Building information for systematic improvement of the prevention of hospital-acquired pressure ulcers with statistical process control charts and regression
- 1Center for Pharmaceutical Outcomes Research, University of Colorado, Aurora, Colorado, USA
- 2Leadership Preventive Medicine Program, Dartmouth-Hitchcock Medical Center, Lebanon, New Hampshire, USA
- 3State University of New York Upstate Medical University, School of Medicine, Syracuse, New York, USA
- 4Harvard Vanguard Medical Associates, Visual Services Department, Boston, Massachusetts, USA
- 5The Dartmouth Institute for Health Policy and Clinical Practice, Center for Leadership and Improvement, Dartmouth College, Hanover, New Hampshire, USA
- Correspondence to Mark E Splaine, Center for Leadership & Improvement, The Dartmouth Institute for Health Policy and Clinical Practice, 35 Centerra Parkway, Lebanon, NH, 03766 USA;
Contributors All authors made significant contributions to the manuscript in accordance with ICMJE Guidelines: study conception and design, data acquisition, analysis and interpretation, drafting and revisions of article content, and final approval of the submitted version. William Padula wrote the Introduction, Methods and Results, and performed SPC chart and regression analyses. Dr Manish Mishra wrote parts of the Introduction and Discussion, collected hospital billing data, and reviewed patient charts for data in SPC chart analyses. Taygan Yilmaz and Christopher Weaver wrote components of the Methods and Results, and performed all SPC p-chart analyses. Dr Mark Splaine, the senior author, developed the study design and wrote components of all sections of the manuscript.
- Accepted 24 February 2012
- Published Online First 23 March 2012
Objectives To demonstrate complementary results of regression and statistical process control (SPC) chart analyses for hospital-acquired pressure ulcers (HAPUs), and identify possible links between changes and opportunities for improvement between hospital microsystems and macrosystems.
Methods Ordinary least squares and panel data regression of retrospective hospital billing data, and SPC charts of prospective patient records for a US tertiary-care facility (2004–2007). A prospective cohort of hospital inpatients at risk for HAPUs was the study population.
Results There were 337 HAPU incidences hospital wide among 43 844 inpatients. A probit regression model predicted the correlation of age, gender and length of stay on HAPU incidence (pseudo R2=0.096). Panel data analysis determined that for each additional day in the hospital, there was a 0.28% increase in the likelihood of HAPU incidence. A p-chart of HAPU incidence showed a mean incidence rate of 1.17% remaining in statistical control. A t-chart showed the average time between events for the last 25 HAPUs was 13.25 days. There was one 57-day period between two incidences during the observation period. A p-chart addressing Braden scale assessments showed that 40.5% of all patients were risk stratified for HAPUs upon admission.
Conclusion SPC charts complement standard regression analysis. SPC amplifies patient outcomes at the microsystem level and is useful for guiding quality improvement. Macrosystems should monitor effective quality improvement initiatives in microsystems and aid the spread of successful initiatives to other microsystems, followed by system-wide analysis with regression. Although HAPU incidence in this study is below the national mean, there is still room to improve HAPU incidence in this hospital setting since 0% incidence is theoretically achievable. Further assessment of pressure ulcer incidence could illustrate improvement in the quality of care and prevent HAPUs.
In 2008, Centers for Medicare & Medicaid (CMS) ceased to cover the cost for a list of preventable illnesses called ‘never events’, which included hospital-acquired pressure ulcers (HAPUs).1 The change in policy immediately strained hospitals to do more to prevent the incidence of these complications.2 HAPUs can range from US$500 to 70 000 per patient and totals US$11 billion per year nationally among an estimated incidence of three million adults.3–6 Stronger efforts to prevent a never event have great cost-saving potential for hospitals.7
The process to improve prevention of never events can begin with an analysis of the current outcomes in the medical setting, including incidence, prevalence and quality measures.8 Yet, even with all the analytical methods available to evaluate the epidemiology and quality measures of never events in the hospital, it often remains unclear where a clinician or health professional team should begin. Combining hospital-level (ie, macrosystem) data with data specific to clinical units (ie, clinical microsystems) for a specified condition can provide complementary information about quality care and patient safety that is actionable by clinicians.9 10 For example, measuring the proper use of a risk-assessment tool in HAPU prevention such as the Braden scale in a microsystem is a useful indicator of quality of care for hospitalised patients.11 Information gleaned from the Braden scale can then be used to decrease the variation in preventive care, thereby reducing the causes of HAPUs.12 This type of information is only available at the microsystem level and requires a system for gathering such data. By coupling this microsystem data and macrosystem data on HAPU incidence and correlated risk factors (eg, age and length of stay) for a local patient population, a hospital can build a bank of information on outcomes and possible causes of never events such as HAPUs. Hospitals may then be able to target improvement of such outcomes based on knowledge of existing risk factors with appropriate interventions throughout separate microsystems based on individual performance.13 14 Thus, a hospital macrosystem experiencing high HAPU incidence among older and extended-stay inpatients should improve targeted Braden scale assessment for these patients.
In this study we provide examples of macrosystem analysis of HAPU outcomes using regression methods. In addition, we analyse HAPU prevention from the level of the clinical microsystem simultaneously with statistical process control (SPC) charts.15–17 By viewing the strengths and limitations of each form of analysis through multiple levels of hospital function, this study offers a guide to clinicians interested in evaluating the outcomes and quality of preventive care for targeted never events.
This study followed an 11-step approach developed by Nelson et al to analyse data on incidence and quality measures.14 We investigated HAPU incidence in four targeted surgical wards at Dartmouth-Hitchcock Medical Center (DHMC), a major tertiary-care facility for the rural region of northern New England located in New Hampshire, USA. Data on the quarterly incidence of HAPUs was obtained using the hospital billing database from 2004 to 2007 for all patients admitted to the four surgical wards, collected before changes in CMS policy for HAPU reimbursement took place.1 These data were used to determine HAPU incidence and the correlation between incidence and patient gender, age, length of stay (LOS), time, ward, previous medical history and HAPU location on the patient's body. LOS was a continuous count variable that increased with each day that patients remained in the hospital. Generally, as LOS increases, a patient's risk of developing a pressure ulcer rises concomitantly at a nonlinear rate.18 Medical history was a binary variable equal to ‘1’ if a medical condition was present in the patient's history or ‘0’ if no condition was diagnosed. We only made note of patients with a previous medical history if they had a primary or secondary diagnosis for any of the following chronic diseases upon admission: chronic obstructive pulmonary disease, diabetes, congestive heart failure, hypertension, depression, renal disease or cardiovascular disease. Any patient diagnosed with a pressure ulcer (ICD-9707.*) during their hospital stay was included in the study.19 Coding was also translated to differentiate the location of the pressure ulcer as elbow, upper back, lower back, hip, buttocks, ankle, heel, sacrum or unspecified. Incidence was then determined relative to the total number of patients admitted to the hospital during a quarter. Once collected, the data were organised longitudinally by individual ward and quarter. For one particular surgical ward, data were prospectively collected from patient records on HAPU incidence and use of the Braden scale upon admission for all patients over 13 quarters during 2004–2007. The data were analysed using multiple methods to determine the magnitude of HAPU incidence and to derive possible sources of causation. A standard approach of regression analyses with cross-sectional and panel data from the macrosystem level were contrasted to a more novel method with SPC charts at the microsystem level.20 21 The results of these evaluations were contrasted for the sensitivity of information to detect quality of preventive HAPU care.
First, we analysed HAPU incidence in all four targeted wards using a regression model. The model was run in Stata (Stata IC V.10, StataCorp) as logit and probit regressions. The regression models were designed to control for multiple parameters, including: age, age2 (the square of a person's age), LOS, gender, medical history and location of the pressure ulcer. Two interaction variables were also included in the model: age and gender, and age and medical history. We did not control for individual wards in this model since we preferred to identify patients at risk for HAPUs by the aforementioned parameters rather than simply comparing wards' performance. β Coefficients and marginal effects were calculated for each variable to determine magnitude of impact on incidence. β Coefficients represent a 1-unit change in the dependent variable, HAPU incidence, with a change in an independent variable (eg, age or LOS) by 1 SD.22 Marginal effects are directly interpretable for the results of this analysis as the change in effect of a dependent variable with a 1-unit change in an independent variable. The probit and logit models each yielded a pseudo R2 calculated with McFadden's formula.23 A joint-hypothesis test was done for the model to determine which variables were statistically significant.
Panel data analysis
With data organised longitudinally by inpatient ward and quarter, we developed a second model with HAPU incidence as the dependent variable, and controlled for the following parameters: age, age2, LOS, gender and medical history. HAPU location was omitted since there were too few patients per ward with a HAPU in each of the seven categories. The longitudinal analysis included multiple steps. First, output of the summary statistics for each variable was obtained. Then, by generating individual-specific means for each ward, residual means were developed to identify between-group and within-group differences for wards across every quarter of observation. In doing so, we could use differences to determine if any particular wards were achieving high-quality care or underperforming during certain quarters compared with each other. Top-performing wards should maintain low HAPU incidence rates; wards identified with high HAPU incidence through between-group differences indicate poor quality performance. By maintaining this panel data analysis with the pre-specified covariates, issues such as higher age or LOS could offer some explanation to poor HAPU prevention despite zero tolerance for never events. A Breusch-Pagan tested if heteroskedasticity existed. Finally, the model was regressed both as a fixed and random effects model, and a Hausman test was performed to determine the more appropriate form of the model through a generalised least squares, logit or probit approach.
Statistical process control analysis
Using incidence data for one inpatient surgical ward at DHMC, we developed three SPC charts to track improvement in HAPU prevention over time. Charts from one ward were most useful to focus on patient quality and safety with regards to HAPUs, whereas multiple wards may have diversified findings with SPC. The charts were categorised as either a p-chart or a t-chart. A p-chart, or ‘proportion chart’ shows count data of non-conforming units for a measurable time, such as disease incidence per quarter among a larger population of patients.24 25 Since the size of both the incident and patient populations can change quarterly, the proportion is always a relative figure; control limits in a p-chart reflect the size of the denominator in each point (equation 1):(equation 1)In equation 1, p̄ represents the proportion of cases to total observations, and n represents the total number of observations at time t; z represents the number of SDs away from either side of the mean that the control limits extend towards, in this case z=3.
A t-chart is a measure of time between disease incidences for a pre-specified period of time.26 The upper control limit for this chart was based on the following formula (equation 2):(equation 2)In equation 2, ŷ is the mean value of the transform of time between occurrences across all observations, such that y=t0.2777. The transform of time, ŷ, is represented in t-charts by a centre line. MR' represents the average moving range of each y. For p-charts and t-charts, there are special and common cause variations. Special cause is indicated by a point above or below a control limit; eight consecutive points above or below the average indicate a process shift. Common cause variation occurs throughout observations at random—it represents the ‘noise’ resulting from factors inherent in the process, such as HAPUs that are occasionally not preventable.25 27
SPC charts of HAPU outcomes at DHMC were created in SPC XL 2007 software (Digital Computations Inc., http://www.sigmazone.com) to analyse several unit-level measures. First, a p-chart showed HAPU incidence for the studied ward. HAPU incidence in this ward was measured against the national mean for HAPU incidence (7%) with a t test for statistical comparison.28 Second, a t-chart illustrated the number of days that occurred between the last 25 HAPU events in the ward during 2007. Third, another p-chart showed the reliability of Braden scale risk assessment and documentation in the nursing charts for all admitted patients in the ward of interest over a period of 20 weeks during the observation period.
There were a total of 43 844 patients analysed in this study in all four inpatient surgical wards at DHMC from 2004 to 2007 (table 1). There were slightly more women than men in these wards. The average patient age was 57.5 years, and most patients were adults and seniors. There were 337 cases of HAPUs between 2004 and 2007 in this population. Unfortunately, there were too few patients per quarter with HAPUs in each body location to include these parameters for the analyses. Nearly half of patients had a medical history with a diagnosis of interest.
Age had a negative marginal effect on incidence in the probit and logit models (table 2). The interaction term between male gender and age had a negative marginal effect on incidence in the probit and logit models. The parameters for medical history and pressure ulcer location were removed from the final model since they were not found to be statistically significant variables.
All included β coefficients for independent variables were statistically significant at the 95% CI (p<0.05) except for age, which was not found to be statistically significant in the probit or logit models. A joint-hypothesis test was done comparing this model to a restricted model (ie, age=0). This joint test showed that although the β coefficient for age was not statistically significant, the impact of age on the regression model was significant (p<0.0001). Therefore, the equation was left unchanged rather than adopting the restricted model.
While there was statistical significance in both methodological approaches, the probit model yielded a higher pseudo R2. Because of this finding, we concluded that the probit regression model is the best-fit model to predict the impact of gender, LOS and age on HAPU incidence.
Panel data results
HAPU incidence ranged from 0 in one ward for a quarter to a high of 13 in a ward over another quarter. Between-group and within-group variation showed no statistically significant differences between wards over time for age, proportion male, LOS or medical history that would have suggested a correlation to the incidence of HAPUs. The Hausman test confirmed use of a generalised least squares random effects model, which showed a statistically significant correlation between LOS and HAPU incidence (online appendix table A1). For each additional day in the hospital, there was a 0.28% (p=0.026, 95% CI 0.0003 to 0.0054) increase in the likelihood of developing a HAPU. No heteroskedasticity was detected among the panel data.
Statistical process control results
A p-chart of HAPU incidence in one ward at DHMC offered valuable information giving context to prevention in a focused patient population. The mean incidence rate was 1.17%, with an upper control limit of 2.50% over the 13 quarters; there was no lower control since 0% incidence is preferable (figure 1). The incidence rate remained in statistical control during the observation period as no special cause variation was detected. In other words, there may be process changes, but none resulted in a special cause variation signal. Compared with the national mean, this HAPU incidence rate was detected as significantly lower using a t test (p<0.0001, 95% CI 0.0001 to 0.025). Overall, the system of care for HAPU prevention remained in statistical control—there was no special cause variation—when taking into account the total number of patient admissions.
A t-chart of the days between HAPU incidences of the same single surgical ward was constructed using data collected from the DHMC billing database (figure 2). The average time between events for the last 25 HAPUs of 2007 was 13.25 days with the upper control limit set at 183.22 days. The centre line was equal to 7.24 days. There were no special cause signals, but there was a 57-day period between the 16th and 17th events. Considering the overall frequency of HAPUs, further investigation for that time period may be warranted. There may have been a change in the microsystem that led to an improvement in HAPU prevention, although it is not indicated by special cause.
Throughout a 20-week observation period, approximately 93% of inpatients in the single surgical ward were risk stratified upon admission with the Braden scale (figure 3). The current process varies between 70% and 100%. Ideally, the support staff in a hospital should strive to attain 100% of Braden scale forms being properly filled out as this will ultimately aid in the prevention of HAPUs, as was done in 8 weeks during the observation period.29 The process appeared to be in statistical control, except in week 6 when risk assessment dropped below 70%. Overall, the mean indicates the need for improvement towards 100% completion of the Braden scale for patients upon admission.
This study demonstrates the utility of combining enumerative and analytical statistical methods to understand complex issues in preventive medicine such as reducing HAPU incidence. Existing literature provides information about the evidence for risk factors related to HAPU outcomes, such as age and LOS. However, supporting evidence can become skewed when translated into a specific context.30 The regression analyses are enumerative statistical methods based on locally collected data showing that HAPU outcomes in DHMC's environment are similar to previous literature. In addition, the regression analyses can be used with macrosystem-level data (eg, billing data) due to their lack of sensitivity for the change in small numbers. Thus, a hospital administration interested in knowing if their macrosystems have actually improved care will observe it in the culmination of data into a regression analysis in which the patient population coming into the hospital has not changed. However, regression analysis should indicate reductions in HAPU incidence when organised properly, such as longitudinally.
Given the purpose of enumerative statistical methods, consider the results of the regression and panel data analyses provided in this study for a large, hospital-wide inpatient population. These analyses correlate HAPU incidence and factors relating to hospitalised patients, such as the 0.28% increase in HAPU risk with additional LOS. This macrosystem analysis should lead to changes in function in hospital microsystems, such as increased Braden scale assessment for patients who remain bedridden for extended duration. Quality improvement interventions, such as alarms, should inform hospital staff to maintain 100% risk assessment after admission for patients under extended stay. Regression can also bring contrast to performance between Microsystems and explain why patient populations in certain microsystems might lead to different outcomes. Issues that are identified at the macrosystem level are valuable since they can be disseminated to all levels of the organisation but must be adapted to the context of individual microsystems within the hospital.9
SPC chart analyses are analytical statistical methodologies that provide examples of the use of measurement at the microsystem level to study and more deeply understand improvement efforts related to HAPU prevention. These charts prove to be an ideal tool for aggregating small amounts of data to determine the level of quality with which clinical staff prevent HAPUs by using recommended components of prevention guidelines (eg, the Braden scale).25 31 32 SPC charts can illustrate the nature of how clinicians are taking care of patients in the microsystem. The p-charts and t-chart show exact periods of time during which the quality of preventive care either improved or reduced. For instance, the p-chart measuring HAPU risk stratification for admitted patients exposed an opportunity for improvement. Although not a special cause, the t-chart contained a signal of 57 days between HAPU incidences. There may have been a brief process change in the ward that led to a favourable outcome. If that were the case, then efforts aimed at recreating those conditions would be an effective strategy. By making hospital clinicians aware of their successes identified through the t-chart, clinicians may emulate their behaviour in this time period as a benchmark for maintaining improved HAPU incidence, and thus higher quality care. However, this signal may also prove to be a false alarm. The 57-day window may have been a function of having fewer high-risk patients in the ward, which could yield a lower likelihood of HAPU incidence. It is therefore important that data on HAPUs are collected prospectively so that as this point of improvement unfolds, clinicians can detect reasons for process change in real time.
The study is limited by several factors, including the size of the dataset and the number of variables available for analysis. The data did not contain hospital-wide data points on completion of the Braden scale to include in the regression analyses for comparison with SPC charts. In fact, we only investigated HAPU outcomes with SPC charts for one ward, due in part to the complex undertaking of manual patient chart reviews. There were not enough HAPUs to sufficiently control for certain pressure ulcer outcomes, such as the body location of the condition, or derive potential causes due to specific medical conditions. Multi-site studies of never events with the use of regression and SPC charts would provide more insightful information about variation between microsystems and macrosystems. In addition, we were unable to determine the stage of the pressure ulcer as reported in the billing records. The importance of this stems from the possibility that only severe (stage III and IV) ulcers may have been reported through billing records, which would likely underestimate the incidence of HAPUs that can cause significant morbidity to patients. Limitations of our analysis are found in the data for the p-chart HAPU incidence. Because SPC relies on the premise of having at least 20 data points, analysis of these data should be viewed with caution.25 The relationship between LOS and HAPU incidence is likely nonlinear based on the results of the probit and logit regression analyses. This study did not further explore the use of other rigorous nonlinear functions to determine the correlation of LOS and may warrant further research.
A further concern of this study is the response to special cause signals. Although this study used prospective data collection for SPC charts, all data were reviewed retrospectively. Because of this, we were unable to identify possible links between special causes and issues in the microsystem as each event occurred. A better system of data collection consists of prospective collection and immediate response to special causes. Issues that arose during the period of data collection included profuse turnover in nursing staff at the hospital, adoption of electronic health record systems and CMS policy intervention regarding never event reimbursement. Any one of these events or another issue could have contributed to special causes at the level of the clinical microsystem, though the issue cannot be pinpointed to one special cause signal. Issues with nursing turnover and electronic health record adoption were more likely to impact signals apparent in the SPC charts since these issues occur at the microsystem level.33 Conversely, CMS policy intervention interacts with the clinical macrosystem and becomes more apparent in longitudinal changes in incidence through regression analysis.34
Ultimately, we can conclude from the regression analyses based on cross-sectional and longitudinal sampling of patients at DHMC—a hospital organisation—that inpatients are disposed to a higher risk for HAPU incidence with certain age, LOS and gender. These results apply to all wards in which the patients receive care at DHMC. While the regression results are aligned with the findings of other studies on HAPU prevention, developing specific actions in response to the findings is challenging.35 The results of SPC charts were drawn from one ward over time, and are straightforward to develop and interpret. These charts allow interventions to be targeted at the level of the clinical microsystem and the results of such changes monitored over time. For example, the SPC analyses of days between HAPU incidences and Braden scale assessment are helpful for informing clinicians of their performance at preventing HAPUs, which is invaluable in achieving lower incidence rates in a microsystem.13 The complementary use of regression and SPC analyses informs clinicians at our hospital that patient risk stratification should reach beyond what is explicitly recorded at the bedside in the Braden scale—mobility, activity, friction and shear, nutrition, moisture and incontinence, and sensory perception.11 Future quality improvement initiatives through clinical microsystems should incorporate these considerations.
In response to these findings, a collaborative team of physicians, nurses and hospital administrators developed quality improvement interventions to reduce HAPU incidence further. Even though statistical analyses indicated that HAPU rates were already low at DHMC, there was still room to reduce pressure ulcer incidence in this hospital setting because 0% incidence is theoretically achievable. We determined that publicly displaying information on HAPU incidence and patient Braden scale assessments by ward was most effective at preventing HAPUs, in conjunction with standard prevention guidelines. The team developed a poster-sized public display of each patient's daily Braden score and quarterly HAPU incidence in p-chart form. By enlarging SPC charts for presentation as data walls, medical staff could efficiently produce a quality improvement intervention for HAPU prevention, which would lead to greater transparency of clinicians' performance and their relationships with patients.36 37 Conclusions drawn from SPC charts can be shared with clinicians to benchmark improvement and sustain high-quality practices. Good practices are contagious and can lead to wide-scale adoption. With further monitoring through the forms of regression highlighted above, successful spread of preventive awareness through microsystem data should be reflected by reductions in HAPU incidence despite steady streams of patients with complications due to age, LOS or comorbidities.
If the goal is to bring the incidence of these preventable illnesses to zero, then the first step is to fully evaluate and understand the clinical microsystem of interest.38 In doing so, an informed clinical microsystem can guide and support innovation through peak performance in healthcare.10
The authors acknowledge Dr Greg Ogrinc and Dr Stephen Plume for their contributions in support of this research. We also thank the editor Dr Kaveh Shojania and the two anonymous reviewers for their comments and edits to improve the manuscript.
Competing interests None.
Ethics approval Ethics approval was provided by Dartmouth College Institutional Review Board.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement All data utilised for the analyses are available by request, providing compliance with HIPAA regulations. Patient charts and billing data are stored on a secured computer in MS Excel. SPC chart analyses are stored in Excel, and coding and logs of regression analyses are stored in Stata data files.