Background: Falls in the elderly are the number one cause of traumatic death in this population. Prevention of falls requires to evaluate which risk factors for fall are present for a person on the basis of available health information. Our objective is to predict the presence or the absence of 12 risk factors for fall in elderly people based on partial observations. Methods: A data set of 1810 patients of the multidisciplinary falls consultation of Lille University Hospital covering fourteen years admissions were used to learn and evaluate a Bayesian network and four usual machine learning classifiers. Variable selection and data pre-processing were achieved on the basis of an ontology and interviews of the experts. The prediction of each target risk factor using the complete set of observations is first compared with the prediction based on a specific subset of variables, and second based on partial observation, from 10 to 90% of the variables. Results: For 7 out of 12 target risk factors, the f1-score of classifiers using complete set of variables is slightly better than the specific subset of variables, with a difference of less than 3%. Bayesian Networks and other classiers perform equivalently in terms of accuracy and f1-score. The best prediction were obtained for the loss of autonomy and osteoporosis with a f1-score from 15 to 20% better than the baseline classifier when using the Bayesian network. At the opposite, for 3 risk factors, no classifier allows to improve the f1-score or the accuracy of more than 1% compared to the baseline classifier. Conclusion: Our results show that the use of specific subsets of variables does not improve the prediction of risk factors, and that no classifier outperform the others. However Bayesian networks perform well and are interesting due to their explainability.