Advance identification of patients with chronic conditions and acute respiratory failure at greatest risk for high-intensity, costly care

Abstract

Context:

Patients with underlying chronic illness requiring mechanical ventilation for acute respiratory failure are at risk for poor outcomes and high costs.

Objectives:

Identify characteristics at time of intensive care unit (ICU) admission that identify patients at highest risk for high-intensity, costly care.

Methods:

Retrospective cohort study using electronic health and financial records (2011–2017) for patients requiring ≥48 hours of mechanical ventilation with ≥1 underlying chronic condition at an academic healthcare system. Main outcome was total cost of index hospitalization. Exposures of interest included number and type of chronic conditions. We used finite mixture models to identify the highest-cost group.

Results:

4,892 patients met study criteria. Median cost for index hospitalization was $135,238 (range, $9,748-$3,176,065). Finite mixture modelling identified three classes with mean costs of $89,980, $150,603, and $277,712. Patients more likely to be in the high-cost class were: 1) < 72 years old (OR: 2.03; 95% CI:1.63, 2.52); 2) with dementia (OR: 1.55; 95% CI:1.17, 2.06) or chronic renal failure (OR: 1.27; 95% CI:1.08, 1.48); 3) weight loss ≥ 5% in year prior to hospital admission (OR: 1.25; 95% CI:1.05, 1.48); and 4) hospitalized during prior year (OR: 1.92; 95% CI:1.58, 2.35).

Conclusion:

Among patients with underlying chronic illness and acute respiratory failure, we identified characteristics associated with the highest costs of care. Identifying these patients may be of interest to healthcare systems and hospitals and serve as one indication to invest resources in palliative and supportive care programs that ensure this care is consistent with patients’ goals.

Introduction

Healthcare spending in the US is highly concentrated, with 5% of patients accounting for about 60% of costs.(1–3) This heterogeneous and medically-complex group of patients often receive high-intensity care that is burdensome, associated with increased pain and suffering, and discordant with their values, goals, and preferences.(4–6) These patients have been characterized as “high-need, high-cost” and may benefit from programs to improve quality and value of care.(5) Recognizing that a small subset of patients account for a vast majority of healthcare spending, these high-need, high-cost patients have elicited interest from clinicians, researchers, and policy makers.(7, 8) Prospectively identifying such patients is challenging and remains a strategic priority for healthcare systems worldwide.(9, 10)

For many high-need, high-cost patients, the majority of whom have non-curable chronic conditions, deterioration in health often results in acute respiratory failure requiring mechanical ventilation in the intensive care unit (ICU). This type of high-intensity care may be ineffective(11) and may place undue emotional, physical, and financial burden on patients and their families. For example, endotracheal tubes, intravascular lines, and restraints reduce mobility, may cause pain, and decrease the ability to communicate and exercise autonomy.(12) Patients receiving this kind of high-intensity care represent a growing population,(11, 13) are especially prone to poor patient- and family-centered outcomes,(14–22) and are at high risk of death, high hospital costs, and unmet palliative care needs.(23–28)

Acknowledging that high-need, high-cost patients represent a diverse group with varied circumstances and needs, we sought to identify patients who are at particularly high risk for high hospital costs among those with underlying chronic illness who develop acute respiratory failure requiring mechanical ventilation. Given the distribution of healthcare costs in seriously-ill populations, we supplemented standard methods that estimate associations between predictors and outcome at the sample mean with finite mixture modelling, which uses patterns to identify sub-groups that account disproportionately for spending.(29)

Methods

Design, setting, and participants

This investigation is a retrospective cohort study, and we followed the STROBE guidelines for observational studies (online supplement). The setting included all ICUs in two hospitals of an academic healthcare system in the Pacific Northwest of the US: a quaternary care facility for surrounding regions (79 ICU beds) and a county-owned safety-net hospital that serves as the only Level 1 Trauma Center for five states (89 ICU beds). The sample comprised all adult patients with at least one chronic condition documented in the electronic health record during the 24 months before ICU admission. Eligible chronic conditions were defined using the International Classification of Diseases (ICD) codes and based on the Dartmouth Atlas’s nine chronic conditions associated with a high probability of death (see Supplement)(30–32). These conditions account for 90% of all Medicare beneficiary deaths(33). Eligible patients received at least 48 hours of mechanical ventilation in the ICU in one of the hospitals between January 1, 2011, and December 31, 2017. For patients with multiple eligible stays, we used the first stay. Data on receipt and duration of mechanical ventilation were abstracted from an RN-charted structured field in the electronic health record. The field has been validated against respiratory therapist-charted structured fields and by manual physician-investigator chart review.(34) Patients were attributed to the facility that provided most of the patient’s care during the 24 months before ICU admission.(30)

Data sources and variables

The primary data sources were the healthcare system’s electronic health and financial records, which were collected for clinical and administrative purposes. Data regarding 6-month mortality status were taken from Washington State death certificates for 2011–2018. The outcome of interest was the total cost of the index hospitalization, including direct and indirect costs associated with ICU care, acute care, and professional services. Costs were extracted from the system’s accounting database and reflect the best available quantification of resources attributable to care of a specific patient.(35) Direct costs vary at the individual patient level, reflecting medications, procedures, tests, consumables, and proportionate staffing costs. Indirect costs (e.g., for management, maintenance, and cleaning) are not attributable to a specific patient but shared proportionately across patients. Costs were adjusted for inflation to represent their value on June 1, 2019.

We examined 19 potential predictors of high-cost care, representing four categories: (1) demographic variables (age, sex [male or female], racial or ethnic minority status [white non-Hispanic vs. minority], and insurance coverage [private, Medicare, Medicaid, military, other type, uninsured]); (2) diagnoses (nine chronic conditions); (3) health status (diagnosis with at least three of chronic conditions; the Deyo-Charlson comorbidity index(36, 37); exploratory markers of frailty(38–43), including a documented weight loss of 5% or more during the year before ICU admission(38, 39) and an albumin test result under 3 gm/dL during the first 48 hours in the ICU(40–43)); and (4) prior healthcare utilization(6) (more than one hospitalization in the year before index admission, and hospitalization costs ≥$27,290 during that year).

In addition, we used three variables for adjustment purposes. Because costs for patients attributed to the safety-net hospital were, on average, significantly lower than for those attributed to other facilities, we used this binary variable to adjust for possible confounding. In addition, we did sensitivity analyses, adjusting for mortality status at hospital discharge (controlling for the shorter average hospital stays and hence lower costs of decedents and detailed further below) and calendar year of admission (controlling for possible temporal changes in intensity of care). The University of Washington institutional review board approved this study (Study ID#:00004132).

Addressing Bias

This is a descriptive observational study reporting associations between observed baseline factors and cost of hospitalization. We considered the principal risk of bias to be that introduced by omitted variables: in particular, important predictor(s) associated with the outcome that were not available at baseline. Omission of such variables may distort the reported associations for available predictors. We hypothesized ex ante that proximity to death was the most important potential source of omitted-variable bias because of its strong association with both outcome and most baseline predictors. Although we tried to control for this by using variables that were available at baseline and typically predictive of hospital mortality (36), these variables had poor prognostic power in our sample. As a result, we conducted a sensitivity analysis of our main results, adjusting for discharge status (alive/dead) in addition to our routine adjustment for hospital site.

Sample size

The sample size for the study was determined by the number of years of data available and judged to be complete in the health system’s electronic data bank. Data were complete and available from both electronic health records and financial records for calendar years 2011 through 2017, inclusive. The initial sample included all patients meeting study eligibility requirements whose ICU admission occurred by January 1, 2011, and whose ICU discharge occurred by December 31, 2017. This provided an initial sample of 5,427 patients.

Statistical methods

Standard regression methods return coefficients for predictors at the sample mean (e.g., what is the average effect on total cost of hospital admission of a subject having a cancer diagnosis versus not having a cancer diagnosis, holding all other predictors constant?) Given that a minority of people account for a majority of costs, this estimated mean effect for the sample may not be the parameter of greatest interest. Instead, we are most often most interested in the high-cost minority: what is the average effect of specific characteristics on costs for this high-cost group, and what are the factors that predict membership of that group? To explore these questions, we supplemented standard cost regressions with finite mixture modelling. Finite mixture models divide observations into classes according to differences of the effects of covariates on outcome.(44) (For example, early research found weak effects between prenatal care and infant health when pooling all pregnancies in analysis, but finite mixture models identified a large sub-class for whom prenatal care was highly beneficial. When pooling all observations in a standard regression, this effect was disguised by a small number of complicated and influential cases for whom the intervention was not effective.(29)) By using finite mixture models in our study, we explored a new approach for managing the substantial latent heterogeneity in observational data on people with advanced medical illness, thereby improving our ability to characterize the high-cost minority and the factors that may contribute to high-intensity healthcare use.

Beginning with the 19 potential baseline predictors, we tested bivariate and multi-predictor linear regression models of log-transformed costs, retaining variables with nonnegligible (p<0.10) associations with outcome, categorizing continuous variables and recoding categorical variables where it was parsimonious to do so. Of the initial predictors, only one had missing data: racial and ethnic minority status (2.3%). We found this variable had no significant association with costs, therefore allowing inclusion of the full sample in final models.

Having determined from the linear regression models a set of promising predictors, we ran a series of generalized linear models (GLMs) using a Huber/White/sandwich estimator of the variance/covariance matrix, determining the best-fitting family and link by comparing values on the Akaike and Baysian information criteria. We then retested predictors in the optimal nonlinear model to rule out collinearity issues, determine whether some of the predictors that were ruled out in the preliminary linear models should be restored, and eliminate predictors with p-values >0.10.

Next, we used the final GLM as the basis for finite mixture models (FMMs), using a GLM estimator with the family and link from the final GLM, to test models with 2, 3, and 4 classes, determine the optimal number of classes (based on information criteria), compare the estimated means for the resulting classes, and calculate the probability of each patient’s assignment to each class. Finally, we computed a binary variable indicating whether the patient’s most likely membership was in the highest-cost class and ran multi-predictor logistic regression models of this binary outcome, beginning with all predictors used in the final GLM, and sequentially removing predictors by descending value of p-value until only the predictors with p<0.05 remained.

Data were analyzed using STATA, version 16.0 (StataCorp., College Station, TX) statistical software.

Study Results

Study participants and characteristics

From our initial sample of 5,427 eligible patients, 535 were dropped because of mismatches between electronic health records and financial records with regard to patients’ medical record numbers. This left 4,892 patients who met study criteria and had both health and financial data available (Table 1). Admissions by calendar year for these 4,892 patients ranged from 648 (in 2013) to 770 (in 2015). A majority of patients were male (64%) and white non-Hispanic (74%), with median age of 60 years (ranging from 18 to 98 years). Over one-third (35%) had Medicaid insurance, with smaller groups having Medicare (24%) or private insurance (30%), and 3% being uninsured. More patients had heart failure (43%), coronary artery disease (41%), or lung disease (35%) than other chronic illnesses; 38% had at least three of the nine chronic conditions; and patients’ median score on the weighted Deyo-Charlson comorbidity index was 3 (ranging from 0 to 16). Almost half (49%) had an albumin test result less than 3 gm/dL during the first 48 hours in the ICU. More than one-fifth of the sample (21%) had documented weight loss of 5% or more during the year before hospital admission, and almost one fourth (24%) had been hospitalized during that year.

Table 1.

Sample Characteristics

Characteristic	Valid n	n (%)
Female	4,892	1,776 (36.3)
Racial/ethnic minority	4,779	1,259 (26.3)
Age	4,892
18–47 years		987 (20.2)
48–56 years		920 (18.8)
57–63 years		1,024 (20.9)
74–71 years		1,002 (20.5)
72–98 years		959 (19.6)
Insurance coverage	4,892
Private		1,444 (29.5)
Medicare		1,196 (24.4)
Medicaid		1,707 (3 4.9)
Military		203 ( 4.1)
Miscellaneous other type of insurance		193 ( 3.9)
Uninsured		1 9 ( 3.0)
Specific chronic conditions	4,892
Cancer		981 (20.1)
Lung disease		1,717 (35.1)
Coronary artery disease		2,023 (41.4)
Heart failure		2,119 (43.3)
Peripheral vascular disease		1,087 (22.2)
Liver disease		737 (15.1)
Diabetes		1,112 (22.7)
Renal failure		1,412 (28.9)
Dementia		315 ( 6.4)
Multi-morbidity: 3 or more chronic conditions	4,892	1,878 (38.4)
Deyo-Charlson comorbidity index	4,892
0–2		1,752 (35.8)
3		702 (14.3)
4		697 (14.2)
5–6		1,026 (21.0)
7–16		715 (14.6)
Albumin test result <3, 1st 48 hours of index ICU stay	4,892	2,386 (48.8)
5% weight loss, year before index ICU admissiona	4,892	1,009 (20.6)
Attribution to the county safety-net hospital	4,892	2,820 (57.6)
Hospitalizations in year before index admission	4,892	1,169 (23.9)
0		3,723
1		597
2–18	223	572
Hospital costs in year before index admission	4,892
$0.00		2,713 (55.5)
$0.01 – $1,890.99		223 ( 4.6)
$1,891 – $27,289.99		978 (20.0)
$27,290 – $1,390,000		978 (20.0)
Cost of index hospitalization	4,892
$9,748 – $67,168		979 (20.0)
$67,168 – $111,437		978 (20.0)
$111,437 – $164,52		979 (20.0)
$164,852 – $260,846		978 (20.0)
$260,846 – $3,176,065		978 (20.0)
Death during index hospitalization	4,892	1,709 (34.9)
Death within 6m of index hospital admissionb	4,892	2,091 (42.7)

^aPresence of evidence in the electronic health record of ≥ 5% weight loss in year prior to ICU admission

^bBased on Washington State death certificate data (2011–2018)

Main results

Median cost for the index hospitalization was $135,238, ranging from $9,748 to $3,176,065. More than one-third of the sample (35%) died during the index hospitalization (median costs for those who died: $102,918; for those who survived: $152,113), and slightly more than one-half (58%) were attributed to the county safety-net hospital (median costs for county hospital: $126, 996; median costs for other hospitals: $146, 341) (Table 1).

Preliminary linear regression analysis of log-transformed total costs reduced the number of promising predictors to 12. GLM analyses of the raw total-cost scores (using gamma family and log link) produced estimates for a final set of 10 predictors with p-values <0.10 (Table 2). We used these 10 predictors to evaluate finite mixture models (FMM) with two, three and four classes, comparing model performance using AIC and BIC, and identifying the three-class model as performing best. All GLM and FMM analyses included the indicator for the lower-cost safety-net hospital as a covariate adjustment. The FMM estimated that 1,076 (22%) of the 4,892 patients were in the low-cost class (with mean total costs of $89,980); 2,104 (43%) were in the middle class (with mean total costs of $150,603); and 1,712 (35%) were in the high-cost class (with mean costs of $277,712). (Table 3). When patients were assigned to the class representing their most-likely membership, 1,222 (25.0%) were in the lowest-cost class; 2,682 (54.8%), in the middle class; and 988 (20.2%), in the highest-cost class.

Table 2.

Final Generalized Linear Model Used as Basis for Finite Mixture Modeling, Adjusted for Attributed Facility^a

Patient Characteristic	Coefficient	p	95% CI
Age
18–47 years	0.000
48–71 years	−0.194	<0.001	−0.264, −0.124
72–98 years	−0.471	<0.001	−0.557, −0.384
Insurance coverage
Private, military, other	0.000
Medicare	−0.062	0.049	−0.124, −0.000
Medicaid, uninsured	−0.043	0.150	−0.102, 0.016
Cancer	−0.074	0.029	−0.140, −0.007
Lung disease	0.054	0.041	0.002, 0.106
Peripheral vascular disease	0.050	0.060	−0.002, 0.102
Renal failure	0.201	<0.001	0.142, 0.260
Dementia	−0.086	0.0 60	−0.177, 0.004
5% or more weight loss, year before admission	−0.(71	0.026	−0.133, −0.009
More than 1 hospitalization in year before admission	−0.194	<0.001	−0.294, −0.093
Hospital costs ≥$27,290 in year before admission	0.303	<0.001	0.223, 0.383

^aThis model resulted from preliminary analyses of 19 potential predictors, with removal of 9 predictors having p-values >0.10. The generalized linear model used robust estimation of the variance-covariance matrix and was based on the gamma family and log link, which produced the best fit according to AIC and BIC values. In addition to the predictors shown in the table, the model was adjusted for whether the patient was attributed to the lower-cost safety-net hospital.

Table 3.

Mean Probability of Membership and Total Costs by Class, Based on an Estimated Finite Mixture Model^a

Class	Probability of Membership	Total Costs
	Mean (95% CI)	Mean (95% CI)
1	0.22 (0.10, 0.42)	$89,980 ($ 70.790, $109,170)
2	0.43 (0.32, 0.55)	$150,603 ($120,752, $180.455)
3	0.35 (0.21, 0.52)	$277,712 ($236,158, $319,266)

^aThe finite mixture model used a GLM estimator with gamma family and log link, using the 10 predictors shown in Table 2, plus adjustment for whether the patient was attributed to the lower-cost safety-net hospital. The mean probability of membership in each class estimates the proportion of the sample in each class, based on the model.

After computing a binary outcome variable indicating whether each patient’s most-likely class membership was in the highest-cost class, we found five predictors with statistically significant independent associations with membership in that class, after adjustment for whether the patient was cared for at the safety-net hospital (Table 4). Patients were more likely to be in the high-cost class if they were younger than 72 years old (OR: 2.03; 95% CI:1.63, 2.52), had dementia (OR: 1.55; 95% CI:1.17, 2.06) or chronic renal failure (OR: 1.27; 95% CI:1.08, 1.48), had experienced weight loss of 5% or more during the year prior (OR: 1.25; 95% CI:1.05, 1.48), or had been hospitalized during that year (OR: 1.92; 95% CI:1.58, 2.35).

Table 4.

Characteristics Associated with Membership in Highest-Cost Class, Adjusted for Attributed Facility and Other Covariates^a

Model #	Characteristic	p	Odds Ratio	95% CI
1	Aged <72 years	<0.001	2.026	1.630, 2.520
	Dementia	0.002	1.553	1.169, 2.063
	Chronic renal failure	0.003	1.265	1.084, 1.476
	5% weight loss, year before index admission	0.011	1.246	1.051, 1.478
	More than 1 hospitalization in year before admission	<0.001	1.923	1.577, 2.345
	Not attributed to “safety net” hospital	<0.001	1.640	1.417,1.899
2	Aged <72 years	<0.001	1.883	1.511, 2.345
	Dementia	0.009	1.460	1.097, 1.943
	Chronic renal failure	0.006	1.247	1.067, 1.457
	5% weight loss, year before index admission	0.003	1.296	1.091, 1.538
	More than 1 hospitalization in year before admission	<0.001	2.084	1.703, 2.549
	Not attributed to “safety net” hospital	<0.001	1.691	1.459,1.959
	Died during hospitalization	<0.001	0.570	0.485, 0.671
3	Aged <72 years	<0.001	2.021	1.625, 2.513
	Dementia	0.002	1.554	1.170, 2.065
	Chronic renal failure	0.002	1.270	1.088, 1.482
	5% weight loss, year before index admission	0.014	1.237	1.043, 1.467
	More than 1 hospitalization in year before admission	<0.001	1.924	1.578, 2.347
	Not attributed to “safety net” hospital	<0.001	1.639	1.416,1.898
	Year of hospital admission (ordinal)	0.038	1.039	1.002,1.077

^aEach of the three models is based on a multi-predictor logistic regression model of the binary outcome indicating whether the patient’s most likely membership was in the highest-cost group from the finite mixture model. The non-italicized variables are predictors from the finite mixture model that had p<0.05 when used as predictors of the binary outcome. The italicized variables were used to adjust for possible confounding between the predictors of interest and membership in the highest-cost class.

Sensitive analyses

We then did two sensitivity tests (Table 4), each adding one adjustment variable to the model with only hospital. The first adjusted for the patient’s mortality status at hospital discharge. The same five predictors remained significant, and in the same direction, although four of the non-exponentiated coefficients changed by >10% with adjustment (weight loss, increasing by 17.6%; dementia decreasing by 14.1%; the top quintile of number of hospitalizations in the year prior to admission increasing by 12.2%; and the lower quintiles of age decreasing by 10.4%).

The second sensitivity test adjusted for calendar year of the index hospital admission, in addition to the adjustment for hospital. Again, the same five predictors remained significant and in the same direction. The largest change in a non-exponentiated coefficient was just 3.3%, suggesting that there was virtually no confounding by year of admission.

Discussion

In this study, we propose a novel analytic approach as a first step in identifying high-need, high-cost patients within a heterogeneous group of patients with serious chronic conditions requiring mechanical ventilation. Our inclusion criteria of at least one serious chronic condition and 48 hours of mechanical ventilation were chosen to identify a sample already at high risk for poor outcomes and high-intensity, costly care.(14–20) To our knowledge, no ICU prognostic models have been developed to predict high costs within a cohort of patients with underlying chronic conditions who develop acute respiratory failure. Our sample had a median hospital length of stay (LOS) of 20 days, ICU LOS of 12 days, 7 days of mechanical ventilation, and 35% hospital mortality. We used finite mixture modeling to understand the cost distribution and to further segment these patients into categories differentiated by the association of costs with characteristics known at ICU admission. Even among a cohort receiving high-intensity care (mechanical ventilation) in a high-cost setting (the ICU), we found a skewed distribution of costs, with 20% of patients accounting for 46% of the costs. Although this degree of skewness is not as dramatic as the often cited “20% of people account for 80% of costs,”(7) we still found a minority of patients accounted for a majority of costs. To the extent that high costs may reflect high-intensity care that some patients and families might consider potentially burdensome, identifying these patients may have important implications for identifying potentially-burdensome care and improving the quality and value of care.(5, 6) We found that dementia, chronic renal failure, age < 72, frailty (suggested by weight loss in the past year), and high-frequency prior hospitalizations were associated with membership in the highest class of hospital costs—a class with mean hospitalization costs of $420,523 (median = $351,868) and median ICU LOS of 31 days and hospital LOS of 46 days.

Teno and colleagues previously reported that among nursing home residents with advanced dementia, from 2000 to 2013, the use of mechanical ventilation steadily increased without improvement in survival.(11) Our finding that dementia was associated with membership in the highest category of costs adds to the evidence base that these patients are at particularly high risk for low value care if admitted to the ICU. Efforts to promote tailor palliative and supportive care interventions to their needs warrant increased attention.

For over a decade, the American Society of Clinical Oncology has recommended that combined standard oncology care and palliative care be considered early in the course of illness for any patient with metastatic cancer and/or high symptom burden.(45) We found that having a diagnosis of cancer with poor prognoses was associated with lower costs, suggesting that perhaps efforts to promote palliative care involvement have been successful in reducing high-intensity, costly care for these patients. More recently, integrating palliative care into the care of patients with other serious illnesses such as advanced heart failure(46) and chronic kidney disease(47) has received increasing attention. Our finding that chronic renal failure was associated with membership in the highest category of costs suggests that ongoing efforts are warranted to ensure this care is goal-concordant.

Recent randomized trials of interventions designed to improve palliative and supportive care in the setting of critical illness have resulted in weakly positive, negative, or even harmful results.(48–50) One explanation for these mixed results is that these studies did not tailor interventions to individual palliative and supportive care needs.(51) Our results may help to identify a subset of patients at particularly high risk for high-intensity, costly care who may have unique needs and might benefit from specific interventions targeting those needs.

Relying exclusively on the diagnosis of chronic conditions to identify patients who would benefit most from targeted interventions has received criticism for not accurately identifying those who are “high-need.” Although many of the patients in our cohort likely have unmet needs and would benefit from supportive and palliative care interventions, efficient use of palliative and supportive care resources are an important component of sustainability and may be facilitated if some of these resources are directed towards patients and families who are at risk for high-cost and potentially goal-discordant care.(51, 52) Kelley and colleagues demonstrated that inclusion of measures of functional limitation and high prior healthcare utilization may improve identification of seriously ill patients with high costs and high needs and likely to benefit from palliative and supportive care interventions.(6, 9) Our exploration of weight loss in the prior year as an exploratory measure of frailty(38–43) and high prior healthcare utilization(6) may be additional risk factors for identifying those at particularly high risk for costly care. Both of these factors were associated with membership in the highest category of costs.

Our study has several strengths. First, we used characteristics present at ICU admission and easily abstracted from electronic health records via automated processes. Leveraging the electronic health record is a more feasible and cost-efficient approach than relying on manual abstraction. Second, we did not rely on a retrospective decedent cohort, as our goal was to identify patients at risk for high-intensity, costly care at the time of ICU admission.(53) Third, our inclusion criteria of acute respiratory failure and underlying chronic illness captured a high morbidity, high mortality, and high cost population; this may serve as a useful denominator population for further research. Lastly, we used rigorous, novel statistical methods to identify class membership for the highest category of costs. We selected these methods because standard regression models to predict high utilization suffer from weak identification – in particular low sensitivity.(9) One possible reason for this is the unusual distribution of healthcare costs complicates the identification of determinants.

Our study has several important limitations. First, our data were collected from one health system in one region of the country and may not generalize to other systems or regions. Second, we found that mortality confounded our results; however, in sensitivity analyses adjusting for survival, the same predictors remained significant. Third, our outcome of interest was total hospital costs; we did not collect data on quality of care and thus cannot directly equate high-cost care with low-value care. However, we selected eligibility criteria that identify a patient population well documented to have high mortality and poor outcomes.(14–20) Fourth, data from the EHR may contain inaccuracies, a problem inherent in all studies utilizing the EHR(54). Lastly, our sensitivity analysis showing strong association with mortality raises a familiar problem in modelling hospital costs in populations with serious illness: costs are strongly positively associated with both mortality and LOS, and mortality and LOS are negatively correlated. Since each of these variables is known only at the end of the care episode it is not possible to adjust for them prospectively.(55) Both new data and new methodological research is needed to advance understanding of these dynamics.

Conclusions

Among a cohort of patients with acute respiratory failure and underlying chronic illnesses, we used a novel analytic approach to identify characteristics present at the time of ICU admission that were associated with very high costs. Depending on the patient’s goals, this high-cost care could potentially place these patients at high risk for low-value health care. Identifying these patients may be of particular interest to healthcare systems and serve as one indication to invest resources in palliative and supportive care programs that ensure this care is consistent with patients’ goals and address the unique needs of this population.

Key Message.

Among patients with underlying chronic illness who develop acute respiratory failure, we identified characteristics associated with high-intensity, costly care. Identifying these patients may be of interest to healthcare systems and hospitals and serve as one indication to invest resources in palliative and supportive care programs that ensure this care is consistent with patients’ goals.