQWT) is an improved version of the basic Discrete Wavelet Transform (DWT), as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and was selected as the preprocessing technique. Similar to the existing DWT, TQWT performs multiple conversion steps by combining high-frequency and low-frequency filters, but is a technology designed to intensively analyze the desired frequency band by introducing a Q-factor that can be selected by the user during the conversion process. Due to these characteristics, TQWT was judged to be suitable for the analysis of EEG, where specific frequency bands have special meaning.\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{y}_{high}\\left[n\\right]=\\left({x}^{*}{h}_{1}\\right)\\left[n\\right]=\\:\\sum\\:_{k=-\\infty\\:}^{\\infty\\:}x\\left[k\\right]{h}_{1}\\left[n-k\\right]$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{y}_{low}\\left[n\\right]=\\left({x}^{*}{h}_{0}\\right)\\left[n\\right]=\\:\\sum\\:_{k=-\\infty\\:}^{\\infty\\:}x\\left[k\\right]{h}_{0}[n-k]$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:{H}_{0}^{\\left(j\\right)}\\left(\\omega\\:\\right)=\\left\\{\\begin{array}{cc}\\:\\:\\prod\\:_{m=0}^{j-1}{H}_{0}\\left(\\frac{\\omega\\:}{{\\alpha\\:}^{m}}\\right),\u0026amp;\\:if\\left|\\omega\\:\\right|\\le\\:{\\alpha\\:}^{j}\\pi\\:\\\\\\:0,\u0026amp;\\:\\:\\:\\:\\:\\:if{\\alpha\\:}^{j}\\pi\\:\\:\\le\\:\\left|\\omega\\:\\right|\\le\\:\\pi\\:\\end{array}\\right.\\:$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:{H}_{1}^{\\left(j\\right)}\\left(\\omega\\:\\right)=\\left\\{\\:\\:\\begin{array}{cc}{H}_{1}\\left(\\frac{\\omega\\:}{{\\alpha\\:}^{j-1}}\\right)\\prod\\:_{m=0}^{j-2}{H}_{0}\\left(\\frac{\\omega\\:}{{\\alpha\\:}^{m}}\\right),\\:\\:\u0026amp;\\:\\:\\:if\\:\\left(1-\\beta\\:\\right){\\alpha\\:}^{j-1}\\pi\\:\\:\\le\\:\\left|\\omega\\:\\right|\\le\\:\\:{\\alpha\\:}^{j-1}\\pi\\:\\\\\\:0,\u0026amp;\\:\\begin{array}{c}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\\\\\:for\\:other\\:\\omega\\:\\in\\:\\left[-\\pi\\:,\\:\\pi\\:\\right]\\end{array}\\end{array}\\right.\\:\\:$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:\\beta\\:=\\frac{2}{Q+1},\\:\\:\\:\\:\\alpha\\:=1-\\frac{\\beta\\:}{r}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDWT uses the results of passing the input signal \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:x\\left(n\\right)\\)\u003c/span\u003e\u003c/span\u003e through a high-frequency filter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{1}\\left(n\\right)\\)\u003c/span\u003e\u003c/span\u003e and a low-frequency filter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{0}\\left(\\text{n}\\right)\\)\u003c/span\u003e\u003c/span\u003e, respectively, as shown in equations (\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) and (\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). This is advanced to TQWT, and scaling of low and high frequencies is applied at the j-th stage, respectively, as shown in equations (\u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), (\u003cspan refid=\"Equ4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), and (5). When the sampling frequency of x(n) is \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{f}_{s}\\)\u003c/span\u003e\u003c/span\u003e, α, and β are scaled to \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\alpha\\:}{\\text{f}}_{\\text{s}}\\)\u003c/span\u003e\u003c/span\u003e(\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:0\u0026lt;{\\alpha\\:}\u0026lt;1\\)\u003c/span\u003e\u003c/span\u003e) and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\beta\\:}{\\text{f}}_{\\text{s}}\\)\u003c/span\u003e\u003c/span\u003e(\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:0\u0026lt;{\\beta\\:}\u0026lt;1\\)\u003c/span\u003e\u003c/span\u003e), respectively, and the condition for lowering the sampling frequency of the filter based on the Q factor is presented in Eq.\u0026nbsp;(\u003cspan refid=\"Equ5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), where r means oversampling. Selesnick recommended that the r value be between 2 and 4 so that the sampling frequency of the low- and high-frequency filters can be changed by adjusting the Q-factor.\u003c/p\u003e\n\u003ch3\u003e• CNN model\u003c/h3\u003e\n\u003cp\u003eThe CNN model used in this study, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, a relatively straightforward architecture, comprising two convolutional layers that integrate convolution and max pooling, followed by two fully connected layers. Due to the nature of the CHB-MIT database, a substantial amount of data is available for classification between pre-ictal and inter-ictal states; therefore, the dropout technique was applied to the fully connected layer. Because the goal was to predict seizures through real-time monitoring, which requires fast processing speed and efficient power consumption in wearable hardware capable of independent operation, the model was designed to be as light as possible.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe input signal was processed by applying a 30-second window to the original signal, preprocessing was then performed using TQWP, and the prediction results were updated through deep learning for new input values ​​every 30 seconds.\u003c/p\u003e\n\u003ch3\u003e• Prediction and validation by time interval\u003c/h3\u003e\n\u003cp\u003eEpileptic seizures are characterized by their sudden and unpredictable nature, posing significant risks to patients\u0026rsquo; daily lives. Accurate and reliable seizure prediction systems can provide alerts before a seizure occurs, as well as give the patient and caregiver provider enough time to take appropriate measures [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The pre-ictal stage is not too close to the actual seizure point, and the period sufficiently secured to respond to subsequent seizures is called the Seizure Prediction Horizon (SPH).\u003c/p\u003e \u003cp\u003eThe main goal of epilepsy prediction is to detect seizures within the range of Seizure Prediction Horizon (SPH). An appropriate SPH should include an appropriate time range for taking adequate intervention or preventive measures before actual seizures. A long prediction range can cause patient anxiety and pose a challenge to using neural network prediction models, while a short prediction range may result in insufficient preparation time for both patients and healthcare providers, ultimately failing to achieve the goal of effective epilepsy prediction [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSeizure symptoms can vary significantly from one patient to another due to various factors, including differences in the body parts where the seizure occurs. Therefore, various drugs and treatments are available, and the response time required for these treatments varies. Instead of the existing method for setting SPH commonly for all patients, an appropriate seizure response time should be established for each patient. However, due to the irregularity of seizure symptoms and the limitations of prediction models, it is not possible to specify a 100% accurate timing for seizures. Therefore, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the seizure response period SPH was commonly set, and within it, the specific interval with the highest probability of seizure for each patient was designated as the specific warning period.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the literature on seizure prediction, the time for seizure prediction, SPH, has been variably set within a range of 5, 15, 30 minutes, or longer [\u003cspan additionalcitationids=\"CR19 CR20\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. In general, there is no definition of how to select SPH, but if the SPH is too long, the patient may remain tense throughout the duration. Conversely, if the SPH is too short, there may not be sufficient time to respond effectively to the seizure, ultimately deviating from the study's objectives. In the case of the CHB-MIT database, many of the data containing seizures have a short pre-seizure period, so if the seizure prediction range is too long, it is difficult to secure data within an appropriate range that can be used. Therefore, it is essential to secure data by setting the seizure prediction horizon (SPH) within an appropriate range that is neither excessively long nor too short in response to seizures, and in this study, the SPH was set to 15 minutes.\u003c/p\u003e \u003cp\u003eIn existing studies, the pre-ictal period indicated the possibility of seizure occurrence in a wide SPH range, so the probability of seizure occurrence at a specific point in time cannot be determined. However, because the symptoms of seizures vary among patients, and different anticonvulsant medications must be administered or treatment methods executed accurately and promptly, it is essential to establish a timeframe during which seizures are likely to occur for each individual patient.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTherefore, we used the k-fold cross-validation structure, an ensemble validation technique of deep learning, and divided it into 6 sections of 2.5 minutes within the range of 0 to 15 minutes (SPH) before the seizure occurred, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. Seizure prediction was performed for each section by alternating learning and testing for these sections six times, and the section with the highest probability of seizure occurrence was set individually for each patient.\u003c/p\u003e"},{"header":"Experimental results","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e\u0026bull; Performance index\u003c/h2\u003e \u003cp\u003eF1 score, sensitivity, and false discovery rate (FDR) were selected as performance indicators to evaluate the effectiveness of seizure prediction. The F1 score is an indicator similar to the commonly used accuracy, specifically designed to prevent distortion of accuracy that can occur when there is a significant imbalance in the number of samples across different classes within a dataset. The F1 score was calculated using sensitivity and precision, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and Eq.\u0026nbsp;(\u003cspan refid=\"Equ6\" class=\"InternalRef\"\u003e6\u003c/span\u003e), making it less sensitive to the total volume of data.\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\:\\mathbf{F}1\\:\\mathbf{s}\\mathbf{c}\\mathbf{o}\\mathbf{r}\\mathbf{e}\\:=\\:2\\varvec{*}\\frac{\\left(\\varvec{S}\\varvec{e}\\varvec{n}\\varvec{s}\\varvec{i}\\varvec{t}\\varvec{i}\\varvec{v}\\varvec{i}\\varvec{t}\\varvec{y}\\varvec{*}\\varvec{P}\\varvec{r}\\varvec{e}\\varvec{c}\\varvec{i}\\varvec{s}\\varvec{i}\\varvec{o}\\varvec{n}\\right)}{(\\varvec{S}\\varvec{e}\\varvec{n}\\varvec{s}\\varvec{i}\\varvec{t}\\varvec{i}\\varvec{v}\\varvec{i}\\varvec{t}\\varvec{y}+\\varvec{P}\\varvec{r}\\varvec{e}\\varvec{c}\\varvec{i}\\varvec{s}\\varvec{i}\\varvec{o}\\varvec{n})}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDue to the nature of ictal EEG, including the CHB-MIT database used in this study, there is relatively more data in the normal state without seizures than data with seizures, so the F1 score is considered to be a more appropriate performance indicator than accuracy.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe calculation of sensitivity and FDR from the confusion matrix is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Sensitivity is the ratio of actual seizures that the classifier correctly identifies as seizures. The probability of predicting an actual seizure in advance is an important performance indicator because it is directly related to patient safety. FDR is the rate at which a seizure is incorrectly identified as occurring when it is not expected. It represents the probability of failing to predict an actual seizure in advance. FDR is an important indicator in the medical field directly related to patient safety and is often used along with sensitivity.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e• Section prediction results\u003c/h3\u003e\n\u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eTable\u0026nbsp;1.\u003c/b\u003e Results of seizure prediction for time intervals\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe shaded time section in Table\u0026nbsp;1, which shows a certain level of prediction performance for each patient, was selected as the section where seizures can be predicted with high probability for that patient. Excluding chb12,15 and chb24 due to differences in sensor configuration and data size, the prediction performance was achieved with an average sensitivity of up to 0.98, an average F1 score of up to 0.93, and an average FDR of 0.13 for the data from 21 patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). In addition, we observed patterns of high-risk seizure timing for each patient, with intervals ranging from a maximum of 12.5 to 15 minutes to a minimum of 0 to 2.5 minutes before the seizure occurred. Since seizure auras vary significantly among patients, they can manifest in various forms, including the possibility of not appearing at all or occurring either early or late in the seizure process. Additionally, substantial individual differences in brain wave patterns complicate the identification of common features.\u003c/p\u003e \u003cp\u003eTherefore, research on predicting seizures relies on a personalized system, and we also generated patient-specific predictions using the CHB-MIT database. However, we categorized the response patterns for each patient into five distinct types, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e. We aim to respond more effectively to each patient by utilizing these patterns as a reference.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e shows five types of probability patterns observed during the time intervals when pre-ictal signals are detected. (a) and (b) illustrate situations in which symptomatic signals are likely to begin either very close to or far from the seizure point, (c) is a situation in which they occur consistently with a high probability, (d) and (e) are situations in which they become weaker or stronger at a certain point in time. For example, patient chb22 belongs to (a), chb09 to (b), chb01 to (c), chb17 to (d), and chb23 to (e), respectively.\u003c/p\u003e \u003cp\u003eIn the experimental results, the seizure prediction at each time interval was represented as a probabilistic value, indicating the likelihood of a seizure occurring at any moment within 15 minutes. Therefore, all patients must prepare for seizure treatment within 0 to 15 minutes from when a seizure is predicted and an alarm is activated. Establishing a response plan for each patient is essential based on their pre-ictal patterns within that timeframe. First, as in previous studies, SPH was established within a 15-minute range. Subsequently, the period with the highest probability of seizures for each patient was designated as a specific warning period to facilitate a more detailed analysis of seizure responses. Therefore, we prevented the patient's state of tension from prolonging, allowing for a more efficient response to seizures through a short-term, intensive approach. This method is similar to weather forecasting based on specific dates and time intervals.\u003c/p\u003e"},{"header":"Conclusion and discussion","content":"\u003cp\u003eAs a result, in this paper, we applied TQWT using the CHB-MIT EEG database to predict seizures with a simple CNN. We achieved an average sensitivity of 0.98, an average F1 score of 0.93, and an average FDR of 0.13 across 21 patients. To provide sufficient time for patients to respond to seizures, a target SPH range of 15 minutes was set for each patient. In addition, by making predictions at 2.5-minute intervals for each patient, we were able to respond to seizures more effectively by presenting a specific warning period with the highest possibility of seizures and a probability pattern for seizure prediction. Therefore, in the future, it is essential to analyze the seizure patterns of each patient and integrate them into a common framework, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, and develop a more accurate seizure prediction model using an algorithm that autonomously identifies these patterns.\u003c/p\u003e \u003cp\u003eThe CHB-MIT database used in this paper consists of 24 data groups collected from 23 patients and encompasses nearly 1,000 hours of EEG recordings, which include more than 150 seizures. However, not all patients had the same EEG channel configuration, and the electrode connection structure of the chb12 and chb15 groups is significantly different from that of the other groups. Therefore, even when the same CHB-MIT database is used, the dataset is frequently configured and utilized in different ways based on the researcher's preferences. It is difficult to compare predictive performances, such as sensitivity, across multiple studies. We anticipate that various approaches can be compared in the future based on the data utilization method outlined in this paper, which employs groups with 18 common channels and a specified data length.\u003c/p\u003e \u003cp\u003eThrough the time section predictions, the results for each patient were displayed. However, this indicates that the probability of a seizure and its associated signs can occur at any moment, potentially leading to an actual seizure. Although this method does not guarantee a perfect prediction of 100% during the special caution period and 0% during other periods, it offers the advantage of helping patients understand their seizure probability patterns and prepare accordingly. In cases where signs of a seizure appear long before the actual event, the patient experiences a prolonged period of tension. In the future, we will also examine cases in which more than 15 minutes are required to respond effectively to a seizure.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003eData availability\u003c/p\u003e\n\u003cp\u003eCHB-MIT Scalp EEG database that support the findings of this study have been deposited in the PhysioNet(https://doi.org/10.13026/C2K01R).\u003c/p\u003e\n\u003cp\u003eData is provided within the manuscript.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThis study was funded by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education(NRF-2021R1I1A3043911)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eThe American Epilepsy Society\u0026rsquo;s seizure prediction challenge, https://www.kaggle.com/c/seizure-prediction, last access 2024/12/10.\u003c/li\u003e\n\u003cli\u003eJana, R., Mukherjee, I. Deep learning based efficient epileptic seizure prediction with EEG channel optimization. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e 68, 102767, DOI: https://doi.org/10.1016/j.bspc.2021.102767 (2021).\u003c/li\u003e\n\u003cli\u003eElgohary, S., Eldawlatly, S. \u0026amp; Khalil, M. I. Epileptic seizure prediction using zero-crossings analysis of EEG wavelet detail coefficients. \u003cem\u003eProc. IEEE Conf. Comput. Intell. Bioinf. Comput. Biol. (CIBCB)\u003c/em\u003e, 1\u0026ndash;8, DOI: https://doi.org/10.1109/CIBCB.2016.7758115 (2016).\u003c/li\u003e\n\u003cli\u003eBandarabadi, M., Teixeira, C. A., Rasekhi, J. \u0026amp; Dourado, A. Epileptic seizure prediction using relative spectral power features. \u003cem\u003eClin. Neurophysiol.\u003c/em\u003e 126, 237\u0026ndash;248, DOI: https://doi.org/10.1016/j.clinph.2014.04.029 (2015).\u003c/li\u003e\n\u003cli\u003eMoghim, N. \u0026amp; Corne, D. W. Predicting epileptic seizures in advance. \u003cem\u003ePloS One\u003c/em\u003e 9, e99334, DOI: https://doi.org/10.1371/journal.pone.0099334 (2014).\u003c/li\u003e\n\u003cli\u003eTeixeira, C. A. et al. Epileptic seizure predictors based on computational intelligence techniques: A comparative study with 278 patients. \u003cem\u003eComput. Methods Programs Biomed.\u003c/em\u003e 114, 324\u0026ndash;336, DOI: https://doi.org/10.1016/j.cmpb.2014.01.004 (2014).\u003c/li\u003e\n\u003cli\u003eShen, M., Wen, P., Song, B. \u0026amp; Li, Y. An EEG-based real-time epilepsy seizure detection approach using discrete wavelet transform and machine learning methods. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e 77, 103820, DOI: https://doi.org/10.1016/j.bspc.2022.103820 (2022).\u003c/li\u003e\n\u003cli\u003eZhang, Z. \u0026amp; Parhi, K. K. Low-complexity seizure prediction from iEEG/sEEG using spectral power and ratios of spectral power. \u003cem\u003eIEEE Trans. Biomed. Circuits Syst.\u003c/em\u003e 10, 693\u0026ndash;706, DOI: https://doi.org/10.1016/10.1109/TBCAS.2015.2477264 (2015).\u003c/li\u003e\n\u003cli\u003eTsiouris, K. M. et al. A long short-term memory deep learning network for the prediction of epileptic seizures using EEG signals. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e 99, 24\u0026ndash;37, DOI: https://doi.org/10.1016/j.compbiomed.2018.05.019 (2018).\u003c/li\u003e\n\u003cli\u003eTruong, N. D. et al. Convolutional neural networks for seizure prediction using intracranial and scalp electroencephalogram. \u003cem\u003eNeural Netw.\u003c/em\u003e 105, 104\u0026ndash;111, DOI: https://doi.org/10.1016/j.neunet.2018.04.018 (2018).\u003c/li\u003e\n\u003cli\u003eLiu, X. et al. Epileptic seizure prediction based on EEG using pseudo-three-dimensional CNN. \u003cem\u003eFront. Neuroinform.\u003c/em\u003e 18, 1354436, DOI: https://doi.org/10.3389/fnins.2024.1354436 (2024).\u003c/li\u003e\n\u003cli\u003eKhan, H., Marcuse, L., Fields, M., Swann, K. \u0026amp; Yener, B. Focal onset seizure prediction using convolutional networks. \u003cem\u003eIEEE Trans. Biomed. Eng.\u003c/em\u003e 65, 2109\u0026ndash;2118, DOI: https://doi.org/10.1109/TBME.2017.2785401 (2017).\u003c/li\u003e\n\u003cli\u003eKitano, L. A. S. et al. Epileptic seizure prediction from EEG signals using unsupervised learning and a polling-based decision process. \u003cem\u003eProc. Int. Conf. Artif. Neural Netw. (ICANN)\u003c/em\u003e, 27, 731\u0026ndash;742, DOI: https://doi.org/10.1007/978-3-030-01424-773 (2018).\u003c/li\u003e\n\u003cli\u003eJana, R., Bhattacharyya, S. \u0026amp; Das, S. Patient-specific seizure prediction using convolutional neural networks. \u003cem\u003eDoSIER Intell. Res.\u003c/em\u003e 51\u0026ndash;60, DOI: https://doi.org/10.1007/978-981-13-2542-06 (2019).\u003c/li\u003e\n\u003cli\u003eJana, R., Bhattacharyya, S. \u0026amp; Das, S. Epileptic seizure prediction from EEG signals using DenseNet. \u003cem\u003eProc. IEEE Symp. Series Comput. Intell. (SSCI)\u003c/em\u003e, 1\u0026ndash;6, DOI: https://doi.org/10.1109/SSCI44817.2019.9003059 (2019).\u003c/li\u003e\n\u003cli\u003eDaoud, H. \u0026amp; Bayoumi, M. A. Efficient epileptic seizure prediction based on deep learning. \u003cem\u003eIEEE Trans. Biomed. Circuits Syst.\u003c/em\u003e 13, 804\u0026ndash;813, DOI: https://doi.org/10.1109/TBCAS.2019.2900723 (2019).\u003c/li\u003e\n\u003cli\u003eSchelter, B. et al. Seizure prediction: The impact of long prediction horizons. \u003cem\u003eEpilepsy Res.\u003c/em\u003e 73, 213\u0026ndash;217, DOI: https://doi.org/10.1016/j.eplepsyres.2006.12.003 (2007).\u003c/li\u003e\n\u003cli\u003eNiederhauser, J. J. et al. Detection of seizure precursors from depth-EEG using a sign periodogram transform. \u003cem\u003eIEEE Trans. Biomed. Eng.\u003c/em\u003e 50, 449\u0026ndash;458, DOI: https://doi.org/10.1109/TBME.2003.809112 (2003).\u003c/li\u003e\n\u003cli\u003eVan Quyen, M. L. et al. Anticipating epileptic seizures in real time by a non-linear analysis of similarity between EEG recordings. \u003cem\u003eNeuroreport\u003c/em\u003e 10, 2149\u0026ndash;2155, DOI: https://doi.org/10.1097/00001756-199907050-00015 (1999).\u003c/li\u003e\n\u003cli\u003ePark, Y., Luo, L., Parhi, K. K. \u0026amp; Netoff, T. Seizure prediction with spectral power of EEG using cost-sensitive support vector machines. \u003cem\u003eEpilepsia\u003c/em\u003e 52, 1761\u0026ndash;1770, DOI: https://doi.org/10.1111/j.1528-1167.2011.03176.x (2011).\u003c/li\u003e\n\u003cli\u003eHowbert, J. J. et al. Forecasting seizures in dogs with naturally occurring epilepsy. \u003cem\u003ePloS One\u003c/em\u003e 9, e81920, DOI: https://doi.org/10.1371/journal.pone.0081920 (2014).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table 1","content":"\u003cp\u003eTable 1 is available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"inter-ictal, pre-ictal, seizure prediction, patient specific, specific warning period","lastPublishedDoi":"10.21203/rs.3.rs-5714799/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5714799/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSeizures are sudden activations of part or all of the brain, which are the primary symptoms of epilepsy. Epileptic seizures are characterized by their sudden and unpredictable nature and pose significant risks to patients\u0026rsquo; daily lives. In many patients with epilepsy, specific symptoms are observed for a short period before a seizure occurs. Research is ongoing to develop technology that can predict seizures by detecting these symptomatic signals. In this paper, we used the CHB-MIT database, which contains more than 150 seizures in approximately 1,000 hours of EEG data collected from 23 children with intractable seizures. Tunable Q-factor Wavelet Transform (TQWT), a specific type of wavelet transform, was applied to the data to predict seizures by identifying inter-ictal and pre-ictal states using a relatively simple deep learning classifier. Each classification was performed individually for each patient, and seizure prediction during a specific period was achieved using k-fold cross-validation, a technique commonly employed in deep learning. By combining the results, the period with the highest probability of seizure occurrence for each patient was designated the specific warning period. Ultimately, it was possible to predict seizures 15 minutes in advance, achieving an average sensitivity of 0.97, an F1 score of 0.90, and a false discovery rate (FDR) of 0.13 for all patients. Additionally, a specific warning period was established for each patient, ranging from a minimum of 2.5 minutes to a maximum of 15 minutes.\u003c/p\u003e \u003cp\u003e.\u003c/p\u003e","manuscriptTitle":"Patient specific seizure prediction for time intervals using TQWT and deep learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-01-08 20:07:18","doi":"10.21203/rs.3.rs-5714799/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a389cf0b-9a2c-4990-ab1b-5b06ed78be87","owner":[],"postedDate":"January 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-02-14T08:23:24+00:00","versionOfRecord":[],"versionCreatedAt":"2025-01-08 20:07:18","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5714799","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5714799","identity":"rs-5714799","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}