- Email: [email protected]
Probability of detection of clinical seizures using heart rate changes
Seizure 30 (2015) 120–123
Contents lists available at ScienceDirect
Seizure journal homepage: www.elsevier.com/locate/yseiz
Probability of detection of clinical seizures using heart rate changes Ivan Osorio a,*, B.F.J. Manly b a b
University of Kansas Medical Center, United States Manly-Biostatistics Ltd., United States
A R T I C L E I N F O
A B S T R A C T
Article history: Received 12 May 2015 Received in revised form 18 June 2015 Accepted 21 June 2015
Purpose: Heart rate-based seizure detection is a viable complement or alternative to ECoG/EEG. This study investigates the role of various biological factors on the probability of clinical seizure detection using heart rate. Methods: Regression models were applied to 266 clinical seizures recorded from 72 subjects to investigate if factors such as age, gender, years with epilepsy, etiology, seizure site origin, seizure class, and data collection centers, among others, shape the probability of EKG-based seizure detection. Results: Clinical seizure detection probability based on heart rate changes, is significantly (p < 0.001) shaped by patients’ age and gender, seizure class, and years with epilepsy. The probability of detecting clinical seizures (>0.8 in the majority of subjects) using heart rate is highest for complex partial seizures, increases with a patient’s years with epilepsy, is lower for females than for males and is unrelated to the side of hemisphere origin. Conclusion: Clinical seizure detection probability using heart rate is multi-factorially dependent and sufficiently high (>0.8) in most cases to be clinically useful. Knowledge of the role that these factors play in shaping said probability will enhance its applicability and usefulness. Heart rate is a reliable and practical signal for extra-cerebral detection of clinical seizures originating from or spreading to central autonomic network structures. ß 2015 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
Keywords: Clinical seizures Heart rate based detection Detection probability Factors/variables Regression models
1. Introduction Automated detection of epileptic seizures originating from or spreading to central autonomic network (CAN) structures [1–6] using heart rate changes [7–15], is gaining acceptance as either a valuable complement or as an alternative to cortical electrical signals (e.g., ECoG) due to their: (a) Higher signal-to-noise ratio (mV for EKG vs. mV ECoG; (b) Greater ease and economy of recording (EKG requires 2 electrodes for indirect but all encompassing monitoring of ictal activity in CAN structures, while the same number of scalp or intracranial electrodes would yield inadequate coverage [16]; (c) Lower processing and computational analysis cost [13] due to EKG’s lower signal complexity than ECoG. Since seizure detection using heart rate is in an early developmental stage, more knowledge about its power and reliability is desirable to meaningfully assess its clinical applicability. The magnitude of relative changes in heart rate at seizure
* Corresponding author at: 3901 Rainbow Blvd, Kansas City, KS 66160, United States. Tel.: +1 913 5884529; fax: +1 913 5884585. E-mail addresses: [email protected], [email protected] (I. Osorio).
onset, is the single most important biological factor governing the probability of detection using this variable [13], which is subject not only to fluctuations in type and level of physical activity, but also to various other factors. The degree to which age, gender, seizure class (e.g., simple vs. complex partial), etiology (e.g., lesional vs. cryptogenic) or hemispheric location (left vs. right) of the epileptogenic zone may impact the probability of clinical seizure detection is unknown. Also, the plasticity (e.g., accommodation or potentiation) of the central and peripheral autonomic system [17] raises relevant questions such as: Does the intensity of the ictal cardiac response depend on the length of the course of epilepsy? If yes, is it enhanced or blunted? The importance of identifying which factors and the degree to which they shape the cardiac response to seizures and through it, the probability of detection of clinical seizures, motivates this investigation. To this end, the probability of detection (which in this study corresponds to True Positive detections) of clinical seizures using heart rate changes was investigated as a function of: Subjects’ age and gender, years with epilepsy, etiology, site of origin, seizure class, electrode type (subdural vs. depth) used for localization of the epileptogenic zone, and data collection epilepsy centers.
http://dx.doi.org/10.1016/j.seizure.2015.06.007 1059-1311/ß 2015 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
I. Osorio, B.F.J. Manly / Seizure 30 (2015) 120–123
121
2. Methods and materials
2.3. EKG-based seizure intensity, duration and severity
2.1. Data sets
Measures of intensity, duration and severity for seizures detected using ECoG [20,21] were applied to seizures detected using heart rate changes, since they are statistically equivalent and thus clinically meaningful [14]. (a) Seizure intensity (Szi) was defined as the maximal ictal heart rate (Max HR; bpm) recorded during an event; (b) Seizure duration (SzDur) was defined as the time (in s) the HR spends above a ‘‘threshold’’), and (c) Seizure severity (SzSev) as the product of MaxHR and SzDur.
This investigation utilizes the data of an extensive validation of an EKG-based seizure detection algorithm published elsewhere [13,14]. Only details deemed necessary to properly frame this work will be provided herein. Seventy-nine ECoGs (72 subjects; 6935 h, 266 clinical seizures) with simultaneous EKG were analyzed with two different seizure detection algorithms: One was applied to ECoG [18,19] and the other to EKG [13,14]. These data, from which all identifiers had been removed, were collected with approval from the Human Subjects Committee of each of the contributing centers. One data set (33 subjects; 98 clinical seizures; 4–8 h recordings; total duration 352 h) was contributed by the Minnesota Epilepsy Group, Montreal Neurological Institute, Stanford Medical Center and the University of Kansas Medical Center (‘‘Multi-Center’’) and the other (39 subjects, 168 clinical seizures; full stay recordings, 6583 h) by the University of Kansas Medical Center (‘‘Full Stay’’). Each subject’s ECoG recorded with depth and/or subdural electrodes had a case report form containing clinical information and the times of onset and termination for each seizure, as scored by the epileptologists at the contributing centers. For analyses purposes, all seizures originating from the same epileptogenic zone were lumped into one ‘‘observation’’ regardless of the total number of seizures from that zone; subjects with more than one independent epileptogenic zone contributed more than one ‘‘observation’’ provided seizures from each zone occurred during the monitoring session (e.g., subject 21 had 2 seizures originating from the right temporal lobe, 1 seizure from the left temporal and 1 seizure from the left frontal lobe for a total of 4 seizures, but contributed only 3 observations to the data because 2/4 seizures originated from the same zone). Only ECoGs containing clinical seizures and with good quality EKGs (determined through expert visual analysis) were included in the final analyses. Clinical seizures were defined as those having electrographic and clinical/behavioral manifestations. This study focuses on clinical seizures, because unlike electrographic or sub-clinical seizures, they pose serious risks to patients which may be managed through automated interventions (warnings and/or therapy delivery) triggered by heart rate-based detections. Seizures originated from mesial temporal structures in the majority of patients; epileptogenic tissue was localized to the frontal neocortices in a small number of subjects. All seizures were ‘‘focal’’ at onset and were not pre-selected based on the presence of heart rate changes. For safety and other reasons (e.g., video-ECoG signal quality), the subjects were sedentary for the duration of the recordings.
2.4. Statistical analyses The dependence of probability of detection (which in this study corresponds to True Positive detections) of clinical seizures using EKG on various factors such as age, seizure class, etc. was investigated by applying a logistic regression model to the seizure time series: P¼
expðb0 þ b1 X 1 þ . . . þ b p X p Þ f1 þ expðb0 þ b1 X 1 þ . . . þ b p X p Þg
(1)
where the detection probability P, is estimated based on the linear combination b0 + b1X1 ++ bpXp, corresponding to the following factors that typify the patient samples: (a) Data set (‘‘Full Stay’’ = 1; ‘‘Multi-Center’’ = 2); (b) Age group [{10–19} = 1; {20–29} = 2; {30– 39} = 3; {40–49} = 4; {50–59 years} = 5]; (c) Gender (Male = 1; Female = 2); (d) Etiology (Congenital = 1; Cryptogenic = 2; Infection = 3; RH factor incompatibility = 4; Trauma = 5; Tumor = 6; Vascular = 7); (e) Seizure class (Not Reported = 1; Complex partial = 2; Generalized Tonic-Clonic = 3; Simple Partial = 4); (f) Seizure topography (Right Temporal = 1; Left Temporal = 2; Right Extra-temporal = 3; Left Extra-temporal = 4; Other = 5); (g) Years with epilepsy, and (h) Electrode type (e.g., subdural). The effect of each of these factors on the variation in the observed proportions of seizures detected was investigated individually using the logistic regression model specified above, starting with models containing only factors such as age and gender that are known to influence heart rate, and increasing their complexity as required to better assess which ones contributed to the observed variation in the proportion of detected seizures, expressed as a probability. 3. Results 3.1. Model 1 Age, Gender and Data set origin were the factors (b) selected for this model since the first two markedly influence heart rate and data heterogeneity (given the differences in sample size and the large number of contributing centers) deserves investigation as a potentially important source of variation.
2.2. EKG-based seizure detection The EKG-based seizure detection algorithm ‘‘localizes’’ R waves and determines R-R sequences thus yielding a relative heart rate that is transformed into detection intervals [13]. EKG-based seizure detections were first classified as either True Positive (TP) or False Positive (FP) based on whether or not they were temporally correlated with visually validated ECoG detections, as determined by an epileptologist at each of the contributing centers. Each ECoG validated clinical seizure was used to define a time interval during which EKG-based detections overlapping this interval would be classified as TPs [13]. Validated ECoG detections that were not detected by the EKG algorithm were classified as False Negatives (FN) detections. Only TPs clinical seizures were included in the regression analysis.
3.1.1. Results Age and gender, not data set, were both significant contributors to variation in the proportion of detected clinical seizures with age being the most significant. Further model fitting revealed that a better model for the probability of detecting a seizure was a quadratic function of Age (Age2) and the factor Seizure Class. This model accounts for 33.5% of the variation in the data, which is very highly significant (p < 0.001), with the probability of detecting a seizure being much lower for simple than for complex partial seizures. 3.2. Model 2 In this model the observed proportions of clinical seizures detected was plotted against the factors Dataset, Age, Gender,
122
I. Osorio, B.F.J. Manly / Seizure 30 (2015) 120–123
Etiology, Seizure Class, Topography of Seizure onset, Electrode Type, and Years with Epilepsy. All possible factors were considered in this model since even though the first model yielded very highly significant (p < 0.001), the magnitude of data variation accounted for by three important factors (Gender, Age, Seizure class) was modest. 3.2.1. Results This model accounted for 50.3% in the variation of the proportion of detected seizures. The plots suggest that the probability of detecting a clinical seizure is, to a lesser or greater degree, a function of each of these variables.
3.4.1. Results The evidence shows that: (a) Average Szi is lower with simple partial than with other seizure classes; (b) the average SzDur is longer and SzSev higher in patients with Depth than with other Electrode types. There was no evidence that: (a) the average values of Szi, SzDur or SzSev change with the Age of patients or is a function of gender; (b) Szi, SzDur and SzSev are independent of etiology. In the ‘‘full stay’’ data set, the probability of detection was greater than 0.9 in 21 cases, between 0.81 and 0.9 in 11, between 0.51 and 0.8 in 3 cases and 0.5 or lower in 3 cases. In the ‘‘multicenter’’ group, the probability of detection was greater than >0.9 in 31 cases, between 0.81 and 0.9 in 3, between 0.5 and 0.8 in 4 and 0.5 or lower in 3 cases.
3.3. Selection of best model by fitting all possible models 4. Discussion The All Possible Subset Selection procedure in GenStat Version 16 was used to determine the model that best fitted the proportion of seizures detected using a validated algorithm [13,14]. To this end, the logistic model of equation (1) was fitted, where the linear combination consists of variables allowing for the effects of different factor levels or variables. For fitting this model, 79 observations were available for different patients and different classes of seizures. The best model was selected with the Akaike Information Criterion (AIC) [22] and the adjusted R2 value, which is the proportion of the variation in the dependent variable accounted for by the variables considered. Low AIC values and high percentages of variation accounted for by the factors under consideration indicate goodness of fit of a model. Since previous analyses revealed that the age of patients and the years with epilepsy are important variables, and that the squares of these variables might also have significant effects, these terms were included in all of the fitted models. 3.3.1. Results The best regression model includes the variables Age, Age2, Gender, Etiology, Seizure Class, Electrode Type and Years with Epilepsy. This model accounts for 62.6% of the variation in the observed proportions of clinical seizures detected, which is very highly significant (p < 0.001) and is much higher than the 33.5% of model 1 or the 50.3% accounted for by model 2. The best model shows that the probability of seizure detection: (a) is lower for females than for males; (b) is highest for infectious etiologies; (c) is highest for complex and lowest for simple partial seizures; (d) is highest for subjects implanted with subdural electrodes; (e) increases with a subject’s Years with epilepsy; and, (f) the effect of the ‘‘Age of subject’’ factor is non-linear. Interestingly, the site of epileptogenesis proved insignificant in shaping the probability of detection. 3.4. Probability of detection as a function of seizure intensity, duration and severity In the final analysis, the probability of seizure detection was related to the: (a) maximal ictal heart rate (Max HR; bpm) a valid surrogate for ECoG-based peak intensity (Szi) [14]; (b) time (in sec.) the HR spends above a ‘‘threshold’’ an indicator of duration (SzDur), and (c) the product of these two variables (MaxHR SzDur), a clinically meaningful measure of severity (SzSev) [14]. This was accomplished via randomization tests to assess the relation between Szi, SzD and SzSev and Age, Gender, Etiology, Seizure Class, Electrode Type and Years with Epilepsy; site of seizure onset was not included since this factor had no significant impact on the probability of detection. These tests compared the observed distribution of the differences in mean values of Szi, SzD and SzSev and Age as a function of the factor under consideration (age, seizure class, etc.) and its ‘‘control’’ set to have zero effect.
These results inform of the factors that shape the probability of detection of clinical seizures using heart rate and provide a framework for enhancing its applicability whose simplicity, ease of implementation into extra-cranial devices and cost-effectiveness, make it a valuable complement or alternative to ECoG/EEG. The significant (p < 0.001) dependency of the probability of detection of clinical seizures using heart rate on factors such as age and gender [23–26], should be taken into account in the design of algorithms for automated seizure detection since they are likely to introduce volatility into their performance. Explicitly, females have higher heart rates, cardiac output and stroke volume [26] than males so that if heart rates applicable to females are used as reference for seizure detection in males or vice-versa, degradation in seizure detection performance is likely to occur. That in this study, in which no adjustments were made for gender differences in heart rate, seizure detection probability was lower in females compared to males, points to the importance of gender-specific adaptations. The contribution of age to the variability in seizure detection probability is difficult to assess independently of the ‘‘years with epilepsy’’ factor, since they are co-variables; that is, the older an epileptic patient, the longer the history of epilepsy. These results reveal that the probability of seizure detection in adult patients increases with age and years with epilepsy, but since aging is associated with a decrease in heart rate [23] and with moderate attenuation of autonomic cardio-vascular responses [24], the length of epilepsy is probably the main factor accounting for the increased detection probability. One plausible interpretation of this finding is that over time, central and/or peripheral autonomic nervous system responses to epileptic seizures are potentiated. Indeed, mechanisms for long-term potentiation of autonomic ganglia that depend on serotonin for their induction and maintenance have been identified [27]. Moreover, neuroplastic changes within the nucleus tractus solitarius and the hypothalamic paraventricular nucleus reciprocal network, are responsible for cardio-vascular response adaptation in sedentary and in trained individuals [28]; this plasticity may also be induced by paroxysmal neuronal activity originating from or invading central autonomic network (CAN) structures. The higher probability of detection of complex partial compared to simple partial seizures using heart rate, likely reflects the larger volume of brain tissue engulfed by paroxysmal neuronal activity in the former [29]. Simply put, the likelihood of ictal involvement of any part of the CAN (and with it, the probability of ictal chronotropic changes) increases with the volume of tissue in the seizure state. Icti emerging from: (a) neuronal masses below a certain critical volume even if belonging to the CAN [for example, seizures restricted to one of the amygdalae do not cause tachycardia [29]; unpublished observations)] or, (b) outside the CAN and not spreading into a CAN ‘‘hub’’ or ‘‘node’’, will not be amenable to detection using heart rate. Judicious selection of
I. Osorio, B.F.J. Manly / Seizure 30 (2015) 120–123
subjects for automated seizure detection and quantification of frequency and severity, using heart rate changes, is important for the effective application of this modality. It is worth remarking that heart rate based seizure duration is overestimated in tonicclonic seizures since tachycardia persists beyond the termination of ictal activity due to profound metabolic acidosis, among other factors. The finding that the type of intracranial electrode (subdural vs. depth) used for surgical evaluation bears on the probability (higher for subjects with subdural than with depth electrodes) of EKGbased seizure detection may be at first glance, puzzling, if not irrelevant. But, the knowledge about the relation between tissue volume in the seizure state and the probability of appearance of chronotropic ictal manifestations provides clues. Selection of electrode type (subdural or depth) in the invasive surgical evaluation process is based at least partly, on the presumed number and size of epileptogenic zones; subjects with abundant epileptogenic tissue are investigated with subdural grids as it was the case in this study, whereas those with ‘‘scarce’’ tissue (e.g., amygdalo-hippocampal epilepsy) receive depth electrodes. The unequal distribution of sympathetic and parasympathetic neurons between the cerebral hemispheres [30–32] was not reflected in this study in inter-hemispheric differences in heart rate-based detection probability of clinical seizures. That the proportion of subjects with detection probabilities > 0.9 was higher in the ‘‘Multicenter’’ (31/33) than in the ‘‘Full Stay’’ dataset (21/39) is deserving of mention. Differences in Age (lower), Years with Epilepsy (higher) and more cases implanted with subdural than with depth electrodes in the ‘‘Multicenter ‘‘compared to the ‘‘Full Stay’’ group, account for the higher detection probability in the former compared to the latter. Being central to the maintenance of adequate blood perfusion to all organs under conditions of increased demand (e.g., from rest to exercise), heart rate increases accordingly. This response confounds the interpretation of changes in heart rate, when they are used to ascertain the occurrence of pathological phenomena with chronotropic effects such as epileptic seizures, since these occur less frequently (even in subjects with pharmaco-resistant epilepsies) than daily life activities also having chronotropic effects. The potential limitations imposed by the ubiquitous nature of heart rate changes in ambulatory human beings on automated seizure detection were not examined in this study since the subjects were, of necessity, sedentary. A recent study reveals that accurate distinction between exertional and ictal tachycardia is feasible [33]. 5. Conclusions The probability of clinical seizure detection using heart rate is shaped by various biological factors. The stability and robustness of ictal tachycardia over the course of epilepsy, makes heart rate changes, a useful signal for extra-cerebral automated detection of seizures and quantification of their frequency and severity. Conflict of interest Flint Hills Scientific LLC (Lawrence, KS) has licensed to Cyberonics Inc. (Houston, TX) intellectual property for seizure detection using EKG. Products sold using this technology generate royalties to Flint Hills and the University of Kansas Medical Center. I. Osorio is a member of Flint Hills and of the University of Kansas Medical Center and receives royalties only from Flint Hills. This work was undertaken without financial, human or technical
123
support from Cyberonics, Flint Hills or the University of Kansas Medical Center. BFJ Manly has no conflicts of interest to disclose. References [1] Bard P. Anatomical organization of the central nervous system in relation to control of the heart and vessels. Physiol Rev 1960;4:3–26. [2] Benarroch EE. The central autonomic network: functional organization, dysfunction and perspective. Mayo Clin Proc 1993;68:988–1001. [3] Buijs RM. The autonomic nervous system: a balancing act. Handb Clin Neurol 2013;117:1–11. [4] Korner PI. Integral neural cardiovascular control. Physiol Rev 1971;51:312–67. [5] Mraovitch S, Calando Y. Interactions between limbic, thalamo-striatal-cortical and central autonomic pathways during epileptic seizure progression. J Comp Neurol 1999;16:145–61. [6] Ulrich-Lai YM, Herman JP. Neural regulation of endocrine and autonomic responses. Nature Rev Neurosci 2009;10:397–409. [7] van Elmpt WJ, Nijsen TM, Griep PA, Arends JB. A model of heart rate changes to detect seizures in severe epilepsy. Seizure 2006;15:366–75. [8] Doyle OM, Greene BR, Marnane W, Lightbody G, Boylan GB. Characterisation of heart rate changes and their correlation with EEG during neonatal seizures. Conf Proc IEEE Eng Med Biol Soc 2008;498:4–7. [9] Greene BR, de Chazal P, Boylan GB, Connolly S, Reilly RB. Electrocardiogram based neonatal seizure detection. IEEE Trans Biomed Eng 2007;54:673–82. [10] Greene BR, Boylan GB, Reilly RB, de Chazal P, Connolly S. Combination of EEG and ECG for improved automatic neonatal seizure detection. Clin Neurophysiol 2007;118:1348–59. [11] Jansen K, Varon C, Van Huffel S, Lagae L. Peri-ictal ECG changes in childhood epilepsy: implications for detection systems. Epilepsy Behav 2013;29:72–6. [12] Malarvili MB, Mesbah M, Boashash B. Time-frequency analysis of heart rate variability for neonatal seizure detection. Australas Phys Eng Sci Med 2006;29:67–72. [13] Osorio I. Automated seizure detection using EKG. Int J Neural Syst Int J Neural Syst 2014;24:1450001. [14] Osorio I, Manly BJ. Is seizure detection based on EKG clinically relevant. Clin Neurophysiol 2014;125:1946–51. [15] Eggleston KS, Olin BD, Fisher RS. Ictal tachycardia: the head-heart connection. Seizure 2014;23:496–505. [16] Schernthaner C, Lindinger G, Po¨tzelberger K, Zeiler K, Baumgartner C. Autonomic epilepsy—the influence of epileptic discharges on heart rate and rhythm. Wien Klin Wochenschr 1999;111:392–401. [17] Hasan W. Autonomic cardiac innervation: development and adult plasticity. Organogenesis 2013;9:176–93. [18] Osorio I, Frei MG, Wilkinson SB. Real-time automated detection and quantitative analysis of seizures and short-term prediction of clinical onset. Epilepsia 1998;39:615–27. [19] Osorio I, Frei MG, Giftakis J, Peters T, Ingram J, Turnbull M, et al. Performance reassessment of real-time seizure-detection algorithm on long ECoG series. Epilepsia 2002;43:1522–35. [20] Osorio I, Frei MG, Sunderam S, Giftakis J, Bhavaraju NC, Schaffner SF, et al. Automated seizure abatement in humans using electrical stimulation. Ann Neurol 2005;57:258–68. [21] Osorio I, Manly BFJ, Sunderam S. Toward a quantitative multivariate analysis of the efficacy of antiseizure therapies. Epilepsy Behav 2010;18:335–43. [22] Akaike H. Information theory and an extension of the maximum likelihood principle. In: Petran BN, Csaaki F, editors. International symposium on information theory, 2nd ed. Budapest: Akadeemiai Kiadi; 1973. p. 267–81. [23] Petrofsky JS, Burse RL, Lind AR. Comparison of physiological responses of women and men to isometric exercise. J Appl Physiol 1975;38:863–8. [24] Kaijser L, Sachs C. Autonomic cardiovascular responses in old age. Clin Physiol 1985;5:347–57. [25] Ryan SM, Goldberger AL, Pincus SM, Mietus J, Lipsitz LA. Gender- and agerelated differences in heart rate dynamics: are women more complex than men. J Am Coll Cardiol 1994;24:1700–7. [26] Wittnich C, Tan L, Wallen J, Belanger M. Sex differences in myocardial metabolism and cardiac function: an emerging concept. Pflugers Arch 2013;465:719–29. [27] Alkadhi KA, Alzoubi KH, Aleisa AM. Plasticity of synaptic transmission in autonomic ganglia. Prog Neurobiol 2005;75:83–108. [28] Michelini LC, Stern JE. Exercise-induced neuronal plasticity in central autonomic networks: role in cardiovascular control. Exp Physiol 2009;94:947–60. [29] Epstein MA, Sperling MR, O’Connor MJ. Cardiac rhythm during temporal lobe seizures. Neurology 1992;42:50–3. [30] Oppenheimer SM, Gelb A, Girvin JP, Hachinski VC. Cardiovascular effects of human insular cortex stimulation. Neurology 1992;42:1727–32. [31] Oppenheimer S. The anatomy and physiology of cortical mechanisms of cardiac control. Stroke 1993;24(12 (Suppl.)):I3–15. [32] Oppenheimer S. Cerebrogenic cardiac arrhythmias: cortical lateralization and clinical significance. Clin Auton Res 2006;16:6–11. [33] Jeppesen J, Beniczky S, Johansen P, Sidenius P, Fuglsang-Frederiken A. Detection of epileptic seizures with a modified heart rate variability algorithm based on Lorenz plot. Seizure 2015;24:1–7.
Contents lists available at ScienceDirect
Seizure journal homepage: www.elsevier.com/locate/yseiz
Probability of detection of clinical seizures using heart rate changes Ivan Osorio a,*, B.F.J. Manly b a b
University of Kansas Medical Center, United States Manly-Biostatistics Ltd., United States
A R T I C L E I N F O
A B S T R A C T
Article history: Received 12 May 2015 Received in revised form 18 June 2015 Accepted 21 June 2015
Purpose: Heart rate-based seizure detection is a viable complement or alternative to ECoG/EEG. This study investigates the role of various biological factors on the probability of clinical seizure detection using heart rate. Methods: Regression models were applied to 266 clinical seizures recorded from 72 subjects to investigate if factors such as age, gender, years with epilepsy, etiology, seizure site origin, seizure class, and data collection centers, among others, shape the probability of EKG-based seizure detection. Results: Clinical seizure detection probability based on heart rate changes, is significantly (p < 0.001) shaped by patients’ age and gender, seizure class, and years with epilepsy. The probability of detecting clinical seizures (>0.8 in the majority of subjects) using heart rate is highest for complex partial seizures, increases with a patient’s years with epilepsy, is lower for females than for males and is unrelated to the side of hemisphere origin. Conclusion: Clinical seizure detection probability using heart rate is multi-factorially dependent and sufficiently high (>0.8) in most cases to be clinically useful. Knowledge of the role that these factors play in shaping said probability will enhance its applicability and usefulness. Heart rate is a reliable and practical signal for extra-cerebral detection of clinical seizures originating from or spreading to central autonomic network structures. ß 2015 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
Keywords: Clinical seizures Heart rate based detection Detection probability Factors/variables Regression models
1. Introduction Automated detection of epileptic seizures originating from or spreading to central autonomic network (CAN) structures [1–6] using heart rate changes [7–15], is gaining acceptance as either a valuable complement or as an alternative to cortical electrical signals (e.g., ECoG) due to their: (a) Higher signal-to-noise ratio (mV for EKG vs. mV ECoG; (b) Greater ease and economy of recording (EKG requires 2 electrodes for indirect but all encompassing monitoring of ictal activity in CAN structures, while the same number of scalp or intracranial electrodes would yield inadequate coverage [16]; (c) Lower processing and computational analysis cost [13] due to EKG’s lower signal complexity than ECoG. Since seizure detection using heart rate is in an early developmental stage, more knowledge about its power and reliability is desirable to meaningfully assess its clinical applicability. The magnitude of relative changes in heart rate at seizure
* Corresponding author at: 3901 Rainbow Blvd, Kansas City, KS 66160, United States. Tel.: +1 913 5884529; fax: +1 913 5884585. E-mail addresses: [email protected], [email protected] (I. Osorio).
onset, is the single most important biological factor governing the probability of detection using this variable [13], which is subject not only to fluctuations in type and level of physical activity, but also to various other factors. The degree to which age, gender, seizure class (e.g., simple vs. complex partial), etiology (e.g., lesional vs. cryptogenic) or hemispheric location (left vs. right) of the epileptogenic zone may impact the probability of clinical seizure detection is unknown. Also, the plasticity (e.g., accommodation or potentiation) of the central and peripheral autonomic system [17] raises relevant questions such as: Does the intensity of the ictal cardiac response depend on the length of the course of epilepsy? If yes, is it enhanced or blunted? The importance of identifying which factors and the degree to which they shape the cardiac response to seizures and through it, the probability of detection of clinical seizures, motivates this investigation. To this end, the probability of detection (which in this study corresponds to True Positive detections) of clinical seizures using heart rate changes was investigated as a function of: Subjects’ age and gender, years with epilepsy, etiology, site of origin, seizure class, electrode type (subdural vs. depth) used for localization of the epileptogenic zone, and data collection epilepsy centers.
http://dx.doi.org/10.1016/j.seizure.2015.06.007 1059-1311/ß 2015 British Epilepsy Association. Published by Elsevier Ltd. All rights reserved.
I. Osorio, B.F.J. Manly / Seizure 30 (2015) 120–123
121
2. Methods and materials
2.3. EKG-based seizure intensity, duration and severity
2.1. Data sets
Measures of intensity, duration and severity for seizures detected using ECoG [20,21] were applied to seizures detected using heart rate changes, since they are statistically equivalent and thus clinically meaningful [14]. (a) Seizure intensity (Szi) was defined as the maximal ictal heart rate (Max HR; bpm) recorded during an event; (b) Seizure duration (SzDur) was defined as the time (in s) the HR spends above a ‘‘threshold’’), and (c) Seizure severity (SzSev) as the product of MaxHR and SzDur.
This investigation utilizes the data of an extensive validation of an EKG-based seizure detection algorithm published elsewhere [13,14]. Only details deemed necessary to properly frame this work will be provided herein. Seventy-nine ECoGs (72 subjects; 6935 h, 266 clinical seizures) with simultaneous EKG were analyzed with two different seizure detection algorithms: One was applied to ECoG [18,19] and the other to EKG [13,14]. These data, from which all identifiers had been removed, were collected with approval from the Human Subjects Committee of each of the contributing centers. One data set (33 subjects; 98 clinical seizures; 4–8 h recordings; total duration 352 h) was contributed by the Minnesota Epilepsy Group, Montreal Neurological Institute, Stanford Medical Center and the University of Kansas Medical Center (‘‘Multi-Center’’) and the other (39 subjects, 168 clinical seizures; full stay recordings, 6583 h) by the University of Kansas Medical Center (‘‘Full Stay’’). Each subject’s ECoG recorded with depth and/or subdural electrodes had a case report form containing clinical information and the times of onset and termination for each seizure, as scored by the epileptologists at the contributing centers. For analyses purposes, all seizures originating from the same epileptogenic zone were lumped into one ‘‘observation’’ regardless of the total number of seizures from that zone; subjects with more than one independent epileptogenic zone contributed more than one ‘‘observation’’ provided seizures from each zone occurred during the monitoring session (e.g., subject 21 had 2 seizures originating from the right temporal lobe, 1 seizure from the left temporal and 1 seizure from the left frontal lobe for a total of 4 seizures, but contributed only 3 observations to the data because 2/4 seizures originated from the same zone). Only ECoGs containing clinical seizures and with good quality EKGs (determined through expert visual analysis) were included in the final analyses. Clinical seizures were defined as those having electrographic and clinical/behavioral manifestations. This study focuses on clinical seizures, because unlike electrographic or sub-clinical seizures, they pose serious risks to patients which may be managed through automated interventions (warnings and/or therapy delivery) triggered by heart rate-based detections. Seizures originated from mesial temporal structures in the majority of patients; epileptogenic tissue was localized to the frontal neocortices in a small number of subjects. All seizures were ‘‘focal’’ at onset and were not pre-selected based on the presence of heart rate changes. For safety and other reasons (e.g., video-ECoG signal quality), the subjects were sedentary for the duration of the recordings.
2.4. Statistical analyses The dependence of probability of detection (which in this study corresponds to True Positive detections) of clinical seizures using EKG on various factors such as age, seizure class, etc. was investigated by applying a logistic regression model to the seizure time series: P¼
expðb0 þ b1 X 1 þ . . . þ b p X p Þ f1 þ expðb0 þ b1 X 1 þ . . . þ b p X p Þg
(1)
where the detection probability P, is estimated based on the linear combination b0 + b1X1 ++ bpXp, corresponding to the following factors that typify the patient samples: (a) Data set (‘‘Full Stay’’ = 1; ‘‘Multi-Center’’ = 2); (b) Age group [{10–19} = 1; {20–29} = 2; {30– 39} = 3; {40–49} = 4; {50–59 years} = 5]; (c) Gender (Male = 1; Female = 2); (d) Etiology (Congenital = 1; Cryptogenic = 2; Infection = 3; RH factor incompatibility = 4; Trauma = 5; Tumor = 6; Vascular = 7); (e) Seizure class (Not Reported = 1; Complex partial = 2; Generalized Tonic-Clonic = 3; Simple Partial = 4); (f) Seizure topography (Right Temporal = 1; Left Temporal = 2; Right Extra-temporal = 3; Left Extra-temporal = 4; Other = 5); (g) Years with epilepsy, and (h) Electrode type (e.g., subdural). The effect of each of these factors on the variation in the observed proportions of seizures detected was investigated individually using the logistic regression model specified above, starting with models containing only factors such as age and gender that are known to influence heart rate, and increasing their complexity as required to better assess which ones contributed to the observed variation in the proportion of detected seizures, expressed as a probability. 3. Results 3.1. Model 1 Age, Gender and Data set origin were the factors (b) selected for this model since the first two markedly influence heart rate and data heterogeneity (given the differences in sample size and the large number of contributing centers) deserves investigation as a potentially important source of variation.
2.2. EKG-based seizure detection The EKG-based seizure detection algorithm ‘‘localizes’’ R waves and determines R-R sequences thus yielding a relative heart rate that is transformed into detection intervals [13]. EKG-based seizure detections were first classified as either True Positive (TP) or False Positive (FP) based on whether or not they were temporally correlated with visually validated ECoG detections, as determined by an epileptologist at each of the contributing centers. Each ECoG validated clinical seizure was used to define a time interval during which EKG-based detections overlapping this interval would be classified as TPs [13]. Validated ECoG detections that were not detected by the EKG algorithm were classified as False Negatives (FN) detections. Only TPs clinical seizures were included in the regression analysis.
3.1.1. Results Age and gender, not data set, were both significant contributors to variation in the proportion of detected clinical seizures with age being the most significant. Further model fitting revealed that a better model for the probability of detecting a seizure was a quadratic function of Age (Age2) and the factor Seizure Class. This model accounts for 33.5% of the variation in the data, which is very highly significant (p < 0.001), with the probability of detecting a seizure being much lower for simple than for complex partial seizures. 3.2. Model 2 In this model the observed proportions of clinical seizures detected was plotted against the factors Dataset, Age, Gender,
122
I. Osorio, B.F.J. Manly / Seizure 30 (2015) 120–123
Etiology, Seizure Class, Topography of Seizure onset, Electrode Type, and Years with Epilepsy. All possible factors were considered in this model since even though the first model yielded very highly significant (p < 0.001), the magnitude of data variation accounted for by three important factors (Gender, Age, Seizure class) was modest. 3.2.1. Results This model accounted for 50.3% in the variation of the proportion of detected seizures. The plots suggest that the probability of detecting a clinical seizure is, to a lesser or greater degree, a function of each of these variables.
3.4.1. Results The evidence shows that: (a) Average Szi is lower with simple partial than with other seizure classes; (b) the average SzDur is longer and SzSev higher in patients with Depth than with other Electrode types. There was no evidence that: (a) the average values of Szi, SzDur or SzSev change with the Age of patients or is a function of gender; (b) Szi, SzDur and SzSev are independent of etiology. In the ‘‘full stay’’ data set, the probability of detection was greater than 0.9 in 21 cases, between 0.81 and 0.9 in 11, between 0.51 and 0.8 in 3 cases and 0.5 or lower in 3 cases. In the ‘‘multicenter’’ group, the probability of detection was greater than >0.9 in 31 cases, between 0.81 and 0.9 in 3, between 0.5 and 0.8 in 4 and 0.5 or lower in 3 cases.
3.3. Selection of best model by fitting all possible models 4. Discussion The All Possible Subset Selection procedure in GenStat Version 16 was used to determine the model that best fitted the proportion of seizures detected using a validated algorithm [13,14]. To this end, the logistic model of equation (1) was fitted, where the linear combination consists of variables allowing for the effects of different factor levels or variables. For fitting this model, 79 observations were available for different patients and different classes of seizures. The best model was selected with the Akaike Information Criterion (AIC) [22] and the adjusted R2 value, which is the proportion of the variation in the dependent variable accounted for by the variables considered. Low AIC values and high percentages of variation accounted for by the factors under consideration indicate goodness of fit of a model. Since previous analyses revealed that the age of patients and the years with epilepsy are important variables, and that the squares of these variables might also have significant effects, these terms were included in all of the fitted models. 3.3.1. Results The best regression model includes the variables Age, Age2, Gender, Etiology, Seizure Class, Electrode Type and Years with Epilepsy. This model accounts for 62.6% of the variation in the observed proportions of clinical seizures detected, which is very highly significant (p < 0.001) and is much higher than the 33.5% of model 1 or the 50.3% accounted for by model 2. The best model shows that the probability of seizure detection: (a) is lower for females than for males; (b) is highest for infectious etiologies; (c) is highest for complex and lowest for simple partial seizures; (d) is highest for subjects implanted with subdural electrodes; (e) increases with a subject’s Years with epilepsy; and, (f) the effect of the ‘‘Age of subject’’ factor is non-linear. Interestingly, the site of epileptogenesis proved insignificant in shaping the probability of detection. 3.4. Probability of detection as a function of seizure intensity, duration and severity In the final analysis, the probability of seizure detection was related to the: (a) maximal ictal heart rate (Max HR; bpm) a valid surrogate for ECoG-based peak intensity (Szi) [14]; (b) time (in sec.) the HR spends above a ‘‘threshold’’ an indicator of duration (SzDur), and (c) the product of these two variables (MaxHR SzDur), a clinically meaningful measure of severity (SzSev) [14]. This was accomplished via randomization tests to assess the relation between Szi, SzD and SzSev and Age, Gender, Etiology, Seizure Class, Electrode Type and Years with Epilepsy; site of seizure onset was not included since this factor had no significant impact on the probability of detection. These tests compared the observed distribution of the differences in mean values of Szi, SzD and SzSev and Age as a function of the factor under consideration (age, seizure class, etc.) and its ‘‘control’’ set to have zero effect.
These results inform of the factors that shape the probability of detection of clinical seizures using heart rate and provide a framework for enhancing its applicability whose simplicity, ease of implementation into extra-cranial devices and cost-effectiveness, make it a valuable complement or alternative to ECoG/EEG. The significant (p < 0.001) dependency of the probability of detection of clinical seizures using heart rate on factors such as age and gender [23–26], should be taken into account in the design of algorithms for automated seizure detection since they are likely to introduce volatility into their performance. Explicitly, females have higher heart rates, cardiac output and stroke volume [26] than males so that if heart rates applicable to females are used as reference for seizure detection in males or vice-versa, degradation in seizure detection performance is likely to occur. That in this study, in which no adjustments were made for gender differences in heart rate, seizure detection probability was lower in females compared to males, points to the importance of gender-specific adaptations. The contribution of age to the variability in seizure detection probability is difficult to assess independently of the ‘‘years with epilepsy’’ factor, since they are co-variables; that is, the older an epileptic patient, the longer the history of epilepsy. These results reveal that the probability of seizure detection in adult patients increases with age and years with epilepsy, but since aging is associated with a decrease in heart rate [23] and with moderate attenuation of autonomic cardio-vascular responses [24], the length of epilepsy is probably the main factor accounting for the increased detection probability. One plausible interpretation of this finding is that over time, central and/or peripheral autonomic nervous system responses to epileptic seizures are potentiated. Indeed, mechanisms for long-term potentiation of autonomic ganglia that depend on serotonin for their induction and maintenance have been identified [27]. Moreover, neuroplastic changes within the nucleus tractus solitarius and the hypothalamic paraventricular nucleus reciprocal network, are responsible for cardio-vascular response adaptation in sedentary and in trained individuals [28]; this plasticity may also be induced by paroxysmal neuronal activity originating from or invading central autonomic network (CAN) structures. The higher probability of detection of complex partial compared to simple partial seizures using heart rate, likely reflects the larger volume of brain tissue engulfed by paroxysmal neuronal activity in the former [29]. Simply put, the likelihood of ictal involvement of any part of the CAN (and with it, the probability of ictal chronotropic changes) increases with the volume of tissue in the seizure state. Icti emerging from: (a) neuronal masses below a certain critical volume even if belonging to the CAN [for example, seizures restricted to one of the amygdalae do not cause tachycardia [29]; unpublished observations)] or, (b) outside the CAN and not spreading into a CAN ‘‘hub’’ or ‘‘node’’, will not be amenable to detection using heart rate. Judicious selection of
I. Osorio, B.F.J. Manly / Seizure 30 (2015) 120–123
subjects for automated seizure detection and quantification of frequency and severity, using heart rate changes, is important for the effective application of this modality. It is worth remarking that heart rate based seizure duration is overestimated in tonicclonic seizures since tachycardia persists beyond the termination of ictal activity due to profound metabolic acidosis, among other factors. The finding that the type of intracranial electrode (subdural vs. depth) used for surgical evaluation bears on the probability (higher for subjects with subdural than with depth electrodes) of EKGbased seizure detection may be at first glance, puzzling, if not irrelevant. But, the knowledge about the relation between tissue volume in the seizure state and the probability of appearance of chronotropic ictal manifestations provides clues. Selection of electrode type (subdural or depth) in the invasive surgical evaluation process is based at least partly, on the presumed number and size of epileptogenic zones; subjects with abundant epileptogenic tissue are investigated with subdural grids as it was the case in this study, whereas those with ‘‘scarce’’ tissue (e.g., amygdalo-hippocampal epilepsy) receive depth electrodes. The unequal distribution of sympathetic and parasympathetic neurons between the cerebral hemispheres [30–32] was not reflected in this study in inter-hemispheric differences in heart rate-based detection probability of clinical seizures. That the proportion of subjects with detection probabilities > 0.9 was higher in the ‘‘Multicenter’’ (31/33) than in the ‘‘Full Stay’’ dataset (21/39) is deserving of mention. Differences in Age (lower), Years with Epilepsy (higher) and more cases implanted with subdural than with depth electrodes in the ‘‘Multicenter ‘‘compared to the ‘‘Full Stay’’ group, account for the higher detection probability in the former compared to the latter. Being central to the maintenance of adequate blood perfusion to all organs under conditions of increased demand (e.g., from rest to exercise), heart rate increases accordingly. This response confounds the interpretation of changes in heart rate, when they are used to ascertain the occurrence of pathological phenomena with chronotropic effects such as epileptic seizures, since these occur less frequently (even in subjects with pharmaco-resistant epilepsies) than daily life activities also having chronotropic effects. The potential limitations imposed by the ubiquitous nature of heart rate changes in ambulatory human beings on automated seizure detection were not examined in this study since the subjects were, of necessity, sedentary. A recent study reveals that accurate distinction between exertional and ictal tachycardia is feasible [33]. 5. Conclusions The probability of clinical seizure detection using heart rate is shaped by various biological factors. The stability and robustness of ictal tachycardia over the course of epilepsy, makes heart rate changes, a useful signal for extra-cerebral automated detection of seizures and quantification of their frequency and severity. Conflict of interest Flint Hills Scientific LLC (Lawrence, KS) has licensed to Cyberonics Inc. (Houston, TX) intellectual property for seizure detection using EKG. Products sold using this technology generate royalties to Flint Hills and the University of Kansas Medical Center. I. Osorio is a member of Flint Hills and of the University of Kansas Medical Center and receives royalties only from Flint Hills. This work was undertaken without financial, human or technical
123
support from Cyberonics, Flint Hills or the University of Kansas Medical Center. BFJ Manly has no conflicts of interest to disclose. References [1] Bard P. Anatomical organization of the central nervous system in relation to control of the heart and vessels. Physiol Rev 1960;4:3–26. [2] Benarroch EE. The central autonomic network: functional organization, dysfunction and perspective. Mayo Clin Proc 1993;68:988–1001. [3] Buijs RM. The autonomic nervous system: a balancing act. Handb Clin Neurol 2013;117:1–11. [4] Korner PI. Integral neural cardiovascular control. Physiol Rev 1971;51:312–67. [5] Mraovitch S, Calando Y. Interactions between limbic, thalamo-striatal-cortical and central autonomic pathways during epileptic seizure progression. J Comp Neurol 1999;16:145–61. [6] Ulrich-Lai YM, Herman JP. Neural regulation of endocrine and autonomic responses. Nature Rev Neurosci 2009;10:397–409. [7] van Elmpt WJ, Nijsen TM, Griep PA, Arends JB. A model of heart rate changes to detect seizures in severe epilepsy. Seizure 2006;15:366–75. [8] Doyle OM, Greene BR, Marnane W, Lightbody G, Boylan GB. Characterisation of heart rate changes and their correlation with EEG during neonatal seizures. Conf Proc IEEE Eng Med Biol Soc 2008;498:4–7. [9] Greene BR, de Chazal P, Boylan GB, Connolly S, Reilly RB. Electrocardiogram based neonatal seizure detection. IEEE Trans Biomed Eng 2007;54:673–82. [10] Greene BR, Boylan GB, Reilly RB, de Chazal P, Connolly S. Combination of EEG and ECG for improved automatic neonatal seizure detection. Clin Neurophysiol 2007;118:1348–59. [11] Jansen K, Varon C, Van Huffel S, Lagae L. Peri-ictal ECG changes in childhood epilepsy: implications for detection systems. Epilepsy Behav 2013;29:72–6. [12] Malarvili MB, Mesbah M, Boashash B. Time-frequency analysis of heart rate variability for neonatal seizure detection. Australas Phys Eng Sci Med 2006;29:67–72. [13] Osorio I. Automated seizure detection using EKG. Int J Neural Syst Int J Neural Syst 2014;24:1450001. [14] Osorio I, Manly BJ. Is seizure detection based on EKG clinically relevant. Clin Neurophysiol 2014;125:1946–51. [15] Eggleston KS, Olin BD, Fisher RS. Ictal tachycardia: the head-heart connection. Seizure 2014;23:496–505. [16] Schernthaner C, Lindinger G, Po¨tzelberger K, Zeiler K, Baumgartner C. Autonomic epilepsy—the influence of epileptic discharges on heart rate and rhythm. Wien Klin Wochenschr 1999;111:392–401. [17] Hasan W. Autonomic cardiac innervation: development and adult plasticity. Organogenesis 2013;9:176–93. [18] Osorio I, Frei MG, Wilkinson SB. Real-time automated detection and quantitative analysis of seizures and short-term prediction of clinical onset. Epilepsia 1998;39:615–27. [19] Osorio I, Frei MG, Giftakis J, Peters T, Ingram J, Turnbull M, et al. Performance reassessment of real-time seizure-detection algorithm on long ECoG series. Epilepsia 2002;43:1522–35. [20] Osorio I, Frei MG, Sunderam S, Giftakis J, Bhavaraju NC, Schaffner SF, et al. Automated seizure abatement in humans using electrical stimulation. Ann Neurol 2005;57:258–68. [21] Osorio I, Manly BFJ, Sunderam S. Toward a quantitative multivariate analysis of the efficacy of antiseizure therapies. Epilepsy Behav 2010;18:335–43. [22] Akaike H. Information theory and an extension of the maximum likelihood principle. In: Petran BN, Csaaki F, editors. International symposium on information theory, 2nd ed. Budapest: Akadeemiai Kiadi; 1973. p. 267–81. [23] Petrofsky JS, Burse RL, Lind AR. Comparison of physiological responses of women and men to isometric exercise. J Appl Physiol 1975;38:863–8. [24] Kaijser L, Sachs C. Autonomic cardiovascular responses in old age. Clin Physiol 1985;5:347–57. [25] Ryan SM, Goldberger AL, Pincus SM, Mietus J, Lipsitz LA. Gender- and agerelated differences in heart rate dynamics: are women more complex than men. J Am Coll Cardiol 1994;24:1700–7. [26] Wittnich C, Tan L, Wallen J, Belanger M. Sex differences in myocardial metabolism and cardiac function: an emerging concept. Pflugers Arch 2013;465:719–29. [27] Alkadhi KA, Alzoubi KH, Aleisa AM. Plasticity of synaptic transmission in autonomic ganglia. Prog Neurobiol 2005;75:83–108. [28] Michelini LC, Stern JE. Exercise-induced neuronal plasticity in central autonomic networks: role in cardiovascular control. Exp Physiol 2009;94:947–60. [29] Epstein MA, Sperling MR, O’Connor MJ. Cardiac rhythm during temporal lobe seizures. Neurology 1992;42:50–3. [30] Oppenheimer SM, Gelb A, Girvin JP, Hachinski VC. Cardiovascular effects of human insular cortex stimulation. Neurology 1992;42:1727–32. [31] Oppenheimer S. The anatomy and physiology of cortical mechanisms of cardiac control. Stroke 1993;24(12 (Suppl.)):I3–15. [32] Oppenheimer S. Cerebrogenic cardiac arrhythmias: cortical lateralization and clinical significance. Clin Auton Res 2006;16:6–11. [33] Jeppesen J, Beniczky S, Johansen P, Sidenius P, Fuglsang-Frederiken A. Detection of epileptic seizures with a modified heart rate variability algorithm based on Lorenz plot. Seizure 2015;24:1–7.