Predicting musculoskeletal disorders risk using tree-based ensemble methods

Abstract

Abstract Background Musculoskeletal disorders (MSD) can cause short-term disorders and permanent disabilities which may all result in serious limitations in activities of daily living. According to the WHO, MSD are the second largest cause of disability worldwide. In France, preventing MSD became a challenge for both the sickness fund and the complementary health insurers as the MSD prevalence increased by more than 50% since 2003, the MSD representing more than 80% of the occupational diseases leading to sick-leaves or financial compensations in 2015. Methods Data from the IRDES 2012 health and social welfare survey (ESPS, enquête santé et protection sociale) was matched with individual care consumptions from the French health insurance. The ESPS survey is a national representative survey including 23,047 CATI/CAPI interviewed individuals (participation rate: 66%). Respondents self-assessed their health status, including MSD among other health troubles. Prediction of MSD risk was designed as a supervised binary classification problem, with the help of Random Forest (RF) and Gradient Boosting Machine (GBM) methods. Results Variable importance scores based on the squared error criterion over all trees were computed for the continuous and categorical variables supposed to be associated with MSD: the consumption of medicines, physiotherapists’ and GPs’ services, medical imaging, biomedical analyses, dental care, and sociodemographic features (occupation, age and income). Slightly better results were obtained with GBM compared to RF (overall test accuracy rate: 70% for both methods). Conclusions Machine learning methods with merged survey and registry data may help health insurers to better identify the individual MSD risk and prevent the occurrence and the disabling consequences of MSD with the help of individualized prevention strategies. This research is financially supported by Malakoff Mederic Humanis, life & health French insurance for employees and individuals. Key messages The use of machine learning methods may be of interest to improve the ability to predict the MSD risk and help to identify insured people whom individualized prevention strategies can be offered to. Matching survey and health insurance liquidation data improves the ability to identify people at improved risk of MSD and gives the opportunity to provide them resources to prevent and to cope with.