A Cross-Sectional Study to Predict Mortality for Medicare Patients Based on the Combined Use of HCUP Tools

Background and Significance

Many studies explain the methods for the prediction of inpatient mortality for patients admitted with acute conditions [1, 2] and discuss data challenges for the prediction of this outcome [3]. Since inpatient mortality is an important metric for quality assessment and external reporting, it is oftentimes studied as a baseline to measure the predictive performance of clinical analytics. The prediction of short-term hospital mortality is not an easy problem. Due to the probabilistic nature of healthcare, similar patients may differ in terms of outcomes. Awad et al. summarized publications between the years 1984 and 2016 that discuss the prediction of mortality [4]. Variations to the risk for negative outcomes of care, such as death, can be warranted, due to inherent clinical characteristics and disease complexity, or unwarranted, due to modifiable factors [5]. Objectively, severe case mixes are at higher risk for hospital death, while at the same time, several hospital structure attributes, such as hospital volume [6, 7], can influence the risk for hospital death.

To examine inpatient mortality, critical variables of the study include the principal diagnosis, the secondary diagnoses that may or may not be associated with the reason for hospitalization (complications and comorbidities), and medical procedures, as well as the age-associated risk [8]. Other hospital mortality predictors have been reported to be various disease-specific laboratory test results [9].

Hospital-acquired conditions, the presence of chronic conditions, and several medical procedures can also be associated with an increased risk for hospital death. Morsy et al. found that the development of bacterial infection, shock, early rebreeding, low serum albumin, low baseline hemoglobin, and the need for endoscopic treatment were independent risk factors of hospital mortality among cirrhotic patients [10]. Other information that can be associated with the risk for hospital death is whether the patient was a readmission case [11] and was directly transferred from a different hospital, as well as the urgency of hospital admission.

Several disease dimensions are important for the thorough study of inpatient mortality. Researchers should be able to distinguish between principal and secondary diagnoses. Studies could also benefit from characterizing the secondary diagnoses as of chronic or acute nature and as hospital-acquired or preexisting. These secondary diagnosis dimensions can contribute to improved predictive performance for inpatient outcomes research.

Beginning in 2015, hospitals in the USA started to use the 10th edition of the clinical modification of the International Classification of Diseases (ICD-10-CM). ICD-10-CM provides a great level of specificity since it includes more than 69,000 unique codes. On the other hand, its out-of-the-box appropriateness in predictive analytics can be problematic for several reasons. Firstly, with such an enormous feature set, the risk for overfitting is high, unless many millions of data records are available. Additionally, when enormous data matrices are analyzed, there is a substantial increase in computational complexity. The clinical classification software (CCS) was developed by the Healthcare Cost and Utilization Project (HCUP) and classifies the thousands of ICD-10-CM and ICD-10-PCS codes into 285 diagnostic and 231 procedure categories. CCS, therefore, provides a compact representation of the medical diagnoses and procedures [12]. In 2016, HCUP released updated versions that are compatible with the 10th editions of ICD-CM and ICD-PCS. Besides, HCUP makes available the Chronic Condition Indicator (CCI), to classify each ICD-10-CM code as chronic or non-chronic [13].

While CCS is a well-studied approach to manually reduce the dimensionality of medical diagnoses [14, 15], other methods include the use of comorbidity indexes. One of these indexes is the Elixhauser Comorbidity Index, which extracts patient comorbidities from the ICD and diagnosis-related group (DRG) codes found in administrative data. The index has been used in hospital outcomes research such as in [16], to predict hospital resource use and hospital mortality. The most recent edition of the index includes 29 categories and is compatible with ICD-10-CM. The Centers for Medicare and Medicaid Services (CMS) provides medical claim datasets to researchers; the HCUP tools can be combined with these datasets since they use the ICD-10-CM and ICD-10-CPT classification systems to represent the clinical diagnoses and procedures.

Objective

The motivation for the present study is to improve the predictive capacity of medical claims data, by extracting and utilizing meaningful dimensions of diagnoses from HCUP tools and other available metadata. The research hypothesis is that the combined use of medical claims data with the HCUP tools (ICD to CCS and ICD to CCI grouper) and other metadata (“present on admission” indicator) can enrich the information about medical conditions, reducing information entropy and therefore improving the capacity to predict the outcome of hospital mortality. To the authors’ knowledge, this is the first study that examines the combined use of the updated (ICD-10 edition) HCUP tools with medical claims data, for inpatient outcomes research.

This study aims to (i) examine the predictive power of the combined use of the HCUP tools and medical claims data for the outcome of inpatient mortality, among Medicare patients, (ii) compare the predictive performance of the recommended secondary diagnosis dimensions against the use of the Elixhauser Comorbidity Index, and (iii) recommend a statistically significant collection of critical CCS diagnosis and procedure predictors for inpatient mortality.

With this work, the authors want to provide to the research community a meaningful reduced multidimensional CCS diagnostic code set to facilitate future research effort that uses HCUP and CMS inpatient databases. This set includes statistically significant predictors. Besides, the statistical model that was developed to extract these statistically significant variables can easily be replicated in future research efforts. This is important since HCUP and CMS databases are frequently utilized in health data science research studies. The motivation is to evaluate the use of the updated HCUP tools on recent inpatient CMS data and provide a set of comorbidities, including chronic and hospital-acquired conditions that increase the likelihood of death, while the study of mortality, in general, would benefit from additional predictors, some of which may be condition-specific and not examined in the current research. The study provides a multidimensional set of high-risk diagnoses for inpatient death by making the distinction as a hospital or nonhospital acquired, chronic, or acute. While the HCUP tools have been used for risk adjustment by other studies, no research relies on the updated CCS to ICD-10 mapping, and this is a contribution of this work: apparently, regrouping diagnostic codes under CCS buckets, with the 10th ICD edition, which follows a different philosophy from ICD-9-CM, may result in CCS groups different from the ones that an ICD-9-CM dataset would generate.

Materials and Methods

Results

Discussion

This is the first study to combine the new HCUP grouper files (for the 10th editions of ICD-CM and ICD-PCT), with medical claims data. The goal is to examine the predictive power of the clinical diagnosis dimensions that this study developed and to also recommend a critical collection of diagnosis and procedure predictors for inpatient mortality. CCS can be a workable way to gather information quickly regarding a healthcare topic [19] and has also been shown to be useful for improved predictions of mortality [20]. This research utilized the CCS system on ICD-10-CM data. While there are previous studies that examined the potential for CCS to be used in predicting hospital outcomes, they used the ICD-9-CM and ICD-9-PCT grouper versions.

Various comorbidity indexes have been developed to improve on measures of comorbidity in administrative databases and are often used as control variables in cross-sectional research. The Elixhauser Comorbidity Index labels secondary diagnoses as comorbidities only when these are disconnected from the principal diagnosis. To achieve this, it uses the DRGs to assess the independence of principal and secondary diagnoses. This is reflected in the comparison results of the present study: When the Elixhauser Comorbidity Index was used instead of the secondary diagnosis dimensions that this research developed, the principal diagnoses that were positively associated with hospital death were different and mostly consisted of malignancies. Illustratively, CCS = 20 (cancer; respiratory and intrathoracic), CCS = 31 (cancer of other genital organs), CCS = 24 (cancer of breast), CCS = 26 (cancer of cervix), and CCS = 41 (cancer, other principal) were the top principal diagnoses associated with an increased likelihood for inpatient mortality. When the secondary diagnosis dimensions were used instead, a majority of the aforementioned principal diagnosis were not found to be significant predictors of inpatient mortality. This can be explained, since, in acute care, death is oftentimes attributed to secondary ailments derivative of the principal diagnoses, and not from the principal diagnosis itself. The use of the secondary diagnosis dimensions that the present research developed was shown to contribute to a better prediction of inpatient mortality in comparison to the representation of secondary diagnoses with the Elixhauser Comorbidity Index. It appears that, while the Elixhauser Index includes a collection of comorbidities that are not related to the principal diagnosis, this data structure may not be optimal for the prediction of inpatient deaths among Medicare patients. The performance of the proposed multidimensional diagnosis collection, against the Elixhauser Comorbidity Index, was found to be slightly better, with a 7% increase to the sensitivity not being a negligible one.

Other comorbidity indexes, designed for long-term mortality, have been used in research to examine inpatient death. Unlike the Elixhauser Index, all these indexes were not originally designed for short-term outcomes in acute settings, D′ Hoore et al. [21] used the Charlson Index to study hospital death among patients with ischemic heart disease and found that the index was strongly associated with death. The Charlson Index has also been used to predict the cost of chronic disease in primary care [22]. Using study-specific weights with the Charlson variables has also been shown to improve the prediction of mortality [23]. Other approaches that the literature discusses include Cumulative Illness Rating Scale (CIRS), which addresses all relevant body systems without using specific diagnoses; the Index of Coexisting Disease (ICED) which has a two-dimensional structure, measuring disease severity and disability, which can be useful when mortality is the outcome of interest; and other disease-specific indexes like the Kaplan Index for diabetes research [24]. The main reason why the proposed set of diagnoses outperforms the Elixhauser Index is that the Elixhauser Index was not designed specifically for the study of inpatient mortality but serves as a collection of comorbidities that are independent of the cause for hospitalization. The additional information that our research feeds the model with was based on the capacity to predict in-hospital mortality with the three dimensions that were generated using the HCUP tool sets and POA information from the CMS LDS file.

Several diagnoses were found to be associated with a lower risk for hospital death (based on the expected risk). What this means in clinical terms is not that, for instance, asthma protects against death but that asthma is a lower risk factor for death than the expected risk for the overall inpatient population. Those associations can be explained case by case. For instance, congestive heart failure (CHF) is not a direct life-threatening condition for inpatient death as much as other heart diseases are, like acute myocardial infarction (AMI). CHF is characterized by chronicity and is very prevalent in elderly patients. When CHF develops more rapidly in-hospital, it can be life-threatening, and this is not the case for a chronic patient who has CHF for years and has been admitted to the hospital for a problem of a different nature. Besides, CHF is managed well with diuretics, beta-blockers, etc., and is not imminently life-threatening, unless it is of a late stage.

When hospital outcomes of care are studied, it is important to differentiate between principal and secondary diagnoses. Studies demonstrated the importance of this distinction by comparing hospital costs and length of stay, among other factors, between patients admitted with the principal and secondary diagnoses. Steward et al. showed that secondary diagnoses often led to longer hospital stays as well as higher hospital costs associated with their diagnoses. They also demonstrated that readmission rates may be higher for individuals with specific principal diagnoses [25].

The POA dimension of the secondary diagnoses is often available in administrative datasets and provides information on whether each diagnosis was acquired/recognized during the hospital stay or not. A non-POA diagnosis may be associated with unique treatment challenges, especially when it is derivative from the principal diagnosis. In some other cases, though, it can be easier to treat a hospital-acquired diagnosis due to a timelier response to a newfound problem. Hospital-acquired conditions (HACs) are associated with negative outcomes of care and many studies have examined predictors of HACs [26]. Glance et al. found that mortality, cost, and length of stay were significantly higher among trauma patients who had developed hospital-acquired infections [27]. According to IBM-Watson Health [28], HACs can increase mortality risk per patient by 72.32%. As in the current study, Lagu et al. [29] also used POA codes to identify other acute conditions that are of concern but that are not recognized in the Elixhauser Comorbidity Index. Other approaches to study comorbidities were proposed in the literature, such as the Romano Index [30].

The current research utilized the chronic condition indicators of HCUP to recognize secondary diagnoses of a chronic nature. CCI has been used in other studies, like in [31], to understand the prevalence of chronic conditions and medical expenditures of the elderly for healthcare planning development of chronic conditions, to study child cancer complications [32], medical morbidity narcolepsy, and more [33]. Some of the previously mentioned studies also adopted CCI and CCS by the AHRQ to determine chronic condition diagnostic codes and classify the diseases.

The prediction of hospital mortality has been a popular research topic [34, 35]. While the statistical approaches may vary, many of the studies used multivariate logistic regression and the development of prediction rules. Most studies focus on specific conditions, though. For instance, among cardiology patients, attributes contributing to inpatient mortality were found to be history of congestive heart failure and the time between the onset of a cardiac event and starting chest compressions [36]. Claxton found that age, neurological lever, and Glasgow Coma Scale can be used independently to predict mortality in patients who suffer traumatic cervical spine [37]. For patients who have very extended stays in the ICU, a prolonged requirement for life support therapies and a limited number of preexisting comorbidities both are predictors of increased hospital mortality [38] while mean platelet volumes and platelet distribution width both have been shown to contribute to the prediction of mortality in ICU patients [39]. For patients following operative management of hip fractures, the list of predictors of mortality includes male gender, admission from a long-term facility, age greater than 90 years, living at home with support, and the presence of a postoperative complication [40]. The present research followed a non-disease-specific approach to find the predictors of all-cause death. Existing studies, including the aforementioned ones, follow more conventional research designs and statistical approaches, without the prediction being their objective but, instead, the knowledge about statistical associations with mortality. It is finally worth mentioning that, in the present study, when examining comorbidities with the NB classifier, the comorbidity interactions are ignored. In our work, the sensitivity to detect hospital death was higher with NB, despite this. One reasonable explanation is that the outcome of death may have predictors that act as predominant and exist as such, regardless of other interactions with coexisting conditions.

Conclusion

When using medical claims data for inpatient outcomes research, the combined use of HCUP tools with other medical claims metadata (POA indicator) can improve the prediction of inpatient death among Medicare patients. Also, this combined use provides improved predictions when compared against the well-known Elixhauser Comorbidity Index. The collection of critical dimensions of diagnoses (principal, secondary, hospital-acquired, chronic) and medical procedures that this study recommends may be used in future research efforts for a better understanding of drivers of inpatient mortality. This is especially important since this is the first study that uses the newly published grouping files of CCS and CCI with the 10th edition of ICD-CM. The findings of this study can contribute to hospitals having access to updated inpatient mortality risk information. This information can make it easier to prioritize proactive care and rank high-risk admissions for improved inpatient care management. The work also provides a foundation to develop a minimum set of inpatient mortality diagnosis predictors for future studies of mortality that utilize medical claim datasets.

Declarations