Surveillance of diseases and associated exposures is a major issue in occupational health especially for identifying new work-related diseases. In addition to classical epidemiology (hypothesis-driven studies), complementary methods relying on data mining of health insurance data must be developed for early detection of work-related diseases, without any prior hypothesis. For this reason, the social security system of French agricultural workers “Mutualité sociale agricole” (MSA) emphasized the potential interest of a surveillance system through the exploitation of their up-to-date databases created for reimbursement of medical expenses. In partnership with the French Agency for Food, Environmental and Occupational Health and Safety (Anses), a pilot study has been set up in order to develop an innovative approach of data mining to look for associations between long-term diseases (LTD) and occupational activities.

The MSA was covering 3.2 million individuals in 2016, with about 1.2 million of active workers. Among them, the study population included all self-employed and steady employees, potentially including individuals who may have retired from 2007, that were registered at least once in MSA databases between 2006 and 2015. MSA holds administrative and medico-administrative databases which include occupational activities, socio-demographic variables (age, sex, incomes), as well as LTD, identified through ICD-10 codes. Following due authorizations especially of the French independent data protection authority (CNIL), MSA databases were cross-linked for the first time and restructured in order to apply logistic models and latent factor models. Obtained p-values and odds ratio (OR) were represented on graphics to highlight the key statistical signals of over-represented statistical associations between occupational activity and LTD.

The population covered by this study accounted for about 2 million individuals (active workers or retirees), with about 900,000 self-employed and 1.2 million steady employees. In both databases, there was a majority of men (65.7%) and the average age was significantly different (P<2.2^E−16) between self-employed (about 53 years) and steady employees (about 40 years). Concerning LTD, about 9.1% of steady employee and 13.4% of self-employed had at least one LTD declaration during the study period. Significant associations were highlighted especially between specific agricultural sectors and particular dementias or specific cancer, which will then be undergo to in-depth expertise. For example, among steady employees, a significant association was found between Parkinson's disease and viticulture (P<0.01). In addition, OR showed that certain occupational activities were concerned by more or fewer LTD reports during the study period. For example, among self-employed, rural artisans seemed to have less reports of malignant tumors (OR=0.3 [0.1–0.5]) and diabetes (OR=0.2 [0.1–0.4]).

This pilot study permitted to prove the feasibility and the relevance of using MSA data for health surveillance of occupational risks among agricultural workers. This approach has the following advantages: 1) enabling systematic evaluations of all disease–occupational activity associations, 2) high statistical power and 3) costless data acquisition. The main drawback is its lack of direct information regarding exposure. For this reason, in partnership with the French national public health agency “Santé publique France”, further work is currently performed to estimate retrospectively pesticides use relying on previous activities and crop-exposure matrices. This data mining approach will later be enriched by identifying diseases using the drugs consumption as a proxy via specific algorithms.