IADR Abstract Archives
Predicting Tooth Loss and Periodontitis Progression with Pre-identified Risk Factors
Objectives: Several works have studied risk factors for tooth loss (TL) and periodontitis progression (PP) in patients undergoing periodontal maintenance therapy. However, the effectiveness of risk factors for predicting TL and PP after the end of therapy (T2) remains unknown. TL is periodontitis-related tooth loss for both patient and tooth levels. PP is a ≥ 3 mm CAL loss at ≥ 2 interproximal sites for the patient level and a ≥ 2 mm increase in PPD for tooth level from the end of therapy (T1) to T2. This study analyses the effectiveness of risk factors identified by Siow et al., 2023 in various predictive machine learning models.
Methods: This study uses 108 patients' periodontal records collected at T1 and T2. For TL, patient-level risk factors are smoking, diabetes, stage, maintenance freq., and the num. of sites with PPD ≥ 5 mm, whereas tooth-level risk factors are residual PPD ≥ 7 mm, Grade 1–2 furcation involvement, Grade 1–3 mobility and max. CAL ≥ 6 mm. For PP, patient-level risk factors are gender, FMBS, compliance, maintenance freq., and the num. of sites with PPD ≥ 5 mm, whereas tooth-level risk factors are residual PPD 5–6 mm, Grade 1–3 furcation involvement, and max. CAL = 6 mm. This study considers five predictive models, namely logistic regression, support vector machine, decision tree, random forest, and neural network, that take the risk factors at T1 as input features and generate binary predictions for TL and PP at T2.
Results: Table 1 shows the macro F1-score of predictive models with risk factors at T1 for predicting TL and PP at T2. Macro F1-score is a suitable evaluation metric for this study due to its sensitivity to a class imbalance problem which exists at the tooth level dataset. The best macro F1-score are 0.76 and 0.60 for patient-level and tooth-level TL, 0.55 and 0.5 for patient-level PP, which are generally low compared to the desired macro F1-score of ≥ 0.90.
Conclusions: This study found that the efficacy of predictive models with pre-identified risk factors is substandard. Despite this undesired performance, the results hold promise for predicting tooth loss and periodontitis progression using predictive models. By considering the contextual nuances and model refinements, the performance and clinical utility of these predictive models could be further improved.
Methods: This study uses 108 patients' periodontal records collected at T1 and T2. For TL, patient-level risk factors are smoking, diabetes, stage, maintenance freq., and the num. of sites with PPD ≥ 5 mm, whereas tooth-level risk factors are residual PPD ≥ 7 mm, Grade 1–2 furcation involvement, Grade 1–3 mobility and max. CAL ≥ 6 mm. For PP, patient-level risk factors are gender, FMBS, compliance, maintenance freq., and the num. of sites with PPD ≥ 5 mm, whereas tooth-level risk factors are residual PPD 5–6 mm, Grade 1–3 furcation involvement, and max. CAL = 6 mm. This study considers five predictive models, namely logistic regression, support vector machine, decision tree, random forest, and neural network, that take the risk factors at T1 as input features and generate binary predictions for TL and PP at T2.
Results: Table 1 shows the macro F1-score of predictive models with risk factors at T1 for predicting TL and PP at T2. Macro F1-score is a suitable evaluation metric for this study due to its sensitivity to a class imbalance problem which exists at the tooth level dataset. The best macro F1-score are 0.76 and 0.60 for patient-level and tooth-level TL, 0.55 and 0.5 for patient-level PP, which are generally low compared to the desired macro F1-score of ≥ 0.90.
Conclusions: This study found that the efficacy of predictive models with pre-identified risk factors is substandard. Despite this undesired performance, the results hold promise for predicting tooth loss and periodontitis progression using predictive models. By considering the contextual nuances and model refinements, the performance and clinical utility of these predictive models could be further improved.