Search

International audience; Abstract Background The use of machine learning techniques is increasing in healthcare which allows to estimate and predict health outcomes from large administrative data sets more efficiently. The main objective of this study was to develop a generic machine learning (ML) algorithm to estimate the incidence of diabetes based on the number of reimbursements over the last 2 years. Methods We selected a final data set from a population-based epidemiological cohort (i.e., CONSTANCES) linked with French National Health Database (i.e., SNDS). To develop this algorithm, we adopted a supervised ML approach. Following steps were performed: i. selection of final data set, ii. target definition, iii. Coding variables for a given window of time, iv. split final data into training and test data sets, v. variables selection, vi. training model, vii. Validation of model with test data set and viii. Selection of the model. We used the area under the receiver operating characteristic curve (AUC) to select the best algorithm. Results The final data set used to develop the algorithm included 44,659 participants from CONSTANCES. Out of 3468 variables from SNDS linked to CONSTANCES cohort were coded, 23 variables were selected to train different algorithms. The final algorithm to estimate the incidence of diabetes was a Linear Discriminant Analysis model based on number of reimbursements of selected variables related to biological tests, drugs, medical acts and hospitalization without a procedure over the last 2 years. This algorithm has a sensitivity of 62%, a specificity of 67% and an accuracy of 67% [95% CI: 0.66–0.68]. Conclusions Supervised ML is an innovative tool for the development of new methods to exploit large health administrative databases. In context of InfAct project, we have developed and applied the first time a generic ML-algorithm to estimate the incidence of diabetes for public health surveillance. The ML-algorithm we have developed, has a moderate performance. The next step is to apply this algorithm on SNDS to estimate the incidence of type 2 diabetes cases. More research is needed to apply various MLTs to estimate the incidence of various health conditions.

Additional file 1. It describes the summary of results with three boosted algorithms [boosted version of logistic regression, boosted C5.0 classification model, and XGBoost] to our set of four models [i.e., 1. Linear discriminant analysis (LDA), 2. Logistic regression (LR), 3. Flexible discriminant analysis (FDA) and 4. Decision tree model (C5)]. It includes Table S 1.1 (main analysis), Table S1.2 (sensitivity analysis: over sampling), Table S1.3 (sensitivity analysis: down sampling), Table S2 (description of models used and the set of hyperparameters explored) and Table S3 (Area under curve: AUC).

International audience; Objectives To determine whether the terrorist attacks occurring in Paris on November 2015 have changed benzodiazepine use in the French population. Design Interrupted time series analysis. Setting National population-based cohort. Participants 90 258 individuals included in the population-based CONSTANCES cohort from 2012 to 2017. Outcome measures Benzodiazepine use was evaluated according to two different indicators using objective data from administrative registries: weekly number of individuals with a benzodiazepine delivered prescriptions (BDP) and weekly number of defined daily dose (DDD). Two sets of analyses were performed according to sex and age (≤50 vs >50). Education, income and area of residence were additional stratification variables to search for at-risk subgroups. Results Among women, those with younger age (incidence rate ratios (IRR)=1.18; 95% CI=1.05 to 1.32 for BDP; IRR=1.14; 95% CI=1.03 to 1.27 for DDD), higher education (IRR=1.23; 95% CI=1.03 to 1.46 for BDP; IRR=1.23; 95% CI=1.01 to 1.51 for DDD) and living in Paris (IRR=1.27; 95% CI=1.05 to 1.54 for BDP) presented increased risks for benzodiazepine use. Among participants under 50, an overall increase in benzodiazepine use was identified (IRR=1.14; 95% CI=1.02 to 1.28 for BDP and IRR=1.12; 95% CI=1.01 to 1.25 for DDD) and in several strata. In addition to women, those with higher education (IRR=1.22; 95% CI=1.02 to 1.47 for BDP), lower income (IRR=1.17; 95% CI=1.02 to 1.35 for BDP) and not Paris residents (IRR=1.13; 95% CI=1.02 to 1.26 for BDP and IRR=1.13; 95% CI=1.03 to 1.26 for DDD) presented increased risks for benzodiazepine use. Conclusion Terrorist attacks might increase benzodiazepine use at a population level, with at-risk subgroups being particularly concerned. Information and prevention strategies are needed to provide appropriate care after such events.

INTRODUCTION - Le Système national des données de santé (SNDS) est une source d’informations majeure pour la surveillance du diabète. L’identification des cas de diabète repose sur des algorithmes basés sur le traitement pharmacologique sans distinction entre type 1 (DT1) et type 2 (DT2). Les objectifs de cette étude étaient le développement d’un algorithme de typage du diabète en utilisant une approche d’intelligence artificielle (IA) et son application pour estimer la prévalence du DT1 et DT2 chez l’adulte en France. METHODES - L’algorithme a été développé à partir des participants traités pharmacologiquement pour diabète dans la cohorte Constances (n= 951, base de données [BdD] de référence). Une méthode d’apprentissage automatique supervisé a été utilisée, déclinée en huit étapes : sélection de la BdD de référence, identification de la cible (DT1), codification des variables SNDS, division de la BdD en base d’entrainement et base de test, sélection des variables et entrainement, validation et sélection des algorithmes. L’algorithme sélectionné a été appliqué sur l’ensemble du SNDS pour estimer, après correction basée sur sa performance, la prévalence des DT1 et DT2 en 2016, déclinée par sexe, chez les adultes âgés de 18 à 70 ans. RESULTATS - Sur 3481 variables codifiées dans le SNDS, seules 14 étaient sélectionnées pour entrainer les différents algorithmes. L’algorithme final est un modèle d’analyse discriminante linéaire basé sur le nombre de remboursements dans l’année précédente : d’insuline à action rapide, d’insuline de longue durée et de biguanides. Cet algorithme a une spécificité de 97,2 % et une sensibilité de 100% pour l’identification des DT1. En 2016, la prévalence du DT1 était 0,32% (femmes 0,29% ; hommes 0,36%) et celle du DT2 était 4,36% (femmes 3,72% ; hommes 5,03%). CONCLUSION - Les perspectives de recherche et prévention offertes par l’IA sont nombreuses et dépassent le champ de la surveillance du diabète.

INTRODUCTION - En 2013-2014, selon les données de la cohorte Constances, 1,6% de la population française âgée de 18 à 69 ans avait un diabète méconnu. L’objectif de notre étude était de développer un algorithme pour identifier les cas de diabète méconnu dans le Système National des données de santé (SNDS) en utilisant l’intelligence artificielle. METHODES - L’algorithme a été développé à partir de la cohorte Constances dans laquelle des données d’auto-questionnaire, de questionnaire médical et des résultats biologiques sont appariés avec les données du SNDS. Nous avons utilisé une méthodologie d’apprentissage automatique supervisé composée de huit étapes. Premièrement, nous avons sélectionné la base de données (BdD) de référence, en excluant les cas de diabète connu. Parmi les 44,185 participants, nous avons identifié comme cible les cas de diabète méconnu - glycémie à jeun ≥7 mmol/l (n=655)-. Les étapes suivantes étaient : codification des variables SNDS, division de la BdD de référence en base d’entrainement et base de test, sélection des variables et entrainement, validation et sélection des algorithmes. RESULTATS - Seules 12 des 3471 variables codées étaient retenues pour leur capacité de discrimination entre la cible : diabète méconnu versus pas de diabète. L’algorithme final est un modèle de régression logistique basé sur les 5 variables les plus discriminantes : âge, sexe et nombre de remboursements (hors hôpital public) dans l’année précédente d’explorations d’une anomalie lipidique, de consultations d’un médecin généraliste et de dosages de glycémie. La spécificité, la sensibilité et la précision de l’algorithme étaient de 70%, 71 % et 69%, respectivement. CONCLUSION - L’intelligence artificielle ouvre de nombreuses perspectives en termes de prévention du diabète. Ainsi, l’identification des personnes à très haut risque permettrait de cibler les personnes à inclure dans les campagnes de prévention et de leur offrir une prise en charge spécifique.

International audience; From an epidemiological standpoint, quantifying the individual and the combined effect of lifestyle factors on uncontrolled blood pressure (BP) deserves further evaluation. We aimed to examine the individual and combined associations between unhealthy behaviors and uncontrolled hypertension among treated hypertensive adults. Cross-sectional analysis was conducted using data from CONSTANCES, an ongoing French population-based cohort study. Uncontrolled BP was defined as mean systolic BP ≥140 mmHg and/or mean diastolic BP ≥90 mmHg. Unhealthy behaviors were considered as heavy alcohol consumption, low or medium adherence to dietary recommendations, sedentary physical activity level, and overweight. A total of 10,710 hypertensive treated volunteer participants were included and 56.1% had uncontrolled hypertension; of them, 2.0%, 24.5%, 54.0% and 19.5% exhibited 0, 1, 2 or ≥3 unhealthy behaviors respectively. In men, there was an increased odds of uncontrolled hypertension with heavy alcohol drinking compared to light-or-never (adjusted odds ratio 1.34, 95% CI 1.10-1.63), with low as well as with medium adherence to dietary recommendations compared to high (p

International audience; Objectives: We aimed to evaluate the individual and combined association between unhealthy behavior and hypertension (HTN).Methods: We conducted cross-sectional analysis using data from the population-based cohort study CONSTANCES. Blood pressure measurements were taken based on standardized operational procedures. Dietary adherence was done following the dietary approach to stop HTN diet. We considered heavy alcohol drinking, sedentary-level physical activity, low/medium dietary adherence and overweight/obesity as unhealthy behaviors. Participants' characteristics were compared according to the number of unhealthy behaviors and the association between HTN and unhealthy behaviors was estimated using logistic regression.Results: A total of 86 448 volunteer participants were included and the prevalence of HTN was of 31.1%. Of those with HTN, 8.2, 33.0, 44.3 and 14.5% exhibited 0, 1, 2 or at least 3 unhealthy behaviors, respectively. In both sexes, the prevalence of HTN increased with low/medium dietary adherence compared with high (P < 0.01), in overweight/obese compared with normal BMI (P < 0.001), with heavy alcohol consumption compared with moderate or never (P < 0.05) and with sedentary physical activity level compared to high in women only (P = 0.049). Combination of several unhealthy behaviors was associated with increased odds of HTN (P trend

Background Parkinson’s disease (PD) is 1.5 times more frequent in men than women. Whether age modifies this ratio is unclear. We examined whether male-to-female (M–F) ratios change with age through a French nationwide prevalence/incidence study (2010) and a meta-analysis of incidence studies. Methods We used French national drug claims databases to identify PD cases using a validated algorithm. We computed M–F prevalence/incidence ratios overall and by age using Poisson regression. Ratios were regressed on age to estimate their annual change. We identified all PD incidence studies with age/sex-specific data, and performed a meta-analysis of M–F ratios. Results On the basis of 149 672 prevalent (50% women) and 25 438 incident (49% women) cases, age-standardised rates were higher in men (prevalence=2.865/1000; incidence=0.490/1000 person-years) than women (prevalence=1.934/1000; incidence=0.328/1000 person-years). The overall M–F ratio was 1.48 for prevalence and 1.49 for incidence. Prevalence and incidence M–F ratios increased by 0.05 and 0.14, respectively, per 10 years of age. Incidence was similar in men and women under 50 years (M–F ratio 0.20), and over 1.6 (p

The role of occupational exposure to pesticides in Parkinson's disease (PD) is recognized, but no studies have evaluated the excess risk of PD among the French agricultural population. In addition, few studies have examined non-occupational exposure. We used databases from the national health insurance information system (SNNIRAM) to identify incident PD cases in metropolitan France (2010-2012). We compared the incidence and prevalence of PD among affiliates of the Mutualité agricole agricole with those among affiliates of the other health insurance schemes and observed an increased frequency of PD among MSA affiliates, farmers in particular. In the French general population, the incidence of PD increased with the proportion of land devoted to agriculture in the cantons. occupational exposure was involved. The strongest association was observed for cantons with a high proportion of land devoted to vineyards.This association was confirmed in non-farmers affiliated to the General health insurance scheme. The association with vineyards may be explained by an important use of pesticides leading to environmental exposure near farms. If this association is confirmed, the fraction of PD attributable to pesticides would be greater than if only occupational exposure was involved.Amyotrophic lateral sclerosis or motor neuron disease (MMD) is a rare disease with a poor prognosis and there are few data on its incidence in France. We have developed an algorithm to identify cases of MMD in the SNIIRAM which allowed us to estimate the incidence of this disease in France (2012-2014) and to study its relationship with the agricultural characteristics. Unlike PD, we did not observe any increase of incidence among MSA members and we did not find any association with agricultural characteristics.; Le rôle de l'exposition professionnelle aux pesticides dans la maladie de Parkinson (MP) est documenté, mais aucune étude n’a évalué l’excès de risque de MP parmi la population agricole française. De plus, peu d’études ont porté sur l’exposition non-professionnelle. A partir des bases de données du système national d’information inter-régimes de l’assurance maladie (SNNIRAM), nous avons identifié l’ensemble des cas incidents de MP en France métropolitaine (2010-2012). Nous avons comparé l’incidence et la prévalence de la MP parmi les affiliés à la Mutualité sociale agricole à celles des affiliés des autres régimes de l’assurance maladie et observé une augmentation de fréquence de MP parmi les affiliés à la MSA, notamment les exploitants agricoles. Parmi la population française métropolitaine, l’incidence de la MP augmentait avec la proportion de terres consacrées à l’agriculture dans les cantons. L’association la plus forte a été observée pour les cantons fortement viticoles. Cette association a été confirmée chez les non-agriculteurs affiliés au Régime général de l’assurance maladie. L’association avec la viticulture pourrait s’expliquer par une utilisation importante de pesticides responsable d’une exposition environnementale à proximité des exploitations. Si cette association est confirmée, la fraction de MP attribuable aux pesticides serait plus importante que si seule l’exposition professionnelle était impliquée.La sclérose latérale amyotrophique ou maladie du motoneurone (MMN) est une maladie rare de pronostic sombre et il existe peu de données sur son incidence en France. Nous avons développé un algorithme permettant d’identifier les cas de MMN à partir du SNIIRAM qui a permis d’estimer l’incidence de cette pathologie en France (2012-2014) et d’étudier sa relation avec les caractéristiques agricoles. A l’inverse de la MP, nous n’avons pas observé d’augmentation d’incidence au sein de la MSA et nous n’avons pas retrouvé d’association avec les caractéristiques agricoles.

The role of occupational exposure to pesticides in Parkinson's disease (PD) is recognized, but no studies have evaluated the excess risk of PD among the French agricultural population. In addition, few studies have examined non-occupational exposure. We used databases from the national health insurance information system (SNNIRAM) to identify incident PD cases in metropolitan France (2010-2012). We compared the incidence and prevalence of PD among affiliates of the Mutualité agricole agricole with those among affiliates of the other health insurance schemes and observed an increased frequency of PD among MSA affiliates, farmers in particular. In the French general population, the incidence of PD increased with the proportion of land devoted to agriculture in the cantons. occupational exposure was involved. The strongest association was observed for cantons with a high proportion of land devoted to vineyards.This association was confirmed in non-farmers affiliated to the General health insurance scheme. The association with vineyards may be explained by an important use of pesticides leading to environmental exposure near farms. If this association is confirmed, the fraction of PD attributable to pesticides would be greater than if only occupational exposure was involved.Amyotrophic lateral sclerosis or motor neuron disease (MMD) is a rare disease with a poor prognosis and there are few data on its incidence in France. We have developed an algorithm to identify cases of MMD in the SNIIRAM which allowed us to estimate the incidence of this disease in France (2012-2014) and to study its relationship with the agricultural characteristics. Unlike PD, we did not observe any increase of incidence among MSA members and we did not find any association with agricultural characteristics.; Le rôle de l'exposition professionnelle aux pesticides dans la maladie de Parkinson (MP) est documenté, mais aucune étude n’a évalué l’excès de risque de MP parmi la population agricole française. De plus, peu d’études ont porté sur l’exposition non-professionnelle. A partir des bases de données du système national d’information inter-régimes de l’assurance maladie (SNNIRAM), nous avons identifié l’ensemble des cas incidents de MP en France métropolitaine (2010-2012). Nous avons comparé l’incidence et la prévalence de la MP parmi les affiliés à la Mutualité sociale agricole à celles des affiliés des autres régimes de l’assurance maladie et observé une augmentation de fréquence de MP parmi les affiliés à la MSA, notamment les exploitants agricoles. Parmi la population française métropolitaine, l’incidence de la MP augmentait avec la proportion de terres consacrées à l’agriculture dans les cantons. L’association la plus forte a été observée pour les cantons fortement viticoles. Cette association a été confirmée chez les non-agriculteurs affiliés au Régime général de l’assurance maladie. L’association avec la viticulture pourrait s’expliquer par une utilisation importante de pesticides responsable d’une exposition environnementale à proximité des exploitations. Si cette association est confirmée, la fraction de MP attribuable aux pesticides serait plus importante que si seule l’exposition professionnelle était impliquée.La sclérose latérale amyotrophique ou maladie du motoneurone (MMN) est une maladie rare de pronostic sombre et il existe peu de données sur son incidence en France. Nous avons développé un algorithme permettant d’identifier les cas de MMN à partir du SNIIRAM qui a permis d’estimer l’incidence de cette pathologie en France (2012-2014) et d’étudier sa relation avec les caractéristiques agricoles. A l’inverse de la MP, nous n’avons pas observé d’augmentation d’incidence au sein de la MSA et nous n’avons pas retrouvé d’association avec les caractéristiques agricoles.

Objective: There are no estimates of the nationwide incidence of motor neuron disease (MND) in France. We used the French health insurance information system to identify incident MND cases (2012–2014), and compared incidence figures to those from three external sources. Methods: We identified incident MND cases (2012–2014) based on three data sources (riluzole claims, hospitalisation records, long-term chronic disease benefits), and computed MND incidence by age, gender, and geographic region. We used French mortality statistics, Limousin ALS registry data, and previous European studies based on administrative databases to perform external comparisons. Results: We identified 6553 MND incident cases. After standardisation to the United States 2010 population, the age/gender-standardised incidence was 2.72/100,000 person-years (males, 3.37; females, 2.17; male:female ratio = 1.53, 95% CI1.46–1.61). There was no major spatial difference in MND distribution. Our data were in agreement with the French death database (standardised mortality ratio = 1.01, 95% CI = 0.96–1.06) and Limousin ALS registry (standardised incidence ratio = 0.92, 95% CI = 0.72–1.15). Incidence estimates were in the same range as those from previous studies. Conclusions: We report French nationwide incidence estimates of MND. Administrative databases including hospital discharge data and riluzole claims offer an interesting approach to identify large population-based samples of patients with MND for epidemiologic studies and surveillance.

Objective: There are no estimates of the nationwide incidence of motor neuron disease (MND) in France. We used the French health insurance information system to identify incident MND cases (2012–2014), and compared incidence figures to those from three external sources. Methods: We identified incident MND cases (2012–2014) based on three data sources (riluzole claims, hospitalisation records, long-term chronic disease benefits), and computed MND incidence by age, gender, and geographic region. We used French mortality statistics, Limousin ALS registry data, and previous European studies based on administrative databases to perform external comparisons. Results: We identified 6553 MND incident cases. After standardisation to the United States 2010 population, the age/gender-standardised incidence was 2.72/100,000 person-years (males, 3.37; females, 2.17; male:female ratio = 1.53, 95% CI1.46–1.61). There was no major spatial difference in MND distribution. Our data were in agreement with the French death database (standardised mortality ratio = 1.01, 95% CI = 0.96–1.06) and Limousin ALS registry (standardised incidence ratio = 0.92, 95% CI = 0.72–1.15). Incidence estimates were in the same range as those from previous studies. Conclusions: We report French nationwide incidence estimates of MND. Administrative databases including hospital discharge data and riluzole claims offer an interesting approach to identify large population-based samples of patients with MND for epidemiologic studies and surveillance.