"The American Journal of Surgery (2015) 209, 509-514 Midwest Surgical Association Prediction models of Medicare 90-day postdischarge deaths, readmissions, and costs in bowel operations a,b,c, a Donald E. Fry, M.D. *, Michael Pine, M.D., M.B.A. , a a David Locke, B.S. , Gregory Pine, B.A. a b Michael Pine and Associates, 1 East Wacker Drive #1210, Department of Surgery, Northwestern c University Feinberg School of Medicine, Chicago, IL, USA; Department of Surgery, University of New Mexico School of Medicine, Albuquerque, NM, USA Abstract KEYWORDS: BACKGROUND: The 90-day postdischarge morbidity and mortality rates following elective and Small bowel surgery; emergent bowel surgery remain poorly defined. Large bowel surgery; METHODS: The2009to2011Medicareinpatientfilesforpatientsundergoingelectiveandemergent Postdischarge small and large bowel operations in 1,024 hospitals that passed present-on-admission coding accuracy readmissions; standards had prediction models designed for inpatient mortality, prolonged postoperative length of Postdischarge deaths; hospital stay (prLOS), 90-day postdischarge mortality and readmissions, and total hospital costs. Risk-adjusted RESULTS: Of 118,758 patients studied, there was a 4.7% inpatient mortality rate and 7.3% prLOS outcomes; amonglivedischarges.Anadditional7,586deathsand26,969readmissionsoccurredwithin90daysof Control charts discharge. Prolonged preoperative and prolonged postoperative hospitalizations were significant (P, .0001) variables in predicting postdischarge deaths and readmissions. Total hospital costs were increased by over $18,000 per adverse outcome. CONCLUSION: Postdischarge deaths and readmissions are more common than inpatient adverse events of death and prLOS in elective and emergent Medicare large and small bowel operations. \u0000 2015 Elsevier Inc. All rights reserved. Acomprehensiveassessmentofoutcomesofsurgicalcare beenmuchmoredif?culttoidentifybecausepatientsarelost must include postdischarge events. Reports on inpatient to follow-up or may seek care from another hospital or mortality rates and speci?c complications of care have been physician. Most hospitals and surgeons do not know the in abundance for most operations. However, postdischarge completeresultsofpostdischargecare.Aswehavepreviously adverseevents(eg,readmissiontoanacutecarehospital)have reported,90-daypostdischargedeathsintheMedicarepopu- lationexceededthoseoccurringduringtheindexhospitaliza- tionforelectiveprocedures,andmanymorearereadmittedto 1 This is developmental research funded internally at MPA (Michael thehospitalduringthissame90-daytimeinterval. Ifoutcome Pine and Associates). No external funding has been used in this research. pro?lesexcludepostdischargeevents,thencareredesignand * Corresponding author. Tel.: 11-312-467-1700; fax: 11-312-467- improvementstrategieswillbesuboptimal. 1705. Major abdominal procedures of the small and large E-mail address: [email protected] bowel have substantial risks of major morbidity. Many of Manuscript received July 22, 2014; revised manuscript December 11, 2014 these complications are not identi?ed until the 0002-9610/$ - see front matter\u0000 2015 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.amjsurg.2014.12.005510 The American Journal of Surgery, Vol 209, No 3, March 2015 postdischarge period of time and, if they are suf?ciently Adverse outcomes are de?ned in this presentation to severe, will lead to readmissions and even deaths. Because include the following: (1) inpatient or 90-daypostdischarge of the evolving strategy of Medicare to impose hospital deaths; (2) risk-adjusted prolonged postoperative length of penalties for ‘‘excessive’’ readmissions in selected medical stay (prLOS) following the operation; and (3) readmission conditions, it is anticipated that these policies will soon be to an acute care hospital with 90-days of discharge. Models 2 extendedintosurgicalcare. However,predictionmodelsto for inpatient deaths following small and large bowel facilitate the identi?cation of preventable events among operations were designed by methods previously reported postdischarge deaths and readmissions have proven very for elective colon surgery except that POA risk factors are 6,7 dif?cult because risk factors appear to be different than included for emergency case evaluation. Because an 3 those identi?ed for inpatient morbidity. The conundrum extended preoperative LOS is associated with increased 8 for modeling the risk of readmissions is further com- postoperative morbidity, these cases have not been pounded in emergency operations because those conditions excluded as we have done previously and a risk factor that are present-on-admission (POA) versus those that are has been added for operations that occurred more than hospital acquired require accurate POA coding to delineate 2 days following admission. Hospital dummy variables risk factors from complications of care. are employed to account for hospital effects on ?nal Our group has developed a screening methodology to models. 4 identify the accuracy of hospital POA coding. This allows Linearrisk-adjustedLOSmodelsaredesignedusinglive the identi?cation of best hospitals for POA accuracy. Using discharges from the index hospitalization and employed the best coding hospitals permit the development of inpa- chronic disease risk factors that we have previously used, tient prediction models of adverse outcomes for both elec- but also incorporated POA variables and the extended tive and emergency small and large bowel operations, but preoperative LOS variable as candidates. The prLOS cases also when used with the Medicare Inpatient Limited Data are de?ned by using a moving average control chart by 9 Set permits the additional development of prediction methods we have previously published. Once prLOS out- models for postdischarge deaths and readmissions with liers are identi?ed, a logistic model is then created to pre- the most accurate administrative data. dict prLOS as an adverse outcome. The research database is then searched for the all-cause 90-day readmissions to acute care hospitals. Rehabilitation Methods and skilled nursing admissions are excluded as readmission events. The 90-day interval is chosen because preliminary The Centers for Medicare and Medicaid Services studies identi?ed readmissions within this time frame to be Inpatient Limited Data Set for 2009 to 2011 was used to reasonably linked to the index hospitalization. Rarely do identify patients undergoing elective and emergent small patients following small and large bowel operations have and large bowel operations that were identi?ed as part of elective admissions scheduled for unrelated indications Medicare Severity Diagnosis-Related Groups (MS-DRG) during this 90-day interval. The patient’s survival status is 329-331. The research database was re?ned to include only in the database for each year,and deaths within 90days but discharges 65 years of age or older with an International without readmission are identi?ed. All deaths that occur Classi?cation of Diseases 9th Revision-Clinical Modi?ca- during or following readmission are included in the tion code for one of the following procedures: 17.31 to readmission cohort for prediction modeling. 17.39,45.61,45.62,45.71to45.79,45.81to45.83,48.50to The 90-day deaths without readmission and the 90-day 48.59, 48.62, 48.63, 48.69, 48.75, 48.76, and 68.8. Cases readmission cases become the dependent variables in a were excluded if they had a missing patient identi?er, forward stepwise regression model that use all index hospi- hospital identi?er, principal diagnosis, or admission or talization risk factors, and also use prLOS from the index discharge date. They were also excluded if the patient was hospitalization as an additional variable. We then identi?ed transferred to another acute care facility, received from the MS-DRG for each readmission that occurred during the another acute care facility (including an ambulatory surgi- 90-daypostdischargeintervaltoestablishthosereadmissions cal center), or left against medical advice. Further re- thatwerenotlikelyrelatedtotheindexhospitalization. ?nements required that each hospital in the ?nal research Costs for each casewere computed by adjusting charges data set had 20 or more evaluable cases for the study for the index hospitalization using cost-to-charge ratios. period, and included only those that met an 80% or better The Medicare paid amount was used as the cost for 4 rate for POA coding accuracy. readmissions. A lower limit for hospital costs was derived In model development, only variables with P less than by computing each hospital’s 10th percentile cost of each .001 were used. Schwarz criterion is employed to avoid case and setting the lower limit equal to the 10th percentile 5 over-?ttingthemodels. The performanceofall?nalpre- of these costs at all hospitals. This lower limit is de?ned as diction models was evaluated by c-statistic computation. the required costs of care and any case with lower costs 2 Cost models were evaluated by R computations. All were winsorized to this level. For routine cases (ie, those analyses used SAS software Version 9.4 (SAS Institute, without any adverse event), the total hospital costs are Cary, NC). subtracted by the required costs and this difference inD.E. Fry et al. Postdischarge prediction models in bowel surgery 511 individualized routine costs became the dependent variable hospitalization.Those nonassociated readmissions included in a forward stepwise linear regression. Hospital dummy chemotherapy (n 5 354), major joint replacement (n 5 variables are used as dummy variables for cases performed 193), and hip/femur procedures (n 5 133) as the most in 2010 and 2011 to adjust for in?ation changes. Excess common. costs attributable to adverse outcomes were computed by Among all readmission patients, 3,856 died with total subtracting predicted total routine hospital costs (required postdischarge deaths of non-readmitted and readmitted 1 individualized costs) from the total costs of those cases patients to 7,586 (6.7% of live discharges). Deaths from with an adverse outcome. A predictive model for excess inpatient care (n 5 5,594), 90-day postdischarge deaths costsofadverseoutcomeswas derivedbyapplyingforward without readmission(n5 3,730), and 90-day postdischarge stepwise linear regression to the previously described deaths among readmitted patients (n 5 3,856) totaled candidate variables. 13,180 patients (11.1%). Total patients with one or more adverse outcomes of inpatient or postdischarge deaths, prLOS, and postdischarge readmissions were 40,108, or Results 33.8% of all cases. The prediction models for each of the 4 components of There were 118,758 patients from 1,024 hospitals in the the adverse outcomes are detailed in Table 2. The inpatient study that met coding standards and case volumes for the mortalitymodelhad42signi?cantriskfactors.Thec-statis- study period. There were 5,594 inpatient deaths (4.7%). tic was .850with hospital variables and .825 without hospi- Therewere8,207prLOSpatientsamonglivedischargesfor tal variables. The prLOS model had 26 signi?cant variables, a c-statistic of .782 with hospital variables, and an overall inpatient adverse outcome rate of 13,801 a c-statistic of .744 after hospital variables were removed. (11.6%). The 90-day postdischarge mortality model had 35 vari- A total of 3,730 patients died within 90 days of discharge but were not readmitted to any acute care ables, a c-statistic of .895, and a c-statistic of .872 after hospital. Of these 3,730 deaths, 786 were prLOS and removal of hospital variables. The 90-day readmission 2,944werenotprLOS.Atotalof26,969patients(23.8%of model for all readmissions had 39 signi?cant variables, a live discharges) were readmitted within 90 days of c-statistic of .668, and a c-statistic of .650 without hospital discharge for a total of 39,394 times. Among readmitted variables. Signi?cant variables identi?ed in all prediction patients, 3,606 were prLOS and 23,363 were not prLOS. models were advancing age, renal failure, liver failure The most common MS-DRGs for readmissions are identi- and obstructive lung disease, total abdominal colectomy, ?ed in Table 1. A total of 1,882 (4.8%) readmissions were and excess preoperative LOS. The prLOS outliers had an identi?ed as not being linked to adverse events of the index odds ratio of 4.26 in the prediction of postdischarge deaths Table 1 These are the most common MS-DRGs that are identi?ed in 39,394 readmissions MS-DRGs of readmission Description of readmission n (% of readmissions) 391–395 Miscellaneous or other GI diagnosis 3,770 (9.6%) 335–337, 388–390 Lysis of adhesions or GI obstruction 2,544 (6.5%) 862–864 Postoperative fever or infection 2,277 (5.8%) 870–872 Septicemia or sepsis 2,064 (5.2%) 682–684 Renal failure 2,003 (5.1%) 329–331 Major small/large bowel procedure 1,933 (4.9%) 640–641 Nutrition, metabolism, ?uid, electrolyte disorders 1,594 (4.0%) 371–373 Major GI disorder and peritoneal infection 1,295 (3.3%) 344–346 Minor small/large bowel procedure 1,242 (3.2%) 177–179, 193–195 Pneumonia/respiratory infection 1,224 (3.1%) 291–293 Heart failure and shock 1,167 (3.0%) 377–379 GI hemorrhage 1,026 (2.6%) 689–690 Kidney/urinary tract infection 967 (2.5%) 853–858 Infection with operating room procedure 949 (2.4%) 280–285, 308–310 Acute myocardial infarction or cardiac arrhythmias 853 (2.2%) 919–921 Complication of treatment 809 (2.1%) 299–301 Peripheral vascular disease (medical MS-DRGs) 633 (1.6%) 175–176 Pulmonary embolism 535 (1.4%) 061–066 Stroke or intracranial hemorrhage 442 (1.1%) 811–812 Red cell disorders 428 (1.1%) All groups are considered to be related to the index hospitalization when the patients underwent major small and large bowel surgery GI5 Gastrointestinal; MS-DRGs5 Medicare Severity Diagnosis-Related Groups.512 The American Journal of Surgery, Vol 209, No 3, March 2015 Table 2 Details of the important characteristics of the 4 measured components of the adverse outcomes of major small and large bowel surgery Prolonged 90-day postdischarge 90-day Inpatient deaths length of stay deaths: no readmission readmissions No. of adverse outcomes 5,594 (4.7%) 8,207 (6.9%) 3,730 (3.1%) 26,969 (22.7%) (3,856 deaths) Signi?cant variables 42 26 35 39 c-statistic with hospital variables .850 .782 .895 .668 c-statistic, no hospital variables .825 .744 .872 .650 Odds ratios Total abdominal colectomy 3.54 1.72 2.78 2.13 Procedure on 31 day of hospitalization 1.47 1.35 1.96 1.31 Age 75–84 2.11 1.22 2.22 1.14 AgeR85 3.88 1.33 5.36 1.33 prLOS for index hospitalization – – 4.26 2.13 prLOS5 postoperative length of hospital stay. andanoddsratioof2.13inthepredictionofreadmissionto assessment of clinical outcomes and risk models have an acute care hospital. beendevelopedforeachcomponentoftheadverseoutcome The required costs were $7,016 for the study population aswehavede?nedit.Witheffectiveriskadjustment,results in 2009. This required cost increased by $720 in 2010 and among different hospitals can be compared and objective by $1,928 in 2011. The ?nal predictive model for routine metricswillallowfordynamicmeasurementofoutcomesto costsinadditiontotherequiredcostshad79riskfactors.Its gauge the effectiveness of quality improvement initiatives. 2 R was.746whenhospitalvariableswereincludedand.348 Risk adjustment models have been developed for each when hospital variables were removed. Routine costs were component of our de?nition of an adverse outcome to increased by $12,532 in cases with greater than or equal to permit identi?cation of signi?cantly associated variables. 3 days of preoperative hospitalization, $9,618 for total When both elective and emergent operations are being abdominal colectomy, and $7,772 for laparoscopic- evaluated, it is critical that accurate POA coding permits assisted total colectomy/abdominoperineal resection. The differentiationofclinicalconditionsatthetimeofadmission ?nal model for the excess cost of adverse outcomes had to be designated as risk factors, and for hospital-acquired 2 17variablesandanR of.625withhospitalvariables.After conditions to be correctly identi?ed as complications of 2 removalofhospitalvariables,theadjustedR was.204.The care.Forthisreason,wehaveanalyzedoutcomesfromonly average excess cost for cases with adverse outcomes was about20%ofacutecarehospitalsthatmetobjectivecriteria $18,429. ofeffectivePOAcoding.Althoughtheuseofadministrative claims data has been criticized as not having the detail that might be achieved with clinical abstraction of the Comments medical record, POA coding has been shown to greatly enhancethediscriminationofpredictivemodels,andwillbe Valid assessments of surgical outcomes must have further enhanced with the use of preoperative numerical 10,11 objectivemetrics,reproducibleanalytic methods,andcover laboratorydata. Claimsdatahavethedecidedadvantage the entire episode of care until clinical recovery has been to identify postdischarge events that have remained elusive achieved. In an era where a premium has been placed on in most clinical abstractions. Claims data are a necessary shorter lengths of hospitalization for major operative resource to de?ne the excess costs of care when adverse procedures, assessments of inpatient mortality events and events occur. inpatient complications of care are not suf?cient measures IntheMedicarepopulation,majorsmallandlargebowel of outcomes. Many untoward events that are direct or operations have nearly a 34% rate of adverse outcomes. indirect consequences of the surgical procedure are not The frequency of deaths and readmissions in the 90 days identi?ed until the postacute period and must be included following discharge exceeded deaths and prLOS from the to cover the complete continuum of care. Thus, in this index hospitalization. Postdischarge adverse outcome pre- study, we have expanded the de?nition of an adverse diction models demonstrate signi?cant associations with outcome as inpatient or 90-day postdischarge death, risk- severe complications of inpatient care (eg, prLOS), but the adjusted LOS outlier, and 90-day readmission to an acute c-statistics clearly point to other risk factors that need to be care hospital. identi?ed in the prediction of readmission. LOS outliers This study has examined adverse outcomes of care in predict readmissions,butthemajorityofreadmissionswere both elective and emergent small/large bowel operations. not prLOS cases. Socioeconomic factors such as patient Effective risk adjustment is essential for the accurate education, distance from the index hospital, patientD.E. Fry et al. Postdischarge prediction models in bowel surgery 513 6. Fry DE, Pine M, Jones BL, et al. Surgical warranties to improve qual- compliance with postdischarge directives, and many other ity and ef?ciency in elective colon surgery. Arch Surg 2010;145: nonmedical variables are likely to in?uence readmission 647–52. rates. Better prediction models will provide evidence to 7. Fry DE, Pine M, Jones BL, et al. The impact of ineffective and inef- de?ne the patient at risk for readmission and enhance the ?cient care on the excess costs of elective surgical procedures. J Am development of focused strategies needed to reduce the Coll Surg 2011;212:779–86. 8. Vogel TR, Dombrovskiy VY, Lowry SF. In-hospital delay of elective expensive and morbidity of readmissions. surgery for high volume procedures: the impact on infectious compli- In a previous study, readmissions among Medicare cations. J Am Coll Surg 2010;211:784–90. surgical and medical inpatients identi?ed that a cumulative 9. Fry DE, Pine M, Jones BL, et al. Adverse outcomes in surgery: redef- total of more than 20% of readmissions had no outpatient inition of post-operative complications. Am J Surg 2009;197:479–84. 12 physician visits before the readmission event. Earlier and 10. PineM,JordanHS,ElixhauserA,etal.Enhancementofclaimsdatato improve risk adjustment of hospital mortality. JAMA 2007;297:71–6. more frequent follow-up with the operating surgeon, effec- 11. Fry DE, Pine M, Jordan HS, et al. Combining administrative and clin- tivehomehealthcaremanagement,improvedpatienteduca- ical data to stratify surgical risk. Ann Surg 2007;246:875–85. tion about avoiding readmission, and enhanced direct 12. Jencks SF, Williams MV, Coleman EA. Rehospitalizations among pa- communication (eg, frequent telephone contacts) may be tients in the Medicare fee-for-service program. N Engl J Med 2009; strategies for earlier recognition of postdischarge problems 360:1418–28. and avoidance of readmission. This study has limitations. First, the inclusion of Discussion emergency small and large bowel operations may include LOS outliers that were in?uenced by disposition issues that delayeddischargeratherthanmajorcomplications.Second, Dr Harry Reynolds (Cleveland, OH). How does your we have included all-cause readmissions in our analyses of model predict and, furthermore, how does Medicare or an prediction models. However, Table 1 identi?es that the insurer predict whataprolongedlengthof stayis?Howcan most frequent causes of readmission are linked to the index we use this model and data to guide us in our practice? Can hospitalization, and detailed examination of the readmis- we predict who will do poorly and alter our care paths to sion MS-DRGs indicates that 95% are reasonably associ- helpdecreaseourriskofperioperativemorbidityandmortal- ated with the index inpatient episode. Finally, our risk ity,orcanwetellsomepatientsthattheirperioperativeriskis adjustment methods are dependent on the completeness of just too prohibitive? Finally, what do you foresee Medicare coded diagnoses and require additional clinical information and insurers penalizing surgeons for in the future? And to enhance accuracy. howwillguidelines becreatedandpenaltiesassessed? In summary, the assessment of outcomes in major small Dr Fry: The method we use for predicting or for identi- and largebowel operations requires that the full duration of fyinglengthofstayoutliershasbeenpresentedbeforethisor- the episode must be included. Each component of the ganizationinthepastandisastatisticalprocesscontrolchart, adverse outcome in surgical care has different signi?cant whichisanestedseriesoflinearlogisticmodelsinacontrol risk factors and each will require individualized strategies chartmethodthat’sreallyquitecomplex,butwebelieveit’s for prevention and management. As the evolution of reallyquiteaccurate.Itidenti?eswhatistheexpectedlength ‘‘bundled’’ payment models evolve, surgeons and hospitals ofstayinpatientswithoutcomplicationandthenappliesitto will have ?nancial risk that will be linked to the entire the entire population of patients. We identi?ed patients that duration of the episode including events following arethreestandarddeviationsoutfromtheirpredictedlength discharge. This changing model of payment will require ofstay,soit’shardlytooexacting.Wehavefoundithasbeen an intensi?ed effort to not only manage the inpatient, but veryaccurateandtheninthenpredictingbadoutcomeswhen also postdischarge events. the patients are discharged. How’s this model going to be usedbythefedsorotherpeople,Godonlyknows.Theyprob- ablywon’tuseitatall.Wethinkthemodelhasvaluetoyouas a clinician, because I think the models allow you to predict References who is going to do badly, who is highest risk to come back inthehospital,whoarethehighriskpatientsfordyingafter 1. FryDE, Pine M,Pine G.Medicare post-discharge deathsand readmis- sions following elective surgery. Am J Surg 2014;207:326–30. discharge.Andbearinmind,3.1%ofthesepatientsdiedpost 2. CentersforMedicareandMedicaidServices:ReadmissionsReduction dischargewithin90dayswithoutbeingreadmittedtothehos- Program. Available at: http://www.cms.gov/Medicare/Medicare-Fee- pital at all, a fairly alarming ?gure to me. How are payers for-Service-Payment/AcuteInpatientPPS/Readmissions-Reduction-Pro going to penalize or reward physicians in the future gets to gram.html#. Accessed June 4, 2014. thecrux ofour modellingof prospectivepayment. Wehope 3. Kansagara D, Englander H, Salanitro A, et al. Risk prediction models for hospital readmission: a systematic review. JAMA 2011;306: therewon’tbepenalties.Wehopethattherewillbeprospec- 1688–98. tive bundled payments for the entire episode of care, 4. Pine M, Fry DE, Jones BL, et al. Screening algorithms to assess the including 90 days of post discharge care, and that if you in accuracy of present-on-admission coding. Perspect Health Inf Manag your hospital can bring in the care underneath that budget 2009;6:2. withgoodresultsandef?cientutilizationofresources,there 5. Schwarz GE. Estimating the dimension of a model. Ann Stat 1978;6: 461–4. will be a margin. And hopefully the hospital would share it"