IADR Abstract Archives
Artificial Neural Networks Can Reliably Predict the Presence of Periodontitis
Objectives: This study aimed to train machine learning classifiers on population-based data from the 2009 – 2014 National Health and Nutritional Examination Surveys (NHANES), in order to develop and validate predictive models for periodontitis.
Methods: Adult participants who had received a periodontal examination, with ≥ 6 teeth examined, and had data collected on various risk factors (socio-demographic, smoking status, biochemical data) qualified for inclusion for model development and validation. Two types of machine learning classifiers were fit: (i) multivariate logistic regression, and (ii) artificial neural networks (ANNs). For each classifier, two types of models were developed: (i) utilising only patient-reported variables, and (ii) incorporating biochemical data. The predictive models were validated using bootstrap sampling, with replacement (n = 1,000). Four outcome measures were used to assess predictive validity: area under the curve for the receiver operator characteristic (AUROC), accuracy, sensitivity, and specificity.
Results: 4,373 participants qualified for inclusion in this study. For both logistic regression and ANNs, models with and without biochemical data were comparable across all four of the outcome measures assessed.
Model 1. Logistic regression, without biochemical data (95% CI); AUROC: 0.800 (0.749 – 0.852), accuracy: 71.08% (69.42 – 72.74), sensitivity: 66.74% (64.73 – 68.76), specificity: 75.38% (72.83 – 77.93).
Model 2. Logistic regression, with biochemical data (95% CI); AUROC: 0.779 (0.730 – 0.827), accuracy: 69.55% (67.98 – 71.10), sensitivity: 67.93% (65.95 – 69.92), specificity: 71.13% (69.02 – 73.24).
Model 3. Artificial neural network, without biochemical data (95% CI); AUROC: 0.801 (0.746 – 0.857), accuracy: 72.70% (70.93 – 74.48), sensitivity: 68.43% (66.29 – 70.56), specificity: 76.99% (74.76 – 79.21).
Model 4. Artificial neural network, with biochemical data (95% CI); AUROC: 0.783 (0.732 – 0.832), accuracy: 69.94% (68.50 – 71.42), sensitivity: 67.22% (65.06 – 69.38), specificity: 72.64% (70.42 – 74.85).
Conclusions: Using variables which may be obtained in a non-clinical setting, ANNs offer the best performance for predicting presence of periodontitis. Additional biochemical data does not improve predictive validity.
Methods: Adult participants who had received a periodontal examination, with ≥ 6 teeth examined, and had data collected on various risk factors (socio-demographic, smoking status, biochemical data) qualified for inclusion for model development and validation. Two types of machine learning classifiers were fit: (i) multivariate logistic regression, and (ii) artificial neural networks (ANNs). For each classifier, two types of models were developed: (i) utilising only patient-reported variables, and (ii) incorporating biochemical data. The predictive models were validated using bootstrap sampling, with replacement (n = 1,000). Four outcome measures were used to assess predictive validity: area under the curve for the receiver operator characteristic (AUROC), accuracy, sensitivity, and specificity.
Results: 4,373 participants qualified for inclusion in this study. For both logistic regression and ANNs, models with and without biochemical data were comparable across all four of the outcome measures assessed.
Model 1. Logistic regression, without biochemical data (95% CI); AUROC: 0.800 (0.749 – 0.852), accuracy: 71.08% (69.42 – 72.74), sensitivity: 66.74% (64.73 – 68.76), specificity: 75.38% (72.83 – 77.93).
Model 2. Logistic regression, with biochemical data (95% CI); AUROC: 0.779 (0.730 – 0.827), accuracy: 69.55% (67.98 – 71.10), sensitivity: 67.93% (65.95 – 69.92), specificity: 71.13% (69.02 – 73.24).
Model 3. Artificial neural network, without biochemical data (95% CI); AUROC: 0.801 (0.746 – 0.857), accuracy: 72.70% (70.93 – 74.48), sensitivity: 68.43% (66.29 – 70.56), specificity: 76.99% (74.76 – 79.21).
Model 4. Artificial neural network, with biochemical data (95% CI); AUROC: 0.783 (0.732 – 0.832), accuracy: 69.94% (68.50 – 71.42), sensitivity: 67.22% (65.06 – 69.38), specificity: 72.64% (70.42 – 74.85).
Conclusions: Using variables which may be obtained in a non-clinical setting, ANNs offer the best performance for predicting presence of periodontitis. Additional biochemical data does not improve predictive validity.