Procedure-Specific Risk Prediction for Recurrence in Patients Undergoing Lobectomy or Sublobar Resection for Small (≤2 cm) Lung Adenocarcinoma: An International Cohort Analysis

Abstract

Purpose:

To develop and validate procedure-specific risk prediction for recurrence following resection for early-stage lung adenocarcinoma (ADC) and investigate risk prediction utility in identifying patients who may benefit from adjuvant chemotherapy (ACT).

Basic Procedures:

In patients who underwent resection for small (≤2 cm) lung ADC (lobectomy, 557; sublobar resection, 352), an association between clinicopathological variables and risk of recurrence was assessed by a competing risks approach. Procedure-specific risk prediction was developed based on multivariable regression for recurrence. External validation was conducted using cohorts (N=708) from Japan, Taiwan, and Germany. The accuracy of risk prediction was measured using a concordance index (C-index). We applied the lobectomy risk prediction approach to a propensity score–matched cohort of patients with stage II-III disease (n=316, after matching) with or without ACT and compared lung cancer-specific survival between groups among low or high-risk scores.

Main Findings:

Micropapillary pattern, solid pattern, lymphovascular invasion, and necrosis were involved in the risk prediction following lobectomy, and micropapillary pattern, spread through air spaces, lymphovascular invasion, and necrosis following sublobar-resection. Both internal and external validation showed good discrimination (C-index in lobectomy and sublobar resection: internal, 0.77 and 0.75; and external, 0.73 and 0.79). In the stage II-III propensity score–matched cohort, among high-risk patients, ACT significantly reduced the risk of lung cancer–specific death (subhazard ratio 0.43, p=0.001), but not among low-risk patients.

Principal Conclusions:

Procedure-specific risk prediction for patients with resected small lung ADC can be used to better prognosticate and stratify patients for further interventions.

INTRODUCTION

Lung adenocarcinoma (ADC) is the most common histological subtype of non-small cell lung cancers (NSCLCs); 25% of which are diagnosed at stage IA.¹ Following the results of the National Lung Screening Trial, the identification of early-stage lung ADC is expected to increase.² The standard of care for early-stage lung ADC is curative-intent anatomic surgical resection by lobectomy; however, sublobar resection is appropriate for selected patients.³ Despite ongoing concerns about the adequacy of sublobar resection for cure,^4–6 the use of sublobar resection is increasing.⁷ Development of a procedure-specific risk prediction model following sublobar resection or lobectomy that takes into consideration-widely-variable patient background and the recurrence risk,^{4, 5, 8} will be useful to predict prognosis. Such a risk model can further help to stratify patients for prospective investigation of potential postoperative interventions (e.g., completion lobectomy and/or adjuvant therapies following sublobar resection or adjuvant therapies following lobectomy).

Based on our literature review, among ten risk prediction algorithms that have been described to predict prognosis following lung resection for all stages of NSCLC (Supplementary Figure S1 and Table S1),^9–18 six have not been externally validated.^13–18 Of the four that have, only two studies externally validated with an international cohort.^{9, 10} No study addressed competing risks, which can bias prognostic assessment especially in early-stage disease.⁸ No study included prognostic histologic subtypes.¹⁹ A recent study by Liang et al.¹⁰ was independently validated using the National Cancer Database (US); however, the tumor, node, metastasis (TNM) staging system still had a superior predictive capability.²⁰

In an effort to expand the prognostication of lung ADC beyond the use of the TNM staging system, a multidisciplinary group comprising experts from the International Association for the Study of Lung Cancer, American Thoracic Society, and European Respiratory Society (IASLC/ATS/ERS) introduced a classification of lung ADC in 2011,¹⁹ which has been validated in independent international cohorts^21–23 and was incorporated into the 2015 World Health Organization (WHO) classification.²⁴ We and others reported the prognostic impact of micropapillary (MIP) and solid (SOL) subtypes in stage I lung ADC^{22, 23, 25–28} and were the first to describe tumor spread through air spaces (STAS),²⁹ which has since been validated by others.^{30, 31} These prognostic pathologic variables including MIP, SOL, and STAS can be associated with each other.²⁹ To our knowledge, no group has investigated how these variables interact and influence outcomes. In order to incorporate multiple, relevant clinicopathologic factors for a given patient, we specifically chose to utilize a risk-based prediction model.³² In this study, we developed two procedure-specific scores for predicting recurrence after curative-intent lobectomy or sublobar resection. Given the significant competing risks in elderly patients with early-stage lung cancer,⁸ we utilized a competing risks analysis. We validated our model with an external data set consisting of three cohorts of patients from Japan, Taiwan and Germany.

The current guidelines do not recommend platinum-based ACT for patients with stage IA NSCLC.^33–35 However, if we could reliably identify stage IA patients with a higher risk of recurrence equal to stage II-III NSCLC patients, we may be able to predict the benefit from ACT. To test this potential utility of this approach, we would ideally use a cohort of patients with stage IA lung ADC. However, since only a minor fraction of stage IA lung ADC (<5%) undergo ACT both at our and other centers, we utilized a propensity-score matched cohort of patients with resected stage II-III lung ADC with or without ACT to investigate the survival benefit from ACT in those patients classified as high risk using the proposed lobectomy risk prediction score.

METHODS

Study cohorts for development and validation of risk prediction for recurrence

This retrospective study was approved by the Institutional Review Board at Memorial Sloan Kettering Cancer Center (MSK). The MSK Thoracic Service’s prospectively maintained lung cancer database was reviewed to identify consecutive patients who had been surgically treated for pathological stage I small (≤2 cm) lung ADC between January 1, 1995, and December 31, 2011. At MSK, lobectomy was considered a standard surgical treatment for patients with stage I NSCLC. Sublobar resection was considered for patients with poor pulmonary reserve or other major comorbidities that relatively contraindicate lobectomy or for patients with a small (≤2 cm) peripheral nodule with ≥50% ground-glass appearance on computed tomography. Pathological stage was based on the eighth edition of the American Joint Committee on Cancer Staging Manual.³⁶ Exclusion criteria included induction therapy, multiple nodules, positive surgical margin (R1 or R2), other lung cancer diagnosis in the past two years, other disease progression, and inadequate tissue available to review. In total, 909 patients at MSK met the inclusion criteria (Supplementary Figure 2A). The development (MSK) cohort of patients was used to create the two procedure-specific risk scores. To externally validate the risk prediction, we used a combined cohort of three independent international sets of patients (Japan, Taiwan, and Germany) (Supplementary Figure 2B). The University of Tokyo (Japan) cohort consisted of 250 consecutive patients treated between September 1, 1998, and December 31, 2012. The Taipei Veterans General Hospital (Taiwan) cohort consisted of 259 consecutive patients treated between April 1, 1996, and December 31, 2010. All tumor slides were reviewed by Y.Y. The Heidelberg University (Germany) cohort consisted of 199 consecutive patients treated between April 1, 2002, and December 31, 2014. All tumor slides were reviewed by A.W. Inclusion and exclusion criteria were the same as those for the primary cohort. Histologic evaluation was performed by experienced pathologists in each institution (A.S. at the University of Tokyo, Y.Y. at Taipei Veterans General Hospital, and A.W. at Heidelberg University) following the same method as the primary cohort, as described later in this Methods section.

Stage II-III cohort

The MSK Thoracic Service’s prospectively maintained database was reviewed to identify consecutive patients who had been surgically treated for pathological stage II and III lung adenocarcinoma (ADC) between January 1, 2000, and December 31, 2013. Exclusion criteria included induction therapy, multiple nodules, positive surgical margin (R1 or R2), other lung cancer diagnosis in the past 2 years, concurrent other disease progression, wedge resection, and no available tumor slides to review. To evaluate benefit from platinum-based (cisplatin or carboplatin plus other drug) adjuvant chemotherapy (ACT), we also excluded patients with unknown ACT status, unknown regimens, nonplatinum-based ACT, adjuvant EGFR-TKI therapy, and <2 cycles of platinum-based ACT. Since perioperative mortality and morbidity can affect treatment decisions as well as outcomes, we excluded patients who died within 90 days of surgery and patients with planned but cancelled ACT due to postoperative morbidity or recurrence. In total, 589 patients with stage II-III lung ADC met our inclusion criteria (Supplementary Figure 1C). In this study, ACT was defined as any additional intravenously administered chemotherapy after the primary surgery, within 3 months after surgery, without recurrence of the resected primary tumor. We evaluated the regimen of ACT, the date of the first dose, and the number of performed cycles of ACT. In patients who did not receive ACT, we evaluated the specific reason for this.

Histological evaluation

Tumor slides were reviewed by at least two experienced thoracic pathologists (K.K., L.S. and W.D.T.) who were blinded to patient clinical outcomes. The percentage of each histological pattern was recorded in 5% increments, and tumors were classified by the predominant subtype (in accordance with the IASLC/ATS/ERS and 2015 WHO classifications): adenocarcinoma in situ; minimally invasive adenocarcinoma; lepidic-, acinar-, papillary-, MIP-, or SOL– predominant invasive adenocarcinoma; invasive mucinous adenocarcinoma; and colloid adenocarcinoma.^{19, 24} Tumor STAS was defined as isolated tumor cells within air spaces surrounding the main tumor.²⁹ The presence of visceral pleural, lymphovascular invasion (LVI) and necrosis was also investigated.

Statistical methods

To develop procedure-specific risk prediction, patients who underwent lobectomy were analyzed separately from those who underwent sublobar resection. The outcome of interest was recurrence. Patients were monitored from date of surgery until recurrence or death, whichever came first. The probability of recurrence was estimated as the cumulative incidence of recurrence (CIR) from the time of surgery and analyzed using a competing risks approach, considering deaths without recurrence as competing events.³⁷ Differences in CIR between groups were assessed using Gray’s method and univariable Fine and Gray’s tests. Definitions of the clinicopathological factors included in the univariable analyses are presented in the Supplementary Materials. Factors that yielded p<0.1 in univariable analyses were considered as candidates in the multivariable models. Technical information on imputation of missing data by predictive mean matching and model building using the adaptive lasso procedure is presented in the Supplementary Materials.

The procedure-specific risk prediction was developed using recurrence probability estimates derived from the final multivariable competing risks models. The predictions of recurrence in both models were projected at 3 years and 5 years. The median follow-up duration for the development and combined external cohorts were computed based on reverse Kaplan-Meier approach. In both models, categorical covariates were included as dummy variables, and nonlinearity of continuous variables was assessed using restricted cubic splines.³⁸ Cube-root transformation was applied to MIP and SOL percentage, given the right-skewed distribution data with a high proportion of zeros. On further investigation of multicollinearity, pleural invasion was excluded from the variable-selection procedure due to its high correlation with LVI. Development of risk prediction was performed with blinding of the external-validation cohort data.

The predictive performance of the procedure-specific risk prediction was assessed by examining discrimination (C-index), calibration (calibration plots), and overall accuracy (Brier score). We also generated decision curves to assess the net benefit of risk prediction assisted decisions. Additional information on the performance measures of the risk prediction is presented in the Supplementary Materials. Internal validations were done with 1000 bootstrap resamples. To complete external validation, we applied the procedure-specific risk prediction to data from Japan, Taiwan, and Germany. The external-validation cohort comprised 708 patients (lobectomy, 551; sublobar resection, 157). Owing to the limited sample size and the incidence of recurrence, the external cohorts were analyzed as a combined cohort. As an exploratory analysis, overall survival (OS) for patients with high and low risk of recurrence was estimated using the Kaplan-Meier approach and compared using log-rank tests (additional information in the Supplementary Materials).

To investigate the potential utility of risk prediction to benefit from ACT in high-risk patients, we developed a propensity score–matched cohort of patients with stage II-III disease who were treated by lobectomy with or without ACT and applied the lobectomy risk prediction score to the matched cohort. The propensity score–matching procedure selects matched pairs with similar baseline probability of being in either the ACT or the no-ACT group.^{39, 40} For matching, we utilized age, sex, smoking status, comorbidities (chronic obstructive pulmonary disease, cardiovascular disease, etc.), serum creatinine level, pulmonary function (forced expiratory volume in one second, diffusion capacity of the lung for carbon monoxide), maximum standard uptake value in (18)F-fluorodeoxyglucose–positron emission tomography, type of resection, surgical approach, pathological tumor size, p-N status, p-stage, pleural invasion, LVI, necrosis, tumor STAS, and morphological grade. To account for the multiple imputations of the missing data, separate logistic regression models were generated from each imputed data set (additional information in the Supplementary Materials). Propensity score–matched pairs were identified without replacement using a 1:1 nearest neighbor matching with caliper width equal to 0.286. The caliper width was determined by the recommendation from Austin (0.2 of the standard deviation of the logit of the propensity scores).⁴¹ Balance of covariates between the groups was assessed by the absolute standardized mean difference (ASMD) before and after the matching procedure. ASMD <0.1 indicates balance in the covariate between the two groups.⁴² We investigated the association between receipt of ACT and various outcomes using the matched cohort: primary outcome, lung cancer–specific cumulative incidence of death (CID); secondary outcomes, CIR and OS. The relationship was characterized by the patient-level score calculated from the lobectomy risk prediction. We categorized as high or low risk based on the predetermined 5-year CIR 18% (lobectomy score 95) that was a previously reported average 5-year CIR in stage I lung ADC.²⁸ Hence, we categorized each patient as “high score” (score is at or above 95) or “low score.” OS was estimated using the Kaplan-Meier approach and compared between the two groups on the basis of the log-rank test, stratified by pathological stage. Statistical analyses were conducted using R 3.3.1 (R Development Core Team, Austria, Vienna), including the “survival,” “cmprsk,” “crrp,” “ClevClinicQHS,” “QHScrnomo,” “rms,” and “pec” packages, downloaded in January 2017.

RESULTS

Patient characteristics

The development (MSK) cohort comprised 909 patients (lobectomy, 557; sublobar resection, 352). The external-validation cohort (N=708) included patients from Japan (N=250; lobectomy/sublobar resection, 165/85), Taiwan (N=259; 221/48), and Germany (N=199; 175/24). Patient clinicopathological characteristics and outcomes for each cohort, as well as a comparison between the development and external-validation cohorts, are shown in Table 1. Among patients who underwent lobectomy, the development cohort was older and had higher proportion of women, larger invasive size tumor, LVI, MIP (≥5%), and lower proportion of low-grade tumors than the external-validation cohorts. The development sublobar resection cohort had higher proportion of women, larger invasive size tumor, LVI, necrosis, MIP (≥5%), SOL (≥5%) and lower proportion of low-grade tumors than the external-validation cohorts. In patients who underwent lobectomy, the number of distant recurrence events was higher than that of locoregional recurrence events in both the development and the external-validation cohorts (number of events for locoregional vs. distant: development, 13 vs. 43; external-validation, 19 vs. 52). In contrast, in patients who underwent sublobar resection, the number of distant recurrence events was lower than that of locoregional recurrence events in both the development and the external-validation cohorts (number of events for locoregional vs. distant: development, 43 vs. 30; external-validation, 15 vs. 8).

Table 1:

Clinicopathological and demographic characteristics for each cohort and comparison between the Memorial Sloan Kettering Cancer Center (development) cohort and the combined external cohort

	Lobectomy						Sublobar resection
	MSK n=557	External cohort				P MSK vs. Comba	MSK n=352	External cohort				P MSK vs. Comba
		Combined n=551	Japan n=165	Taiwan n=211	Germany n=175			Combined n=157	Japan n=85	Taiwan n=48	Germany n=24
VARIABLES
Age at surgery						<0.001						0.7
≤65	246 (44)	313 (57)	83 (50)	130 (62)	100 (57)		120 (34)	59 (38)	34 (40)	18 (38)	7 (29)
>65	311 (56)	238 (43)	82 (50)	81 (38)	75 (43)		232 (66)	98 (62)	51 (60)	30 (63)	17 (71)
Sex						0.001						<0.001
Female	340 (61)	283 (51)	83 (50)	114 (54)	86 (49)		224 (64)	61 (39)	37 (44)	17 (35)	7 (29)
Male	217 (39)	268 (49)	82 (50)	97 (46)	89 (51)		128 (36)	96 (61)	48 (56)	31 (65)	17 (71)
Smoking history						<0.001						<0.001
Never	93 (17)	151 (27)	84 (51)	46 (22)	21 (12)		52 (15)	43 (27)	37 (44)	5 (10)	1 (4)
Former/current	464 (83)	257 (47)	81 (49)	33 (16)	143 (82)		300 (85)	79 (50)	48 (56)	9 (19)	22 (92)
Unknown	0	143	0	132	11		0	35	0	34	1
p-Stage						<0.001						1
IA	498 (89)	417 (76)	152 (92)	147 (70)	118 (67)		296 (84)	132 (84)	79 (93)	34 (71)	19 (79)
IB	59 (11)	134 (24)	13 (8)	64 (30)	57 (33)		56 (16)	25 (16)	6 (7)	14 (29)	5 (21)
Tumor size						<0.001						0.11
≤1 cm	88 (16)	110 (20)	40 (24)	43 (20)	27 (15)		114 (32)	63 (40)	43 (51)	13 (27)	7 (29)
>1 cm	469 (84)	441 (80)	125 (76)	168 (80)	148 (85)		238 (68)	94 (60)	42 (49)	35 (73)	17 (71)
Invasive tumor size						<0.001						0.003
≤1 cm	198 (36)	270 (49)	109 (66)	104 (49)	57 (33)		191 (54)	107 (68)	71 (84)	26 (54)	10 (42)
>1 cm	359 (64)	281 (51)	56 (34)	107 (51)	118 (67)		161 (46)	50 (32)	14 (16)	22 (46)	14 (58)
LVI						<0.001						<0.001
Absent	338 (61)	397 (72)	130 (79)	180 (85)	87 (50)		229 (65)	130 (83)	75 (88)	43 (90)	12 (50)
Present	219 (39)	154 (28)	35 (21)	31 (15)	88 (50)		123 (35)	27 (17)	10 (12)	5 (10)	12 (50)
Pleural invasion						<0.001						1
Absent	498 (89)	417 (76)	152 (92)	147 (70)	118 (67)		296 (84)	132 (84)	79 (93)	34 (71)	19 (79)
Present	59 (11)	134 (24)	13 (8)	64 (30)	57 (33)		56 (16)	25 (16)	6 (7)	14 (29)	5 (21)
Necrosis						0. 3						<0.001
Absent	444 (80)	457 (83)	145 (88)	160 (76)	152 (87)		274 (78)	143 (91)	79 (93)	45 (94)	19 (79)
Present	108 (19)	94 (17)	20 (12)	51 (24)	23 (13)		76 (22)	14 (9)	6 (7)	3 (6)	5 (21)
Unkonwn	5	0					2	0
Histologic gradeb						<0.001						<0.001
Low	80 (14)	169 (31)	83 (50)	71 (34)	15 (9)		86 (24)	85 (54)	62 (73)	21 (44)	2 (8)
Intermediate	352 (63)	264 (48)	55 (33)	95 (45)	114 (65)		189 (54)	43 (27)	12 (14)	16 (33)	15 (63)
High	125 (22)	118 (21)	27 (16)	45 (21)	46 (26)		77 (22)	29 (18)	11 (13)	11 (23)	7 (29)
MIP						<0.001						<0.001
Absent (<5%)	287 (52)	365 (66)	147 (89)	114 (54)	104 (59)		197 (56)	121 (77)	79 (93)	27 (56)	15 (63)
Present (≥5%)	270 (48)	186 (34)	18 (11)	97 (46)	71 (41)		155 (44)	36 (23)	6 (7)	21 (44)	9 (38)
SOL						0. 3						0.006
Absent (<5%)	353 (63)	367 (67)	113 (68)	156 (74)	98 (56)		220 (63)	118 (75)	69 (81)	37 (77)	12 (50)
Present (≥5%)	204 (37)	184 (33)	52 (32)	55 (26)	77 (44)		132 (38)	39 (25)	16 (19)	11 (23)	12 (50)
STAS						0. 5						0.1
Absent	366 (66)	350 (64)	121 (73)	133 (63)	96 (55)		226 (64)	113 (72)	73 (86)	27 (56)	13 (54)
Present	191 (34)	201 (36)	44 (27)	78 (37)	79 (45)		126 (36)	44 (28)	12 (14)	21 (44)	11 (46)
OUTCOMES
No. of recurrence
Any	56	71	14	30	27		73	23	10	10	3
Locoregionalc	13	19	3	8	8		43	15	6	7	2
Distant	43	52	11	22	19		30	8	4	3	1
No. of death
Any	124	74	15	31	28		129	27	11	10	6
Recurrence (+)	39	38	7	15	16		54	11	5	5	1
Recurrence (−)	85	36	8	16	12		75	16	6	5	5
5-yr CIR (%)d	11 (81, 13)	13 (10, 17)	9 (5, 15)	13 (9, 18)	18 (12, 26)	0.025e	21 (17, 26)	15 (9, 21)	12 (6, 20)	18 (10, 31)	19 (6, 45)	0.3e
5-yr OS (%)d	88 (85, 91)	89 (86, 93)	92 (88, 97)	91 (86, 95)	85 (78, 92)	0.085f	72 (67, 77)	84 (77, 90)	90 (83, 97)	81 (70, 94)	61 (39, 97)	0.004f
Median follow-up (months)g	80 (76, 83)	61 (57, 63)	63 (54, 68)	78 (59, 86)	51 (44, 57)		70 (67, 76)	56 (50, 63)	56 (43, 67)	63 (49, 79)	35 (20, 47)

Data are number (%). CIR, cumulative incidence of recurrence; LVI, lymphovascular invasion; MIP, micropapillary; MSK, Memorial Sloan Kettering Cancer Center; OS, overall survival; SOL, solid; STAS, spread through air spaces.

^aComparison between MSK versus combined external cohorts.

^bHistologic grade based on predominant subtypes (low grade, lepidic; intermediate, acinar or papillary; high, micropapillary, solid, invasive mucinous, or colloid).

^cOnly locoregional recurrence without distant recurrence.

^dData are shown as percentage (95% confidence interval).

^eGray’s test.

^fLog-rank test.

^gEstimated median follow-up time (95% confidence interval).

Estimated 5-year CIR and OS are available in Table 1. The CIR and OS curves for each cohort are shown in the Supplementary Figure S3.

Lobectomy risk prediction score

In the multivariable model for recurrence following lobectomy, four risk factors (MIP percentage, SOL percentage, LVI, necrosis) were significantly associated with hazard of recurrence lobectomy (Table 2). The resulting lobectomy risk prediction (Figure 1) and scoring systems (Supplementary Table S2) were generated using the model estimates. In internal validation, the optimism-corrected C-index was 0.77 (95% confidence interval [CI] 0.72–0.83), indicating good discrimination. In external validation, the C-index was 0.73 (95% CI, 0.67–0.79), indicating good discrimination (Supplementary Table S3). The C-indices of the lobectomy risk prediction were higher than those of the TNM classification (risk prediction score vs. TNM: internal 0.74 vs. 0.58; and external 0.73 vs. 0.68) (Supplementary Table S4). The calibration plots at 5 years for the development (MSK) and external-validation cohorts indicate moderate calibration (Figure 2). Brier scores (Supplementary Table S3) and decision curves (Supplementary Figure S4) are presented in the Supplementary Materials.

Table 2:

Results of the univariable and multivariable competing risks analyses for any recurrence, by lobectomy and sublobar resection

	Lobectomy							Sublobar resection
		Univariable analysis			Final multivariable model				Univariable analysis			Final multivariable model
Variables	5-yr CIR	SHR	95% CI	P	SHR	95% CI	P	5-yr CIR	SHR	95% CI	P	SHR	95% CI	P
Age >65	10%	0.78	(0.46 to 1.32)	0.4				21%	0.98	(0.61 to 1.57)	0.9
ref. ≤65	12%							22%
Age (per 1 year increase)	N/A	0.99	(0.97 to 1.02)	0. 6				N/A	1.00	(0.98 to 1.03)	0. 7
Male sex	11%	1.03	(0.60 to 1.76)	0.9				25%	1.40	(0.88 to 2.22)	0.15
ref. Female	10%							19%
Smoker	12%	3.75	(1.17 to 11.97)	0.026				22%	1.26	(0.64 to 2.49)	0. 5
ref. Never	4%							17%
p-Stage IB	26%	2.90	(1.56 to 5.41)	0.001				37%	2.38	(1.39 to 4.05)	0.001
ref. IA	9%							18%
Tumor size >1 cm	11%	1.16	(0.55 to 2.45)	0. 7				23%	1.40	(0.83 to 2.37)	0. 2
ref. ≤1 cm	8%							17%
Tumor size (per 1-cm increase)	N/A	1.42	(0.69 to 2.92)	0.3				N/A	1.42	(0.85 to 2.36)	0.177
Invasive tumor size >1 cm	13%	2.13	(1.13 to 4.04)	0.020				29%	2.16	(1.34 to 3.47)	0.001
ref. ≤1 cm	6%							15%
Invasive tumor size (per 1-cm increase)	N/A	2.39	(1.31 to 4.38)	0.005				N/A	2.87	(1.90 to 4.34)	<0.001
Lymphovascular invasion	20%	4.27	(2.40 to 7.59)	<0.001	2.14	(1.11 to 4.14)	0.024	38%	3.86	(2.40 to 6.21)	<0.001	2.02	(1.16 to 3.51)	0.013
ref. Absent	5%							12%
Pleural invasion	26%	2.90	(1.56 to 5.41)	0.001				37%	2.38	(1.39 to 4.05)	0.0 01
ref. Absent	9%							18%
Necrosis present	24%	3.49	(2.06 to 5.93)	<0.001	1.75	(1.00 to 3.04)	0.050	35%	2.59	(1.59 to 4.20)	<0.001	1.72	(1.06 to 2.82)	0.031
ref. Absent	7%							18%
Histologic grade
High	21%	6.37	(1.94 to 20.88)	0.002				25%	6.32	(2.16 to 18.45)	0.0 01
Intermediate	9%	2.36	(0.73 to 7.65)	0.15				27%	6.04	(2.20 to 16.61)	<0.001
ref. Low	4%							4%
MIP percentage (linear, per 10% increase)	N/A	1.32	(1.13 to 1.53)	<0.001	1.21	(1.04 to 1.43)	0.021
MIP percentage (nonlinear, 10% vs. 0%)a								N/A	3.10	(2.17 to 4.42)	0.001	1.77	(1.17 to 2.67)	0.007
SOL percentage (nonlinear, 15% vs. 0%)a	N/A	2.89	(2.04 to 4.09)	0.002	2.12	(1.39 to 3.22)	<0.001	N/A	1.63	(1.23 to 2.16)	<0.001
STAS present	16%	2.14	(1.27 to 3.60)	0.004				42%	5.22	(3.14 to 8.68)	<0.001	2.48	(1.26 to 4.89)	0.009
ref. Absent	8%							10%

CI, confidence interval; CIR, cumulative incidence of recurrence; N/A, not applicable; MIP, micropapillary; OS, overall survival; SHR, subhazard ratio; SOL, solid; STAS, spread through air spaces.

^aThe two values are the third and first quartiles, respectively, of the variable distribution.

Figure 1: Risk prediction for patients who underwent lobectomy (A) or sublobar resection (B).

These algorithms present a method to calculate the 3-year or 5-year cumulative incidence of recurrence (CIR) after resection on the basis of a patient’s combination of characteristics. To calculate the probability of recurrence, locate the patient’s micropapillary level (%) and draw a straight line up to the “Scores” axis to derive the score associated with the level. Repeat for the other three covariates (solid, lymphovascular invasion [LVI], and necrosis). Add the scores for all covariates and locate the sum on the “Total Scores” axis. Draw a vertical line down from “Total Scores” to the last two axes to obtain the corresponding 3-year and 5-year CIR following lobectomy. STAS, spread through air spaces.

Figure 2: Internal and external calibration for 5-year cumulative incidence of recurrence (CIR) derived from the risk prediction for patients who underwent lobectomy (A, B) or sublobar resection (C, D).

The calibration of the risk score for lobectomy (A, development cohort; B, validation cohort) and sublobar resection (B, development cohort; D, validation cohort) is shown. The horizontal axis is the algorithm prediction of recurrence at 5 years. The vertical axis is the corresponding observed 5-year CIR, based on the cumulative incidence function with the competing risks approach. The dashed line is the reference line on which an ideal algorithm would lie. The solid line indicates the performance of the current risk prediction (dots represent the average predicted 5-year CIR). X’s indicate the optimism-corrected estimate of the risk predictive algorithm performance based on 500 bootstrap resamples. The vertical bars represent 95% confidence intervals.

Sublobar-resection risk prediction score

In the multivariable model for recurrence following sublobar resection, four risk factors (MIP percentage, STAS, LVI, necrosis) were significantly associated with hazard of recurrence after LR (Table 2). The resulting sublobar-resection risk prediction (Figure 1) and scoring system (Supplementary Table S2) were generated using the model estimates. In internal validation, the optimism-corrected C-index was 0.75 (95% CI 0.72–0.81), indicating good discrimination. In external validation, the C-index was 0.79 (95% CI 0.73–0.87), indicating good discrimination (Supplementary Table S3). The C-indices of the sublobar resection risk prediction score were higher than those of the TNM classification (risk prediction score vs. TNM: internal 0.75 vs. 0.60; and external 0.79 vs. 0.78) (Supplementary Table S4). The calibration plots at 5 years for the development and external-validation cohorts indicate moderate calibration (Figure 2). Brier scores (Supplementary Table S3) and decision curves (Supplementary Figure S4) are presented in the Supplementary Materials.

Distribution of patients in each cohort by type of procedure and risk of recurrence

The distribution of patients who underwent sublobar resection was highest in the MSK cohort (39%), followed by the Japan (34%), Taiwan (19%), and Germany (12%) cohorts (Supplementary Figure S5A).

The distribution of patients categorized by probability of 5-year CIR estimated by procedure-specific risk prediction in each cohort is shown in Supplementary Figure S5. The recurrence risk-based patient distribution was different between cohorts especially in patients who were categorized into the lowest risk category. The Japan and Taiwan cohorts were composed more heavily of patients in the lowest risk category (lobectomy, 59% and 47%; and sublobar resection, 76% and 52%) than the composition of the MSK and Germany cohorts (lobectomy, 29 % and 26%; and sublobar resection, 37% and 29%).

Exploratory OS analysis by high/low risk of recurrence

OS curves in the development and external validation cohorts, by high/low risk of recurrence (based on risk prediction score), are shown in Supplementary Figure S6. In the development cohort, patients with a high risk of recurrence had worse OS than those with lower risk, for both procedures (5-year OS [95% CI], low risk vs. high risk: lobectomy, vs. 91% [89–94] vs. 79% [73–86], p<0.001; and sublobar resection, 81% [76–87] vs. 58% [50–68], p<0.001). Similarly in the external-validation cohort, patients with a high risk of recurrence had worse OS than those with lower risk, for both procedures (5-year OS [95% CI], low risk vs. high risk: lobectomy, 90% [87–94] vs. 88% [82–94], p=0.008; and sublobar resection, 90% [84–96] vs. 62% [46–73], p<0.001).

Lobectomy risk prediction score for predicting benefit from ACT using stage II-III propensity score-matched cohort

The 1:1 matching for ACT versus no ACT resulted in 158 matched pairs (n=316), with balanced covariates between ACT groups (Table 3). Analysis of lung cancer–specific CID curves revealed a significant survival benefit for ACT only in the high score group (p=0.003) (Figure 3A). A similar pattern was observed for recurrence and OS (Figure 3B, 3C). In a multivariable competing risks regression analysis for lung cancer–specific death, a significant interaction was observed between high/low risk score and ACT (p=0.026), after adjustment for stage and STAS. Specifically, for high-risk patients, ACT significantly reduced the risk of lung cancer–specific death (SHR 0.43, 95% CI 0.26–0.72, p=0.001) (Supplementary Table S5). No significant benefit of ACT was observed among low-risk patients.

Table 3:

Clinicopathological demographic characteristics in the stage II-III cohort and difference between adjuvant chemotherapy versus no adjuvant chemotherapy before/after propensity-score matching

		Before matching			After matching
Variables		Adjuvant (−) n=312	Adjuvant (+) n=277	ASMDa	Adjuvant (−) n=158	Adjuvant (+) n=158	ASMDa
Background
Age		71 (63, 78)	66 (59, 72)	0.527	68 (62, 74)	68 (63, 74)	0.046
Sex	female	188 (60)	181 (65)	0.105	99 (63)	103 (65)	0.053
	male	124 (40)	96 (35)		59 (37)	55 (35)
Smoking	never	49 (16)	51 (18)	0.076	25 (16)	24 (15)	0.024
	former	222 (71)	193 (70)		113 (72)	113 (72)
	current	41 (13)	33 (12)		20 (13)	21 (13)
COPD		54 (17)	46 (17)	0.019	25 (16)	24 (15)	0.017
CVD		58 (19)	33 (12)	0.187	23 (15)	26 (16)	0.052
Other comorbidityb		97 (31)	81 (29)	0.040	43 (27)	49 (31)	0.084
Cr (mg/dl)		1–0 (0.9, 1.2)	1.0 (0.8, 1.1)	0.177	1.0 (0.8, 1.1)	1.0 (0.9, 1.2)	0.089
FEV1 (%)		87 (74, 99)	88 (72, 101)	0.031	87 (73, 100)	85 (70, 99)	0.079
DLCO (%)		76 (63, 90)	81 (69, 95)	0.238	80 (64, 95)	78 (67, 90)	0.020
Radiology
SUVmax		6.8 (3.7, 11.0)	7.0 (3.5, 9.9)	0.084	6.2 (3.3, 10.8)	5.8 (3.0, 10.2)	0.002
Surgery
Resection type	Pneumonectomy	15 (5)	11 (4)	0.053	9 (6)	9 (6)	0.050
	Bilobectomy	8 (3)	6 (2)		3 (2)	4 (3)
	Lobectomy	271 (87)	245 (88)		136 (86)	136 (86)
	Segmentectomy	18 (6)	15 (5)		10 (6)	9 (6)
Approach	MIS	57 (18)	86 (31)	0.342	33 (21)	29 (18)	0.064
	Thoracotomyc	260 (83)	191 (69)		125 (79)	129 (82)
Pathology
Tumor size (cm)		4.1 (2.5, 5.0)	2.8 (2.0, 4.1)	0.433	2.7 (2.0, 4.5)	2.8 (1.9, 4.5)	0.044
pN	0	131 (42)	37 (13)	0.700	35 (22)	33 (21)	0.041
	1	108 (35)	119 (43)		68 (43)	71 (45)
	2	73 (23)	121 (44)		55 (35)	54 (34)
p-Stage	II	196 (63)	139 (50)	0.257	88 (56)	91 (58)	0.038
	III	116 (37)	138 (50)		70 (44)	67 (42)
LVI		182 (58)	220 (79)	0.468	108 (68)	112 (71)	0.055
Pleural invasion	0	230 (74)	180 (65)	0.213	112 (71)	112 (71)	0.083
	1 or 2	73 (23)	91 (33)		41 (26)	43 (27)
	3	9 (3)	6 (2)		5 (3)	3 (2)
Necrosis		112 (36)	55 (20)	0.377	45 (28)	42 (27)	0.043
STAS		144 (46)	183 (72)	0.412	104 (66)	108 (68)	0.054
Histologic graded	Low	14 (4)	6 (2)	0.142	4 (3)	5 (3)	0.043
	Intermediate	153 (49)	147 (53)		81 (51)	79 (50)
	High	145 (46)	124 (45)		73 (46)	74 (47)

Data are shown as number (%) or median (25^th, 75^th percentile).

^aASMD (absolute standardized mean difference) <0.1 indicates balance in the covariate between the two groups. Bold represents ASMD ≥0.1.

^bComorbidities other than COPD or CVD included in Charlson comorbidity index evaluation.

^cIncluded conversion from MIS to thoracotomy.

^dHistologic grade based on predominant subtypes (low grade, lepidic; intermediate, acinar or papillary; high, micropapillary, solid, invasive mucinous, or colloid).

ASMD, absolute standardized mean difference; COPD, chronic obstructive pulmonary disease; Cr, serum creatinine; CVD, cardiovascular disease; DLCO, diffusion capacity of the lung for carbon monoxide; FEV1, forced expiratory volume in one second; LVI, lymphovascular invasion; MIS, minimally invasive surgery; pN, pathological lymph node metastasis status; PS, propensity score; p-Stage, pathological stage; STAS, spread through air spaces; SUVmax, maximum standardized uptake value.

Figure 3: Lung cancer–specific cumulative incidence of death (CID) (A), cumulative incidence of recurrence (CIR) (B), and overall survival (C) in the propensity score–matched stage II-III cohort -a comparison between no adjuvant chemotherapy (ACT) and ACT by lobectomy algorithm-derived risk score.

Lung cancer–specific CID curves in four groups are shown: the black solid line represents low algorithm-derived risk score without ACT; black dashed line, low score with ACT; red solid line, high score without ACT; and red dashed line, high score with ACT. In the comparison between high score/ACT and high score/no ACT (red dashed vs. solid lines), lung cancer–specific CID is significantly lower for ACT than for no ACT (p=0.003), whereas in the comparison between low score/ACT and low score/no ACT (black dashed vs. solid lines), no significant difference was observed between the two groups. High or low score categories were determined by whether the score was above or below the median. A similar pattern was observed for CIR and OS: in high score, lower CIR and higher OS for ACT than for no ACT (p=0.005 and p=0.002, respectively), whereas in low score, no difference in CIR and OS between ACT and no ACT.

In the propensity score–matched stage II-III cohort, 14% of patients did not undergo lobectomy (included were pneumonectomy, bilobectomy, and segmentectomy). Survival analyses for only patients who underwent lobectomy (N=272) are provided in Supplementary Figure S7. Similar to the results for all matched patients (Figure 3), a significant survival benefit for ACT was observed only in the high-score group.

DISCUSSION

We developed a clinical tool to predict recurrence in patients with resected small lung ADC. Our risk prediction algorithms are distinct from previously designed algorithms for the following reasons: (1) we performed a multivariable analysis including recently recognized prognostic markers such as MIP, SOL, and STAS; (2) we created surgical procedure–specific risk prediction for lobectomy and sublobar resection; (3) we selected recurrence as an outcome and performed competing risks analyses; (4) we performed external validation using a data set comprised of three separate international cohorts, confirming the applicability of the scores to patients with varying ethnic backgrounds, histological subtypes, and clinical presentations; (5) we applied the lobectomy risk prediction algorithm to a propensity score–matched cohort of patients—in which all clinical, radiological, surgical, and pathological variables known to contribute to patient lung cancer and noncancer survival were matched—and demonstrated a survival benefit for platinum-based ACT only in high-risk patients (according to lobectomy risk prediction score), and not in low-risk patients, which suggests our lobectomy risk prediction score has the potential to predict benefit from ACT.

The use of cohorts with information available on prognostic histological variables enabled a comprehensive analysis and facilitated the development of scores with high accuracy. The C-index was higher than those for previously designed lung cancer risk scores (Supplementary Table S4). One reason for these high C-indices is that we included several recently validated—and IASLC- and WHO-recommended—histopathological factors that can influence outcomes following lobectomy or sublobar resection; other studies have included only tumor size as a prognostic variable.

Prognostic stratification based on the new lung ADC classification has been reported to identify high-risk patients with SOL- or MIP-predominant tumors.^{22, 23, 28} In addition to the associated worse outcomes following surgery, Tsao et al. recently demonstrated that high-grade (MIP and SOL) predominant subtypes can help predict patients’ benefit from platinum-based ACT.⁴³ On the basis of our and others’ data,^{21–23, 26, 27, 29, 30} we focused on the continuous percentage of the high-grade subtypes (MIP and SOL) and the presence of STAS, LVI, and necrosis to develop scores to predict risk of recurrence. The subsequent analysis revealed that the proposed lobectomy risk score has the potential to predict benefit from ACT.

As recurrence is strongly correlated with lung cancer–specific death (Supplementary Figure S8 and Table S6), we used recurrence as our primary outcome. In early-stage disease, particularly among elderly patients, OS, recurrence-free survival are affected by competing events, which underlines the need to perform competing risks analysis.^{44, 45} In this study, death without recurrence was treated as a competing event. Competing risks analysis addressed the occurrence of noncancer-related events when predicting cancer-specific outcomes to improve predictions.

One limitation of our study is that the external cohorts had relatively fewer patients than the development cohort. Nevertheless, the inclusion of cohorts from Eastern and Western continents (with inherent variability in demographic factors and histological subtypes), the proportions of patients who underwent lobectomy and sublobar resection, and the use of multiple pathologists to confirm pathological interpretations add strength to our approach. There is significant variation in the prevalence of certain pathological features—such as LVI, pleural invasion, and the presence of MIP—in different cohorts. This could be due to inherent differences in ethnic backgrounds but could also be secondary to interobserver variation among pathologists. Although all known histological variables were utilized in the risk prediction development, due to lack of availability of imaging and molecular tumor analysis from patients who underwent resection in earlier time period, we did not include them in our analysis. These algorithms do not provide assistance in selecting patients preoperatively for lobectomy versus sublobar resection. Future studies that improve the diagnostic accuracy of histological subtypes during intraoperative frozen section may extend the use of our risk prediction scores intraoperatively. Another limitation is that we did not include pathological lymph node evaluation in this analysis, and this might have affected outcomes (Supplementary Method and Table S7).

Both the frequency of ACT use and the reasons for not administering ACT for patients with stage II-III disease differed by year of surgery, especially before and after 2004 (Supplementary Table S8). Since this change in treatment strategy might have affected our results, we performed the same prognostic analysis in patients who underwent surgery between 2004 and 2013 (n=215; Supplementary Figure S9): the difference in survival between the ACT and the no-ACT groups was similar to that of the entire cohort.

In conclusion, we have developed and validated the two procedure-specific risk prediction algorithms. Given that 80% of patients with stage IA lung ADC survive for at least 5 years following surgical resection,¹ it is imperative that postoperative interventions intended to improve outcomes identify patients at high risk of recurrence and that optimal treatments are investigated. Despite the success of curative-intent surgery, high rates of recurrence are seen in a subgroup of early-stage patients, particularly those with MIP, SOL, STAS, LVI, and necrosis. The two procedure-specific risk prediction algorithms can be used postoperatively to provide patients with more-accurate prognostic information that is based on their individual clinicopathological status and can also assist in stratifying patients for further interventions. In particular, the lobectomy risk prediction algorithm—with its demonstrated ability to predict benefit from platinum-based ACT in stage II-III ADC—has the potential to be used for patient selection in future trials of ACT for stage I lung ADC.

Abbreviation list:

ACT

adjuvant chemotherapy

ADC

adenocarcinoma

ASMD

absolute standardized mean difference

CI

confidence interval

C-index

concordance index

CID

cumulative incidence of death

CIR

cumulative incidence of recurrence

IASLC/ATS/ERS

International Association for the Study of Lung Cancer, American Thoracic Society, and European Respiratory Society

LVI

lymphovascular invasion

MIP

micropapillary

MSK

Memorial Sloan Kettering Cancer Center

NSCLC

non-small cell lung cancer

OS

overall survival

SOL

solid

STAS

spread through air spaces

TNM

tumor, node, metastasis

WHO

World Health Organization