We developed an early warning discharge disposition prediction tool to facilitate discharge planning and coordination, potentially reducing length of hospital stay and improving patient experience.
Objectives: To develop an early warning discharge disposition prediction tool based on clinical and health services factors for hospitalized patients. Recent study results suggest that early prediction of discharge disposition (ie, whether patients can return home or require placement in a facility) can improve care coordination, expedite care planning, and reduce length of stay.
Study Design: Retrospective analysis of inpatient data; development of multiple logistic regression model and an easy-to-use score.
Methods: We used retrospective data from all patients who were admitted in 2013 to the general medical service at the Veterans Affairs Boston Healthcare System and discharged alive. A derivation-validation approach was used to build a multiple logistic regression model, which was transformed into a score for potential implementation.
Results: Of the 4760 patients discharged in 2013, 485 (10.2%) were discharged to a facility other than home. Correlates of discharge to a facility included a primary admission diagnosis of neoplasm (odds ratio [OR], 2.71; 95% CI, 1.73-4.25), diseases of the nervous system (OR, 2.53; 95% CI, 1.26-5.08), and musculoskeletal diseases (OR, 2.55; 95% CI, 1.52-4.27), as well as discharge to a facility during previous hospitalization. Patients with a prior primary diagnosis of circulatory disorder and those with comorbidity of hypertension, either complicated or uncomplicated, were less likely to be discharged to a facility. A value of 5 or greater on the 20-point scale indicated discharge to a facility with 83% sensitivity and 48% specificity.
Conclusions: A validated, easy-to-use score can assist providers in identifying upon admission those patients who may not be able to go directly home after hospitalization, thus facilitating early discharge planning and coordination, potentially reducing length of hospital stay and improving patient experience.
Am J Manag Care. 2018;24(10):e325-e331Takeaway Points
We developed an early warning discharge disposition prediction tool to facilitate discharge planning and coordination, potentially reducing length of hospital stay and improving patient experience.
Although the Department of Veterans Affairs (VA) health system is not necessarily reimbursement-based, it strives to achieve high quality care and efficient resource utilization via evidence-based practice and continuous measurement and improvement. Effectively managing patient flow through VA medical centers requires the proactive identification of appropriate level of care and services. Early planning and coordination by the healthcare team, including physicians, social workers, rehabilitation specialists, and postacute care services, improves inpatient care quality and access while reducing costs.1,2 The ability to predict discharge disposition—whether a patient can return home or requires placement in a care facility—could expedite rehabilitation, improve coordination of care among consultants, prepare caregivers, and help community agencies plan for needed resources. It also helps reduce length of hospital stay, which, in turn, may mitigate the risks of hospitalization and improve patient recovery.3-7 Consequently, early, accurate, and effective discharge planning has emerged as a high priority for both patients and hospital systems.8
The primary objective of this study was to develop an approach that could serve as an early warning decision aid to care providers for predicting, within 24 hours of admission, the discharge disposition of hospitalized veterans based on the available clinical and health utilization factors at index and previous hospitalizations.
The study setting was the inpatient medical service at the West Roxbury campus of the VA Boston Healthcare System (VA-BHS). At the time of the study, discharge planning occurred within an interdisciplinary group consisting of a nurse case manager, a social worker, and a clinician representing the medical team. Consultations with other services, such as physical and/or occupational therapy, were made based on the individual assessments of these groups, mostly prompted by clinical judgment.
We obtained 4817 records (3187 were unique patients) from the corporate data warehouse of the VA-BHS for January 1, 2013, through December 31, 2013. After excluding 57 records (1.18%) with missing discharge disposition, the final data set consisted of 4760 records. We considered factors for inclusion in this study that would be readily available to the care team within 24 hours of admission.
Index admission. For the index admission, we ascertained demographic information (ie, age, sex, race, presence of non-VA health insurance, and marital status), clinical factors (primary diagnosis, number of diagnoses), source of admission, and specialty of the admitting ward. The primary diagnosis at the time of admission is not an available field in VA administrative data; however, the primary hospitalization diagnosis, recorded clinically at the time of discharge, is routinely transcribed from the admission history and physical notes and thus represents the likely primary diagnosis at the time of admission. These diagnoses were grouped into 19 categories based on International Classification of Diseases, Ninth Revision standard code groups. We counted all recorded discharge diagnoses associated with that admission and included this sum as the number of diagnoses. Admission sources included VA nursing home, VA domiciliary, transfer from other VA hospital, outpatient treatment, and other direct admissions (eg, walk-ins, directly admitted from home, transfers from non-VA facilities). Admitting ward specialties included general (acute medicine), cardiac intensive care unit (ICU), medical ICU, medical step down, telemetry, and hospice for acute care. We included ICU patients because the variables considered in predicting their disposition are comparable with those of patients admitted to non—critical care units. We further derived the 31 Elixhauser comorbidities using the updated International Classification of Diseases, Ninth Revision, Clinical Modification coding schema of Quan et al.9 This study was approved by the VA-BHS Institutional Review Board.
Historical factors. We also derived several historical clinical and health services factors using a 12-month “look back” into 2012 records of each of the unique patients. The clinical factors included the primary diagnosis of the immediately preceding hospital admission, the number of diagnoses on the immediately preceding admission, and the discharge disposition for the immediately preceding admission. The health services/utilization factors included the number of admissions in the 12 months prior to the index admission and an indicator of whether the index admission was an all-cause readmission within 30 days.
Main outcome measure. The main outcome was the patient’s discharge disposition, home versus facility. Patients discharged to the community, including those who were homeless and referred to a shelter, were considered discharged to “home.” Facility locations included VA nursing home care units (NHCUs), also known as Community Living Centers, and non-VA nursing homes.
Due to pragmatic limitations to conducting a prospective validation, we adopted the standard derivation-validation approach. Accordingly, we split the records randomly into 2 subsets of 70% and 30%. Using the 70% data subset (the derivation set; n = 3351), we built a case-mix—adjusted multiple logistic regression model of discharge to facility (using backward stepwise selection; P = .05), with sex forced into the model. Robustness of the selected predictors was confirmed through combined backward-forward and forward-only stepwise methods, and other P values.
After the first model was built, admission from a VA NHCU was found to have a disproportionately large odds ratio (OR) of approximately 145 for facility discharge, attributable to the fact that nearly all patients admitted from an NHCU (40/42; 95%) were discharged to a facility. Consequently, admission from an NHCU was deemed the major, and sufficient, factor for the hospital care team to deem a facility discharge. The final multiple logistic regression model we present next does not include NHCU admission as a predictor.
Once the final model was derived, we used the remaining 30% of the data (the validation data set; n = 1409) to estimate the model’s predictive power based on area under the operating curve (AUC) values. A score for clinical application was derived from the final model. Using the standardized logistic regression coefficients, each factor in the final model was assigned a relative weight on a 20-point scale; the score for a patient would then be the sum of these weights, if the corresponding factors were present for that patient. The 20-point scale was chosen as the best balance between discrimination (to allow enough separation between factors) and ease of use. These weights were rounded to integer values for ease of addition in practice. In deciding whether to round each weight up or down, our objective was to maximize the correlation between the total score for a patient and the predicted probability of discharge to facility for that patient from the logistic regression model (across all records in the derivation data set). Continuous predictors (age and number of diagnoses) were separated into categories based on the distribution of the data; the weight assigned to each of these was divided among the corresponding categories, and these divisions were considered as additional decision variables in the optimization. The sum of all the weights in the score was constrained to be 20 (ie, if a patient exhibited all predictors from the final model, then the patient would be assigned a score of 20). Details of the scoring methodology and optimization are presented in eAppendix A (eAppendices available at ajmc.com).
Because the coefficients in the logistic regression model, and the corresponding factor score weights, could be either positive (predictive of discharge to a facility) or negative (protective factors), we separated them into 2 positive additive subtotals in the scoring tool, 1 for the predictive factors and 1 for the protective factors. The latter was then subtracted from the former to get the final score for a patient. Clinical practitioners were consulted in order to ensure that the questions relating to each factor from the model were phrased in a way that would make sense to a care provider. We then determined a threshold value beyond which the patient was likely to go to a facility, considering an acceptable sensitivity and specificity of the score suggested by our medical collaborators. SAS version 9.4 (SAS Institute; Cary, North Carolina) was used for all statistical analyses.
Of the 4760 patients in the full data set, 485 (10.2%) were discharged to a facility, which included a VA nursing home (n = 301), VA medical center (n = 129), community nursing home (n = 53), VA domiciliary (n = 1), and other government hospital (n = 1).
Table 1 [part A and part B] indicates that demographic variables such as age, marital status, and white race were independent predictors of the patient’s disposition location. Several other factors related to admitting source, admitting ward, clinical diagnosis, and comorbidities were also significant.
Table 2 highlights that a previous admission diagnosis of diseases of the nervous system, diseases of the circulatory system, or external causes of injury and supplemental classification, as well as total number of diagnoses, were independent historical clinical predictors of the discharge disposition for the index admission. Number of previous admissions in the past 6 months and previous discharge to a community hospital, VA medical center, or VA NHCU were independent health services predictors of the discharge disposition for the index admission.
Table 3 shows the case-mix—adjusted logistic regression model, along with the ORs and 95% CIs based on the derivation data set; ORs greater than 1 indicate higher chance of discharge to a facility, whereas ORs less than 1 indicate higher chance of going home. The AUC of this model was 0.75 for the derivation data set; using the validation set, the AUC was 0.74.
Among diagnoses on admission of the index hospitalization, neoplasms (OR, 2.71; 95% CI, 1.73-4.25), diseases of the nervous system (OR, 2.53; 95% CI, 1.26-5.08), and diseases of the musculoskeletal system and connective tissue (OR, 2.55; 95% CI, 1.52-4.27) were associated with discharge to a facility. In contrast, historical primary diagnosis of circulatory system disease was associated with lower likelihood of discharge to a facility (OR, 0.54; 95% CI, 0.35-0.83), as were the presence of both uncomplicated hypertension (OR, 0.62; 95% CI, 0.47-0.80) and complicated hypertension (OR, 0.31; 95% CI, 0.20-0.48) during prior hospitalization. The previous primary diagnosis of external causes of injury and supplemental classification indicated a higher likelihood of patient discharge to a facility (OR, 2.58; 95% CI, 1.73-3.84), as did the comorbidity of other neurological disorders (OR, 1.70; 95% CI, 1.16-2.50).
The 3 previous discharge disposition locations of community hospital, VA medical center, and VA NHCU were associated with high ORs (10.33, 4.21, and 3.59, respectively), indicating that if the patient had been discharged to one of these locations after the prior admission, then this patient was very likely to go to a facility again upon discharge from the index hospitalization.
The score developed for clinical application is shown in eAppendix B. The weights assigned to the factors in the score achieved an 84% correlation with the logistic regression model probabilities for the derivation set. At a classification threshold of 5 points, the score achieved a sensitivity (number of discharges to a facility correctly identified as such) and specificity (number of home discharges correctly identified as such) of 83% and 46%, respectively (Figure). When tested on the validation set, the score with this threshold achieved 82% sensitivity and 48% specificity, suggesting that the score is robust in predicting discharge to a facility.
In this retrospective study of an entire year’s acute care hospitalizations at a tertiary care VA medical center, we identified variables that predict, at the time of admission, a patient’s likely discharge to a facility. We found that nearly all patients admitted from a VA NHCU were discharged to a facility. Using a derivation-validation approach, we determined that older patients; those admitted with neurologic, oncologic, and musculoskeletal primary diagnoses; those with higher numbers of diagnoses; and those previously hospitalized and discharged to a VA medical center were more likely to be discharged to a facility during the index hospitalization. In contrast, those with hypertension and those with a prior hospitalization with a primary diagnosis of a circulatory disorder were less likely to be discharged to a facility in the index hospitalization. These findings, validated in a subsequent logistic regression model, led to the creation of a clinically relevant score that can be used to predict at the time of admission, with good sensitivity and acceptable specificity, discharge to a facility.
Several studies have focused on predicting discharge among specific patient populations, such as stroke, traumatic brain injury, total joint arthroplasty, geriatric, trauma, or cardiac surgery2,10-14; some have proposed various prediction tools.15-18 The Barthel scale19 was developed to assess improvement of stroke patients throughout rehabilitation by periodic repetitions of the test, whereas the Blaylock Risk Assessment Screening Score index20 was developed for use in elderly patients within 48 hours of admission. Recently, Simonet et al developed a validated score predicting risk of discharge to a postacute care facility for general medicine patients, both on day 1 (admission) and day 3 of hospital stay.5 Although their study had fewer than 400 patients, most likely due to its prospective nature, 3 of 5 significant factors in their day 1 model were similar to our findings: age, number of diagnoses, and admission source. Unlike the present study, they did not find historical health utilization factors (ie, hospital and emergency department visits in the past 3 months) to be independent predictors. In further contrast, they used an aggregate comorbidity index (Charlson index) not found to predict discharge disposition, whereas we used the more granular set of 31 Elixhauser comorbidities, 3 of which were part of the final adjusted model. We acknowledge that some of the comorbid conditions occurred with relatively low prevalence in this population and, as such, the comorbid conditions in the final model should be validated in future study.
Our study is among the first to predict discharge disposition among general internal medicine VA patients. It is also among only a few to deliver an implementable predictive tool based on the statistical analysis conducted in the study. Although current clinical factors, such as diagnosis and comorbidities, typically form the basis of a care provider’s instinctive prediction of the eventual disposition location, we uncovered a few historical clinical and health utilization factors that also seem to play an important role in this decision-making process in practice.
This tool could be used by the patient’s care team at admission to classify the patient as one to potentially be discharged to a facility. This information would provide an early indicator to the patient and the care team at the beginning phase of the patient’s stay, when clinical information is sparse and before other clinical bedside data and functional status assessments have been collected. Having this “early warning” would facilitate discharge planning, as the discharge coordinator could initiate discussion with the patient soon after admission regarding the potential for placement in a facility and begin evaluating alternatives. It is worth noting that the purpose of such an early warning system is to help the care team anticipate the potential workload of discharge planning, to promote and enhance communication between providers and the patient, and to expedite overall care coordination associated with discharge to a facility, and not to identify an optimal discharge within those 24 hours upon admission.
In terms of the diagnoses found in this study to be significantly associated with discharge to a facility, many were consistent with what clinicians would expect among very ill patients who would likely need postacute care in a nursing, rehabilitation, or specialized facility, such as those with metastatic cancer, paralysis, and psychoses. Conversely, the factors that seemed to be protective (ie, predictive of a discharge to home)—for example, having a diagnosis of hypertension, either complicated or uncomplicated—may simply be an indicator of the absence of more serious or complicated diagnoses.
Lowering the threshold, or cut point, in the prediction model allows for increased sensitivity at the expense of decreased specificity. The implications of these trade-offs are worth considering further. For instance, although an increase in sensitivity would predict accurately at admission a larger proportion of patients to be discharged to a facility, a number of these would also be false positives (predicting discharge to a facility when actually the patient goes home). This may be due to the evolving clinical and functional status information, along with other factors (eg, patient/family choice, patient’s social network, availability of beds at postacute care settings, etc), during the patient’s stay. In all of those false-positive cases, the care provider team would have begun planning for a discharge to a facility on day 1 and continued their effort until it became apparent that the patient may be able to realize a discharge to home. This process induces overutilization of scarce resources, including expert discharge planning and physical therapy consultations. In general, however, our sense is that most clinicians and case managers would favor being surprised by a patient going home rather than being surprised by a patient requiring discharge to a rehabilitation or skilled nursing facility, as planning for the latter normally requires a longer lead time.
Our study findings must be viewed in the context of several limitations. First, although we accessed rich clinical data from the corporate data warehouse, this data source did not include patients’ activities of daily living, income level, or level of social support. We used the number of diagnoses as a proxy to estimate patient acuity and marital status as a proxy to estimate available social support. Second, in determining the primary diagnosis and number of diagnoses for a patient admission, the recorded discharge diagnoses associated with that admission were used. At this facility, the primary diagnosis at admission is routinely transcribed as the discharge diagnosis; however, this approach may have resulted in our using some diagnoses that were not actually known at the time of admission. Because discharge diagnosis by definition is not available on admission, the model and scoring system use the active diagnosis identified on admission as a proxy. Third, although we could not identify the eventual disposition of 129 patients transferred to another VA medical center, we classified them as discharged to a facility, as the eventual disposition was not germane to the disposition outcome of the index hospitalization. Finally, to develop and validate our early warning approach, we focused on a specific cohort of veterans hospitalized on a general medicine unit at one VA facility. We believe that the tool may be potentially generalizable to similarly sized VA hospitals, but further study is needed for validation in this population as well as for generalizability to other settings and populations.
In order to preempt potential delays in discharging patients from an acute inpatient facility, we attempted to identify factors apparent at the time of a patient’s admission to a general medicine unit that are statistically significant predictors of eventual discharge to a facility, as opposed to discharge to home. Using a year’s worth of clinical data from a tertiary care VA medical center, we used a derivation-validation approach and identified several such factors. From the final model, we developed a point-based scoring tool that can be readily implemented in clinical practice with minimal disruption.
Our study is both confirmatory and exploratory. We confirm previous research conducted on nonveteran populations and among patients with specific medical conditions that demonstrated the ability to develop a model to predict discharge disposition at admission using readily available factors. We explored the statistical significance of several historical clinical and health services factors that are frequently considered in clinical practice; many of these variables were independent predictors, and 2 were indeed part of the final adjusted model. We also developed a predictive score; future study should explore the validation of this score in actual clinical practice.
This research was supported by a contract from the New England Veterans Engineering Resource Center (NE-VERC). The authors appreciate the support from Maria Lebedeva, PhD, who was the primary contact at NE-VERC and facilitated data collection and collaborations with the VA Boston collaborators.
Coauthor Michael Donlin, MSN, ACNP-BC, FHM, died April 3, 2017.Author Affiliations: Department of Biomedical, Industrial, and Human Factors Engineering (NB, PJP), and Department of Surgery, Boonshoft School of Medicine (PJP), Wright State University, Dayton, OH; Medical Service (MD, SRS), and Geriatrics and Extended Care Service (SRS), VA Boston Healthcare System, Boston, MA; New England Veterans Engineering Resource Center (EKM), Boston, MA.
Source of Funding: New England Veterans Engineering Resource Center.
Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article. The views expressed in this article are those of the authors and do not necessarily represent the views of the Department of Veterans Affairs.
Authorship Information: Concept and design (NB, PJP, MD, EKM, SRS); acquisition of data (NB, MD, EKM, SRS); analysis and interpretation of data (NB, PJP, MD, EKM); drafting of the manuscript (NB, PJP, MD); critical revision of the manuscript for important intellectual content (NB, PJP, SRS); statistical analysis (NB, PJP); obtaining funding (PJP); administrative, technical, or logistic support (SRS); and supervision (PJP).
Address Correspondence to: Pratik J. Parikh, PhD, Wright State University, 207 Russ Center, 3640 Col Glenn Hwy, Dayton, OH 45435. Email: firstname.lastname@example.org.REFERENCES
1. Schlegel DJ, Tanne D, Demchuk AM, Levine SR, Kasner SE; Multicenter rt-PA Stroke Survey Group. Prediction of hospital disposition after thrombolysis for acute ischemic stroke using the National Institutes of Health Stroke Scale. Arch Neurol. 2004;61(7):1061-1064. doi: 10.1001/archneur.61.7.1061.
2. de Guise E, LeBlanc J, Feyz M, Lamoureux J. Prediction of outcome at discharge from acute care following traumatic brain injury. J Head Trauma Rehabil. 2006;21(6):527-536.
3. Meijer R, van Limbeek J, Kriek B, Ihnenfeldt D, Vermeulen M, de Haan R. Prognostic social factors in the subacute phase after a stroke for the discharge destination from the hospital stroke-unit: a systematic review of the literature. Disabil Rehabil. 2004;26(4):191-197.
4. Ekstrand E, Ringsberg KA, Pessah-Rasmussen H. The physiotherapy clinical outcome variables scale predicts length of hospital stay, discharge destination and future home facility in the acute comprehensive stroke unit. J Rehabil Med. 2008;40(7):524-528. doi: 10.2340/16501977-0210.
5. Simonet ML, Kossovsky MP, Chopard P, Sigaud P, Perneger TV, Gaspoz JM. A predictive score to identify hospitalized patients’ risk of discharge to a post-acute care facility. BMC Health Serv Res. 2008;8:154-162. doi: 10.1186/1472-6963-8-154.
6. Tanaka M, Yamamoto H, Kita T, Yokode M. Early prediction of the need for non-routine discharge planning for the elderly. Arch Gerontol Geriatr. 2008;47(1):1-7. doi: 10.1016/j.archger.2007.06.002.
7. Hagino T, Ochiai S, Sato E, Watanabe Y, Senga S, Haro H. Prognostic prediction in patients with hip fracture: risk factors predicting difficulties with discharge to own home. J Orthop Traumatol. 2011;12(2):77-80. doi: 10.1007/s10195-011-0138-y.
8. Cherlin EJ, Curry LA, Thompson JW, et al. Features of high quality discharge planning for patients following acute myocardial infarction. J Gen Intern Med. 2013;28(3):436-443. doi: 10.1007/s11606-012-2234-y.
9. Quan H, Sundararajan V, Halfon P, et al. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med Care. 2005;43(11):1130-1139. doi: 10.1097/01.mlr.0000182534.19832.83.
10. Brauer SG, Bew PG, Kuys SS, Lynch MR, Morrison G. Prediction of discharge destination after stroke using the Motor Assessment Scale on admission: a prospective, multisite study. Arch Phys Med Rehabil. 2008;89(6):1061-1065. doi: 10.1016/j.apmr.2007.10.042.
11. Pohl PS, Billinger SA, Lentz A, Gajewski B. The role of patient demographics and clinical presentation in predicting discharge placement after inpatient stroke rehabilitation: analysis of a large, US data base. Disabil Rehabil. 2013;35(12):990-994. doi: 10.3109/09638288.2012.717587.
12. Cuthbert JP, Corrigan JD, Harrison-Felix C, et al. Factors that predict acute hospitalization discharge disposition for adults with moderate to severe traumatic brain injury. Arch Phys Med Rehabil. 2011;92(5):721-730.e3. doi: 10.1016/j.apmr.2010.12.023.
13. Sharareh B, Le NB, Hoang MT, Schwarzkopf R. Factors determining discharge destination for patients undergoing total joint arthroplasty. J Arthroplasty. 2014;29(7):1355-1358.e1. doi: 10.1016/j.arth.2014.02.001.
14. Gabbe BJ, Cameron PA, Wolfe R, Simpson P, Smith KL, McNeil JJ. Predictors of mortality, length of stay and discharge destination in blunt trauma. ANZ J Surg. 2005;75(8):650-656. doi: 10.1111/j.1445-2197.2005.03484.x.
15. Barsoum WK, Murray TG, Klika AK, et al. Predicting patient discharge disposition after total joint arthroplasty in the United States. J Arthroplasty. 2010;25(6):885-892. doi: 10.1016/j.arth.2009.06.022.
16. Wachtel TJ, Fulton JP, Goldfarb J. Early prediction of discharge disposition after hospitalization. Gerontologist. 1987;27(1):98-103.
17. Beaulieu RA, McCarthy MC, Markert RJ, Parikh PJ, Ekeh AP, Parikh PP. Predictive factors and models for trauma patient disposition. J Surg Res. 2014;190(1):264-269. doi: 10.1016/j.jss.2014.02.032.
18. Pattakos G, Johnston DR, Houghtaling PL, Nowicki ER, Blackstone EH. Preoperative prediction of non-home discharge: a strategy to reduce resource use after cardiac surgery. J Am Coll Surg. 2012;214(2):140-147. doi: 10.1016/j.jamcollsurg.2011.11.003.
19. Mahoney FI, Barthel DW. Functional evaluation: the Barthel Index. Md State Med J. 1965;14:61-65.
20. Blaylock A, Cason CL. Discharge planning: predicting patients’ needs. J Gerontol Nurs. 1992;18(7):5-9.