Prediction of Return-to-original-work after an Industrial Accident Using Machine Learning and Comparison of Techniques

A Scoping Review on the Use of Machine Learning in Return-to-Work Studies: Strengths and Weaknesses

PURPOSE

Decisions to increase work participation must be informed and timely to improve return to work (RTW). The implementation of research into clinical practice relies on sophisticated yet practical approaches such as machine learning (ML). The objective of this study is to explore the evidence of machine learning in vocational rehabilitation and discuss the strengths and areas for improvement in the field.

METHODS

We used the PRISMA guidelines and the Arksey and O'Malley framework. We searched Ovid Medline, CINAHL, and PsycINFO; with hand-searching and use of the Web of Science for the final articles. We included studies that are peer-reviewed, published within the last 10 years to consider contemporary material, implemented a form of "machine learning" or "learning health system", undertaken in a vocational rehabilitation setting, and has employment as a specific outcome.

RESULTS

12 studies were analyzed. The most commonly studied population was musculoskeletal injuries or health conditions. Most of the studies came from Europe and most were retrospective studies. The interventions were not always reported or specified. ML was used to identify different work-related variables that were predictive of return to work. However, ML approaches were varied and no standard or predominant ML approach was evident.

CONCLUSIONS

ML offers a potentially beneficial approach to identifying predictors of RTW. While ML uses a complex calculation and estimation, ML complements other elements of evidence-based practice such as the clinician's expertise, the worker's preference and values, and contextual factors around RTW in an efficient and timely manner.

Comparison of Machine Learning Methods in the Study of Cancer Survivors' Return to Work: An Example of Breast Cancer Survivors with Work-Related Factors in the CONSTANCES Cohort

PURPOSE

Machine learning (ML) methods showed a higher accuracy in identifying individuals without cancer who were unable to return to work (RTW) compared to the classical methods (e.g. logistic regression models). We therefore aim to discuss the value of these methods in relation to RTW for cancer survivors.

METHODS

Breast cancer (BC) survivors who were working at diagnosis within the CONSTANCES cohort were included in the study. RTW was assessed five years after the BC diagnosis (early retirement was considered as non-RTW). Age and occupation at diagnosis, and physical occupational job exposures assessed using the Job Exposure Matrix, JEM-CONSTANCES, were evaluated as predictors of RTW five years after BC diagnosis. The following four ML methods were used: (i) k-nearest neighbors; (ii) random forest; (iii) neural network; and (iv) elastic net.

RESULTS

The training sample included 683 BC survivors (RTW: 85.7%), and the test sample 171 (RTW: 85.4%). The elastic net method had the best results despite low sensitivity (accuracy = 76.6%; sensitivity = 31.7%; specificity = 90.8%), and the random forest model was the most accurate (= 79.5%) but also the least sensitive (= 14.3%).

CONCLUSION

This study takes a first step towards opening up new possibilities for identifying the occupational determinants of cancer survivors' RTW. Further work, including a larger sample size, and more predictor variables, is now needed.

Exploring predictors of welfare dependency 1, 3, and 5 years after mental health-related absence in danish municipalities between 2010 and 2012 using flexible machine learning modelling

BACKGROUND

Using XGBoost (XGB), this study demonstrates how flexible machine learning modelling can complement traditional statistical modelling (multinomial logistic regression) as a sensitivity analysis and predictive modelling tool in occupational health research.

DESIGN

The study predicts welfare dependency for a cohort at 1, 3, and 5 years of follow-up using XGB and multinomial logistic regression (MLR). The models' predictive ability is evaluated using tenfold cross-validation (internal validation) and geographical validation (semi-external validation). In addition, we calculate and graphically assess Shapley additive explanation (SHAP) values from the XGB model to examine deviation from linearity assumptions, including interactions. The study population consists of all 20-54 years old on long-term sickness absence leave due to self-reported common mental disorders (CMD) between April 26, 2010, and September 2012 in 21 (of 98) Danish municipalities that participated in the Danish Return to Work program. The total sample of 19.664 observations is split geospatially into a development set (n = 9.756) and a test set (n = 9.908).

RESULTS

There were no practical differences in the XGB and MLR models' predictive ability. Industry, job skills, citizenship, unemployment insurance, gender, and period had limited importance in predicting welfare dependency in both models. On the other hand, welfare dependency history and reason for sickness absence were strong predictors. Graphical SHAP-analysis of the XGB model did not indicate substantial deviations from linearity assumptions implied by the multinomial regression model.

CONCLUSION

Flexible machine learning models like XGB can supplement traditional statistical methods like multinomial logistic regression in occupational health research by providing a benchmark for predictive performance and traditional statistical models' ability to capture important associations for a given set of predictors as well as potential violations of linearity.

TRIAL REGISTRATION

ISRCTN43004323.

Return-to-Work Predictions for Chinese Patients With Occupational Upper Extremity Injury: A Prospective Cohort Study

Objective

We created predictive models using machine learning algorithms for return-to-work (RTW) in patients with traumatic upper extremity injuries.

Methods

Data were obtained immediately before patient discharge and patients were followed up for 1 year. K-nearest neighbor, logistic regression, support vector machine, and decision tree algorithms were used to create our predictive models for RTW.

Results

In total, 163 patients with traumatic upper extremity injury were enrolled, and 107/163 (65.6%) had successfully returned to work at 1-year of follow-up. The decision tree model had a lower F1-score than any of the other models (t values: 7.93–8.67, p < 0.001), while the others had comparable F1-scores. Furthermore, the logistic regression and support vector machine models were significantly superior to the k-nearest neighbors and decision tree models in the area under the receiver operating characteristic curve (t values: 6.64–13.71, p < 0.001). Compared with the support vector machine, logistical regression selected only two essential factors, namely, the patient's expectation of RTW and carrying strength at the waist, suggesting its superior efficiency in the prediction of RTW.

Conclusion

Our study demonstrated that high predictability for RTW can be achieved through use of machine learning models, which is helpful development of individualized vocational rehabilitation strategies and relevant policymaking.

Smart Work Injury Management (SWIM) System: Artificial Intelligence in Work Disability Management

PURPOSE

This paper aims to illustrate an example of how to set up a work injury database: the Smart Work Injury Management (SWIM) system. It is a secure and centralized cloud platform containing a set of management tools for data storage, data analytics, and machine learning. It employs artificial intelligence to perform in-depth analysis via text-mining techniques in order to extract both dynamic and static data from work injury case files. When it is fully developed, this system can provide a more accurate prediction model for cost of work injuries. It can also predict return-to-work (RTW) trajectory and provide advice on medical care and RTW interventions to all RTW stakeholders. The project will comprise three stages. Stage one: to identify human factors in terms of both facilitators and barriers RTW through face-to-face interviews and focus group discussions with different RTW stakeholders in order to collect opinions related to facilitators, barriers, and essential interventions for RTW of injured workers; Stage two: to develop a machine learning model which employs artificial intelligence to perform in-depth analysis. The technologies used will include: 1. Text-mining techniques including English and Chinese work segmentation as well as N-Gram to extract both dynamic and static data from free-style text as well as sociodemographic information from work injury case files; 2. Principle component/independent component analysis to identify features of significant relationships with RTW outcomes or combine raw features into new features; 3. A machine learning model that combines Variational Autoencoder, Long and Short Term Memory, and Neural Turning Machines. Stage two will also include the development of an interactive dashboard and website to query the trained machine learning model. Stage three: to field test the SWIM system.

CONCLUSION

SWIM ia secure and centralized cloud platform containing a set of management tools for data storage, data analytics, and machine learning. When it is fully developed, SWIM can provide a more accurate prediction model for the cost of work injuries and advice on medical care and RTW interventions to all RTW stakeholders.

ETHICS

The project has been approved by the Ethics Committee for Human Subjects at the Hong Kong Polytechnic University and is funded by the Innovation and Technology Commission (Grant # ITS/249/18FX).

Assessment of a new tool to improve case manager identification of delayed return to work in the first two weeks of a workers' compensation claim

OBJECTIVE

To determine whether the Plan of Action for a Case (PACE) tool improved identification of workers at risk of delayed return to work.

DESIGN

Prospective cohort of workers with accepted workers' compensation claims in the state of New South Wales, Australia.

INTERVENTIONS

The 41-item PACE tool was completed by the case manager within the first two weeks of a claim. The tool gathered information from the worker, employer and treating practitioner. Multivariate logistic regression models predicted work time loss of at least one and three months.

RESULTS

There were 524 claimants with complete PACE information. A total of 195 (37.2%) had work time loss of at least one month and 83 (15.8%) had time loss of at least three months. Being male, injury location, an Orebro Musculoskeletal Pain Screening Questionnaire-Short Form score >50, having a small employer, suitable duties not being available, being certified unfit, and the worker having low one-month recovery expectations predicted time loss of over one month. For three months, injury location, a Short Form Orebro score >50, no return-to-work coordinator, and being certified unfit were significant predictors. The model incorporating PACE information provided a significantly better prediction of both one- and three-month outcomes than baseline information (area-under-the-curve statistics-one month: 0.85 and 0.68, respectively; three months: 0.85 and 0.69, respectively; both P < 0.001).

CONCLUSION

The PACE tool improved the ability to identify workers at risk of ongoing work disability and identified modifiable factors suited to case manager-led intervention.