Multicenter prospective testing to predict malignancy in adnexal masses
using Bayesian network models
O Gevaert1, C Van Holsbeke2,3, R Fruscio4, S Guerriero5, A Czekierdowski6, L Savelli7, A
Testa8, D Fischerova9, D Jurkovic10, T Bourne2,11, P Neven2, L Valentin12, B De Moor1, D
Timmerman2
1 Dept Electrical Engineering (ESAT-SCD), Katholieke Universiteit Leuven, Leuven, Belgium; 2 University
Hospitals, Leuven, Belgium; 3 Ziekenhuis Oost-Limburg, Genk, Belgium; 4 San Gerardo Hospital, Monza, Italy;
5 Ospedale San Giovanni di Dio, Cagliari, Italy; 6 Medical University, Lublin, Poland; 7 Reproductive Medicine
Unit, Bologna, Italy; 8 Università Cattolica del Sacro Cuore, Rome, Italy; 9 General Faculty Hospital of Charles
University, Prague, Czech Republic; 10 King's College Hospital, London, UK; 11 St Georges Hospital Medical
School, London, UK; 12 University Hospital, Malmö, Sweden;
Objective
To investigate prospectively the performance of Bayesian networks built using data from the
International Ovarian Tumor Analysis (IOTA) study phase 1.
Methods
The first phase of IOTA resulted in a data set of 1066 patients from 9 centers in 5 countries.
The data were randomly stratified using 70% of patient data (i.e. the training set) to construct
a Bayesian network and 30% of the patient data to estimate the generalization performance
(i.e. the test set). After building a Bayesian network model (referred to as BN1) on the
training set, this resulted in an Area Under the ROC curve (AUC) of 0.942 (SE 0.017).
Subsequently, new data were gathered in IOTA phase 2. This data set was used to estimate
the prospective performance of the BN1 model.
Results
In total 1940 patients from 19 centers, including old and new centers from 8 countries, were
recruited. The BN1 model has an AUC of 0.940 (SE 0.006) on all IOTA phase 2 data. After
splitting the IOTA 2 data into data from the old centers and new centers, the AUC was 0.939
(SE 0.008) and 0.940 (SE 0.009) respectively.
Conclusions
The prospective performance of the BN1 model is similar to the initial performance on the
IOTA phase 1 data and thus BN1 is able to generalize to new data. Analyis of the
performance observed in old and new centers, showed almost no difference in the prospective
performance of new and old centers, indicating that BN1 can be used in other centers in
addition to the ones that were used to train the model.