Although it is well known that hip fractures (HFx) in older adults are a leading cause of morbidity, long-term functional impairment, and mortality [1], these outcomes are exacerbated when such patients are readmitted to the hospital within 30 days of hospital discharge after surgery, in addition to doubling their risk of 1-year mortality [2].

While many readmissions are unavoidable, unplanned hospital readmissions can easily negate the functional gains painstakingly achieved through weeks of post–acute rehabilitation and can increase the risk of infections acquired during hospital stays [3]. This loss is over and above the costs to caregivers and relatives who have to relive the stress of the original HFx episode, reorganize their work schedules to care for the patient, resulting in loss of productivity, and restart rehabilitation after discharge [3]. Across all conditions, unplanned readmissions cost almost US $17 billion annually in the United States [4], making them an ineffective use of costly resources and therefore closely scrutinized as a marker for poor quality by the Centers for Medicare & Medicaid Services (CMS) [5]. Consequently, the CMS instituted the Hospital Readmissions Reduction Program (HRRP) [6], which has imparted penalties on hospitals if their 30-day readmission rates exceeded the national average.

Although such incentives initially appeared to improve the readmission rates in US hospitals [7], recent reports argue that the start of the HRRP coincided with an increase in mortality among older adults [6,8]. This could have been because, as hospitals tightened their policies for readmission, many older adult patients were denied care, resulting in increased mortality. Furthermore, the decrease in readmission rates might merely reflect changes in the administrative and billing practices rather than an improvement in care [9]. These results suggest a need for more targeted research to comprehend the processes that precipitate readmission and clinical interventions that address the underlying causes of hospital readmission.

As hospital readmissions in the older adult HFx population are predominantly for reasons not related to the HFx surgery [10], several studies have focused on using supervised machine learning methods to determine how pre-existing comorbidities (defined as one or more conditions or diseases co-occurring with a primary condition such as HFx) increased the risk of readmission [2,10-14]. Most of these studies have focused on using logistic regression to analyze the risk of readmission of single comorbidities. For example, a recent study using Medicare data conducted for the CMS, analyzed patients with total hip or total knee arthroplasty to construct a logistic regression model with variables including 29 comorbidities to predict readmission [14]. Although the above descriptive and predictive approaches have provided important insights into the role of comorbidities in the readmission of patients with HFx, such studies do not focus on understanding how multiple comorbidities co-occur within and across subgroups of patients, a critical step in the design of targeted interventions to reduce readmissions.

Although the co-occurrence of pre-existing comorbidities has not yet been analyzed in readmitted patients with HFx, it has been analyzed in other index conditions [15-18], such as chronic obstructive pulmonary disease (COPD), and in patient populations, such as in older adults [19-21]. Such studies have focused on using unsupervised machine learning methods such as clustering (eg, hierarchical and partitioning clustering), dimensional reduction methods (eg, principal component analysis), and visual analytics (eg, network visualization and analysis). These include a recent questionnaire-based study of senior Australians that compared several unsupervised clustering methods to analyze patterns of multimorbidities (2 or more co-occurring conditions or diseases irrespective of an index condition) in the population [20]. The results found frequent co-occurrences, such as high blood pressure and diabetes, across the study population. Another study used unipartite networks (where nodes represented comorbidities, and edges between pairs of comorbidities represented the frequency of co-occurrence in patients) to identify clusters of frequently co-occurring comorbidities [21].

Although these studies have revealed the feasibility and appropriateness of using unsupervised methods to analyze the co-occurrences of comorbidities, they have typically focused on a unipartite analysis (clustering of only comorbidities) of the data and therefore cannot reveal complex patterns of patient heterogeneity hidden within those co-occurrences. Furthermore, such analyses cannot reveal the nature and degree of overlap among such subgroups. Understanding the complexities in such overlapping patient subgroups and their risk for readmission has direct relevance to clinician stakeholders in inferring the underlying processes involved in precipitating readmission and for the design of targeted interventions to reduce the risk of readmission.

Therefore, we explored an approach that integrates a supervised combinatorial method with an unsupervised bipartite network to address 3 questions: (1) Which combinations of comorbidities confer high risk for readmission in patients with HFx? (2) How do high-risk comorbidities co-occur within and across subgroups of readmitted patients with HFx? (3) What is the association between comorbidity risk, comorbidity co-occurrence, and patient subgroups?

As shown in Figure 1, we addressed our 3 research questions by using a supervised machine learning method to address the first question, an unsupervised visual analytical method to address the second question, and an integrated visualization of both results to address the third question. Our goal was to analyze which combinations of comorbidities confer high risk for readmission and how those high-risk comorbidities co-occur within and across patient subgroups. This integrated visual analytical approach was designed to explicitly enable clinician stakeholders using a team-centered informatics [22] approach to comprehend the complex association of comorbidity risk, comorbidity co-occurrence, and patient subgroups, with the goal of designing targeted interventions, a cornerstone of precision medicine.

Our data consisted of a training data set extracted from the 2010 inpatient Medicare claims data (the most current Medicare data set to which we had access) and a replication data set extracted from the 2009 inpatient Medicare claims data (the next most current dataset). In 2010, Medicare provided health insurance to approximately 48 million Americans, of which 40 million were older adults (≥65 years), representing 93% of all older adult Americans. Furthermore, the eligible claims were from 6204 medical institutions from across the United States, thereby confirming this to be one of the few data sets that are highly representative of the US older adult population and its care.

As is commonly done in analytical studies of claims data, we used Medicare Severity-Diagnosis Related Group (MS-DRG) codes to define our population. The MS-DRG codes are used by physicians to categorize Medicare beneficiaries into payment groups for the purposes of billing. We operationally defined patients with HFx as those who were discharged from an acute care hospital with the MS-DRG codes 480, 481, or 482. To isolate the association of pre-existing comorbidities with the risk of readmission and to maintain homogeneity of our study population, we included only patients without hospital complications. Furthermore, we included only patients who were enrolled in Medicare part A but not in a health maintenance organization (a type of health insurance that limits coverage to care from contracted doctors) during the period of 90 days after discharge, in addition to patients who survived 90 days after hospital discharge.

For both the training and replication data sets, we extracted (1) the data of all patients with HFx without hospital-acquired complications who were readmitted within 30 days of discharge and (2) an equal number of controls matched for age, gender, and race, who were not readmitted within 90 days of discharge. This 90-day window of no readmittance represents an episode of care proposed by CMS for patients with HFx [23], indicating that the controls are substantially free from complications that result in readmission during this period, thereby allowing an effective comparison with the cases.

The above inclusion and exclusion criteria for patients resulted in a training data set consisting of 16,886 patients (8443 cases and 8443 controls), and a replication data set consisting of 16,222 patients (8111 cases and 8111 controls) for a total of 33,108 patients with HFx (Multimedia Appendix 1). For each of the above patients, we extracted their status on 70 high-level comorbidities (Multimedia Appendix 2) as defined by hierarchical condition categories (HCCs) [24], which represent the range of conditions typically encountered in older adults. As our index condition was HFx, we excluded it as a comorbidity, resulting in the status of 69 HCC comorbidities across 33,108 patients with HFx in the Medicare database. This retrospective study was approved by the Institutional Review Board of the University of Texas Medical Branch. The Medicare data files used for the study were in the research identifiable format, and the records were anonymized and deidentified before the analysis. Therefore, the analysis of the data did not require informed consent. Furthermore, a data use agreement was completed, which met all CMS privacy and confidentiality requirements.

As shown in Table 1, the pairwise overall test identified 24 pairs (all rows shown in the table) that were significant in the training data set. Furthermore, the pairwise directionality test identified 10 pairs that were significant in both directions, 13 that were significant only in one direction, and 1 that was not significant in either direction (for clarity, only significant results are shown for the pairwise directionality test in Table 1).

Next, we identified which of the above 24 pairs replicated by identifying comorbidity pairs that were identical in their significance and direction in the replication data set. As shown in Table 1, of the 24 pairs, 11 pairs (highlighted in blue and pink) replicated in the replication data set. Of these, 2 pairs (serial number pairs 1-2) were significant in both directions, and 9 pairs (serial number pairs 3-11) were significant only in one direction. The overlapping pairs resulted in 8 unique comorbidities: congestive heart failure (CHF), COPD, major complications of medical care and trauma (MCMCT), RF I-V, stroke, diabetes without complications (diabetes), arrhythmias, and vascular disease.

As shown in Figure 4, the CoRisk network revealed how the high-risk pairs were (1) related to each other, (2) their directionality, and (3) how they were related to the patient subgroups. This integrated network enabled stakeholders to identify 2 sets of comorbidities. The first set (diabetes and RF) consisted of comorbidities that can have multi-organ consequences and is therefore referred to as systemic diseases. In contrast, the second set (CHF, arrhythmia, stroke, MCMCT, COPD, and vascular disease) consisted of comorbidities that had mainly organ-specific consequences. For example, while cardiac arrhythmia could potentially have systemic consequences, this comorbidity is specific to the electrophysiological properties of the heart.

As the clinician stakeholders were most interested in the interrelationship of risk between multi-organ and organ-specific comorbidities, we bolded all the edges that started from an organ-specific comorbidity (CHF, arrhythmia, stroke, MCMCT, and COPD) and ended at a multi-organ comorbidity (RF or diabetes). As shown in Figure 4, all the remaining edges pointed toward RF and diabetes, forming an asymmetrical hub (more edges pointing in than those pointing out). This meant that the implicated pairs connecting the nodes had significantly higher risk compared with RF and diabetes alone, but not significantly higher risk compared with the other members of the pair. For example, the directed edge starting from COPD and pointing to diabetes indicated that patients with COPD and diabetes have a significantly higher risk compared with diabetes alone, but not a significantly higher risk compared with COPD alone.

The asymmetrical risk hubs of RF and diabetes suggested that because they have multi-organ consequences, their outcomes are largely chronic and therefore require considerable severity on their own before they become the sole risk factors for readmission (note that the HCC definition of RF has a wide range from Stage I to V, possibly resulting in several patients with RF being in the early stages). However, when they co-occur with an organ-specific disease such as CHF, arrhythmia, stroke, MCMCT, or COPD, it can exacerbate those pre-existing conditions leading to a significantly higher risk of readmission. This pattern of asymmetrical risks resulted in the following hypothesis for a 2-tiered comorbidity exacerbation risk model with significantly higher risk at each subsequent tier:

Combining the above risk model with their own domain experience, the physician and the occupational therapist on the stakeholder team inferred hypotheses for the processes precipitating readmission in patients with HFx and provided corroborative evidence from the literature to support their inferences. They noted that when a patient is discharged from a hospital after an HFx surgery, the standard-of-care in generating discharge notes and order sets is focused on wound healing, postoperative delirium, mobilization, rehabilitation, and nutritional needs [31,32]. However, despite these guidelines, older adult patients, particularly those in skilled nursing facilities, regularly suffer from dehydration and malnutrition [33-37]. These conditions can worsen compromised renal function [38] as well as glycemic control in diabetics, ultimately triggering the deterioration of existing organ-specific comorbidities such as CHF and COPD [39-41]. Unfortunately, by the time symptoms of exacerbation in comorbidities are detected, the patient’s health may have considerably declined, requiring urgent care, triggering an unplanned hospital readmission.

The nurse practitioner on the stakeholder team further stated that a contributing factor to the above cascade of events could be the lack of multidisciplinary care when patients with HFx are discharged after surgery. As stated in a recent review [42], HFx management “requires physicians to anticipate problems that may arise during recovery, whether the complications are from hip fracture and immobility, exacerbations of chronic diseases, or problems with social and psychological support...it takes a team of dedicated professionals working together seamlessly to deliver care appropriate for patient goals, and to maximize recovery.” In fact, although an increase in the number of registered nurses and multidisciplinary care teams has been associated with reduced 30-day readmission rates and improved health outcomes [43-45], such post–acute care is not yet widespread. The results of the analysis, combined with domain experience and corroborating evidence, enabled the clinician stakeholders to infer that the generation of discharge notes and order sets before discharge and the level of multidisciplinary care after discharge could be prime targets for reducing the risk of hospital readmission in specific subgroups of patients with HFx.

Our approach to integrate the results from supervised and unsupervised approaches into the CoRisk network helped to reveal (1) the overlap among the high-risk pairs, resulting in the stakeholders identifying the asymmetrical hub, and (2) the relationship of high-risk pairs to the network-wide and bicluster-specific patient heterogeneity. These results enabled the clinician stakeholders to infer hypotheses about the processes that precipitate readmission through a comorbidity exacerbation risk model. These results have the following implications for the design of interventions and predictive modeling.

The strength of this study is that we integrated the results from well-known methods with novel approaches, which together enabled a deeper understanding of the associations between risk, co-occurrence, and subgroups. This in turn led to insights related to targeted interventions (a critical goal of precision medicine), in addition to the design of predictive models. Furthermore, the analytical results were replicated in another year, demonstrating its generalizability. Critical to this process was the team-centered informatics approach [22] we pursued at each step of the project, which used intuitive visual analytical representations to span the disciplinary boundaries of clinicians and methodology stakeholders, enabling them to comprehend and address the complexity in a large data set.

A limitation of this study is that we tested the method on just one index condition, and our ongoing research [49] is testing the approach on other index conditions. Furthermore, the interpretability of the clusters could be enhanced by constructing additional figures wherein the patient nodes are colored based on covariates important to hospital readmission (eg, age, gender, race, length of hospital stay, and reason for readmission), in addition to determining which of them are significantly higher and lower across the clusters. Finally, the prevalence and severity of comorbidities may vary in patients receiving care in clinics, acute care hospitals, skilled nursing facilities, and nursing home settings. Therefore, future research should analyze whether the results vary across different care settings.

Fully cognizant that few data sets are without limitations, we consciously chose to analyze Medicare data because of its scale (enabling us to have adequate numbers of patients when analyzing patient heterogeneity), availability of data over multiple years (enabling us to test external replicability), and generalizability (enabling us to analyze data from patients and hospitals across the United States). However, given that Medicare data are collected mainly for administrative purposes, it has well-known limitations, including the lack of test results, which could enable a finer understanding of the severity of comorbidities. Furthermore, comorbidities associated with mental health are known to be undercoded in the Medicare database, which could bias our results. Therefore, when clinical data across hospitals become available in future (eg, through the PCORnet [50] funded by the Patient-Centered Outcomes Research Institute [PCORI] and through the Accrual to Clinical Trials network [51] funded by the National Center for Advancing Translational Sciences [NCATS]), we intend to repeat our analysis using clinical data, but we fully realize that such data might have limitations that are yet unknown.

Methodologically, our approach of integrating supervised and unsupervised visual analytical approaches is just one of the many possible ways such integration can be achieved [52]. In our future research, we plan to explore other integration strategies with a specific focus on enabling clinician stakeholders to go beyond the analyses of prevalence and risk, enabling inferences for the underlying processes precipitating readmission. Such improvements in data and methods should enable discharge planners and providers in rehabilitation facilities to more accurately predict which patients will be readmitted and to select targeted interventions to reduce the risk of readmission and, consequently, the concomitant burden on patients and caregivers.