Introduction:
Unlike statistical inference, the purpose of Machine Learning (ML) is to obtain a model that can make repeatable predictions without prior assumptions about the relationships among variables.

Methods:
ML techniques were applied to the Full Data Set (FDS) of the European Injury DataBase (EU-IDB) which provides information on the external causes and diagnoses of injury observed at the Emergency Departments (ED). The IDB-FDS provides more than 3.800.000 ED records, for the period 2008-19 in 19 Countries. LASSO (Least Absolute Shrinkage and Selection Operator) cross-validated linearized regression technique was used for variable selection and parameter regularization. Inpatients were considered those admitted, transferred to other hospitals or deceased during hospitalization. Cross-validation was performed randomly assigning the records on 5 folds. A cross-validated logistic model was performed on 5 folds which were randomly sampled assigning 80% of records to training and 20% to testing samples.

Results:
The strongest predictors of hospital admission risk selected by the model were in order of importance: EUROCOST-39 diagnoses categories, AgeGroup, Intent, MechanismOfInjury, ActivityWhenInjured, TransportInjuryEvent, SexOfPatient, PlaceOf Occurrence. EUROCOST-39 categories represent 61,9% of explained variability and Age Groups 19,4%. In applying a cross-validated logistic regression with these independent variables we obtain an average root mean square error of 0,338662 ranging from 0,3384615 in fold 3 up to 0,3392664 in fold 4. The estimated Odds Ratios of admission risk for instance in the median sample (fold 1) are: MechanismOfInjury=1177,10 (95%CI: 1090,55-1270,52); EUROCOST-39=401,36 (95%CI: 390,82-412,18); Intent=32,23 (95%CI: 31,17-33,32); PlaceOfOccurrence=8,99 (95%CI: 8,52-9,47); ActivityWhenInjured=4,62 (95%CI: 4,47-4,76); AgeGroup=2,19 (95%CI: 2,17-2,23); TransportInjuryEvent=1,90 (95%CI: 1,88-1,91); SexOfPatient=0,38 (95%CI: 0,37-0,39).

Discussion:
The ML model explains a significant part of the hospitalization risk variability and this measure is stable in the different training and testing samples used to cross validate the estimates. The most of variability is explained by the diagnoses reclassified according to a disability standardization method. For instance, in the maximum fold sample risk of hospitalization ranges from odd 0,76% for hand/fingers sprain up to 154,02% for brain concussion. The respective figures in the minimum fold are odds 0,55% and 149,90%. Conclusion. LASSO technique has proven useful to enhance the prediction accuracy of hospital admission risk. A combination of more disabling injury, older age, self-harm intent, exposure to chemicals or threat to breathing increases enormously the risk of hospitalization. The EU-IDB databank can provide estimates of predictors of hospital admission risk for targeting preventive measures and organizing health care.

Keywords: Machine Learning, Injury, Prediction models

B2C Innovation - Milan - Italy - ItalyGianni Fondi, Carlo Mamo, Marco Giustini, IDB-FDS Reference Group