Chronic pain, defined as pain that recurs or persists for at least three months, affects between 12.7 and 33.7 percent of adults in Europe. People living with it often adapt their patterns of activity and rest to cope with both pain and daily demands, and these adaptations carry clinical weight. Fear of activity-related pain can lead to avoidance, while excessive overexertion produces a characteristic sawtooth pattern: individuals push through rising pain, then collapse into prolonged rest. Pacing, by contrast, is considered adaptive, involving the deliberate regulation of activity and rest. Because these patterns may reflect psychological state, physical function and pain interference, clinicians want objective ways to measure them. Yet in rehabilitation practice, activity patterns are still assessed mainly through questionnaires and interviews, which rely on recall and perception rather than actual behavior, and self-reported activity correlates only weakly with what activity monitors record.

Wrist-worn tri-axial accelerometers are user-friendly and yield valid and reliable data in the general population, but converting their raw signals into activity intensity classes has traditionally relied on cut-off point thresholds or regression-based approaches. Recent evidence suggests that deep learning models may outperform these conventional methods, offering high accuracy, adaptability across populations and the potential for continuous improvement as new data arrive. Two knowledge gaps, however, have limited their substantiated use in chronic pain care. First, the clinimetric properties of deep learning models have been investigated primarily in laboratory settings, leaving their construct validity under real-world conditions underexplored. Second, pain is known to alter motor behavior, from subtle stiffness and guarding to outright avoidance of painful movements, and these changes may distort accelerometer signals enough to reduce the accuracy of models trained on data from people without chronic pain. Until now, no study had investigated whether such models require adaptation to chronic pain-specific motor behavior.

To build the model, the researchers, led by Annet Doomen of HU University of Applied Sciences Utrecht and Maastricht University, assembled three datasets. Thirty healthy participants and thirty participants with chronic pain, recruited from interdisciplinary pain rehabilitation clinics in Eindhoven and Hoensbroek, each performed twenty-one predefined activities in a semi-controlled setting while wearing a tri-axial accelerometer sampling at 12.5 hertz on the lateral side of the dominant wrist. The activities, ranging from sitting reading a book to running and cycling, were labeled with metabolic equivalent of task values from the Adult Compendium of Physical Activities and classified into three intensity categories: sedentary, light and moderate-to-vigorous physical activity. Participants marked the start and end of each activity with three rapid forearm shakes, producing distinct acceleration peaks that allowed precise segmentation. The third dataset, Capture-24, is a large open-source collection of annotated real-world accelerometer data from 151 participants in Oxfordshire, United Kingdom, who wore an Axivity AX3 accelerometer and a neck-worn camera for approximately twenty-four hours in free-living conditions, with time-stamped camera images used to annotate the data.

After processing, the training data from 208 participants contained 211,864 ten-second windows, equivalent to 588 hours of labeled movement. The researchers tested four initial deep learning architectures and selected a convolutional neural network for its balance between performance and model size. They then trained this network with seven different strategies to evaluate performance across populations: training and testing on a single source, training on healthy data and validating on chronic pain data, and training on the combined datasets while validating separately on each. To address class imbalance, category-specific weights were applied during training, and the model was evaluated with repeated ten-fold cross-validation, with data split at the subject level so that each participant appeared exactly once in the validation set across the ten iterations. The Capture-24 data were downsampled from 100 to 12.5 hertz to match the laboratory data, a step that reduced accuracy when trained exclusively on Capture data but was more than compensated for by incorporating the laboratory data.

The results showed that classification accuracy reached 76.5 percent in individuals with chronic pain and 80.5 percent in individuals without chronic pain, with Macro-F1 scores of 73.4 and 78.3 percent respectively. The highest performance, 81.3 percent accuracy, was achieved when the model was trained, tested and validated exclusively on healthy participants, while strategies validated on the Capture data yielded the lowest scores. Notably, strategies incorporating the large Capture-24 dataset showed lower standard deviations and narrower performance ranges, suggesting that its inclusion enhanced the stability of the model’s output. Crucially, performance was comparable whether the model was trained on the pooled dataset, on healthy data alone, or exclusively on chronic pain data, indicating that chronic pain-specific training conferred no clear advantage within the populations studied. Sedentary intensity was correctly classified in 89 percent of chronic pain cases, light intensity in 64 percent and moderate-to-vigorous activity in 66 percent, with both light and moderate-to-vigorous activity more frequently underestimated than overestimated.

To explore construct validity in real-world settings, the researchers observed ten healthy participants and five individuals with chronic pain for approximately four hours each in their daily-life environments while they wore the accelerometer and followed their usual routines without modification. Observers recorded activity descriptions every minute, and two researchers independently assigned metabolic equivalent values to each recorded activity, achieving 81 percent inter-rater agreement before resolving all discrepancies. The deep learning model’s minute-by-minute predictions were then compared with these observations. Agreement was 78.6 percent in healthy participants and 73.6 percent in those with chronic pain, with weighted Cohen’s kappa values of 0.66 and 0.60 respectively, interpreted as substantial and moderate agreement. Individual agreement varied widely, ranging from 57.1 to 94.3 percent, and the lower agreement in chronic pain was primarily due to the model’s overclassification of sedentary activity in 26 percent of cases.

A closer inspection of the disagreements revealed multiple sources of error. Observer-based intensity classifications were sometimes inherently uncertain, particularly during minutes containing multiple activities with different metabolic demands, where assigning a single intensity category required averaging. The comparator instrument also relied on a predefined list of activities and corresponding metabolic equivalent values, irrespective of common individual variations in task execution; for example, the compendium assigns sitting and speaking a value of 1.3 without distinguishing between speaking with or without gesticulation, so when substantial arm movements occur during speech, the model may legitimately classify a higher intensity than the annotation suggests. More tellingly, several participants demonstrated systematic over- or underestimation across specific activities, indicating that individual differences in movement characteristics, rather than random classification errors, reduced construct validity. Post hoc analyses showed that two participants with chronic pain, diagnosed with fibromyalgia, patellofemoral pain syndrome, neuroma and non-specific low back pain, showed markedly lower performance metrics and substantially higher acceleration magnitudes during sedentary and light-intensity activities than the rest of the sample, suggesting elevated motor activity that the annotation procedure did not capture. Activities near the boundaries of the intensity categories posed additional challenges, as small differences around the metabolic cut-off values of 1.5 and 3.0 can result in different labels despite very similar movement patterns.

The authors acknowledge several limitations. Performance varied substantially between individuals, particularly within the chronic pain sample, so the overall results may have differed with a different group of participants. The awareness of being observed may have altered participants’ usual activities and movement patterns, and the five individuals with chronic pain included in the construct validity analysis reported lower pain intensity and less disability than those in the training data, meaning that individuals with more severe symptoms, who may display greater movement pattern deviations, were underrepresented. The study also focused specifically on activity intensities rather than the detection of pain-related movement behavior such as guarding, avoiding or rubbing. Before clinical or research application, the researchers recommend further investigation of construct validity, for example in individuals with chronic pain who exhibit pronounced differences in motor behavior, and of factors such as hypervigilance, obesity, pain location or the use of walking aids that may influence individual model performance. They also suggest exploring a continuous measure of activity intensity rather than discrete categories, which may help address borderline classification errors and enhance sensitivity to clinically meaningful change.

Nevertheless, the study represents a significant step toward standardizing activity intensity measurement in chronic pain research. All four initial deep learning models outperformed the conventional cut-points approach, consistent with previous findings in the general population, and the accompanying scripts and datasets enable accurate processing of raw accelerometer time series into activity intensities in the chronic pain populations studied. By making everything openly available on Zenodo, the team has lowered the barrier for other researchers to apply, scrutinize and improve the model. The work also delivers a valuable insight for the growing field of digital health: the inter-individual variability in agreement between real-world observations and model predictions highlights that the value of further investigating construct validity within a larger and more diverse sample of people with chronic pain is not just warranted but essential, if wearable-based activity monitoring is ever to deliver on its promise of truly personalized care.

Subject of Research: An open-source deep learning model for classifying real-world activity intensities in persons with chronic pain using a wrist-worn accelerometer

Article Title: An open-source deep learning model for classifying real-world activity intensities in persons with chronic pain using a wrist-worn accelerometer

Article References: Doomen, A., Felius, R., Wittink, H., Huijnen, I., Punt, M., & Smeets, R. (2026). An open-source deep learning model for classifying real-world activity intensities in persons with chronic pain using a wrist-worn accelerometer. Frontiers in Digital Health, 8, Article 1783507. https://doi.org/10.3389/fdgth.2026.1783507

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1783507

Keywords: open-source, deep, learning, model, classifying, real-world, activity, intensities, persons, chronic, pain, wrist-worn