From Accelerometer Data to Activity Patterns in Chronic Pain: Methodical Reasoning Is Key

1. Introduction

Chronic pain (CP), defined as pain persisting or recurring for at least three months [1], significantly impacts various aspects of daily life, including physical, mental, social, economic [2,3] and spiritual [4] domains. Individuals with CP often experience negative emotions, fatigue, depression, deconditioning, sleep dysfunctions and reduced self-efficacy [5]. These challenges can lead to limitations in expressing the self, work, leisure and family life and to social isolation [2,3]. For healthcare providers, CP presents a complex challenge due to the interplay of biological, social, and psychological factors. Often, there is no definitive intervention to resolve CP, leading healthcare interventions to focus on self-management and mitigating the negative consequences, such as limitations in activities, participation, and quality of life [6,7].

It is hypothesized that CP interferes with activity patterns (AP) within daily living [8,9,10,11,12,13]. AP have been defined as the temporal structure of physical activity and sedentary behavior accumulated over a specified period during waking hours [14]. Therefore, AP capture the succession of activity and rest, rather than cumulative measures like total activity per day. It is theorized that AP in individuals with CP differ from those in healthy populations, are influenced by coping styles and are associated with various health-related outcomes [15,16,17,18,19]. Several theoretical models have been proposed to characterize AP in CP, including the avoidance-endurance model [13].

Traditionally, AP have been assessed using self-report questionnaires. More recently, objective measurement using accelerometry has gained attention [20,21,22]. However, consensus on optimal assessment methods remains lacking. Comparisons of questionnaires with accelerometer outcomes have yielded no or inconsistent associations [23,24,25]. This evokes uncertainty about the validity of questionnaire outcomes for measuring AP on the one hand, and the effects of methodical choices in accelerometry on the other hand [23,26,27,28,29]. Self-report instruments are inherently subjective, relying, amongst others, on retrospective recall, personal perceptions and emotional state, which limits their ability in capturing actual behavior. The validity of accelerometry is influenced by incidental user-related factors, such as variations in the wear angle of the sensor.

More importantly, validity of accelerometry is structurally affected by methodical choices, including variables in data processing, and outcome variable selection. For example, very short epoch lengths may introduce unnecessary noise, whereas overly long epoch lengths may obscure temporal detail. Moreover, accelerometer studies often rely on coarse metrics such as total time per day [30,31,32] which fail to capture the nuanced temporal structure of activity and rest.

To interpret findings from previous studies and relate them to methodical choices, detailed insight into these methods is needed. AP are constructs – abstract representations that require clear definitions and operationalization for empirical investigation. This process involves thoroughly defining the concept followed by its operationalization with measurable variable and outcome measures.

Although recent reviews have explored behavior-related activity parameters from accelerometry[33,34] and cumulative activity in CP populations [35], no review has specifically focused on AP parameters relevant to CP. Therefore, this scoping review aims to provide an overview of methods for extracting AP parameters relevant to CP care from accelerometer time series. The primary aim is to elucidate the process of methodical reasoning for extracting AP from accelerometer data in patients with CP. This will inform researchers on the usability and availability of methods to investigate AP in patients with CP, and on the underlying concepts and definitions. The secondary aim is to interpret and compare clinically relevant findings in light of methods employed.

The review is structured around the methodical reasoning process, encompassing: (1) selection of the AP-related concept, (2) its definition or specification (conceptualization), (3) the definition of variables and indicators that can be observed and measured (operationalization), (4) the measurement properties and the methods for extracting the indicators from raw accelerometer data, and (5) any associations of the AP-outcome measures with symptoms or disabilities.

This scoping review [36] includes studies employing triaxial accelerometers to assess AP in individuals with chronic primary musculoskeletal pain. This condition is defined as CP in the muscles, bones, joints, or tendons that is characterized by significant emotional distress or functional disability, that cannot be accounted for by another condition[1,37].

2. Materials and Methods

The research protocol has been published in the Open Science Framework (OSF; OSF | Accelerometry in chronic pain). This scoping review is conducted and reported according to the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews PRISMA-ScR guidelines [38]. A scoping review is the most appropriate method to fit our aim, as it allows us to provide an overview of the volume and the focus of available literature, examine how research has been conducted, identify key characteristics, and identify knowledge gaps [36,39].

2.1. Eligibility Criteria

According to the recommendations of JBI Manual of Evidence Synthesis (Chapter 11, Scoping Reviews) [39], population, concept and context of interest were defined. This review focusses on the population of adults with chronic primary musculoskeletal pain as a main condition. The concept of study is the investigation of AP with triaxial accelerometry. The context of study is physical activity in daily life. All analytical methods were included if AP were involved, with emphasis on the temporal structure of activity versus sedentary behavior.

Full-text publications were included with (1) populations over 18 years, (2) primary musculoskeletal CP, (3) AP as a primary outcome measure, (4) triaxial accelerometers and (5) at least five consecutive days of physical activity measurement. In line with the definition of primary CP, publications were excluded if the sample consisted of individuals with musculoskeletal CP resulting from identifiable underlying conditions. Primary CP was selected as the focus of this review because it emphasizes the interaction between pain and AP, rather than the interaction between physical impairments and AP, as seen in conditions such as M. Parkinson or severe osteoarthritis. With mixed samples, papers were included when at least 50% of the sample fulfilled the second inclusion criterium. When subgroups were investigated separately, only data pertaining to the subgroup that met the inclusion criteria were considered.

The minimum of five consecutive measurement days is based on the recommendations of Verbunt et al. (2012; [40]) for assessing physical activity in CP, as well as prior reliability studies in the general population [41], older adults[42], and individuals with rheumatoid arthritis [43]. Eligible publications included peer-reviewed articles, gray literature and dissertations, provided they were written in English, German, Dutch, or French.

2.2. Search Strategy

PubMed, Embase and Cinahl databases were searched from inception until November 2024. For gray literature the first ten pages of Google Scholar were scanned and dissertations were searched in ProQuest via PsycInfo. The search strategies were developed in cooperation with a specialized information specialist (TE). The search strings for PubMed, Embase, Cinahl, GoogleScholar and PsycInfo were composed by TE. The search strings are included in Appendix A. Search was restricted to title, abstract and MeSh terms and included CP, physical behavior, accelerometry and synonyms. Synonyms were based on keyword, MeSh terms (PubMed) and Emtree terms (Embase) of relevant publications. The PubMed, Embase and Cinahl searches were performed by TE. Dissertations were searched by AD in consultation with TE. Finally, references of the included articles were scanned for missed publications.

Six key references [44,45,46,47,48,49] were selected in advance and in the search results it was verified that the six pre-selected key publications were included.

2.3. Processing of Search Results and Selection

References found in PubMed were removed from the Embase and Cinahl results by subtracting PMID-numbers. The outcome was deduplicated in RefWorks Legacy via the close deduplication method and this was double-checked with SR-accelerator Deduplicator.

Selection on inclusion criteria was done in two steps, with (1) a screening phase and (2) a selection phase. Both phases were performed by AD and HW independently with all search results. Inclusions after both screening and selection were compared, discrepancies were discussed, and decisions were made. A third researcher (MV) could be consulted in case of persistent doubt.

For the screening phase title and abstract of all references were imported into Active Learning for Systematic Reviews (ASReview) Lab software [50]. With ASReview the screening phase was assisted with an AI-approach, namely Active Learning with different Machine Learning (ML)-algorithms. With this approach, the number of references that need to be manually labeled as relevant or irrelevant is reduced by approximately 90 percent, while maintaining or even improving reliability [51]. The active learning method (the ‘ASReview Pipeline’) is summarized in box 1 and is extensively elucidated by Van der Schoot et al. (2021) [50] and Boetje and Van der Schoot (2024) [51]. The 6 preselected key references were used to check the results of the ASReview-assisted screening phase.

For step 1, ‘adding prior knowledge for training’, AD and HW independently screened and labelled 100 references manually and compared, discussed and adjusted the selection. Steps 2 and 3 were performed by AD and HW independently as well. The resulting two shortlists of included references were compared, and discrepancies were discussed which resulted in a final shortlist. These papers were read full-text and accordance with all inclusion criteria led to final inclusion for this scoping review.

As a final step titles of all reference lists of the included papers were screened independently by HW and AD. The selection phase was similar as with the other databases.

Box 1. Active Learning for Systematic Reviews, the ASReview pipeline.

Step 1: Prior knowledge is added for training the ML-model by manually screening and labelling the first 100 references as ‘include’ or ‘exclude’. With this prior knowledge the ML-learning classifier Term Frequency-Inverted Document Frequency (TF-IDF) with Naive Bayes, is trained to predict study relevance. This results in a ranking of all references in the order of relevance.

Step 2: In the active learning part, the references are manually labelled as relevant or irrelevant one-by-one. Each decision is used to train the ML-model after which the ranking of relevance is adjusted, and a new reference is presented. Decision rules are available to decide whether sufficient references have been assessed to yield a reliable ranking: when (1) all key references are selected, (2) at least twice the expected number of relevant references has been screened, (3) at least 10% of the total dataset has been screened and (4) screening of at least 50 successive records does not reveal new relevant records.

Step 3: The resulting labelling and ranking of references are used to train the deep learning model Sentence BERT with Fully Connected Neural Network (FCNN). Applying this model results in another ranking of relevance. After this ranking, unlabeled records are presented one-by-one in the order of relevance and judged manually until 50 successive records are labeled as irrelevant.

2.4. Data Extraction

The study characteristics were extracted from the included articles by the researchers (AD and HW), along with the information necessary to understand the process of methodical reasoning for the quantification of AP. This process involved the successive steps of conceptualizing and operationalizing the concept, followed by selecting measurement properties and data processing. Methodical reasoning concludes with the interpretation of results in relation to the study’s aims.

To find evidence for the existence of concepts or associations with a concept, the concept must be clearly stated and quantifiably defined. The operational definition should include information on the variables (the properties or characteristics) of the concept and its indicators (the methods of quantifying the variables).

This process of conceptualization and operationalization was extracted from the included papers and consisted of four successive steps: (1) formulating the theoretical concept of the study related to AP within CP, (2) defining the concept as precisely as possible (conceptualization), (3) operationalizing the concept with concrete and specific variables, which are the properties and characteristics of the concept, and (4) identifying the concomitant indicators, which quantify variables.

This process is clarified by an example in Figure 1. Additionally, measurement properties and data processing were summarized, along with any association with symptoms or disabilities, differences between groups and differences pre- and post-treatment.

3. Results

As illustrated in Figure 2, after deduplication, the search in PubMed, Cinahl and Embase yielded 11366 records. With ranking and labelling according to the ASReview pipeline, in total 2300 of 11366 references were screened and labelled by HW and AD (in step 1 100 by HW and 100 by AD, in step 2 2 times 1000, and in step 3 2 times 50). Forty-two references were labelled as relevant, which included all 6 key references. These 42 papers were read full text and inventoried on the inclusion criteria after which 14 papers were included. Ten publications were excluded because they were posters, meeting abstracts or commentaries instead of full text peer reviewed papers. 17 papers were excluded because they did not meet at least one of the inclusion criteria, of which 16 did not investigate AP according to the definition of ‘the temporal structure of physical activity and sedentary behavior accumulated over a specified time period during waking hours,’ and one had a measurement period of one day for investigating activity variability during the day. Of these 17 excluded papers one used a biaxial accelerometer instead of triaxial. In one paper [52] the number of accelerometer-axes was not specified, and, on inquiry, the author could not clarify the number of axes. Paraschiv et al. (2008) used a biaxial and a uniaxial accelerometer on different body locations and Paraschiv et al. (2012) used three biaxial accelerometers which was assumed to deliver at least the same dimensionality and order of detail as one triaxial accelerometer. Therefore, these two studies by Paraschiv et al. were included. One publication could not be found.

With Google Scholar, 14 papers were ranked as relevant of which 13 could be excluded by reading title and abstract. One publication was selected for full text reading and was excluded because it did not investigate AP. Screening of the titles of ProQuest results yielded 6 records for reading the abstract after which one dissertation was screened full text [53]. This dissertation was excluded based on the inclusion criteria. After scanning the reference lists of the 14 included papers 26 titles and abstracts were screened and all were excluded because they did not meet the inclusion criteria. One publication [54] that was not found in the databases was added from the personal archives, which yielded 15 papers for this scoping review.

3.1. Study Characteristics

The included studies are described in Table 1. Most studies had a cross-sectional set-up. Three studies had a longitudinal set up to investigate treatment outcomes [49], associations between changed behavior-type and disability and quality of life [46] or associations between changed activity level and changed pain intensity [55]. CP sample size ranged from 15 to 292. Some studies included a sample without pain [22,56,57] or with acute pain [55] for comparison. Most participants were recruited from multidisciplinary pain centers and hospital departments of rehabilitation. In two studies participants were recruited from the general population through their physician [58,59] and one study recruited participants with advertisement on a university campus [60]. In most studies women were predominant and age ranged from 20 to over 74 years (SD 8). In most studies the type of CP was not specified. We assumed that CP in tertiary multidisciplinary pain centers, as in Andrews et al. (2023, 2015 and 2014), Liszka-Hackzell et al. (2004), Paraschiv et al. (2012 and 2008) predominantly comprised primary musculoskeletal CP. In Fanning et al. (2024 and 2023) inclusion criteria were CP in at least two sites of neck, shoulder, back, hip or knee, without specification on the cause of pain. We assumed that this pain mainly comprised chronic primary musculoskeletal pain.

3.2. Concepts of Investigation, Definitions, Variables and Indicators

In most papers a quantifiable definition of the concept of study was available or could be derived from reasoning. As a next step, operational definitions should describe how the concept of study was measured and how the measurements were interpreted. Concepts, their definitions and the definition of variables and their indicators are summarized in Table 2 and elucidated in the following sections.

Concepts and Definitions (Conceptualization)

The concepts used to investigate AP were diverse (Table 2), and the meaning of these concepts could be extracted from all papers. Definitions and specifications of these concepts are presented in Table 2. A distinction could be made between concepts related to behavioral patterns and those derived from physics.

Six papers [44,45,46,48,49,61] utilized existing models of behavioral patterns related to activity and rest as their research concept, analyzing accelerometer time series, sometimes combined with pain intensity time series.

Huijnen et al. (2011_1) defined persistence as continuing activities despite pain until task completion, resulting in forced rest due to increased pain. Additionally, a longer daily uptime was attributed to persistence because persisters tend to postpone rest. This definition was also adopted in Huijnen et al. (2020). Besides persistence, Huijnen et al. (2011_1, 2011_2 and 2020) investigated avoidance. In Huijnen et al. (2011_1) an avoider was defined as “an avoider will try to escape from activities for which they expect an increase of pain or injury.” Although definitions of avoidance and persistence were lacking in Huijnen et al. (2011_2), it can be deduced from the analyses that they were consistent with Huijnen et al. (2011_1), as verbally confirmed by the first author.

In line with the definition of persistence in these Huijnen-papers, Andrews et al. (2014) defined overactivity as engaging in high levels of activity that result in severe pain aggravation and a subsequent period of inactivity where an individual is unable to function. Daily tasks are resumed when pain subsides or when frustration stimulates new activity. Consequently, overactive persons exhibit a “sawtooth” AP with large fluctuations in pain and activity. Andrews et al. (2015 and 2023) expanded this definition to include pain aggravation after prolonged sedentary periods. Andrews et al. (2023) also introduced the concept of pacing, defined as a lower frequency of overactivity.

Time series of pain were included in four of the six papers on persistence and overactivity due to the hypothesized relationship between pain and activity in these behavioral patterns. Andrews et al. (2014) and Huijnen et al. (2011_2) did not include pain levels, focussing on fluctuation values of physical activity intensities (Andrews et al. 2014 and Huijnen et al. 2011_2), daily uptime and mean activity (Huijnen et al. 2011_2).

The remaining nine papers used physics derived measures to capture temporal patterns or complexity. Fanning et al. (2024) investigated activity intensity patterns using fitted Fourier functions on time series of steps per minute. Fanning et al. (2023) examined the pattern of activity and rest accumulation, described by the total time of activity and rest per day and the breaks within bouts of rest and activity. Liszka-Hackzell et al. (2004) aimed to examine the causal relation between activity level and pain by calculating the cross-correlation between time series of activity level and pain with different time lags. They hypothesized that activity correlates with pain, possibly with some time lag, in acute pain but not in CP.

Neikrug et al. (2017) and Sarwar et al. (2022) focused on the concept of rhythmicity of rest and activity during the day. This focus was chosen because consistent functioning might be hampered by fluctuating symptoms of pain, fatigue, mood and physical impairment (Neikrug et al., 2010). Moreover, reduced activity levels, sleep disturbances and circadian dysregulation have been observed in populations with CP (Sarwar et al. 2022). Both papers a fitted a cosine curve on accelerometry time series and extracted parameters from this curve to describe the diurnal rhythmicity of rest and activity.

Paraschiv et al. (2008) choose the dynamics of human activity as the study concept. This was more clearly defined in Paraschiv et al. (2012) as the temporal and dynamical structure of human physical activity. Paraschiv et al. (2016) adopted the same concept, and all three papers build on the method of constructing time series of activity type and intensity of walking derived from accelerometry as presented in Paraschiv et al. (2004). Paraschiv et al. (2008) focused on dynamics of these physical activity parameters, Paraschiv et al. (2012) added intensity classes to the PA parameters and focused on the complexity of these time series and Paraschiv et al. (2016) focused on deriving and applying a composite score from the outcome scores of the different deduced time series. All three Paraschiv-papers used advanced and mainly non-linear statistical methods to capture dynamics of physical activity.

Zheng et al. (2023) addressed the concept of physical activity intensity patterns in chronic low back pain. Physical activity intensity patterns were defined as the temporal organization of intensity levels and the transitions between intensity levels. This concept was chosen because it was recognized that, for instance, with a certain amount of accumulated sedentary time, it makes a difference whether someone alternates sedentary time with small bouts of activity or not. They used the data driven method of unsupervised learning (Hidden semi-Markow Model) to detect physical activity intensity levels from the vector magnitude time series.

Variables and Indicators (Operationalization)

In all the papers, observable and measurable variables for operationalizing the concept of study were defined, and descriptions of the procedures for quantifying these variables were included.

In the six papers that used behavioral concepts [44,45,46,48,49,61], variables were formulated to inventory the occurrence or nonoccurrence of the behavior. Due to the expected sawtooth pattern with overactivity, Andrews et al. (2014) focused on the fluctuation of activity levels. They calculated a fluctuation value from time series of activity counts per minute, which is the root mean square of the difference of two successive cumulative 15-minute vector magnitudes. In Andrews et al. (2015 and 2023), the outcome measures emphasized the concomitant pain increase with overactivity (Figure 3A), rather than focusing on the fluctuation values related to sawtooth pattern. They used conditional statements to count the occurrence of significant pain increases after activity or prolonged sedentary tasks.

Huijnen et al (2011_1) compared variables related to avoidance and persistence between persons classified as avoider or persister with the POAM-P self-report questionnaire. Persisters were expected to have a longer daily uptime, a higher average activity level and more fluctuations in activity level compared to avoiders. Daily uptime was quantified as the daily wear-time of the accelerometer. Activity level was quantified as mean activity counts per day, and 80% power of highest activity score. Fluctuations in activity levels were quantified as the root mean square of the difference of two subsequent 15-minute activity counts. Additionally, it was hypothesized that in persisters, the time series of activity level and pain are associated.

To quantify the level of persistence, Huijnen et al. (2011_2) calculated the daily uptime and a linear composite score of daily uptime, mean total activity score and a fluctuation score. These parameters were calculated according to the methods of Huijnen et al. (2011_1). To investigate differences in activity behavior between individuals classified as avoider, persister, mixed performer or healthy performer, Huijnen et al. (2020) used cumulative variables such as overall daily activity level and total sedentary time, as well as variables representing the distribution of activity and rest over time. The distribution variables included the number of bouts of physical activity and sedentary behavior, the median duration of bouts, the variance of bout length, the number of bouts divided by total duration for activity and rest separately as a measure of fragmentation and cumulative time of bouts above median bout length divided by total time of rest or activity (W-index), expressing the relatively higher contribution of longer bouts with a higher value.

To investigate the timing of activity and rest as well as the amplitude of activity intensity, Fanning et al. (2024) fitted a 9-basis Fourier-function on each participant’s time series of steps per minute. Subsequently, they performed a functional Principal Component Analysis to identify a set of Fourier functions capturing the most variability. In Fanning et al. (2023) the concept of accumulation of rest and physical activity was operationalized using a combination of cumulative measures and breaks within bouts of rest and activity. These measures included minutes per day in a seated or lying position, average number of steps per day, minutes per day at different classes of stepping frequency, the number of shifts from sitting to standing and the frequency of classes of bout lengths for different activity intensities.

Neikrug et al. (2017) extracted day to day rhythmic features of activity and rest using cosinor-based techniques and determined correlations with fibromyalgia symptoms. Parameters extracted from the cosine curves were mesor (mean activity level), amplitude (distance between mesor and peak level), Phi (time of day of peak activity level) and the standard errors. Besides these three parameters, Sarwar et al. (2022) extracted six other rhythmic features with a cosinor model, quantifying the most and least active minutes, complexity of the time series, nocturnal activity and intradaily variability. Some of these parameters and a fitted cosinor model are shown in Figure 3B and 3C.

Liszka-Hackzell et al. (2004) calculated the cross-correlation of interpolated time series of pain levels with time series of activity counts per minute. Cross-correlation was determined with a time lag of the pain time series of -60, -30, 0, 30 and 60 minutes, with which they investigated whether a pain level increase was ahead of activity level increase, synchronous, or delayed.

Paraschiv et al. derived different time series types of postures and activity intensity from accelerometer data, refining their methods in successive papers (2008, 2012, 2016), based on the method presented in Paraschiv et al. (2004). In the 2004 paper, four different postures (lying, sitting, standing and walking) and the intensity of walking were derived from accelerometer data using discrete wavelet transformation, Savitzky-Golay filters, vector functions and gait analysis parameters. In Paraschiv et al. (2008) different time series were constructed using these methods, including the sequence of posture allocation (Figure 3D), the sequence of the duration of walking episodes, the timing of transitions from rest to activity and vice versa, and the duration of activity relative to the duration of rest before and after activity, represented as a symbolic sequence. Nonlinear analyses were applied to these time series to investigate the physical activity pattern, including three types of fractal analysis and symbolic dynamic statistics.

Paraschiv et al. (2012) added intensity to the four activity types (lying, sitting, standing and walking) based on different acceleration thresholds for each activity type. This results in 18 possible physical activity states: two classes of sitting and lying, four classes of standing and twelve classes of walking. This classification resulted in a time series of 18 possible symbols (Figure 3E). From these time series complexity metrics, quantitative metrics, a composite deterministic score and a composite statistical score were derived.

Exploring the ability of a composite score to characterize physical behavior, Paraschiv et al. (2016) performed a factor analysis on outcome variables. These outcome variables were based on the same procedure as Paraschiv et al. (2012) to classify 18 physical activity states. Variables were percentage of time being active, percentage of time walking, 0,975^th quartile of length of activity periods, duration of sedentary periods following activity (expressed as Kolmogorov-Smirnov distance) and three types of entropy. Moreover, the association between pain intensity and activity behavior was assessed with multiple regression and discriminant analysis.

Zheng et al. (2023) applied unsupervised learning (Hidden Semi-Markow modelling, HSMM) where ML-algorithms were used to discover a set of hidden states in unlabeled accelerometer data. These hidden states were derived from time series of vector magnitude acceleration averaged in 5-second bouts. The modelling identified five hidden states corresponding to five different activity levels. Bout lengths, total time per day in these five states and transition frequencies between all five activity levels were compared between groups with and without chronic low back pain. The results were also compared with the same outcome measures derived from time series constructed with the conventional cut-points approach.

In summary, AP related research concepts were operationalized in many ways. Figure 3 provides an impression of the variation of methods used. Methods included a measure of fluctuation by subtracting two successive bouts of activity intensity, timing and amplitude of activity intensity quantified by fitted Fourier-functions, parameters derived from a fitted cosine curve, non-linear analyses, complexity metrics and variables that quantify fluctuations and distribution like transition frequency and W-index.

3.3. Measurement Properties and Data Processing

Measurement properties provide information on how data were collected to quantify the required variables. Data processing refers to the conversion of raw accelerometer data into the outcome measure needed for further analysis (see Table 3).

Measurement properties and data processing were heterogeneous (Table 3). Thirteen papers used triaxial accelerometers and two papers used a combination of multiple biaxial and/or uniaxial accelerometers [22,56]. In two studies the accelerometer was worn on the non-dominant arm [55] or wrist [54] and in one study the side of the wrist-worn accelerometer was not specified [60]. In one study the accelerometer was attached to the waist [49], in two studies to the upper midline of the thigh [58,59], and in another study to the front right hip [47]. Five studies used multiple accelerometers on multiple wear locations (chest and both thighs [44], sternum and mediolateral axis of thigh [57], sternum with mediolateral axis of thigh and shank [56] and chest with thigh and shank [22]. In four studies the wear-location was not reported [45,46,48,61].

Measurement frequency ranged from 30 to 128 Hz, with frequencies from 30 to 40 Hz being predominant. Measurement frequency was not reported in seven studies [45,46,49,55,58,59,60]. Sampling duration ranged from 5 to 21 days, with a duration of five days being predominant. Of the studies that sampled five days, two studies included one weekend day [48,61], four studies only included weekdays [22,56,57,60] and four studies didn’t specify which weekdays were covered [44,49,58,59]. Valid data was defined in most studies. Epoch length ranged from one second to one hour, with one minute being predominant.

Four studies sampled pain intensity during the day and used these time series to investigate AP since pain was part of the definition and operationalization of the concept of research [45], [46,48,49,61]. Time series of pain were measured with 11-point visual analogue scale [49,61] or a 7-point likert-scale [45]. The pain measurement instrument was not specified in Andrews et al. 2023. Many other variables were sampled to investigate associations, associations through time and differences between subgroups. These variables include age, sex, self-reported measures of pain intensity, pain duration, pain interference, health-related quality of life, medication intake, approach to activity, depression, anxiety, self-discrepancy type, fatigue, impairment, functioning, mood, sleep and central sensitization symptoms.

Triaxial accelerometry results in time series of acceleration around x-, y- and z- axes, with the number of values per second dependent on sampling frequency. These time series were converted to another parameter in most of the included papers. In general, the description of conversion methods was limited. Some authors only refer to software packages and some to manuals or websites of the accelerometer manufacturer that were no longer available on the web.

For data processing six studies transformed the acceleration time series to activity counts per minute, also named vector counts per minute [45,46,48,49,55,61]. This method operates on the assumption that counts per minute is associated with the energy expenditure of activities and therefore with activity intensity. The vector magnitude of acceleration was calculated from the triaxial acceleration values and to yield counts per minute the number of times per minute of exceedance of a predefined threshold value was counted. The threshold for a count was not specified in the papers. Two studies transformed the acceleration time series to steps per minute and postures of lying, sitting or standing [58,59].

Paraschiv et al. [22,56,57] processed accelerometer data with discrete wavelet transformation, Savitzky-Golay filters and a numerical gradient. Subsequently they constructed time series of activity type with a previously developed algorithm [62]. They used different methods to detect different activities, postures and intensity of walking. Zheng et al. (2023) used raw accelerometer data from which the gravity effects were removed and then computed vector magnitude. This acceleration was averaged over five seconds. Unsupervised learning was applied to these time series which resulted in time series of five activity intensity classes. For comparison of result with a traditional method, they applied the cut-off points approach as well. With this approach the tri-axial acceleration signal was converted to vector magnitude and thresholds were defined for different activity levels.

3.4. Associations, Differences Between Groups and Differences Pre and Post Treatment

Understanding the associations between selected indicators and related behavior types or relevant clinical outcome parameters, as well as differences between groups and changes through interventions, may provide valuable insights into the usefulness and clinical value of these indicators.

All papers had a primary or secondary aim to examine associations, differences between groups or changes from pre- to post-treatment. Due to the heterogeneity of concepts, methods and outcome measures used, no firm conclusions can be drawn from the included papers. Results of some papers can be compared because of overlapping concepts, definitions and outcome measures.

Papers examined associations of objective AP parameters with self-reported behavior types such as avoidance or persistence [45,48] and behavior types classified by their treating consultant [44] (Figure 4). They also examined associations of objective AP parameters with self-discrepancy type [44] (Figure 5) and clinical outcome scores [45,54,56,57,58,59,60] (Figure 6). Since the quality of evidence is not rated, only the direction of these associations is summarized in Figure 4 to 6 as negative (-), no significant association (No), or positive (+).

Most associations of self-reported behavior-type with objective AP parameters were not significant, except for a positive association of self-reported persistence with activity fluctuation[48] and daily uptime [45], and of daily uptime with mental quality of life within objective avoiders [45]. This suggests that individuals with higher self-reported persistence have more activity fluctuations and longer daily uptime. This is understandable, as persisters probably are likely more active throughout the day than avoiders, increasing their chances of intensity fluctuations and longer daily uptime. Additionally, among individuals reporting activity avoidance, quality of life might improve with a longer daily uptime. However, given the large number of associations investigated in some studies, the risk of a type I error is considerable.

One paper examined associations of objective AP parameters with self-discrepancy types [46] (Figure 5). This study found more objective persistence with higher scores on self-discrepancy type ideal-other, which suggests that persons who are close to the ideal person from the perspective of a significant other, showed more objective persistence behavior. Furthermore, patients with increasing distance in time from their ideal-self from their own perspective had increasing daily uptime and patients getting closer to their ideal-self had decreasing daily uptime. In addition, patients with increasing distance in time between the actual-self and the ideal-self from the perspective of a significant other had decreasing daily uptime.

Seven papers examined associations of objective AP parameters with clinical outcome measures. These outcome measures were pain intensity [45,54,56,57,58,59,60], pain interference [58,59,60], health-related quality of life [58], fatigue [54], mood [54] and disability [54,60] (Figure 6). Associations were investigated with AP parameters that capture (1) the amount of activity, (2) timing of activity, (3) activity intensity, (4) distribution over different activity intensities, (5) complexity and (6) variability. The diversity of AP parameters and outcome measures complicated comparison and interpretation and when comparable parameters were used, associations were inconsistent between studies.

Amount of activity was represented with (maximum) step count, stepping time, percentage of time being active, activity counts, change (∆) in steps per day and change in stepping time. Maximum step counts and percentage of time being active were negatively correlated with pain intensity [56,60], indicating decreasing pain intensity with increasing active time. None of the other parameters for amount of activity correlated significantly with pain intensity.

Timing of activity was represented by phi, ‘Later start of activity’ and ‘∆ later start of activity. Phi was positively correlated with pain intensity, indicating higher pain levels with later activity peak, while start of activity did not have a significant association.

Activity intensity was expressed as mesor, change in sedentary intensity time, change in light intensity time, change in moderate intensity time. Mesor and change in moderate intensity time were negatively associated with pain intensity, indicating decreasing pain intensity when average activity intensity increases and when daily moderate intensity time increases after intervention. Change in light intensity time was positively correlated with pain intensity, indicating more pain with more light intensity time. Amplitude was positively correlated, indicating more pain with higher maximum intensity.

Distribution over activity intensities was indicated with bouts per intensity class, change of number of bouts of classified lengths per intensity class and percentage of time per intensity class. Bouts per intensity class did not correlate with pain intensity and changed bout lengths for the different intensity classes did not associate consistently with changed pain intensity. The change of bouts of classified length per intensity class is hard to interpret and its clinical meaning is unsure.

Complexity measures correlated negatively with pain intensity and intradaily variability correlated positively. These two outcomes seem contradictory since level of complexity is partly dependent on variability.

Most of these measures for amount, intensity, timing, distribution, complexity and variability were used to investigate associations with pain interference as well. Increased intradaily variability was found to increase pain interference. Moreover, a later start of activity was not associated with pain interference at baseline, but retarding the start of activity after a 12-week intervention program was found to increase pain interference. Decreasing the number of 5-10 minutes bouts of light intensity and 10-20 minutes bouts of moderate intensity seemed to increase pain. None of the other parameters correlated significantly with pain interference.

Scores on fatigue, mood and disability increased with later timing of peak activity (phi) [54], while another study found that fatigue was not significantly associated with timing of activity[58]. In one study disability increased with increasing intradaily variability [60]. Scores on fatigue, mood and disability decreased with higher amplitude and higher average activity level [54], but this was refuted by another study in which fatigue increased with increasing activity levels expressed as steps/day[58]. Retarding the start of activity was associated with decreased quality of life and increasing the steps per day improved quality of life.

Other studies investigated differences between groups or changes pre- and post-treatment. Andrews et al. (2023) found that a 15-week occupational therapy intervention resulted in significant reductions in objective overactivity periods. Six other studies compared differences between subgroups, grouped by different characteristics. One study [44] compared objective activity parameters in persisters and avoiders as classified by their treating consultant and found no differences. Another study [45] compared objective activity parameters of persisters and avoiders as classified by a self-report questionnaire and found longer daily uptime in persisters, but no differences in activity level.

One study [55] examined the cross-correlation between time series of activity level and pain between subjects with acute pain and CP. They found a positive association in acute pain and no association in CP.

Paraschiv et al (2008) compared groups with CP and no pain and found differences in the temporal organization of daily life with a scaling component from detrended fluctuation analysis, a different distribution of walking episodes and a different timing of rest-activity transitions. Moreover, they found that short activity with subsequent long rest appeared more frequently in CP and long activity with subsequent short rest appeared more frequently in no pain. Paraschiv et al. (2016) compared groups with different pain levels and found that the number and duration of activity periods differed between groups, as well as the complexity of the temporal pattern of activity.

Lastly, Zheng et al. (2023) compared groups with and without central sensitization (CS) using both a traditional cut-off points approach, and a time series generated with ML. They found no differences between individuals with and without CS when activity level time series were generated with the cut-off points approach. However, many differences were observed with the ML-generated time series. Participants without CS had (1) shorter bout durations of activities with lower intensity and longer bout durations of activities with higher intensity, (2) less accumulated time per day at rest and light physical activity, (3) more time per day at moderate at vigorous activity, (4) more frequent and shorter sedentary states, (5) more frequent transitions from rest, light, moderate-vigorous to sedentary, suggesting more rest after activity, (6) fewer transitions from light to moderate activity, suggesting less persistence and (7) more continuous time in both active and inactive states.