EARLY CHILDHOOD CARIES (ECC) PREDICTION MODELS USING MACHINE LEARNING

Abstract

BACKGROUND: TO EVALUATE THE PERFORMANCE OF DIFFERENT PREDICTION MODELS BASED ON MACHINE LEARNING TO PREDICT THE PRESENCE OF EARLY CHILDHOOD CARIES. MATERIAL AND METHODS: CROSS-SECTIONAL ANALYTICAL STUDY. THE SOCIODEMOGRAPHIC AND CLINICAL DATA USED CAME FROM A SAMPLE OF 186 CHILDREN AGED 3 TO 6 YEARS AND THEIR RESPECTIVE PARENTS OR GUARDIANS TREATED AT A HOSPITAL IN ICA, PERU. THE DATABASE WITH SIGNIFICANT VARIABLES WAS LOADED INTO THE ORANGE DATA MINING SOFTWARE TO BE PROCESSED WITH DIFFERENT PREDICTION MODELS BASED ON MACHINE LEARNING. TO EVALUATE THE PERFORMANCE OF THE PREDICTION MODELS, THE FOLLOWING INDICATORS WERE USED: PRECISION, RECALL, F1-SCORE AND ACCURACY. THE DISCRIMINATORY POWER OF THE MODEL WAS DETERMINED BY THE VALUE OF THE ROC CURVE. RESULTS: 76.88% OF THE CHILDREN EVALUATED HAD CAVITIES. THE SUPPORT VECTOR MACHINE (SVM) AND NEURAL NETWORK (NN) MODELS OBTAINED THE BEST PERFORMANCE VALUES, SHOWING SIMILAR VALUES OF ACCURACY, F1-SCORE AND RECALL (0.927, 0.950 AND 0.974; RESPECTIVELY). THE PROBABILITY OF CORRECTLY DISTINGUISHING A CHILD WITH ECC WAS 90.40% FOR THE SVM MODEL AND 86.68% FOR THE NN MODEL. CONCLUSIONS: THE MACHINE LEARNING-BASED CARIES PREDICTION MODELS WITH THE BEST PERFORMANCE WERE SUPPORT VECTOR MACHINE (SVM) AND NEURAL NETWORKS (NN). © MEDICINA ORAL S. L. C.I.F. B 96689336 - EISSN: 1989–5488