Predicting hospital admissions by looking backwards: an alternative perspective

Despite the increasing age and co-morbidity of incident patients, overall survival on renal replacement therapy (RRT) has improved over recent decades [1, 2]. Part of the explanation for this may be the drive to adopt healthcare strategies which minimize the complication burden associated with dialysis, such as facilitating the formation of arterio-venous fistulae [3] and incentivizing their usage [4]. But has this reduction in mortality been matched by a similar reduction in morbidity? Data from the United States Renal Data System suggest that there has been little improvement in the rates of hospitalization in prevalent haemodialysis and peritoneal dialysis patients in over a decade, with the average admission rate being two per year [1]. In North America, the delivery of inpatient care has previously been shown to consume a third of the Medicare dialysis expenditure, equating to an estimated ten billion dollars per annum in 2008 [1]. In addition to its monetary cost, hospitalization is associated with considerable risk to the patient. Access to routine dialysis treatments and physicians experienced in end-stage renal disease care will vary depending on where the patient is admitted [5]. Within 30 days of hospitalization, 25% of haemodialysis patients are readmitted and 2.4% died, exceeding the rates seen in the general population by 76 and 200%, respectively [6]. Hospital admission rates therefore act as both an intermediate outcome and a surrogate measure for quality improvement interventions. So what is already known about the factors associated with hospital admission? When looking at this topic, most studies start by identifying the cohort of patients they are interested in, such as patients incident on dialysis during a defined period or prevalent on dialysis on a particular date, and go on to examine counts of admissions or the relative hazard in time to first hospitalization. Admission rates have been shown to be heavily influenced by co-morbidity, previous RRT modalities and the attainment and maintenance of laboratory clinical targets such as haemoglobin, calcium and phosphate [7–9]. Medicare-reported standardized hospitalization rates, which are designed to compare facilities, employ these methods by calculating expected admission rates over particular periods and comparing these with the observed admission rates [10]. Methods that predict when an admission will occur (rather than if or how often) have not been forthcoming. It is in this context that Usvyat et al. report in this edition of NDT their experience of using data from a large haemodialysis provider database to offer insight into the events leading up to hospitalization and death. As for cases in a case–control study, an individual's eligibility for inclusion is their attainment of an outcome (e.g. hospitalization or death); unlike a case–control study, however, Usvyat's approach looks at within-individual changes in variables in the time leading up to the attainment of the outcome (not just differences in characteristics and exposures). They highlight that the rate of admission to hospital doubles from 4 months before death, and that admissions themselves (in this case the first admission since commencing dialysis) are preceded by changes in clinical and laboratory variables routinely collected by most providers. These findings build on previous work by the authors. By looking backwards from key events they have previously identified deflections or trends in the variables of interdialytic weight gain, systolic blood pressure, serum albumin and C-reactive protein (CRP) levels up to 1 year prior to death [11, 12]. Changes in CRP and albumin, variables one might more commonly associate with inflammation and infection, were even identified in patients who died suddenly in whom one might assume a cardiovascular cause of death. These findings are important in the context of reports from other investigators highlighting changes in the functional status in elderly haemodialysis patients that only occur within 1 month of death [13], suggesting that changes in biochemical values may actually significantly predate changes in symptoms. So what signs should the clinician look for to anticipate a forthcoming admission? Using a linear mixed-effects model (LMM), Usvyat et al. direct us to changes in interdialytic weight gain, intradialytic change in systolic blood pressure, neutrophil to lymphocyte ratio, haemoglobin, treatment time and saline usage over the 6 months prior to admission. LMM allows (for instance) an individual patient's baseline haemoglobin at the start of the study to vary randomly, but collects information on changes relative to this baseline to inform the overall cohort's trend. By this technique and with a relatively large sample size (4204 patients receiving treatment in eight Renal Research Institute clinics in North America), they appear to have identified statistically significant but potentially clinically undetectable changes. A change in haemoglobin by 0.2 g/dL is smaller for instance than the deflections associated with haemoglobin cycling, previously identified as having a strong association with hospitalization [14]. A patient's treatment time being shortened by 2 min or receiving an additional 7 mL of saline may not trigger a clinical review despite their association with haemodynamic instability and inferior outcomes [15, 16]. The application of LMM used by Usvyat et al. is, by its nature, dependent on observations from multiple patients and would not be the optimal statistical approach with which to develop an early warning tool. Further, it examines only changes from baseline in those admitted to hospital. To allow for this, the authors helpfully provide a ‘control’ population which enables comparison of the variables in those admitted and not admitted. This highlights that throughout the observation period, many of the variables in those not admitted are already at significantly different levels to those subsequently admitted, although these levels themselves are above what has previously been shown to be associated with excess hospitalization [17, 18]. This perhaps argues that the absolute value is as important as the trend. Indeed, differences in absolute values appear to be bigger than the differences over time and might therefore be easier to detect clinically. Such knowledge about both differences and deflections will certainly be very important in constructing algorithms to identify those at risk of hospital admission. One way to inform the construction of predictive models using prior knowledge is to use a reasoning process called temporal abstraction. This has been used in a number of other settings, such as critical care, where repeated measurements have been analysed to aid a decision support system rather than using regressions to determine the risk of the event [19]. These methods have also been applied to haemodialysis patients using albumin, haemoglobin, platelet count triglycerides and age with encouraging results [20]. Generating rules which inform a decision tree that was subsequently interpreted by clinical experts resulted in a sensitivity of 82% and a specificity of 77% in predicting a hospitalization in the forthcoming month from the preceding 6 months laboratory variables. Lessons may also be available from other chronic diseases where many tools have been developed that use routine comorbidity and socioeconomic data to predict readmission to hospital, with c-statistic model performance in the region of 0.68–0.83 [21]. Recognizing that deflections from a patient's own physiological norm are relevant, some of the more evolved screening methods in heart failure have used deflections in bioimpedance and shallowness of breathing to predict future decompensations or diuretic dose adjustments [22]. The use of bioimpedance alone resulted in a false-positive predictive rate for heart failure events of five per year, but its combination with other indicators achieved a false-positive rate of just 0.9 per patient year with 65% sensitivity and 92% specificity. Intuitively, changes in several clinical variables would also likely be needed to reliably predict impending admission to hospital in haemodialysis patients. The transferability of knowledge from other disease settings to dialysis patients has limitations, however. Staying with the heart failure example, telemonitoring incorporating the routine collection of physiological variables has been repeatedly trialled to prevent mortality and hospitalization. Although generally found to be associated with a reduction in all-cause mortality and predictive of admission to hospital with heart failure, studies have tended to be less successful at predicting admission to hospital from ‘all-causes’ [23]. This suggests, perhaps not surprisingly, that the optimal performance of these algorithms is most likely to be achieved with disease-specific endpoints. Aside from the associations with small deviations in clinical parameters, issues with Usvyat et al.'s study design should be highlighted. Patients needed to miss a haemodialysis session in order for the facility to explore if the patient was hospitalized, meaning that some short stay admissions may not have been detected. Some patients with deteriorating parameters will have received clinical intervention preventing admission, and the declining laboratory variables in these patients will exist in the control group meaning their exclusion may amplify differences further. The practice patterns in the US facilities that inform this analysis may not be generalizable to other countries. Indeed in international work by the same authors, deflections in these key variables before death were less marked or remained unaltered in patients dialysing in Asia than in Europe, North and South America [12]. Baseline admission rates in these countries have previously been shown to differ [7]. The paper in this issue tracks hospitalization in patients who approach death and clinical variables in patients approaching hospitalization, but offers little insight into early warning signs in those who die unexpectedly at home. UK studies have shown that ∼35% of RRT patients die outside of an acute admitting hospital [24]; deaths out of hospital may be as high as 50% in North America [25]. Collecting clinical parameters is engrained in the nephrology community, with the purpose of benchmarking performance and tailoring therapy. There is increasing evidence that we should be putting this information to use elsewhere: in identifying patients at risk of morbidity-associated hospitalization and prompting intervention, improving the patient experience and reducing the healthcare cost. Haemodialysis patients, who have their physiological parameters measured and usually electronically recorded several times a week, provide the ideal opportunity to develop tools that can focus clinical efforts in a cost-effective sustainable way on the patients that are in greatest need. None declared. The results presented in this paper have not been published previously in whole or part, except in abstract format. (See related article by Usvyat et al. Dynamics of hospitalizations in hemodialysis patients: results from a large US provider. Nephrol Dial Transplant 2014; 29: 442–448.)