ARATIONAL approach to the provision of urgent

Transcription

1 48 URGENT CARE Batal et al. PREDICTING URGENT CARE VISITS Predicting Patient Visits to an Urgent Care Clinic Using Calendar Variables HOLLY BATAL, MD, JEFF TENCH, BS, SEAN MCMILLAN, BA, JILL ADAMS, BA, PHILLIP S. MEHLER, MD Abstract. Objective: To develop a prediction equation for the number of patients seeking urgent care. Methods: In the first phase, daily patient volume from February 1998 to January 1999 was matched with calendar and weather variables, and stepwise linear regression analysis was performed. This model was used to match staffing to patient volume. The effects were measured through patient complaint and left without being seen rates. The second phase was undertaken to develop a model to account for the continual yearly increase in patient volume. For this phase daily patient volume from February 1998 to April 2000 was used; the patient volume from May 2000 to July 2000 was used as a validation set. Results: First-phase prediction equation was: daily patient volume = January 4.56 winter 47.2 Monday 37.3 Tuesday 35.6 Wednesday 28.2 Thursday 24.2 Friday 7.96 Saturday 10.1 day after a holiday. This equation accounted for 75.2% of daily patient volume (p < 0.01). Inclusion of significant weather variables only minimally improved the predictive ability (r 2 = 0.786). The secondphase final model was: daily patient volume = Newdate 52.0 Monday 44.2 Tuesday 39.2 Wednesday 30.2 Thursday 26.5 Friday 10.9 Saturday 12.2 February 3.9 March, which accounted for 72.7% of the daily variation (p < 0.01). The model predicted the patient volume in the validation set within 11%. When the first-phase model was used to predict patient volume and thus staffing, the percentage of patients who left without being seen decreased by 18.5% and the number of patient complaints dropped by 30%. Conclusions: Use of a prediction equation allowed for improved accuracy in staffing patterns with associated improvement in measures of patient satisfaction. Key words: forecasting; utilization; walk-in clinics; regression analysis. ACADEMIC EMERGENCY MEDICINE 2001; 8: ARATIONAL approach to the provision of urgent care services must include consideration of trends in the demand for these services. According to the National Ambulatory Medical Care Survey, 17.6% of health care contacts take place at sites that do not require an appointment; this method of accessing care appears to be on the increase. 1 The ability to predict the number of patients seeking care on any given day in an urgent care clinic is essential in order to allow for optimization of staffing patterns. Matching staffing levels to the variation in daily patient demand can improve cost efficiency as well as patients satisfaction with care by decreasing waiting times. The major reason for dissatisfaction with emergency care and patients leaving without being seen is ex- From the Urgent Care Clinic at Denver Health (HB, JA), the University of Colorado, Denver (JT, SMcM), and Denver Health and the University of Colorado Health Sciences Center, Denver, CO (HB, PSM). Received February 14, 2000; revision received August 15, 2000; accepted August 24, Address for correspondence and reprints: Holly Batal, MD, Denver Health Medical Center, 777 Bannock Street, MC0107, Denver, CO Fax: ; hbatal@ dhha.org cessive length of wait. 2 Patients are more likely to leave without being seen as waiting times increase, with more than a fourth of these patients returning later for evaluation. 3 Other studies have also demonstrated this inverse relationship between waiting times and satisfaction with care. 4 6 Simply attempting to divert patients with apparently minor complaints out of the urgent care clinic because of inadequate staff availability is fraught with problems and may be an unsafe practice with serious conditions being overlooked. 7 To the best of our knowledge, few studies have been conducted in an attempt to quantitate the relationship between patient volume and calendar and weather variables. A study of the use of a pediatric emergency department (ED) found no significant difference in the number of ED visits or the number of resulting hospital admissions during unfavorable weather vs favorable weather. 8 An older study examined patterns of ED use, including ambulance use, in three Boston hospitals as they related to weather conditions. 9 In this study, there was a tendency in each of the three hospitals to have an increase in nonambulance patient traffic during more favorable weather. Only two previous studies have used calendar

2 ACADEMIC EMERGENCY MEDICINE January 2001, Volume 8, Number 1 49 and weather variables to develop prediction formulas for patient volume, and only one used a validation set of data to test their predictions. 10,11 Using calendar variables, Diehl et al. were able to account for 74.5% of the variation in the square root of daily visits; this number increased slightly to 77.3% when weather variables were also included. 11 Holleman et al. used a Veterans Administration hospital as the basis for their study. 10 The significant factors that predicted patient volume in this model were calendar variables (season, week of month, day of week, holidays, and federal check delivery day) and weather variables (high temperature and snowfall). They were able to use their model to predict patient volume, with their model accounting for 84% of the daily variance. Within their validation set, when their model was used to predict staffing, it would have produced adequate staffing on 67% of days, overstaffing on 15% of days, and understaffing on 18% of days. Using these prior studies as a guide, we undertook this project to identify the predictors of patient volume at the Denver Health Medical Center urgent care clinic and to develop a prediction formula to project the number of visits on future days. The ultimate goal was to allow for maximization of staffing to positively impact efficiency of patient care. METHODS Study Design. Stepwise linear regression was performed on existing data to determine significant variables for urgent care volume. A prediction equation was developed and used to predict patient volume presenting for urgent care, and validated with an existing data set. Study Setting and Population. Denver Health Medical Center is an urban safety-net public hospital serving the indigent residents of Denver, Colorado. This fully integrated health care system comprises the state s only Level 1 trauma center and 11 outlying community clinics. Each of these clinics provides primary care services for the residents of the community, as well as urgent evaluation of patients as availability allows. The urgent care center, known as the walk-in clinic (WIC), is located within the hospital and is open 15.5 hours per day, seven days a week. The WIC provides unscheduled outpatient care for adult patients whose clinical conditions are not of significant acuity to warrant ED care but who have chosen not to access primary care in community clinics. During the clinic s hours of operation patients either present directly to the clinic for care, are referred from the hospital s network of outlying community clinics, or are triaged into the clinic from the ED. Approximately 37,000 patient visits were recorded in 1999; in 2000 the clinic is expected to have exceeded 40,000 patient visits. Study Protocol. In the initial phase of the study, the number of patients treated daily from February 1, 1998, through January 31, 1999, was used. Daily patient volume was matched with calendar variables such as day of the week, month of the year, season, holiday, and the day before and after a holiday. Designated holidays were Labor Day, Memorial Day, Fourth of July, Martin Luther King Day, Thanksgiving, New Year s Eve, New Year s Day, and Christmas Eve. Christmas Day was the only day of the year that the clinic was closed due to historical data that showed a distinct paucity of utilization that day. On three of the holidays, Thanksgiving, New Year s Eve, and Christmas Eve, the length of time the clinic was open was less than 15.5 hours. For these days, the patient volume was corrected by extrapolating the actual volume to what would have been expected for a hour day. This was done by determining the number of patients per hour actually seen and then multiplying by 15.5 to calculate the projected number of patients who would have been seen in a normal-length day. Since a holiday can affect patient volume both on the day of and the days before and after a holiday, we also included variables if a given day was the day before a holiday, the day of the holiday, or the day after the holiday. To account for the effect of season, we included a variable winter, defined as November through February. These months were chosen based on the statistical distribution of patient volume. In addition to season, recent historical volumes may also be an accurate predictor of current volume. That is, the volume in the clinic tends to trend up and down over time, perhaps related to periodic outbreaks of upper respiratory infection or other communicable diseases or seasonal illnesses. To attempt to capture this effect, we created an independent variable for prior volume. To create this variable, we averaged the daily volume over the prior seven-day period. The local climatological data during the study period were obtained from the National Weather Service. We considered maximum, minimum, and average temperatures (in degrees Celsius) and precipitation amount. Precipitation was recorded both as total (water equivalent precipitation in inches) and as the amount of snowfall (also in inches). Whenever the National Weather Service listed precipitation as trace, we assigned this a value of inches. The second phase of the study used the daily patient volume from January 1998 to May 2000 to develop a prediction equation that captured the significant calendar variables as well as the con-

3 50 URGENT CARE Batal et al. PREDICTING URGENT CARE VISITS TABLE 1. Patient Volume by Day of the Week Day Average Daily Patient Volume 95% CI Monday , 118 Tuesday , 111 Wednesday , 108 Thursday 95 91, 99 Friday 93 90, 96 Saturday 77 74, 80 Sunday 69 66, 72 tinual rise in patient volume that we were experiencing (an approximately 7 8% increase per year since 1997). This new independent variable was labeled Newdate. Weather variables were not included in this second-phase prediction equation development since results of the first phase showed that they added little. Beginning September 1, 1999, we began to use the equation developed in the first phase of the study (inclusive of only calendar variables) to predict patient volume. Since the first-phase prediction equation did not account for the continual increase in patient volume that we were experiencing, 7% was added to the predicted values. Since the average number of patients seen by each provider class per hour (resident, nurse practitioner, attending physician) has been consistent and is well documented within our clinic, necessary staffing patterns and provider mix were determined by matching them to the predicted patient volume. In order to test our second-phase model, we used a validation set of data for the time period May 1, 2000, to July 31, For each of these days we calculated the predicted patient volume, with comparison made with the actual number of patients seen. Accurate waiting time data are not historically available within our clinic. In order to measure the effects of prediction equation utilization, patient complaint and left without being seen rates were used as proxy measures. Patient complaint and left without being seen rates were calculated as percentages of patient volume. These statistics are internally kept, and methods of data collection have been stable since January Data Analysis. In the first phase of the study, using both weather and calendar variables, stepwise linear regression analysis was performed on data from February 1, 1998, to January 31, 1999, in order to determine how much additional variance we have accounted for by including different variables. We initially included all variables in the equation, and eliminated insignificant variables one at a time. During model development, F statistics were calculated for all variables during each step. Variables were deleted from the model sequentially as dictated by the values of these statistics, until all variables were significant at the 5% level. Because calendar variables are predetermined, whereas weather variables are much less predictable, we developed a parallel model that excluded weather variables. A model without weather variables is more practical for advance calculation of day-to-day staffing needs. All analyses were performed using Mini-tab software (Release 12, State College, PA). In the second phase of the study, stepwise linear regression analysis was performed on data from February 1, 1998, to May 31, 2000, with the addition of a variable allowing us to account for the continual increase in patient volume over time (Newdate). All calendar variables were initially included, and insignificant variables were eliminated one at a time until all remaining variables were significant at the p < 0.01 level. Analyses in this second phase were performed using SAS software (SAS Institute, Cary, NC). RESULTS As shown in Table 1, day of the week is the greatest consistent predictor of patient volume. The highest daily patient volume occurs on Monday, with a fairly linear decrease until Sunday. On any given day the patient volume shows a normal distribution. Month of the year is much less predictive of patient volume (Table 2). However, volume is consistently lowest in the spring and summer months (April August). First-phase Model Development. In the first phase of the study, the final regression equation using all significant calendar and weather variables was: daily patient volume = January 6.63 winter 46.7 Monday 37.2 Tuesday 36.0 Wednesday 29.4 Thursday 24.1 Friday 8.67 Saturday 10.1 day after a holiday 5.28 July 5.57 August maximum temperature (C) 3.88 snowfall (inches). This equation accounted for almost 79% of the daily patient variability (r 2 = 0.786, p < 0.01). When weather variables were excluded from the model, the final regression equation became: daily patient volume = January 4.56 winter 47.2 Monday 37.3 Tuesday 35.6 Wednesday 28.2 Thursday 24.2 Friday 7.96 Saturday 10.1 day after a holiday. This equation still accounted for approximately three-fourths of the variation in daily patient volume (r 2 = 0.752, p < 0.01). The coefficients before each variable determine the number of additional patients who will be expected on that day given that variable being

4 ACADEMIC EMERGENCY MEDICINE January 2001, Volume 8, Number 1 51 present. For instance, if the day to be predicted is a Monday in January that is not the day after a holiday, the expected patient volume would be 129 ( ). Comparison of the two models demonstrated that inclusion of weather variables diminished the impact of day of the week and of winter season. However, it also included the months of July and August as predictive of lower patient volume. The final model, excluding the weather variables, contained only nine easily measured and predicted variables, in comparison with 13 variables in the alternative model. The equation using only calendar variables from the first phase of the study, with 7% added on for the increase in subsequent yearly volume, was then used to predict patient volume starting in September Based on this, the staffing pattern of providers was altered to allow for the predicted patient volume to be seen, given current provider productivity. Model predictions for September allowed the clinic to be appropriately staffed on 22 days (73%), understaffed on four days (13.5%), and overstaffed on four days (13.5%). On the four days that the clinic was understaffed, it was only by an average of 0.9 patients/hour. This small excess volume of patients can usually be easily managed by the providers who are working throughout the day. For the days that the clinic was overstaffed, it was only by an average of 1 patient/hour. This translated into an only 0.3 FTE (full-time equivalent) per week that was not being fully utilized. The prediction equation was continuously used to develop staffing patterns. It was expected that waiting times would decrease as staffing was better matched to patient volume. This would then improve patient satisfaction as well as decrease the number of patients who left without being seen. When the time period June 1998 to August 1999 (prior to prediction utilization) was compared with September 1999 to May 2000 (after prediction utilization), the percentage of patients who left without being seen dropped from 2.71% to 2.21%, an approximately 18.5% decrease. In addition, the formal patient complaint rate dropped from 0.208% to 0.145%, a 30% decrease. TABLE 2. Daily Patient Volume by Month Month Average Daily Patient Volume 95% CI January , 115 February , 108 March 90 84, 96 April 92 86, 99 May 89 81, 97 June 94 87, 101 July 91 85, 98 August 89 83, 95 September 96 89, 104 October 93 86, 99 November 94 87, 102 December 94 86, 103 Second-phase Model Development. In the second phase of the study, a model was developed with the intent of providing the ability to continue to predict patient volume while accounting for the ongoing increase in patients seeking care. The final regression equation was: daily patient volume = Newdate 52.0 Monday 44.2 Tuesday 39.2 Wednesday 30.2 Thursday 26.5 Friday 10.9 Saturday 12.2 February 3.9 March. The value for Newdate was created by sequential numbering of dates starting with a value of 0 at the first date in the data set, February 1, For example, October 28, 2000, would have 1,000 as a value for Newdate, since it is the 1000th date after February 1, Calculation of the Durbin Watson statistic 12 showed that there was not significant autocorrelation among these variables during model development. In addition, the final model was chosen based on Schwarz s Bayesian criterion (SBC) statistics 13 so as not to over fit the data, which would have decreased the forecasting accuracy. This final equation accounted for 72.7% of the variation in daily patient volume and had a p-value of <0.01. Using the second-phase model to test our validation set of data (May 1, 2000, to July 31, 2000) demonstrated that the model on average underpredicted patient load only by an average of two patients per day, with a standard deviation of 11 patients. The mean difference between predicted and actual patients was 3.0%, with a standard deviation of 11.6%. DISCUSSION The weekly pattern of visits that we observed, highest on Monday and thereafter declining steadily throughout the week, with the fewest patients seen on Sunday, is consistent with that demonstrated in other studies. 10,11,14 The study by Glass and Friedman found a 42% difference in average patient volume from the high on Monday to the low on Sunday. 14 This is virtually identical to the 41% difference that we detected. However, the monthly distribution of visits in our data differed from that in other studies. Diehl et al. noted increasing patient visits from May through August. 11 The study by Glass and Friedman, which studied trends in ED services, also found the monthly volume to be greatest in late summer and early fall and lowest throughout the winter, except for January. 14 In contrast, our peak months were November through

5 52 URGENT CARE Batal et al. PREDICTING URGENT CARE VISITS February, similar to Holleman et al., who found that the winter months were the busiest, aside from December. 10 We found an approximately 11% increase in patient volume after a holiday, identical to that in the study by Holleman et al. 10 The presumption is that more patients present to the clinic the day after a holiday because they prefer to delay needed care rather than disrupt their holiday time. Other explanations include illness resulting directly from holiday excess or family concern for previously neglected medical problems. Our study design did not allow us to distinguish between these alternate explanations. Weather variables contributed little to our firstphase model, only marginally increasing the amount of variability that could accurately be accounted for. It is interesting that the two weather variables that were significant contributors to our model (daily high temperature and snowfall) were the same variables that were significant in the model of Holleman et al. 10 Both this study and that of Diehl et al. 11 conclude that weather-related variables add little predictive value to the regression equation after consideration of calendar variables. This is probably due to the strong association of adverse weather with winter season, which was significant in both of our first-phase models. Our more streamlined model promotes ease of use of the prediction equation for determining staffing patterns since adequate prediction can occur without actually checking weather forecasts. Although the studies by Holleman et al. 10 and Diehl et al. 11 allude to using their derived equations to determine staffing levels, to the best of our knowledge, ours is the first published report to demonstrate the applicability of this as well as measure the effects of this on patients. Appropriate staffing can theoretically decrease waiting times. Prior ED studies have shown the relationship between increasing waiting times and the number of patients who leave without being seen. 3 Although our study was not able to directly measure patient waiting times, the accurate matching of provider staffing patterns with predicted patient volume did result in an 18.5% decrease in the number of patients who left without being seen (used as a surrogate marker of waiting times). This is similar to the results found in another study; reducing the length of stay can decrease by more than half the number of patients who leave an ED without being seen. 15 In addition, we were able to decrease by almost one-third our patient complaints. LIMITATIONS AND FUTURE QUESTIONS One limitation of this study is the generalizability to other clinical facilities. However, the similarities between our study s conclusions and that of others that have investigated these issues demonstrate the consistent and likely generalizability of the patterns patients show in presenting for care. In addition, a model whereby routinely collected data are continuously used to develop projects for process improvement (such as staffing patterns) is applicable to all clinical settings. However, evidence against generalizability is the study by Noble et al., 9 which showed discrepancies between institutions in the same city in regard to the effect of calendar and meteorological variables on patient volume. Thus, each institution needs to use the overall principles expounded here, but hone its applicability with data from its own experience. In order to further refine this model, it will be necessary to continually reevaluate the association between calendar and weather variables and patient volume. Future study in this area will also require focusing on uncovering additional determinants of patient volume to be included in the model. These independent variables will need to be easily measured far in advance to facilitate their utilization in future predictions. The next step will be to determine a model to predict the hourly variability in patient volume during a given day. This would allow for additional refinement of staffing patterns to meet patient needs. In addition, models could be developed to determine the projected need for various ancillary services, such as radiology. Increasing attention to the ability to predict patients presentation patterns and their needs will allow for continued improvement in patient satisfaction. CONCLUSIONS Our study demonstrates that careful consideration of the relationship of calendar variables to patient volume allows for fairly accurate prediction of future patient volume. These predictions can be used to optimize staffing patterns to allow timely access for the burgeoning number of patients who tend to receive their health care in these types of urgent care settings. 16 Emergency departments and urgent care clinics currently serve as America s health care safety net of last resort 17 and, as such, need to have adequate staff capacity to care for all who seek care. Accurate anticipation of this need will allow for better use of available personnel resources and improved efficiency of care. References 1. Schappert SM. National Ambulatory Medical Care Survey: 1991 summary. Vital Health Stat ; 116: Fernandes CMB, Dayor MR, Barry S, et al. Emergency department patients who leave without seeing a physician: the

6 ACADEMIC EMERGENCY MEDICINE January 2001, Volume 8, Number 1 53 Toronto Hospital experience. Ann Emerg Med. 1994; 24: Bindman AA, Grumbach K, Keane D, Rauch L, Luce JM. Consequences of queuing for care at a public hospital emergency department. JAMA. 1991; 266: Hall MF. Keys to patient satisfaction in the emergency department: results of a multiple facility study. Hosp Health Serv Admin. 1996; 4: Rhee K, Bind J. Perceptions and satisfaction with emergency department care. J Emerg Med. 1996; 14: Thompson DA, Yarnola PR. Relating patient satisfaction to waiting time perceptions and expectations: the disconfirmation paradigm. Acad Emerg Med. 1995; 2: Birnbaum A, Gallagher EJ, Utkewicz M, Gennis P, Carter W. Failure to validate a predictive model for refusal of care to emergency department patients. Acad Emerg Med. 1994; 1: Attia MW, Edward R. Effect of weather on the number and the nature of visits to a pediatric ED. Am J Emerg Med. 1998; 16: Noble JH, La Montagne ME, Bellotti C, Wechsler H. Variations in visits to hospital emergency care facilities: ritualistic and meteorological factors affecting supply and demand. Med Care. 1971; 9: Holleman DR, Bowling RL, Gathy C. Predicting daily visits to a walk-in clinic and emergency department using calendar and weather data. J Gen Intern Med. 1996; 11: Diehl AK, Morris MD, Mannis SA. Use of calendar and weather data to predict walk-in attendance. South Med J. 1981; 74: Durbin J, Watson GS. Testing for serial correlation in least squares regression. Biometrika. 1951; 37: Schwarz G. Estimating the dimension of a model. Ann Stat. 1978; 6: Glass R, Friedman D. Trends in the demand for emergency room services: the Mount Sinai Hospital. Mount Sinai J Med. 1977; 44: Fernandes CM, Price A, Christenson JM. Does reduced length of stay decrease the number of emergency department patients who leave without seeing a physician? J Emerg Med. 1997; 15: Rask KJ, Williams MV, Parker RM, McNagny SE. Obstacles predicting lack of a regular provider and delays in seeking care for patients in an urban public hospital. JAMA. 1994; 271: Johnson LA, Derlet RM. Conflicts between managed care organizations and emergency departments in Los Angeles. West J Med. 1996; 164: