1. Abstract Tooth loss data from 100 periodontal patients (2509 teeth) under maintenance care for up to 15 years were analyzed to develop a system for determination of periodontal prognostic indicators based on actual tooth survival and commonly used periodontal clinical parameters. The goal was to identify the proper classification of each tooth into one of the five commonly used categories of periodontal prognosis: good, fair, poor, questionable, and hopeless). The tree-based method of survival analysis was employed. In order to account for correlation among teeth within each patient’s mouth, the robust log rank statistic was used as the splitting statistic. A simulation study showed that the robust log rank statistic split censored correlated survival data adequately. The method, analogous to CART, included algorithms for growing a tree, pruning the tree and selecting the final tree. Three approaches were used to select the best-sized tree: test sample, cross validation, and bootstrap. An amalgamation procedure was also conducted to form the five prognostic indicator groups. The best fitting tree using the test-sample approach included probing depth, furcation involvement, crown-to-root ratio, hygiene level, fixed abutment status, biteguard use, and percent of bone loss as significant factors in assignment of periodontal prognosis. No tree was found to fit better than other trees using the cross-validation approach, as indicated by a clear pattern of the splitting statistic first increasing and then decreasing upon pruning. The best tree obtained using the bootstrap method included age, probing depth, furcation involvement, crown-to-root ratio, mobility, fixed abutment status, and percent bone loss. Based on tree-based multivariate survival analysis applied to these data, probing depth, furcation involvement, crown-to-root ratio, fixed abutment status, and percent bone loss were consistently included in the optimal tree-based survival model, regardless of the method of pruning. Introduction Previous studies have investigated the relationship of commonly-used clinical parameters to progression of periodontal disease and tooth loss from periodontal disease. However, these studies failed to clearly delineate objective criteria for assignment of prognostic indicators based on actual tooth loss from periodontal disease because of limitations of existing statistical techniques. In order to develop a system of criteria of assignment of prognostic indicators based on tooth loss, the method of Classification And Regression Trees (CART) was extended to the multivariate case so that this methodology could be applied to tooth loss data. In the current study, we apply this new methodology to long-term tooth loss data from 100 well-maintained periodontal patients with moderate-to-severe periodontal disease. The subjects in this study had all completed active treatment and were in maintenance for at least 5 years following active treatment with follow-up for up to 15 years. Study Population The data for this study were obtained from the clinical records of a private periodontal practice. 100 consecutive patients with at least 5 years of maintenance care were selected from the practitioner’s appointment book over a 2-month period. All had been diagnosed initially as having chronic generalized moderate to severe adult periodontitis. Clinical Parameters Tooth-Level Parameters: 1.Prognosis (Good, Fair, Poor, Questionable, Hopeless) 2.Probing Depth (deepest site of each tooth) 3.Mobility (Miller 1,2,3) 4.Furcation involvement (Glickman class I, II, III) 5.Bone Loss Type (horizontal or vertical) 6.Percentage Bone Loss 7.Crown-to-Root Ratio (CR ratio) 8.Root Form 9.Root Proximity 10.Pulpal Involvement 11.Caries 12.Abutment Status 13.Tooth Position Statistical Methods The method of CART for survival that was originally proposed by LeBlanc and Crowley was extended to accommodate multiple failures (tooth loss) within each subject. “Best split” algorithm was used to grow the tree where the robust log-rank test was maximized. Pruning was accomplished by maximizing the complexity factor, as defined by LeBlanc and Crowley. Three methods for choosing the best-sized subtree were utilized: 1. Test Sample – Used with large samples; sample is divided into 2 parts: a) A learning sample L 1 b) A test sample L 2 2. Cross-Validation – Used with small to moderate samples; divide sample L into V equal size samples, L v, v=1,…,V 3. Bootstrap – Draw B bootstrap samples, L b, b=1,…,B from the entire sample L Determination of periodontal prognostic indicators based on multivariate survival trees Nunn ME 1, Fan J 2, McGuire MK 3. 1 Boston University, 2 University of California, Davis 3 Private Practice, Houston, TX Conclusions  Model selection procedures resulted in some variation in classification.  Common clinical parameters were found to be significant in all modeling. In particular, probing depth, furcation involvement, CR ratio, %bone loss, and fixed abutment status were factors common to all methods of model selection.  Further work should be conducted on a large database. No. 2756 Clinical Parameters Patient-Level Parameters: 1.Age 2.Gender 3.Smoking Status 4.Negative Health History 5.Family History of Periodontal Disease 6.Oral Hygiene (Good, Fair, Poor) 7.Parafunctional Habit Sample Characteristics Gender – 65 Females and 35 Males Smoking Status – 60 Nonsmokers and 40 Smokers Age – 44.8  9.44 years (Range: 21 to 64 years of age) Number of Teeth – 2,509 Teeth at Baseline Teeth Lost from Periodontal Disease – 131 Teeth Results Good ■ No Furcation Involvement ■ Not a Fixed Abutment ■ % Bone Loss  37.5% or ■ Furcation Involvement of 1 or 2 ■ Probing Depth  4 Fair ■ No Furcation Involvement ■ Fixed Abutment or ■ No Furcation Involvement ■ Not a Fixed Abutment ■ % Bone Loss > 37.5% ■ Good Oral Hygiene Test Sample Method: Poor ■ No Furcation Involvement ■ Not a Fixed Abutment ■ % Bone Loss > 37.5% ■ Fair-Poor Oral Hygiene ■ Probing Depth  5.5 or ■ Furcation Involvement of 1 or 2 ■ Probing Depth > 4 ■ Biteguard Questionable ■ Furcation Involvement of 1 or 2 ■ Probing Depth > 4 ■ No Biteguard ■ Satisfactory CR Ratio Hopeless ■ No Furcation Involvement ■ Not a Fixed Abutment ■ % Bone Loss > 37.5% ■ Fair-Poor Oral Hygiene ■ Probing Depth > 5.5 or ■ Furcation Involvement of 1 or 2 ■ Probing Depth > 4 ■ No Biteguard ■ Unsatisfactory CR Ratio or ■ Furcation Involvement of 3 Cross-Validation Method: No optimal classification found. Bootstrap Method: Good  No Furcation Involvement  Probing Depth  4.5  Age  39.5 or  No Furcation Involvement  Probing Depth  4.5  Age > 39.5  Fixed Abutment or  Furcation Involvement of 1  Probing Depth  6.5 Fair  No Furcation Involvement  Probing Depth  4.5  Age > 39.5  Not a Fixed Abutment  % Bone Loss  37.5 Poor  No Furcation Involvement  Probing Depth  4.5  Age > 39.5  Not a Fixed Abutment  % Bone Loss > 37.5 or  Furcation Involvement of 1  Probing Depth of 7 or  No Furcation Involvement  Probing Depth > 4.5  Mobility of 0 or 1 or  Furcation Involvement of 2 or 3  Satisfactory CR Ratio  Probing Depth  7.5 Questionable  No Furcation Involvement  Probing Depth > 4.5  Mobility of 2 or 3 or  Furcation Involvement of 1  Probing Depth > 7.5 or  Furcation Involvement of 2 or 3  Satisfactory CR Ratio  Probing Depth > 7.5 Hopeless  Furcation Involvement of 2 or 3  Unsatisfactory CR Ratio