Introduction

In North America, approximately 300 000 planned or emergent colectomies (large bowel resections) are performed yearly [1, 2], with over 50% performed in patients aged 65 and older [3]. Indications for elective colectomies include colorectal cancer (41%), diverticulitis (23%), inflammatory bowel disease (8%), and non-malignant colorectal polyps (12%) [4]. The median length of hospital stay is 6 days [5, 6], with a mean cost of stay of $25 900 United States dollars [1]. Postoperative mortality and morbidity rates within 30 days of operation are 1.4% and 12.4%, respectively [7]. Common postoperative surgical complications include ileus (12%), bleeding (8%), surgical site infection (7%), and anastomotic leakage (5%) [4], while major medical complications include pneumonia (2%), venous thromboembolism (1 to 2%), myocardial infarction (< 1%), acute renal failure (0.4%), and cerebrovascular accident (0.2%) [8]. Postoperative complications are associated with increased lengths of stay, readmissions, and costs, and prognosticate both short- and long-term mortality [9–12].

As the population ages and becomes more frail with increased risk of perioperative morbidity [13], prediction models that identify high-risk patients allow both preoperative shared decision-making and postoperative targeted enhanced monitoring of complications. Accurate risk prediction may facilitate timely treatment and reduce failure to rescue [14]. A systematic review found 26 models developed specifically for patients aged 65 years or older undergoing colorectal cancer surgery [15]. Other prediction models include the American College of Surgeons National Surgical Quality Improvement Program (NSQIP) surgical risk calculator derived from the multicentered North American dataset [16] and the Codman score derived from a single-centre South African dataset [17]. However, existing models are limited by moderate to high risks of bias [15], incomplete reporting of calibration [15], or poor performance during external validation [18, 19]. Furthermore, existing risk prediction models consider outcome variables as dichotomous (i.e., occurred vs did not occur within a specified period) without predicting when such complications may occur. It may be overwhelming in clinical practice for clinicians, patients/caregivers, and health systems to monitor for all possible complications throughout the postoperative period. Accurate prediction of the timing of when patients are at the highest risks for different complication allows for tailored monitoring strategies and efficient resource use [12, 20], particularly during the post-discharge period.

Our study objective was to develop and internally validate distinct time-to-event risk prediction models for each of the following outcomes within 30 days after elective colectomy: mortality, hospital readmission, myocardial infarction, cerebral vascular event, pneumonia, venous thromboembolism, acute renal failure, and sepsis or septic shock.

Materials and methods

Following approval from the Providence Health Care Research Ethics Board with waiver of informed consent (H21-02670; November 16, 2021), we conducted survival analysis modelling using a retrospective linked cohort from the prospectively collected multicentered NSQIP general and Procedure Targeted Colectomy datasets [21]. Before data analysis, the study protocol was registered at clinicaltrials.gov (NCT05150548; November 25, 2021). Data were accessed for analysis on January 25, 2022, and are reported according to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines [22].

Results

Cohort characteristics

After applying exclusion criteria, the final cohort contained 132 145 patients (Fig 1). Cohort characteristics are shown in Table 1; 47.8% were male (63 199/132 145), with a mean ± SD age of 61 ± 14 years. The median [quartiles] length of stay was 4 [3, 6] days. Incidence and timing of postoperative morbidity and mortality are reported in Fig 2 and S2.1 Table in S2 Appendix. The incidence of death within 30 days was 0.5% (714/132 145), with the median [quartiles] time to death at 10 [6, 18] days postoperatively. The incidences of readmission, bleeding requiring transfusion within 72 hours, and sepsis or septic shock were 8.6%, 5.6%, and 2.6%, respectively, while other model endpoints had incidences less than 2%. Cumulative hazard plots for each model endpoint are shown in S1 Fig in S3 Appendix. Heat map indicated highest co-occurrences between sepsis and readmission, venous thromboembolism and readmission, and sepsis and pneumonia (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study cohort flow diagram.

https://doi.org/10.1371/journal.pone.0314526.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Incidence and timeline of postoperative morbidity and mortality.

The incidence and median [quartiles] of postoperative day of events are presented for each postoperative complication. See S2.1 Table in S2 Appendix.

https://doi.org/10.1371/journal.pone.0314526.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Dendrogram of co-occurrence of model endpoints.

High co-occurrence is represented by dark red while low co-occurrence is represented by a light red and no co-occurrence by white.

https://doi.org/10.1371/journal.pone.0314526.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Cohort characteristics.

https://doi.org/10.1371/journal.pone.0314526.t001

Model results

The formula, baseline hazard function, and coefficients for each model are reported in S2.3-S2.11 Tables in S2 Appendix. The performance measures for the Cox proportional hazards models are listed in Table 2 with calibration plots shown in S3 Appendix.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Performance measures of Cox proportional-hazard models.

https://doi.org/10.1371/journal.pone.0314526.t002

For discrimination, the best-performing models are for mortality, myocardial infarction, renal failure, and cerebrovascular accident, with high time-dependent AUC at 30 days (0.835, 0.802, 0.800, and 0.796, respectively) and high Uno’s concordance indices (0.845, 0.813, 0.805, and 0.802 respectively). The worst performing models are for readmission, venous thromboembolism, and sepsis, with Uno’s concordance indices of 0.617, 0.655, and 0.679, respectively. Uno’s and Harrell’s concordance indices are similar for all models.

All models are well-calibrated (S2(a) to S2(h) Figs in S3 Appendix), though the predictive ranges are limited to low probability areas due to the low incidence of outcomes. All models had low time-dependent Brier scores, scaled time-dependent Brier scores, and pseudo R-squares.

During sensitivity analysis, the performance measures of the exploratory accelerated failure time models were similar to those of the Cox proportional hazards models (S2.12 Table in S2 Appendix). To provide an example of time-to-event prediction, we plotted the risks over time for different postoperative outcomes for a theoretical patient (S4 Appendix). Analysis using AIC-based stepwise elimination for the sepsis and readmission models did not change alter model performance metrics (S5 Appendix).

Discussion

The implementation of targeted, resource-efficient monitoring of postoperative complications requires accurate prediction of when a patient may be at highest risks of certain complications. In this NSQIP cohort of 132 145 patients, we developed and internally validated Cox proportional hazards models to predict time-to-event for major medical complications, readmission, and mortality within 30 days after elective colectomy. We found that while models for mortality, myocardial infarction, pneumonia, and renal failure performed well by concordance (> 0.8), the scaled Brier scores and pseudo R-squared values were poor. The models for readmission, venous thromboembolism, and sepsis performed poorly. All models displayed good calibration though with a limited range of prediction due to low incidence of complications.

Postoperative complications contribute to short- and long-term mortality, increased length of stay, readmission, and increased healthcare costs [4, 10–12].Since major complications have distinct temporal patterns [12, 20], accurate prediction of both if and when a complication may occur allows for targeted patient education and monitoring for specific complications at specific periods. We observed that the risk periods of each investigated complication differ, which has important implications for clinical management and targeted allocation of resources. Alarmingly, the median [quartiles] periods of diagnoses for most complications (ranging from 3 [1, 7] for myocardial infarction and 12 [6, 18] for venous thromboembolism) fall outside of the median [quartiles] for the length of stay of 4 [3, 6] days, highlighting the need for enhanced postoperative (and particularly post-discharge) monitoring, particularly after discharge.

Our time-to-event models add to the literature by facilitating efficient, targeted monitoring for complications in high risk patients during high risk periods. The NSQIP surgical risk calculator consists of 21 preoperative factors (demographics, comorbidities, and procedure) and was developed using 1 414 006 patients from 393 hospitals encompassing 1557 unique Current Procedural Terminology codes (i.e. procedures) [16]. However, its external validation in colectomy patients has been poor [18, 19]. The Codman score developed using a NSQIP dataset in South Africa is simpler with six predictors [17], with similar performances compared to the NSQIP morbidity and mortality algorithms when validated in 40 589 colectomy patients [19]. There are 26 models for elderly patients undergoing colorectal cancer surgery predicting outcomes, including mortality (n = 10), anastomotic leakage (n = 7), and surgical site infections (n = 3), with moderate to high bias in most studies [15]. Overall, the clinical utility of existing models is limited by the inability to predict the timing of complications for tailored monitoring strategies and prompt clinical intervention.

Among perioperative factors, postoperative complications have the greatest influence on both short-term and long-term mortality, compared to other preoperative and intraoperative determinants [10, 11]. Our study highlights the importance of incorporating temporality into postoperative risk prediction, providing a proof-of-concept of time-to-event risk prediction models that can be applied to other surgical populations to improve the detection and treatment of postoperative complications. Institutional perioperative data can be used to develop and continuously improve time-to-event postoperative risk prediction models [34]. This can guide informed consent with personalized risk calculators, disposition planning, and protocolized screening of postoperative complications. Importantly, since different complications tend to occur at different time ranges after surgery, spanning both during hospitalization and post-discharge, continuity of care and targeted monitoring are required to support patients. Hospital systems may leverage virtual care and remote monitoring for patients after discharge [35]. In a large randomized trial of patients undergoing noncardiac surgery, post-discharge remote monitoring did not increase the days alive and at home within 30 days of surgery [36]. However, there are signals of benefits for pain, reduced medication errors, and outcomes in centers with higher clinical response to the abnormalities detected [36]. Perioperative teams can use the graphs of risks for different complications over time (example in S4 Appendix) to educate patients about postoperative complications, identify patient-centered methods for self-screening, and empower patients to seek help across the risk periods.

There are several strengths in this study. The multicentered NSQIP dataset provides prospectively collected data that is likely generalizable to many other similar centers. There are clear definitions for each variable for the data collectors, with high inter-rater reliability in previous audits [37]. In terms of analysis, we performed time-to-event modelling, which provided additional temporal information compared to previous binary prediction models. Moreover, sensitivity analysis using accelerated failure time models was performed to ensure that the choice of modelling did not limit the model performances.

Our study has several limitations. First, there may have been measurement errors in the collected variables, particularly for the timing of the diagnosis of postoperative time-to-event outcomes. Some of the complications may be initially asymptomatic, and different patients experiencing complications may present with different levels of severity. Second, the severity of the outcomes was not available, but it is likely important for patients and clinicians to contextualize the predicted risk. Third, the NSQIP lacks intraoperative and many postoperative variables that may augment the models [38] and provide a real-time prediction that modifies risk prediction according to the patient’s postoperative course. Fourth, each model performance metric has limitations requiring nuanced interpretation [27]. The discrepancy between a high Harrell’s concordance index and low R-squared and scaled Brier scores may be due to low incidences of outcomes and a high percentage of censored data [32, 39]. The range of calibration was limited likely due to the low incidence of outcomes. Finally, the proportional hazard assumption was not satisfied for certain predictors. Nevertheless, the close correspondence in outcomes when comparing with results from bootstrapping and accelerated failure time models indicate that this did not significantly impact the results [33].

Future directions include validating the model using more recent NSQIP data and external validation using other prospective perioperative datasets, as well as optimizing predictors and modeling techniques. The incorporation of long-term outcome data would be important, as the 30-day mortality may underestimate the 90-day mortality rate [40]. Once refined and externally validated, the models could be incorporated into personalized risk calculators and postoperative screening guidelines. Finally, patients, caregivers, and clinicians should be engaged to optimize how risk information is conveyed for patient education and perioperative planning.

In conclusion, we developed and internally validated Cox proportional hazards models to predict time-to-event for major medical complications, readmission, and mortality elective colectomy. There are distinct temporal patterns for each postoperative complication. Once further validated, the models can facilitate tailored monitoring of high risk patients during high risk periods.

Supporting information

Acknowledgments

We thank Ming Yuan, Wei Tang, and Rama Sreepada for statistical advice for the protocol. Drs. Janny Ke, Tim Jen, Alana Flexman, and Stephan Schwarz thank the St. Paul’s Hospital/Providence Health Care Department of Anesthesia for research support. We thank our patient partner, Jeff Reuger, for reviewing our protocol and manuscript to provide patient perspectives.

Disclaimer: The American College of Surgeons National Surgical Quality Improvement Program and its participating hospitals are the sources of the data used herein; they have not been verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors.