Résumé : The considerable overlap of distributions of values for different parameters between control and ovarian hyperstimulation syndrome (OHSS) populations makes any single variable inefficient for risk prediction. Combinations of variables were studied in a discriminant function in order to increase predictivity and decrease the false negative rate. Such analyses were performed on two groups of in-vitro fertilization (IVF) patients: all OHSS cases (n = 128) (group A) and only severe OHSS cases (n = 92) (group B). Progressive introduction and automated stepwise selection of variables were applied to both groups. The best prediction (78.5%) was obtained in group A under post-oocyte retrieval conditions using log oestradiol, slope of log oestradiol increment, human menopausal gonadotrophin (HMG) dosage, number of oocytes retrieved and ratio of luteinizing hormone/follicle stimulating hormone (LH/FSH), in the formula. The corresponding false negative rate was 18.1%. However, effective prevention of OHSS implies the ability to withhold human chorionic gonadotrophin injection. Therefore a formula for pre-oocyte retrieval conditions was established yielding a prediction rate of 76.1% with a false negative rate of 18.1%. To be validated, such formulae have to be applied to another population of IVF cases used as a 'testing-set'.