Study sample

We used clinical records and insurance claims data from patients seen in HealthPartners clinics. HealthPartners provides care and insures more than 1 million members through its hospitals and medical and dental clinics in Minnesota and Wisconsin. We searched records from January 2004 through June 2017, inclusive, to identify patients who: 1) were > = 18 years of age; 2) had ≥1 periodontal disease diagnosis code and ≥1 diagnosis code(s) for coronary artery disease (CAD; ICD-9 codes 410–414; ICD-10 codes I20-I25), cerebrovascular disease (CBVD; ICD-9 codes: 433–436, 437.1, 437.4, 437.6, 437.9, 438; ICD-10 codes: I60-I69), or type 2 diabetes (T2D; ICD-9 codes: 250, 357.2, 362.0, 366.41; ICD-10 codes: E11); 3) had not opted out of research; and 4) had continuous (allowing gaps ≤30 days) medical, pharmacy, and dental insurance coverage for at least one year prior to, and one year beyond the date at which both a qualifying medical and periodontitis diagnoses were first observed. HealthPartners’ periodontal diagnosis codes map to the 1999 American Academy of Periodontology’s classifications [26]. We classified an individual based on their most severe periodontal disease diagnosis (i.e., severe over moderate over slight). Individuals were also categorized based on the extent of their most severe disease: generalized, affecting roughly ≥30% of the remaining teeth; or localized, affecting less than 30% of remaining teeth. Finally, the extent of gingival inflammation (i.e., the gingiva was visibly inflamed or bled when probed) was similarly scored as generalized or localized based on the above criteria.

Exposure and covariates

Periodontal treatment was ascertained based on American Dental Association’s Current Dental Treatment (CDT) codes, abstracted from dental records: D4341, D4342, D4355, D4210, D4211, D4240, D4241, D4245, D4260, D4261, D4263-7, D4264, D4274 and D4910. D4341, D4342 and D4355 denote non-surgical procedures (scaling or “deep cleaning”). D4910 denotes a maintenance procedure (cleaning) that typically follows treatment. The other codes denote surgical procedures. We categorized individuals as having received active therapy alone (non-surgical and/or surgical procedures), active therapy plus maintenance care, maintenance care only, or no care following the index date. Extractions were not considered as periodontal treatment, although tooth count (from dental records) was considered as a baseline covariate and updated with subsequent extractions.

We abstracted demographic information, laboratory results, vital signs, and smoking history from medical records. Smoking status [current (some days, every day, heavy smoker, light smoker); former (in the past but not currently); never (never)] was determined from self-reports at baseline (I.e., the index date). We computed the Charlson Co-Morbidity Index from ICD-9 or -10 codes present at the index date/baseline [27]. Dental insurance coverage was ascertained from health plan data.

Outcomes

We tracked clinical outcomes and medical care costs after an individual’s index date, which was the earliest date the individual had both a periodontitis and CAD, CBVD or T2D diagnosis. For the clinical outcomes, individuals were followed (retrospectively) from their index date until death, censoring due to discontinued enrollment in insurance plans, or the end of the follow-up period. Medical care costs were tracked starting at the index date and followed for up to 4 years following the index date. Hospitalizations, total medical care costs, and change in HbA1c (for the T2D cohort) were the pre-specified primary outcomes. Secondary outcomes included cardiovascular-related hospitalizations, composite cardiovascular events, all-cause death, and component medical care costs (i.e., inpatient, outpatient, and pharmacy costs). Inpatient hospitalizations were determined from insurance claims, cardiovascular-related hospitalizations from diagnosis-related group (DRG) codes (215–320), all-cause death from medical records and Minnesota Death Records, composite cardiovascular events (myocardial infarction, stroke, angioplasty and bypass graft procedure) from ICD, CPT or HCPCS codes (Table 1), and HbA1c from medical records, for patients with T2D.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Components of the composite cardiovascular study outcome.

https://doi.org/10.1371/journal.pone.0290028.t001

We assessed utilization of non-dental healthcare services from insurance claims. Specifically, we classified outpatient care using Current Procedural Terminology, 4th Edition, and Healthcare Common Procedure Coding System procedure codes; inpatient care using Medicare Severity DRG coding; and pharmacy utilization for medication fills using 11-digit National Drug Codes. Overall costs consisted of outpatient (utilization related to primary care, urgent care, emergency care, specialty care, laboratory procedures, and radiology), inpatient, and pharmacy costs.

We converted medical and pharmacy claims to dollar amounts using Total Care Relative Resource Values^™ (TCRRVs) [28]. TCRRVs are a nationally standardized set of pricing measures that have been endorsed by the National Quality Forum and are derived from Centers for Medicare and Medicaid Services (CMS) relative value units (RVUs) [29]. We used 2017 TCRRVs to convert all claims costs to 2017 U.S. dollars.

Statistical analyses

For analyses of clinical outcomes, we constructed a longitudinal dataset with patient-month as the unit of analysis. Outcome, treatment, and covariate values were carried forward and updated as new measures were available. Cohort-specific data sets were drawn from this master set. Patients were followed (retrospectively) from baseline until death, censoring due to discontinued enrollment in insurance plans, or the end of the follow-up period.

We tabulated descriptive statistics and evaluated bivariate associations using Kruskal-Wallis or chi-square tests as appropriate. We constructed generalized linear models (GLM), with an independent covariance structure for repeated measures, assuming a binomial distribution with a logit link function (except for HbA1c, a continuous outcome for which a normal distribution and identity link were specified). The GLM methods we used were equivalent to Cox survival models using pooled logistic regression [30–32].

Additionally, we estimated time-dependent, stabilized inverse probability weights (IPW) for likelihood of active periodontal treatment and censoring using the R package ‘ipw’. For the IPW weighting, a binomial family with a logit link was used. Time-related variables included: Age and Years Since 1st Diagnosis of Periodontal Disease. Time-varying variables included: Charlson Comorbidity Index, Tooth Count, Cumulative Months of Continuous Enrollment in Dental Plan, Generalized Periodontal Disease, Severe Periodontal Disease, Comprehensive Dental Insurance, Current Smoker, and History of Coronary Artery Disease. Time-invariant variables included: Male, Non-White.

We included covariates for the treatment and censoring models if they were specified a priori or were selected via LASSO (least absolute shrinkage and selection operator) [33, 34]. Table 2 contains additional details on these propensity models. In these models we considered an individual’s periodontitis diagnosis (slight, moderate or severe), extent of disease (generalized or localized) and extent of gingival inflammation (generalized or localized) as separate covariates. Final adjusted models were weighted by the time-dependent product of treatment and censoring weights. We truncated weights at the 1^st and 99^th percentiles to attenuate potential influence of extreme values, and report adjusted effect estimates as odds ratios (beta estimates for A1C) with 95% confidence intervals.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Propensity of active periodontal therapy model.

https://doi.org/10.1371/journal.pone.0290028.t002

Next, we calculated cumulative total and component costs at 1, 2, 3, and 4 years from baseline, which we modeled as a function of baseline covariates. We truncated extreme values at the 99^th percentile. For cost models, we considered individuals as treated if they had codes for active periodontal therapy (active treatment with or without maintenance care) by the end of year 1. Values for total, outpatient, and pharmacy cost outcomes were log-transformed and modeled assuming a normal distribution on the log scale. Because of a high proportion of zero values (76.7% of study patients did not incur inpatient costs in the 1^st year, and 28.1% in 4 years), and a right-skewed distribution, we modeled inpatient costs using a 2-stage process [35]. We report estimates as the difference in mean costs by treatment status, with 95% confidence intervals.

We examined the timing of an individual’s initial periodontal and medical diagnoses and considered patients as “incident” cases if the first periodontitis diagnosis appeared after the medical diagnosis, and “prevalent” if the periodontitis diagnosis preceded the initial medical diagnosis of interest. We conducted stratified analyses separately within each disease cohort (“prevalent” and “incident” cases). These results were qualitatively similar to those using the entire cohort and are not reported.

We compiled and analyzed the data in 2019–21 and used SAS v9.4 (SAS Institute, Cary, NC) and R v4.0.3 (cran.r-project.org). We report two-sided p-values.

Ethics statement

This research was approved by the HealthPartners IRB. Individuals who opted out of using their medical record data for research were not included in this study. Only medical record data was utilized so consenting of individuals was not required. Prior to the release of the data sets to investigators, private health information was removed and a random identification number was assigned.

Limitations

Along with previous studies, we did not consider an individual’s response to periodontal treatment. Periodontal treatment improves clinical measures of disease and oral health quality of life, but rarely eliminates all inflammation [46–48]. Poor responders to treatment may be at greater risk for subsequent cardiovascular disease and T2D, [49, 50] and it is possible that periodontal treatment positivity impacts general health outcomes only in patients who respond most favorably to treatment. In the current study, patients were assessed by >100 care providers not specifically trained to measure and record clinical indices in a standardized manner. Thus, we considered only the receipt of treatment as the “exposure.” Finally, there are no widely-accepted and precise criteria to determine treatment success, [51] other than prevention of tooth loss, which tends to occur over decades. We also did not distinguish between treatment provided by a dental specialist (periodontist), general dentist or dental hygienist. Although it is possible that the average response to periodontal treatment may have differed among these provider types, our goal was to examine the effects of periodontal treatment on medical care costs within a healthcare system, and not within provider subtypes. Additionally, non-surgical periodontal treatments and maintenance care are commonly performed in general dental offices, frequently by dental hygienists [52].

We established the 3 cohorts using ICD codes alone, which is a common assessment method in reports of heart disease, stroke, and type 2 diabetes [53–55]. We did not use medication data to identify affected individuals. In a large study of insured US adults with T2D, only 77% were found to be taking diabetes-related medications, suggesting that diagnosis codes, and not medication use, was the more comprehensive assessment method [56]. Nonetheless, an estimated 27–28% of the insured adult population with type 2 diabetes remains undiagnosed. Although our reliance on ICD codes alone may have missed a substantial fraction of those with T2DM, we compared outcomes only among those with an ICD T2D diagnosis, and not between those with and without a T2D diagnosis.

This was not a randomized trial and treated and untreated groups differed substantially in terms of baseline characteristics. Adjustment for covariates markedly attenuated associations seen in the unadjusted analyses of the clinical outcomes. These findings highlight the myriad factors that affect a patient’s decision to undergo treatment [57–59], and the impact of confounders on the relationship between periodontal treatment and non-oral health outcomes. Using simulation, Alshihayb et al. showed that adjusted associations between periodontitis and incident T2D could be further attenuated by considering unmeasured confounders [60]. Other potential confounders include attitudes toward health and oral hygiene, diet and exercise habits, and the dental care provider. We did not have information about the type of dental care provider, including his or her propensity to recommend treatment, and recognize this as a limitation. Information about diet and exercise was available for only a small fraction of our patients and therefore could not be used in the current work. We used IPW to balance treatment groups for confounders, but did not utilize a “doubly robust” approach, which would have produced an unbiased treatment effect estimator if our statistical model (of the causal effect of periodontal treatment on the medical outcomes) was misspecified [61]. Further, our adjusted analyses implicitly assume that that the subgroup of ’no active treatment’ (i.e., ’no care’ and ’maintenance only’) have a similar distribution of confounders for the comparison with ’active treatment’, which may be too strong an assumption. Nonetheless, the substantial differences between unadjusted and IPW (adjusted) results suggest that our models addressed many–but certainly not all–confounders.

Our study included more complete patient data but was considerably smaller than many previous studies. For some outcomes, our samples lacked statistical power to detect small to moderate treatment effects. Although we excluded large treatment effects for some clinical and cost outcomes, larger samples are needed to rule out smaller but potentially clinically meaningful effects.

We examined electronic health record data from 2004–2017. We did not consider the calendar year of an individual’s index date in the analysis. It is possible that, over time, patients with periodontitis are more likely to seek care because of a growing awareness of associations between periodontal disease and other health conditions. Associations between periodontitis and heart disease, stroke, and type 2 diabetes [62–64], however, have been publicized since the early-1990s, which is well before the start of our study period.

Finally, we included only patients with continuous medical and dental insurance one year prior to and at least one year following their index date, but did not have access to measures of socioeconomic status (SES). Insurance plan type, however, was a component of the propensity score for receiving periodontal treatment. Individuals had either comprehensive insurance, which included a benefit for periodontal treatment, or a public plan, which at the time of this study included no such benefit for adults. In this regard, the comprehensiveness of one’s insurance plan might be considered a measure of SES. In 2014–7, only about half of U.S. adults with medical insurance also had dental insurance in the previous 12 months [65]. Individuals without dental insurance are less likely to visit a dentist and undergo treatment [66–68]. Given all individuals in this study had some level of continuous dental insurance, it is unclear why a relatively small fraction of each cohort received any periodontal treatment (see Table 4). Patients accept or refuse periodontal treatment for many reasons, including fear, anxiety, and trust of the provider [69]. Thus, our results may not be fully generalizable to the broader at-risk adult population.

It is not possible to fully elucidate the nature of associations between periodontitis and non-oral diseases through observational studies alone. Patients who pursue periodontal treatment and maintenance care differ from others in a wide variety of characteristics [57, 58, 70], for which statistical adjustments or matching alone are unlikely to fully address. Evidence from large, randomized clinical trials, when feasible, are needed to help determine if periodontal treatment improves medical outcomes and reduces medical care costs in individuals with CAD, CBVD or T2D. Otherwise, studies that use large, comprehensive, and integrated clinical record systems and quasi-experimental observational study designs, such as those that leverage natural experiments or instrumental variables, remain warranted.