Introduction
Methods
Results
Study | Autonomic parameter | Measurement | Device | Algorithm | Prospective/retrospective validation | Population (N) | No. of seizures/TRTa | Type of seizures (N) | Mean age (years) [range] | Performance (mean of values per person) |
---|---|---|---|---|---|---|---|---|---|---|
De Cooman et al. [22] | Cardiac | HRV | Single-lead ECG | Noise filtering: High- and low-pass Butterworth filters. HRI-extract algorithm and HRI feature extraction with patient-independent SVM classifier: LOPO-CT. Fastened HRI-SVM seizure detection | Retrospective | Refractory temporal lobe epilepsy (17) | 127/918 h | FOIA, FOBTC | 33.5 [9–54] | Sens: 83.2% [50–100%] (overall: 81.9%) PPV: 7.9% [0.4–21%] (overall: 5.4%) FAR: 2.01/h [0.88–3.52/h] (overall: 1.97/h) DL: 13.3 s [− 18.2–54.3] (overall:17.8 s) |
De Cooman et al. [23] | Cardiac | HRV | Single-lead ECG | See De Cooman [22]. Patient-specific heuristic adaptive classifier | Retrospective | 1) Children (14) 2) Other group of children (14) | 107/695 h | GOS (30), FOS (77) | NA | Patient-independent: Sens: (overall:81.3%) PPV: NA FAR (overall: 0.75/h) DL: NA Patient-specific: Sens: (overall: 77.6%) PPV: (overall: 30.7%) FAR (overall: 0.33/h) DL: 19.1 s |
De Cooman et al. [24] | Cardiac | HRV | Single-lead ECG | See De Cooman [22]. Patient-specific heuristic adaptive classifier and real-time adaptive classifier | Retrospective | Temporal lobe epilepsy (19) | 153/2833 h | FOS, FOIA, FOBTC, U, subclinical | NA | Patient-independent: Sens: (overall: 78.4%) PPV: (overall: 2.4%) FAR: (overall: 1.73/h) DL: NA Patient-specific: Sens: (overall: 76.5%) PPV: (overall: 3.7%) FAR: (overall: 1.09/h) DL: NA Adaptive: Sens: (overall: 77.1%) PPV: (overall: 3.3%) FAR: (overall: 1.24/h) DL: NA |
Fujiwara et al. [25] | Cardiac | HRV | ECG | Time domain analysis: mean NN, SDNN, RMSSD, TP, NN50. Frequency domain analysis: LF, HF, LF/HF. Analysis over 2-5 min. Algorithm 2: T2 and Q statistics exceed limit > 10 s continuously | Retrospective | Refractory focal epilepsy (14) | 11/69 h | FOS, awakening seizures (11) | 30.6 [14–63] | T2 (Overall): Sens: 55% PPV: NA FAR: 1.2/h DL: − 524 ± 216 s Q (Overall): Sens: 91% PPV: NA FAR: 0.7/h DL: − 494 ± 262 s |
Jeppesen et al. [26] | Cardiac | HRV | Single-lead ECG | Noise filtering: High-pass filter + manual edit. Automatic R-peak detection. Lorenz plotb analysis: SD1, SD2, CSI, mCSI, CVI | Retrospective | Temporal lobe epilepsy (5) | 11/13 h | FOIA (11) | NA | Sens: 88% (CSI-30), (overall: 73%, CSI-30, mCSI-50) PPV: NA FAR: NA DL: − 5–60 s |
Jeppesen et al. [8] | Cardiac | HRV | Single-lead ECG | Noise filtering: High-pass filter + manual edit. Automatic R-peak detection. Frequency domain analysis: HF-power (using FFT) HR-diff. Lorenz plotb analysis: CSI, mCSI | Retrospective | Focal epilepsy (17) | 47/±27 h | FOS (44), FOBTC (3) | 39 [20–55] | Sens: 81% (mCSI-100) (overall: 74%, mCSI-100) PPV: NA FAR: NA DL: 16 s [6–50] |
Moridani et al. [27] | Cardiac | HRV | ECG | R-peak detection by Pan and Tompkins’ algorithm. Time domain analysis: SDNN, RMSSD, NN50, pNN50. Frequency domain analysis: LF, HF, VLF, LF/HF. Poincaré plot analysis: SD1, SD2, SD2/SD1 | Retrospective | Focal epilepsy (7) | 11/±6 h | NA | NA | Sens: (overall: 88.3%) PPV: NA FAR: NA DL: NA |
Pavei et al. [28] | Cardiac | HRV | ECG | Noise filtering: Visual artifact inspection, high- and low-pass Butterworth filters. QRS detection algorithm by Kohler. Time domain analysis: SDNN, RMSSD. Frequency domain analysis: LF, HF (using FFT). SampEn: Entropy changes. Lorenz plot analysis: CSI, CVI | Retrospective | Temporal lobe epilepsy (12) | 34/171 h | FOIA (34) | 34.5 SD 7.5 | Sens: (overall: 94.1%) PPV: (overall: 95.6%) FAR: (overall: 0.49/h) DL: NA |
Qaraqe et al. [29] | Cardiac | HRV | Single-lead ECG | Noise filtering: Baseline estimation and denoising with sparsity. QRS detection algorithm. Outlier removal, linear interpolation. Time–frequency analysis: MP-WVD algorithm. SVM classifier for EEG features | Retrospective | Focal epilepsy (7) | 68/NA | FOA, FOIA, FOBTC | 43.6 [26–65] | ECG: Sens: 96.4% [75–100%] PPV: NA FAR: 5.4/h [1.5–9.5/h] DL: 13.1 s [8–20.5] ECG + EEG: Sens: 100% PPV: NA FAR: 1.6/h [0–3.5/h] DL: 12.3 s [3–26] |
Vandecasteele et al. [30] | Cardiac | HRV/PRV | 180° eMotion Faros and Empatica E4 smart-watch | R-peak detection by Pan–Tompkins’ algorithm. HRV analysis: Method of Varon. PRV analysis: Method of Lázaro. Seizure detection algorithm with SVM classifier by de Cooman. Feature extraction: HRpeak, HRbase, and STDHRbase | Retrospective | Temporal lobe epilepsy (11) | 47/701 h | NA | 42.7 [19-67] | Wearable ECG: Sens: 64% (overall: 70%) PPV: 2.03% (overall: 2.15%) FAR: 2.35/h (overall: 2.11/h) DL: NA Hospital ECG: Sens: 57% (overall: 57%) PPV: 2.22% (overall: 1.93%) FAR: 2.05/h (overall: 1.92/h) DL:NA PPG: Sens: 33% (overall:32%) PPV: 1.43% (overall: 1.12%) FAR: 1.88/h (overall:1.80/h) DL: NA |
Varon et al. [43] | Cardiac | QRS morphology | Single-lead ECG | R-peak detection via Pan–Tomkins’ algorithm. Algorithm 1: principal component analysis for changes in QRS morphology. Algorithm 2: ictal acceleration of HR quantified by using phase-rectified signal averaging (PRSA) | Retrospective | 1) Children with refractory epilepsy (37) 2) Women with epilepsy (5) | 1) 98 2) 10/±5 hc | 1) FOS (48) (28 frontal, 20 temporal) GOS (50) (29 T/TC 11 MC, 10 absences) 2) FOS (10) | 1) 9.2 [3–16] 2) [31–48] | Algorithm 1d: Sens: 89.5% (F1), 86% (G1), 100% (F2) PPV: 85.7% (F1), 57.3% (G1), 52.6% (F2) FAR: NA DL: NA Algorithm 2: Sens: 100%(F1), 90%(G1), 100%(F2) PPV: 90.5% (F1), 77.5% (G1), 71.4% (F2) FAR: NA DL: NA |
Elmpt, van et al. [32] | Cardiac | HR | 2-lead ECG | R identification (increase signal > 250 µV in 10 ms). HR: baseline (60 s before increase), periictal period (120 s after), SD, min and max. Detection threshold: +2SD from baseline for ≥ 5QRS complexes. Curve-fitting algorithm and onset detection algorithm | Retrospective | Severe epilepsy (10) | 104/9 h | T, TC, MC, atypical absences | 34.1 [21–50] | Sens: NA PPV: NA FAR: NA DL: NA HR increase in 48.1% of seizures. Great variability in sens and PPV. Better performance when combined with ACC |
Osorio et al. [34] | Cardiac | HR | ECG | True beat range determined (30–180 bpm). 5 s moving window: RRI determination. Time of beat sequence (TOBS) : RHR, 4 threshold values (T) and 3 duration values (D) | Retrospective | Focal onset epilepsy (81) Dataset 1 (41), 2 (40) | 241/6935 h | FOS | NA | Lowest settings T, D Sens: 98.8% PPV: NA FAR: 9.5/h (1), 7.2/h (2) DL: NA Highest settings T, D Sens: 85.5% PPV: NA FAR: 1.1/h (1), 0.7/h (2) DL: NA |
Boon et al. [31] | Cardiac | HR | VNS—AspireSR | Relative HR increase > 1 s above threshold (≥ 20%, ≥ 40%, ≥ 60% above baseline HR) | Prospective | Refractory epilepsy (16) | 66/NA | FOS (8), FOA (26), FOIA (31), FOBTC (17), U (5)e | 39.6 SD 13.4 [19–66] | Threshold > 20%: Sens: 16/27 = 59.3% PPV: NA FAR: 7.2/h [95% CI 5.31–9.94] DL: 6.0 s [− 112–105] Threshold > 40%: Sens: 8/23 = 34.8% PPV: NA FAR: 2.7/h [95% CI 1.70–3.91] DL: 27.5 s [0–57] Threshold > 60%: Sens: 3/16 = 18.8% PPV: NA FAR: 0.5/h [95% CI 0.20–0.96] DL: 35.0 s [4–40] |
Hampel et al. [33] | Cardiac | HR | VNS—AspireSR | Heartbeat sensitivity threshold 50% compared to baseline | Prospective | Refractory epilepsy (1) | 12/68 h | FOS with hyperkinetic movements | 29 | Sens: 92% PPV: 8% FAR: 1.88/h (n = 128) DL: 7.4 s (± 5) |
Andel, van et al. [36] | Combined | HR, ACC | Shimmer sensor (chest ECG + 3D ACC) | Algorithm 1: No. of s in which summed waveform length > fixed threshold within a fixed window. Detection if no. > window length/4. Algorithm 2: HR > threshold. Algorithm 3: combination of summed waveform length OR HR | Retrospective | Epilepsy (43) | 86/402 hf | Major motor (86) (18 TC, 41 T, 18 HM, 9 Cluster) | Median 15 [2–65] | All seizures: Sens: 60% (A1), 56% (A2), 71% (A3) PPV: NA FAR: 0.5/h (A1), 0.3/h (A2), 0.7/h (A3). DL: NA Clinically urgent seizures: Sens: 74% (A1), 71% (A2), 87% (A3) PPV: NA FAR: 0.6/h (A1), 0.3/h (A2), 0.8 (A3). DL: NA |
Cogan et al. [2] | Combined | HR, SpO2, EDA | Affectiva Q-curve and Nonin WristOx2 sensor | Seizure pattern analysis (HR↑, SpO2↓, EDA↑). Biosignal algorithm. Personalized parameters (P) | Retrospective | Focal epilepsy (10) | 26/340 h | FOIA (23), FOBTC (2), GTCS (1) | 41.8 [21–64] | 3 Sensors (n = 6): Sens: 100%, 100% (P) PPV: 86%, 100% (P) FAR: 0.015/h, 0.000/h (P) DL: NA |
Goldenholz et al. [37] | Combined | HR, QTc, SpO2 | Single-lead ECG, Radical-7 | SpO2: ictal drop, provided it remained > 50%. Optimal balance 80–86%. Calculation of HR and QTc by Bazette method | Retrospective | Refractory epilepsy (45) | 151/7104 h | FOS (119), FOBTC (32) | 40 [14–68] | Sensg: (overall: 81–94% (FOBTC), 25–36%(FOS)) PPV: NA FAR: (overall: 0.4–2.4/h) DL: NA |
Heldberg et al. [38] | Combined | EDA, ACC | Empathica E3 wristband | EDA: low-pass filter, cutoff frequency of 1.5 Hz. Time windows: 10 s 50% overlap and 5 min 80% overlap. Decomposing signal (Ledalab algorithm). Feature extraction (56) of EDA and ACC. kNN (11 features) and random forest (26 features) classifiers | Retrospective | Epilepsy (8) | 55/540 h | Motor seizures (21), nonmotor (34) | NA | kNN classifier: Sens: 76.2% (M), 97.1% (non-M) PPV: 4.6% (M), 9.7% (non-M) FAR: NA DL: NA Random forest: Sens: 90.5% (M), 85.3% (non-M) PPV: 5.6% (M), 12.3% (non-M) FAR: NA DL: NA |
Onorati et al. [39] | Combined | EDA, ACC | Empatica E3, E4, iCALM | 10 s sliding epochs (75% overlap), feature extraction, classifier, decision thresholds. (3 sets: Poh’s (19), larger (46), reduced (25)) | Retrospective | Epilepsy (69) (24 children, 45 adults) | 55/5928 h | FOBTC (49), FOTC (6) | Median 14/37 [4–60] | Sens: 83.6% (C1), 92.7% (C2), 94.6 (C3) PPV: 39% (C1), 50% (C2), 51% (C3) FAR: 0.29/day (C1), 0.21 (C2), 0.20 (C3) DL: 31.2 s (C1), 29.3 s (C2), 29.3 s (C3) |
Poh et al. [40] | Combined | EDA, ACC | Custom-built wrist-worn biosensors | 10 s epochs, sliding window with 75% overlap, preprocessing, 19 time, frequency, and nonlinear features extracted to form feature vectors. SVM to classify vectors as (non)seizure. Cross-validation | Retrospective | Focal epilepsy (7) | 16/688 hh | FOBTC (16) | 10 SD 4.6 | Nonpatient-specific: Sens: 14/16 = 88% PPV: NA FAR: 0.04/h (n = 28) DL: NA Semi-patient specific: Sens: 15/16 = 94% PPV: NA FAR: 0.04/h DL: NA |
Heart rate analysis
Heart rate variability (HRV)
Combining autonomic parameters
Quality of the included studies
Study | Subjects | Recordings | Analysis and alarms | Reference standard | Study phase | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Simulation/healthy subjects | No. of people with seizures | No. of seizures | Conventional methods | Dedicated device | Continuous | Multicenter | Offline/retrospective | Training and testing using the dataset | Predefined algorithm and cutoff values | Real time | Blinded | Video or video-EEG recordings | Information from pt and caregivers | ||
Van Andel et al. [36] | – | 20–50 | ≥ 75 | – | + | + | + | + | – | + | – | – | + | – | 1 |
Boon et al. [31] | – | 10–20 | 30–75 | – | + | + | + | – | – | + | + | – | + | + | 2 |
Cogan et al. [2] | – | 10–20 | 15–30 | – | +a | + | – | + | + | – | – | – | + | – | 1 |
De Cooman et al. [22] | – | 10–20 | ≥ 75 | + | – | + | – | + | + | – | – | – | + | – | 1 |
De Cooman et al. [23] | – | 20–50 | ≥ 75 | + | – | + | + | + | – | – | – | – | + | – | 1 |
De Cooman et al. [24] | – | 10–20 | ≥ 75 | + | – | + | – | + | – | – | – | – | –b | –c | 1 |
van Elmpt et al. [32] | – | 10–20 | ≥ 75 | + | – | + | – | + | + | – | – | – | + | – | 1 |
Fujiwara et al. [25] | – | 10–20 | 1–15 | + | – | – | + | + | – | – | – | – | + | – | 1 |
Goldenholz et al. [37] | – | 20–50 | ≥ 75 | + | – | + | – | + | + | – | – | – | + | – | 1 |
Hampel et al. [33] | – | ≥ 1 | 1–15 | – | + | + | – | – | – | + | + | – | + | + | 1 |
Heldberg et al. [38] | – | 1–10 | 30–75 | – | + | + | – | + | – | – | – | – | + | – | 1 |
Jeppesen et al. [26] | – | 1–10 | 1–15 | + | – | – | – | + | + | – | – | – | + | – | 1 |
Jeppesen et al. [8] | – | 10–20 | 30–75 | + | – | – | – | + | + | – | – | – | + | – | 1 |
Moridani et al. [27] | – | 1–10 | 1–15 | + | – | + | – | + | + | – | – | – | –b | – | 0 |
Onorati et al. [39] | – | ≥ 50 | 30–75 | – | + | + | + | + | – | – | – | – | + | – | 1 |
Osorio et al. [34] | – | ≥ 50 | ≥ 75 | + | – | + | + | + | + | – | – | – | –d | – | 0 |
Pavei et al. [28] | – | 10–20 | 30–75 | + | – | – | – | + | – | – | – | – | + | – | 1 |
Poh et al. [40] | – | 1–10 | 15–30 | – | + | – | – | + | + | – | – | – | + | – | 1 |
Qaraqe et al. [29] | – | 1–10 | 30–75 | + | – | – | – | + | + | – | – | – | –b | – | 0 |
Vandecasteele et al. [30] | – | 10–20 | 30–75 | + | + | + | – | + | – | + | – | – | + | – | 1 |
Varon et al. [43] | – | 20–50 | ≥ 75 | + | – | – | – | + | – | – | – | – | + | – | 1 |
Discussion
Study | Validation of algorithm | Performance of algorithm | ||||||
---|---|---|---|---|---|---|---|---|
No. of subjects | No. of seizures/TRT | Type of seizures | Algorithm | Sensitivity (%) | FAR | PPV (%) | DL (s) [range] | |
Large datasets | ||||||||
van Andel et al. [36] | 23 | 86/402 ha | All major motorb | Heart rate | 60 | 0.5/h | NA | NA |
Movement | 56 | 0.3/h | NA | NA | ||||
Hart rate or movement | 71 | 0.7/h | NA | NA | ||||
59 | Clinically urgent seizuresc | Heart rate | 74 | 0.6/h | NA | NA | ||
Movement | 71 | 0.3/h | NA | NA | ||||
Hart rate or movement | 87 | 0.8/h | NA | NA | ||||
De Cooman et al. [22] | 17 | 127/918 h | FOS, including TCs | 83.2 [50–100] | 2.01/h [0.88–3.52/h] | 7.9% [0.4–21%] | 13.3[− 18.2–54.3] | |
De Cooman et al. [23] | 28 | 107/695 h | Convulsive and clinical subtle seizures | Patient-independent | Overall: 81.3 | Overall: 0.75/h | NA | NA |
Patient-specific | Overall: 77.6 | Overall: 0.33/h | Overall: 30.7 | 19.1 | ||||
De Cooman et al. [24] | 19 | 153/2833 h | FOS, including TCs (only clinical seizures) | Patient-independent | Overall: 78.4 | Overall: 1.73/h | Overall: 2.4 | NA |
Patient-specific | Overall: 76.5 | Overall: 1.09/h | Overall: 3.7 | NA | ||||
Adaptive | Overall: 77.1 | Overall: 1.24/h | Overall: 3.3 | NA | ||||
Goldenholz et al. [37] | 45 | 151/7104 h | FOS, including TCs | Overall: 81–94 (FOBTC) 25–36 (FOS)e | Overall: 0.4–2.4/h | NA | NA | |
Onorati et al. [39] | 69 | 55/5928 h | FOS, all TCs | Classifier 1 | 83.6 | 0.29/day | 39 | 31.2 |
Classifier 2 | 92.7 | 0.21/day | 50 | 29.3 | ||||
Classifier 3 | 94.6 | 0.20/day | 51 | 29.3 | ||||
Medium datasets | ||||||||
Boon et al. [31] | 16 | 66/NA | Different types of FOS, including TCs | Threshold > 20% | 59.3 | 7.2/h [95% CI 5.31–9.94] | NA | 6.0 [-112–105] |
Threshold > 40% | 34.8 | 2.7/h [95% CI 1.70–3.91] | NA | 27.5 [0–57] | ||||
Threshold > 60% | 18.8 | 0.5/h [95% CI 0.20–0.96] | NA | 35.0 [4–40] | ||||
Heldberg et al. [38] | 8 | 55/540 h | Motor (M) and non-M seizures | kNN classifier | 76.2 (M) 97.1 (non-M) | NA | 4.6 (M) 9.7 (non-M) | NA |
Random forest | 90.5 (M) 85.3 (non-M) | NA | 5.6 (M) 12.3(non-M) | NA | ||||
Jeppesen et al. [8] | 17 | 47/ ± 27 h | FOS, including TCs | 81: (mCSI-100) (overall: 74, mCSI-100) | NA | NA | 16 [6–50] | |
Osorio et al. [34] | 81 | 241/6935 h | FOS | Lowest settings T,D Datasets (1) and (2) | 98.8 | 9.5/h (1) 7.2/h (2) | NA | NA |
Highest settings T,D Datasets (1) and (2) | 85.5 | 1.1/h (1) 0.7/h (2) | NA | NA | ||||
Pavei et al. [28] | 12 | 34/171 h | FOIA | Overall: 94.1 | Overall: 0.49/h | Overall: 95.6 | NA | |
Poh et al. [40] | 7 | 16/688 hf | FOS, all TCs | Non-patient-specific | 88 | 0.04/h (n = 28) | NA | NA |
Semi-patient-specific | 94 | 0.04/h | NA | NA | ||||
Qaraqe et al. [29] | 7 | 68/NA | FOS, including TCs | ECG | 96.4 [75–100] | 5.4/h [1.5–9.5/h] | NA | 13.1 [8–20.5] |
ECG + EEG | 100 | 1.6/h [0–3.5/h] | NA | 12.3 [3–26] | ||||
Vandecasteele et al. [30] | 11 | 47/701 h | FOIA | Wearable ECG | 64 (overall: 70) | 2.35/h (overall: 2.11/h) | 2.03 (overall: 2.15) | NA |
Hospital ECG | 57 (overall: 57) | 2.05/h (overall: 1.92/h) | 2.22 (overall: 1.93) | NA | ||||
PPG | 33 (overall:32) | 1.88/h (overall:1.80/h) | 1.43 (overall: 1.12) | NA | ||||
Small datasets | ||||||||
Cogan et al. [2] | 6 | 10/340 h | FOIA and TCs | 3 Sensors | 100 | 0.015/h | 86 | NA |
Personalized | 100 | 0.000/h | 100 | NA | ||||
Elmpt, van et [32] | 10 | 104/9 h | Motor seizures (T, TC, MC) and atypical absences | NAd | NA | NA | NA | |
Fujiwara et al. [25] | 14 | 11/69 h | FOS (awake) | T2 statistics | Overall: 55 | Overall: 1.2/h | NA | − 524 ± 216 |
Q statistics | Overall: 91 | Overall: 0.7/h | NA | − 494 ± 262 | ||||
Hampel et al. [33] | 1 | 12/68 h | FOS with hyperkinetic movements | 92 | 1.88/h | 8 | 7.4 (± 5) | |
Jeppesen et al. [26] | 5 | 11/13 h | FOIA | 88 (CSI-30) (overall: 73, CSI-30, mCSI-50) | NA | NA | − 5–60 | |
Moridani et al. [27] | 7 | 11/± 6 h | FOS | Overall: 88.3 | NA | NA | NA | |
Varon et al. [43] | 42 | 108/± 5 h | FOS and GOS, including T, TC, MC, and absences | Algorithm 1g | 89.5 (F1) | NA | 85.7 (F1) | NA |
86 (G1) | 57.3 (G1) | |||||||
100 (F2) | 52.6 (F2) | |||||||
Algorithm 2f | 100 (F1) | NA | 90.5 (F1) | NA | ||||
90 (G1) | 77.5 (G1) | |||||||
100 (F2) | 71.4 (F2) |