Natural Language Processing to Assess Palliative Care Processes and Health Care Utilization in Seriously Ill Older Adults with Severe Trauma

Abstract

Background:

National guidelines recommend palliative care (PC) alongside life-sustaining treatment for older adults with severe trauma. However, outcomes associated with PC for these patients are not well-defined.

Objectives:

To determine frequency of inpatient PC process documentation in older adults with severe trauma and test associations with postdischarge health care utilization.

Design:

Retrospective cohort study using electronic health record data linked to Medicare claims.

Setting/Subjects:

We included adults ≥66 years old admitted to a large, regional U.S. health care system with severe trauma (2016–2018) using consensus criteria for serious illness in trauma.

Measurements:

Natural language processing was used to measure documentation of five inpatient PC processes: code status limitations, goals-of-care (GOC) conversations, hospice discussions, PC consultations, and health care proxy designations. Associations between PC processes and postdischarge health care utilization were tested using multivariable regression.

Results:

Among 1267 admissions, the median age was 82 years (interquartile range [IQR] 75–88), and median injury severity score (0–75, higher is worse) was 16 (IQR 9–21); ≥1 PC process was documented in 81%. Among those surviving hospitalization (87%), one-year mortality was 26%. Documentation of ≥1 PC process was not significantly associated with differences in mean hospital days (16 vs. 19), home days (306 vs. 307), emergency department visits (2.3 vs. 2.2), or intensive care unit days (0.6 vs. 0.9) at one year. PC processes were significantly associated with subsequent hospice enrollment (p < 0.01).

Conclusions:

PC was not associated with reduced health care utilization in older adults after trauma but was associated with one-year hospice enrollment. GOC conversations, specialty PC, and inpatient hospice discussions had low utilization, highlighting target areas for improvements in care delivery.

Keywords

natural language processing palliative care serious illness surgery trauma

Key Message

This study leveraged a novel methodology using natural language processing of electronic health record data linked to Medicare claims to ascertain delivery of PC processes for older adults admitted with trauma. Despite national guidelines, several elements of PC were underutilized, highlighting potential target areas for future research.

Introduction

Traumatic injury in older adults leads to more than 1,500,000 hospitalizations, 85,000 deaths, and $1 trillion in cost in the United States annually.¹ Older adults have worse functional and quality-of-life outcomes after trauma compared with young adults due to frailty and comorbidity burden.^2–5 Palliative care (PC), which focuses on symptom relief and quality of life for patients and families with serious illness,⁶ has been recognized as a critical component of geriatric traumatology.⁷ A consensus definition of serious illness in surgery has outlined criteria for trauma patients eligible for PC, including those with severe traumatic brain injury (TBI) or critical injury.⁸ In 2017, the American College of Surgeons Trauma Quality Improvement Program (ACS TQIP) published best practice guidelines that championed PC delivery in parallel with life-sustaining trauma interventions.⁷ These recommendations include elements of primary PC as well as screening criteria for specialty PC involvement.⁹

Despite the established benefits of PC on quality of life and health care utilization in nonsurgical patients,^10–12 the scope of its impact in older trauma patients is not well-defined. Associations between PC and decreased use of invasive procedures, shorter hospital stays, and decreased health care utilization in older trauma patients have been inconsistently demonstrated in preliminary studies.^13–18 These studies are limited to specialty PC consultations, which are underutilized in trauma patients^19,20 and tend to be reserved for end-of-life scenarios.²¹ A retrospective analysis of seriously ill trauma patients found that early goals-of-care (GOC) conversations—a core primary PC process embraced by TQIP––were infrequently documented (<20%) and mainly conducted for sicker patients on ventilators.²² Efforts to incorporate TQIP guidelines via institutional trauma quality improvement protocols^23,24 as well as programs to support primary PC education^25,26 have demonstrated feasibility; however, their frequency of use and associated impact on outcomes remain unclear. This gap was highlighted by an expert review that found insufficient evidence for the impact of routine PC processes on outcomes in geriatric trauma.²⁷ Consensus trauma research agendas have emphasized it as a priority area for future investigation.²⁸

To address this gap, we sought to (1) determine the frequency of documentation of five PC processes in a cohort of older adults (≥66 years old) admitted to a large academic health system with an injury that met the consensus definition of serious illness in trauma, and (2) test associations between PC processes and postdischarge health care utilization. We hypothesized that patients who received PC processes would have reduced health care utilization and increased hospice enrollments in the year after discharge.

Methods

Study design and data source

We conducted a retrospective cohort study of trauma admissions from 2016 to 2018 at Mass General Brigham (MGB), a large regional health care system in Massachusetts, identified through institutional trauma registries at two level I trauma centers and one level III trauma center. Trauma admissions were included for patients ≥66 years old who met a previously described consensus definition for serious illness in trauma: (1) severe TBI with Head Abbreviated Injury Scale (AIS) ≥3 (0–6 scale; higher is worse) or (2) critical injury defined as >24-hour intensive care unit (ICU) stay or Injury Severity Score (ISS) >25 (0–75 scale; <9 mild, 9–15 moderate, 16–25 severe, >25 very severe).⁸ Patients transferred to a non-MGB trauma center after initial stabilization were excluded. The cohort was linked to the MGB Research Patient Data Registry (RPDR), a centralized institutional repository that hosts clinical and administrative data, including all free-text clinical notes associated with the admission. Finally, the cohort was linked to Medicare claims data to assess both preexisting serious illness diagnoses and post-trauma health care utilization outcomes. Patients without fee-for-service Medicare data one year before trauma and one year following trauma or until death were excluded. The study period of 2016–2018 was chosen based on availability at the time of study conception of both electronic health record (EHR) data and linked Medicare claims data to ensure adequate one-year look back and follow-up. This study was approved by the MGB Institutional Review Board. We followed the Strengthening the Reporting of Observational Studies in Epidemiology guidelines for reporting cohort studies.²⁹

Natural language processing to identify PC processes

All clinical notes associated with the index trauma admission for the final cohort were analyzed using natural language processing (NLP) software (ClinicalRegex, version 1.0.3; Lindvall Lab, Dana-Farber Cancer Institute), as previously reported.^30–33 This technique leverages predefined keyword and phrase libraries (Supplementary Table S1) to support NLP-assisted human adjudication for identifying the presence or absence of documentation (yes/no) for five distinct PC processes: (1) GOC conversations, (2) code status limitations, (3) specialty PC consultations, (4) hospice assessment, and (5) surrogate decision maker or health care proxy identification. These processes are endorsed by the National Consensus Project for Quality Palliative Care and ACS TQIP Guidelines as quality indicators of both primary and specialty PC delivery.^7,34 Two human coders (D.I.H. and S.M.) used a previously described and validated NLP codebook to identify documented PC processes.³⁵ Inter-rater reliability between the two coders was assessed via a Cohen’s Kappa statistic on a random test cohort of N = 50 admissions. The final cohort was stratified by receipt of any PC process (defined as documentation of at least one of the five processes identified via NLP) or none. An exploratory NLP analysis was conducted to determine the timing of GOC conversations (earlier defined as within first 72 hours of admission).

Independent variable and covariates

Our primary independent variable was defined as binary receipt of any (at least one) PC process during index admission as measured by NLP. For our secondary analysis, each of the five PC processes was treated individually as independent variables. The following patient, injury, and admission covariates were collected using Medicare claims, RPDR, and institutional trauma registries: age, sex, race/ethnicity, marital status, domicile, Charlson comorbidity index, hospital length-of-stay (LOS), ICU utilization, days in ICU, head AIS, ISS, in-hospital mortality, and discharge location. In addition, using Medicare claims data from one year before trauma, the presence of a preexisting serious illness was identified with a previously validated technique³⁶ that uses International Classification of Diseases 10th edition (ICD-10) codes operationalized from a consensus definition of 10 serious illnesses or health states for surgical patients.⁸

Health care utilization and health outcomes

The following health care utilization and health outcomes were measured from the trauma admission discharge date up to one year of follow-up or until death using Medicare claims data: emergency department (ED) visits, days in hospital, days at home, hospice enrollments, ICU admissions, days in ICU, readmissions, and mortality. Days at home was considered the primary outcome and was determined by days alive in the one year after hospital discharge from index admission, excluding inpatient admissions, ED visits, and days at skilled nursing facilities or other nonacute institutional stays. Patients who died during their index admission were excluded from outcomes analysis.

Statistical approach

Descriptive statistics on baseline covariates in the final cohort were calculated between patient admissions stratified by those who received any PC process versus those with no PC processes, using the Mann–Whitney tests for continuous variables, Pearson chi-square tests for categorical variables, and Fisher exact tests for categorical measures when cell counts were less than five. Exact values of cells with counts <11 were suppressed in accordance with the Centers for Medicare and Medicaid Services guidelines.

For the primary outcome analysis, associations between receipt of any PC process during an index admission and one-year health care utilization outcomes were assessed with multivariable Poisson regression models with time-alive as an offset. Clinically relevant baseline patient or trauma admission characteristics with a p value ≤0.1 on univariable analysis were included as covariates in our model. Final covariates included in the model included age (65–74 vs. 75–84 vs. 85+), sex, marital status (married vs. unmarried), domicile (home/self-care vs. facility), preexisting serious illness (none vs. 1–2 vs. 3+), ISS (<9 vs. 9–15 vs. 16–24 vs. 25+), hospital LOS (continuous log transformed), and ICU utilization during index admission (binary). Outcomes were reported both as adjusted mean number of events or cumulative probability per 365 patient-days of follow-up and as incidence rate ratios with no PC as the reference group. Values were reported with 95% confidence intervals (CI). P-values <0.05 were considered statistically significant. As a secondary analysis, these regressions were also run in a similar manner to test the associations between exposure to each of the five PC processes individually and health care utilization outcomes. All computations were conducted using SAS version 9.4 (SAS Institute) and R version 4.2.1.

Results

Patient and trauma admission characteristics

Our final cohort yielded 1267 severe trauma admissions from 2016 to 2018 (Supplementary Fig. S1). Baseline patient and injury characteristics are summarized in Table 1. The cohort had a median age of 82 years, was mostly White (94.1%), and less than half presented from home/self-care (43.5%). Most patients had preexisting serious illness (76.1%). The most common qualifying severe trauma criterion was critical injury with ICU stay >24 hours (65.3%), followed by severe TBI (60.7%). A total of 30.9% of patients met more than one of the inclusion criteria. The median ISS was 16, and 11.0% of patients had very severe injury with ISS ≥25.

Table 1.

Patient and Hospitalization Characteristics of Severe Trauma Admissions at a Regional Health System (2016–2018)

	Total (N = 1267)	Any palliative care process		p Value
	Total (N = 1267)	No (N = 240)	Yes (N = 1027)	p Value
Patient characteristics
Age (years), median (IQR)	82 (75–88)	79 (73–84)	83 (76–88)	<0.001
Sex (female), %	50.4	42.9	52.1	0.01
Race and ethnicity, %				0.18
White	94.1	95.0	93.9
Black	2.0	<4.6	>1.1
Hispanic	>0.9	<4.6	>1.1
Asian	2.2	<4.6	>1.1
North American Native	<0.9	<4.6	<1.1
Marital status, %				0.10
Married	44.8	52.1	43.0
Divorced	7.3	>4.6	>1.1
Widowed	32.0	25.4	33.6
Single	13.7	13.3	13.8
Unknown	2.2	<4.6	>1.1
Domicile, %				0.24
Home/self-care	43.5	48.8	42.3
Other acute inpatient	53.5	>4.6	>1.1
Skilled nursing facility	1.8	<4.6	>1.1
Other/unknown	1.2	<4.6	>1.1
Preexisting serious illness, %
Vulnerable elder	63.4	43.3	68.1	<0.001
Advanced cancer	5.9	4.6	6.2	0.33
Pulmonary disease	5.2	<4.6	>1.1	0.04
Heart failure	29.8	21.3	31.7	0.001
Cirrhosis	3.9	<4.6	>1.1	0.22
End-stage renal disease	7.0	<4.6	>1.1	0.05
Dementia	13.4	6.3	15.1	<0.001
Frailty	31.3	20.8	33.7	<0.001
Prior severe trauma	23.2	18.3	24.3	0.05
Nursing home resident	2.7	0	3.3	0.001
≥1 Preexisting serious illness	76.1	58.8	80.1	<0.001
≥2 Preexisting serious illness	49.8	32.5	53.8	<0.001
≥3 Preexisting serious illness	32.4	19.6	35.4	<0.001
Charlson Comorbidity Index, %				0.002
<2	24.2	32.1	22.4
≥2	75.8	67.9	77.6
Trauma admission characteristics
Qualifying severe trauma type, %
Severe TBI with AIS 3+	60.7	70.4	58.4	<0.001
ICU stay >24 hours	65.3	44.2	70.2	<0.001
ISS >25	11.0	7.1	12.0	0.03
≥2 factors	30.9	19.6	33.5	<0.001
ISS^a, median (IQR)	16 (9–21)	15 (10–19)	16 (9–21)	0.83
Length of index admission (days), median (IQR)	5 (3–9)	4 (2–6)	6 (3–10)	<0.001
Any ICU stay during index admission, %	71.7	52.1	76.3	<0.001
Days in ICU during index admission, median (IQR)	2 (0–4)	1 (0–3)	2 (1–4)	<0.001
No. of palliative care processes, %
0	18.9	—	—	—
1	39.0	—	—	—
2	5.1	—	—	—
3	18.2	—	—	—
4	14.3	—	—	—
5	4.5	—	—	—
Discharge location, %				<0.001
Home/self-care	26.6	49.2	21.3
Other acute inpatient or rehabilitation	21.6	26.3	20.5
Long-term care facility	35.0	22.5	37.9
Hospice	3.5	0	4.3
In-hospital mortality	13.3	<4.6	>15.4

Cohort is stratified by documentation of any versus no palliative care processes.

Injury severity score: 0–75 scale; <9 mild, 9–15 moderate, 16–25 severe, >25 very severe.

AIS, Abbreviated Injury Scale; ICU, intensive care unit; IQR, interquartile range; ISS, Injury Severity Score; TBI, traumatic brain injury.

NLP identification of inpatient PC processes

A random sample of 50 admissions was double-coded by both the human coders to ensure adequate inter-rater reliability; Cohen’s Kappa coefficient in this sample was 0.93 (95% CI 0.88–0.98), indicative of near-perfect agreement.

There were 55,604 inpatient notes analyzed using NLP. Median notes per admission was 29 (interquartile range [IQR] 19–50). Among 1267 admissions, 81.1% had any documented PC process. Documentation frequency of each PC process is illustrated in Figure 1. The most common process was health care proxy designation (74.7%), and 42.1% of admissions had two or more processes documented. Of patients who had a GOC conversation (38.3%), 75.3% of these conversations were documented within the first 72 hours of admission.

FIG. 1.

Frequency of palliative care (PC) processes in trauma admissions. Documentation of five PC processes in N = 1267 severe trauma admissions in adults ≥66 years old from 2016 to 2018 was analyzed via natural language processing: (1) code status limitations, (2) goals of care (GOC) conversation, (3) hospice discussion, (4) specialty PC consultation, and (5) health care proxy designation.

Compared with patients without PC process documentation, those with any PC process documentation were older (83 vs. 79 years) and more often female (52.1% vs. 42.9%), with higher rates of preexisting serious illness (80.1% vs. 58.8%), ICU utilization (76.3% vs. 52.1%), discharge to facility (58.4% vs. 48.8%), and in-hospital mortality (>15.4% vs. <4.6%); all ps < 0.05.

Health care utilization within one year of discharge

There were 1098 patients who survived index hospitalization and were included in the postdischarge health care utilization analysis. Unadjusted one-year health care utilization and health outcomes are summarized in Supplementary Table S2. The combined in-hospital and one-year mortality rate was 39.4%. At one year, 63.2% had at least one hospitalization, and 16.4% had at least one ICU stay with a median length of ICU stay of three days (IQR 2–5). Median days at home within one year was 321 (IQR 193–350), and 17.7% of patients enrolled in hospice.

Multivariable Poisson regression controlling for age, sex, marital status, domicile, preexisting serious illness, ISS, hospital LOS, and index ICU use was used to test associations between first, receiving any PC processes and outcomes, and second, each of the individual five PC processes and outcomes. Model results and adjusted utilization outcomes are summarized in Table 2. For our primary analysis, aggregate receipt of any PC process was not associated with differences in one-year health care utilization (all ps > 0.05). GOC conversations, code status limitations, and specialty PC consults were each independently associated with increased hospice enrollment and fewer days at home (all ps < 0.05). Two sensitivity analyses were conducted: (1) repeating the model to test each PC process individually in a subset of patients who received at least one PC process (N = 863), and (2) testing differential effects of early (within 72 hours) versus late GOC conversations. Neither of these exploratory analyses resulted in clinically or statistically significant changes from our primary analysis.

Table 2.

Adjusted One-Year Health Care Utilization After Severe Trauma in Older Adults by Receipt of Palliative Care Processes

Days at home	Mean^a (days)	Incidence rate ratio (95% CI)	p Value
Any palliative care process vs. none	305.8 vs. 306.9	1.00 (0.97, 1.03)	0.81
GOC conversation vs. none	290.2 vs. 308.9	0.94 (0.90, 0.98)	0.002
Code status limitation vs. none	287.7 vs. 311.7	0.92 (0.89, 0.95)	<0.001
Health care proxy vs. none	305.9 vs. 306.6	1.00 (0.97, 1.03)	0.87
Specialty palliative care vs. none	268.1 vs. 306.8	0.87 (0.78, 0.98)	0.02
Hospice assessment vs. none	281.9 vs. 306.7	0.92 (0.84, 1.00)	0.06

Days in hospital	Mean^a (days)	Incidence rate ratio (95% CI)	p Value
Any palliative care process vs. none	16.1 vs. 18.6	0.87 (0.69, 1.09)	0.23
GOC conversation vs. none	17.1 vs. 16.6	1.03 (0.82, 1.29)	0.80
Code status limitation vs. none	16.4 vs. 16.7	0.98 (0.78, 1.23)	0.87
Health care proxy vs. none	16.0 vs. 18.6	0.86 (0.69, 1.07)	0.18
Specialty palliative care vs. none	15.8 vs. 16.7	0.95 (0.57, 1.59)	0.84
Hospice assessment vs. none	13.1 vs. 16.7	0.78 (0.46, 1.34)	0.38

ED visits	Mean^a (N)	Incidence rate ratio (95% CI)	p Value
Any palliative care process vs. none	2.2 vs. 2.1	1.05 (0.85, 1.28)	0.67
GOC conversation vs. none	2.3 vs. 2.2	1.05 (0.84, 1.30)	0.67
Code status limitation vs. none	2.6 vs. 2.1	1.22 (1.01, 1.47)	0.04
Health care proxy vs. none	2.3 vs. 2.2	1.01 (0.84, 1.23)	0.88
Specialty palliative care vs. none	2.0 vs. 2.2	0.90 (0.52, 1.56)	0.71
Hospice assessment vs. none	2.1 vs. 2.2	0.93 (0.58, 1.49)	0.77

ICU days	Mean^a (days)	Incidence rate ratio (95% CI)	p Value
Any palliative care process vs. none	0.60 vs. 0.91	0.66 (0.31, 1.39)	0.27
GOC conversation vs. none	0.55 vs. 0.69	0.79 (0.36, 1.72)	0.55
Code status limitation vs. none	0.49 vs. 0.73	0.68 (0.30, 1.50)	0.34
Health care proxy vs. none	0.58 vs. 0.34	0.61 (0.31, 1.22)	0.16
Specialty palliative care vs. none	0.44 vs. 0.67	0.65 (0.10, 4.08)	0.64
Hospice assessment vs. none	0.28 vs. 0.67	0.41 (0.06, 3.03)	0.38

Hospice enrollment	Probability^a	Incidence rate ratio (95% CI)	p Value
Any palliative care process vs. none	0.17 vs. 0.07	2.53 (0.42, 15.19)	0.31
GOC conversation vs. none	0.39 vs. 0.10	4.70 (2.16, 10.25)	<0.001
Code status limitation vs. none	0.37 vs. 0.08	5.43 (2.11, 13.98)	<0.001
Health care proxy vs. none	0.16 vs. 0.11	1.54 (0.38, 6.25)	0.54
Specialty palliative care vs. none	0.63 vs. 0.13	6.95 (1.75, 27.59)	0.006
Hospice assessment vs. none	0.79 vs. 0.12	12.58 (6.07, 26.07)	<0.001

Multivariable Poisson regression controlled for age, sex, marital status, domicile, preexisting serious illness, Injury Severity Score, hospital length of stay, and index ICU use, with time-alive as an offset. Only patients surviving index trauma admission were included in the analysis. Model tested exposure of aggregate measure of palliative care processes (any vs. none) as well as individual processes.

Per 365 patient-days of follow-up.

CI, confidence intervals; ED, emergency department; GOC, goals of care.

Discussion

This retrospective cohort study of seriously ill older adults admitted to a large academic health system for severe trauma leveraged a granular NLP methodology using EHR data linked to Medicare claims to demonstrate modest rates of PC process documentation during inpatient admissions, with notable variation in documentation rates between processes. Contrary to our hypothesis, receipt of any PC process was not associated with reductions in postdischarge health care utilization. However, most PC processes were independently associated with hospice enrollment within one year postdischarge, suggesting that despite low rates of documentation, clinicians who deliver PC are doing so appropriately.

This study highlights the feasibility of combining NLP analysis of institutional EHR data with Medicare claims, offering broad applications in health services research to capture “real-world” PC documentation and its association with postdischarge outcomes. NLP is superior to the traditional measurement of PC services via claims data, which underestimates PC delivery outside of specialized services because of underbilling, under coding, or both.³⁷ Although the rates of PC documentation in this study were lower than would be expected in accordance with guidelines, they were higher than those documented in prior studies using administrative data.^19,20 For example, Fakhry et al. demonstrated very low (<10%) rates of inpatient PC utilization for older adult trauma patients.¹⁹ In contrast, and similar to our study that showed a 31% rate of GOC documentation, Pierce et al. used chart review to demonstrate a 26% rate of inpatient GOC documentation in a retrospective cohort of seriously ill older adult trauma patients.²² Together, these studies support ongoing use of more granular methods such as NLP to capture PC processes and emphasize the feasibility of linkage with Medicare claims, augmenting the robustness of health care utilization measurements.

The rates of PC documentation in our cohort suggest underutilization for hospitalized seriously ill older adults and highlight potential targets for improved delivery. ACS TQIP guidelines and Geriatric Trauma Best Practice Guidelines³⁸ both recommend early GOC conversations in all older adult trauma patients with serious illness, especially those with severe or critical injuries. As our study population comprised older adults with a severe trauma diagnosis, the entirety of our cohort would screen positive for “Category II” PC screening criteria as delineated by ACS TQIP,⁷ defined as patients with either severe injury burden or chronic serious illness, conferring high risk of in-hospital mortality or discharge to dependent care. For Category II patients, TQIP recommends not only early GOC conversations but also more intensive PC processes such as code status reevaluation, hospice discussion, and consideration of specialist PC consultation. Our data show that PC processes are largely underutilized in this group of patients after injury, despite a resource-rich setting with readily available specialist PC. High (39%) overall one-year mortality underscores the appropriateness of PC for these patients. This gap suggests that expanding implementation of screening tools for older trauma patients based on TQIP criteria may improve not only awareness of eligibility for PC but also trigger a range of specific PC processes that can be routinely initiated.

Consistent with previous studies,^14,21,39 PC processes in our cohort were more commonly delivered to older patients with more severe injuries, longer hospitalizations, increased serious illness burden, higher ICU utilization, and higher mortality. We also found independent associations between four PC processes and subsequent hospice enrollment within one year postdischarge. Together, these findings highlight patient-level and hospital-level factors that may influence PC delivery. Unsurprisingly, clinicians are more likely to document PC processes in sicker patients, intuiting which subgroups of patients are nearing end of life. Furthermore, the ICU setting may represent a hospital-level factor that facilitates documentation of PC processes. Our group previously conducted a similar study in seriously ill older adults undergoing major elective surgery at MGB, finding even lower rates of PC process documentation, with <3% GOC conversation, <1% PC consultation, and <60% health care proxy designation (article under review). In contrast to elective surgical patients, the relatively high utilization of the ICU in our cohort (>70%) suggests that this element of care may play a role in delivery. Therefore, improving early PC screening and implementation efforts in seriously ill older trauma patients who do not require ICU admission may be a distinct target to expand appropriate PC utilization. Prior single-institution efforts have demonstrated success of internal quality improvement programs in reaching benchmarks for advance care planning process documentation in older adult trauma patients,²⁴ highlighting that interventions at the level of the trauma system may be an avenue to bridge this gap.

Associations between PC and health care utilization in trauma patients remain equivocal. Our multivariable analysis did not demonstrate an association between PC processes and reduced one-year health care utilization. Early studies of these patients have demonstrated mixed outcomes. A 2023 study by Spencer et al. used propensity-score matching to demonstrate that PC consults in trauma patients were associated with decreased health care costs; however, they were also associated with increased ICU utilization and longer length of stays.⁴⁰ Studies in both trauma¹⁶ and nontrauma patients¹² have found that the timing of PC intervention was an important factor, as early GOC conversations were associated with reduced health care utilization. To address this possibility, we did explore differential effects of early versus late GOC conversations in a sensitivity analysis, which did not lead to any significant reductions in utilization outcomes. Taken together, our results suggest that health care utilization may be a challenging outcome to study retrospectively and is biased by increased PC delivery in the most complex patients.

There are several important limitations to this study. First, the five processes evaluated here, while endorsed as quality indicators of PC delivery,^7,34 are not inclusive of all elements of PC (i.e., symptom management, caregiver assessments, psychological/spiritual distress). However, the novelty of an NLP approach linked to Medicare claims offers improved granularity in comparison with traditional claims-based analyses, as discussed. Second, this retrospective study is unable to measure any causal link between PC delivery and health care outcomes or account for all possible confounders. Third, this study in a large, academic health system with two level I and one level III trauma centers and well-established PC divisions is not generalizable to other settings. Generalizability is further limited by the racial and ethnic homogeneity of this cohort. However, given the underutilization seen in this setting, our results underscore the need for improved strategies to better deploy PC for older adults admitted to hospital after injury.

Authors’ Contributions

D.I.H.: Methodology, data curation, formal analysis, investigation, visualization, and writing (original draft). S.M.: Data curation, formal analysis, and writing (review/editing). M.R.: Methodology, data curation, formal analysis, software, resources, and writing (review/editing). M.T.-K.: Methodology, investigation, writing (review/editing), project administration, conceptualization, and funding acquisition. K.S.: Methodology, investigation, and writing (review/editing). T.F.G.: Methodology, investigation, and writing (review/editing). S.R.L.: Methodology, supervision, conceptualization, and writing (review/editing). C.S.R.: Conceptualization, investigation, and writing (review/editing). C.L.: Methodology, investigation, conceptualization, supervision, software, and writing (review/editing). Z.C.: Methodology, investigation, conceptualization, supervision, software, writing (review/editing), and funding acquisition.

Footnotes

Author Disclosure Statement

The authors declare no relevant conflicts of interest.

Funding Information

Z.C. and D.I.H. receive funding from NIA Grant #R01AG070252A. T.F.G. receives funding from the RWJF Harold Amos Medical Faculty Development Program and the Betty Irene Moore Fellows for Nurse Leaders and Innovators. Funding sponsors took no part in the design, methods, data collection, analysis, or preparation of the article.

Supplemental Material

References

Centers for Disease Control and Prevention. WISQARS Fatal and Nonfatal Injury Infographics. Available from: https://wisqars.cdc.gov/

Adams

, Holcomb

. Geriatric trauma. Curr Opin Crit Care, 2015; 21(6):520–526; doi: 10.1097/MCC.0000000000000246

Hashmi

, Ibrahim-Zada

, Rhee

, et al. Predictors of mortality in geriatric trauma patients: A systematic review and meta-analysis. J Trauma Acute Care Surg, 2014; 76(3):894–901; doi: 10.1097/TA.0b013e3182ab0763

Joseph

, Pandit

, Zangbar

, et al. Superiority of frailty over age in predicting outcomes among geriatric trauma patients: A prospective analysis. JAMA Surg, 2014; 149(8):766–772; doi: 10.1001/jamasurg.2014.296

Pollack

, Liao

, Powelson

, et al. Long-term health-related quality of life and independence among older survivors of serious injury. J Trauma Acute Care Surg, 2023; 94(4):624–631; doi: 10.1097/TA.0000000000003864

Palliative Care Definition. What is Palliative Care. Available from: https://www.capc.org/about/palliative-care/ [Last accessed: October 30, 2024].

American College of Surgeons Trauma Quality Improvement Program. ACS TQIP Palliative Care Best Practice Guidelines. 2017. Available from: https://www.facs.org/media/g3rfegcn/palliative_guidelines.pdf

Lee

, Walling

, Senglaub

, et al. Defining serious illness among adult surgical patients. J Pain Symptom Manage, 2019; 58(5):844–850.e2; doi: 10.1016/j.jpainsymman.2019.08.003

Miner

. Understanding the core principles of primary and specialty surgical palliative care. J Am Coll Surg, 2020; 231(1):185–186; doi: 10.1016/j.jamcollsurg.2020.03.032

10.

Temel

, Greer

, Muzikansky

, et al. Early palliative care for patients with metastatic non-small-cell lung cancer. N Engl J Med, 2010; 363(8):733–742; doi: 10.1056/NEJMoa1000678

11.

Triplett

, LeBrett

, Bryant

, et al. Effect of palliative care on aggressiveness of end-of-life care among patients with advanced cancer. J Oncol Pract, 2017; 13(9):e760–e769; doi: 10.1200/JOP.2017.020883

12.

Piscitello

, Arnold

, Schell

. Cost-related outcomes associated with documented goals of care conversations. J Pain Symptom Manage, 2025:S0885; doi: 10.1016/j.jpainsymman.2025.09.017

13.

Duncan

, Holkup

, Sang

, et al. Benefits of early utilization of palliative care consultation in trauma patients. Crit Care Explor, 2023; 5(9):e0963; doi: 10.1097/CCE.0000000000000963

14.

Wycech

, Fokin

, Katz

, et al. Reduction in potentially inappropriate interventions in trauma patients following a palliative care consultation. J Palliat Med, 2021; 24(5):705–711; doi: 10.1089/jpm.2020.0218

15.

, Zhou

, Quinlan

, et al. Early palliative care consultation offsets hospitalization duration and costs for elderly patients with traumatic brain injuries: Insights from a Level 1 trauma center. J Clin Neurosci, 2023; 108:1–5; doi: 10.1016/j.jocn.2022.12.013

16.

Spencer

, Miller

PRI

, Russell

, et al. Timing is everything: Early versus late palliative care consults in trauma. J Trauma Acute Care Surg, 2023; 94(5):652–658; doi: 10.1097/TA.0000000000003881

17.

Rotundo

, Braunreuther

, Dale

, et al. Retrospective review of trauma ICU patients with and without palliative care intervention. J Am Coll Surg, 2022; 235(2):278–284; doi: 10.1097/XCS.0000000000000220

18.

Fokin

, Wycech

, Katz

, et al. Palliative care consultations in trauma patients and role of do-not-resuscitate orders: Propensity-matched study. Am J Hosp Palliat Care, 2020; 37(12):1068–1075; doi: 10.1177/1049909120919672

19.

Fakhry

, Carrick

, Hoffman

, et al. Hospice and palliative care utilization in 16 004 232 medicare claims: Comparing trauma to surgical and medical inpatients. Trauma Surg Acute Care Open, 2024; 9(1):e001329; doi: 10.1136/tsaco-2023-001329

20.

Lilley

, Lee

, Scott

, et al. The impact of inpatient palliative care on end-of-life care among older trauma patients who die after hospital discharge. In: Journal of Trauma and Acute Care Surgery. Vol 85. NIH Public Access; 2018:992–998. doi:10.1097/TA.0000000000002000

21.

Rivet

, Ferrada

, Albrecht

, et al. Characteristics of palliative care consultation at an academic level one trauma center. Am J Surg, 2017; 214(4):657–660; doi: 10.1016/j.amjsurg.2017.06.010

22.

Pierce

, Ricon

, Rukmangadhan

, et al. Adherence to the TQIP palliative care guidelines among patients with serious illness at a level I trauma center in the US. JAMA Surg, 2022; 157(12):1125–1132; doi: 10.1001/jamasurg.2022.4718

23.

Vogel

, McGraw

, Redmond

, et al. The ACS-TQIP palliative care guidelines at two level I trauma centres: A prospective study of patient and caregiver satisfaction. BMJ Support Palliat Care, 2022; 12(e1):e120–e128; doi: 10.1136/bmjspcare-2020-002229

24.

Moran

, Soltis

, Politis

, et al. Improving a mature palliative care program at a level I trauma center. J Trauma Nurs, 2021; 28(2):119–125; doi: 10.1097/JTN.0000000000000569

25.

Schockett

, Prather

, Benjenk

, et al. Integrating palliative care on an adult trauma service. J Palliat Med, 2021; 24(5):668–672; doi: 10.1089/jpm.2020.0378

26.

Amen

, Berndtson

, Cain

, et al. Communication and palliation in trauma critical care: Impact of trainee education and mentorship. J Surg Res, 2021; 266:236–244; doi: 10.1016/j.jss.2021.03.005

27.

Aziz

, Lunde

, Barraco

, et al. Evidence-based review of trauma center care and routine palliative care processes for geriatric trauma patients; A collaboration from the American Association for the Surgery of Trauma Patient Assessment Committee, the American Association for the Surgery of Trauma Geriatric Trauma Committee, and the Eastern Association for the Surgery of Trauma Guidelines Committee. J Trauma Acute Care Surg, 2019; 86(4):737–743; doi: 10.1097/TA.0000000000002155

28.

Kim

, Lissauer

, Martin

, et al.; and the Critical Care Committee of the AAST. Defining the surgical critical care research agenda: Results of a gaps analysis from the Critical Care Committee of the American Association for the Surgery of Trauma. J Trauma Acute Care Surg, 2020; 88(2):320–329; doi: 10.1097/TA.0000000000002532

29.

von Elm

, Altman

, Egger

, et al.; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: Guidelines for reporting observational studies. Ann Intern Med, 2007; 147(8):573–577; doi: 10.7326/0003-4819-147-8-200710160-00010

30.

Lindvall Lab, Dana Farber Cancer Institute. Clinical Regex. Available from: https://lindvalllab.github.io/clinical-regex/

31.

Volandes

, Chang

, Lakin

, et al. An intervention to increase advance care planning among older adults with advanced cancer: A randomized clinical trial. JAMA Netw Open, 2025; 8(5):e259150; doi: 10.1001/jamanetworkopen.2025.9150

32.

Lindsay

, de Oliveira

, Sciacca

, et al. Harnessing natural language processing to assess quality of end-of-life care for children with cancer. JCO Clin Cancer Inform, 2024; 8:e2400134; doi: 10.1200/CCI.24.00134

33.

Volandes

, Zupanc

, Paasche-Orlow

, et al. Association of an advance care planning video and communication intervention with documentation of advance care planning among older adults: A nonrandomized controlled trial. JAMA Netw Open, 2022; 5(2):e220354; doi: 10.1001/jamanetworkopen.2022.0354

34.

Ferrell

, Twaddle

, Melnick

, et al. National consensus project clinical practice guidelines for quality palliative care guidelines, 4th Edition. J Palliat Med, 2018; 21(12):1684–1689; doi: 10.1089/jpm.2018.0431

35.

Lee

, Udelsman

, Streid

, et al. Natural language processing accurately measures adherence to best practice guidelines for palliative care in trauma. J Pain Symptom Manage, 2020; 59(2):225–232.e2; doi: 10.1016/j.jpainsymman.2019.09.017

36.

Kelly

, Sturgeon

, Harlow

, et al. Using Medicare data to identify serious illness in older surgical patients. J Pain Symptom Manage, 2020; 60(2):e101–e103; doi: 10.1016/j.jpainsymman.2020.04.002

37.

Lilley

, Lindvall

, Lillemoe

, et al. Measuring processes of care in palliative surgery. Ann Surg, 2018; 267(5):823–825; doi: 10.1097/SLA.0000000000002579

38.

American College of Surgeons Trauma Quality Programs. Best Practice Guidelines: Geriatric Trauma Managment. 2023. Available from: https://www.facs.org/media/ubyj2ubl/best-practices-guidelines-geriatric-trauma.pdf

39.

Fiorentino

, Hwang

, Pentakota

, et al. Palliative Care in trauma: Not just for the dying. J Trauma Acute Care Surg, 2019; 87(5):1156–1163; doi: 10.1097/TA.0000000000002440

40.

Spencer

, Nunn

, Miller

, et al. The value of compassion: Healthcare savings of palliative care consults in trauma. Injury, 2023; 54(1):249–255; doi: 10.1016/j.injury.2022.10.021

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.18 MB

0.00 MB