Identification of a Predictive Decision Model Using Different Data Mining Algorithms for Diagnosing Peri-implant Health and Disease: A Cross-Sectional Study


Cetiner D., İŞLER S. Ç., BAKIRARAR B., URAZ A.

INTERNATIONAL JOURNAL OF ORAL & MAXILLOFACIAL IMPLANTS, vol.36, no.5, pp.952-965, 2021 (SCI-Expanded) identifier identifier identifier

  • Publication Type: Article / Article
  • Volume: 36 Issue: 5
  • Publication Date: 2021
  • Doi Number: 10.11607/jomi.8965
  • Journal Name: INTERNATIONAL JOURNAL OF ORAL & MAXILLOFACIAL IMPLANTS
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, CINAHL, EMBASE, MEDLINE, DIALNET
  • Page Numbers: pp.952-965
  • Keywords: data mining, diagnosis, peri-implantitis, risk assessment, risk factors, RISK-FACTORS, CONSENSUS REPORT, DENTAL IMPLANTS, INDICATOR, THICKNESS
  • Gazi University Affiliated: Yes

Abstract

Purpose: The aim of this study was to determine a predictive decision model for peri-implant health and disease and to reveal the highest accuracy of prediction using three different data mining methods. Materials and Methods: This cross-sectional study included a total of 216 patients with 542 dental implants from the Periodontology Department of Gazi University. The implants were classified into peri-implant health, peri-implant mucositis, and peri-implantitis groups based on established clinical and radiographic assessments. Prediction models were created using clinical variables in combination with possible risk factors for peri-implant diseases. Different data mining methods (decision-tree [DT]; J48), logistic regression, and artificial neural network (multilayer perceptron [MLP]) were compared to yield a better predictive decision model based on predictor variables with the highest potential of effect. Results: The prevalence of peri-implant mucositis and peri-implantitis among the participants of the specialist referral periodontal practice of the university was 36.1% (95% CI: 29.7 to 42.5) and 34.7% (95% CI: 28.4 to 41.0) at the patient level, respectively. The J48 method revealed a higher prediction of peri-implant health and disease with an accuracy of 0.871 compared with the logistic regression and MLP methods (0.832 and 0.852, respectively) for the present data set. In this specific patient population, the J48 model revealed the top-level node as "bleeding on probing (BOP)." "Maintenance therapy" and "medication use" were noted as powerful predictors in the next split-levels. Conclusion: The J48 model presented an acceptable predictive accuracy of peri-implant health and disease. The model revealed BOP as a major predictive clinical parameter when evaluated with possible risk factors for this patient population.