Usefulness of Bayesian modeling in risk analysis and prevention of Home Leisure and Sport Injuries (HLIs)

Usefulness of Bayesian modeling in risk analysis and prevention of Home Leisure and Sport Injuries (HLIs) Madelyn Rojas Castro, Marina Travanca, Marta Avalos, David Valentin Conesa, Emmanuel Lagarde To cite this version: Madelyn Rojas Castro, Marina Travanca, Marta Avalos, David Valentin Conesa, Emmanuel Lagarde. Usefulness of Bayesian modeling in risk analysis and prevention of Home Leisure and Sport Injuries (HLIs). III Jornadas Científicas de Estudiantes de la Sociedad Española de Biometría, 2018, Bilbao, Spain. <hal-01964484> HAL Id: hal-01964484 https://hal.archives-ouvertes.fr/hal-01964484 Submitted on 22 Dec 2018 HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Usefulness of Bayesian modeling in risk analysis and prevention of Home Leisure and Sport Injuries (HLIs) Madelyn Rojas Castro 1, Marina Travanca 1, Marta Avalos 1, David Conesa 2, Emmanuel Lagarde 1 1 Bordeaux Population Health Research Center INSERM U1219, Injury epidemiology, transport, occupation (IETO). 2 Departamento de Investigación Operativa, Universidad de Valencia 18 January 2018 Bayesian Modeling of HLIs 1 / 27

Home Leisure and Sport Injuries (HLIs) Bayesian Modeling of HLIs 2 / 27

Injuries Epidemiological Context World Health Organization (WHO) Injuries are the 4th cause of mortality in the EU. 230 thousand annual deaths (EU), 58 % are HLIs. In France : HLIs are the 3th cause of mortality. Leading cause of childhood mortality. Each year: 20 thousand deaths, 5 times more than traffic accidents. 5 million emergencies. 11 million injuries Bayesian Modeling of HLIs 3 / 27

L Observatoire MAVIE Prospective online cohort study of HLIs Currently, MAVIE has more than 26 thousand volunteers in France during 3 years of recruitment (target sample 100 thousand volunteers). Objectives Identify the risk factors associated with the HLIs occurrence and severity. Implement prevention measures to reduce the number of victims. Bayesian Modeling of HLIs 4 / 27

MAVIE-Lab (mhealth) Mobile app including a DSS (Decision Support System) To self-management of HLIs risks (Evaluation). To experience personalized prevention solutions to reduce the risk of injury (Mitigation). Bayesian Modeling of HLIs 5 / 27

MAVIE-Lab Development Figure: Steps for MAVIE-Lab Development Bayesian Modeling of HLIs 6 / 27

Modeling Problems MAVIE data 1 Reduced number of injuries declared by each activity. 2 Missing values. 3 Under-representation between injuries reported and those occurred. 4 Complex relationships between risk factors. Bayesian Modeling of HLIs 7 / 27

Bayesian Generalized Lineal Models (Logistic Regression) Bayesian Generalized Lineal Models (Logistic Regressión) p(θ X, y) p(θ) l(β, σ X, y) posterior prior likelihood Bayesian Modeling of HLIs 8 / 27

Bayesian Generalized Lineal Models (Logistic Regression) Bayesian Modeling of HLIs 9 / 27

Bayesian Generalized Lineal Models (Logistic Regression) Methodological Proposal Priors experts elicitation Reason for using Bayesian experts elicitation 1 The knowledge is limited or incomplete. 2 The evidence is inconsistent missing or ambiguous. 3 The questions are complex and the relations between variables. 4 To deal with bias and uncertainties. 5 To integrate different sources of knowledge. (A. B. Knol et al. 2010) Bayesian Modeling of HLIs 10 / 27

Bayesian Generalized Lineal Models (Logistic Regression) Bayesian experts elicitation (MAVIE-Lab) Bayesian Modeling of HLIs 11 / 27

Bayesian Generalized Lineal Models (Logistic Regression) Exploratory analysis objectives Principal Objective To explore Bayesian modeling methodologies for being used in MAVIE-Lab development To explore experts elicitation to improve the estimation of model parameters. To explore the use of automatic selection models methods since the Bayesian approach. Bayesian Modeling of HLIs 12 / 27

Bayesian Generalized Lineal Models (Logistic Regression) Model Selection BMA Bayesian Model Averaging The BMA is a weighted average of the posterior distributions for each parameter for all possible models. Figure taken from (FitzGerald et al. 2014). Bayesian Modeling of HLIs 13 / 27

Database Work Database Sample of MAVIE Cohort: N = 4,345 (March 2017). Volunteers over 18 years old, who had completed and validated the questionnaire. Injuries: 603 Reported Injuries (13.87 %). Variables: Explanatory variables: 20 categorical or categorized variables (Factors associated with HLIs occurrence). Demographic: Age and sex. Previous Injuries. Physical and mental health: BMI, Health problems, depression, anxiety, hyperactivity, drowsiness and concentration. Consumption: Medicines, alcohol, tobacco and cannabis. Sport Practice: Sports, use of compression and maintenance accessories. Bayesian Modeling of HLIs 14 / 27

Database Methodology MAVIE Data Multiple Imputation BY Logistic Model (Non informative priors) Model Selection (BMA Bayesian Model Averaging) Elicitation Prior Distributions BY Logistic Model (Informative priors) Bayesian Modeling of HLIs 15 / 27

1.0 0.5 0.0 0.5 1.0 Introduction Methodology Results Discussion References Results BY Logistic Model (non informative priors) beta[22] beta[6] beta[8] beta[9] beta[10] beta[21] beta[20] beta[18] beta[23] 1.00 1.00 beta[11] 0.75 0.75 Parameter beta[17] beta[4] beta[24] beta[14] beta[15] beta[16] beta[7] beta[12] beta[2] beta[3] beta[5] beta[25] Sensitivity 0.50 beta[1] 0.25 Chain 1 2 3 Predicted 0.50 0.25 beta[19] beta[13] 0.00 0.00 HPD 0.00 0.25 0.50 0.75 1.00 1 Specificity (a) CI 95% (b) ROC (AUC=0,58) (c) Separation Graphic Bayesian Modeling of HLIs 16 / 27

Model Selection BMA Model Selection BMA: Better models according to BIC criteria Models selected by BMA IND_NAIS2 IND_NAIS3 IND_NAIS4 IND_NAIS5 IND_SEXE1 IMC1 pb_sante1 etat_sante_physique1 etat_sante_mental1 depression1 anxiete1 hyperactif1 medic_psy1 tabac2 tabac3 cannabis2 cannabis3 alcool1 freq_6verres_alcool1 hospit_acc1 RAD_PRATIQUE_SPORTIVE1 LST_MAINTIEN1 LST_COMPRESSION1 MWscore1 somnolence1 1 2 3 4 5 6 7 8 9 11 Model # Selected Variables: Sex, Health Prob., Previous HLIs, Sports. Bayesian Modeling of HLIs 17 / 27

0.25 0.00 0.25 0.50 0.75 1.00 Introduction Methodology Results Discussion References Model BMA Variables Model BMA Variables: BY Logistic Model (Non informative priors) beta[14] 1.00 1.00 beta[6] beta[7] beta[8] beta[9] 0.75 0.75 beta[10] beta[13] beta[11] Chain Parameter beta[12] Sensitivity 0.50 1 2 3 Predicted 0.50 beta[4] beta[15] beta[1] 0.25 0.25 beta[2] beta[3] beta[5] 0.00 0.00 HPD 0.00 0.25 0.50 0.75 1.00 1 Specificity (d) CI 95% (e) ROC (AUC=0.64) (f) Separation Graphic Bayesian Modeling of HLIs 18 / 27

Informative Prior Distributions Informative Prior Distributions: Elicited values prior distributions models. Reference Study (Injuries in France) (Lefèvre & Mhiri 2015). Variables OR Mean β Variance β Sex (F) 0.44-0.36 0.04 Age 50-60 1.20 0.08 0.05 Age 40-50 1.40 0.15 0.04 Age 30-40 1.40 0.15 0.05 Age 18-30 2.10 0.32 0.04 Health Prob. 1.74 0.24 0.07 Smoker 1.64 0.21 0.03 Ex-smoker 1.34 0.13 0.03 Alcohol 5.72 0.76 0.28 Bayesian Modeling of HLIs 19 / 27

0.0 0.5 1.0 Introduction Methodology Results Discussion References Informative Prior Distributions Informative Prior Distributions beta[14] 1.00 1.00 beta[6] beta[7] beta[8] beta[9] 0.75 0.75 beta[10] beta[13] beta[12] Chain Parameter beta[4] Sensitivity 0.50 1 2 3 Predicted 0.50 beta[11] beta[15] beta[1] 0.25 0.25 beta[2] beta[3] beta[5] 0.00 0.00 HPD 0.00 0.25 0.50 0.75 1.00 1 Specificity (g) CI 95% (h) ROC (AUC=0.6) (i) Separation Graphic Bayesian Modeling of HLIs 20 / 27

Informative Prior Distributions Results BY model: (BMA, Prior Information) OR Devest 2.5% 97.5% Intercept - - - - Age (50-60) 0.95 0.035 0.883 1.020 Age (40-50) 0.93 0.037 0.855 1.002 Age (30-40) 0.89 0.043 0.812 1.973 Age (18-30) 1.29 0.076 1.154 1.443 Sex (F) 0.80 0.024 0.760 0.851 Sport 1.58 0.065 1.467 1.719 Health Prob. 1.57 0.050 1.476 1.668 Smokers 1.54 0.074 1.397 1.688 Ex-smokers 1.44 0.060 1.331 1.567 Physic Health 1.35 0.052 1.252 1.453 Psc.Medic. 1.22 0.038 1.146 1.294 Mantenim. Acc. 1.29 0.072 1.150 1.434 Comp. Acc. 1.31 0.065 1.188 1.440 Previous HLIs 2.41 0.095 2.236 2.605 Alcohol 1.18 0.037 1.112 1.254 Bayesian Modeling of HLIs 21 / 27

Discussion The model does not allow separating profiles of greater or lesser injury risk. The variables included do not explain the occurrence of injuries. Bayesian approach remains appropriate for the development of MAVIE-Lab. Bayesian Modeling of HLIs 22 / 27

Solutions and ongoing work 1 To make more specific models by type of injury, including the most important variables in each case (Example: Sport Injuries PhD Project). 2 To perform a formal elicitation (Devilee & A. Knol 2011): Experts selection Uncertain evaluation Elicitation protocol 3 To perform Bayesian Network models including besides the relationships between variables (probabilistic and graphical modeling) (Mujalli et al. 2016). Bayesian Modeling of HLIs 23 / 27

Variables relations in Running Injuries (DAG) Bayesian Modeling of HLIs 24 / 27

Conclusion In conclusion, Bayesian statistic and the expert s elicitation are powerful tools for the construction of expert system to be included in mhealth. This methodology makes possible to combine statistical data and experts information as for example medical advice. Figure: mhealth (Figure taken from web-site UNC Gillinds School of Global Public Health) Bayesian Modeling of HLIs 25 / 27

Bibliography Devilee, J. & Knol, A. Software to support expert elicitation An exploratory study of existing software Software to support expert elicitation. Tech. rep. (2011), 1 100. FitzGerald, T. H. B., Dolan, R. J. & Friston, K. J. Model averaging, optimal inference, and habit formation. Frontiers in Human Neuroscience 8, 457. ISSN: 1662-5161 (June 2014). Knol, A. B., Slottje, P., Van Der Sluijs, J. P. & Lebret, E. The use of expert elicitation in environmental health impact assessment: a seven step procedure. Environmental Health 9. doi:10.1186/1476-069x-9-19. http://www.ehjournal.net/content/9/1/19 (2010). Lefèvre, B. & Mhiri, E. F. Facteurs sociodémographiques et pratiques associés aux accidents liés à la pratique physique et sportive. Science et Sports 30, 126 133 (2015). Mujalli, R. O., Lopez, G. & Garach, L. Bayes classifiers for imbalanced traffic accidents datasets. Accident Analysis and Prevention. doi:10.1016/j.aap.2015.12.003 (2016). Bayesian Modeling of HLIs 26 / 27

Usefulness of Bayesian modeling in risk analysis and prevention of Home Leisure and Sport Injuries (HLIs) Madelyn Rojas Castro 1, Marina Travanca 1, Marta Avalos 1, David Conesa 2, Emmanuel Lagarde 1 1 Bordeaux Population Health Research Center INSERM U1219, Injury epidemiology, transport, occupation (IETO). 2 Departamento de Investigación Operativa, Universidad de Valencia 18 January 2018 Bayesian Modeling of HLIs 27 / 27