Studying Heart Rhythms With a Wearable Sensor to Detect and Predict Seizures in People With Epilepsy

Mean heart rate will be calculated in beats per minute for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

The standard deviation of normal-to-normal heartbeat intervals (SDNN) will be calculated in milliseconds for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

The root mean square of successive differences between normal-to-normal heartbeat intervals (RMSSD) will be calculated in milliseconds for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

pNN50 is the percentage of successive normal-to-normal heartbeat interval pairs that differ by more than 50 milliseconds. It will be calculated for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

Low-frequency (LF) power of normal-to-normal heartbeat interval variability will be calculated in milliseconds squared for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

High-frequency (HF) power of normal-to-normal heartbeat interval variability will be calculated in milliseconds squared for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

The ratio of low-frequency power to high-frequency power (LF/HF) of normal-to-normal heartbeat interval variability will be calculated for each analyzed recording period. The ratio is dimensionless. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

Poincare plot SD1 is the standard deviation of normal-to-normal heartbeat interval pairs perpendicular to the line of identity. It will be calculated in milliseconds for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

Poincare plot SD2 is the standard deviation of normal-to-normal heartbeat interval pairs along the line of identity. It will be calculated in milliseconds for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

Sample entropy, a dimensionless measure of the irregularity of normal-to-normal heartbeat interval sequences, will be calculated for each analyzed recording period. Textile-based wearable electrocardiography (ECG) recordings will be used. Values from the standardized one-hour awake resting recording will be compared between the epilepsy and healthy volunteer groups. Sleep and wakefulness will also be compared between groups, with sleep stage analyses limited to reliably classified periods. Within epilepsy participants, interictal, preictal, ictal and postictal values will be compared using video-EEG reference annotations and accounting for repeated observations. Equal-length, technically usable segments will be analyzed under prespecified rules. Effect sizes and confidence intervals will be reported.

The percentage of eligible focal to bilateral tonic-clonic seizures confirmed by video-EEG that are correctly detected by each model: correctly detected seizures divided by all eligible reference seizures in the test recordings, multiplied by 100. Reference events will be annotated independently of ECG changes and model outputs. Alarm-to-event matching rules will be specified before outcome analysis. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.

The number of detection alarms not matched to a reference seizure, divided by the total evaluable monitoring time in hours. Reference seizures will be established by independent video-EEG annotation. Alarm-to-event matching and alarm-merging rules will be specified before outcome analysis and applied consistently to held-out recordings. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.

The percentage of eligible lead focal to bilateral tonic-clonic seizures correctly preceded by a model warning within the prespecified prediction timing rules. Correctly predicted lead seizures will be divided by all eligible lead seizures in the test recordings and multiplied by 100. Independent targets must be preceded by at least two seizure-free hours. The prediction horizon, seizure occurrence period and alarm-to-event matching rules will be specified before outcome analysis. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.

The number of prediction alarms not followed by an eligible reference seizure within the prespecified seizure occurrence period, divided by the total evaluable monitoring time in hours. Lead seizure eligibility, prediction horizon, seizure occurrence period and alarm-matching rules will be specified before outcome analysis. Video-EEG annotations will provide the reference events. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.

The percentage of reference-positive analysis windows classified as positive: true-positive windows divided by all reference-positive windows, multiplied by 100. Reference-positive windows represent seizure activity for detection and the prespecified preictal target for prediction. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The percentage of reference-negative analysis windows classified as negative: true-negative windows divided by all reference-negative windows, multiplied by 100. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The percentage of positive model classifications that are reference-positive: true-positive windows divided by all predicted-positive windows, multiplied by 100. This measure is also called precision. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The percentage of negative model classifications that are reference-negative: true-negative windows divided by all predicted-negative windows, multiplied by 100. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The F1 score is the harmonic mean of window-level precision and sensitivity. It is dimensionless, ranges from 0 to 1, and is higher when classification performance is better. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

Balanced accuracy is the arithmetic mean of window-level sensitivity and specificity, expressed as a percentage. It gives equal weight to performance on positive and negative reference classes. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The area under the receiver operating characteristic curve (ROC-AUC) will summarize discrimination across classification thresholds using window-level prediction scores. It is dimensionless and ranges from 0 to 1. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The area under the precision-recall curve (AUPRC) will summarize the relationship between precision and sensitivity across classification thresholds using window-level prediction scores. It is dimensionless and ranges from 0 to 1. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

The percentage of all analysis windows classified correctly: the sum of true-positive and true-negative windows divided by all evaluated windows, multiplied by 100. Accuracy will be interpreted alongside the other metrics because seizures are much less frequent than non-seizure periods. Calculated separately for seizure detection and seizure prediction classification tasks and for each model approach, using prespecified reference window labels derived from video-EEG. Evaluation will use held-out recordings, preserving participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Values will be summarized across validation folds or repeated runs, with standard deviations or appropriate confidence intervals.

For correctly detected focal to bilateral tonic-clonic seizures, the interval from the clinical seizure onset annotated using video-EEG to the corresponding model alarm will be calculated in seconds. Mean latency will be reported for each detection model, with variability or confidence intervals. Alarm matching will follow prespecified rules. Timing will be assessed offline from time-stamped recordings and algorithm outputs; it does not represent prospective deployment of a clinical alarm. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.

For correctly predicted eligible lead focal to bilateral tonic-clonic seizures, the interval from the corresponding model warning to reference seizure onset will be calculated in minutes. Mean lead time will be reported for each prediction model, with variability or confidence intervals. Reference onset, lead seizure eligibility, prediction horizon, seizure occurrence period and the rule for selecting the matched warning will be specified before outcome analysis. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.

The total duration of the model-defined seizure warning state divided by the total evaluable recording duration, multiplied by 100. Overlapping warning intervals will contribute only once to total warning time. The duration and timing rules defining the warning state will be specified before outcome analysis and applied consistently. This measure will be reported separately for each prediction model alongside event-level sensitivity and false alarms per hour. Performance will be evaluated by internal validation on held-out recordings, with participant-level separation for patient-independent models and separation by seizure event and time for patient-specific models. Results will be reported separately by model approach, with variability across validation folds or repeated runs.