PaperlayerView exact revision on arXiv

Source-linked AI summary

Non-invasive Seizure Detection Using Wearable Wrist-worn Accelerometry and Deep Learning

Nilushika Udayangani Hewa Dehigahawattage, Kishor Nandakishor, Marimuthu Palaniswami

arXiv:2610.05919v1cs.LGcs.AI

TL;DR

Continuous seizure monitoring is needed, but existing monitoring approaches can be resource-intensive, limited in availability, or intrusive. This paper evaluates a single wrist-worn accelerometer with a deep neural network, detecting 44 of 46 convulsive seizures at 95.65% sensitivity and an overall false alarm rate of 0.15/24 h.

Abstract

from arXiv · show

Seizure monitoring and detection are crucial for reducing the morbidity and mortality associated with seizures. Current epilepsy care, often involving expensive video-electroencephalography (VEEG) monitoring, requires specialized expertise and is limited to in-hospital settings, and intrusive in nature. Seizure diaries, on the other hand, suffer from unreliability due to under-reporting, leading to incorrect therapeutic decisions. Wearable non-invasive seizure detection may offer a more tolerable and feasible solution for long-term ambulatory monitoring. This study explores a wearable remote monitoring system utilizing a single wrist-worn accelerometer device and capable of detecting multiple types of seizures, including shorter duration events. We enrolled 79 patients under video-electroencephalography monitoring to wear accelerometer devices and collect data. Concurrent VEEG recordings were reviewed by board-certified epileptologists to produce annotations, including seizure onset, offset, and seizure type. Using this data, we constructed a deep neural network based on the time-series ResNet architecture, which could discriminate among seizure and non-seizure events. Our proposed approach achieved a seizure detection sensitivity of 95.65% and an overall false alarm rate of 0.15/24 hours during the evaluation, which spanned 5576 hours of total recording. Additionally, it resulted in an area under the receiver operating characteristic curve (AUC-ROC) of 0.98 and an area under the precision-re call curve (AUC-PRC) of 0.67 when averaged over 20 patients who experienced 46 convulsive seizures. These promising results suggest that the proposed seizure detection system can be effectively used for long-term ambulatory seizure monitoring. Future steps include validating our findings in larger datasets and assessing the utility of detection for additional seizure types.

I. INTRODUCTION

Seizures create serious health and quality-of-life risks, motivating continuous monitoring and early detection. Existing monitoring and seizure diaries have practical and reliability limitations, while non-invasive wearable biosensors are presented as a potential alternative.

A. Background E

  • Seizure-related risks include injury, status epilepticus, and SUDEP, supporting the need for early detection and continuous monitoring.The paper notes that prompt recognition may help reduce injuries or fatalities and support clinical management.Source: A. Background · p. 1

B. Limitations in current seizure monitoring

  • VEM and iEEG are resource-intensive and typically limited to tertiary hospitals, while ambulatory EEG is also described as obtrusive for long-term monitoring.Seizure diaries are unreliable: fewer than half of seizures are accurately documented, and under-reporting can hinder diagnostic and therapeutic decisions.Source: B. Limitations in current seizure… · pp. 1–2

C. Non-invasive seizure monitoring using wearable devices

  • Non-invasive wearable biosensors paired with machine learning are presented as having potential to address barriers to long-term seizure monitoring.The passage describes recent wearables that acquire signals including ECG, EDA, sEMG, accelerometry, PPG, skin temperature, and oxygen saturation.Source: C. Non-invasive seizure monitoring… · p. 2

1) seizure detection modalities:

Wearable seizure-detection modalities trade off comfort, signal characteristics, and seizure-type coverage. The discussion highlights accelerometry’s wrist-worn practicality while noting the limitations of single modalities and the burdens of combining sensors.

1) seizure detection modalities:

  • ECG-derived HRV methods can detect motor and non-motor seizures, but patient-specific tuning and variation in cardiac responses challenge generalization.One cited study reported 87% sensitivity among responders and 28% among non-responders.Source: 1) seizure detection modalities · p. 2
  • PPG can be affected by motion artifacts, sEMG placement and attachment can constrain long-term use, and EDA has delayed responses and is often used multimodally.The passages also describe low sensitivity and higher false alarm rates for PPG alone, and discomfort or obtrusiveness for sEMG and EDA.Source: 1) seizure detection modalities · p. 2
  • Wrist-worn accelerometry measures movement in three dimensions and is described as comfortable, familiar, cost-effective, and low-power for long-term monitoring.The passage reports acceptance above 70% for smartphone- and watch-based devices, compared with below 50% for devices placed on the head, leg, upper arm, or chest.Source: 1) seizure detection modalities · p. 2
  • Accelerometry and sEMG are better suited to motor seizures, while ECG and PPG can cover non-motor seizures; multi-sensor systems add practical burdens.Those burdens include portability, data processing, power consumption, signal synchronization, commercialization, and clinical implementation.Source: 1) seizure detection modalities · pp. 2–3

2) seizure detection algorithms:

The section contrasts thresholding, traditional machine learning, and deep learning, alongside personalized and generalized detection strategies. It emphasizes class imbalance and the challenge of generalizing across patients and seizure durations.

2) seizure detection algorithms:

  • Threshold-based algorithms use selected signal thresholds but can be noise-sensitive, generalize poorly to unseen data, and adapt weakly to changing patient signals.Examples of threshold inputs include heart-rate rise, zero crossings, and z-scores.Source: 2) seizure detection algorithms · p. 3
  • Traditional machine-learning methods require hand-crafted features that may miss complex nonlinear patterns and depend on the study dataset and targeted seizure types.Feature engineering can include extracting time- or frequency-domain characteristics from preprocessed raw data.Source: 2) seizure detection algorithms · p. 3
  • Deep learning supports end-to-end feature learning, but reported results vary and many existing networks emphasize multimodal signals rather than single-modal multivariate analysis.The passage describes further exploration of deep learning for single-modal signals as warranted.Source: 2) seizure detection algorithms · p. 3
  • Personalized algorithms can report higher sensitivity and lower false alarm rates, but require patient-specific data and tuning; generalized models avoid that cost while facing patient variability.The passage identifies learning features not influenced by individual patient characteristics as key to generalization.Source: 2) seizure detection algorithms · p. 3
  • Seizures are rare in long-term monitoring, and imbalance worsens for shorter targeted durations; some prior imbalance-aware methods are unsupervised.The authors note that real-world imbalance can exceed training imbalance and that supervised learning could help control sensitivity-specificity balance.Source: 2) seizure detection algorithms · pp. 3–4

3) limitations in current ACM-based seizure detection and monitoring:

The section describes limitations of accelerometry-based detection and the study’s data collection and cohort. Its dataset comprises VEM recordings from 79 patients, with the analysis focused on convulsive events across three seizure types.

3) limitations in current ACM-based seizure detection and monitoring:

  • Accelerometry focuses on motor behavior, so everyday rhythmic movements can mimic convulsions, while many existing algorithms focus on GTCS and ignore focal seizures.The passage notes that algorithms failing to capture these distinctions may produce false alarms and that false alarm rates matter to wearable acceptance.Source: 3) limitations in current ACM-based… · p. 4

A. Data Acquisition

  • Apple iPod Touch devices with built-in accelerometers were strapped to patients’ forearms and recorded 3D motion at 50 Hz during VEM.Recordings ran in 12-hour sessions, with devices replaced to maintain coverage; patients were encouraged to continue usual activities.Source: A. Data Acquisition · p. 4

B. Recorded Dataset

  • The dataset includes 79 patients and 5,576 recording hours; the study focuses on 46 convulsive seizures across GTCS, PNES, and FIAS.The convulsive-seizure cohort comprised 20 patients; the paper also groups patients by combinations of these seizure types and notes that observed counts decrease as minimum duration increases.Source: B. Recorded Dataset · pp. 4–5

C. Pre-Processing

Preprocessing converts three-axis accelerometry into fixed-length events by filtering and merging active signal windows. The resulting dataset contains 8,362 events, including 46 seizure events and 8,316 non-seizure events.

C. Pre-Processing

  • 10-second windows with 50% overlap were used, and activity filtering thresholded the signal at 0.2 g, a value based on the lower bound of collected seizure data.The window size was selected based on the 10-second minimum seizure duration considered in the study.Source: C. Pre-Processing · p. 5
  • A sixth-order Butterworth bandpass filter used 2–25 Hz cutoffs to filter spurious spikes, based on prior observations of typical seizure-event frequencies.The resultant signal was derived from the three-dimensional accelerometer data.Source: C. Pre-Processing · p. 5
  • 8,362 events were formed from 79 patients: 46 seizure events and 8,316 non-seizure events.Preprocessing also reduced data volume, computational complexity, and class imbalance to some extent.Source: C. Pre-Processing · p. 5
  • Windows with more than 50% activity were treated as active, and active windows separated by less than 20 seconds were merged into events.Events shorter than 10 seconds were then removed, and each remaining event was resized to 500 samples using Fourier upsampling or downsampling.Source: C. Pre-Processing · p. 5

D. Time Series Residual Network

The seizure classifier uses a time-series ResNet whose shortcut connections and residual blocks address vanishing gradients and network degradation. Its implementation stacks convolutional blocks in three residual blocks, followed by global average pooling and a two-class softmax.

D. Time Series Residual Network

  • Shortcut connections in ResNet allow gradients to flow directly through bottom layers and transmit inputs from low to high levels, addressing vanishing gradients and network degradation [34].The network follows the Residual Network architecture proposed in [78].Source: D. Time Series Residual Network · pp. 5–6
  • The implemented ResNet has three residual blocks, each adding the output of three stacked convolutional blocks to its input before passing the result onward.Each convolutional block consists of convolution, batch normalization, and ReLU; pooling layers were omitted to prevent overfitting.Source: D. Time Series Residual Network · p. 6
  • The network ends with global average pooling and a softmax layer for seizure versus non-seizure classification; global average pooling reduces the number of weights compared with a fully connected layer.The first residual block uses 64 filters per convolution, while later residual blocks use 128.Source: D. Time Series Residual Network · p. 6

E. Training DNN

Training addresses class imbalance with weighted binary cross-entropy and uses a one-cycle learning-rate policy. The maximum learning rate is selected through a range test, with early stopping used to limit overfitting.

E. Training DNN

  • Weighted binary cross-entropy uses positive-class weight β, tuned to 3, to adjust the balance between true-positive and true-negative rewards under class imbalance.The loss is intended to discourage the model from favoring negative predictions when negative samples substantially outnumber positive samples.Source: E. Training DNN · p. 6
  • The 1cycle policy raises the learning rate from low to high and then lowers it below its initial value, using one cyclical learning-rate cycle during training.The authors adopted it for a shallow ResNet trained with limited labeled examples.Source: E. Training DNN · pp. 6–7
  • The maximum learning rate was selected using an LR range test, and early stopping used validation loss with a patience of 3 epochs.The range test increased learning rate after each mini-batch until loss began to explode; the selected value was between that point and the steepest-slope point.Source: E. Training DNN · p. 7

F. Datasets Formation and Experimental Setup

The evaluation uses leave-one-patient-out cross-validation among patients who experienced seizures, while patients without seizures form a separate population for assessing false alarm rate. Patient data are kept separate across training, validation, and testing.

F. Datasets Formation and Experimental Setup

  • Leave-one-patient-out models train on seizure-experiencing patients other than the held-out patient, with group sizes of 12, 14, 18, and 20 for PG1–PG4, respectively.The held-out patient supplies the test data for each model.Source: F. Datasets Formation and Experimental… · p. 7
  • Within each LOPO training set, two-thirds of the data are allocated to learning and one-third to validation, without shuffling data across patients.Each model is trained in ten randomizations, and the best-performing model is selected.Source: F. Datasets Formation and Experimental… · p. 7
  • Four additional models, one per patient group, train on seizure-experiencing patients and test on patients without seizures to assess false alarm rate.The training data are split into learning and validation sets using the same setup, and ten randomizations are run for each model.Source: F. Datasets Formation and Experimental… · p. 7

G. Performance Metrics and Evaluation

Models were selected using validation AUC-PRC and evaluated on held-out patients and non-seizure cohorts. Across the four patient groups, PG4 detected 44 of 46 seizures with an overall FAR of 0.15/24h.

G. Performance Metrics and Evaluation

  • The best of ten training randomizations was selected by maximum validation AUC-PRC because AUC-ROC was less discriminative for these imbalanced data.AUC-PRC incorporates precision and recall, whereas AUC-ROC uses false positive rate.Source: G. Performance Metrics and Evaluation · p. 7
  • The evaluation recorded patient-specific and group-level sensitivity, specificity, F1 score, PPV, FAR/24h, AUC-ROC, and AUC-PRC.Overall FARs were aggregated across seizure and non-seizure patient cohorts and averaged over 5576 recording hours.Source: G. Performance Metrics and Evaluation · p. 7

III. RESULTS

  • PG1 correctly classified all GTCS events (100%) and had an overall FAR of 0.06/24h across 79 patients.Its mean FAR was 0.34/24h among 12 patients.Source: §III · pp. 7–8, Table III · p. 8
  • PG2 detected 25 (96.15%) of 26 seizures and had an overall FAR of 0.05/24h across 79 patients.One FIAS event went undetected; the mean FAR among 14 seizure patients was 0.2/24h.Source: §III · p. 8, Table IV · p. 8
  • In PG3, 37 (90.24%) of 41 seizures were detected, and the overall FAR across 79 patients was 0.14/24h.Four seizures were missed; the 18 seizure patients had a mean FAR of 0.54/24h.Source: §III · p. 8, Table V · p. 8
  • 44 (95.65%) of 46 seizures were detected in PG4, with an overall FAR of 0.15/24h across 79 patients and 5576 recording hours.The detected events spanned GTCS, FIAS, and PNES.Source: §III · p. 8, Table VI · p. 8

IV. DISCUSSION

The discussion evaluates a single-accelerometer model across seizure types, including brief events, and reports strong PG4 performance. Results vary by seizure type and patient, with limited FIAS representation constraining firm conclusions.

IV. DISCUSSION

  • AUC 0.98 and AUC-PRC 0.67 were achieved for PG4, which includes all seizure types.The discussion also notes the system detects seizures lasting ≥10 seconds.Source: §IV · p. 9

A. Performance of the Proposed Method

  • All GTCS events were detected in PG1; eight of 12 patients had AUC and AUC-PRC values of 1.0 with no false alarms.The missed GTCS event for P4 had dispersed rather than concentrated spikes and lacked clearly discernible clonic features.Source: A. Performance of the Proposed Method · p. 9
  • PG2 missed one FIAS event in P12, while all FIAS events in the two FIAS-only patients, P19 and P20, were classified with zero false alarms.For P20, only two FIAS training samples were available in the LOPO model.Source: A. Performance of the Proposed Method · p. 9
  • PG3 missed four seizures, three of them PNES; overall detection rates were 85% for PNES and 95% for GTCS.Ten of 14 patients had AUC-PRC and AUC-ROC values of 1.0 with no false alarms.Source: A. Performance of the Proposed Method · p. 9
  • PG4 missed two seizures, one GTCS and one FIAS; PNES-only patients contributed 85% of false alarms.The discussion reports improved FIAS-versus-normal discrimination when PNES was included in training.Source: A. Performance of the Proposed Method · p. 10
  • Only five FIAS events were represented, so the authors caution that firm inferences about FIAS cannot be drawn.They include FIAS as an important component of a more generalized epileptic cohort.Source: A. Performance of the Proposed Method · p. 10

B. Comparison with Existing Litterateur

The discussion compares the proposed ResNet with a same-dataset baseline and single-sensor systems, while noting that study conditions and data sources affect how directly results can be compared.

B. Comparison with Existing Litterateur

  • Sensitivity improved by nearly 9% in PG2 and PG4 versus the baseline; FAR was reduced by 50% in PG1, 16% in PG2, 34% in PG3, and 30% in PG4.The baseline used the same dataset; the table summarizes results for all patient groups.Source: B. Comparison with Existing Litterateur · p. 10
  • The literature comparison fixed seizure type, minimum duration, sensor count, and modality one factor at a time to support a more justifiable comparison.Source: B. Comparison with Existing Litterateur · p. 10
  • With one sensor fixed, PG1 performed better in sensitivity and FAR than the systems of Kusmakar et al., Beniczky et al., Lockman et al., Kramer et al., Beniczky et al. 2018, and Jeppesen et al..The comparison holds sensor count constant.Source: B. Comparison with Existing Litterateur · p. 10
  • The system detected shorter-duration seizures than Kramer et al.; Conradsen et al. reported 100% sensitivity but a 0.96/24h FAR and a 50-second minimum duration.The authors note uncertainty about whether Conradsen et al. would retain similar sensitivity on shorter seizures.Source: B. Comparison with Existing Litterateur · p. 10
  • Szabó et al. reported lower FAR but also lower sensitivity, with about one-fourth the recording duration; Motahar et al. used synthetic seizure data.The authors raise generalizability concerns for the synthetic-data evaluation.Source: B. Comparison with Existing Litterateur · p. 10
  • Agrahri et al. reported 95.65% sensitivity and 14.80% FAR, but their recording duration was unavailable for a direct FAR comparison.Their 451 false alarms contrasted with 13 generated by the proposed model.Source: B. Comparison with Existing Litterateur · p. 10

V. CONCLUSION

The study concludes that a single wrist-worn accelerometer with a time-series ResNet can detect multiple seizure types, including short events. In VEM recordings, it detected 44 of 46 convulsive seizures with 95.65% sensitivity and a 0.15/24 h FAR.

V. CONCLUSION

  • 44 of 46 convulsive seizures were detected, with 95.65% sensitivity and an overall FAR of 0.15/24 h.The system was validated on accelerometer data from patients undergoing VEM.Source: §V · p. 11
  • The proposed remote system uses a single wrist-worn accelerometer to detect three major seizure types with durations ≥10 s.The authors describe it as a generalized model applicable to patients without personalized data collection.Source: §V · p. 11
  • The time-series ResNet autonomously extracts features without manual feature engineering; preprocessing and weighted loss address class imbalance.Source: §V · p. 11
  • The authors say the approach could accommodate other seizure types as relevant training data becomes available.Source: §V · p. 11

Related papers

Loading 2610.05919v1…