Key Points

Question听 What do long-term postoperative recovery trajectories look like?

Findings听 This cohort study (378鈥�682 cases) included 5 postoperative recovery trajectories. Trajectories included routine, slow, and protracted recoveries, loss of independence, and chronic dependence; these included trajectories were heavily associated with days elsewhere than home and various preoperative and operative risk factors.

Meaning听 In this study, trajectory models demonstrated clinically meaningful differences in postoperative recovery that should inform surgical decision-making.

Importance听 Major surgery sometimes involves long recovery or even permanent institutionalization. Little is known about long-term trajectories of postoperative recovery, as surgical registries are limited to 30-day outcomes and care can occur across various institutions.

Objective听 To characterize long-term postoperative recovery trajectories.

Design, Setting, and Participants听 This retrospective cohort study used Veterans Affairs (VA) Surgical Quality Improvement Program data (2016 through 2019) linked to the Residential History File, combining data from the VA, Medicare/Medicaid, and other sources to capture most health care utilization by days. Patients were divided into younger (younger than 65 years) or older (65 years or older) subgroups, as Medicare eligibility is age dependent. Latent-class, group-based trajectory models were developed for each group. These data were analyzed from February 2023 through August 2024.

Exposure听 Surgical care in VA hospitals.

Main Outcomes and Measures听 Days elsewhere than home (DEH) were counted in 30-day periods for 275 days presurgery and 365 days postsurgery.

Results听 A 5-trajectory solution was optimal and visually similar for both age groups (cases: 179鈥�879 younger [mean age (SD) 51.2 (10.8) years; most were male [154鈥�542 (83.0%)] and 198鈥�803 older [mean (SD) age, 72.2 (6.0) years; 187鈥�996 were male (97.6%)]). Most cases were in trajectories 1 and 2 (T1 and T2). T1 cases returned home within 30 days (younger, 74.0%; older, 54.2%), while T2 described delayed recovery within 30 to 60 days (younger, 21.6%; older, 35.5%). Trajectory 3 (T3) and trajectory 4 (T4) were similar for the first 30 days postsurgery, but subsequently separated with T3 representing protracted recovery of 6 months or longer (younger, 2.7%; older, 3.8%) and T4 indicating long-term loss of independence (younger, 1.3%; older, 5.2%). Few (trajectory 5) were chronically dependent, with 20 to 30 DEH per month before and after surgery (younger, 0.4%; older, 1.3%).

Conclusions and Relevance听 In this study, trajectory models demonstrated clinically meaningful differences in postoperative recovery that should inform surgical decision-making. Registries should include longer-term outcomes to enable future research to distinguish patients prone to long-term loss of independence vs protracted, but meaningful recovery.

Surgical outcome studies often focus on mortality, complications, or hospital readmissions. These outcomes, while important, may not adequately capture what matters most to patients.^1,2 For example, while long-term functioning is likely extremely important to a patient鈥檚 decision to have surgery, most outcomes are limited to 30 days after surgery^3,4 or discharge.⁵ Furthermore, patients may require physical rehabilitation or skilled nursing before eventually being discharged home. Distinguishing between patients ultimately returning home (ie, a long recovery but ultimately a successful surgery) and those permanently institutionalized is critically important to patients. Days at home⁶ is an alternative, patient-centered measure of quality with 2 major strengths: (1) postsurgical care can be spent in various institutions; days at home captures time in various kinds of care and (2) data to calculate days at home has longer follow-up periods. We propose the inverse of days at home, days elsewhere than home (DEH), as the patient鈥檚 focus is the length of postoperative recovery.

Studies using days at home have been limited to administrative claims or utilization data.⁶ While these data are useful to track postsurgical care, there are known limitations: variation between facilities in comorbidity capture using International Classification of Diseases, Ninth Revision/Tenth Revision codes^7,8 and difficulty distinguishing between preoperative risk factors and postoperative complications.⁹ While administrative claims/utilization data can adequately capture postsurgical independence, these data limit the ability to assess preoperative and operative factors associated with recovery or permanent institutionalization.

To overcome these limitations, we merged data from the Residential History File (RHF), a dataset combining Veterans Affairs (VA), US Centers for Medicare and Medicaid Services, Minimum Data Set, and other data to track nearly all health care utilization on a daily basis¹⁰ with Veterans Affairs Surgical Quality Improvement Program (VASQIP) data, an audited, nurse-abstracted surgical registry.¹¹ Using these data, we summarized total DEH in the year after surgery and isolated and described trajectories of postoperative recovery. We hypothesized that frailty,¹² preoperative acute serious conditions (PASC),^13-15 higher operative stress,¹³ and urgent or emergent case status would be associated with longer recovery trajectories. We also hypothesized that longer recovery trajectories would be more common among older veterans than younger veterans.

This retrospective cohort study used 2016 to 2019 VASQIP data, followed Strengthening the Reporting of Observational Studies in Epidemiology () reporting guidelines,¹⁶ and was determined exempt by the VA鈥檚 Central Institutional Review Board. While VASQIP has recently been criticized for its systematic sampling procedure,¹⁷ it contains reliable, nurse-adjudicated data necessary to accurately identify surgical risk factors.¹⁸ The VASQIP cohort was linked to the RHF, a combined dataset using advanced algorithms to arrange episodes of care in chronological sequence, specifying each episode鈥檚 duration and setting¹⁰ to track preoperative and postoperative care. All care reimbursed by the VA or the US Centers for Medicare and Medicaid Services is captured by the RHF, including data pertaining to nursing homes found in the Minimum Data Set.

We selected a single VASQIP case at random for patients with multiple cases over the study period, except if the cases were on the same day, then the case with the highest operative stress was selected. The VA Corporate Data Warehouse was then used to capture race, ethnicity, and surgical case status (elective, urgent, or emergent). Mortality was derived from the Vital Status File, which contains data from multiple VA and non-VA data sources, such as the Social Security Administration Death Master File and the Integrated Benefits System Death File.

Using RHF data, each day within 275 days before and 365 days after surgery was classified as either a DEH or a day at home. Periods of 30 days were used to represent an asynchronous duration of a month, with each month summarizing the sum of DEH for that month. Because data began 275 days before surgery, the date of surgery fell on the 6th day of the 10th month, which was landmarked to month 0. Patients were then classified into several trajectories of postsurgical dependence, using methods described below. We also summarized the total count of DEH during the postoperative year and examined bivariate associations between total DEH and various patient-level and surgical factors.

We replaced up to 30 days after death with the DEH location from the day preceding death, allowing for complete 30-day periods. Otherwise, valid DEH data would need to be excluded for any 30-day period in which death occurred. We used the DEH location from the day before death because sensitivity analyses (eMethods, eFigures 1 through 4 in Supplement 1) suggested that some patients are discharged home shortly before death, possibly with home health or home hospice services, appearing more independent immediately before death than might be warranted. This replacement also allows future studies to analyze DEH and mortality independently. After completing the nearest 30-day period, DEH for days after death were coded as missing.

Frailty was assessed using the Risk Analysis Index (RAI),¹² which uses VASQIP variables to render a score from 0 to 81, categorized as robust (20 or less), normal (21 to 29), frail (30 to 39), and very frail (40 or more).¹² The RAI has been validated across multiple datasets^4,12 and provides a single, patient-level estimate that overcomes model fit issues with less parsimonious models.^13,19-21

Patients presenting with PASC¹³ were assessed using 6 preoperative VASQIP鈥揷aptured conditions (ventilator use, pneumonia, coma, sepsis, large transfusions, or renal failure) within 30 days of surgery (eTable 1 in Supplement 1).

Operative stress was assessed using the expanded Operative Stress Score (OSS),¹³ which estimates surgical-induced physiologic stress by assigning scores of 1 (very low stress) to 5 (very high stress) to Current Procedural Terminology (CPT) codes. OSS was coded using the highest score for all available CPT codes within each case. If the principal CPT code did not have an assigned OSS, we reported the OSS as missing. Case status quantified the acuity of the surgical procedure, categorized as elective, urgent, or emergent (eMethods in Supplement 1).

We approximated Clavien-Dindo²² IV (CDIV) complications using 9 VASQIP variables (reintubation, pulmonary embolism, ventilator for longer than 48 hours, acute renal failure, stroke, cardiac arrest, myocardial infarction, septic shock, and coma for longer than 48 hours) (eTable 1 in the Supplement 1) and methods previously described with private sector data.¹³

We used group-based trajectory modeling²³ to explore clinically meaningful trajectories of independence, examining data in separate younger (younger than 65 years) and older (65 years or older) subgroups because veterans younger than 65 years old may lack Medicare data. We use these terms throughout, consistent with 糖心vlog guidelines.²⁴ Group-based trajectory modeling explores variation in longitudinal outcomes, by separating cases into subgroups.²³ This method is data driven, with model-fit statistics to determine the optimum number of groups²³ with 500 or more cases required for reliable estimation.²⁵ DEH across 30-day panels were analyzed using a continuous group-based trajectory model (using a normal distribution for model fitting) as quadratic trajectories over time. Linear trajectories had poor model fit and cubic trajectories were visually similar to quadratic trajectories. We used the PROC TRAJ macro²⁶ in SAS version 9.4 (SAS Institute) to evaluate all solutions with 2 to 7 groups and selected the optimal number of groups for both the younger and older subgroups using clinical judgment and model-fit statistics (Akaike information criterion and Bayesian information criterion; eTable 2 in Supplement 1). Based on the probabilities of group membership from PROC TRAJ, almost all cases were cleanly sorted into trajectories. Because group-based trajectory models assume homogeneity within the identified subgroups, we performed a sensitivity analysis using growth curves with k-means clustering²⁷ on a subsample of the data (eFigure 5 in Supplement 1).

We described each patient trajectory using the preoperative, operative, and postoperative variables listed above, each defined before data collection, without excluding or imputing data for any non-DEH variable. We also described the most common procedure families for each trajectory, grouping procedures by principal procedure CPT codes using the hierarchical structure of CPT codes (eTable 3 in Supplement 1).²⁸ Analyses outside group-based trajectory modeling were performed using R version 4.3.1 (The R Project). Categorical data were summarized using counts and percentages and continuous data using mean and SD. 蠂² Tests and F tests were used to test for differences between groups for categorical and continuous variables, respectively.

Since total DEH was positively skewed and drastically zero inflated, we used zero-inflated negative binomial regression to examine bivariate associations between total DEH and various preoperative, operative, and postoperative variables, as other work has found this to be the optimal statistical method for variables like total DEH.²⁹ In these bivariate models, missing variables were excluded pairwise.

VASQIP included 463鈥�160 cases, which were limited to 378鈥�702 patient-unique cases. A total of 192鈥�527 were 65 years or older and 186鈥�175 were 64 years or younger (Table 1; eTable 4 in Supplement 1). RHF data were successfully linked to 99.7% of the sample. The older subgroup had a mean age of 72.2 (SD, 6.0) years; most were male (187鈥�996 [97.6%]; female, 4531 [2.4%]). There were 562 patients who identified as Asian (0.3%), 24鈥�209 patients identified as Black (12.6%), 9950 patients identified as Hispanic (5.3%), 1085 patients identified as Native American (0.6%), 1013 patients identified as Pacific Islander (0.5%), 147 254 patients identified as White (76.5%), 8354 patients had missing data for race and ethnicity and most had normal RAI/frailty (134鈥�804 [70.0%]). Cases were most commonly elective (165鈥�540 [86.0%]), with few patients presenting with PASC (4531 [2.4%]). The younger subgroup had a mean (SD) age of 51.2 (10.8) years; most were male (154鈥�542 [83.0%]). Most identified as White (117鈥�950 [63.4%]) and had robust RAI/frailty (126鈥�588 [68.0%]). Cases were most commonly elective (164鈥�570 [88.4%]), with few patients presenting with PASC (109 [1.5%]).

For both the older and younger subgroups (Figures 1 and 2), a 5-group solution was optimal because each trajectory described a distinct and plausible clinical outcome and the model-fit statistics had the lowest Akaike information criterion and Bayesian information criterion (eTable 2 in Supplement 1). The trajectories were visually similar between the older and younger subgroups. Most cases were either routine recovery (T1) or slow recovery (T2). T1 cases returned home within 30 days (older, 54.2%; younger, 74.0%), while T2 cases recovered within 30 to 60 days (older, 35.5%; younger, 21.6%). T3 and T4 were similar for the first 30 days postsurgery, but subsequently separated; T3 indicating protracted recovery with patients not returning home until at least 6 months postsurgery (older, 3.8%; younger, 2.7%) and T4 representing long-term loss of independence (older, 5.2%; younger, 1.3%). Chronically dependent (T5) cases had 20 to 30 DEH per month before and after surgery (older, 1.3%; younger, 0.4%).

The older patients with longer recoveries or who did not recover at all had higher mean RAI scores (T3: 33.1, T4: 34.6, T5: 37.4) than those with quicker recoveries (T1: 26.2, T2: 28.2, Table 1). Trajectories of longer recovery also had higher proportions of patients presenting with PASC (T3: 753 [10.2%], T4: 1285 [12.9%], T5: 271 [10.8%]) or urgent/emergent cases status (T3: 43.4%, T4: 44.5%, T5: 41.1%). Lastly, total DEH for the routine recovery trajectory (T1) was typically 2 to 3 days, while 2 weeks for slow recovery (T2), 2 to 3 months for the protracted recovery (T3), and almost the entire postoperative year for the loss of independence (T4) and chronically dependent (T5) groups. Similar patterns were observed in the younger subgroup (eTable 4 in Supplement 1). Sensitivity analyses using growth curves with k-means clustering showed routine recovery (T1), protracted recovery (T3), and loss of independence (T4) groups that were similar to those in Figures 1 and 2, confirming the robustness of our primary findings (eFigure 5 in Supplement 1).

For both the older and younger subgroups, cases with quicker recovery (T1 to T2) had high frequencies of hernia repairs, colectomies, and joint repairs in the knee, hip, or pelvis (Table 2). While knee, hip, and pelvis repairs were also frequent in the trajectories of longer recovery (T3 to T5), these groups also had a high rate of lower extremity amputations and lower extremity revascularizations.

Higher RAI scores were associated with more DEH in both the older (relative risk [RR], 1.14; 95% CI, 1.14-1.15; P鈥�&濒迟;鈥�.001, Table 3) and younger (RR, 1.16; 95% CI, 1.16-1.16; P鈥�<鈥�.001) subgroups. Lower operative stress (OSS 1 to 2) was associated with fewer DEH (RRs range, 0.35-0.42), while high stress (OSS 4) and very high stress (OSS 5) procedures were associated with more DEH (RRs range, 1.12-3.01) vs moderate stress (OSS 3). Cases presenting with PASC had higher DEH (RRs range, 5.33-9.18), as did urgent (RRs range, 4.16-4.52) and emergent cases (RRs range, 4.70-4.90) vs elective. Cases experiencing CDIV complications had higher DEH (RRs range, 6.45-12.04) than those who did not.

This study is among the first to describe trajectories of long-term recovery in the year after surgery, operationalized by the continuous outcome measure of DEH. We described patterns of 12-month postoperative recovery, using both a total count of DEH and empirically derived DEH trajectories. Group-based latent trajectory modeling revealed 5 meaningful trajectories representing routine, slow, and protracted recovery, loss of independence, and chronic dependence. We also found that frailty, PASC, higher operative stress, and urgent or emergent case status were associated worse DEH trajectories and more DEH. Of note, the loss of independence and protracted recovery trajectories intersected in-between the month of surgery and the first postoperative month, suggesting that the 30-day follow-up typical for surgical outcomes does not adequately distinguish between these 2 distinct populations.

Although the shape of these trajectories was remarkably similar to those observed in prior studies of health care utilization,^30,31 our focus on DEH improves on these studies in several ways. The first study uses hospital cost data,³⁰ which is not a particularly patient-centered outcome. In addition, the values are scaled to the average cost of the low trajectory.³⁰ While this was done to obscure the proprietary financial data of a particular health care system, it leaves the outcome dimensionless and difficult to interpret directly. The second study examines months with 鈥渁t least one inpatient admission.鈥�³¹ While this outcome better represents patient long-term functioning, it treats single-night admissions the same as an entire month in the hospital. By tracking health care utilization on a daily basis, we modeled DEH with granularity that allows for rough estimates of the time from surgery to baseline functioning. In this way, our quasi-continuous outcome contains more information than a binary one. Additionally, the second study鈥檚 focus on acute care admissions does not account for utilization in outpatient and postacute settings鈥攃are that can be substantial, burdensome, and incredibly relevant to surgical decision-making. Our approach focuses more on what matters to patients鈥攍ong-term functioning^1,2鈥攁nd it provides intuitive and granular detail of the experience within each trajectory.

Group-based trajectory models can overestimate the number of groups for various reasons: (1) models are estimated assuming zero variation within groups (ie, groups with enough heterogeneity will sometimes be estimated as different groups),³² (2) Akaike information criterion and Bayesian information criterion heavily rely on this assumption such that better statistics in models with more subgroups may represent a statistical artifact rather than better model fit, and (3) modeling assumes that outcomes within each group are normally distributed: heavily skewed data can sometimes be estimated as a mixture of several normal distributions. Alternate modeling approaches (like latent growth mixture modeling and growth curves with k-means clustering)²⁷ do not share these limitations but have computational demands that are impractical for larger datasets. While sensitivity analyses provide support for our primary findings, we advise caution regarding the protracted recovery and chronically dependent groups, the smallest groups within the data.

Loss of independence is critically important to patients and should inform shared decision-making.³³ Prior work using 30-day outcomes defined short-term loss of independence to describe those patients who presented from home for surgery but were discharged elsewhere after surgery.³⁴ Such short-term loss of independence was associated with both advancing age³⁴ and increasing frailty³⁵ with 30% to 42% of patients with frailty discharged to skilled nursing or assisted-living facilities after major surgery.^36,37 However, enduring challenges of capturing postacute health care utilization limited attempts to determine whether these patients regained their independence or remained institutionalized.

Other studies have examined total days away from home or days alive and out of the hospital as postoperative outcomes,^29,38 but their sample limits generalizability. For example, one study excluded patients based on life expectancy, biasing their cohort to healthier, lower-risk patients.²⁹ In addition, the focus on overall days away from home does not account for distinct patterns of recovery in our analysis that correspond to clinical experience. Longer periods of follow-up, that encompass the full range of postacute care, are essential for patients and surgeons as they consider whether surgery will restore patients to better flourishing rather than whether a surgery will be mortality, complication, and readmission free.

We did not attempt to develop predictive models for group membership in our trajectories, but we did demonstrate that each trajectory is heavily associated with known predictors of postoperative outcomes, suggesting that future work might develop effective tools to predict trajectory membership with preoperative data. Although most patients experience routine recovery, 1.3% to 5.2% of patients experience long-term and possibly permanent loss of independence. The possibility of this adverse outcome is critical information for shared decision-making in elective, urgent, and emergent contexts because many patients, especially those with limited life span, prefer less or no treatment when treatment is associated with high functional or cognitive impairment risks. For example, 99% of seriously ill patients 60 years or older stated they would agree to routine treatment that was likely to restore health.³⁹ However, 74% to 94% of these patients reported that they would forgo such treatments if there were a significant chance of functional or cognitive impairment, accepting a shortened life span to preserve quality of life.³⁹ Furthermore, the SUPPORT trial, an investigation of patient preferences and life-limiting disease, showed that critically ill patients often received more invasive treatment (eg, surgery) than they preferred⁴⁰ and that patients often receive treatment inconsistent with their values and preferences.⁴¹ Older adults undergoing elective surgery identified social events, recreational events, and activities of daily living as most important to preserve after surgery; however, as few as 65% achieved this recovery in 6 months.⁴²

While we consider these trajectories, particularly loss of independence, to be important information for surgical decision-making, we did not include data from patients choosing nonoperative treatments, which would be useful for comparison. Additionally, although the younger and older subgroups had visually similar trajectories, patients younger than 65 years might have private sector care not captured by the US Centers for Medicare and Medicaid Services, lowering their DEH and sorting them into shorter recovery trajectories. The low prevalence of women receiving VA care might make these findings less applicable to highly female populations. Trajectories derived in private sector samples might differ from those presented here due to well-characterized differences between veterans and the general US population, especially regarding comorbidities, mental illness, social support, and health care coverage. In particular, VA health care coverage might increase access to skilled nursing facilities or nursing homes compared with private sector patients. Furthermore, regional and individual differences in nursing care and family caretaker availability may create trajectory misclassifications and could be the focus of further research regarding these trajectories. In patients who died during the 1-year postoperative follow-up, we used the DEH location from the day before death to finish that month. While we consider this to be the best option given the results of our sensitivity analysis, we acknowledge the possibility for misclassification. Although our data are now 5 years old, they end with the onset of COVID-19 pandemic, which introduced substantial confounding to usual operative care until May 2023 (the end of the pandemic). This confounding suggests that replication of these findings in more recent data must be deferred several years to accrue appropriate data. Lastly, the cross-sectional design of our study limits our ability to draw strong causal inference from the longitudinal data.

We isolated 5 trajectories of 12-month postoperative recovery, finding that preoperative and operative risk factors (such as frailty, PASC, operative stress, and urgent/emergent case status) were heavily associated with these trajectories. Cases with protracted recoveries, loss of independence, or chronic dependence were most commonly procedures that impact mobility and physical or social functioning. Future work should focus on developing effective risk calculators to predict recovery trajectories with preoperative data to inform surgical decision-making.

Accepted for Publication: August 22, 2024.

Published Online: October 23, 2024. doi:10.1001/jamasurg.2024.4691

Corresponding Author: Daniel E. Hall, MD, MDiv, MHSc, UPMC Presbyterian, 200 Lothrop St, Ste F1264, Pittsburgh, PA 15213 (hallde@upmc.edu).

Author Contributions: Mr Jacobs and Dr Hall had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Concept and design: M. Jacobs, Intrator, Makineni, Youk, McCoy, Kinosian, Hall.

Acquisition, analysis, or interpretation of data: M. Jacobs, C. Jacobs, Intrator, Makineni, Youk, Boudreaux-Kelly, McCoy, Kinosian, Shireman, Hall.

Drafting of the manuscript: M. Jacobs, Boudreaux-Kelly, McCoy, Hall.

Critical review of the manuscript for important intellectual content: M. Jacobs, C. Jacobs, Intrator, Makineni, Youk, McCoy, Kinosian, Shireman, Hall.

Statistical analysis: M. Jacobs, Intrator, Makineni, Youk, McCoy, Kinosian.

Obtained funding: Intrator, Shireman, Hall.

Administrative, technical, or material support: C. Jacobs, Makineni, Boudreaux-Kelly.

Supervision: Youk, Hall.

Conflict of Interest Disclosures: None reported.

Funding/Support: This research was supported by a grant support from the VHA Office of Research and Development (HSR&D I01HX003322). The authors disclose other grant funding from the National Institutes of Health and Veterans Health Administration Office of Research and Development outside the scope of this work. Dr Hall discloses a consulting relationship with FutureAssure.

Role of the Funder/Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Disclaimer: The opinions expressed here are those of the authors and do not necessarily reflect the position of the US government.

Meeting Presentation: A subset of the study data was presented at the 2023 American College of Surgeons Clinical Congress; October 23, 2023; Boston, Massachusetts.

Data Sharing Statement: See Supplement 2.