<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>SZLoc: A Multi-resolution Architecture for Automated Epileptic Seizure Localization from Scalp EEG</title></titleStmt>
			<publicationStmt>
				<publisher></publisher>
				<date>2022</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10400612</idno>
					<idno type="doi"></idno>
					<title level='j'>Medical Imaging with Deep Learning</title>
<idno></idno>
<biblScope unit="volume"></biblScope>
<biblScope unit="issue"></biblScope>					

					<author>Jeff Craley</author><author>Emily Johnson</author><author>Christophe C. Jouny</author><author>David Hsu</author><author>Raheel Ahmed</author><author>Archana Venkataraman</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[Arch a na Venkataraman 1]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="1.">Introduction</head><p>Epilepsy is a heterogeneous neurological disorder characterized by recurrent and unprovoked seizures <ref type="bibr">(Fisher et al., 2014)</ref> that aects roughly 1.2% of the population <ref type="bibr">(Zach and et al., 2017)</ref>. An estimated 20{40% of epilepsy patients are medically refractory <ref type="bibr">(French, 2007)</ref> and do not respond to anti-epileptic drugs. In patients with focal epilepsy, whose seizures originate a discrete Seizure Onset Zone (SOZ) <ref type="bibr">(Lu &#8226;ders et al., 2006)</ref>, subsequent therapeutics include resective surgery and neurostimulation. Here, accurate SOZ localization is a key determining factor of treatment outcome <ref type="bibr">(Rosenow and Lu &#8226;ders, 2001)</ref>. Scalp E E G monitoring is often the rst modality used in diagnosis and treatment planning for focal epilepsy. Clinical SOZ localization based on E E G relies on visual inspection, a process that is time consuming and often has low inter-rater agreement <ref type="bibr">(van Donselaar et al., 1992)</ref>.</p><p>Machine learning methods for scalp E E G have largely overlooked the challenging prob-lem of SOZ localization, instead focusing on the simpler problem of seizure detection. Traditional approaches used a standard feature extraction and classication pipeline to categorize short windows (1{30 seconds) of the E E G as seizure or baseline <ref type="bibr">(Osorio et al., 2016)</ref>. More recent works have turned to deep learning via convolutional <ref type="bibr">(Wei et al., 2019;</ref><ref type="bibr">Zou et al., 2018)</ref>, recurrent <ref type="bibr">(Vidyaratne et al., 2016;</ref><ref type="bibr">Hu et al., 2020)</ref>, and hybrid CNN-RNN architectures <ref type="bibr">(Craley et al., 2021;</ref><ref type="bibr">Aes et al., 2019;</ref><ref type="bibr">Liang et al., 2020)</ref>. While these methods report high detection accuracies, simply determining the onset and oset times of the seizure has limited clinical utility, as it fails to provide localization information for therapeutic planning.</p><p>A few prior studies have explored the problem of SOZ localization from scalp E E G . Such methods have leveraged both cross-channel and spatio-temporal information. For example, the work of <ref type="bibr">(Craley et al., 2019</ref>) uses a graphical model with related detection and localization variables to learn a spatial onset distribution for each patient. An alternative approach by <ref type="bibr">(Covert et al., 2019)</ref> relies on graph convolutional networks (GCNs); the authors demonstrate that omitting E E G channels from within the SOZ greatly reduces downstream seizure prediction, suggesting a possible localization scheme based on \explainable AI". Likewise, <ref type="bibr">(Dissanayake et al., 2021)</ref> posits that localization information may be revealed by analyzing the deep network trained for seizure detection. While promising, the latter two methods are geared towards a patient-specic analysis, rather than drawing inferences for new patients.</p><p>The strategy of mining cross-channel and spatio-temporal information have appeared in other E E G applications. Notably, G C N architectures have been applied to seizure detection, as in <ref type="bibr">(Wagh and Varatharajah, 2020)</ref> and <ref type="bibr">(Lian et al., 2020)</ref>. Transformer architectures go a step further by relating sequential and spatially distributed features. The works of <ref type="bibr">(Cisotto et al., 2020)</ref> and <ref type="bibr">(Kostas et al., 2021)</ref> investigate transformer architectures for E E G classication tasks. In <ref type="bibr">(Liu et al., 2021)</ref>, spatial and temporal attention are combined into one transformer layer and evaluated on an emotion recognition task. <ref type="bibr">(Qu et al., 2020)</ref> uses a convolutional feature extractor and apply a transformer for sleep staging. Finally, <ref type="bibr">(Sun et al., 2021)</ref> and <ref type="bibr">(Bagchi and Bathula, 2021)</ref> incorporate convolutional layers within the transformer layers for motor imagery and visual stimuli classication, respectively.</p><p>In this paper, we present the rst end-to-end framework for cross-patient SOZ localization from scalp E E G . Our novel architecture, S Z L o c , combines a convolutional, transformer, and recurrent module to eciently learn the relevant (and heterogeneous) seizure patterns from limited training examples. Crucially, we introduce an ensemble of weakly supervised loss functions to train SZLo c to produce accurate SOZ localizations from inex-act spatial and temporal onset labels. We validate the SOZ localization maps produced by SZLo c on two clinically annotated datasets. Our results demonstrate the synergy of our architecture, loss function, and training strategy for robust cross patient localization using scalp E E G . To our knowledge, this is the rst result of its kind in the epilepsy literature.  </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.">Methods</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.2.">Generating Seizure L e ve l Onset Maps</head><p>The datasets used in this work are recorded in the 10-20 international system, corresponding to 19 E E G electrodes. During training, we have access to clinical annotations of the lobe and hemisphere of seizure onset in each patient. This coarse SOZ information can be used to derive a binary localization map O 2 f0; 1g 19 , where electrodes within the annotated lobe and hemisphere are set to 1 and the remaining electrodes are set to 0. This labeling scheme is demonstrated in Figure <ref type="figure">2</ref>, where channels in the clinically provided annotation of \Left Temporal" are enclosed in a separate region, denoting O i = 1.</p><p>The onset attentions a </p><p>Notice that the numerator in Eq. (1) will be large if electrode i is manifesting seizure activity during the learned patient-level seizure onset time, as captured by the attention variable a[t]. The denominator of Eq. (1) normalizes the SOZ map, such that O i = 1. Thus, SZLo c provides interpretable labels of seizure activity used in its nal SOZ predictions.</p><p>As is standard in epilepsy monitoring, each patient may have multiple seizures. We average the predictions O across these seizures to create a patient-level SOZ map incorporating localization from each seizure. This process mirrors the SOZ information harmonization used in clinical practice, where congruent information across multiple recordings provides greater condence in the nal SOZ location. This strategy is depicted in Figure <ref type="figure">2</ref>, where two correct and one incorrect localization combine to form a correct patient level prediction. </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.3.">We a k l y Sup ervised Loss Functions</head><p>Our data contains coarse and noisy detection and localization labels, from which SZLo c must learn precise, high resolution information. To learn from these labels, we separate detection and localization into separate losses and further divide these tasks based on the signal paths used. In detection, the annotated seizure interval is on a patient level and tends to be overly generous compared to the underlying electrographic signatures. To compensate, we train SZLo c on 45 s E E G snippets centered around the annotated seizure onset. Using a cross-entropy loss, we enforce that S[t] should be zero (baseline) for the rst 15 s, one (seizure) for the last 15 s, and normalize over these 30 s. However, we do not enforce any label during the middle 15 s, allowing SZLo c to learn the appropriate onset time. Formally,</p><p>where S may take the superscript g or e denoting global or electrode-level information.</p><p>On the localization front, only region-level (lobe and hemisphere) information is provided. While the SOZ lies somewhere within this region, notice that not all electrodes will necessarily manifest seizure activity at the onset. In addition seizure activity may appear outside the onset zone due to propagation eects. To accommodate these factors, we construct three complementary loss functions. Mathematically, let P (O ) be the set of electrodes within the clinician annotated region and let O , O= max i O i be SOZ predictions normalized such that the maximum is 1. To reward correct onset predictions in P (O ) and penalize those outside of it, we apply a square loss to positive and negative SOZ regions.</p><p>We additionally maximize the margin between regions, reecting our evaluation strategy based on the region containing the maximum predicted electrode location. We train SZLo c with global and aggregated electrode detection and localization losses. We apply a scaling factor to Eqs. (3-4), such that the loss terms lie in [0; 1], ensuring (5)</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.4.">Validation Strategy</head><p>We train the SZLo c architecture using leave-one-patient-out cross validation (LOPO-CV).</p><p>Using this strategy, we are able to quantify generalization ability to new patients. SZLo c and competing baselines are evaluated for ve random initializations and results are then averaged across these test runs. We evaluation localization accuracy based on the maximum value in the predicted onset map O. If the maximum predicted weight arg max i O is within the labeled SOZ for a seizure or patient P (O), the localization is considered correct.</p><p>Evaluating Onset Att ent i on We validate our multi-signal path framework by evaluating our architectures on subsets of weakly supervised loss functions. By setting loss factors for O e and O g to 0, we evaluate the performance of the models using electrode and global, respectively, onset attention individually. In addition, we train the same architecture using an ' 2 reconstruction loss between O and O, rather than the loss terms in Eqs. (3-4). Thus, we quantify the improvement due to each source of information.</p><p>Baseline Mo dels Similar to the loss ablations, we evaluate SZLo c against ablated and reordered versions of the architecture. Hidden representations in each layer are of length 160, allowing modications without the need to adjust network dimensions. The C G T baseline reorders the layers such that the G RU follows the CNN and precedes the transformer layer. SZLoc-No Connect and C G T No-Connect omit global features in the transformer input. In C G , the transformer layer is omitted, while C T omits the GRUs. SZLoc-Final State predicts SOZ using the nal state of the GRU, rather than based on the attentions. Finally, the T G C N uses Architecture I I of <ref type="bibr">(Covert et al., 2019)</ref> modied to generate electrode level predictions. The SZLoc-Final State and T G C N models are trained using only one set of localization losses, as these models contain only one source of SOZ maps.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.">Results</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.1.">C l i n i c a l E E G Datasets</head><p>Datasets of adult patients from the Johns Hopkins Hospital (JHH) and pediatric patients from the University of Wisconsin-Madison (UWM) were used in this work. Results are presented as the total number of correctly localized patients or seizures averaged over all 5 seeds. The JHH dataset contains 201 seizures from 34 adult patients with focal epilepsy and is used in our main L OP O -CV evaluation. JHH models are evaluated for generalization in the UWM dataset, consisting of 101 seizures across 16 pediatric patients. Due to dierences between pediatric and adult E E G , we expect JHH model performance to degrade in the UWM dataset. Each 45 s seizure recording contains 15 s of pre-seizure, an onset period from 15 s to 30 s, and 15 s of post-onset seizure. E E G recordings are ltered from 0.5{30.0 Hz and normalized to mean 0 with S T D 1, and clipped at 2 STDs. 1 s windows with no overlap were used with data augmentation as described in the appendix in Section A.5. </p><p>A l l Losses Patient Seizure 6:4 1:6 37:8 4:4 6:1 1:6 32:1 4:5 6:5 1:8 37:5 5:0 5:8 1:7 32:7 5:1 5:9 1:7 32:7 4:8 6:8 1:5</p><p>38:1 4:3 5:1 1:7<ref type="foot">foot_1</ref> 31:8 4:3 1 5:8 1:0 1 37:3 4:9 1</p><p>Electrode Losses Patient Seizure 6:7 1:6 39:3 4:9 5:7 1:7 34:6 5:0 6:7 1:6 39:7 5:0 5:4 1:8 33:6 4:7 5:3 1:7 33:6 4:8 6:9 1:3 37:</p><p>Seizure 6:6 1:3 35:4 5:1 5:0 1:5 32:8 4:5 6:7 1:4 37:0 5:1 4:9 1:4 31:9 5:6 5:2 1:5 31:9 4:6 5:9 1:5 32:8 5:1 4:6 1:6 29:2 6:3 5:1 1:0 35:6 4:0</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.2.">Lo calization Results</head><p>Localization in the JHH dataset is shown in Table <ref type="table">1</ref> using electrode onset attention a e [t].</p><p>Performance with global attention is reported in Table <ref type="table">4</ref> in the appendix. Across the board, aggregated electrode attention outperforms global attention. SZLo c outperforms all competing baselines at the patient level, achieving its best performance of 24:2 1:0 of 34 (71.1%), though this result does not rise to the level of statistical signicance due the size of the dataset. For comparison, random chance prediction would result in a patient-level accuracy of 9.6 (28.2%). T G C N achieves a higher seizure accuracy of 118.0 (58.7%) but underperforms at the patient level. Competing baselines show similar performance with no global attention applied during training, suggesting that only SZLo c is able to leverage information from both onset attention sources eectively in the multi-signal path loss paradigm. All models show performance degradation when trained with the ' 2 reconstruction loss, demonstrating the ecacy of our ensemble of weakly supervised loss functions.</p><p>Table <ref type="table">2</ref> shows generalization in the UWM dataset using JHH trained models. While performance is lower, all models show localization ecacy as random chance assignment would result in 4.2 (26.2%) and 25.2 (24.9%) at the patient and seizure level, respectively. Similar trends are evident, with a statistically signicant SZLo c result over the T G C N . Transformer models generalize well, underscoring the benets of the multi-scale architecture. Figure <ref type="figure">3</ref> shows patient SOZ localization maps for the best performing random initialization of SZLo c. SZLo c correctly identied the SOZ in 26 out of 34 patients. Secondary modes are observed when localizations for a subset of seizures diverge from the majority. This behavior conforms with clinical expectations that some seizure presentations contain better localization information than others, highlighting the benets of patient level aggregation. We note that in most missed patients, either the mode is adjacent to the annotated region, or a secondary mode occurs within the annotated region. This demonstrates that SZLo c provides valuable clinical information even in cases where the prediction is incorrect.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="4.">Conclusion</head><p>In this paper we present SZLo c for automated epileptic seizure localization. Fusing information across multi-scale signal paths, SZLo c combines global and electrode-level spatiotemporal seizure activity leading to robust SOZ localization. An ensemble of weakly supervised loss functions allows SZLo c to compensate for coarse and noisy training labels. Namely, desirable properties of the SOZ are balanced to generate spatial maps at the seizure and patient level. By leveraging the information rich output of the SZLo c model, E E G activity contributing to the nal localization can be visualized in the original signal space. We also demonstrate that SZLo c is capable of cross-patient localization, a relevant clinical use case. Taken together, SZLo c provides clinically useful information at multiple scales to aid in the localization of focal epileptic seizures for therapeutic planning.    <ref type="figure">4</ref> (a). We use two separate CNN architectures for the global and electrode signal paths. Each CNN consists of an embedding layer, followed by a cascade of residual blocks and convolutional projection layers. Each residual block rst applies a convolution layer followed by batchnorm and a P R e L U activation. Another application of convolution and batchnorm follow, with a residual connection between the input of the block prior to the nal P R e L U activation.</p><p>Convolutions within the block use the same number of channels as the block's input and kernels with a length 7 and a stride of 1. The projection layers between convolutional blocks double the number of channels while halving the length of the representation; they employ a kernel size of 3 and a stride of 2 unless otherwise noted. We apply global average pooling at the end of the CNN to reduce the time-series channels to a length 160 feature vector.</p><p>The global CNN starts with an embedding convolution on the 19-channel E E G signal with kernel size 7 and an output of 80 channels. Next, a residual convolution block with 80 channels is applied, followed by a projection layer to increase the number of channels to 160, followed by another convolution block. This representation is then passed through a convolution layer with 160 channels of input and output, kernel size 1, and a stride of 2 to reduce the (temporal) length of the representation. A nal residual block is applied before global average pooling. We use dropout on the nal feature representation during training.</p><p>The electrode CNN uses a similar architecture but increases the network depth. We train the same CNN across E E G channels to ensure consistency in the nal feature representation for subsequent multi-channel analyses. The individual electrode signal is passed through an embedding convolution with kernel size 7 and 20 output channels. Next 3 residual blocks and projection layers are applied, increasing the representation from 20 to 40, 80, and 160, while halving the length of the representation after each projection. A nal residual block is applied with 160 features before global average pooling and dropout are applied.</p><p>By adopting an encoder-decoder transformer structure, dierent sets of feature representations can be considered within each transformer component. In the encoder, global and electrode level CNN representations are used as input features while only electrode features are included in the decoder. This structure allows features from the global CNN and from each individual electrode to contribute to the nal representation for each electrode. The decoder and encoder layers of the transformer follow <ref type="bibr">(Vaswani et al., 2017)</ref>, however only one layer is used in both the encoder and decoder. In preliminary experimentation, increased transformer depth led to poorer generalization, likely due to overtting. The feedforward dimension is set to 256 and dropout is applied. The 19 length 160 representations from from the electrode CNN and 160 dimensional representation from the global CNN are fed into the encoder of the transformer. The 19 electrode representations are input into the decoder, resulting in a length 160 feature representation of each electrode channel. Thus each electrode channel representation incorporates information from the other channels as well as the global E E G signal for eective multi-scale information fusion.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>A.2 . Global G R U</head><p>In the G RU cell, reset and update gates, r t and z t , are computed based on the previous value of the hidden state, h (t 1) , and the input x t . The sigmoid nonlinearity ensures that these gate values range between 0 and 1. The gates control how the G RU weighs (1) incoming information from the previous hidden state, and (2) the value of the new observed datapoint. An update to the hidden state n t is computed based on the input data x t and the previous state h (t 1) , as weighted by the value of the reset gate. For example, if the reset gate is near zero, only information from the the new datapoint will be considered in the update. Finally a linear combination of the previous hidden state and the update n t ^P T ^&#348; Z L o c : M u lt i -r e s o l u t i o n Seizure L o c a l i z at i o n f r o m S c a l p E E G are computed using the update gate z t . Mathematically, the G RU is governed by r t = W ir x t + b ir + W hr h (t 1) + b hr z t = W iz x t + b iz + W hz h (t 1) + b hz</p><p>At a high level, the G RU unit weighs incoming data according to the previous values of the hidden state to produce a new hidden representation.</p><p>The bidirectional G RU in SZLo c uses 2 layers with a hidden size of 80. Thus the combined output from the forward and backward directions is 160 dimensional. To generate predictions S g [t], the length 160 representations for each time window are fed into a linear layer with an output of 2 followed by a softmax. Global onset attention scores a g [t] are generated using a linear layer which reduces the 160 length representation to a scalar at each time window. The softmax operation is applied across time windows so that t = 1 a g [t] = 1.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>A.3 . Electro de G R U</head><p>The electrode G RU cells follow the same formulation as the Global G RU cells given in the previous section. Following the global signal path, the electrode signal path is fed into a 2 layer bidirectional G RU with a hidden size of 80. The 160 dimensional output is passed through a linear classication layer and a softmax is applied to generate electrode level seizure activity predictions Y i <ref type="bibr">[t]</ref>. Similar to the electrode CNN, the same G RU and linear layers are trained across channels to eciently leverage multi-channel information. We construct the T G C N baseline according to architecture I I of <ref type="bibr">(Covert et al., 2019)</ref>. In this network, a graph of connectivity is constructed over the E E G electrodes based on local proximity. Using this graph, a series of spatio-temporal convolutional ( S T C ) layers are applied. In each S T C layer, a graph convolution is rst applied across E E G electrodes to share features between neighboring electrodes. Second, a 1D convolution is applied temporally as in a temporal CNN <ref type="bibr">(Lea et al., 2016)</ref>. A cascade of 8 S T C layers is applied. In the original formulation, features are pooled across E E G channels for seizure detection (i.e., a single prediction for each time window). Thus, we modify the architecture for application to seizure localization. Instead of pooling features, we append a linear classication layer to the nal nodes at every point in time to generate predictions Y i [t] as in SZLo c. E E G channel, let M i [t] Bernoulli(0:5) be a random variable indicating whether or not Gaussian noise will be added. Similarly, the degree of additive noise for each window is sampled from a Gaussian distribution V i [t] N (0; 0:1) with variance 0.1. Mathematically, noise augmentation G ( ) applied in channel i at time t can be written as</p><p>where each entry of 2 R 200 is generated from a standard normal distribution. This data augmentation technique applies noise to each channel and window individually, thus prompting robustness to changes in noise conditions across channels and time.</p><p>In addition to additive noise, signal time reversal and cross-hemispheric signal ipping is applied. For signal time reversal, the E E G signal in each window is reversed for all channels simultaneously. However, the order of the windows remains consistent, ensuring that seizure onset and propagation information is preserved. Cross-hemispheric ipping interchanges left and right channels, creating a mirror image of the original E E G signal and onset labels. This ipping ensures that the global signal path is invariant to lateralized eects. The electrode signal path remains unaected as no sense of spatial position is preserved inherently in this part of the network. In contrast to the window based additive noise augmentation, to preserve important phase and spatial relationships between channels, both of these techniques are applied for all electrodes and time windows for a given sequence. Each augmentation technique is applied with probability 0.5 to each sequence.</p><p>A.6 . Tr a i n i n g Details SZLo c is implemented in Pytorch 1.9. We train the architecture for 100 epochs using Adam (Kingma and Ba, 2014), a batch size of 5, and a learning rate of 0.001. Weight decay is set to 0.001 and dropout in the CNN, transformer, and G RU are set to 0.1. </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>Appendix B . Supplemental Results and Discussion</head><p>When comparing SZLo c to the ablated baselines in Table <ref type="table">1</ref>, we observe SZLo c outperforms all of them. Interestingly, the C T baseline, which contains no recurrent element, performs best among ablated baselines with a patient level electrode score of 22:0 0:7. This result suggests that while the G RU enhances performance in the overall SZLo c architecture, when elements of the network are removed, then the G RU may overt to the training data. This hypothesis is reinforced by the fact that the C G T baseline performs the worst of all baselines with a patient electrode performance of 15:6 2:6. By placing the G RU before the transformer, it is likely that the model can t to the data more directly than when the transformer occurs rst in the signal path. Interestingly, when the connection between the global and the electrode signal paths are severed (SZLoc-No Connect), the performance degrades. Finally, the the CGT-No Connect performance increases when multi-scale information is not considered in the transformer, which further supports our hypothesis that placing the G RU before the transformer increases the likelihood of overtting.</p><p>Figure <ref type="figure">6</ref> shows the localization maps O g generated using the global attention variable for the same model as pictured in Figure <ref type="figure">3</ref>. Patients are presented in the same order to allow for easy comparison between the gures. While the overall localization behavior is similar for both attention variables, we note that two patients with correct channel localizations are incorrectly localized in Figure <ref type="figure">6</ref>. As noted in the main text, we hypothesize that the decrease in performance can be attributed to the tendency of the global attention variable to remain constant in the period before the onset of a seizure.</p><p>Figures <ref type="figure">7</ref> and<ref type="figure">8</ref> show localization maps from the two next best performing baselines, the C T and T G C N , respectively. As in Figure <ref type="figure">6</ref>, the patients are preseted in the same order as Figure <ref type="figure">3</ref> for easy comparison. In its best fold performance, the C T baseline correctly identies 23 of 34 localization zones. While the quantitative performance of the C T model is decent, the localization maps from this model are notably more diuse, possibly due to the lack of sequence modeling. As shown in Figure <ref type="figure">8</ref>, the T G C N correctly localizes 27 of 34 patients in its best performance, exceeding the maximum of SZLo c. However, the T G C N baseline shows a strong preference for temporal and central electrodes T7, C3, C4, and T8. The local connectivities encoded within the structured graph layers of the T G C N encourage the SOZ predictions within these electrodes, as they are common to nearly all SOZ annotations. Eectively, this allows the model to reduce the SOZ localization problem to the simpler problem of lateralization. In summary, while the T G C N baseline achieves the best quantitative performance of any of the competing baselines with 23:2 2:4 (68.2%) at the patient level and 118:0 9:2 (58.7%), it is apparent that the T G C N does this by overtting &#348;   We observe that SZLo c learns to maintain a constant value for a g before fading to zero during the seizure while a e is more sharply positioned near the onset, indicating a potential explanation for electrode onset outperforming global onset in localization metrics.  </p></div><note xmlns="http://www.tei-c.org/ns/1.0" place="foot" xml:id="foot_0"><p>&#169; 2022 J .Craley et al.   </p></note>
			<note xmlns="http://www.tei-c.org/ns/1.0" place="foot" n="1" xml:id="foot_1"><p>Baseline models with one source of attention are trained with only one set of localization loss functions.</p></note>
		</body>
		</text>
</TEI>
