Development and validation of the VA Lung Cancer Mortality (VALCAN-M) score for 90-day mortality following surgical treatment of clinical stage I lung cancer

Structured Abstract

Objective:

To develop and validate the Veterans Administration (VA) Lung Cancer Mortality (VALCAN-M) score, a risk prediction model for 90-day mortality following surgical treatment of clinical stage I non-small cell lung cancer (NSCLC).

Background:

While surgery remains the preferred treatment for functionally fit patients with early-stage NSCLC, less invasive, non-surgical treatments have emerged for high-risk patients. Accurate risk prediction models for post-operative mortality may aid surgeons and other providers in optimizing patient-centered treatment plans.

Methods:

We performed a retrospective cohort study using a uniquely compiled VA dataset including all Veterans with clinical stage I NSCLC undergoing surgical treatment between 2006 and 2016. Patients were randomly split into derivation and validation cohorts. We derived the VALCAN-M score based on multivariable logistic regression modeling of patient-and treatment-variables and 90-day mortality.

Results:

A total of 9749 patients were included (derivation cohort: n=6825, 70.0%; validation cohort: n=2924, 30.0%). The 90-day mortality rate was 4.0% (n=390). The final multivariable model included 11 factors that were associated with 90-day mortality: age, body mass index, history of heart failure, forced expiratory volume (FEV1, % predicted), history of peripheral vascular disease, functional status, delayed surgery, American Society of Anesthesiology performance status, tumor histology, extent of resection (lobectomy, wedge, segmentectomy, or pneumonectomy), and surgical approach (minimally invasive or open). The c-statistic was 0.739 (95% CI=0.708-0.771) in the derivation cohort.

Conclusions:

The VALCAN-M score uses readily available treatment-related variables to reliably predict 90-day operative mortality. This score can aid surgeons and other providers in objectively discussing operative risk among high-risk patients with clinical stage I NSCLC considering surgery versus other definitive therapies.

Mini-Abstract

Less invasive, non-surgical treatments have emerged for patients with lung cancer who are considered high-risk for surgical treatment. In this study, we used a uniquely compiled cohort of 9749 Veterans receiving lung cancer resection to develop an easy-to-use, 11-variable score that reliably predicts post-operative mortality within 90 days after surgery. The score was also associated with risk of post-operative major complications and overall survival. The VA Lung Cancer Mortality (VALCAN-M) score may aid surgeons and other providers in objectively discussing operative risk among high-risk patients with clinical stage I NSCLC considering surgery versus other definitive therapies.

Introduction

Non-small cell lung cancer is the leading cause of cancer related mortality in the United States¹. The relative incidence of stage I lung cancer is increasing in part due to stage migration associated with the more widespread adoption of lung cancer screening programs^2,3. Surgical treatment remains the preferred therapy for functionally fit patients with stage I disease⁴; however, other definitive treatments (i.e., stereotactic body radiotherapy [SBRT]) have emerged as alternatives for high-risk surgical candidates who carry an elevated risk of short-term morbidity or mortality following surgery⁵. Unfortunately, direct comparisons of SBRT and surgery are lacking, though ongoing studies, including the Veterans Affairs Lung cancer surgery Or stereotactic Radiotherapy (VALOR) trial, hope to assess this⁶. A challenge of such comparisons, however, is the complexity of assigning treatments based on objective, individualized risk assessment⁷.

Contemporary scores that accurately stratify operative risk for patients with newly diagnosed lung cancer can help to inform the patient-centered decision process for choosing between these treatment approaches. While previous studies have proposed models for predicting operative risk, the datasets used for model development have been limited by important factors (e.g., atypical patient populations, stage heterogeneity, limited availability of covariates, access to care disparities, non-uniform insurance coverage, etc.), and have produced scores with suboptimal performance^8,9. Additionally, these tools have been rarely adopted into routine clinical practice¹⁰. Furthermore, prior models have assessed 30-day mortality despite a growing body of evidence suggesting that broader time horizons – like 90-day mortality – are more accurate metrics of operative risk¹¹. Indeed, studies routinely demonstrate that mortality rates following pulmonary resection approximately double from 30- to 90-days post-operatively, suggesting that analyzing only the former may fail to accurately capture perioperative risk^12-14.

Operative risk scores are most likely to alter practice in high-risk candidates, such as patients with significant comorbidities or patients who are smoking^13,15,16. The Veterans Health Administration (VHA) serves a patient population which carries elevated operative risk; despite this, lung cancer outcomes are similar in Veterans compared to the general US population¹³. Such a high-risk population would theoretically be ideally suited for developing a risk prediction model for operative mortality.

In this study, we employed a uniquely compiled VHA dataset consisting of Veterans with clinical stage I NSCLC to develop and validate the VA Lung Cancer Mortality (VALCAN-M) score for 90-day mortality following surgical treatment.

Methods

Study Design and Patient Population

We performed a retrospective cohort study using a VHA dataset consisting of all Veterans with clinical stage I NSCLC receiving surgery (2006-2016). Patients with lung cancer were identified via ICD-O-3 (International Classification of Diseases for Oncology, Third Edition) codes. Surgical treatments were confirmed by using ICD 9/10 procedure or CPT codes and by manual review of pathology and operative reports. The dataset was assembled using the VHA Informatics and Computing Infrastructure which includes multiple administrative and clinical data sources within the Corporate Data Warehouse (CDW)¹⁷.

The St. Louis VHA Research and Development Committee and Institutional Review Board granted a waiver for consent given the deidentified nature of the analysis. The study was performed in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines¹⁸.

Outcomes

The primary outcome was all-cause mortality (binary) within 90 days of surgery. In-hospital and out-of-hospital deaths were included and verified using the VHA Vital Status File (VSF)¹⁹. We chose 90-day mortality as our outcome since this time horizon better captures peri-operative risk compared to shorter periods^12,13. In exploratory analyses, the relationship between our developed score and several secondary outcomes was assessed, including major complications (pneumonia, empyema, myocardial infarction, cardiorespiratory failure, renal failure, or cerebrovascular accident)^16,20, readmission, prolonged length of stay (LOS) (>14 days), and overall survival. Patients were censored at 5-years or end of study follow-up (May 1, 2020).

Predictor covariates

VHA datasets provide a wide-variety of patient-, treatment-, and tumor-related covariates, several of which are unavailable in other clinical or administrative databases²¹. We obtained candidate predictor variables for this study from various datasets within the CDW, including Oncology Raw and VA Surgical Quality Improvement Program (VASQIP)²². While these datasets are considered highly accurate and regularly audited, a dedicated team of researchers at the St. Louis VHA augmented certain data elements through chart review to ensure maximal accuracy and low missingness^13,15,16,23. Since the score was designed for use in pre-operative risk assessment, we excluded variables that would not be reasonably available prior to an operation (occurrence of complication, occurrence of readmission, etc.). We also excluded some non-modifiable sociodemographic variables. For example, patient race was excluded since inclusion of race as a variable in risk prediction modeling has been widely criticized as propagating racial and racist disparities in health^24-26.

We evaluated several covariates in our score which are routinely available or reasonably known in the pre-operative period, including: age, sex, body mass index (BMI, kg/m²), smoking status (current, former, never; assessed at several different timepoints prior to surgery)^15,16, Charlson comorbidity score (with each comorbidity assessed separately in the model), unique prescription classes, forced-expiratory volume at 1 second (FEV1) percent predicted, functional status (independent, partially dependent, or totally dependent as defined by VASQIP)^27,28, area deprivation index (a county-based measure of socioeconomic deprivation based on 17 poverty, education, housing, and employment indicators from the US census)²³, distance from treating facility, hospital volume, American Society of Anesthesiology (ASA) performance status, delayed surgery (>12 weeks from diagnosis)²³, tumor size, tumor location, histology (adenocarcinoma, squamous cell carcinoma, or other), extent of resection (lobectomy, wedge, segmentectomy, or pneumonectomy), extent of lymph node sampling, and surgical approach (minimally invasive or open).

Statistical Analysis

Cohort descriptive statistics were presented as mean (standard deviation, SD) for continuous variables or absolute numbers (percent) for categorical variables. Median (interquartile range, IQR) were presented for non-normally distributed variables. Missing covariates were reported in univariate analysis and were imputed (n=20) using the Proc MI MCMC function in SAS (SAS Institute, Cary, NC). The linearity of the relationship between the continuous candidate predictor variable and outcome was assessed using restricted cubic spline functions and likelihood ratio tests for non-linearity (lgtphcurv9 macro in SAS²⁹).

To derive the model, we randomly split the cohort into derivation (70%) and validation (30%) cohorts. Using multivariable logistic regression, we modeled the relationship between candidate variables and 90-day mortality in the derivation cohort. We used stepwise backward selection with a significance value of 0.15 to remove non-significant variables from the model to obtain the most parsimonious model (i.e., easy-to-use) while maintaining optimal model performance³⁰. The β coefficients from the derivation cohort were then fixed to develop an integer-based score³¹. Model performance was evaluated in the validation cohort using calibration and discrimination parameters. Calibration was assessed by measuring the agreement between predicted versus observed mortality by decile and analyzing with the calibration slope, intercept, and Hosmer-Lemeshow goodness-of-fit test. Discrimination was assessed using receiver operator characteristic (ROC) curves and c-statistics (with 0.5 representing no discrimination and 1.0 representing perfect discrimination).

Discrete risk groups were created based on the model output. In exploratory analyses, the association between the VALCAN-M risk group and short-term outcomes (major complication, prolonged LOS, readmission) were assessed. Similarly, the association between risk group and overall survival was displayed using Kaplan Meier curves. All tests were 2-tailed and p-values less than 0.05 were considered statistically significant. All analyses were performed in SAS version 9.3 (SAS Institute, Cary, NC).

Results

Study cohort

The study cohort included 9749 Veterans with clinical stage I NSCLC receiving surgical treatment (eFigure 1). The mean (SD) age was 67.6 (7.9) years old, 9,391 (96.3%) were male, and 8,067 (82.8%) identified as white race (Table 1). Current smoking at the time of surgery was reported by 5,701 (58.5%) Veterans. The median (IQR) Charlson score was 7 (5-8), with the most common comorbidities being chronic obstructive pulmonary disease (n=6,155, 63.1%) and peripheral vascular disease (n=4,372, 44.9%). The median (IQR) FEV1 % predicted was 78 (66-95). The most common procedure was lobectomy (n=6,913, 70.9%) and the most common surgical approach was via thoracotomy (n=5,690, 58.5%). A majority of tumors were adenocarcinomas (n=5,195, 53.3%) and most tumors exhibited higher-grade features (grade II-IV, n=7,927, 81.3%). The rates of major post-operative complication and readmission after surgery were 13.9% (n=1,351) and 7.8% (n=759), respectively (eTable 1). The rates of 30- and 90-day mortality were 2.1% (n=207) and 4.0% (n=390), respectively.

Table 1.

Patient characteristics in derivation and validation cohort

Demographics	Full Cohort N=9749	Derivation Cohort N=6825	Validation Cohort N=2924	P-value
Age, years (SD)	67.6 (7.9)	67.6 (8.0)	67.6 (7.8)	0.908
Sex (%)				0.223
Male	9,391 (96.3)	6,564 (96.2)	2,827 (96.7)
Female	358 (3.7)	261 (3.8)	97 (3.3)
Race (%)a				0.399
White	8,067 (82.8)	5,636 (82.8)	2,431 (83.1)
Black	1,457 (15.0)	1,028 (15.1)	429 (14.7)
Other	131 (1.3)	99 (1.5)	32 (1.1)
Unknown	94 (1.0)	62 (0.9)	32 (1.1)
Smoking status (%)b				0.906
Current	5,701 (58.5)	4,001 (58.6)	1,700 (58.1)
Former	3,915 (40.2)	2,731 (40.0)	1,184 (40.5)
Never	133 (1.4)	93 (1.4)	40 (1.4)
BMI, kg/m2 (SD)c	27.3 (5.3)	27.2 (5.4)	27.2 (5.3)	0.994
BMI category (%)				0.784
<18.5	292 (3.2)	206 (3.2)	86 (3.1)
18.5-24.9	3,125 (33.8)	2,199 (34.1)	926 (33.3)
25.0-29.9	3,328 (36.0)	2,310 (35.8)	1,018 (36.6)
30.0-34.9	1,777 (19.2)	1,228 (19.0)	549 (19.7)
≥35.0	712 (7.7)	506 (7.9)	206 (7.4)
FEV1% (IQR)c	78 (66-95)	78 (65-95)	79 (66-96)	0.204
Charlson/Deyo score (IQR)	6.9 (2.2)	6.9 (2.2)	6.9 (2.2)	0.542
Functional Status (%)d,c				0.622
Independent	7,449 (96.9)	5,210 (96.8)	2,239 (97.0)
Partially or Fully Dependent	238 (3.1)	170 (3.2)	68 (3.0)
Treatment Characteristics
Wait time to surgery, days
Median (IQR)	63 (41-96)	62 (41-95)	63 (41-97)	0.161
>12 weeks (%)	3,046 (31.2)	2,105 (30.8)	941 (32.2)	0.191
Tumor size, mm (%)				0.554
≤10	861 (8.8)	591 (8.7)	270 (9.2)
11-20mm	3,831 (39.3)	2,656 (38.9)	1,175 (40.2)
21-30mm	2,653 (27.2)	1,871 (27.4)	782 (26.7)
31-40mm	1,746 (15.1)	1,041 (15.3)	435 (14.9)
40+ mm	719 (7.4)	520 (7.6)	199 (6.8)
Unknown	209 (2.1)	146 (2.1)	63 (2.2)
Resection (%)				0.876
Lobectomy	6,913 (70.9)	4,850 (71.1)	2,059 (70.4)
Wedge	2,139 (22.0)	1,489 (21.8)	650 (22.3)
Segment	541 (5.6)	376 (5.5)	165 (5.7)
Pneumonectomy	156 (1.6)	106 (1.6)	50 (1.7)
Surgical approach (%)c				0.436
VATS	4,032 (41.5)	2,840 (41.7)	1,192 (40.9)
Thoracotomy	5,690 (58.5)	3,966 (58.3)	1,724 (59.1)
Histology (%)				0.627
Adenocarcinoma	5,195 (53.3)	3,618 (53.0)	1,577 (53.9)
Squamous cell	3,294 (33.8)	2,313 (33.9)	981 (33.6)
Other	1,260 (12.9)	894 (13.1)	366 (12.5)

BMI=body mass index; FEV=forced expiratory volume; VATS=video-assisted thoracoscopic surgery

^aDefined according to American College of Surgery Facility Oncology Registry Data Standards manual.

^bSmoking status (current, former, or never) was assessed at the time of surgery.

^cMissing data were present for BMI (n=515, 5.3%), FEV1 (2333, 23.9%), functional status (2062, 21.2%) and surgical approach (n=27, 0.3%).

^dDefined according to VASQIP as independent, partially dependent (patient requires assistance for some activities of daily living) and totally dependent (patient requires assistance for all activities of daily living)

Model development and performance

The dataset was randomly divided into derivation (n=6825, 70.0%) and validation (n=2924, 30.0%) cohorts. Factors associated with higher likelihood of 90-day mortality (Table 2) included older age (multivariable-adjusted odds ratio [aOR] 1.057, 95% CI 1.039-1.074, p<0.001, Table 1), lower BMI (e.g., <18.5 vs 18.5-24.9 kg/m², aOR 3.849, 95% CI 2.365-6.266, p<0.001), congestive heart failure (aOR 1.626, 95% CI 1.169-2.262, p=0.005), peripheral vascular disease (aOR 1.394, 95% CI 1.072-1.813, p=0.01), dependent functional status (aOR 2.881, 95% CI 1.753-4.733, p<0.001), delayed surgery (aOR 1.793, 95% CI 1.398-2.302, p<0.001), ASA classification (for every 1-unit increase, aOR 1.384, 95% CI 1.031-1.860, p=0.03), squamous cell histology (vs. adenocarcinoma, aOR 1.431, 95% CI 1.093-1.873, p=0.009), and pneumonectomy (vs. lobectomy, aOR 5.305, 95% CI 3.096-9.091, p<0.001). Factors associated with lower likelihood of 90-day mortality included higher FEV1 (for every 1-unit increase, aOR 0.993, 95% CI 0.986-1.000, p=0.05), minimally invasive approach (aOR 0.584, 95% CI 0.439-0.776, p<0.001), and wedge resection (vs. lobectomy, aOR 0.653, 95% CI 0.465-0.918, p=0.01).

Table 2.

Final multivariable model of factors associated with mortality at 90 days following surgery

Variable	aOR, 95% CI	β	p-value
Age, y	1.057 (1.039-1.074)	0.055	<0.001
BMI, kg/m2
<18.5	3.849 (2.365-6.266)	1.348	<0.001
18.5-24.9	[1 ref]
25.0-29.9	0.879 (0.645-1.197)	−0.129	0.41
30.0-34.9	0.734 (0.493-1.092)	−0.309	0.13
≥35.0	0.920 (0.558-1.519)	−0.083	0.75
FEV1, % pred.	0.993 (0.986-1.000)	−0.007	0.05
CHF
No	[1 ref]
Yes	1.626 (1.169-2.262)	0.486	0.004
PVD
No	[1 ref]
Yes	1.394 (1.072-1.813)	0.332	0.01
Functional statusa
Independent	[1 ref]
Dependent	2.881 (1.753-4.733)	1.058	<0.001
Delayed surgery
≤ 12 weeks	[1 ref]
>12 weeks	1.793 (1.398-2.302)	0.584	<0.001
ASA classification	1.384 (1.031-1.860)	0.325	0.03
Histology
Adenocarcinoma	[1 ref]
Squamous cell carcinoma	1.431 (1.093-1.873)	0.358	0.009
Other	1.274 (0.866-1.875)	0.242	0.22
Surgical approach
Open	[1 ref]
VATS	0.584 (0.439-0.776)	−0.538	<0.001
Extent of resection
Lobectomy	[1 ref]
Pneumonectomy	5.305 (3.096-9.091)	1.669	<0.001
Segmentectomy	0.674 (0.360-1.260)	−0.395	0.22
Wedge	0.653 (0.465-0.918)	−0.426	0.01

BMI=body mass index; FEV=forced expiratory volume; CHF=congestive heart failure; PVD=peripheral vascular disease; VATS=video-assisted thoracoscopic surgery

^aDefined according to VASQIP as independent, partially dependent (patient requires assistance for some activities of daily living), and totally dependent (patient requires assistance for all activities of daily living)

The c-statistic was 0.739 (95% CI 0.708-0.771) in the derivation cohort and 0.669 (95% CI 0.615-722) in the validation cohort (eFigure 2-3). The model was well calibrated with non-significant Hosmer-Lemeshow goodness-of-fit tests in both the derivation (p=0.49) and validation (p=0.51) cohorts (eFigure 4-5).

Development of VALCAN-M score

We next developed an easy-to-use, integer-based score to predict the likelihood of 90-day mortality. Details for calculating this score are available in Figure 1. The final score ranged from −10 to 42, with most patients having scores between −5 and 15. As shown in eFigure 6, the predicted probability of mortality increases with higher scores. For example, a score of −5 corresponds to a 0.4% risk of mortality while a score of 10 corresponds to a 7.3% risk of mortality. Score distributions were similar between the derivation and validation cohorts (eFigure 7). Based on this score, the cohort was subsequently divided into clinically relevant categories based on operative risk: very-low risk (score ≤−1, predicted mortality <1%; n=886, 9.1%); low risk (score=0-6, predicted mortality 1-4%; n=5589, 57.3%); moderate risk (score=7-11, predicted mortality 4-10%; n=2597, 26.6%); and high risk (score ≥12, predicted mortality >10%; n=677, 6.9%). The rate of 90-day mortality in each risk category was similar between the derivation and validation cohorts (Figure 2). For example, compared to individuals in the very low risk category, individuals in the moderate risk category exhibited a 5-fold higher likelihood of death within 90 days of surgery (eTable 2).

Figure 1. Calculation of VA Lung Cancer Mortality (VALCAN-M) Score.

The integer-based score ranges from −10 to 42, with higher scores representing higher risk of 90-day mortality. The c-statistic was 0.739 (95% CI 0.708-0.771) in the derivation cohort and 0.669 (95% CI 0.615-722) in the validation cohort. The model was well calibrated with non-significant Hosmer-Lemeshow goodness-of-fit tests in both the derivation (p=0.49) and validation (p=0.51) cohorts.

Figure 2. Rates of 90-day mortality by group (very-low, low, medium, high).

Patients were sub-divided into discrete categories based on the predicted probability of 90-day mortality: very-low risk (score ≤−1, predicted mortality <1%); low risk (score=0-6, predicted mortality 1-4%); moderate risk (score=7-11, predicted mortality 4-10%); and high risk (score ≥12, predicted mortality >10%).

In exploratory analyses, we assessed whether the VALCAN-M score was also predictive of other short- and long-term post-operative outcomes. As displayed in eTable 3, higher VALCAN-M risk category was associated with higher likelihood of major complications (e.g., moderate vs. very low, OR 2.698, 95% CI 2.059-3.535, p<0.001), prolonged length of stay (e.g., moderate vs. very low, OR 2.818, 95% CI 2.115-3.756, p<0.001), and readmission (e.g., high vs. very low, OR 1.464, 95% CI 1.029-2.082, p=0.03). In terms of overall survival, higher risk category was also associated with significantly lower rates of 5-year overall survival (very-low=71.1%; low=61.1%; moderate=49.4%; high=34.7%; p<0.001, Figure 3), even after controlling for pathologic stage and receipt of adjuvant therapy (eFigure 8 and eTable 4).

Figure 3. Association between VALCAN-M Score and 5-year Overall Survival.

Overall survival in very-low risk, low risk, moderate risk, and high risk subgroups.

Discussion

In this study, we developed the VALCAN-M score, a tool for predicting 90-day mortality following the surgical treatment of clinical stage I NSCLC. By leveraging data from a uniquely compiled dataset in the VHA, the largest integrated healthcare system in the U.S., we developed an 11-variable, parsimonious, easy-to-use model for assessing operative risk. The score, which considers routinely available data in the pre-operative period, accurately subdivides patients into various risk categories. Further, while the score was developed to assess operative mortality, it is also highly correlated with other short-term outcomes (readmission, major complications, prolonged length of stay) and overall survival. This score can objectively inform both patients and providers about operative risk among high-risk patients with clinical stage I NSCLC considering surgery versus other definitive therapies.

The optimal treatment (e.g., surgery vs. SBRT) for clinical stage I NSCLC in medically high-risk patients remains a heavily contested topic, including in the VHA with the highly anticipated VALOR trial⁶. While prospective studies comparing SBRT and surgery are critically lacking, most contemporary analyses suggest that surgery is associated with significantly longer overall and cancer-specific survival, including lower rates of locoregional recurrence, in functionally fit patients compared to primary radiotherapy^4,6,32-35. A frequent observation in various retrospective analyses, however, is that SBRT may have similar outcomes to surgery in non-functionally fit (i.e., “high-risk”) patients^4,32. For example, Puri and colleagues performed a propensity-matched analysis of patients with clinical stage I NSCLC receiving surgery or SBRT using the National Cancer Database⁴. They observed that patients in the highest quintile of propensity score (i.e., highest surgical risk) had similar survival outcomes compared to patients who received SBRT. These findings suggest that the short-term risk of post-operative mortality may supersede the long-term benefits of surgery among high-risk patients. Consequently, the modern practice landscape largely reserves SBRT for patients who are deemed to be poor candidates for surgery (in addition to patients who decline surgery). Unfortunately, the threshold of “high-risk” is variable and large surgical series have shown good outcomes even in these patients³⁶. This uncertainty, at least partly, explains the low accrual rates and subsequent premature closure of several prospective trials examining surgery versus SBRT⁶. The VALCAN-M score is a mechanism of objectively quantifying operative risk. This easily calculable personalized risk-assessment tool can help surgeons when objectively discussing the short-term risks of surgery with patients.

The relative value of short- versus long-term outcomes is important to discuss with patients. For example, we found that patients in the moderate- and high-risk categories (VALCAN-M score ≥7) represented a sizable proportion of the population receiving surgery (33.6%). While the rates of 90-day mortality were elevated in these groups (6.0% and 15.7%, respectively), the more striking finding was the dismal 5-year overall survival in these groups (49.4% and 34.7%, respectively). It is important to discuss these findings with high-risk patients considering surgery versus SBRT since the presumed long-term benefits of surgery may not be realized by patients in these high-risk groups given their significantly elevated risk of dying from non-cancer-related causes. Prospective studies examining SBRT versus surgery would be most useful in these moderate and high-risk groups.

It is important to reflect on several of the patient- and treatment-related factors in this model, particularly the modifiable surgical factors. For example, we found that sublobar resection and minimally invasive approach were associated with significantly reduced risk of post-operative mortality. Surgeons should discuss such factors with patients, while also acknowledging the caveat that sublobar resections may carry higher risk of locoregional recurrence³⁷. It is also important to discuss factors that were excluded from our model. For example, smoking did not appear in the final model which is perhaps due to the collinearity between smoking status and other variables in the model (i.e., pulmonary function tests, extent of resection, etc.). Surgeons should nonetheless continue to evaluate such factors when designing personalized treatment plans as the relationship between smoking and peri-operative morbidity is well established^15,16.

Prior studies have attempted to quantify operative risk in the setting of lung cancer resection, including the American College of Surgeons NSQIP calculator⁸, the European Society Objective Score (ESOS)³⁸, the Thoracoscore³⁹, and the Society of Thoracic Surgeons-derived calculators⁴⁰. Unfortunately, these scores are infrequently used in modern practice¹⁰. This is due in part to significant methodological limitations of these models, such as stage heterogeneity and the inclusion of non-malignant indications altogether^8,38-40. Further, the databases used to create these scores do not control for access to care barriers, such as variability in insurance coverage. Our score, conversely, was developed using a relatively homogenous cohort of Veterans with clinical stage I NSCLC with theoretically equal access to care. This should allow for more focused application of the model, particularly among high-risk Veterans with clinical stage I NSCLC considering SBRT versus surgery.

Several studies and regulatory agencies view post-operative mortality as a quality measure of surgical care. The VALCAN-M score is as a standardized metric through which to assess operative risk among Veterans, and could potentially aid in quality improvement efforts within the VHA system. However, it is also essential to recognize that 90-day mortality may be a myopic metric of surgical quality, particularly in the setting of lung cancer where a paradox exists between short- and long-term quality⁴¹. For example, it is widely recognized that wedge resections hold a lower risk of post-operative morbidity and mortality²⁰. It is similarly known that such resections carry a higher risk of recurrence and diminished overall survival⁴². Therefore, it is very challenging to assess operative quality when certain factors (like receiving a wedge resection) antithetically affect short- and long-term outcomes⁴³. In other words, surgeons should consider several factors in addition to the VALCAN-M score for designing patient-centered treatment plans.

This study has several strengths. For example, we compiled a unique VHA dataset that reflects a homogenous cohort of patients with clinical stage I NSCLC. Also, given that this score was derived among Veterans with heavy comorbidity burdens, the score is particularly useful in high-risk patients considering surgery. The score additionally reflects the largest and most modern risk-assessment tool to our knowledge among surgical patients with clinical stage I NSCLC. The score was further correlated with various short- and long-term outcomes. Conversely, this study has some limitations. For example, the VHA treats a unique patient population (predominantly male, heavy smoking histories, etc.). Therefore, this score should initially be implemented within the VHA. Validating the VALCAN-M score in non-Veterans would provide evidence for its application in the general population. Another limitation is that the model performance, while similar to prior scores, is modest. This is a common reality of scores that assess operative mortality and reflects the complexity of trying to predict rare post-operative events¹¹. Finally, certain data elements, like CHF severity, were unavailable.

Conclusions

This study describes the development of the VALCAN-M score, a risk stratification tool for Veterans with clinical stage I NSCLC considering surgery. The VALCAN-M score was also associated with risk of major complication, prolonged length of stay, readmission, and overall survival. This score can aid surgeons and other providers in objectively discussing operative risk among high-risk patients with clinical stage I NSCLC considering surgery versus other definitive therapies.