Journal of Neuroscience Methods 221 (2014) 189–195
Contents lists available at ScienceDirect
Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth
Computational Neuroscience
Semi-automatic sleep EEG scoring based on the hypnospectrogram Andreas M. Koupparis, Vasileios Kokkinos, George K. Kostopoulos ∗ Neurophysiology Unit, Medical School, University of Patras, Greece
h i g h l i g h t s • We explore the possibility of sleep scoring using the whole-night time-frequency analysis, termed hypnospectrogram, with a computer-assisted K-means clustering method.
• Hypnograms were derived from 10 whole-night sleep EEG recordings using either standard visual scoring under the Rechtshaffen and Kales criteria or semi-automated analysis of the hypnospectrogram derived from a single EEG electrode.
• We measured substantial agreement between the two approaches with Cohen’s kappa considering all 7 stages at 0.61. • The hypnospectrogram approach offers the scorer the opportunity to exploit the information-rich graphic representation of the whole night sleep.
a r t i c l e
i n f o
Article history: Received 29 July 2013 Received in revised form 15 October 2013 Accepted 16 October 2013 Keywords: Sleep EEG Sleep scoring Time-frequency analysis K-means clustering
a b s t r a c t Background: Sleep EEG organization is revealed by sleep scoring, a time-consuming process based on strictly defined visual criteria. New method: We explore the possibility of sleep scoring using the whole-night time-frequency analysis, termed hypnospectrogram, with a computer-assisted K-means clustering method. Results: Hypnograms were derived from 10 whole-night sleep EEG recordings using either standard visual scoring under the Rechtshaffen and Kales criteria or semi-automated analysis of the hypnospectrogram derived from a single EEG electrode. We measured substantial agreement between the two approaches with Cohen’s kappa considering all 7 stages at 0.61. Comparison with existing methods: A number of existing automated procedures have reached the level of human inter-rater agreement using the standard criteria. However, our approach offers the scorer the opportunity to exploit the information-rich graphic representation of the whole night sleep upon which the automated method works. Conclusion: This work suggests that the hypnospectrogram can be used as an objective graphical representation of sleep architecture upon which sleep scoring can be performed with computer-assisted methods. © 2013 Elsevier B.V. All rights reserved.
1. Introduction The organization of sleep has been revealed by means of the EEG (Loomis et al., 1937; Carskadon and Dement, 2011). Even since the early days of sleep EEG recordings, it was clear that major patterns of EEG activity are repeated throughout the night, in a sequential manner. Periods of rapid eye-movement (REM) sleep interchange periodically with non-REM (NREM) periods, creating a sequence of about 5–6 cycles in every whole night’s sleep (Aserinsky and Kleitman, 1953). In addition, NREM is divided into 4 stages corresponding to increasing depth of sleep (Gibbs and Gibbs, 1950;
∗ Corresponding author at: Medical School, University of Patras, 26500 Patras, Greece. Tel.: +30 2610 969157/969155; fax: +30 2610 969176. E-mail addresses: [email protected] (A.M. Koupparis), [email protected] (G.K. Kostopoulos). 0165-0270/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jneumeth.2013.10.010
Dement and Kleitman, 1957). The depiction of this macroscopic architecture of sleep is performed by the process of sleep scoring. Sleep scoring based on EEG as well as EOG (to detect eye movements) and EMG (to evaluate muscle tone) has been the most widely used tool of sleep research since the formulation of scoring criteria by Rechtschaffen and Kales (1968), which were updated by the AASM Visual Scoring Task Force (Silber et al., 2007) with the combination of stages 3 and 4 into slow wave sleep (SWS). The resulting graphical representation of sleep architecture, called hypnogram, has offered a macroscopic overview of the sleep organization into alternating states of brain activity with recurrent cycles of NREM and REM stages. Besides providing a frame for research and understanding of sleep, sleep scoring is a valuable tool in several clinical investigations including that of objectively establishing sleep quality, sleep apnea, mutual relations between sleep and epilepsy, etc. The release of several hormones and other physiological markers as
190
A.M. Koupparis et al. / Journal of Neuroscience Methods 221 (2014) 189–195
well as the clinical expression of several diseases are increasingly realized to occur at specific times of a hypnogram and so their study requires sleep scoring. Sleep scoring can be performed by sleep researchers, clinicians and trained neurophysiologists (Rodenbeck et al., 2006; Silber et al., 2007). Hypnograms can be produced either manually (Halász et al., 2002), or by means of the spectral power distribution (Gross and Gotman, 1999; Ferri et al., 2001; Tan et al., 2003); the former being a severely time-consuming process, the latter being biased by parametric selection (band limits, thresholds, etc.) (Huupponen et al., 2007). Our group has proposed guidelines for the use of the hypnospectrogram – the three dimensional plot of the time-frequency analysis of the EEG signal – as a rough visual means of estimation of the sleep staging and quality of sleep (Kokkinos et al., 2009). In this work we extend the use of the hypnospectrogram toward semiautomated formulation of hypnograms using templates from the 3-dimensional image data, in an attempt to evaluate its capabilities to provide structured information about sleep organization.
2. Methods 2.1. Subjects and recordings Ten individuals (6 females and 4 males), between 23 and 35 years of age (mean age 26.5 ± 3.8) participated in this sleep study. All subjects were good sleepers, without particular difficulty in falling and remaining asleep throughout the night. They were all in good health, free from medication at the time of study, without any history of neurological or psychiatric disorder, or disordered sleep. The participants were instructed to keep a 7-day sleep diary, to follow their regular sleep schedule, and refrain from alcohol and caffeine for at least 3 and 1 days, respectively, prior to the scheduled sleep recording. Menstrual phase was not controlled for in female subjects. All subjects read and signed an informed consent form regarding the procedures and the purpose of the sleep EEG study. Our participants were scheduled to arrive at the laboratory for electrode preparation approximately 1 h prior to their usual nocturnal bedtime; the latter calculated as an average of reported bedtimes from their sleep diaries. Each spent a whole night in the laboratory, in an air-conditioned soundproof temperaturecontrolled Faraday-cage dark room. Sleep was spontaneous, without having to administer any pharmacological substance. The sleep EEG recording begun after the subjects willingly switched off the room lights, as instructed to do when they would feel like falling asleep, and ended with their spontaneous full arousal in the morning. The recorded electrophysiological signals were continuously monitored in the adjacent room, accompanied by an overnight bidirectional vocal communication system through a microphonespeaker console system (video-monitoring was intentionally not used to avoid potential sleep disturbances). Upon awakening in the morning, all subjects reported to have had a comfortable and undisturbed sleep. Subjects that failed to maintain adequate quality of sleep were not included in this study. All procedures described including the informed consent form were approved by the Ethics Committee of the Medical School of University of Patras. Whole-night sleep was recorded using 58 EEG tin electrodes, placed on an electrode cap (ElectroCap International Inc., Eaton, OH, USA) according to the extended international 10–20 system (FP1, FPz, FP2, AF3, AF4, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FC5, FC3, FC1, FCz, FC2, FC4, FC6, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP7, CP5, CP 3, CP 1, CPz, CP 2, CP 4, CP6, TP8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, PO1, POz, PO2, PO4, PO8, O1, Oz, O2). In 2 subjects, a similar Ag–AgCl electrode cap was used. Electrode inputs were linked-ear referenced and grounded over the AFz position. A bipolar derivation of oblique EOG was used to detect eye-movements,
with the electrodes placed 1 cm above the right outer canthus and 1 cm below the left outer canthus. A bipolar EMG from the upper masseter muscle was used to track muscle tone changes. Impedances for all electrodes were kept below 10 k for most of the night. All electrophysiological signals were AC recorded, amplified at a total gain of 1000, band-pass filtered at 0.05–500 Hz, and digitized through a 16bit-resolution A/D converter, providing an accuracy of 0.084 V/LSB, at a sampling frequency of 2500 Hz by a Synamps system (Neuroscan Inc., Charlotte, NC, USA), and stored on hard disk. No 50 Hz line notch filter was used during recording. Subject nocturnal body movements were detected by a sensitive motion-detector placed over the bed area, that produced a 2 s transistor–transistor logic (TTL) signal every time movement occurred. The latter was fed into the recording system as an external trigger on a separate event channel and was stored along with the rest of the signals. 2.2. Analysis All data were analyzed in two ways, manually and semiautomatically: Manual sleep staging was performed by visual inspection of the EEG recordings along with EOG and EMG channels using the criteria of Rechtschaffen and Kales (1968), taking into consideration the propositions of the AASM Visual Scoring Task Force (Silber et al., 2007) and the DGSM Task Force (Rodenbeck et al., 2006), and keeping a time resolution of one second. Microarousals were scored using the guidelines of the ASDA report (Bonnet et al., 1992). For the semi-automatic sleep staging, hypnospectrograms for each subject were produced according to the procedures described previously (Kokkinos et al., 2009). Fig. 1 shows a hypnospectrogram (A), a manually produced hypnogram (B), a hypnogram resulting from semiautomatic procedure (C) and an example of raw EEG with characteristics of NREM stage 2 (K-complex and spindles, D). The whole-night time-frequency analysis (TFA) was derived using short-time Fourier transformation (STFT) with a 4096-point window and 4000 overlap on the signal of Cz downsampled at 500 Hz, for frequencies in the range 0.05–45 Hz at a step of 0.05 Hz, which produced 192 ms time bins. The algorithmic steps used to produce the automatic scoring from the TFA image were: (1) Reduction of the time resolution to 30-s using the mean spectral power of the corresponding time bins. (2) Normalization of the power values of every time-frequency bin using the average for every 0.05 Hz frequency bin over the whole night. (3) Clustering the derived time bins according to their distribution of EEG power in the different frequencies using the k-means algorithm (Hartigan and Wong, 1979). The corresponding sleep stage for each cluster was selected by visual inspection of the average EEG spectral power from all the elements in a same cluster (Awake, Microarousal, REM, NonREM I, II, III and IV). Though ideally seven clusters would suffice, twelve clusters were used to account for spectral differences within a stage that could be separated into different clusters. The resulting hypnogram was expanded to 1 s resolution by 30-fold repetition of each value, so that corrections could be made on a finer time scale, and plotted along the original hypnospectrogram. Manual corrections to both specific clusters or specific time periods – where ever deemed necessary – were used to derive the final sleep scoring. The manually derived and the semi-automatically derived hypnograms were compared on the basis of agreement between stages and Cohen’s kappa (Landis and Koch, 1977). For the semiautomatically derived hypnograms, two scorings were used; one where the manual intervention was minimal and focused on the scoring of clusters and small corrections on the hypnogram (“minimal effort”), and another where the manual corrections included zooming into the hypnospectrogram and were more “detailed”. In each case, agreement values were calculated (a) for awake, REM
A.M. Koupparis et al. / Journal of Neuroscience Methods 221 (2014) 189–195
191
Fig. 1. Hypnospectrogram, hypnograms and raw EEG from the same sleep recording of one subject. (A) Hypnospectrogram derived from Cz. (B) Hypnogram derived by visual scoring. (C) Hypnogram derived by full-effort semi-automated analysis of the hypnospectrogram. This subject had trouble falling asleep for the first hour of the recording. (D) Raw EEG traces for 10 s of NREM2 exhibiting a K-complex and two sleep spindles. The corresponding time point in the hypnograms and the hypnospectrogram is shown with a vertical green line. MA: Microarousals, AW: Awake, NR1-4: Non-REM stages 1–4. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
sleep and NREM sleep, (b) for awake, REM, light sleep (NREM I + II) and slow-wave sleep (NREM III + IV), and (c) for all sleep stages. All the steps of the analysis were carried out using custom built software on MATLAB (The MathWorks, Natick, MA). 3. Results Three hypnograms were produced for each of the 10 subjects, one using visual scoring on the raw EEG (central derivations along with EMG and EOG) recordings, one derived automatically from the hypnospectrogram with minor manual interventions and another with detailed interventions – on both latter cases using only the hypnospectrogram as a reference. For all subjects, the hypnospectrogram was able to provide enough information for the researchers to identify sleep cycles and macrostructure features. To further quantitatively appreciate these results using hypnograms, the mean time required by the scorer to produce the whole night hypnograms was 5 min per subject for the minimal effort semiautomated scoring, about 45 min for the detailed semi-automated scoring and 3 h for the visual scoring using the EEG, EMG and EOG traces. All the resulting hypnograms are presented in a so-called lasagna plot (Swihart et al., 2010) in Fig. 2 and the quantitative sleep
parameters are summarized in Table 1. Comparison between the different hypnograms of each subject was quantified using Cohen’s kappa (Fig. 3). The minimal effort semi-automated scoring focused mainly on identifying the corresponding sleep stage for the automatically produced clusters using their spectral information (Fig. 4). In this step, 12 clusters were requested and those belonging to the same stage were manually merged. The difficulties encountered included inadequate discrimination of NREM stages 3 and 4, as well as differentiation between awake, REM and NREM 1 stages. In these cases, different hypnograms were created by labeling problematic clusters as different stages and concurrent visual inspection of the hypnospectrogram was used to decide which hypnogram fits the hypnospectrogram best, thus providing enough clues to decide whether a specific cluster should be labeled as belonging to one sleep stage or the other. The resulting hypnograms did, however, exhibit frequent shifts between similar, in spectral contents, stages. Also as expected, awake and microarousals clusters were frequently intermixed. The detailed intervention on the automatically derived hypnogram used zooming into the hypnospectrogram (resolution up to 200 ms), and allowed a time resolution of 1 s for the corrections on the hypnograms so that microarousals could be accurately scored.
192
A.M. Koupparis et al. / Journal of Neuroscience Methods 221 (2014) 189–195
Fig. 2. “Lasagna” plot of 30 hypnograms for 10 subjects. For each subject, the top row represents the visual scoring, the medium the semi-automated scoring with full effort and the bottom row the automated scoring with almost no intervention. In this way the degree of concordance between the three ways of analysis can be easily appreciated. Epochs scored as microarousals are not shown in this plot. Some microarousals are seen in the automatically derived hypnograms where they were scored as awake epochs.
The corrections were based on the interpretation of the hypnospectrogram as described earlier (Kokkinos et al., 2009). The main difficulty was the decision for the point of transition between NREM 2 and 3, and between NREM 3 and 4. The time required to produce the hypnogram was on average 5 min per subject to identify the stage represented by each original cluster and create the initial hypnogram, and another 40 min on average to modify it by zooming into the hypnospectrogram and deciding the stage on a finer time scale. Stable periods of sleep without many transitions were checked faster using larger time scales, whereas scoring of transitions and microarousals needed more time and zooming into a few seconds scale to identify the transition more accurately. Comparison between the visually scored and the semiautomatic hypnograms, revealed that the sleep cycles are identifiable by all three approaches (Fig. 2). Concerning the discrimination between awake, REM and NREM stage 1, issues arose when NREM 1 appeared after the first cycle near REM periods. Agreement
on the transitions between NREM stages 2–4 was also problematic. These results are quantified in Fig. 3. For the assessment of Cohen’s kappa values, Landis and Koch (1977) have provided arbitrary guidelines as follows:
Contents lists available at ScienceDirect
Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth
Computational Neuroscience
Semi-automatic sleep EEG scoring based on the hypnospectrogram Andreas M. Koupparis, Vasileios Kokkinos, George K. Kostopoulos ∗ Neurophysiology Unit, Medical School, University of Patras, Greece
h i g h l i g h t s • We explore the possibility of sleep scoring using the whole-night time-frequency analysis, termed hypnospectrogram, with a computer-assisted K-means clustering method.
• Hypnograms were derived from 10 whole-night sleep EEG recordings using either standard visual scoring under the Rechtshaffen and Kales criteria or semi-automated analysis of the hypnospectrogram derived from a single EEG electrode.
• We measured substantial agreement between the two approaches with Cohen’s kappa considering all 7 stages at 0.61. • The hypnospectrogram approach offers the scorer the opportunity to exploit the information-rich graphic representation of the whole night sleep.
a r t i c l e
i n f o
Article history: Received 29 July 2013 Received in revised form 15 October 2013 Accepted 16 October 2013 Keywords: Sleep EEG Sleep scoring Time-frequency analysis K-means clustering
a b s t r a c t Background: Sleep EEG organization is revealed by sleep scoring, a time-consuming process based on strictly defined visual criteria. New method: We explore the possibility of sleep scoring using the whole-night time-frequency analysis, termed hypnospectrogram, with a computer-assisted K-means clustering method. Results: Hypnograms were derived from 10 whole-night sleep EEG recordings using either standard visual scoring under the Rechtshaffen and Kales criteria or semi-automated analysis of the hypnospectrogram derived from a single EEG electrode. We measured substantial agreement between the two approaches with Cohen’s kappa considering all 7 stages at 0.61. Comparison with existing methods: A number of existing automated procedures have reached the level of human inter-rater agreement using the standard criteria. However, our approach offers the scorer the opportunity to exploit the information-rich graphic representation of the whole night sleep upon which the automated method works. Conclusion: This work suggests that the hypnospectrogram can be used as an objective graphical representation of sleep architecture upon which sleep scoring can be performed with computer-assisted methods. © 2013 Elsevier B.V. All rights reserved.
1. Introduction The organization of sleep has been revealed by means of the EEG (Loomis et al., 1937; Carskadon and Dement, 2011). Even since the early days of sleep EEG recordings, it was clear that major patterns of EEG activity are repeated throughout the night, in a sequential manner. Periods of rapid eye-movement (REM) sleep interchange periodically with non-REM (NREM) periods, creating a sequence of about 5–6 cycles in every whole night’s sleep (Aserinsky and Kleitman, 1953). In addition, NREM is divided into 4 stages corresponding to increasing depth of sleep (Gibbs and Gibbs, 1950;
∗ Corresponding author at: Medical School, University of Patras, 26500 Patras, Greece. Tel.: +30 2610 969157/969155; fax: +30 2610 969176. E-mail addresses: [email protected] (A.M. Koupparis), [email protected] (G.K. Kostopoulos). 0165-0270/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jneumeth.2013.10.010
Dement and Kleitman, 1957). The depiction of this macroscopic architecture of sleep is performed by the process of sleep scoring. Sleep scoring based on EEG as well as EOG (to detect eye movements) and EMG (to evaluate muscle tone) has been the most widely used tool of sleep research since the formulation of scoring criteria by Rechtschaffen and Kales (1968), which were updated by the AASM Visual Scoring Task Force (Silber et al., 2007) with the combination of stages 3 and 4 into slow wave sleep (SWS). The resulting graphical representation of sleep architecture, called hypnogram, has offered a macroscopic overview of the sleep organization into alternating states of brain activity with recurrent cycles of NREM and REM stages. Besides providing a frame for research and understanding of sleep, sleep scoring is a valuable tool in several clinical investigations including that of objectively establishing sleep quality, sleep apnea, mutual relations between sleep and epilepsy, etc. The release of several hormones and other physiological markers as
190
A.M. Koupparis et al. / Journal of Neuroscience Methods 221 (2014) 189–195
well as the clinical expression of several diseases are increasingly realized to occur at specific times of a hypnogram and so their study requires sleep scoring. Sleep scoring can be performed by sleep researchers, clinicians and trained neurophysiologists (Rodenbeck et al., 2006; Silber et al., 2007). Hypnograms can be produced either manually (Halász et al., 2002), or by means of the spectral power distribution (Gross and Gotman, 1999; Ferri et al., 2001; Tan et al., 2003); the former being a severely time-consuming process, the latter being biased by parametric selection (band limits, thresholds, etc.) (Huupponen et al., 2007). Our group has proposed guidelines for the use of the hypnospectrogram – the three dimensional plot of the time-frequency analysis of the EEG signal – as a rough visual means of estimation of the sleep staging and quality of sleep (Kokkinos et al., 2009). In this work we extend the use of the hypnospectrogram toward semiautomated formulation of hypnograms using templates from the 3-dimensional image data, in an attempt to evaluate its capabilities to provide structured information about sleep organization.
2. Methods 2.1. Subjects and recordings Ten individuals (6 females and 4 males), between 23 and 35 years of age (mean age 26.5 ± 3.8) participated in this sleep study. All subjects were good sleepers, without particular difficulty in falling and remaining asleep throughout the night. They were all in good health, free from medication at the time of study, without any history of neurological or psychiatric disorder, or disordered sleep. The participants were instructed to keep a 7-day sleep diary, to follow their regular sleep schedule, and refrain from alcohol and caffeine for at least 3 and 1 days, respectively, prior to the scheduled sleep recording. Menstrual phase was not controlled for in female subjects. All subjects read and signed an informed consent form regarding the procedures and the purpose of the sleep EEG study. Our participants were scheduled to arrive at the laboratory for electrode preparation approximately 1 h prior to their usual nocturnal bedtime; the latter calculated as an average of reported bedtimes from their sleep diaries. Each spent a whole night in the laboratory, in an air-conditioned soundproof temperaturecontrolled Faraday-cage dark room. Sleep was spontaneous, without having to administer any pharmacological substance. The sleep EEG recording begun after the subjects willingly switched off the room lights, as instructed to do when they would feel like falling asleep, and ended with their spontaneous full arousal in the morning. The recorded electrophysiological signals were continuously monitored in the adjacent room, accompanied by an overnight bidirectional vocal communication system through a microphonespeaker console system (video-monitoring was intentionally not used to avoid potential sleep disturbances). Upon awakening in the morning, all subjects reported to have had a comfortable and undisturbed sleep. Subjects that failed to maintain adequate quality of sleep were not included in this study. All procedures described including the informed consent form were approved by the Ethics Committee of the Medical School of University of Patras. Whole-night sleep was recorded using 58 EEG tin electrodes, placed on an electrode cap (ElectroCap International Inc., Eaton, OH, USA) according to the extended international 10–20 system (FP1, FPz, FP2, AF3, AF4, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FC5, FC3, FC1, FCz, FC2, FC4, FC6, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP7, CP5, CP 3, CP 1, CPz, CP 2, CP 4, CP6, TP8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, PO7, PO3, PO1, POz, PO2, PO4, PO8, O1, Oz, O2). In 2 subjects, a similar Ag–AgCl electrode cap was used. Electrode inputs were linked-ear referenced and grounded over the AFz position. A bipolar derivation of oblique EOG was used to detect eye-movements,
with the electrodes placed 1 cm above the right outer canthus and 1 cm below the left outer canthus. A bipolar EMG from the upper masseter muscle was used to track muscle tone changes. Impedances for all electrodes were kept below 10 k for most of the night. All electrophysiological signals were AC recorded, amplified at a total gain of 1000, band-pass filtered at 0.05–500 Hz, and digitized through a 16bit-resolution A/D converter, providing an accuracy of 0.084 V/LSB, at a sampling frequency of 2500 Hz by a Synamps system (Neuroscan Inc., Charlotte, NC, USA), and stored on hard disk. No 50 Hz line notch filter was used during recording. Subject nocturnal body movements were detected by a sensitive motion-detector placed over the bed area, that produced a 2 s transistor–transistor logic (TTL) signal every time movement occurred. The latter was fed into the recording system as an external trigger on a separate event channel and was stored along with the rest of the signals. 2.2. Analysis All data were analyzed in two ways, manually and semiautomatically: Manual sleep staging was performed by visual inspection of the EEG recordings along with EOG and EMG channels using the criteria of Rechtschaffen and Kales (1968), taking into consideration the propositions of the AASM Visual Scoring Task Force (Silber et al., 2007) and the DGSM Task Force (Rodenbeck et al., 2006), and keeping a time resolution of one second. Microarousals were scored using the guidelines of the ASDA report (Bonnet et al., 1992). For the semi-automatic sleep staging, hypnospectrograms for each subject were produced according to the procedures described previously (Kokkinos et al., 2009). Fig. 1 shows a hypnospectrogram (A), a manually produced hypnogram (B), a hypnogram resulting from semiautomatic procedure (C) and an example of raw EEG with characteristics of NREM stage 2 (K-complex and spindles, D). The whole-night time-frequency analysis (TFA) was derived using short-time Fourier transformation (STFT) with a 4096-point window and 4000 overlap on the signal of Cz downsampled at 500 Hz, for frequencies in the range 0.05–45 Hz at a step of 0.05 Hz, which produced 192 ms time bins. The algorithmic steps used to produce the automatic scoring from the TFA image were: (1) Reduction of the time resolution to 30-s using the mean spectral power of the corresponding time bins. (2) Normalization of the power values of every time-frequency bin using the average for every 0.05 Hz frequency bin over the whole night. (3) Clustering the derived time bins according to their distribution of EEG power in the different frequencies using the k-means algorithm (Hartigan and Wong, 1979). The corresponding sleep stage for each cluster was selected by visual inspection of the average EEG spectral power from all the elements in a same cluster (Awake, Microarousal, REM, NonREM I, II, III and IV). Though ideally seven clusters would suffice, twelve clusters were used to account for spectral differences within a stage that could be separated into different clusters. The resulting hypnogram was expanded to 1 s resolution by 30-fold repetition of each value, so that corrections could be made on a finer time scale, and plotted along the original hypnospectrogram. Manual corrections to both specific clusters or specific time periods – where ever deemed necessary – were used to derive the final sleep scoring. The manually derived and the semi-automatically derived hypnograms were compared on the basis of agreement between stages and Cohen’s kappa (Landis and Koch, 1977). For the semiautomatically derived hypnograms, two scorings were used; one where the manual intervention was minimal and focused on the scoring of clusters and small corrections on the hypnogram (“minimal effort”), and another where the manual corrections included zooming into the hypnospectrogram and were more “detailed”. In each case, agreement values were calculated (a) for awake, REM
A.M. Koupparis et al. / Journal of Neuroscience Methods 221 (2014) 189–195
191
Fig. 1. Hypnospectrogram, hypnograms and raw EEG from the same sleep recording of one subject. (A) Hypnospectrogram derived from Cz. (B) Hypnogram derived by visual scoring. (C) Hypnogram derived by full-effort semi-automated analysis of the hypnospectrogram. This subject had trouble falling asleep for the first hour of the recording. (D) Raw EEG traces for 10 s of NREM2 exhibiting a K-complex and two sleep spindles. The corresponding time point in the hypnograms and the hypnospectrogram is shown with a vertical green line. MA: Microarousals, AW: Awake, NR1-4: Non-REM stages 1–4. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)
sleep and NREM sleep, (b) for awake, REM, light sleep (NREM I + II) and slow-wave sleep (NREM III + IV), and (c) for all sleep stages. All the steps of the analysis were carried out using custom built software on MATLAB (The MathWorks, Natick, MA). 3. Results Three hypnograms were produced for each of the 10 subjects, one using visual scoring on the raw EEG (central derivations along with EMG and EOG) recordings, one derived automatically from the hypnospectrogram with minor manual interventions and another with detailed interventions – on both latter cases using only the hypnospectrogram as a reference. For all subjects, the hypnospectrogram was able to provide enough information for the researchers to identify sleep cycles and macrostructure features. To further quantitatively appreciate these results using hypnograms, the mean time required by the scorer to produce the whole night hypnograms was 5 min per subject for the minimal effort semiautomated scoring, about 45 min for the detailed semi-automated scoring and 3 h for the visual scoring using the EEG, EMG and EOG traces. All the resulting hypnograms are presented in a so-called lasagna plot (Swihart et al., 2010) in Fig. 2 and the quantitative sleep
parameters are summarized in Table 1. Comparison between the different hypnograms of each subject was quantified using Cohen’s kappa (Fig. 3). The minimal effort semi-automated scoring focused mainly on identifying the corresponding sleep stage for the automatically produced clusters using their spectral information (Fig. 4). In this step, 12 clusters were requested and those belonging to the same stage were manually merged. The difficulties encountered included inadequate discrimination of NREM stages 3 and 4, as well as differentiation between awake, REM and NREM 1 stages. In these cases, different hypnograms were created by labeling problematic clusters as different stages and concurrent visual inspection of the hypnospectrogram was used to decide which hypnogram fits the hypnospectrogram best, thus providing enough clues to decide whether a specific cluster should be labeled as belonging to one sleep stage or the other. The resulting hypnograms did, however, exhibit frequent shifts between similar, in spectral contents, stages. Also as expected, awake and microarousals clusters were frequently intermixed. The detailed intervention on the automatically derived hypnogram used zooming into the hypnospectrogram (resolution up to 200 ms), and allowed a time resolution of 1 s for the corrections on the hypnograms so that microarousals could be accurately scored.
192
A.M. Koupparis et al. / Journal of Neuroscience Methods 221 (2014) 189–195
Fig. 2. “Lasagna” plot of 30 hypnograms for 10 subjects. For each subject, the top row represents the visual scoring, the medium the semi-automated scoring with full effort and the bottom row the automated scoring with almost no intervention. In this way the degree of concordance between the three ways of analysis can be easily appreciated. Epochs scored as microarousals are not shown in this plot. Some microarousals are seen in the automatically derived hypnograms where they were scored as awake epochs.
The corrections were based on the interpretation of the hypnospectrogram as described earlier (Kokkinos et al., 2009). The main difficulty was the decision for the point of transition between NREM 2 and 3, and between NREM 3 and 4. The time required to produce the hypnogram was on average 5 min per subject to identify the stage represented by each original cluster and create the initial hypnogram, and another 40 min on average to modify it by zooming into the hypnospectrogram and deciding the stage on a finer time scale. Stable periods of sleep without many transitions were checked faster using larger time scales, whereas scoring of transitions and microarousals needed more time and zooming into a few seconds scale to identify the transition more accurately. Comparison between the visually scored and the semiautomatic hypnograms, revealed that the sleep cycles are identifiable by all three approaches (Fig. 2). Concerning the discrimination between awake, REM and NREM stage 1, issues arose when NREM 1 appeared after the first cycle near REM periods. Agreement
on the transitions between NREM stages 2–4 was also problematic. These results are quantified in Fig. 3. For the assessment of Cohen’s kappa values, Landis and Koch (1977) have provided arbitrary guidelines as follows:
