Clinically Meaningful Laboratory Protocols Reduce Hospital Charges Based on Institutional and ACS-NSQIP® Risk Calculators in Hepatopancreatobiliary Surgery

MOTIVATED BY THE exponentially rising cost of health care and increased awareness of wasted resources as a substantial contributor, the modern health-care system has seen a dramatic shift toward high-quality value-based care.1 Bundled payments for health-care conditions and tiered reimbursements based on hospital performance metrics are becoming quickly adopted, driven in large part by the establishment of the Hospital Value-Based Purchasing Program in 2013 for Medicare reimbursement by the Patient Protection and Affordable Care Act. Whereas system-wide changes to standardize patient care pathways and reduce waste are often emphasized, there is gaining recognition in the hidden waste and potential for significant cost savings in simple dayto-day decision-making.

Routine daily laboratory testing is an underrecognized but substantial contributor to a patient's daily hospital charges.2' 3 The previous literature has shown that unnecessary expenditure often results from overutilization through both repetitive testing without clinical indication and prompting of further testing for clinically insignificant results.2,4'5 Routine daily laboratory testing has been proven to lead to overdiagnosis and unnecessary intervention, cause clinically significant iatrogenic anemia, and reduce patient rest and satisfaction due to frequent and often early morning interruptions.6, 7

Multiple studies have demonstrated reduction in laboratory testing and associated costs after implementing protocols for different patient populations; however, these systems are almost universally applied to an entire cohort of patients without targeted risk assessment to determine a subset of patients who could appropriately receive a limited-laboratory protocol.2, 5' 8' 9 The explosion of predictive analytical tools for clinical outcomes has suggested its expansion to improve economic outcomes by associating patient risk with resource allocation.10 Unfortunately, this application remains underexplored and specifically has not been applied to routine postoperative laboratory testing.

At our institution, patients undergoing major hepatopancreatobiliary (HPB) operations have historically had full laboratory tests ordered every day until discharge, regardless of health or recovery status. A routine battery of laboratories included a complete blood count (CBC), basic metabolic panel (BMP), liver function tests (LFTs), magnesium and phosphorus levels, and a coagulation survey. An inquiry for institutional charge data for this set of laboratories revealed a daily charge of $1,463 (raw charge before insurance or other cost adjustments). With an average length of stay (LOS) of 9.8 days and 5.5 days for pancreaticoduodenectomy (PD) and hepatectomy, respectively, this would result in an estimate of $14,337 for total laboratory charges per patient undergoing PD and $8,046.50 per patient undergoing hepatectomy. Therefore, an interdisciplinary team worked to develop the clinically meaningful laboratory (CML) protocol to standardize routine testing and reduce laboratory-associated charges with patient-centered risk stratification using predictive analytics.

The ACS-NSQIP® Surgical Risk Calculator is a well-accepted risk tool for predictive postoperative risk and is used by many surgeons for patient selection as well as to guide perioperative management for elective and emergent surgery.13 Unfortunately, for high-risk HPB operations, the ACS-NSQIP® surgical calculator has been shown to underestimate certain postoperative outcomes for individual patient risk.16 Our center recently published comparisons of institutionally derived predictive tools to the ACSNSQIP® surgical calculator for PD and hepatectomy17'18 The brier scores results from these studies showed strong and equivalent aggregate predictive ability between the institutional models and the NSQIP® calculator; however, the institutional models demonstrated statistically superior prediction for individual patient risk of nearly all studied outcomes including renal failure, cardiac complication, urinary tract infection, any complication, discharge to a nursing home/rehabilitation facility, and death. These models have been integrated into an institutional Research Electronic Data Capture (REDCap)™ database for HPB surgery as a novel method of integrating risk stratification and tracking of outcomes for outcome auditing.19, 20

The purpose of this study was to evaluate the effectiveness of predictive analytics to risk stratify patients into the tiered CML protocol to reduce patient charges for laboratory testing. Because of the need for generalizability as well as population-specific risk assessment in different health-care environments, we aimed to demonstrate the application of both the NSQIP®-based and institutionally derived predictive models to target appropriately low-risk patients for a limited laboratory protocol.

Methods

Study Design

The CML protocols were developed as prospective single-arm quality improvement initiatives for HPB surgery patients. Before the study, patients undergoing PD or hepatectomy under an enhanced recovery after surgery (ERAS)® program were being assessed for postoperative risk based on the institutionally derived predictive models previously described.17,18 Therefore, beginning in March 2016, these risk models were implemented into the CML protocol for risk stratification. For patients undergoing distal pancreatectomy (DP), however, an independent predictive model had not yet been developed. Therefore, this patient population was selected for implementation of the ACS-NSQIP®-based model for all patients undergoing scheduled DP beginning in January 2018.

Institutionally Based Predictive Analytics for PD and Hepatectomy

Data collection

The institutionally based predictive models for PD and hepatectomy have been previously described in detail.17,18 These models were developed from a historical cohort of 400 PD and 136 hepatectomy cases performed at our institution from 2008 to 2014. Patient demographic and preoperative health data as well as postoperative surgical outcomes were input into an institutional REDCap™ database, a secured web-based application designed to support data capture for research studies.12

Institutional-based predictive model

Major surgical outcomes were identified for each operation as endpoints for the risk calculator, including 30-day major complications (defined as grade 3 or greater on the Clavien-Dindo classification scale21, 22), 30-day readmissions, and 30-day mortality, as well as 10 to 12 specific surgical outcomes for each operation. For each operation-specific surgical outcome, predictive models were developed based on preoperative variables and multiple independent logistic regression models were constructed to assess the model's predictive capacity. Preoperative variables were defined as either binary, categorical, or continuous, as appropriate. All surgical outcomes were coded as binary outcomes. Bivariate analyses were performed using STATA 13 statistical analysis software (StataCorp, College Station, TX). Candidate variables were excluded from the model at a significance level of P < 0.25. All remaining significant variables were then used to construct a multiple logistic regression model for one dependent surgical outcome. This process was repeated for each of the 12 PD and the 10 hepatectomy surgical outcomes to generate individualized logistic regression models. Stepwise backward elimination was used to eliminate nonsignificant variables from the models at P < 0.10. Hosmer-Lemeshow goodness-of-fit tests were performed on each model to determine whether the model was a good fit for the data at P < 0.05. Models were analyzed using bootstrapping analysis in STATA to assess for internal validity. To assess the model's predictive and discriminative ability, Brier and receiver operating characteristic values were calculated.17, 18

REDCap™ integrated risk assessment platform

The mathematical equations derived from the regression models were programmed into the REDCap™ platform by using a data dictionary. For each series of patient variables input, the predicted risk of all surgical outcomes for PD or hepatectomy was generated and displayed. To transform these multiple risk models into simple decision-making, three major categorical risk levels were defined based on risks of 1) 30-day mortality, 2) 30-day readmission, and 3) any major complication represented by each of the other defined surgical outcomes. For each outcome category "low risk" and "high risk" were defined as a predicted risk of outside of two standard deviations from the mean, respectively, and any predicted risk between the cutoff as "average risk." All patients scheduled for PD or hepatectomy during the study period were predicted to be of low risk in all three categories qualified for stratification to the CML protocol.

Prospective risk category assignment

Qualifying patients were assigned to a color-coded tier based on health condition severity according to the HPB Surgery Laboratory Criteria tool (Fig. 1). Based on this algorithm, patients who were without major comorbidity were assigned to the Green Laboratories category and those with a health condition amenable to limited daily laboratories were assigned to Yellow Laboratories. Red Laboratories were assigned to any patient with specific operative factors according to the protocol. A full description of the CML protocol is described shortly.

ACS-NSQIP®-Based Predictive Analytics for Distal Pancreatectomy

Data collection

A prospective single-arm interventional study was then started in January 2018 to evaluate the use of the ACS-NSQIP® Risk Calculator to stratify patients undergoing DP into a similar CML protocol. Data were prospectively collected for a 12-month study period after implementation. All patients scheduled for DP as the primary operation were prospectively enrolled to the NSQIP®-based laboratory protocol. No patients were excluded. Patient data were collected by the enhanced recovery clinical nurse leader from the electronic medical record and input into the freely accessible webbased ACS-NSQIP® Surgical Risk calculator.23 The website was accessed by using on-site secure computers with no input of patient identifiers.

NSQIP®-based Predictive model

For a select operation, the NSQIP® calculator generates risk percentages for multiple independent and combined outcomes, including any complication, serious complications, various cardiopulmonary or infectious complications, readmission, and death. Each risk category generates a quantitative risk per cent as well as a qualitative risk assessment compared with the average risk for that operation, such that each category is assigned "above average risk," "average risk," or "below average risk." For each patient in the study, patient characteristics were input into the ACS-NSQIP® risk calculator based on the operation "DP with or without splenectomy; without pancreaticojejunostomy" (Current Procedural Code 48140).

NSQIP®-based laboratory protocol

For the purposes of this study, three select outcomes provided by the NSQIP® Surgical Risk Calculator were used to stratify risk for the laboratory protocol: serious complication, readmission, and death (Fig. 2). Patients assigned "below average risk" for all three select categories were assigned to Purple Laboratories and received no scheduled laboratories. Any patients predicted to be "average risk" for at least one category were assigned to limited Green Laboratories and any patients assigned at least one "above average risk" were assigned to full daily Red laboratories. Of note, all patients stratified to the Purple or Green Laboratory pathway based on overall risk, but who had a chronic comorbidity (e.g., renal insufficiency, diabetes, and congestive heart failure) which would require specific laboratory monitoring, were reassigned to the comorbidity-guided Yellow Laboratories according to the HPB Surgery Lab Criteria tool (Fig. 2).

CML Protocol

Protocol design and implementation

For both predictive models, a similar CML protocol was used based on the color-coded risk category assigned to each patient (Fig. 3). Patients stratified to Purple laboratories (available for DP patients only) received no scheduled laboratories postoperatively. Patients assigned to Green Laboratories were scheduled to receive no laboratories on postoperative day (POD) 1, a full set of laboratory tests on POD 2, and a creatinine level check alone on POD 5. A full set of laboratory tests consisted of a CBC, BMP, LFT, magnesium level, phosphorus level, and coagulation survey. Patients assigned to Yellow Laboratories received the limited set of daily laboratories according to their chronic health condition in addition to a full laboratory series on POD 2 and creatinine on POD 5. Patients assigned to Red Laboratories from either predictive model received a full set of laboratories every day beginning on POD one. The CML categories were able to be escalated at any point during hospitalization because of a change in the patient condition, warranting additional laboratory evaluation. All laboratory tier changes were tracked by POD.

Patient charges comparison

Each patient's color-based risk category was tracked and recorded once daily until hospital discharge, including any change in color category, reason for change, hospital LOS, and laboratory tests ordered. At the conclusion of the study period, estimated laboratory charges were calculated for each patient based on daily CML color category and hospital LOS. Patient charge data for each component of a full set of laboratories (CBC, BMP, LFT, magnesium, phosphorus, and coagulation panel) as well as select limited laboratories (e.g., potassium, bilirubin, and creatinine) were acquired from our institution's laboratory and financial department. A "CML charge" was calculated based on the cost of the actual laboratories obtained each day as prospectively tracked by the nursing leader. A "historical standard laboratory charge" for each patient was then calculated based on a set of full laboratories being performed daily for their entire LOS. The difference between the historical standard laboratory charge and the CML charge was defined as the "estimated cost savings" for each patient.

Clinical outcomes comparison

All patients in the study participated in the ERAS® protocol, which includes prospective comprehensive tracking of patient outcomes including complications, readmissions, and mortality. Outcomes are input into an institutional ERAS® Interactive Audit System (EIAS) database, which is the officially recognized outcomes tracking and auditing system developed by the ERAS® Society in 2006 for enhanced recovery programs worldwide.24 For the purpose of this study, three clinical outcomes were analyzed: major complications (defined as grade 3 or greater by the Clavien-Dindo classification scale21' 22), morality within 30 days postoperatively, and hospital readmission within 30 days after discharge. The ERAS® program for all three operations in this study were well established approximately one to two years before the CMLs protocols were begun. Therefore, outcomes for the intervention group (under CML protocol) for each operation were compared with a matching number of patients undergoing each operation immediately before the CML protocol initiation without bias from the impact of ERAS® on outcomes.

Results

Institutional-Based Model

Of the 568 patients scheduled for PD (n = 254) or hepatectomy (n = 314) from March 2016 through December 2018, the institutional-based predictive model assigned 394 patients to be of low risk (168 PD and 226 hepatectomy). Seventy-seven patients (33 PD and 44 hepatectomy) were removed from the CML protocol and final analysis because of aborted or changed operation. Table 1 displays the distribution of patients to each laboratory protocol tier including patients requiring escalation due to a clinical change. Cost comparisons between the historical standard and CML charges are shown in Figure 4. Before practice change, these patients would have been charged approximately $1,809,439 for laboratory tests ($1,068,564 PD and $740,875 hepatectomy). Ordering laboratory tests based on our institutionally derived predictive model, however, resulted in CML charges totaling $1,006,048 ($594,121 PD and $411,926 hepatectomy). This represents a total savings of 44.4 per cent, equaling $803,391 in potential patient cost savings ($474,442 PD and $328,948 hepatectomy).

NSQIP®-Based Model

Sixty patients were scheduled for DP under the CML protocol during the study period between December 2017 and December 2018. Nine patients were excluded from analysis because their operations were aborted or changed from DP intraoperatively. Table 1 shows the stratification of patients to each color-coded protocol tier as well as the number of patients requiring escalation to Yellow or Red Laboratories based on a clinical change or operative factors. Cost comparisons are again shown in Figure 4. Before practice change based on the historical standard, these patients would have been charged a cumulative $266,424 for full daily laboratory tests. Using the NSQIP®-assigned risk category and calculating charges for the actual laboratories ordered each day of hospitalization, the CML charges totaled $185,165. This represents an estimated $81,259 in patient savings from laboratory charges compared with the historical standard of daily full laboratories which amounts to a 30.5 per cent reduction.

Clinical Outcome Tracking

A clinical outcomes comparison between the CML intervention group and a historical control immediately before CML implementation is shown in Table 2. There was no statistical difference in either the institutional or NSQIP®-based laboratory protocols on major complications, readmissions, or mortality at 30-day follow-up compared with historical controls for both cohorts. On subgroup analysis of the institutionally based protocol, major complications were only significantly higher for PD patients (P = 0.040); however, the comparison cohort was substantially smaller than the intervention cohort because of the limited number of patients enrolled in ERAS® before the CML initiative. There was no statistical change in 30-day readmissions, or mortality after CML implementation for any subgroup.

Discussion

This study shows that postoperative laboratory order sets and patient charges can be reduced by using predictive analytics to eliminate pathways that reduce variation. Similar cost savings can be realized, regardless of whether an institutional or generalizable model (such as the ACS-NSQIP® calculator) is used. Since implementation of the protocols, we estimate a cumulative savings of $884,651 for all included patients (a total reduction of 42.6% compared with historical laboratory charges) See Figure 5. There was no statistical impact on overall clinical outcomes (major complications, readmissions, and mortality) after either protocol implementation in our patient population at the 30-day follow-up.

The role of predictive analytics has become integral in the drive toward value-based care. Parikh et al.10 described this natural union and emphasized the need for "precision delivery" to personalize care and improve value. Predictive analytics using "big data" to specifically improve economic outcomes has been explored, yet distilling this information into daily patient-specific decision-making remains challenging. Chen et al.25 highlighted this challenge of transforming traditional health economic outcomes research to "precision health economic outcomes research" to match the rise of precision health care. Importantly, the methods used in this study do not apply predictive models to directly anticipate cost outcomes but rather link risk stratification of surgical outcomes indirectly to lowering patient charges.

Although previous protocols in medical and surgical services have demonstrated success in reducing unnecessary laboratory testing, most interventions have been focused on increasing provider awareness, eliminating automatic order sets, and implementing universal guidelines.2' 5' 8' 9 Some programs implemented for high-risk patients in postoperative care units have attempted to consider adjusting ordering patterns based on patient comorbidities and clinical status, yet decision-making is typically based on provider opinion or consensus without a specific tool to risk stratify patients into appropriate laboratory testing tiers.26-28 This study therefore represents a new implementation of predictive analytics to guide a laboratory costreducing protocol in a patient-centered model.

This study illustrates the broad applicability of this cost-saving strategy with implications from individual surgical services to health-care systems of all sizes. High-volume centers with adequate infrastructure and resources can customize their cost savings by investing in population-specific predictive algorithms as demonstrated by our models for PD and hepatectomy. On the other hand, the results from our NSQIP®-based model highlight how smaller centers can still realize resource savings with little investment through the established NSQIP® platform to risk stratify patients to a simple laboratory protocol.

Limitations of this study include the study design for the economic outcomes as patient charges which do not represent actual patient cost. Although the financial impact of our intervention on our patient population is limited, the variability in patient insurance coverage and negotiated payments would similarly obscure total cost savings. Thus, we believe that estimating savings by standard laboratory charges provides a more generalizable representation of the economic impact. We recognized the inherent limitation that institutionally based predictive models cannot be directly translated to other surgical services and constructing separate population-specific models requires resources that many providers do not have access to. Therefore, we aimed from the outset to demonstrate similar results with the ACS-NSQIP® Risk Calculator in a separate patient population. The use of different patient populations for these two models is also acceptable because our aim was not to demonstrate superiority of one model over the other but rather to demonstrate their general capacity and utility for cost reduction with different degrees of investment, given an institution's resources and capability. Finally, the patient sample for the historical comparison group was limited by the number of patients captured by the EiAS data system prior CML implementation for PD and hepatectomy because these protocols were instituted earlier than the DP protocol. These outcome comparisons are at risk of both Type I and II errors despite the larger intervention groups.

Future Direction

Predictive analytics are powerful instruments, yet, as with all tools, their impact depends on careful guidance and fine-tuning. It is important to recognize that the thresholds for assigning risk categories in any predictive model are initially arbitrary and typically conservative. Iterative auditing of outcomes is imperative not simply to ensure a qualitative improvement in the desired direction (e.g., achieving any reduction in charges) but more importantly to recalibrate the predictive levels and titrate the quantity of the improvement even further with simultaneous outcome tracking. Because the risk thresholds were arbitrarily set for all cohorts in this study, we will adjust the risk calculators for hepatectomy, and DP patients broaden the low-risk tier and track outcomes after each iterative change. Conversely, the small but statistical increase in major complications within the PD cohort does not represent a failure of the predictive model but rather instructs us to sequentially tighten the window of low-risk patients until the difference in adverse outcomes is eliminated.

Once a predictive model is developed, the potential application becomes essentially limitless. Opportunities to reduce variation and waste should be explored for other routine practices in patient care such as follow-up axial imaging, daily chest radiographs, or timing of postoperative visits. By collecting big data on even the most mundane clinical activities, we can generate comprehensive predictive models which can drastically change health-care delivery and cost.

Finally, the historical role of risk assessment tools must be fundamentally changed. The availability of real-time collection and analysis of population-specific data allow surgeons to move from predictive analytics (simply anticipating the risk of a poor outcome) to prescriptive analytics, by which patient care is specifically tailored in a proactive manner to reduce risk. Through continual auditing of outcomes, these prescriptive models can be adjusted to intentionally guide patients through the entire perioperative experience with the best possible clinical and financial outcomes.

Conclusion

As the focus of health care continues to move toward increasing personalized value-based care, predictive analytics must be used to integrate clinical and economic outcomes through precision health-care delivery. Risk assessment models can be tailored based on available resources and clinical volume to drastically impact underappreciated costs such as laboratory testing, which has the potential to broadly change practice and health care in the United States. Most importantly, we strongly propose that the framework undergirding the role of predictive models in health care be shifted from static predictive analytics to dynamic prescriptive analytics through which we can make deliberate and continuously improving changes to the delivery of health care.

Acknowledgments

Gerri Chadwick, AuD - Medical Writer, Carolinas Center for Surgical Outcomes Science, Carolinas Medical Center.