Epileptic Seizure Detection Using Wristworn Biosensors* D. Cogan1 , M. Nourani1 , J. Harvey2 and V. Nagaraddi2
Abstract— Single signal seizure detection algorithms suffer from high false positive rates. We have found a set of signals which can be easily monitored by a wristworn device and which produce a distinctive pattern during seizure for patients in an epilepsy monitoring unit (EMU). This pattern is much less likely to be reproduced by nonseizure events in the patient’s daily life than are changes in heart rate alone. We collected 108 hours of data from three EMU patients who suffered a combined total of seven seizures, then developed a time series analysis/pattern recognition based algorithm which distinguishes the seizures from nonseizure events with 100% accuracy.
I. I NTRODUCTION A. Background Seizure prediction or detection would improve the quality of life of millions of patients who suffer from uncontrollable seizures. Prediction would be ideal, as the patient would have opportunity to remove himself from dangerous situations; but with the current state of knowledge, it has proven to be an elusive goal [1][2][3]. Detection is a challenging but more attainable goal and offers two benefits. First, a caregiver can be notified, which is important because seizures can lead to injury or other complications including SUDEP (sudden unexpected death in epilepsy). Second, an electronic diary of seizure events can be created, which would be a boon for physicians as self-reporting of seizures has proven to be inaccurate [4][5]. EEG is the gold standard for seizure detection, but it is impractical for use in daily life because it requires the use of either a special headset worn in just the right way or sensors implanted in the brain. Neither of these options is convenient, and implanting a device in the brain is both dangerous and costly. There are a number of devices on the market designed to detect convulsive seizures using extracerebral (nonEEG) signals [4]. To the best of our knowledge, however, there is no wearable device available to detect complex partial seizures, as they are nonconvulsive. Recent work using heart rate (HR) to detect seizures has achieved fairly accurate results for patients in epilepsy monitoring units (EMUs), but the authors concede that there is much room for improvement to their algorithms [5][6]. In normal daily life, we must expect a greater variation in patient activity, and resulting heart rate changes greater than those typically found in an EMU. *This work is supported in part by Texas Medical Research Collaborative (TxMRC) 1 D. Cogan and M. Nourani are with Quality of Life Technology Laboratory, The University of Texas at Dallas, Richardson, TX 75080,
{diana.cogan, nourani}@utdallas.edu 2 J. Texas
Harvey, Epilepsy
DO and V. Group, Dallas,
Nagaraddi, MD are with TX 75230 {jhharvey,
dr.nagaraddi}@texasepilepsy.org
978-1-4244-9270-1/15/$31.00 ©2015 IEEE
B. Motivation and Contribution We chose a set of five biosignals which can be monitored at the wrist and are known or believed to be affected by seizures: heart rate (HR), arterial oxygenation (SpO2 ), accelerometry (ACC), electrodermal activity (EDA) and temperature (Temp). A number of researchers have observed extreme HR changes at the beginning of some types of seizures [7]. Researchers have also found indications that seizures may cause changes in SpO2 levels [8]. Their findings are supported by studies showing that seizures can cause disruption in patient respiration [9]. Motion measured by ACC and changes in EDA have been found to effectively indicate the onset of convulsive seizures [10]. Temp is associated with febrile seizures but remains to be investigated as a possible means for detection [4]. Several seizure sensitive biosignals, such as respiration and electromyography, cannot be monitored at the wrist and so are excluded in this work. Previously, we reported using these five signals to distinguish physical, cognitive and emotional stresses [11]. Recently, we found a pattern created by HR, SpO2 and EDA during all seven of the seizures suffered by three patients monitored in an EMU. This signal pattern is less likely to be reproduced by nonseizure events during daily life than are changes in HR alone. We developed an algorithm which (i) extracts features via a time series analysis of each signal, (ii) locates times of possible seizures and (iii) uses pattern recognition techniques to distinguish seizures from nonseizure events. Our algorithm has processed over 108 hours of data collected from the three patients and distinguished the seizures from nonseizure events with 100% accuracy. II. S EIZURE D ETECTION U SING A W EARABLE B IOSENSOR S YSTEM The flowchart for our time series analysis/pattern recognition based algorithm described below is shown in Fig. 1. A. Phase 1 - Preprocessing Each of the three signals of interest (HR, SpO2 and EDA) is analyzed separately using a 5 second window with 80% overlap. The mean of each window is compared to a moving baseline (THR , TSpO2 , and TEDA in Fig. 1), then used to update the baseline. If the mean of a window varies from the baseline by a prespecified parameter level (∆HR , ∆SpO2 , ∆EDA ), the time of the window is tagged and recorded. The parameter levels considered are: • 15%, 20%, 25% and 30% increases for HR, based on Osorio’s finding that these four levels are useful for detecting seizures [5].
5086
Fig. 2.
Windows for Event Search and Pattern Recognition.
The Event Search finds times when all three signals were active, lists them as potential seizure events and passes them to Phase 3 for comparison with the HR↑⇒ SpO2 ↓⇒ EDA↑ seizure pattern. The search methodology is illustrated in Fig. 2. SpO2 is the most stable of the three signals; therefore, for each time the selected ∆SpO2 was recorded in the SpO2 activity list from Phase 1 (each t∆SpO2 in Fig. 2), the HR and EDA activity lists are searched for matches. If ∆HR is found within the ∆HR Search Window and ∆EDA is found within the ∆EDA Search Window, all three signals were active simultaneously and t∆SpO2 represents a potential seizure. C. Phase 3 - Pattern Recognition
Fig. 1.
Seizure Detection Algorithm Flowchart.
5%, 7.5%, 10% and 12.5% decreases for SpO2 , chosen because apnea is defined as a 4% drop in SpO2 for 8 seconds or more. Normal SpO2 levels are between 95% and 100%; readings below 90% are low [12]. Consequently, a 5% drop in SpO2 is significant. • 60%, 80%, 100% and 120% increases for EDA, selected by analysis of the available data. Future findings may require modification of some or all parameter levels. These biosignals take time to settle after a seizure. Hence, a long baseline is needed to detect seizures that are close together. We are currently using a 60 second baseline for all three signals (THR = TSpO2 = TEDA = 60). The first and last time of a series of sequential windows which shows a signal change at the selected parameter level is passed to the Event Search in a signal activity list. •
B. Phase 2 - Personalization and Event Search Personalization is done during training, when EEG based seizure timing is available. The highest parameter level for each signal which places it in its activity list (Phase 1) at all seizure times is chosen to optimize algorithm performance. The parameters are selected by an iterative process through the Training Loop shown in Fig. 1.
The methodology of this phase is illustrated in Fig. 2 and Fig. 3. Fig. 3 illustrates the HR↑⇒ SpO2 ↓⇒ EDA↑ up seizure pattern using data from one seizure. For each potential seizure event: 1) A 30 minute Event Window centered about t∆SpO2 is selected. A Local SpO2 Baseline is calculated using the first 10 minutes of this window. 2) tHR , the time HR begins rising, is found within the HR Search Window. 3) td , the time SpO2 drops 2% below its Local Baseline and tr , the time it returns to halfway between its min value and its value at td are found within the SpO2 Search Window. The value of 2% was empirically selected based on the behavior of SpO2 during seizures. 4) tEDA , the time of the EDA max value is found within the EDA Search Window. Note that t∆SpO2 is shown in both Fig. 2 and Fig. 3. As the HR rise is not smooth, the data is windowed to backtrack from the event time (t∆SpO2 ) found in Phase 2 to tHR . tHR is the time of the window whose mean is greater than or equal to the means of both subsequent windows. The size of the window was adjusted so that the time found by the algorithm matched the time found by a visual inspection of the HR curve for all seven seizures. The SpO2 drop occurs so quickly that our algorithm can follow the curve without windowing the data. The bottom of the curve is not always smooth, however, so we search for a lower SpO2 value within 30 seconds of the apparent bottom and replace it if appropriate. We selected a 30 second search window because in all seven seizures the SpO2 begins rising within 30 seconds.
5087
[13]. Data is being collected from patients by means of two commercially available devices: (i) a Nonin WristOx 3150, which monitors HR and SpO2 [14] and (ii) an Affectiva Q Curve, which monitors EDA, Temp and ACC [15]. Our devices are time synced to the EEG equipment so we are certain of seizure onset and offset times relative to our biosignals. We collected data from five patients who had a total of 12 seizures during their stay in the EMU. The three patients who experienced the HR↑⇒ SpO2 ↓⇒ EDA↑ seizure pattern suffered, collectively, 2 secondarily generalized and 5 complex partial seizures. We implemented the algorithm to analyze their 108 hours of data in Matlab [16]. Fig. 3.
The HR↑⇒ SpO2 ↓⇒ EDA↑ Seizure Pattern.
B. Observations After finding the critical points, we calculate (i) tHR−SpO2 = td − tHR and (ii) tSpO2−EDA = tEDA − tr . If either of these calculations returns a negative number, the pattern for the event does not match the seizure pattern and the event is classified as a nonseizure event. If there is sufficient data available and both calculations return positive numbers, the event is classified as a seizure. If the percentage of missing data (caused on occasion by a loose finger cuff) for either the HR (between tHR and the time of HR recovery) or the SpO2 (between td and tr ) is greater than 90%, or if the percentages of missing data for both signals is 80% or more, the event is classified as indeterminate. D. Training vs. Test Each patient’s biosignal response to seizure is slightly different, even within the HR↑⇒ SpO2 ↓⇒ EDA↑ pattern. Consequently, a training phase is required to tune parameters and optimize seizure detection accuracy. During training: • The multiple ∆HR , ∆SpO2 , ∆EDA levels discussed in Subsection II. A. are all calculated so that we can determine the highest levels for which a patient’s seizures will be detected. During this phase the baseline lengths (THR , TSpO2 , TEDA ) and the length of time the SpO2 remains low are evaluated and optimized. • Because we are working with signals that do not immediately recover after a seizure - indeed, EDA peaks after the seizure is over (see Fig. 3) - it is difficult to distinguish seizures that are close together. During training, we evaluate how close together an individual’s seizures can be and still be distinguishable based on how quickly his biosignals recover. In our current model, events that occur within 3 minutes of each other are considered one event. During the test phase, the parameter values chosen during training are used to evaluate the test data.
For our initial analysis, data from all three patients was grouped together, so we used the minimum parameter levels (∆HR = 15%, ∆SpO2 = 5%, ∆EDA = 60%). Our algorithm found 13 potential seizure events in the 108 hours of data: all seven seizures and six additional events. All seven seizures were correctly classified. Four of the remaining six events exhibited a nonseizure pattern and were classified as such. One of the events had insufficient data for proper assessment so it was classified as indeterminate. The last event was classified as a seizure. Since the patient was not on EEG or video monitor at the time, we don’t know its proper classification; however, the patient who experienced this event had very large SpO2 drops during each of his 4 known seizures. When the algorithm was personalized for this patient’s data (∆SpO2 = 10% instead of 5%), the unknown event was not selected as a potential seizure. A visual inspection of the data reveals that although the event fit the overall seizure pattern, it did not fit the variant for this particular patient. Our results are summarized in Table I. Even though our data set is small, we believe our finding is significant because the number of occasions on which all three biosignals are active at the same time is so small relative to the hours of data we collected: In our worst case scenario (not personalized), 13 events (including seizures) out of 108 hours of data collected equates to less than 0.125 events/hour; 1 potential false positive out of 108 hours of data equates to less than 0.01 potential false positives/hour. Again, we are basing our analysis on the fusion of 3 signals rather than depending on HR alone as in [5] and [6]. Our results are compared to theirs in Table II. Although our current algorithm works for the limited data we have analyzed so far, we plan to evaluate more sophisticated methods to improve our results as we go forward:
III. E XPERIMENTAL R ESULTS A. Setup and Data Collection Collection of seizure data is being done under IRB protocol at an EMU in Dallas, Texas. EEG seizure annotation is done by our medical consultants using NeuroWork software
5088
1) Wavelets [17], which are useful for finding abrupt changes in signals, and cepstral coefficients [18], which are used to recognize speech patterns, are likely to be better tools for our pattern recognition phase. 2) We have found machine learning techniques (nearest neighbor, neural networks and support vector machines) [19] useful for analyzing our data in the
TABLE I
R ESULTS FOR DATA A NALYZED BY O UR HR↑⇒ S P O2 ↓⇒ EDA↑ PATTERN S EARCH A LGORITHM . Results without Personalization Results with Personalized SpO2 Levels Class Seizure Nonseizure Indeterminate Seizure Nonseizure Indeterminate Seizure 7 (100%) 0 0 7 (100%) 0 0 1 (16.67%)* 4 (66.67%) 1 (16.67%) 0 3 (75%) 1 (25%) Nonseizure Sensitivity 7/7 (100%) 7/7 (100%) 5/6 (83.33%) 4/4 (100%) Specificity Accuracy 12/13 (92.31%) 11/11 (100%) *The actual class of this event is unknown because the patient was off the EEG monitor at the time it occurred. TABLE II
D ETECTION U SING OUR 3-B IOSIGNAL PATTERN VS . D ETECTION U SING H EART R ATE A LONE .
Sensitivity Specificity Accuracy Potential False Positive
Our Results Without With Personalization Personalization (Worst Case) (Best Case) 100% 100% 83.33% 100% 92.31% 100% 0.01/hour 0.00/hour
past [11]. Consequently, we will continue looking for appropriate applications of these techniques. 3) The discovery of a specific biosignal pattern associated with some patient’s seizures suggests that other patterns also exist. To date, we have not researched all seizure sensitive biosignals, but only those that can be monitored at the wrist using off-the-shelf products. Other biosignals such as electromyography and electrooculography may also form patterns when combined with the signals we have been monitoring. IV. C ONCLUSIONS The HR↑⇒ SpO2 ↓⇒ EDA↑ pattern has been observed in three epileptic patients whose data were collected in the EMU to date; we need to collect more data from epileptic patients in daily life as well as in EMU settings before we can be confident we have found a seizure detection methodology that is useful for a significant number of people. Data collection is continuing, so we will have opportunity to further evaluate our findings. We expect that the more sophisticated methodologies mentioned in Subsection III-B. will improve our results as we move forward. ACKNOWLEDGEMENT The authors thank Mehrdad Heydarzadeh and Maziyar Baran Pouyan for consulting on this project. Their insight into options for analyzing the EMU data was especially valuable. R EFERENCES [1] Mayo Clinic, “Epilepsy,” Accessed March 3, 2014, http://www.mayo.edu/research/departments-divisions/departmentneurology/programs/epilepsy. [2] Litt, Brian and Echauz, Javier, “Prediction of epileptic seizures,” Lancet Neurology 2002;1:22-30. [3] Cook, M., O’Brien, T., Berkovic, S., Murphy, M., Motokoff, A., Fabinyi, G., D’Souza, W., Yerra, R., Archer, J., Litewka, L., Hosking, S., Lightfoot, P., Ruedebusch, V., Sheffield, W., Snyder, D., Leyde, K., Himes, D., “Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study,” Lancet Neural 2013;12:563-71.
Osorio[5] Lowest Highest Settings Settings (Worst Case) (Best Case) N/A
N/A
9.5/hour
1.1/hour
Behbahani[6] Secondarily Complex Generalized Partial Seizures Seizures 88.66% 83.33% 90% 86.11% 88.33% 84.72% N/A N/A
[4] Van de Vel, A., Cuppens, K., Bonroy, B., Milosevic, M., Jansen, K., Van Huffel, S., Vanrumste, B., Lagae, L., Ceulemans, B., “Non-EEG seizure-detection systems and potential SUDEP prevention: State of the art,” Seizure 2013;22:345-355. [5] Osorio, Ivan, “Automated Seizure Detection Using EKG,” International Journal of Neural Systems 2014;24(2):1450001. [6] Behbahani, Socroor, Dabanloo, Nader Jafarnia, Nasrabadi, Ali Motie,Teixeira, Cesar A., Dourado, Antonio, “A new algorithm for detection of epileptic seizures based on HRV signal,” Journal of Experimental & Theoretical Artificial Intelligence 2014;26(2):251265. [7] Leutmezer, F.; Schernthaner, C.; Lurger, S.; Potzelberger, K., “Electrocardiographic Changes at the Onset of Epileptic Seizures,” Epilipsia, 2003;44(3):348-354. [8] Ives, J., Al-Aweel, I., Blum, A., Krishnamurthy, K., Goldberger, A., Schomer, D., “SpO2 Changes Precede EEG Changes during Seizures,” Journal of Clinical Neurophysiology 1997;14(2):162. [9] Regan, Mary E. and Brown, J. Keith, “Abnormalities in cardiac and respiratory function observed during seizures in childhood,” Development Medicine & Child Neurology 2005;47:4-9. [10] Poh, Ming-Zher, “Continuous Assessment of Epileptic Seizures with Wrist-worn Biosensors,” PhD diss., Massachusetts Institute of Technology, 2011. [11] Cogan, Diana, Pouyan, M. Baran, Nourani, M., Harvey, J., “A WristWorn Biosensor System for Assessment of Neurological Status,” Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE 2014: 5748 - 5751. [12] Mayo Clinic, “Hypoxemia (low blood oxygen),” Accessed January 18, 2015, http://www.mayoclinic.org/symptoms/hypoxemia/basics/ definition/sym-20050930? ga=1.191685801.2062893337.14250492450. [13] Neurowork, “Solutions Aligned with Business,” Accessed May, 2014, http://www.neurowork.net/en/products/. [14] Turner Medical Inc., http://www.turnermedical.com/. [15] Affectiva Inc., http://www.affectiva.com/q-sensor/. [16] MathWorks Inc., http://www.mathworks.com. [17] Burrus, C. Sidney, Ramesh A. Gopinath, and Haitao Guo, Introduction to Wavelets and Wavelet Transforms, Vol. 998, New Jersey: Prentice Hall, 1998. [18] Chougule, S.V., Chavan, M.S.,Gaikwad, M.S., “Filter bank based cepstral features for speaker recognition,” Proceedings - 2014 IEEE Global Conference on Wireless Computing and Networking 2014: 102-106. [19] Witten, I., Frank, E., Hall, M., Data Mining: Practical Machine Learning Tools and Techniques, 3rd ed., Elsevier Inc., 2011.
5089
Abstract— Single signal seizure detection algorithms suffer from high false positive rates. We have found a set of signals which can be easily monitored by a wristworn device and which produce a distinctive pattern during seizure for patients in an epilepsy monitoring unit (EMU). This pattern is much less likely to be reproduced by nonseizure events in the patient’s daily life than are changes in heart rate alone. We collected 108 hours of data from three EMU patients who suffered a combined total of seven seizures, then developed a time series analysis/pattern recognition based algorithm which distinguishes the seizures from nonseizure events with 100% accuracy.
I. I NTRODUCTION A. Background Seizure prediction or detection would improve the quality of life of millions of patients who suffer from uncontrollable seizures. Prediction would be ideal, as the patient would have opportunity to remove himself from dangerous situations; but with the current state of knowledge, it has proven to be an elusive goal [1][2][3]. Detection is a challenging but more attainable goal and offers two benefits. First, a caregiver can be notified, which is important because seizures can lead to injury or other complications including SUDEP (sudden unexpected death in epilepsy). Second, an electronic diary of seizure events can be created, which would be a boon for physicians as self-reporting of seizures has proven to be inaccurate [4][5]. EEG is the gold standard for seizure detection, but it is impractical for use in daily life because it requires the use of either a special headset worn in just the right way or sensors implanted in the brain. Neither of these options is convenient, and implanting a device in the brain is both dangerous and costly. There are a number of devices on the market designed to detect convulsive seizures using extracerebral (nonEEG) signals [4]. To the best of our knowledge, however, there is no wearable device available to detect complex partial seizures, as they are nonconvulsive. Recent work using heart rate (HR) to detect seizures has achieved fairly accurate results for patients in epilepsy monitoring units (EMUs), but the authors concede that there is much room for improvement to their algorithms [5][6]. In normal daily life, we must expect a greater variation in patient activity, and resulting heart rate changes greater than those typically found in an EMU. *This work is supported in part by Texas Medical Research Collaborative (TxMRC) 1 D. Cogan and M. Nourani are with Quality of Life Technology Laboratory, The University of Texas at Dallas, Richardson, TX 75080,
{diana.cogan, nourani}@utdallas.edu 2 J. Texas
Harvey, Epilepsy
DO and V. Group, Dallas,
Nagaraddi, MD are with TX 75230 {jhharvey,
dr.nagaraddi}@texasepilepsy.org
978-1-4244-9270-1/15/$31.00 ©2015 IEEE
B. Motivation and Contribution We chose a set of five biosignals which can be monitored at the wrist and are known or believed to be affected by seizures: heart rate (HR), arterial oxygenation (SpO2 ), accelerometry (ACC), electrodermal activity (EDA) and temperature (Temp). A number of researchers have observed extreme HR changes at the beginning of some types of seizures [7]. Researchers have also found indications that seizures may cause changes in SpO2 levels [8]. Their findings are supported by studies showing that seizures can cause disruption in patient respiration [9]. Motion measured by ACC and changes in EDA have been found to effectively indicate the onset of convulsive seizures [10]. Temp is associated with febrile seizures but remains to be investigated as a possible means for detection [4]. Several seizure sensitive biosignals, such as respiration and electromyography, cannot be monitored at the wrist and so are excluded in this work. Previously, we reported using these five signals to distinguish physical, cognitive and emotional stresses [11]. Recently, we found a pattern created by HR, SpO2 and EDA during all seven of the seizures suffered by three patients monitored in an EMU. This signal pattern is less likely to be reproduced by nonseizure events during daily life than are changes in HR alone. We developed an algorithm which (i) extracts features via a time series analysis of each signal, (ii) locates times of possible seizures and (iii) uses pattern recognition techniques to distinguish seizures from nonseizure events. Our algorithm has processed over 108 hours of data collected from the three patients and distinguished the seizures from nonseizure events with 100% accuracy. II. S EIZURE D ETECTION U SING A W EARABLE B IOSENSOR S YSTEM The flowchart for our time series analysis/pattern recognition based algorithm described below is shown in Fig. 1. A. Phase 1 - Preprocessing Each of the three signals of interest (HR, SpO2 and EDA) is analyzed separately using a 5 second window with 80% overlap. The mean of each window is compared to a moving baseline (THR , TSpO2 , and TEDA in Fig. 1), then used to update the baseline. If the mean of a window varies from the baseline by a prespecified parameter level (∆HR , ∆SpO2 , ∆EDA ), the time of the window is tagged and recorded. The parameter levels considered are: • 15%, 20%, 25% and 30% increases for HR, based on Osorio’s finding that these four levels are useful for detecting seizures [5].
5086
Fig. 2.
Windows for Event Search and Pattern Recognition.
The Event Search finds times when all three signals were active, lists them as potential seizure events and passes them to Phase 3 for comparison with the HR↑⇒ SpO2 ↓⇒ EDA↑ seizure pattern. The search methodology is illustrated in Fig. 2. SpO2 is the most stable of the three signals; therefore, for each time the selected ∆SpO2 was recorded in the SpO2 activity list from Phase 1 (each t∆SpO2 in Fig. 2), the HR and EDA activity lists are searched for matches. If ∆HR is found within the ∆HR Search Window and ∆EDA is found within the ∆EDA Search Window, all three signals were active simultaneously and t∆SpO2 represents a potential seizure. C. Phase 3 - Pattern Recognition
Fig. 1.
Seizure Detection Algorithm Flowchart.
5%, 7.5%, 10% and 12.5% decreases for SpO2 , chosen because apnea is defined as a 4% drop in SpO2 for 8 seconds or more. Normal SpO2 levels are between 95% and 100%; readings below 90% are low [12]. Consequently, a 5% drop in SpO2 is significant. • 60%, 80%, 100% and 120% increases for EDA, selected by analysis of the available data. Future findings may require modification of some or all parameter levels. These biosignals take time to settle after a seizure. Hence, a long baseline is needed to detect seizures that are close together. We are currently using a 60 second baseline for all three signals (THR = TSpO2 = TEDA = 60). The first and last time of a series of sequential windows which shows a signal change at the selected parameter level is passed to the Event Search in a signal activity list. •
B. Phase 2 - Personalization and Event Search Personalization is done during training, when EEG based seizure timing is available. The highest parameter level for each signal which places it in its activity list (Phase 1) at all seizure times is chosen to optimize algorithm performance. The parameters are selected by an iterative process through the Training Loop shown in Fig. 1.
The methodology of this phase is illustrated in Fig. 2 and Fig. 3. Fig. 3 illustrates the HR↑⇒ SpO2 ↓⇒ EDA↑ up seizure pattern using data from one seizure. For each potential seizure event: 1) A 30 minute Event Window centered about t∆SpO2 is selected. A Local SpO2 Baseline is calculated using the first 10 minutes of this window. 2) tHR , the time HR begins rising, is found within the HR Search Window. 3) td , the time SpO2 drops 2% below its Local Baseline and tr , the time it returns to halfway between its min value and its value at td are found within the SpO2 Search Window. The value of 2% was empirically selected based on the behavior of SpO2 during seizures. 4) tEDA , the time of the EDA max value is found within the EDA Search Window. Note that t∆SpO2 is shown in both Fig. 2 and Fig. 3. As the HR rise is not smooth, the data is windowed to backtrack from the event time (t∆SpO2 ) found in Phase 2 to tHR . tHR is the time of the window whose mean is greater than or equal to the means of both subsequent windows. The size of the window was adjusted so that the time found by the algorithm matched the time found by a visual inspection of the HR curve for all seven seizures. The SpO2 drop occurs so quickly that our algorithm can follow the curve without windowing the data. The bottom of the curve is not always smooth, however, so we search for a lower SpO2 value within 30 seconds of the apparent bottom and replace it if appropriate. We selected a 30 second search window because in all seven seizures the SpO2 begins rising within 30 seconds.
5087
[13]. Data is being collected from patients by means of two commercially available devices: (i) a Nonin WristOx 3150, which monitors HR and SpO2 [14] and (ii) an Affectiva Q Curve, which monitors EDA, Temp and ACC [15]. Our devices are time synced to the EEG equipment so we are certain of seizure onset and offset times relative to our biosignals. We collected data from five patients who had a total of 12 seizures during their stay in the EMU. The three patients who experienced the HR↑⇒ SpO2 ↓⇒ EDA↑ seizure pattern suffered, collectively, 2 secondarily generalized and 5 complex partial seizures. We implemented the algorithm to analyze their 108 hours of data in Matlab [16]. Fig. 3.
The HR↑⇒ SpO2 ↓⇒ EDA↑ Seizure Pattern.
B. Observations After finding the critical points, we calculate (i) tHR−SpO2 = td − tHR and (ii) tSpO2−EDA = tEDA − tr . If either of these calculations returns a negative number, the pattern for the event does not match the seizure pattern and the event is classified as a nonseizure event. If there is sufficient data available and both calculations return positive numbers, the event is classified as a seizure. If the percentage of missing data (caused on occasion by a loose finger cuff) for either the HR (between tHR and the time of HR recovery) or the SpO2 (between td and tr ) is greater than 90%, or if the percentages of missing data for both signals is 80% or more, the event is classified as indeterminate. D. Training vs. Test Each patient’s biosignal response to seizure is slightly different, even within the HR↑⇒ SpO2 ↓⇒ EDA↑ pattern. Consequently, a training phase is required to tune parameters and optimize seizure detection accuracy. During training: • The multiple ∆HR , ∆SpO2 , ∆EDA levels discussed in Subsection II. A. are all calculated so that we can determine the highest levels for which a patient’s seizures will be detected. During this phase the baseline lengths (THR , TSpO2 , TEDA ) and the length of time the SpO2 remains low are evaluated and optimized. • Because we are working with signals that do not immediately recover after a seizure - indeed, EDA peaks after the seizure is over (see Fig. 3) - it is difficult to distinguish seizures that are close together. During training, we evaluate how close together an individual’s seizures can be and still be distinguishable based on how quickly his biosignals recover. In our current model, events that occur within 3 minutes of each other are considered one event. During the test phase, the parameter values chosen during training are used to evaluate the test data.
For our initial analysis, data from all three patients was grouped together, so we used the minimum parameter levels (∆HR = 15%, ∆SpO2 = 5%, ∆EDA = 60%). Our algorithm found 13 potential seizure events in the 108 hours of data: all seven seizures and six additional events. All seven seizures were correctly classified. Four of the remaining six events exhibited a nonseizure pattern and were classified as such. One of the events had insufficient data for proper assessment so it was classified as indeterminate. The last event was classified as a seizure. Since the patient was not on EEG or video monitor at the time, we don’t know its proper classification; however, the patient who experienced this event had very large SpO2 drops during each of his 4 known seizures. When the algorithm was personalized for this patient’s data (∆SpO2 = 10% instead of 5%), the unknown event was not selected as a potential seizure. A visual inspection of the data reveals that although the event fit the overall seizure pattern, it did not fit the variant for this particular patient. Our results are summarized in Table I. Even though our data set is small, we believe our finding is significant because the number of occasions on which all three biosignals are active at the same time is so small relative to the hours of data we collected: In our worst case scenario (not personalized), 13 events (including seizures) out of 108 hours of data collected equates to less than 0.125 events/hour; 1 potential false positive out of 108 hours of data equates to less than 0.01 potential false positives/hour. Again, we are basing our analysis on the fusion of 3 signals rather than depending on HR alone as in [5] and [6]. Our results are compared to theirs in Table II. Although our current algorithm works for the limited data we have analyzed so far, we plan to evaluate more sophisticated methods to improve our results as we go forward:
III. E XPERIMENTAL R ESULTS A. Setup and Data Collection Collection of seizure data is being done under IRB protocol at an EMU in Dallas, Texas. EEG seizure annotation is done by our medical consultants using NeuroWork software
5088
1) Wavelets [17], which are useful for finding abrupt changes in signals, and cepstral coefficients [18], which are used to recognize speech patterns, are likely to be better tools for our pattern recognition phase. 2) We have found machine learning techniques (nearest neighbor, neural networks and support vector machines) [19] useful for analyzing our data in the
TABLE I
R ESULTS FOR DATA A NALYZED BY O UR HR↑⇒ S P O2 ↓⇒ EDA↑ PATTERN S EARCH A LGORITHM . Results without Personalization Results with Personalized SpO2 Levels Class Seizure Nonseizure Indeterminate Seizure Nonseizure Indeterminate Seizure 7 (100%) 0 0 7 (100%) 0 0 1 (16.67%)* 4 (66.67%) 1 (16.67%) 0 3 (75%) 1 (25%) Nonseizure Sensitivity 7/7 (100%) 7/7 (100%) 5/6 (83.33%) 4/4 (100%) Specificity Accuracy 12/13 (92.31%) 11/11 (100%) *The actual class of this event is unknown because the patient was off the EEG monitor at the time it occurred. TABLE II
D ETECTION U SING OUR 3-B IOSIGNAL PATTERN VS . D ETECTION U SING H EART R ATE A LONE .
Sensitivity Specificity Accuracy Potential False Positive
Our Results Without With Personalization Personalization (Worst Case) (Best Case) 100% 100% 83.33% 100% 92.31% 100% 0.01/hour 0.00/hour
past [11]. Consequently, we will continue looking for appropriate applications of these techniques. 3) The discovery of a specific biosignal pattern associated with some patient’s seizures suggests that other patterns also exist. To date, we have not researched all seizure sensitive biosignals, but only those that can be monitored at the wrist using off-the-shelf products. Other biosignals such as electromyography and electrooculography may also form patterns when combined with the signals we have been monitoring. IV. C ONCLUSIONS The HR↑⇒ SpO2 ↓⇒ EDA↑ pattern has been observed in three epileptic patients whose data were collected in the EMU to date; we need to collect more data from epileptic patients in daily life as well as in EMU settings before we can be confident we have found a seizure detection methodology that is useful for a significant number of people. Data collection is continuing, so we will have opportunity to further evaluate our findings. We expect that the more sophisticated methodologies mentioned in Subsection III-B. will improve our results as we move forward. ACKNOWLEDGEMENT The authors thank Mehrdad Heydarzadeh and Maziyar Baran Pouyan for consulting on this project. Their insight into options for analyzing the EMU data was especially valuable. R EFERENCES [1] Mayo Clinic, “Epilepsy,” Accessed March 3, 2014, http://www.mayo.edu/research/departments-divisions/departmentneurology/programs/epilepsy. [2] Litt, Brian and Echauz, Javier, “Prediction of epileptic seizures,” Lancet Neurology 2002;1:22-30. [3] Cook, M., O’Brien, T., Berkovic, S., Murphy, M., Motokoff, A., Fabinyi, G., D’Souza, W., Yerra, R., Archer, J., Litewka, L., Hosking, S., Lightfoot, P., Ruedebusch, V., Sheffield, W., Snyder, D., Leyde, K., Himes, D., “Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study,” Lancet Neural 2013;12:563-71.
Osorio[5] Lowest Highest Settings Settings (Worst Case) (Best Case) N/A
N/A
9.5/hour
1.1/hour
Behbahani[6] Secondarily Complex Generalized Partial Seizures Seizures 88.66% 83.33% 90% 86.11% 88.33% 84.72% N/A N/A
[4] Van de Vel, A., Cuppens, K., Bonroy, B., Milosevic, M., Jansen, K., Van Huffel, S., Vanrumste, B., Lagae, L., Ceulemans, B., “Non-EEG seizure-detection systems and potential SUDEP prevention: State of the art,” Seizure 2013;22:345-355. [5] Osorio, Ivan, “Automated Seizure Detection Using EKG,” International Journal of Neural Systems 2014;24(2):1450001. [6] Behbahani, Socroor, Dabanloo, Nader Jafarnia, Nasrabadi, Ali Motie,Teixeira, Cesar A., Dourado, Antonio, “A new algorithm for detection of epileptic seizures based on HRV signal,” Journal of Experimental & Theoretical Artificial Intelligence 2014;26(2):251265. [7] Leutmezer, F.; Schernthaner, C.; Lurger, S.; Potzelberger, K., “Electrocardiographic Changes at the Onset of Epileptic Seizures,” Epilipsia, 2003;44(3):348-354. [8] Ives, J., Al-Aweel, I., Blum, A., Krishnamurthy, K., Goldberger, A., Schomer, D., “SpO2 Changes Precede EEG Changes during Seizures,” Journal of Clinical Neurophysiology 1997;14(2):162. [9] Regan, Mary E. and Brown, J. Keith, “Abnormalities in cardiac and respiratory function observed during seizures in childhood,” Development Medicine & Child Neurology 2005;47:4-9. [10] Poh, Ming-Zher, “Continuous Assessment of Epileptic Seizures with Wrist-worn Biosensors,” PhD diss., Massachusetts Institute of Technology, 2011. [11] Cogan, Diana, Pouyan, M. Baran, Nourani, M., Harvey, J., “A WristWorn Biosensor System for Assessment of Neurological Status,” Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE 2014: 5748 - 5751. [12] Mayo Clinic, “Hypoxemia (low blood oxygen),” Accessed January 18, 2015, http://www.mayoclinic.org/symptoms/hypoxemia/basics/ definition/sym-20050930? ga=1.191685801.2062893337.14250492450. [13] Neurowork, “Solutions Aligned with Business,” Accessed May, 2014, http://www.neurowork.net/en/products/. [14] Turner Medical Inc., http://www.turnermedical.com/. [15] Affectiva Inc., http://www.affectiva.com/q-sensor/. [16] MathWorks Inc., http://www.mathworks.com. [17] Burrus, C. Sidney, Ramesh A. Gopinath, and Haitao Guo, Introduction to Wavelets and Wavelet Transforms, Vol. 998, New Jersey: Prentice Hall, 1998. [18] Chougule, S.V., Chavan, M.S.,Gaikwad, M.S., “Filter bank based cepstral features for speaker recognition,” Proceedings - 2014 IEEE Global Conference on Wireless Computing and Networking 2014: 102-106. [19] Witten, I., Frank, E., Hall, M., Data Mining: Practical Machine Learning Tools and Techniques, 3rd ed., Elsevier Inc., 2011.
5089
