Abstract
Background
The mechanisms behind the severe pain of cluster headache remain enigmatic. A distinguishing feature of the attacks is the striking rhythms with which they occur. We investigated whether statistical modelling can be used to describe 24-hour attack distributions and identify differences between subgroups.
Methods
Common hours of attacks for 351 cluster headache patients were collected. Probability distributions of attacks throughout the day (chronorisk) was calculated. These 24-hour distributions were analysed with a multimodal Gaussian fit identifying periods of elevated attack risk and a spectral analysis identifying oscillations in risk.
Results
The Gaussian model fit for the chronorisk distribution for all patients reporting diurnal rhythmicity (n = 286) had a goodness of fit R2 value of 0.97 and identified three times of increased risk peaking at 21:41, 02:02 and 06:23 hours. In subgroups, three to five modes of increased risk were found and goodness of fit values ranged from 0.85–0.99. Spectral analysis revealed multiple distinct oscillation frequencies in chronorisk in subgroups including a dominant circadian oscillation in episodic patients and an ultradian in chronic.
Conclusions
Chronorisk in cluster headache can be characterised as a sum of individual, timed events of increased risk, each having a Gaussian distribution. In episodic cluster headache, attacks follow a circadian rhythmicity whereas, in the chronic variant, ultradian oscillations are dominant reflecting a loss of association with sleep and perhaps explaining observed differences in the effects of specific treatments. The results demonstrate the ability to accurately model chronobiological patterns in a primary headache.
Introduction
Despite some recent advances, the pathophysiology of cluster headache (CH) remains imperfectly understood. The sub-classification of episodic (eCH) or chronic (cCH) CH is purely diagnostic and not based on known differences besides duration of the remission periods. Other than the severe pain and autonomic symptoms, the most distinguishing feature is arguably the chronobiological rhythms in which the attacks occur. The presence of particular diurnal attack patterns is well documented (1,2), but no attempts have been made to identify factors that influence these patterns or to characterise them beyond time of occurrence.
The chronobiological features of CH have been systematically studied since the 1980s and clear times of an increased number of attacks, perhaps influenced by cultural factors, have been identified (3,4). In a recent publication, our group demonstrated the influence of daylight on the occurrence of cluster periods (5) and a phase shift between men and women (6). A better understanding of when attacks occur and what affects this will allow better timing of treatment. The optimisation of the timing of treatment with the aim of maximising effect and minimising side effects is known as chronotherapy. This concept has not been systematically explored in the field of headache, but is well known in oncology (7) and other fields of medicine. Before optimisation of the timing of treatment can be undertaken, a better understanding of the rhythms and natural fluctuations of attacks is necessary. This knowledge can also give insight into the pathology behind the attacks.
How triggering factors and conditions aggregate temporally and trigger medical events is encompassed in the concept of chronorisk (8). This has previously been considered in myocardial infarction (9), takutsubo syndrome (10) and epilepsy (11), among others (12,13), and knowledge of fluctuations in disease activity has improved therapeutic regimes in asthma (14), hypertension (15), stroke (16) and epilepsy (17). With its prominent chronobiological traits, CH is a prime target for individualised chronotherapy.
It is altogether likely that the timing of the administration of preventive treatment in CH patients has a significant impact on efficacy and side effects. For example, administration of verapamil, aiming at a constant 24-hour blood concentration, may not be optimal in a disorder characterised by highly specific times of increased chronorisk. On this basis, our aim was to advance the understanding of chronorisk in CH with the hope of providing a basis for future studies on chronotherapy. Specifically, we aimed to investigate whether Gaussian modelling and spectral analysis can be used to describe chronorisk in CH and identify factors that affect this.
Methods
Self-reported hours of attack occurrence for 351 CH patients were retrospectively collected using a questionnaire. The probability distribution of attacks throughout the day (chronorisk) was created for each subgroup by calculating the percentage of all attacks reported for each hour of the day. This 24-hour distribution of chronorisk was analysed with a multimodal Gaussian fit to identify periods of elevated attack risk, and a spectral analysis to identify oscillations. Gaussian modeling can identify and extract single events from a combined dataset. Furthermore, it can quantify these (peak time, amplitude and spread) disentangling them from the aggregation of several events superimposed over one another. Additionally, the timing of these fluctuations can be broken down using spectral analysis. This process identifies periodic oscillations in chronorisk and estimates to which degree fluctuations in risk throughout the day can be attributed to one or more oscillating processes.
Patients
The Danish Cluster Headache Survey has been described previously (6). Briefly, patients diagnosed with CH according to ICHD-II (18) were recruited from June 2012 to March 2016 from The Danish Headache Centre and neurologists in Denmark. A notice was posted on the website and in the newsletter of the Danish patient organisation for CH (www.hortonforeningen.dk). Patients were contacted by mail with information about the study, the questionnaire, and a written consent form. The inclusion criteria for this study were to be between 18–65 and to have CH. Respondents were excluded if they suffered from other chronic primary and secondary headache disorders or if they were unable to differentiate other headache attacks from CH. Both patients in active cluster (attacks within the past month) and outside were included. The study was first approved in a protocol along with two other studies (H-2-2012-016), but after these were finalised, this questionnaire-based study did not need formal approval according to Danish law. The local ethics committee was informed of the continued data collection and had no objections to it (file number: 17008910).
Questionnaire
The questionnaire consisted of 362 items concerning demographics, CH history and attack patterns, sleep, lifestyle and treatment as well as others, and took roughly one hour to complete. Answers were verified by a structured interview conducted by a physician (MB, NL), trained medical student (AP), or trained study nurse (MFF). All questions had to be answered for the questionnaire to be accepted. Patients reported the hours of the day where attacks were most likely to occur. They could also report “no rhythmicity”. The time frame was midnight to midnight. Patients also completed the Pittsburgh Sleep Quality Index (PSQI) and Morningness-Eveningness Questionnaire (MEQ). The PSQI (19) is a validated (20) measure of subjective sleep quality during the past month. It provides a global score; a value > 5 yields high sensitivity and specificity in distinguishing good and poor sleepers. As the PSQI evaluates sleep quality within the past month, only PSQI data from patients with attacks within this period were included. The MEQ (21,22) chronotypes patients into three categories: morning (“lark”), intermediate and evening (“owl”). Chronotype reflects at what time of the day a person prefers to be active, often reduced to sleeping habits. Coffee consumption was recorded as number of cups per day. Tobacco consumption was analysed as number of cigarettes or equivalent per day.
Data analysis
Three analyses of the data were made for the whole population as well as subgroups: a) Gaussian modeling, b) spectral analysis and c) comparison of chronorisk for individual hours. Subgrouping was done according to clinical characteristics (diagnosis (eCH/cCH), PSQI and MEQ score) and medication and lifestyle factors (verapamil, smoking, alcohol, coffee). Chronorisk distributions were created for both the entire patient population and for subgroups. These chronodistributions were fitted with a multimodal Gaussian model with unbound parameters for number of modes, peak times, sigma values and constant (Figure 1(a) and (b)). Goodness of fit was determined by R2 statistics and visual inspection for each subgroup analysis. Obtained values from the multimodal fit were then compared between subgroups. Only modes with an amplitude > 3 were analysed.
Example of analysis of chronorisk. (a) The chronorisk distribution is created by calculating the percentage of all reported CH attacks occurring in each one-hour interval. (b) A multimodal Gaussian fit is performed on the chronorisk distribution and tested for goodness of fit. Black lines represent identified modes. The blue dashed line represents the whole model. Discrete Fourier transformation is applied to the chronorisk distribution, producing (c) a periodogram of the identified frequencies and (d) a visualization of the three most prominent oscillations. These are summed, producing (e) a visualisation of how much oscillations at each given frequency contribute to the total variation of chronorisk. One peak would represent a circadian rhythm; more than one, an ultradian rhythm. Time: Military time (0–24 h).
Spectral analysis of the chronorisk distributions was performed using a discrete Fourier transformation of the distribution resampled to 48 time points using nearest neighbour interpolation. For each chronodistribution, the transformation produced a periodogram representing how much individual oscillation frequencies contribute to the total variance in the chronorisk distribution (Figure 1(c)). For each subgroup, the three most prominent frequencies were identified and plotted as a sum to display how much they each account for of the total chronorisk distribution for the group (Figure 1(d)). This allowed evaluation of circadian (24 hours) and ultradian ( < 24 hours) oscillations. Time asleep was calculated by adding “time to fall asleep” to the reported “bedtime”, obtained in the PSQI.
Statistical analysis
Reported nocturnal (22:00–08:00) differences in chronorisk between two groups for given time intervals were tested for significance using a t-test. Daytime intervals were not tested in this manner, as assumptions about which peaks correspond across groups would be difficult. Differences in reported demographics were compared using a t-test, except for the MEQ where an analysis of variance (ANOVA, Welch's) comparison was made. Normal distribution of data was verified by visual inspection of histograms. Homogeneity of variances was checked with Levene's test. No statistical tests were made to compare power spectra. All data and statistical analyses were performed using Matlab version 2015a (MathWorks, NB, USA) or SAS 9.4 (SAS Institute Inc., NC, USA). All significance levels are presented with p-values corrected for multiple comparisons (Bonferroni).
Results
Patient demographics and clinical characteristics. Data for patients without rhythmicity are presented in the right hand column. Unless otherwise specified, data are presented as mean (SD).
Numerical rating scale 0–4.
Only eCH patients.
Only patients with attacks within the past month (n = 214).
eCH: episodic cluster headache; cCH: chronic cluster headache; CH: cluster headache; PSQI: Pittsburgh Sleep Quality Index.
Gaussian analysis of chronorisk.
Only patients in cluster at the time of answering (n = 214).
p < 0.001.
eCH: episodic cluster headache; cCH: chronic cluster headache; MEQ: morningness-eveningness questionnaire (chronotype); PSQI: Pittsburgh Sleep Quality Index. Time: Military time (0–24h).
Influence of clinical characteristics on attack rhythm
In eCH, patients’ circadian rhythmicity dominated, as opposed to cCH, where ultradian rhythms were more pronounced (Figure 2, Table 3). Further, eCH patients had their primary nocturnal peak over one hour earlier than cCH patients (01:28 vs. 02:33, Table 2). There was no apparent difference in the morning peak. Chronorisk in eCH at 01:00 was higher than in cCH (p = 0.003).
Oscillations and spectral analysis for episodic (n = 190) and chronic (n = 96) patients. Time: Military time (0–24 h). Spectral analyses of chronorisk. Data are presented as total power (%). Only patients in cluster at the time of answering (n = 214). eCH: episodic cluster headache; cCH: chronic cluster headache; MEQ: morningness-eveningness questionnaire (chronotype); PSQI: Pittsburgh Sleep Quality Index.
Patient chronotype influenced the occurrence of the primary nocturnal peak, the morning type being earliest (00:50), the neither type later (01:02) and the evening type later still (02:11) (Supplementary Figure 1). The morning subgroup had a very early morning peak, and the two other groups a peak around 06:00. In the spectral analysis, the morning group presented more power in the ultradian frequencies, which were not prominent in the neither and evening groups (Table 3). In the hour-specific analysis, there were no differences in chronorisk after correction for multiple comparisons.
Poor sleepers (PSQI > 5) had a chronodistribution with a nocturnal focus, with three prominent peaks (21:46, 02:16, 06:03), whereas good sleepers (PSQI ≤ 5) had distinct peaks early in the night and throughout the day (Table 2, Supplementary Figure 2). This was reflected in the spectral analysis, where good sleepers presented higher power in the ultradian frequencies (52.3%) than poor sleepers (32.4%). In the hour-specific comparison, there were no differences in chronorisk after correction for multiple comparisons.
Patients were treated with a variety of medications including verapamil (n = 164), lithium (n = 14), gabapentin (n = 10), GON-blocks (n = 10), topiramate (n = 7), prednisone (n = 5), indomethacin (n = 5), methysergide (n = 3), melatonin (n = 1). Only patients taking verapamil had a large enough sample size for analysis. Patients not taking verapamil had an earlier nocturnal peak (01:53) than patients taking this preventive medication (02:43). The difference in the morning peak followed the same pattern but was not as pronounced (06:04 vs. 06:51). There were no apparent differences in the spectral analysis of these groups.
Consuming tobacco, alcohol and coffee was associated with an earlier primary nocturnal chronorisk peak. Smoking and alcohol did not seem to affect the morning peak, but coffee did, with abstainers' morning peak occurring later (07:05 vs. 06:07). Smoking and alcohol did not affect chronorisk oscillation in any apparent manner. However, drinking coffee was associated with lower power in the ultradian frequencies compared to abstaining (24.2% vs. 47.9%), indicating that coffee consumption may be associated with a lower daytime chronorisk (Supplementary Figure 3).
Influence of time asleep and attack frequency
Differences in the timing of the nocturnal peak could be caused by differences not just in bedtime, but also in how long it takes to fall asleep. For patients in cluster (n = 214), time asleep differed significantly between the chronotypes and smokers and non-smokers (Table 2). There were no significant differences in the number of attacks reported by any of the subgroups (Supplementary Table 1).
Discussion
In 286 CH patients with diurnal rhythmicity in attack occurrence, we found that Gaussian modeling accurately characterised individual events of elevated chronorisk and the sum of these events could reliably account for the entire 24-hour chronorisk distribution. Several of the analysed subgroups differed in chronorisk patterns, also when taking differences in bedtime into account. Of the analysed factors, those associated with earlier nocturnal peaks in chronorisk were: Being episodic, morning chronotype, a good sleeper, not taking verapamil, not smoking, consuming alcohol during the bout, and drinking coffee. Being chronic, morning chronotype, a good sleeper and abstaining from coffee was associated with stronger ultradian oscillations. The findings are significant, and have implications for future research as they indicate that there are different pathophysiological mechanisms at play in the episodic and chronic subtypes that need to be explored further.
The chronobiological features of CH are perhaps the most pronounced of any of the primary headaches, and also the best studied. In an Italian population, Manzoni et al. found the highest peak during the afternoon, not the night (3). However, in a Norwegian population, Russell did not identify this afternoon peak: instead, he found that 75% of attacks occurred between 21:00 and 10:00 hours (4). Compared to these studies, the findings presented here, and earlier by our group (1,6) are more congruent with Russell's findings. These early findings showed that the temporal profile of attacks depends on the population studied, and hinted that different factors may influence this timing. This study represents the first attempt to assess chronorisk in a population of headache patients and to use Gaussian modeling and Fourier transformations to do so. Existing literature has focused intensely on the frequency of attacks in subgroups of patients, but has never attempted to identify how timing and rhythmicity of attacks differ. This seems especially relevant since research efforts are headed towards investigating patients, not in the ictal phase, but in the pre- and post-ictal phases (23). Therefore, a better understanding of patterns of chronorisk is important for future studies.
With the very distinct chronobiological features, and the influence of a variety of factors, CH is a ripe target for individualised chronotherapy – the administration of therapeutic interventions governed by knowledge of chronobiological processes for maximum therapeutic effect and minimal side-effects. In the optimisation of treatment, not only should chronorisk be taken into account, but also the varying clearance speed of these drugs. For example, constant rate administration of valproic acid, a third-choice preventive drug for CH, showed significant 24-hour differences in clearance speed (24). In our data, the bedtime of patients taking verapamil did not differ from other patients. However, patients taking verapamil had their nocturnal attack peak later, raising the question of a delaying effect of verapamil and possible advantages of an evening dose. Alternatively, it could be that verapamil affects the progression of the sleep cycle and any attacks triggered by this cycle (25), or another reason altogether. Thus, further studies on the efficacy of preventives taking into account the chronodistribution of attacks are warranted.
The issue of medications and chronobiology is especially relevant in CH. For example, corticosteroids are known to have a number of neuropsychiatric side effects that may in turn affect sleep (26). They may also affect sleep quality and nocturnal urine production, leading to nocturia, again disturbing sleep (27). Unfortunately, the sample size of patients taking preventive medications other than verapamil in this study did not allow for detailed analysis. Nor was the number of these patients high enough that it seems likely that they may have influenced the results significantly.
Chronorisk analysis can also be used to navigate cause and effect of the many factors suspected to influence CH attacks. Sleep factors alone include sleep duration, sleep cycle, sleep fragmentation and hypnophobia (28,29). For example, in this study, the subgroup of patients who smoke had a different nocturnal peak than other patients, but they also reported different times asleep. Isolating the specific factors that affect nocturnal attacks requires the sensitivity that large amounts of data can provide. In this case, a chronorisk study with more detailed sleep data could disentangle these factors, giving clinicians a more nuanced tool when advising patients.
cCH is characterised by a relatively higher daytime chronorisk of attacks in patients reporting diurnal rhythmicity. A previous case report analysing > 5000 attacks also found ultradian oscillations in a cCH patient (30). This indicates that preventive treatment for cCH patients should cover the whole day whereas eCH patients are more vulnerable at night and therapeutic efforts should be targeted at this time of day, however, still on an individual basis. The pathophysiological implications of this finding regarding differences between eCH and cCH patients are interesting. A possible explanation could be that there is less influence of the suprachiasmatic nucleus in cCH patients as this structure is a strong circadian, and not ultradian, oscillator. This is substantiated by a stronger response to melatonin in eCH than cCH (31,32). The hypocretins have been found to modulate the activity of the suprachiasmatic nucleus (33), and the previously identified changes in this signaling in CH may play a role (34). Social factors may also be part of the explanation, as cCH patients are more impaired and may be on social disability pension or compensations whereas eCH patients to a greater extent may have full-time jobs. On this note, a previous Italian study identified a peak in the early afternoon, possibly attributed to napping, which was not found here (3). This suggests that cultural factors influence the timing of attacks.
Lastly, the identified differences between coffee drinkers and coffee abstainers are particularly interesting. CH patients have a reputation for consuming large amounts of coffee, perhaps reinforced by high nicotine consumption. It could be theorised that it is a form of self-medication, as caffeine is effective in the treatment of other primary headache disorders (35). The timing of caffeine consumption is important, as this has been shown to affect sleep, although caffeine kinetics are highly individual (36). Caffeine affects sleep even if taken up to 6 hours before by reducing the duration of non-REM sleep and decreasing the overall duration of sleep (37), possibly through increased sympathetic activity (38). Taking into consideration the poor sleep quality in CH patients, caffeine consumption may be a double-edged sword, on one hand possibly reducing pain, on the other hand disrupting sleep.
Methodological considerations
This study is strengthened by a large sample size of well characterised patients diagnosed by headache experts. Limitations are partially similar to other questionnaire-based studies, which are prone to recall bias. In this study, recall bias could lead to patients not remembering the timing of their attacks correctly. Also, the nature of questionnaire studies may limit patients in their ability to report individual nuances. An element of digit preference exceeding one hour (i.e. over inflation of standard times) cannot be excluded. However, we believe the influence of this to be minimal, as we did not observe such standard times in our analyses – for example, noon, midnight or 18:00 hours.
A further limitation may be unique to CH: Treating attacks with oxygen could result in a rebound phenomenon. Therefore, patients using this treatment may report more attacks than those not using it. Here, most of the patients were in bout when answering, which reduces this influence, and we took into consideration when the patient fell asleep. We have attempted to separate the effect of different factors which may influence chronorisk, but it is likely that individual factors also affect each other – for example, smoking and coffee or alcohol consumption. The employed methodology, using retrospective data, where peak times represent pooled data, not individual data, does not allow for analysis of how these different factors may affect each other. However, future studies using prospective data could investigate, for example, the timing of the first nocturnal attack. Another possibility, using such data, would be to investigate the development of chronorisk over the course of the cluster period. It is possible that patients may phase shift forward or backward in relation to the exposure to changing zeitgebers, time cues for the body's circadian rhythms, over the course of the cluster – for example bed time. Contradicting this theory is a case report where cyclic nocturnal awakening occurred at the same hours of impending cluster attacks (39).
Quantitative analysis of chronorisk also has its limitations. Multimodal Gaussian fits, like Fourier analysis, by nature can accurately describe any distribution if a sufficiently large number of modes or frequencies are used. The limits of five modes and three frequencies used in this study prevent over-fitting but are based on personal judgment. Analysis of the residuals did not find any systematic variance, and controls where a higher number of modes were included did not improve the goodness of fit. Power estimates to determine necessary sample sizes for accurate modelling are not possible, again requiring an element of qualitative judgement. We did not model any groups with fewer than 130 attacks for this reason. The discrete Fourier transform also has the limitation that its output is in complete oscillations such as 1, 2, 3, and so on, per day. Other oscillations, such as 1.5 a day or weekly oscillations cannot be detected. Even though the analysis is quantitative, the biological interpretation of oscillations is qualitative. This is partly reflected in the hour-by-hour analysis for significance which, after correction for multiple comparisons, only detected significant differences in hourly chronorisk between two subgroups. This study, like other questionnaire-based studies, does not retain nuances in attack data due to recall bias, digit bias and other forms of generalisation. Such factors are important when modelling the data, as variation in both the number and timing of a patient's daily attacks is minimised by the reporting method, which in turn may artificially reduce the standard errors of model fits. It also limits the ability to properly address the issue of the attacks from several patients being combined in an analysis (non-independence). This study employed patient interviews to assess the degree of variation for individual patients and normalised attack data from individual patients in subgroup analyses to minimise, though not remove, the aggregated effect. A method where single attacks are registered with a timestamp to quantify intra-subject variation would be preferable and would allow more extensive statistical assessments than the primarily subjective assessments presented in this study. Such diary records would also better enable the use of multilevel mixture models where approaches to properly address the non-independence of attacks within individuals can be employed.
Future studies should employ prospective recordings to identify ultra- and infradian oscillations across populations and individual patients. Studies taking into account these attack rhythms in the timing of preventive treatment are warranted. Using this method, it may be possible to predict times of increased chronorisk, also in patients who think there is no clear rhythm in their attacks. Preventive therapy could be administered just before these times for increased effectiveness, and perhaps lead to a better side-effect profile with decreased dosage during times of decreased risk.
In conclusion, chronorisk in patients suffering from CH follows a multimodal Gaussian distribution and is influenced by a variety of factors. eCH patients have a relatively lower daytime risk and prominent circadian rhythms compared to cCH patients, in whom ultradian rhythms are stronger reflecting a loss of association with sleep. Further studies are needed to investigate the variation over time, the underlying pathophysiological mechanisms and, most importantly, the possible influence of timing of the administration of preventive medications, which could in turn lead to individualised therapy. The findings demonstrate the possibilities in applying Gaussian modelling and spectral analysis to chronobiological signals and its potential should be validated in other diseases.
Supplemental Material
Supplementary Figure 1 -Supplemental material for Chronorisk in cluster headache: A tool for individualised therapy?
Supplemental material, Supplementary Figure 1 for Chronorisk in cluster headache: A tool for individualised therapy? by Mads Barloese, Bryan Haddock, Nunu T Lund, Anja Petersen and Rigmor Jensen in Cephalalgia
Supplemental Material
Supplementary Figure 2 -Supplemental material for Chronorisk in cluster headache: A tool for individualised therapy?
Supplemental material, Supplementary Figure 2 for Chronorisk in cluster headache: A tool for individualised therapy? by Mads Barloese, Bryan Haddock, Nunu T Lund, Anja Petersen and Rigmor Jensen in Cephalalgia
Supplemental Material
Supplementary Figure 3 -Supplemental material for Chronorisk in cluster headache: A tool for individualised therapy?
Supplemental material, Supplementary Figure 3 for Chronorisk in cluster headache: A tool for individualised therapy? by Mads Barloese, Bryan Haddock, Nunu T Lund, Anja Petersen and Rigmor Jensen in Cephalalgia
Supplemental Material
Supplementary Table 1 -Supplemental material for Chronorisk in cluster headache: A tool for individualised therapy?
Supplemental material, Supplementary Table 1 for Chronorisk in cluster headache: A tool for individualised therapy? by Mads Barloese, Bryan Haddock, Nunu T Lund, Anja Petersen and Rigmor Jensen in Cephalalgia
Footnotes
Key findings
Chronorisk in cluster headache can be described using a multimodal Gaussian distribution.
Episodic cluster headache patients have a relatively lower daytime chronorisk than chronic cluster headache patients.
Attack patterns in chronic cluster headache are more strongly influenced by ultradian rhythms than in the episodic variant.
Several factors seem to influence the timing of attacks and these must be considered, should chronotherapy be tested in cluster headache.
Acknowledgements
The authors wish to thank study nurse Mette Fisker (MFF) for her efforts with the questionnaires.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
