Machine Learning for Work Disability Prevention: Introduction to the Special Series.

Rapid development in computer technology has led to sophisticated methods of analyzing large datasets with the aim of improving human decision making. Artificial Intelligence and Machine Learning (ML) approaches hold tremendous potential for solving complex real-world problems such as those faced by stakeholders attempting to prevent work disability. These techniques are especially appealing in work disability contexts that collect large amounts of data such as workers' compensation settings, insurance companies, large corporations, and health care organizations, among others. However, the approaches require thorough evaluation to determine if they add value to traditional statistical approaches. In this special series of articles, we examine the role and value of ML in the field of work disability prevention and occupational rehabilitation.

Validity of the Work Assessment Triage Tool for Selecting Rehabilitation Interventions for Workers' Compensation Claimants with Musculoskeletal Conditions.

Purpose The Work Assessment Triage Tool (WATT) is a clinical decision support tool developed using machine learning to help select interventions for patients with musculoskeletal disorders. The WATT categorizes patients based on individual characteristics according to likelihood of successful return to work following rehabilitation. A previous validation showed acceptable classification accuracy, but we re-examined accuracy using a new dataset drawn from the same system 2 years later. Methods A population-based cohort design was used, with data extracted from a Canadian compensation database on workers considered for rehabilitation between January 2013 and December 2016. Data were obtained on demographic, clinical, and occupational characteristics, type of rehabilitation undertaken, and return to work outcomes. Analysis included classification accuracy statistics of WATT recommendations. Results The sample included 28,919 workers (mean age 43.9 years, median duration 56 days), of whom 23,124 experienced a positive outcome within 30 days following return to work assessment. Sensitivity of the WATT for selecting successful programs was 0.13 while specificity was 0.87. Overall accuracy was 0.60 while human recommendations were higher at 0.72. Conclusions Overall accuracy of the WATT for selecting successful rehabilitation programs declined in a more recent cohort and proved less accurate than human clinical recommendations. Algorithm revision and further validation is needed.

Developing leading indicators from OHS management audit data: Determining the measurement properties of audit data from the field.

INTRODUCTION: OHS management audits are one means of obtaining data that may serve as leading indicators. The measurement properties of such data are therefore important. This study used data from Workwell audit program in Ontario, a Canadian province. The audit instrument consisted of 122 items related to 17 OHS management elements. The study sought answers regarding (a) the ability of audit-based scores to predict workers' compensation claims outcomes, (b) structural characteristics of the data in relation to the organization of the audit instrument, and (c) internal consistency of items within audit elements. METHOD: The sample consisted of audit and claims data from 1240 unique firms that had completed one or two OHS management audits during 2007-2010. Predictors derived from the audit results were used in multivariable negative binomial regression modeling of workers' compensation claims outcomes. Confirmatory factor analyses were used to examine the instrument's structural characteristics. Kuder-Richardson coefficients of internal consistency were calculated for each audit element. RESULTS: The ability of audit scores to predict subsequent claims data could not be established. Factor analysis supported the audit instrument's element-based structure. KR-20 values were high (≥0.83). CONCLUSIONS: The Workwell audit data display structural validity and high internal consistency, but not, to date, construct validity, since the audit scores are generally not predictive of subsequent firm claim experience. Audit scores should not be treated as leading indicators of workplace OHS performance without supporting empirical data. PRACTICAL APPLICATIONS: Analyses of the measurement properties of audit data can inform decisionmakers about the operation of an audit program, possible future directions in audit instrument development, and the appropriate use of audit data. In particular, decision-makers should be cautious in their use of audit scores as leading indicators, in the absence of supporting empirical data.

Predictive value of the DASH tool for predicting return to work of injured workers with musculoskeletal disorders of the upper extremity.

OBJECTIVES: To determine whether the Disabilities of the Arm, Shoulder, and Hand (DASH) tool added to the predictive ability of established prognostic factors, including patient demographic and clinical outcomes, to predict return to work (RTW) in injured workers with musculoskeletal (MSK) disorders of the upper extremity. METHODS: A retrospective cohort study using a population-based database from the Workers' Compensation Board of Alberta (WCB-Alberta) that focused on claimants with upper extremity injuries was used. Besides the DASH, potential predictors included demographic, occupational, clinical and health usage variables. Outcome was receipt of compensation benefits after 3 months. To identify RTW predictors, a purposeful logistic modelling strategy was used. A series of receiver operating curve analyses were performed to determine which model provided the best discriminative ability. RESULTS: The sample included 3036 claimants with upper extremity injuries. The final model for predicting RTW included the total DASH score in addition to other established predictors. The area under the curve for this model was 0.77, which is interpreted as fair discrimination. This model was statistically significantly different than the model of established predictors alone (p<0.001). When comparing the DASH total score versus DASH item 23, a non-significant difference was obtained between the models (p=0.34). CONCLUSIONS: The DASH tool together with other established predictors significantly helped predict RTW after 3 months in participants with upper extremity MSK disorders. An appealing result for clinicians and busy researchers is that DASH item 23 has equal predictive ability to the total DASH score.

The Added Value of Collecting Information on Pain Experience When Predicting Time on Benefits for Injured Workers with Back Pain.

Objectives Some injured workers with work-related, compensated back pain experience a troubling course in return to work. A prediction tool was developed in an earlier study, using administrative data only. This study explored the added value of worker reported data in identifying those workers with back pain at higher risk of being on benefits for a longer period of time. Methods This was a cohort study of workers with compensated back pain in 2005 in Ontario. Workplace Safety and Insurance Board (WSIB) data was used. As well, we examined the added value of patient-reported prognostic factors obtained from a prospective cohort study. Improvement of model fit was determined by comparing area under the curve (AUC) statistics. The outcome measure was time on benefits during a first workers' compensation claim for back pain. Follow-up was 2 years. Results Among 1442 workers with WSIB data still on full benefits at 4 weeks, 113 were also part of the prospective cohort study. Model fit of an established rule in the smaller dataset of 113 workers was comparable to the fit previously established in the larger dataset. Adding worker rating of pain at baseline improved the rule substantially (AUC = 0.80, 95 % CI 0.68, 0.91 compared to benefit status at 180 days, AUC = 0.88, 95 % CI 0.74, 1.00 compared to benefits status at 360 days). Conclusion Although data routinely collected by workers' compensation boards show some ability to predict prolonged time on benefits, adding information on experienced pain reported by the worker improves the predictive ability of the model from 'fairly good' to 'good'. In this study, a combination of prognostic factors, reported by multiple stakeholders, including the worker, could identify those at high risk of extended duration on disability benefits and in potentially in need of additional support at the individual level.

Predicting time on prolonged benefits for injured workers with acute back pain.

INTRODUCTION: Some workers with work-related compensated back pain (BP) experience a troubling course of disability. Factors associated with delayed recovery among workers with work-related compensated BP were explored. METHODS: This is a cohort study of workers with compensated BP in 2005 in Ontario, Canada. Follow up was 2 years. Data was collected from employers, employees and health-care providers by the Workplace Safety and Insurance Board (WSIB). Exclusion criteria were: (1) no-lost-time claims, (2) >30 days between injury and claim filing, (3) <4 weeks benefits duration, and (4) age >65 years. Using proportional hazard models, we examined the prognostic value of information collected in the first 4 weeks after injury. Outcome measures were time on benefits during the first episode and time until recurrence after the first episode. RESULTS: Of 6,657 workers, 1,442 were still on full benefits after 4 weeks. Our final model containing age, physical demands, opioid prescription, union membership, availability of a return-to-work program, employer doubt about work-relatedness of injury, worker's recovery expectations, participation in a rehabilitation program and communication of functional ability was able to identify prolonged claims to a fair degree [area under the curve (AUC) = .79, 95% confidence interval (CI) .74-.84]. A model containing age, sex, physical demands, opioid prescription and communication of functional ability was less successful at predicting time until recurrence (AUC = .61, 95% CI .57, .65). CONCLUSIONS: Factors contained in information currently collected by the WSIB during the first 4 weeks on benefits can predict prolonged claims, but not recurrent claims.

Systematic review and network meta-analysis of interventions for fibromyalgia: a protocol.

BACKGROUND: Fibromyalgia is associated with substantial socioeconomic loss and, despite considerable research including numerous randomized controlled trials (RCTs) and systematic reviews, there exists uncertainty regarding what treatments are effective. No review has evaluated all interventional studies for fibromyalgia, which limits attempts to make inferences regarding the relative effectiveness of treatments. METHODS/DESIGN: We will conduct a network meta-analysis of all RCTs evaluating therapies for fibromyalgia to determine which therapies show evidence of effectiveness, and the relative effectiveness of these treatments. We will acquire eligible studies through a systematic search of CINAHL, EMBASE, MEDLINE, AMED, HealthSTAR, PsychINFO, PapersFirst, ProceedingsFirst, and the Cochrane Central Registry of Controlled Trials. Eligible studies will randomly allocate patients presenting with fibromyalgia or a related condition to an intervention or a control. Teams of reviewers will, independently and in duplicate, screen titles and abstracts and complete full text reviews to determine eligibility, and subsequently perform data abstraction and assess risk of bias of eligible trials. We will conduct meta-analyses to establish the effect of all reported therapies on patient-important outcomes when possible. To assess relative effects of treatments, we will construct a random effects model within the Bayesian framework using Markov chain Monte Carlo methods. DISCUSSION: Our review will be the first to evaluate all treatments for fibromyalgia, provide relative effectiveness of treatments, and prioritize patient-important outcomes with a focus on functional gains. Our review will facilitate evidence-based management of patients with fibromyalgia, identify key areas for future research, and provide a framework for conducting large systematic reviews involving indirect comparisons.

Incidence of work and non-work related disability claims in Brazil.

BACKGROUND: Sickness benefit claims are an important economic burden to society. This study aims to determine the incidence of sickness benefit claims in Brazil in 2008, exploring the role of economic activity. METHODS: Population-based study on sickness claims lasting longer than 15 days of sickness absence granted to private sector employees. Data on gender, age, economic activity, diagnosis, and work-relatedness were collected. RESULTS: The annual incidence of sickness benefits was 421.8/10,000 jobs, 435.4 for males and 452.0 for females. There were 3.5 times more non-work-related than work-related claims. The main diagnoses were injuries, musculoskeletal disorders, and mental disorders. Rates increased with age up to 59 years. Economic activity 37-Sewage had the highest incidence of non-work-related and work-related claims. CONCLUSION: The incidence of sickness benefits is higher among female and older workers. Economic activities show great variability of sickness benefit rates, work-relatedness, diagnostic categories, and gender.

Validation of a risk factor-based intervention strategy model using data from the readiness for return to work cohort study.

BACKGROUND: Low back pain (LBP) is a common and in some cases disabling condition. Until recently, workers presenting with non-specific LBP have generally been regarded as a homogeneous population. If this population is not homogeneous, different interventions might be appropriate for different subgroups. We hypothesized that (1) Clusters of individuals could be identified based on risk factors, (2) These clusters would predict duration and recurrences 6 months post-injury. METHODS: The study focuses on the 442 LBP claimants in the Readiness for Return-to-Work Cohort Study. Claimants (n = 259) who had already returned to work, approximately 1 month post-injury were categorized as the low risk group. A latent class analysis was performed on 183 workers absent from work, categorized as the high risk group. Groups were classified based on: pain, disability, fear avoidance beliefs, physical demands, people-oriented culture and disability management practice at the workplace, and depressive symptoms. RESULTS: Three classes were identified; (1) workers with 'workplace issues', (2) workers with a 'no workplace issues, but back pain', and (3) workers having 'multiple issues' (the most negative values on every scale, notably depressive symptoms). Classes 2 and 3 had a similar rate of return to work, both worse than the rate of class 1. Return-to-work status and recurrences at 6 months were similar in all 3 groups. CONCLUSION: This study largely confirms that several subgroups could be identified based on previously defined risk factors as suggested by an earlier theoretical model by Shaw et al. (J Occup Rehab 16(4):591-605, 2006). Different groups of workers might be identified and might benefit from different interventions.

Economic evaluation of a multi-stage return to work program for workers on sick-leave due to low back pain.

OBJECTIVE: To evaluate the cost-effectiveness and cost-utility of a return to work (RTW) program for workers on sick-leave due to low back pain (LBP), comparing a workplace intervention implemented between 2 to 8 weeks of sick-leave with usual care, and a clinical intervention after 8 weeks of sick-leave with usual care. DESIGN: Economic evaluation alongside a randomised controlled trial (RCT). STUDY POPULATION: Workers sick-listed for a period of 2 to 6 weeks due to LBP. INTERVENTIONS: 1. workplace assessment, work modifications and case management). 2. physiotherapy based on operant behavioural principles. 3. usual care: provided by an occupational physician. OUTCOMES: The primary outcome was return to work (RTW). Other outcomes were pain intensity, functional status, quality of life and general health. The economic evaluation was conducted from a societal perspective. Outcomes were assessed at baseline (after 2-6 weeks on sick-leave), and 12 weeks, 26 weeks, and 52 weeks after the first day of sick-leave. RESULTS: The workplace intervention group returned to work 30.0 days (95% CI=[3.1, 51.3]) earlier on average than the usual care group at slightly higher direct costs (ratio of 1 day: 19 euro). Workers in the clinical intervention group that had received usual care in the first 8 weeks returned to work 21.3 days (95% CI= [-74.1, 29.2]) later on average. The group that had received the workplace intervention in the first 8 weeks and the clinical intervention after 8 weeks returned to work 50.9 days (95% CI=[-89.4, -2.7]) later on average. A workplace intervention was more effective than usual care in RTW at slightly higher costs and was equally effective as usual care at equal costs on other outcomes. A clinical intervention was less effective than usual care and associated with higher costs. CONCLUSION: The workplace intervention results in a safe and faster RTW than usual care at reasonable costs for workers on sick-leave for two to six weeks due to LBP.

Cost effectiveness of a multi-stage return to work program for workers on sick leave due to low back pain, design of a population based controlled trial [ISRCTN60233560].

BACKGROUND: To describe the design of a population based randomized controlled trial (RCT), including a cost-effectiveness analysis, comparing participative ergonomics interventions between 2-8 weeks of sick leave and Graded Activity after 8 weeks of sick leave with usual care, in occupational back pain management. DESIGN: An RCT and cost-effectiveness evaluation in employees sick-listed for a period of 2 to 6 weeks due to low back pain. Interventions used are 1. Communication between general practitioner and occupational physician plus Participative Ergonomics protocol performed by an ergonomist. 2. Graded Activity based on cognitive behavioural principles by a physiotherapist. 3. Usual care, provided by an occupational physician according to the Dutch guidelines for the occupational health management of workers with low back pain. The primary outcome measure is return to work. Secondary outcome measures are pain intensity, functional status and general improvement. Intermediate variables are kinesiophobia and pain coping. The cost-effectiveness analysis includes the direct and indirect costs due to low back pain. The outcome measures are assessed before randomization (after 2-6 weeks on sick leave) and 12 weeks, 26 weeks and 52 weeks after first day of sick leave. DISCUSSION: The combination of these interventions has been subject of earlier research in Canada. The results of the current RCT will: 1. crossvalidate the Canadian findings in an different sociocultural environment; 2. add to the cost-effectiveness on treatment options for workers in the sub acute phase of low back pain. Results might lead to alterations of existing (inter)national guidelines.