Implant Failure Prediction Using Discriminant Analysis.

Abstract

Electronic dental records (EDR) provide access to a vast repository of clinical information which may be used for analyzing dental care delivery. The goal of this study was identification of determinants of implant survival and development of implant failure prediction model using large data set of intact and failed implant parameters extracted from EDR. A retrospective analysis of 19 variables reflecting patient, surgeon and dental treatment characteristics of 800 dental implants was performed using discriminant analysis to develop a predictive model identifying potential implant failure based on characteristics routinely available in a clinical care setting. The intact and failed implant characteristics were compared using the Goodman and Kruskal's lambda test, the point-biserial test, the chi-square test, and ANOVA test. A stepwise discriminant analysis reduced model dimensionality from 19 to 4 features. The final discriminant analysis model included the following variables: non-temporary implant, pre-op antibiotics, immunocompromised status, and gender. Overall, 72% of implant failure cases and 62% of intact implants were correctly identified by the resulting discriminant function. As the final predictive feature set is readily available in EDR, the resulting algorithm may be implemented as a clinical decision support module embedded into EDR to promote personalized approach in dental implant patients.