Pain thresholds and suprathreshold pain after sleep restriction in migraine – A blinded crossover study

Objective There is an unexplained association between disturbed sleep and migraine. In this blinded crossover study, we investigate if experimental sleep restriction has a different effect on pain thresholds and suprathreshold pain in interictal migraineurs and controls. Methods Forearm heat pain thresholds and tolerance thresholds, and trapezius pressure pain thresholds and suprathreshold pain were measured in 39 interictal migraineurs and 31 healthy controls after two consecutive nights of partial sleep restriction and after habitual sleep. Results The effect of sleep restriction was not significantly different between interictal migraineurs and controls in the primary analyses. Pressure pain thresholds tended to be lower (i.e., increased pain sensitivity) after sleep restriction in interictal migraineurs compared to controls with a 48-hour preictal-interictal cut-off (p = 0.061). We found decreased pain thresholds after sleep restriction in two of seven migraine subgroup comparisons: heat pain thresholds decreased in migraineurs with lower pain intensity during attacks (p = 0.005) and pressure pain thresholds decreased in migraineurs with higher severity of photophobia during attacks (p = 0.031). Heat pain thresholds tended to decrease after sleep restriction in sleep-related migraine (p = 0.060). Sleep restriction did not affect suprathreshold pain measurements in either group. Conclusion This study could not provide strong evidence for an increased effect of sleep restriction on pain sensitivity in migraineurs compared to healthy controls. There might be a slightly increased effect of sleep restriction in migraineurs, detectable using large samples or more pronounced in certain migraine subgroups.

. Example of pressure measurements using an algometer from the left trapezius of one test subject. Left: force in Newton (N) with time elapsed during one recording (using a 1 cm 2 probe, 10 N correspond to 100 kPa). The solid line is the actual force applied plotted against time elapsed, superimposed on a fitted line (dashed). Right: subjective pain on a visual analogue scale (VAS) in response to force. The solid line shows the VAS scored by the test subject on a manual VAS device as force against the trapezius muscle was increased with a constant rate. The dashed line is the fitted regression lines from measurements of force and VAS. These regression lines with their corresponding correlation coefficients are used to calculate pressure pain threshold (PPT) and pressure at VAS = 50/100 (PP5). The recordings were excluded in the case of R 2 < 0.80. In this case, R 2 was 0.95 for both the left and right sides. We used a FDMIX digital hand-held force gauge instrument (Wagner instruments, Greenwich, U.S.A., probe size 1 cm 2 ) to apply pressure to the trapezius muscles, at points 1/3 from the posterior edge of the acromion to the C7. A custom-written program (National Institute of Occupational Health, Norway) providing real-time visual feedback of force, was used to ensure an increment of 50 kPa/sec 16 .
Supplementary Table S2. Pain and tolerance thresholds in interictal migraineurs and controls after habitual sleep and restricted sleep for the primary response variables in the secondary analysis, using a 48-hour cutoff for the preictal phase. Means are shown as difference from baseline temperature of 32 °C, and absolute force (N). (using a 1cm 2 probe, 10 N correspond to 100 kPa). PP5 was calculated based on a linear fit for force plotted against pain from a visual analogue scale (VAS). a,b P-values of the difference between the two sleep conditions in the control and migraine groups, respectively. c,d P-values of the difference between migraineurs and controls for habitual sleep and restricted sleep, respectively. e P-value of the interaction between group and sleep condition. Random parameters are shown in Table S5. There was a significantly higher suprathreshold heat pain sensitivity in the migraineurs for the SR condition, and an analogous trend for the habitual sleep condition. Except for the fluctuations relative to the thresholds for significance, these findings are quite similar to those of the primary analysis. E.g. the difference in point estimates between migraine and control group for HPTT was 1.1 °C in the primary analysis, and 1.3 °C in the secondary analysis. In this analysis, there were trends toward higher heat and pressure pain sensitivity after SR (PPT and HPT), but only for the migraine subgroup. For pressure pain, there was also a trend toward a different effect of SR between the migraineurs and controls. Suprathreshold pressure pain sensitivity (PP5) seems to be lower in migraineurs based on the point estimates, but were not significant for neither sleep conditions, due to large variance. 0.160 0.001 Level 1: residuals 0.045 0.000 HPT: heat pain threshold. HPTT: heat pain tolerance threshold. CI: confidence interval. PPT: pressure pain threshold. PP5: pressure at VAS = 50/100. Ln: natural logarithm. N: Newton. HPT was square-root transformed, while HPTT and PP5 were power transformed, and PPT was log transformed, to improve normality of residuals. Coefficients presented are transformed and should be interpreted as such. The constant refers to mean thresholds for the control group for the habitual sleep condition. Simple effect of group: change in thresholds in migraineurs compared with controls for the habitual sleep condition. Simple effect of sleep: change in threshold after sleep restriction (SR) compared to habitual sleep for controls. The interaction refers to the difference between the effect of SR compared to habitual sleep in migraine compared to controls. Random effects are presented as variances. Random parameters and covariance matrices were included based on likelihood ratio (LR) tests. An unstructured variance-covariance matrix was used for HPT. We used maximum likelihood estimation (MLE) for HPT and HPTT, whereas restricted maximum-likelihood estimation (REML) was used for PPT and PP5. For HPT and HPTT we used the sandwich estimator, to account for less than normally distributed residuals. a p = 0.051. Table S5. Model specifications for the primary response variables in the secondary analysis, using a 48-hour cutoff for the preictal phase.
[ 0.165 0.001 Level 1: residuals 0.046 0.000 HPT: heat pain threshold. HPTT: heat pain tolerance threshold. CI: confidence interval. PPT: pressure pain threshold. PP5: pressure at VAS = 50/100. Ln: natural logarithm. N: Newton. HPT was square-root transformed, while HPTT and PP5 were power transformed, and PPT was log transformed, to improve normality of residuals. Coefficients presented are transformed and should be interpreted as such. The constant refers to mean thresholds for the control group for the habitual sleep condition. Simple effect of group: change in thresholds in migraineurs compared with controls for the habitual sleep condition. Simple effect of sleep: change in threshold after sleep restriction (SR) compared to habitual sleep for controls. The interaction refers to the difference between the effect of SR compared to habitual sleep in migraine compared to controls. Random effects are presented as variances. Random parameters and covariance matrices were included based on likelihood ratio (LR) tests. We used maximum likelihood estimation (MLE) and an unstructured variance-covariance matrix for HPT, whereas restricted maximum-likelihood estimation (REML) was used for HPTT, PPT, and PP5. We used the sandwich estimator for HPT, to account for less than normally distributed residuals. a p = 0.071. b p = 0.061.
Supplementary Table S6. Model specifications for the warm detection threshold (WDT, secondary response) in the primary and secondary analyses, using 24-and 48-hour cutoffs for the preictal phase, respectively. WDT (•C -0.2 ) Primary analysis Secondary analysis Coef.
[ WDT was power transformed to improve normality of residuals. Coefficients presented are transformed and should be interpreted as such. The constant refers to mean thresholds for the control group for the habitual sleep condition. Simple effect of group: change in thresholds in migraineurs compared with controls for the habitual sleep condition. Simple effect of sleep: change in threshold after SR compared to habitual sleep for controls. The interaction refers to the difference between the effect of SR compared to habitual sleep in migraine compared to controls. Random effects are presented as variances. Random parameters and covariance matrices were included based on likelihood ratio (LR) tests. We used maximum likelihood estimation (MLE) and the sandwich estimator for WDT. Main effects of group and sleep condition were included in the models, but are omitted here, as the main effect of the clinical variables was the focus of these analyses. Hence, significant effects represent an effect of the clinical variable on thresholds. For the TST analysis, the interaction between TST and group (migraine vs. control) was also included. Random effects are presented as variances. Random parameters and covariance matrices were included based on likelihood ratio (LR) tests. An unstructured variancecovariance matrix was used for all cases of HPT. We used maximum likelihood estimation (MLE) for HPT and HPTT except for 'NSM vs. SM', whereas restricted maximum-likelihood estimation (REML) was used for PPT and HPT and HPTT in 'NSM vs. SM'. For HPT and HPTT we used the sandwich estimator in all models, to account for less than normally distributed residuals, except for 'NSM vs. SM'. The Kenward-Roger approximation for small sample inference was used in the 'NSM vs-SM'-analysis. TST was recorded using actigraphy. Rest intervals defined by the actigraphy software were corrected semi-manually in a hierarchal manor, using the rest intervals, actigraphy event marker for 'lights off' and 'lights on', 'lights off' and 'lights on' from sleep diary, and light and activity levels from actigraph. d Interaction for TST: The interaction refers to the combined effect of migraine and TST on thresholds. CI: confidence interval. HPT: heat pain threshold. PPT: pressure pain threshold. Ln: natural logarithm. N: Newton. HPTT: heat pain tolerance threshold. Model specifications for exploratory analyses, were we extended the models on HPT, PPT, HPTT. To improve normality of residuals, HPT was square-root transformed, PPT was log transformed, and HPTT was power transformed. Coefficients presented are transformed and should be interpreted as such. The constant refers to mean thresholds for the control group for the habitual sleep condition. Main effect of clinical variable: change in thresholds between the groups for both sleep conditions. The interaction refers to the combined effect of the clinical variable and sleep restriction (SR) on thresholds. Main effects of sleep condition were included in the models, but are omitted here, as the main effects (and simple effects in case of significant main effects or interactions) of the clinical variables and the interaction were the foci of these analyses. Hence, significant effects represent an effect of the clinical variable or a different effect of sleep restriction for each of the levels of the clinical variable. Random effects are presented as variances. Random parameters and covariance matrices were included based on likelihood ratio (LR) tests. An unstructured variancecovariance matrix was used for all cases of HPT, except for the 'intensity of photophobia'-model. We used maximum likelihood estimation (MLE) for HPT and HPTT, whereas restricted maximum-likelihood estimation (REML) was used for PPT. For HPT, HPTT, as well as the model including intensity of photophobia during attacks and PPT, we used the sandwich estimator, to account for less than normally distributed residuals. a Average duration of headache with or without use of medication. b Age was included as a covariate in these analyses but did not produce significant results. c 1: mild/moderate, 2: severe. d 1: < 4 days/month 2: > 4 days/month. e 1: < ¾ of attacks 2: almost always. 1 p = 0.085. 2 p = 0.049. 3 p = 0.059. 4 p = 0.025. 5 p = 0.045. 6 p = 0.058. 7 p = 0.019. Figure S2. Heat pain threshold (HPT). Graphical display of estimated margins with 95 % confidence intervals from a multilevel model, with the effect of habitual vs. restricted sleep (SR) in migraineurs with non-sleep-related migraine (NSM) compared to sleep-related migraine (SM). There was a main effect of reduced HPT after SR (p = 0.040), related to lower HPT after SR in the SM group (p = 0.0595). HPT tended to be lower in the NSM group for the habitual sleep condition (p = 0.071). The SM group was small (see Supplementary Table S9) and the results should be interpreted with caution. Supplementary Figure S3. Example of one day/night from the sleep diary used in our study. The sleep diary is written in Norwegian (Dato = date; onsdag = Wednesday). The participants were instructed to fill out information about the previous night every morning. The figure below is an example of one day/night from the sleep diary. ↓: Lied down in bed. ↑ Woke up. o: Turned off the lights. Black squares: Sleep. White squares: Awake.

Headache diary
The headache diary was a form with one row for each day and columns for the following SM-analysis.

Correcting the rest interval in actigraphy
Total sleep time was collected using actigraphy. The rest intervals defined by the actigraphy software were corrected semi-manually in a hierarchal manor, using the rest intervals, actigraphy event marker for 'lights off' and 'lights on', 'lights off' and 'lights on' from sleep diary, and light and activity levels from actigraph.

Model specifications
We used STATA version 15.1 (StataCorp LLC) and multilevel models in this study. Multilevel models are often used in data with hierarchal structures, where data are nested in levels or clusters (i.e. students in classes in schools, or in our case repeated measurements in subjects). Observations in students in the same class (level, cluster) cannot be assumed to be independent, and multilevel models accounts for this by considering some coefficients as random in addition to traditional fixed coefficients 23 . These can be random intercepts and random slopes for groups, i.e. allowing intercepts and slopes to be different for different groups within levels. Hence, multilevel models are also termed hierarchal models, mixed models, random-effects models, random-coefficient models, and mixed effects models 23 .
One advantage of multilevel models is their handling of missing values.
Fixed effects were decided a priori on the basis of hypotheses, while random parameters and covariance matrices were included based on likelihood ratio (LR) tests 23 . We categorised the analyses in a primary, a secondary, and an exploratory analysis: Using a 24hour cutoff for the preictal phase in the primary analysis 11, 12 , a 48-hour cutoff in the secondary analysis 14 , and an exploratory analysis investigating addition of clinical migraine and sleep parameters in models from the primary analysis. Marginal means with 95 % confidence intervals and p-values from the multilevel models are shown in Table 4 and 5 for the primary analysis, and in Supplementary Table S2 and S3 for the secondary analysis.
We used exponential transformation on WDT, HPTT, and PP5, square-root transformation on HPT, and logarithmic transformation on PPT. The transformation yielding the best fit for each of the dependent variables was decided based on the Box-Cox transformation.
In selected cases in the primary, secondary, and exploratory analyses we used the sandwich estimator to estimate robust variances. This was done to account for cases with less than normally distributed residuals 23 . We aimed to use restricted maximum-likelihood estimation (REML) in all analyses, as REML is less biased for balanced data compared to maximum-likelihood estimation (MLE) 23 . However, REML is incompatible with the sandwich estimator and REML was therefore only used in cases where the sandwich estimator was not used, i.e. in cases with satisfactory normal distribution. Model specifications, random parameters and use of MLE or REML for each of the models in the primary, secondary, and exploratory analyses are shown in Supplementary Table S4-S8, respectively.
In the primary analysis, MLE and the sandwich estimator were used for WDT, HPT, and estimator were used for WDT and HPT, whereas REML was used for HPTT, PPT, and PP5. In the exploratory analyses, MLE and the sandwich estimator was used for HPT and HPTT, except for the intensity of photophobia and non-sleep-related migraine (NSM) vs. sleeprelated migraine (SM) analyses, where REML was used. REML was used for PPT except in the intensity of photophobia analysis, where MLE and the sandwich estimator was used.
Due to safety considerations, the upper limit of the HPTT was set at 52 °C. Several of the included recordings reached this limit (8.9 %), and therefore we do to know what the 'true' thresholds would be in the absence of an upper limit. Such ceiling effects can lead to biased parameter estimation, perhaps especially in longitudinal studies. Consequently, we conducted an additional regression analysis on HPTT, using the Tobit regression model, and defining measures ≥ 52 °C as censored. The Tobit regression model assumes a specific distribution of the 'true values' and performs better in the case of censored data in a longitudinal study design 22 . There is some added difficulty in estimation of parameters when using Tobit regression models, that becomes relevant with complex models and/or in the presence of several random coefficients 22 . Our models are quite simple, including few random coefficients, so the use of a Tobit regression model seems appropriate in our case.
The usefulness is also dependent on the proportion of censored measurements, where a low proportion will lead to increased similarity between the Tobit regression model and a standard multilevel model 22 . There were no relevant differences between results from the standard multilevel model and the Tobit regression model for HPTT.