Seizure Detection Algorithm Using Heart Rate Variability
Overview
This study aimed to validate a novel seizure detection algorithm, based on heart rate variability (HRV) and logistic regression machine learning (LRML), in a Brazilian patient cohort undergoing long-term video-electroencephalography (LTM) monitoring. The LRML algorithm had shown promise in a Danish patient cohort, and this research sought to assess its generalizability.
A retrospective analysis was conducted on ECG data epochs from LTM recordings, focusing on 107 seizures (79 focal, 28 generalized tonic-clonic [GTC]) from 34 patients, totaling 185.5 hours of recording. Responders, defined as patients with >50 beats per minute change in heart rate during seizures, were selected for analysis. The LRML algorithm was applied to the ECG data, and results were analyzed separately for responders and non-responders.
In the responder group (22 patients, 59 seizures), the patient-adaptive LRML seizure detection algorithm demonstrated a sensitivity of 84.8% (95% CI: 75.6–93.9) with a false alarm rate of 0.25 per 24 hours. Notably, 96.2% of GTC seizures and 75.8% of focal seizures without bilateral convulsions were detected.
These findings indicate the potential efficacy of the LRML algorithm in detecting focal epileptic seizures, particularly in patients exhibiting marked ictal autonomic changes. The algorithm’s performance in this Brazilian patient cohort suggests its promising generalizability beyond the initial Danish study, offering a valuable tool for objective seizure quantification and optimization of epilepsy treatment. Further research and validation in diverse patient populations may enhance the algorithm’s utility and reliability in clinical practice.
The research affirms the effectiveness of a seizure detection method in a novel, external dataset, indicating its robustness and potential for widespread application. Notably, the method demonstrates proficiency in identifying both generalized and focal epileptic seizures, showcasing its versatility. With the capability to be integrated into wearable seizure detection systems, this algorithm holds promise in providing timely alerts to patients and caregivers, thereby facilitating prompt intervention during seizure episodes. Furthermore, its ability to generate objective seizure counts presents an opportunity to enhance patient care by optimizing treatment strategies based on accurate seizure monitoring data. Overall, these findings underscore the significance of this method in improving the management of epilepsy and enhancing the quality of life for affected individuals.
Introduction
Seizure management in patients with uncontrolled seizures often relies on seizure diaries, but these can be unreliable due to difficulties in recognizing and reporting seizures accurately. This leads to suboptimal treatment adjustments and affects patient quality of life. Wearable seizure detection devices offer a promising solution by providing objective seizure counts and alerts to patients and caregivers. However, detecting focal seizures without convulsions remains challenging with existing wearable devices. The International League Against Epilepsy has emphasized the need for further research in this area.
Previous studies have shown that heart rate variability (HRV) derived from wearable ECG devices can reliably detect focal seizures in patients with significant autonomic changes during seizures. However, there is a demand for lower false alarm rates. A patient adaptive logistic regression machine learning (LRML) method has been developed to lower false alarm rates while maintaining detection sensitivity. This method adapts the seizure detection threshold to individual patients, reducing the need for manual threshold setting and prerecording.
In this study, the generalizability of the LRML-seizure detection algorithm is assessed using an independent dataset from Brazilian patients. Previous studies have primarily focused on Danish patients, and it is essential to determine if the algorithm performs consistently across different populations. Assessing the algorithm’s performance in diverse patient populations will enhance its utility and applicability in clinical practice.
Method
The electrocardiogram (ECG) dataset utilized in this study was collected from patients with pharmacoresistant epilepsy undergoing presurgical long-term video-EEG monitoring at the Epilepsy Surgery Center of Santa Catarina, Brazil. The dataset included segments of both ictal (seizure) and inter-ictal (non-seizure) periods, recorded with wired electrodes at sample frequencies of 256 or 512 Hz. To ensure data quality, noisy ECG recordings and those shorter than 10 minutes were excluded, except for recordings containing seizures.
To analyze heart rate variability (HRV), a fully automated R-peak detection algorithm was employed. Seizure timing was marked by a neurophysiology expert, and responders were identified based on a predefined increase in heart rate during seizures.
The study utilized training and test datasets from Danish and Brazilian cohorts, respectively, for the seizure detection algorithm. The algorithm, based on logistic regression machine learning, identified seizure candidates by detecting heart rate changes above a threshold and classified them as seizures or non-seizures based on HRV parameters.
The decision boundary of the classifier was initially established using a training dataset and gradually adapted to individual patients based on their data. Statistics were employed to compare demographic and clinical characteristics between responder and non-responder patients, utilizing chi-square tests and student t-tests.
Overall, the study employed sophisticated methodologies to analyze ECG data and develop a seizure detection algorithm, aiming to improve the understanding and management of epilepsy.
Result
The study examined data from 34 patients who experienced a total of 107 seizures. Among these patients, 22 demonstrated a positive response to seizure detection, indicating a response rate of 64.7%. These responders exhibited a total of 118 “seizure candidate” epochs, with nearly all seizures being identified as candidates for detection, highlighting the effectiveness of the detection algorithm. The overall sensitivity of seizure detection in responders was found to be 84.8%, with a relatively low false alarm rate of .25 per 24 hours. Specifically, focal seizures without bilateral convulsions were detected in 75.8% of cases, while generalized tonic-clonic seizures (GTC) were identified in 96.2% of instances.
However, it’s important to note that when considering technically flawed recordings or those with excessive noise, the sensitivity of detection decreased to 52.6%. This underscores the importance of ensuring data quality and mitigating technical issues in seizure detection systems.
Conversely, among the non-responder group comprising 12 patients, the sensitivity was notably lower at 27.1%, accompanied by a relatively higher false alarm rate of 1.1 per 24 hours. Despite efforts to detect seizures in this group, the results suggest a less effective performance compared to responders.
Moreover, the study did not identify any significant differences in demographic or clinical characteristics between responders and non-responders, indicating that individual patient factors may not strongly influence the efficacy of seizure detection.
Additionally, the F1 score, a measure of the accuracy of the detection algorithm, was relatively high at .91 for the responder group after excluding technically compromised recordings. However, including these flawed recordings resulted in a lower F1 score of .68, highlighting the impact of data quality on algorithm performance.
Furthermore, the study evaluated the use of tachycardia as a potential marker for seizure detection, but found that while it yielded higher sensitivity rates, it also led to substantially increased false alarm rates compared to heart rate variability-based algorithms.
Overall, these findings provide valuable insights into the efficacy and limitations of seizure detection algorithms, emphasizing the importance of data quality, algorithm refinement, and individualized approaches in optimizing seizure detection systems for patients with epilepsy.
Conclusion
In this validation study, a novel personal adaptive machine learning algorithm was employed to detect seizures using heart rate variability (HRV) measurements. The results demonstrated a sensitivity of 84.4% and a low false alarm rate of .25/24h, confirming the efficacy of the seizure detection algorithm across both focal and generalized tonic–clonic seizures. When compared with a previous study conducted on Danish patients using the same algorithm, a slightly higher detection sensitivity was observed in Brazilian patients, though not statistically significant. Additionally, a higher percentage of responders was found in the Brazilian cohort, potentially influenced by socio-demographic differences and healthcare system disparities between the two countries.
The study highlighted the impact of disease duration on HRV indices and emphasized the role of limbic structures in seizure detection using autonomic biomarkers like HRV changes. However, limitations such as data quality issues due to retrospective design and non-continuous dataset of interictal data were acknowledged. Despite these limitations, the study underscores the importance of pre-selecting patients for optimal performance of personalized seizure detection algorithms, particularly those who exhibit marked autonomic changes during seizures.
Furthermore, the study identified the need for patient feedback mechanisms to improve algorithm performance and highlighted the potential of embedding the seizure detection algorithm into wearable devices for real-time seizure monitoring and treatment optimization. Overall, the study contributes to the validation of HRV-based seizure detection algorithms and their potential clinical utility in enhancing patient care and treatment outcomes.
Oncology Related Tools
Other
Latest Research
