Machine Learning Classifies Musculoskeletal Risk in Garment Workers

The research team, led by Md. Zaid Hossain and colleagues, utilized a leakage-controlled evaluation methodology to determine whether dynamic postural exposure metrics could improve risk stratification beyond standard demographic and work-organization variables. The study involved 60 sewing-machine operators, evenly divided by sex, who underwent bilateral video recording and questionnaire-based symptom assessment.

In Bangladesh, where the ready-made garment sector is the largest export industry, sewing-machine operators face a biomechanically challenging exposure profile characterized by prolonged constrained sitting, sustained forward neck flexion, and high-frequency repetitive upper-limb movements. Previous epidemiological studies have reported high prevalence rates of WMSD in this population, with neck, shoulder, and back complaints being the most common. However, traditional risk-assessment tools often rely on static snapshots of posture, which may fail to capture the temporal variability and cumulative load inherent in sewing tasks.

The study aimed to address these limitations by integrating dynamic postural assessment with machine learning classifiers. The researchers hypothesized that time-weighted Rapid Upper Limb Assessment (RULA) scores, derived from continuous video analysis, could provide a more accurate measure of postural exposure than single-point observations. To test this, the team processed 10-minute bilateral task videos for each participant, extracting joint-angle series through a semi-automated workflow. These data were then aggregated into time-weighted indices to represent cumulative postural exposure over the work cycle. The study design was explicitly cross-sectional, meaning it assessed current risk status rather than predicting future disease incidence or establishing causal relationships between exposure and outcome.

Data collection took place between June 2024 and February 2025, involving 60 full-time operators with at least one year of job tenure. Participants completed a modified Nordic Musculoskeletal Questionnaire (mNMQ), which recorded symptom severity, exposure duration, and probability of recurrence for nine specific body regions. These responses were used to calculate a Severity × Exposure × Probability (SEP) based Musculoskeletal Disorder (MSD) Index, which served as the target variable for classification. The index was normalized and categorized into three risk strata: Low, Medium, and High. The sample exhibited a high burden of musculoskeletal symptoms, with 96.7% of participants reporting pain in at least one body region. Neck pain was the most prevalent complaint, affecting 91.7% of the operators, followed by shoulder pain at 58.3%.

The machine learning framework evaluated eight different classifiers, including logistic regression, naïve Bayes, k-nearest neighbors, decision trees, random forests, support vector machines, AdaBoost, and gradient-boosted trees. To ensure methodological rigor and prevent data leakage, the researchers employed a nested cross-validation strategy. This involved ten repeats of stratified five-fold cross-validation, with hyperparameter tuning performed within the inner training partitions. Preprocessing steps, such as median imputation and one-hot encoding, were strictly contained within the training folds to avoid inflating performance estimates. The primary metrics for model selection were balanced accuracy, macro-F1 score, and Matthews correlation coefficient (MCC), which are particularly useful for handling imbalanced datasets.

Gradient-boosted trees (GBT) emerged as the top-performing model, achieving a mean balanced accuracy of 0.940, a macro-F1 of 0.951, and an MCC of 0.906. Random forests followed closely with a balanced accuracy of 0.936. However, statistical comparisons using exact McNemar tests with Holm correction revealed that only the difference between GBT and k-nearest neighbors was significant. The other six comparisons were not statistically significant, indicating that the numerical differences in performance among the top-performing models may not be robust. This suggests that while GBT had the highest observed accuracy, it cannot be definitively declared superior to other ensemble methods based on this internal evaluation alone.

A critical aspect of the study was the feature ablation analysis, which compared the performance of models using postural variables only, non-postural variables only, and a combined set. Contrary to the initial hypothesis, the non-postural feature set yielded numerically higher mean performance metrics than the combined set. The non-postural model achieved a balanced accuracy of 0.953, whereas the combined model achieved 0.940. Postural variables alone showed moderate discrimination with a balanced accuracy of 0.681. This finding implies that in this specific sample, the demographic, physiological, and work-organization factors were sufficient to classify risk without the added complexity of dynamic postural data.

The study also identified a strong proxy pattern involving physical exercise. All six participants classified as Low risk reported regular physical exercise, while no Medium or High-risk participants did. Removing the physical exercise variable from the model reduced the balanced accuracy from 0.940 to 0.795. This perfect separation indicates that the high internal performance may partly reflect sample-specific structure rather than a generalizable predictive signal. The researchers caution that this relationship should not be interpreted as evidence that exercise alone determines WMSD risk, as the exercise item and symptom data were collected during the same session, potentially introducing common-method bias.

Several limitations constrain the interpretation of these findings. The sample size of 60 records is relatively small, and the Low-risk class contained only six participants, making class-specific estimates unstable. The cross-sectional design precludes causal inference, and the self-reported nature of the symptom data may be subject to recall and reporting bias. Additionally, the use of two-dimensional video for postural analysis may underestimate out-of-plane movements such as shoulder abduction and wrist deviation. The semi-automated RULA workflow required manual correction for tracking failures, and some RULA modifiers were fixed at the task level rather than measured per frame, which may have underestimated postural risk.

The framework is intended as an offline proof of concept for periodic ergonomic decision support, not for real-time clinical diagnosis or routine deployment. The researchers emphasize that independent external validation in larger, multi-site cohorts is required before the model can be considered ready for operational use. Future work should focus on validating the dynamic postural pipeline against three-dimensional motion capture standards and conducting longitudinal studies to establish causal links between ergonomic exposures and WMSD progression. The study contributes a methodologically transparent evaluation framework that highlights the importance of leakage-controlled validation in machine learning applications for occupational health.

The study’s findings align with broader epidemiological evidence indicating that musculoskeletal conditions affect approximately 1.71 billion people worldwide, with low back pain identified as the most common cause of years lived with disability. In the garment and textile industry, a systematic review of 14 studies involving 15,029 workers across nine countries reported a pooled WMSD prevalence estimate of 65.6%. Sewing-machine operators are considered a high-risk group within this sector, with previous research in various countries finding mean RULA scores at action level 3 or higher and 12-month neck-pain prevalences exceeding 40%. The biomechanical exposure profile of sewing work includes prolonged constrained sitting, sustained forward neck flexion for visual guidance of fabric, and high-frequency repetitive upper-limb movements, which collectively increase mechanical loading on the cervical spine, shoulder girdle, trunk, and lumbar region.

Traditional ergonomic assessment tools, such as the Rapid Upper Limb Assessment (RULA), are typically designed as static or snapshot observational instruments that rely on an analyst’s selection of one or a few representative postures. This approach may fail to capture the postural variability, repetitive micro-cycles, and temporal loading patterns characteristic of actual sewing tasks, where exposure is primarily a function of posture over time. To address this, the current study utilized a semi-automated workflow involving Kinovea software to track joint angles from bilateral 10-minute video recordings at 25 frames per second. This process yielded time-weighted RULA indices that summarize cumulative postural exposure across the work cycle, rather than relying on single-point observations. The mean weighted RULA scores for the right and left sides were 4.897 and 5.017, respectively, indicating moderate to high ergonomic risk, with 73.3% and 76.7% of records scoring above 4 on the right and left sides, respectively.

Strong positive correlations were observed between the dynamic weighted RULA scores and the SEP-based MSD Index, with Pearson correlation coefficients of r = 0.77 for the right side and r = 0.75 for the left side. These associations suggest that dynamic posture summaries are relevant indicators of postural exposure related to symptom burden in this dataset. However, the study explicitly notes that correlation does not validate RULA as a causal or independent predictor. The feature ablation analysis further demonstrated that adding dynamic RULA scores to the non-postural variables did not improve mean classification performance, implying that the incremental predictive contribution of these dynamic metrics requires testing in larger, independent samples to establish their value in risk stratification.

The diagnostic learning curve analysis revealed a persistent training-validation gap at smaller sample sizes, with mean outer-validation balanced accuracy increasing from 0.802 at a median of 19 training records to 0.953 at 48 records. Training performance remained near 1.00 across the curve, a pattern compatible with decreasing variance but indicative of continuing overfitting risk. The high training scores and wide split-to-split variability, particularly at smaller sizes, suggest that additional data improved generalisation within the sample, but apparent convergence at the largest available training size does not establish external validity. These findings reinforce the necessity for independent multi-site validation to determine whether the framework’s performance holds in other populations and operational settings.

Subject of Research: Machine learning classification of work-related musculoskeletal disorder risk in garment sewing-machine operators

Article Title: Leakage-controlled machine learning for work-related musculoskeletal disorder risk classification among garment sewing-machine operators

Article References: Hossain, M. Z., Kibria, M. G., Arif, B. U., & Ratul, S. H. (2026). Leakage-controlled machine learning for work-related musculoskeletal disorder risk classification among garment sewing-machine operators. Results in Engineering, 32, Article 113789. https://doi.org/10.1016/j.rineng.2026.113789

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113789

Keywords: musculoskeletal disorders, machine learning, ergonomics, garment industry, occupational health, Bangladesh, RULA, cross-sectional study, risk classification, postural analysis, Leakage-controlled, machine