Better big data

Abstract

By 2018, Medicare payments will be tied to quality of care. The Centers for Medicare and Medicaid Services currently use quality-based metric for some reimbursements through their different programs. Existing and future quality metrics will rely on risk adjustment to avoid unfairly punishing those who see the sickest, highest-risk patients. Despite the limitations of the data used for risk adjustment, there are potential solutions to improve the accuracy of these codes by calibrating data by merging databases and compiling information collected for multiple reporting programs to improve accuracy. In addition, healthcare staff should be informed about the importance of risk adjustment for quality of care assessment and reimbursement. As the number of encounters tied to value-based reimbursements increases in inpatient and outpatient care, coupled with accurate data collection and utilization, the methods used for risk adjustment could be expanded to better account for differences in the care delivered in diverse settings.

Big data is the future of healthcare

In January 2015, Health and Human Services Director Sylvia Mathews Burwell announced a program to tie Medicare payments to quality of care [1]. By 2018, 90% of Medicare payments will be tied to performance, as opposed to the current system which primarily reimburses volume. The Health Care Payment Learning and Action Network will measure quality of care and rely heavily on accurate recording of encounters in billing codes and encounters recorded in the electronic medical records adopted by 48% of hospitals [2] and 78% of private practices, as of 2013 [3], through the Centers for Medicare and Medicaid Services Electronic Health Records Incentive Programs.

The Centers for Medicare and Medicaid Services currently uses quality-based metrics for some reimbursements through the Readmissions Reduction Program [4], Hospital-Acquired Conditions Reduction Program [5] and Hospital Value-Based Purchasing Program [6]. The six measures in the readmissions program include all medical and surgical inpatient readmissions and additional specific readmissions for acute myocardial infarction, chronic obstructive pulmonary disease, heart failure, pneumonia, total hip replacement and total knee arthroplasty. The 14 Hospital-Acquired Conditions range from infectious conditions to non-communicable conditions. The Hospital Value-Based Purchasing Program has 24 measures including survey results on patient satisfaction, hospital-based process of care measures and outcome measures including patient safety indicators, myocardial infarction, heart failure, pneumonia and surgical site infections. The quality-based measures included in these three programs cover 39% of inpatient Medicare encounters [7]. When at least 90% of all Medicare reimbursements, including both inpatient and outpatient encounters, are tied to quality in 2018, the encounters not covered by the existing programs will require validated metrics that can be compared across medical centers, like readmissions that are not linked to the currently monitored conditions and high-volume and expensive procedures such as colonoscopies.

Can the current data differentiate risk-adjusted quality of care?

The existing and future quality metrics will rely on risk adjustment to avoid unfairly punishing those who see the sickest and highest-risk patients [8]. One often-cited limitation of billing data and Electronic Health Records is missing data. If there is missing information that correlates highly with the quality metrics, the risk adjustment will be insufficient and may misclassify the highest and lowest quality providers. For example, the patient safety indicators used in the Hospital-Acquired Conditions and the Hospital Value-Based Purchasing programs rely on 30 measures to adjust for patient severity, using a score originally developed by Elixhauser et al. in 1998 [9]. The score includes International Classification of Diseases, Ninth Revision, Clinical Modification codes for cardiovascular health, diabetes, cancer, obesity, alcohol abuse and other factors associated with chronic disease. However, more than 40% of inpatient billing records from individuals with known, chart-confirmed cardiovascular disease, diabetes and cancer do not have these codes listed in the discharge summary [10,11]. Furthermore, despite 34% of the US population being obese and 33% being overweight [12], only 9.6% of hospitalizations in the Nationwide Inpatient Sample have codes indicating obesity and less than 1% have codes for overweight [13]. For alcohol abuse, 25.0% of the US population consists of binge drinkers (five or more drinks on the same occasion on at least 1 day in the past 30 days), while only 4.6% of patients have codes indicating alcohol abuse in the Nationwide Inpatient Sample (NIS) [13]. If value-based reimbursements use the Elixhauser score for risk adjustment, the missing information may create inaccurate adjustment that could lead to inaccurate assessment of the highest and lowest quality providers. When payments are tied to quality, it will be essential to accurately classify and adjust for patient risk to ensure fair payment to more than 830,000 healthcare providers and 4974 hospitals providing care for Medicare patients [14].

Short-term & long-term solutions to improve the data used for risk adjustment

Despite the limitations of the underlying data used for risk adjustment, there are potential short-term and long-term solutions to improve the accuracy of the codes used for risk adjustment. Because it is not feasible to recode all past Medicare encounters to improve accuracy, there are at least two immediate possibilities to improve the data used for risk adjustment using existing information to calibrate missing data and to combine information from multiple existing programs into new individual-level datasets.

Calibrating data by merging databases

The Department of Health & Human Services produces over 1800 publicly available databases that could be used to complete missing information [15]. For example, the Behavioral Risk Factor Surveillance System (BRFSS) and the National Health and Nutrition Examination Survey (NHANES) provide state-level and regional-level estimates for many of the comorbidities measured in the Elixhauser score, including diabetes, cardiovascular disease, obesity and alcohol abuse. Even with the challenge of being imputed from relatively small databases, this nationally representative information can be used to complete missing information from the medical coding based datasets. Multiple imputation can be used to calibrate, or complete, the missing data using the external data, after combining the datasets at the most granular level of linkage possible [16]. External adjustment could be used to create bounds of plausible values if a single variable summarizing risk was available or a single variable was proposed to be responsible for the majority of the risks [17].

Compiling the information collected for multiple reporting programs to improve accuracy

Another option to improve the data is to use information captured about the individuals in the dataset from another data source. For example, incorporating Meaningful Use measurements into datasets that comprise billing codes could enhance the completeness without adding additional coding burden [18]. Meaningful Use measures include reports on vital signs, such as height and weight, smoking status, demographics, medication lists and active diagnoses [19]. Rather than relying on billing codes for obesity and overweight, for example, the directly measured height and weight that are already collected for the Meaningful Use program as specific structured fields improves data capture and consistency. Including these standardized data in the billing dataset could improve the quality of information on risk factors without having to use International Classification of Diseases codes at all, thus improving the risk adjustment. Data imputation will then allow completing the missing fields on the community level rather than on an individual level in such databases. This method will allow for database validation among the different health databases, especially when more and more US hospitals and private practices have attested to Meaningful Use programs, which would translate into more extensive collection of the aforementioned measurements.

Long-term improvement of risk assessment

To improve the quality of risk adjustment in the future, the coding system should also be incentivized to match the suggested reforms. The current coding system would also benefit from a change of financial reimbursement that would be linked to quality metrics to match the suggested reforms of reimbursement, rather than for description of volume that was used in the old compensation system. A second option is to include the measures contributing to the risk-adjustment score in the currently released Medicare records. For example, the public release of Medicare de-identified patient-level data with their corresponding Elixhauser score allows the public to understand who sees the sickest patients and how it affects reimbursement. A third option could include audits of the codes used to calculate risk, just as the audits are performed to identify improper Medicare payments [20]. Combining creation of financial incentives in coding, publicly available information on items used to adjust for risk in the existing Medicare products and random audits could increase the accuracy of information used for risk adjustment within existing Medicare programs.

Improving risk adjustment can lead to better personalized medicine

As the number of encounters tied to value-based reimbursements increases and includes both inpatient and outpatient care, coupled with accurate data collection and utilization, the methods used for risk adjustment could be expanded to better account for differences in the care delivered in diverse settings. For example, a high-dimensional data mining approach, originally developed to predict readmission for chronic pancreatitis, was implemented to predict all-cause readmissions. The predictors were not the same, but the analytic framework allowed plasticity to rapidly adapt to a new setting. This same framework, if built into an electronic medical record system, could not only rapidly adapt to changing medical conditions but also account for changes in the patient mix with time within a hospital [21]. Data mining is also being used to predict treatments for cancer and help understand rare diseases. IBM launched its Watson cognitive learning machine to mine all structured and unstructured data in Electronic Health Records [22]. On the other hand, BioExcel launched its Orphan Disease Suit for the study of rare and ultra-rare disease conditions by integrating big data in the pipeline of diagnosis and drug discovery for these conditions [23]. Improving the accuracy of information for risk-adjusting quality may have a downstream effect of improving the ability to predict treatment response for patients.

Improving big data to better account for patient risk can improve the assessment of quality care, can prevent patient harm by accurately identifying the hospitals that are badly in need of interventions and can save money by allowing for fair payments to the providers delivering the safest care at the right cost.

Financial & competing interests disclosure

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

No writing assistance was utilized in the production of this manuscript.