This analysis was part of the mBrain21 study, a comprehensive and longitudinal research initiative as detailed in De Brouwer et al. [22]. The goal of the mBrain21 study was to provide a profound understanding of migraine manifestations within ambulatory environments. To achieve this, study participants were equipped with wrist-worn wearable devices, specifically the Empatica E4®, for on average 90 days, to gather relevant physiological data. The physiological modalities acquired via the E4 device include skin temperature (4 Hz), skin conductance (4 Hz), acceleration (32 Hz), and blood volume pulse (64 Hz) from which the heart rate is derived. The Empatica devices were connected via Bluetooth to a dedicated headache diary smartphone application we developed. Through this application, shown in Fig. 1, participants could log their headache occurrences, characteristics (intensity, location, medication usage and effectiveness), and associated symptoms consistent with ICHD-3 criteria. In addition, specific medication usage, responses to daily morning and evening questionnaires, and behavioral actions including food intake moments are also questioned [22]. Meanwhile, other behaviors, such as physical activities, sleep, and stress-related events, were automatically inferred from the wearable and smartphone sensor data. This integrated setup enabled continuous, real-time, and objective data collection under free-living conditions over an extended period.

Fig. 1

Headache registration interface of the mBrain21 phone application. Figure 1 [a] shows the overview of headache events registered in the timeline view of the application. Panel [b] visualizes the headache registry module with different steps to be completed

Participants

Study participants were eligible for the study if they were aged 18 to 65, had a migraine history of more than one year as diagnosed by neurologists specializing in headache disorders (NV, KP), and met the ICHD-3 criteria for episodic or chronic migraine. Additional requirements included experiencing at least five days per month free from headaches or related symptoms, migraine onset before age 50, and owning a smartphone with Android Operating System version 8.0 or higher.

Exclusion criteria encompassed any diagnosed headache disorder other than TTH, medication-overuse headaches (as defined by ICHD-3 8.2 and its subsections), daily headaches without pain-free moments, presence of other chronic pain syndromes, significant medical comorbidity that could affect the study outcomes, opioid or barbiturate use, illicit drug or alcohol abuse, or current/planned pregnancy. Participants could also not simultaneously participate in any other academic or commercial medical study.

Study design

Participants were recruited between July 2021 and August 2023. The study involved two in-hospital visits: an initial baseline visit and a final study visit after 90 days. During the baseline visit, all study participants were interviewed by a physician researcher and neurologist with expertise in the field of clinical headache disorders (NV). Participants received instructions on how to wear the Empatica E4® actigraph on their non-dominant wrist, connecting the device to their personal smartphone via Bluetooth, and utilizing the accompanying smartphone application for registering headache events and completing daily morning and evening questionnaires.

During the study period, participants were instructed to wear the device as much as possible (daytime and nighttime), to ensure consistent data collection and minimal disruption to their daily routines. They were also advised to charge the wearable at least once a day, ideally in the evening, before bedtime. Participants were encouraged to perform their regular daily activities during the measurement period. However, they were asked to take off the Empatica E4 device when engaging in activities involving water (e.g., showering and swimming), heat (e.g., barbecuing), or cold (e.g., working in freezers).

Data processingWrist-worn accelerometer data processing

In this section, we describe the methodology, illustrated in Fig. 2, for converting raw accelerometer data from the Empatica device (measured in gravitational (g) units) into delta features for our analysis.

Fig. 2

Flowchart of the wearable data processing pipeline. Abbreviations: g = gravitational unit, ACC = accelerometer, Hz = Hertz; w = window, s = stride, AIABS = absolute activity index, Δ = delta

The first processing step involves identifying and eliminating periods of wearable non-wear. These are intervals where the wrist-worn device is not being worn, but continues to record data. Addressing non-wear periods is a recognized challenge in real-world actigraphy research [25]. To address this, a custom algorithm was developed that identifies non-wear periods by analyzing the device’s movement, skin-temperature, and skin conductance signals, and combines the signal quality indices (SQIs) of these individual signals to create an on-body status signal. Further details can be found in Van Der Donckt et al. (2024) [26]. Figure 3 provides a visual example of the resulting output of the employed non-wear detection algorithm applied to the physiological signals captured by the wearable.

Fig. 3

Visual overview of the non-wear detection algorithm used on an Empatica E4 excerpt, derived from Van Der Donckt et al. [26]. The red-shaded area in each subplot highlights a manually labeled non-wear interval. Subplots (i) and (ii) display signal-specific signal quality indices (SQIs) for the skin conductance (“EDA”) and temperature (“TMP”), respectively. Subplot (iii) shows the standard deviation of the ACC x-axis and the corresponding ACC-SD SQI. Subplot (iv) presents the raw three-axis accelerometer data along with the resulting “Wrist_SQI”, which combines the signal-specific SQIs from the above subplots. A low Wrist_SQI value between 08:55 and 09:00 denotes non-wear. Abbreviations: ACC = accelerometer; SD = standard deviation; SQI = signal quality index, EDA = electrodermal activity, TMP = skin temperature

After excluding non-wear intervals, the remaining accelerometer data is transformed into two distinct signal groups. The first group consists of the Absolute Activity Index (AIABS), calculated using a 1-s window and 1-s stride, as recommended by Bai et al. [27]. This signal approximates activity energy expenditure (AEE). For a visual overview and further details on the AIABS transformation, we refer to Fig. 4 and the work of Vandenbussche et al. [28]. The second signal group is derived from a Human Activity Recognition (HAR) ML model, which uses a 15-s window and 7.5-s stride – the default configuration for this model. The utilized HAR ML model was specifically developed for the mBrain21 study, and was trained on an in-house dataset, as described in Van Der Donckt et al. (2022) [29]. At each stride step (7.5 s), the model predicts the probabilities of six distinct activity types: [“Lying”, “Sitting”, “Standing”, “Walking”, “Running”, “Cycling”], see subplot b (III) of Fig. 4.

Fig. 4

Detailed overview of AIABS computation, its distribution characteristics, and HAR predictions with the derived movement ratio. To calculate the AIABS, the simplified equation given in panel [a] was used, which assumes a systematic noise-variance σi of zero. In panel [b] an Empatica E4 accelerometer excerpt of a study participant (i) is transformed into AIABS values (with window-length ? = 1 s) (ii). Panel [c] displays the distribution of the AIABS values of subplot (ii) alongside the threshold values, whose ratio features are utilized for the eventual analysis. In addition, panel [b] demonstrates in subplot (iii) the HAR ML activity prediction probabilities of each class. Lastly, subplot (iv) visualizes the derived “movement ratio” signal from these HAR ML predictions. Abbreviations: AIABS = absolute activity index; ? = variance window length; Hz = Hertz, HAR = human activity recognition; ML = machine learning; t = time point; σm = signal variance of wrist acceleration over axis x, y or z

The HAR analysis in this work primarily focuses on two key activity classifications: "Lying" and "Walking". These activities were chosen for several reasons (i) the HAR ML demonstrates a high validation accuracy for “Walking”, and previous research has identified “Lying” as a common behavioral response during migraine headaches [5], (ii) “Walking” is more commonly performed by the general population throughout the day (unlike “Cycling” and “Running”), and (iii) activities such as “Standing” and “Sitting” may cover a wide spectrum of movement intensities and behaviors resulting in varying AEE, which may obscure the results [30]. As such, we hypothesize that differences in movement behavior are most likely to be observed in the “Lying” and “Walking” activities. Additionally, we introduce a "Movement Ratio" signal which aggregates the prediction probabilities for "Walking," "Running," and "Cycling" at each time step, serving as a comprehensive measure of all global body movement activities. Figure 4, panel [b], displays the HAR prediction signals and the “Movement Ratio” signal in subplots (iii) and (iv), respectively.

Eligibility criteria for analysis of headaches and corresponding non-headache periods

To ensure accurate analysis of the registered headache periods, we utilized the entry time metadata—the timestamp when the event was logged in the mobile application—to identify potential reporting biases. Headache events were flagged for a high probability of recall bias if they were logged more than 24 h after the reported end time. Similarly, events were flagged for a high probability of predictive bias if their final update occurred more than two hours before the reported end time. Headache records that were flagged for any of the two criteria were excluded from all subsequent analyses. Table 5. shows the number of remaining bias free headache events after this exclusion process.

For the intervals deemed eligible for analysis, a signal data availability threshold of 95% was applied, which was determined using the methodology outlined in Van Der Donckt et al. (2024) [26]. This threshold strikes a pragmatic balance between metric stability (affected by missing data) and sample retention. Supplementary Fig. 1 provides an overview of the distribution spread of signal metric features across various data retention ratios.

Non-headache periods were matched to each bias-free headache period based on specific criteria. To factor out any implicitness, non-headache periods were selected only from days explicitly reported as headache-free in the following morning’s questionnaire. These non-headache periods (i) intersect with the time interval of the corresponding headache period, (ii) must contain wearable data (regardless of the amount as they will be concatenated further on and the above outlined availability ratio criteria will be applied on that), (iii) occur on the same type of day (i.e., weekday or weekend) as the headache period, and (iv) fall within 14 days before or after of the headache period. This 14-day proximity threshold was chosen to increase the likelihood of similar time-of-day behavior patterns, as movement patterns are typically more consistent on days close to each other [31]. Furthermore, to minimize the impact of prodromal and postdromal symptoms, non-headache periods are required to be at least 24 h distant from the end of previous headaches and the beginning of a future headache. Daytime periods are defined as the interval between 8h30 and 22h30.

We are aware that prodromal or postdromal phases may overlap with our 24-h criterion for non-headache period eligibility, as some studies report prodromal durations of up to 72 h [31, 32]. Conversely, Kelman L. (2004) [33] found that only 13.2% of 893 migraine patients experienced prodromal durations lasting over 12 h. Blau J.N. (1980) [34] noted prodromal phases lasting up to 24 h in a study of 50 participants. Regarding postdrome duration, Kelman L. (2005) [35] observed that 88% of postdromes in 827 headache clinic patients lasted less than 24 h, and Blau J.N. (1991) [36] reported an average postdrome duration of 18 h across 40 patients with migraine. Given these findings, this duration threshold was chosen to ensure a substantial number of closely occurring non-headache periods, making it a pragmatic choice for our study.

Additionally, headache and non-headache periods overlapping with Belgian public holidays were filtered out to avoid the potential impact of atypical behavior during holidays.

When multiple eligible non-headache periods were identified for a single headache period, the metric signals from these intervals were concatenated to create a single non-headache metric distribution. Note that the 95% data availability ratio was applied to the concatenation of the eligible non-headache intervals rather than to each individual interval.

From this concatenated distribution, non-headache features were computed and paired with corresponding features derived from the accompanying headache attacks. Table 1 summarizes the features that are computed for each interval. Supplemental Fig. 2 illustrates the empirically determined AIABS threshold ratio features, presenting the AIABS distribution and the threshold ratios across the different activities as predicted by the HAR model.

Table 1 Overview of utilized features for each metric signal and their descriptions

Lastly, feature deltas (Δ) were computed for each feature by subtracting the corresponding non-headache interval feature from those of the headache interval. Consequently, a positive Δ indicates a higher feature value during the headache period, while a negative Δ indicates a lower value.

Data imputation was not applied; instead, features were computed directly from the metric signals, which may include small bouts of missing data (fewer than 5%). This approach was chosen to avoid further complicating the methodology, especially considering the assumptions required for identifying suitable periods for imputation and aggregation in the context of headache events.

Table 2 provides an overview of all the applied eligibility criteria for headache and non-headache period selection.

Table 2 Overview of headache and non-headache interval eligibility criteriaEthics approval and participants’ consent

The study was approved by the Ethics Committee of University Hospital Ghent (BC-10031). The study was preregistered at clinicaltrials.gov (NCT04983186). All participants gave informed consent for the collection, analysis, and publication of their data.

Analysis and statistics

Demographic and participant-specific data, including age, sex, duration of headache syndrome, and headache treatment regimens, are provided descriptively as proportions and means with standard deviations (SD). The compliance rate with the daily questionnaire, which serves as an indicator of daily app engagement, is expressed as the percentage of days an individual participant filled out either the morning or evening questionnaire. This rate is reported using the median, along with the first (Q1) and third quartile (Q3). Registered headache episodes are described with average duration (in hours and minutes) and attack intensity, both accompanied by their respective SDs. Lastly, the proportion of attacks treated with acute therapy and the proportion of these acute-treated attacks that were successfully managed are described.

Four main analyses were conducted, each targeting different intervals of the eligible headache events. Table 3 provides an overview of the criteria applied to each analysis.

Table 3 Summary of criteria applied to different analyses

The first analysis examines movement differences during headache attacks, focussing on the entire ictal phase. Specifically, AIABS threshold ratios and average HAR prediction values were calculated for daytime headache intervals and compared with corresponding non-headache daytime data. Importantly, this analysis only included full headache intervals that did not overlap with nighttime hours (22:00 to 10:30), as also assessing movement differences during nighttime periods might skew the results. Consequently, considering the typical duration of (untreated) migraine attacks (i.e. 4–72 h) and the frequent onset of headaches in the early morning (Supplemental Fig. 3), a limited number of headache instances are retained for this analysis (N = 32 out of 505 bias-free headache events). In addition, this full daytime headache duration analysis also examines the impact of acute treatment, reported movement sensitivity symptoms, and headache intensity on changes in PA. Acute treatment use and effectiveness was determined by the input of the participant within the headache registration view of the application, as shown in panel [b] of Fig. 1.These variables are visually represented using color hues in three supplementary subplots. Movement sensitivity is identified when either “motion sensitivity” or “pain aggravation during routine activity” was reported.

The second analysis categorizes days as headache or non-headache days based on participants’ morning questionnaire responses, which queries whether the previous day was headache-free or not. Eligible headache days were required to also contain the onset of at least one bias-free headache event. Days, marked as headache free by the morning questionnaire, were excluded for analysis if they overlap with a registered headache event or fall on the day before or after a headache event’s start or end. No further filtering of the headache days was performed, meaning headache days could include one or multiple headache events, with onset times ranging from early morning or late evening. Each day was segmented into eight two-hour intervals: morning (8–10 h), pre-noon (10–12 h), noon (12–14 h), afternoon (14–16 h), early evening (16–18 h), evening (18 h–20 h), late evening (20–22 h), and night (22 h–24 h). For each interval on a headache day, corresponding non-headache intervals are identified using the criteria listed in Table 2, followed by metric feature Δ computation. This independent evaluation leads to a varying number of attack pairs across intervals. Subgroup analyses (e.g., based on acute treatment) were not performed due to the complexity of attributing specific headache events to entire days, such as determining the minimum overlap between an event and a headache day or accounting for multiple events occurring on the same day.

The third and fourth analyses investigate movement differences during specific time intervals around the onset and end of headaches, respectively. These analyses focus on five one-hour intervals: from three hours before headache onset (prodromal phase) to the second hour of the headache (ictal phase), and from the last two hours of the headache (ictal phase) to three hours after its end (postdromal phase). These intervals were chosen in an attempt to capture the most pronounced changes in physical activity, hypothesizing that participants might adjust their behavior, such as increasing activity, during the premonitory phase in anticipation of a headache. Expanding these intervals to encompass the full prodromal and postdromal phases would reduce data availability due to strict inclusion criteria (> 95% data coverage and alignment with daytime activity). Similar to the previous analyses, delta features for these periods were computed by subtracting features of non-headache intervals from those of corresponding headache intervals. The number of events considered for each interval may vary, as both headache and non-headache interval pairs must meet the ≥ 95% data ratio requirement and fall within daytime hours (8h30–22h30). For intervals within the ictal phase (i.e., the one-hour intervals up to 2 h after headache onset or before its end), only those fully contained within the ictal phase of the corresponding headache were included. This excludes intervals of shorter headaches (< 2 h) where this condition cannot be met. By independently applying the > = 95% data ratio and daytime criteria to each 1-h interval pair (as detailed in Table 3), we retain approximately 2.5 times more pairs than in our first full headache duration analysis (N = 32), while maintaining the same eligibility criteria for headache and non-headache pairs. In alignment with the first analysis, the second, third, and fourth sub-analyses also examine the presence of motion sensitivity symptoms, acute treatment, and headache intensity.

Statistical testing was performed across all four analyses on the paired signal metric features of headache and their corresponding non-headache intervals. Statistical significance is marked with asterisks (*) in the visualizations. Normality testing, conducted using D’Agostino and Pearson’s normality test [37], did not reject the null hypothesis of the samples originating from a normal distribution for the different sample sets. Consequently, we employed the paired sample t-test to determine whether the Δ feature values are symmetrically distributed around zero, considering a two-tailed distribution as the alternative hypothesis. Subsequently, Bonferroni correction was applied for each subplot family to adjust for multiple testing. Reflecting the exploratory nature of this study, both unadjusted and adjusted significance values are presented in the visualizations. Additionally, due to this exploratory nature and the absence of prior data, we did not perform a formal sample size calculation before starting the study.

Data analysis software and visualization tools

All data processing and analysis were conducted using Python version 3.9. Exploratory data analysis of the raw wearable data was performed with Plotly-Resampler [38]. During this exploratory analysis phase, we verified that daytime headaches (adhering to the 22.00–8.30 nighttime filter) did not overlap with participants’ typical sleep times, as observed during manual sleep period annotation (see Supplemental Table 1 for sleep period statistics). Statistical testing was conducted using the SciPy library, and the seaborn toolkit was utilized for scientific visualization [39, 40]. To efficiently compute the AIABS, vectorized numPy functions were leveraged through the tsflex library [41, 42].