Multilevel and mixture modeling of prognostic factors for tooth loss in patients under periodontal maintenance

Abstract

Background Prognosis is an essential element of medical and dental practice to provide a rational basis for treatment decisions. The judgement is based on a set of prognostic factors that can best predict the outcome of interest such as tooth loss in supportive periodontal care (SPC). Conventional statistical models to identify prognostic factors usually include the information only from the initial (baseline) examination and this yields poor prediction accuracy. Inclusion of updated information (such as information after active periodontal treatment, APT) can model the dynamic process of disease progression and help to improve accuracy of prognosis. This study aimed to identify important prognostic factors for tooth loss during SPC from updated information. The study also compared the predictive power of baseline information and updated information. Methods A retrospective cohort of 58 patients aged between 28 and 73 years old, who had been undergoing periodontal maintenance for 5.1 – 16.5 years in a university-based Periodontology Clinic, was included. Data on the patients’ demographic characteristics, treatment information, oral health behaviors, and clinical parameters per charted oral examination were retrieved. The data were then fitted to three multilevel models: (1) logistic mixed effect model; (2) shared frailty model; (3) time-varying shared frailty model; and one single-level model: (4) mixture cure model. The first model predicted the occurrence of tooth loss whereas the other three predicted the time to tooth loss due to periodontal reasons during SPC. Multicollinearity was checked before fitting. The multilevel assumption was tested after fitting for models (1) – (3). For models (1), (2) and (4), predictive strength of the models based on the updated information and the baseline information was compared by either AIC and BIC, or a heuristic inspection. For model (3), time-varying information of selected clinical parameters were incorporated and compared with model (2). Results For the multilevel models, after APT, a patient with teeth having a deeper pocket depth, a deeper gingival recession, a higher % bleeding sites and a more severe mobility was found to be significantly associated with the odds and the relative risk of having tooth loss during SPC. With the time-varying information, gingival recession and % bleeding sites became not informative for the relative risk of tooth loss whereas more lost teeth in APT became significantly associated. For the mixture cure model, an older patient with fewer teeth at baseline and teeth after APT having a deeper pocket, a deeper gingival recession, a higher % bleeding sites and a more severe mobility was likely to be uncured. For the comparison on predictive strength, all models with updated information were found to have a smaller AIC or a stronger association with the risk of having a tooth loss during SPC. Conclusions This study identifies the significant prognostic factors for tooth loss during SPC using updated information and illustrates the importance of updated information in improving the accuracy of prognosis. The use of updated information in prognosis can account for the treatment response of the patient and help periodontists to assign a better prognosis.