Tobacco related illness remains one of the leading causes of preventable death nationally [1, 2], and is one of the top drivers of healthcare costs among military Veterans [3]. Because of the importance of addressing tobacco use in the clinical setting, smoking status is routinely recorded as a field in the electronic health record (EHR) [4, 5]. This has led to interest in utilizing these fields for research and evaluation purposes. Accurate smoking status is key for a number of research purposes, including the ability to assess the outcome of smoking cessation programs. Healthcare organizations such as the Veterans Health Administration (VA) or academic medical centers are a common setting to test tobacco cessation interventions [6–8], as they allow researchers to capitalize on contacts with high-risk patients that occur as part of routine care. Traditional research methods of capturing smoking status, which require at a minimum that the patient be contacted by phone or mail and have their reported information entered into a database, can be costly and time consuming. This potentially limits their utility for evaluating system-wide tobacco interventions, or tobacco interventions offered in the setting of clinical care. Despite the efforts of study staff, some patients will remain non-responders to surveys, which may bias the results. Finally, patient reports of tobacco status may also be subject to biases such as social desirability or stigma, which may differ between research and routine clinical settings. Computer-based approaches such as natural language processing require significant programmer expertise, and still require a large investment of time on the part of investigators at each institution to ensure accuracy [9]. Use of a validated, EHR-based method could streamline this process.

Within the VA EHR, tobacco use status is recorded in the form of Health Factors (HFs)—electronic data fields that are entered at the time of clinical care and can be readily retrieved. The VA remains the largest single healthcare system in the United States, with 144 hospitals and over 8 million patients [10]. Examples of common tobacco health factors include “CURRENT SMOKER”, “CURRENTLY USES TOBACCO” and “NON-TOBACCO USER.” Smoking status gathered by HF does not require additional resources, such as research staff, to obtain it. Because they are tagged data fields, similar to diagnostic codes, they can be easily retrieved. However, because they are entered by clinical staff and not study personnel, they are not subject to the more rigorous definitions of tobacco use or abstinence employed in research settings, making their use and limitations for research purposes less clear. For example, staff may erroneously identify a patient as a smoker who uses non-combustible tobacco products. HFs are updated frequently during hospitalizations, primary care visits, and mental health appointments, with providers prompted to enter them at least annually. Most VA facilities prompt nursing staff to enter HFs at least annually when a patient presents to the primary care clinic. These automatic prompts do not rely on each provider to remember to make the assessment, which increases the likelihood that HFs will capture changes in smoking status. A previous study compared past year HFs with baseline tobacco status obtained by survey in two cohorts including over 18,000 patients and found high agreement with survey data, particularly among never smokers and current smokers [11]. Subsequently, the VA performed a detailed examination of the utility of HFs for performing Health Services Research. The resulting technical report, produced by the VA Health Economic Resource Center (HERC), examined the records of 5.7 million unique Veterans from 2009 to 2011, and found that the majority of Veterans (70.3%) had a tobacco HF, with significant variation in availability by site. Of patients with an available baseline measure, almost 90% had a follow-up measure within 2 years, with an apparent quit rate of 12.3% among current smokers. The HERC technical report concluded that HFs provide a useful tool for epidemiologic studies and long-term follow-up on quit behaviors within the VA. A more recent analysis used these methods to assess over 700,000 current and 400,000 former smokers within the VA, and found similar rates of the availability of health factors, and an annual quit rate of approximately 12% [12]. However, while both the HERC report and the recent publication by Barnett et al. demonstrate conclusively that HFs are widely available and change over time, the authors were unable to assess the accuracy of the change in status, only that the change had occurred [13]. To date, no studies have reported the accuracy of HFs for assessing change in tobacco status over time by comparing them to another validated method.

Because of these two features—high baseline validity and the possibility of tracking changes in smoking status over time—HFs provide an attractive EHR-based means to capture changes in tobacco status in response to interventions. HFs could be particularly useful for evaluating large scale, multicenter or system-wide interventions that would be cost-prohibitive to assess by conventional means. EHR-based smoking status could also fill in results for survey nonrespondents. These missing data elements are commonly accounted for by statistical means [14], but HFs may prove to be more accurate. We aimed to examine the utility of HFs as an alternative method for obtaining follow-up data on smoking status using the EHR. Specifically, we assessed whether HFs can provide an accurate, cost and labor-saving means of assessing the response to tobacco interventions. To accomplish this goal, we performed a secondary analysis of a randomized controlled trial of proactive outreach for tobacco cessation, comparing the survey tobacco outcomes from the trial with HF data from the EHR. The primary goal of our analysis was to determine the proportion of subjects with a change in HF tobacco status, and assess the accuracy of this information in comparison to survey responses. Our secondary goals were to assess the optimum timeframe for obtaining follow-up HFs, and examine the variability in accuracy and availability of HF data by site. In addition, we performed a stratified analysis of patient groups with frequent healthcare contacts to see whether HFs were more accurate among these subgroups of participants.

6400 participants were randomly selected and assigned to either proactive care or usual care and mailed a baseline survey to confirm smoking status. 1277 subjects were excluded for not meeting eligibility criteria: 428 declined to participate; 201 were misclassified as cigarette smokers (had never used cigarettes or smokeless tobacco user); 444 were former smokers (rather than current smokers) at the time of contact; 179 had incorrect mailing addresses; and 25 were deceased (Fig. 1).

When examining differences between responders and nonresponders to the survey, responders were older (57.8 vs 53.2), and less likely to have mental illness (46.1% vs 51.8%). There were minor differences in response rates by site. When examining differences between subjects with HFs available and those without, subjects with health factors were more likely to be white (63.6% vs 52.6%) and more likely to have been hospitalized in the past year (11.7% vs 9.6%). There were large differences in availability of HFs by site. These analyses are available in Additional file 2.

Agreement and availability by site (Table 2)

Agreement differed by site with large variability in Kappa (Site A: 0.39, Site B: 0.59, Site C: excluded due to low numbers with HF data, Site D: 0.48) and sensitivity (Site A: 43.2, Site B: 76.6, Site C: excluded, Site D: 49.2) among the sites. There was also large variability in data completion status by site. Site A had a large proportion of patients with both sources of data (52.9%) and HF data overall (81.5%), and Site C had very few patients with both sources of data (1.7%) or HF data (3.6%) (Fig. 2).

We compared the smoking cessation behaviors of participants in a randomized controlled trial, analyzing the agreement between survey methods and HFs drawn from the VA EHR. We found moderate agreement for quit behavior by Kappa statistics and sensitivity, but overall high concordance and similar population level quit rates. Though published data are few, our results are similar to the previous study comparing EHR and survey methods, in which the best agreement is found among continuous smokers and long-term non-smokers [11]. Although differences in the definition of smoking abstinence between survey and clinical assessments may account for some of the discrepancies in smoking status, we believe that the observed disagreement is predominantly due to the nature of tobacco cessation attempts. Recent quitters cycle in and out of attempts in relatively short periods of time, with most ultimately relapsing to tobacco use [19–21]. In this case, a discrepancy of a few weeks or months between when the HF and survey data were obtained would account for the difference. This is supported by the very high specificities and NPVs found, suggesting that for patients who remained smokers throughout, HFs are extremely accurate.

Despite these issues, the population level quit rates were very similar between the survey and HF methods. The observed discrepancies between methods of data collection may limit the validity of cessation data for individual patient-level inference, and researchers should be mindful of this when assessing the utility of HFs for answering a specific research question. However, the population rates of abstinence may yield very similar results when evaluated by either survey or HFs, and so HFs may be of greatest utility when this is the outcome of interest. The agreement in population level quit rates mean that HFs would likely provide a valid option for assessing population level abstinence or temporal changes as part of program evaluations, implementation studies, or pragmatic trials. They also provide a useful means for assessing epidemiologic trends in tobacco use, and examining outcomes of clinical innovations.

We did not find a decline in the agreement between HF and survey between wider and narrower timeframes. An even longer timeframe would likely increase availability, but we are unable to comment on the effect on accuracy. Site, more than any subgroup analyzed, had the greatest influence on the availability and accuracy of HFs, which is a major consideration for investigators considering using HFs in a research context. Similar to the HERC technical report [13], sites varied in the availability of follow-up HFs from over 80% of subjects to less than 5%. One site had more subjects with HF tobacco status available than subjects who responded to the survey, while the other three sites had fewer. The large variability in the availability of HFs by site may be related to differences in processes of who records these data, and how. For example, not all inpatient assessments of tobacco status generate HFs. Some HFs are standardized throughout the VA, but unique HFs can be created and used by any individual site. It is possible that some very uncommon HFs were used at a particular site that were not identified by our search algorithm, though we feel this is unlikely. For investigators considering using HFs for research or quality improvement purposes, a key first step would be assessing the availability at the institution in question.

We found some important demographic variation among patients who responded to the survey in comparison to those with HFs, as well as some variation in agreement among patient subgroups. Compared to survey responders, patients with HFs were younger, with higher rates of mental illness—two groups that may be contribute excessively to missing data gathered by survey. HFs are likely a more accurate means for filling in for non-response to survey than statistical methods, and can be employed in this capacity as an adjunct to survey administration. Nearly half of nonresponders to the survey had a HF status, and this number is likely to increase as HFs are employed more systematically throughout the VA [22]. We found that patients who were admitted to the hospital had the lowest agreement, though population level quit rates were again similar throughout. The low agreement for this group may be due to the fact that hospitalizations are a common time for a quit attempt [23]. Patients may indicate that they are a smoker during the admission, only to make a quit attempt shortly thereafter, which could have been captured by the survey. The highest agreement was found among patients over 65 years of age. Importantly, there was no difference in accuracy among patients with mental health diagnoses—a group with high smoking rates that is a priority for cessation interventions [24]. There were some moderate differences in population level quit rates between subjects with both sources of data vs only one, in which subjects with only survey or HF data tended to report a slightly higher quit rate. This suggests that there may be some unmeasured differences between patients who are nonresponders to surveys or are not captured by routine healthcare assessments.

Our study has some limitations. HF queries are made as part of face-to-face clinical interactions, and are not collected according to strict research definitions of tobacco abstinence, rather focusing on current smoking status only. HFs more often assess tobacco use as a whole, not just cigarette smoking. It is possible that patients who were dual users of both combustible and noncombustible tobacco may have been misclassified, as HFs may not allow capture of both forms of tobacco use. We do not have information on site and provider specific differences in how these queries were made that might explain the large variation by site. We compared the HF smoking status to smoking status by survey as our gold standard. While generally accurate [25], survey data may also be subject to inaccurate reporting, and we did see some variation in Kappa by study arm. This may be related to variations in how data was collected, or in recall bias on the part of patients who were or were not participating in a tobacco cessation program. Finally, these data refer to a Veteran population, that was largely white and male. This may limit the generalizability outside of the VA healthcare system, as both the EHR and patient population may not be representative of the US as a whole. However, as the largest single healthcare organization in the US, the data are likely to be applicable to other managed care organizations.

In summary, we found that HFs provide a reasonably accurate and available means to measure change in smoking status over time among current smokers at baseline using the EHR, though their utility is strongly dependent on site. A majority of patients had a follow up tobacco status available in a similar timeframe to a 12 month survey, with substantial agreement with population level quit rates as determined by surveys or by HFs. The specificity for sustained smokers was very high. HFs appear to be a valid alternative for assessing quit behaviors on an administrative level, and could be used both as a primary measure of quit behaviors in the context of tobacco research, as a secondary source of data for patients who do not respond to surveys and lastly, as a source of data to examine epidemiologic trends within the VA.