Objectives: To cluster implants placed in patients of a private practice and identify possible implant "phenotypes" and predictors of individual implant mean bone levels (IIMBL).
Materials and methods: Clinical and radiographical variables were collected from 72 implant-treated patients with 237 implants and a mean 7.4 ± 3.5 years of function. We clustered implants using the k-means method guided by multidimensional unfolding. For predicting IIMBL, we used principal component analysis (PCA) as a variable reduction method for an ensemble selection (ES) and a support vector machines models (SVMs). Network analysis investigated variable interactions.
Results: We identified a cluster of implants susceptible to peri-implantitis (96% of the implants in the cluster were affected by peri-implantitis) and two overlapping clusters of implants resistant to peri-implantitis. The cluster susceptible to peri-implantitis showed a mean IIMBL of 5.2 mm and included implants placed mainly in the lower front jaw and in mouths having a mean of eight teeth. PCA extracted the parameters such as number of teeth, full-mouth plaque scores, implant surface, periodontitis severity, age and diabetes as significant in explaining the data variability. ES and SVMs showed good results in predicting IIMBL (root-mean-squared error of 0.133 and 0.149, 10-fold cross-validation error of 0.147 and 0.150, respectively). Network analysis revealed limited interdependencies of variables among peri-implantitis-affected and non-affected implants and supported the hypothesis of the existence of distinct implant "phenotypes."
Conclusion: Two implant "phenotypes" were identified, one with susceptibility and another with resistance to peri-implantitis. Prediction of IIMBL could be achieved by using six variables.
Keywords: alveolar bone; complex disease; computational biology; dental implant; network analysis; peri-implantitis; statistical learning theory.
© 2016 John Wiley & Sons A/S. Published by John Wiley & Sons Ltd.