Pretreatment prediction of early severe treatment-related toxicity in older patients with lung adenocarcinoma: same-center temporal validation of a risk model

Abstract

Background

Older adults with lung adenocarcinoma have heterogeneous tolerance of chemotherapy-based treatment. Early severe toxicity can disrupt planned therapy, require unplanned admission, and limit treatment intensity. We developed a pretreatment risk model for severe treatment-related toxicity within the first two chemotherapy-based cycles and examined whether Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight (FRAIL) category and inflammatory-nutritional markers add risk-stratification information.

Methods

This single-center retrospective cohort included patients aged ≥60 years with pathologically confirmed lung adenocarcinoma who received cytotoxic chemotherapy-based systemic treatment between September 2015 and September 2025. Among 1,106 screened patients, 1,019 formed the primary complete-case cohort; a multiple-imputation sensitivity analysis included 1,045 patients with eligible treatment exposure and adjudicable outcomes. Patients treated from September 2015 to December 2021 formed the development cohort (n=679), and those treated from January 2022 to September 2025 formed the same-center temporal validation cohort (n=340). The endpoint was Common Terminology Criteria for Adverse Events (CTCAE) version 5.0 grade ≥3 treatment-related toxicity within the first two cycles. Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the neutrophil-to-lymphocyte ratio (NLR) and prognostic nutritional index (PNI). Performance was assessed using discrimination, calibration, Brier score, decision curve analysis, and risk grouping.

Results

Severe treatment-related toxicity occurred in 316 of 1,019 patients (31.0%): 202 in the development cohort and 114 in the validation cohort. Toxicity rates increased across FRAIL categories: 13.5% in robust, 28.5% in pre-frail, and 51.4% in frail patients (P for trend <0.001). In Model C, frailty, higher Eastern Cooperative Oncology Group (ECOG) performance status, tumor-node-metastasis (TNM) stage IV disease, and lower PNI indicated greater risk. Model C yielded an area under the receiver operating characteristic curve (AUC) of 0.770 [95% confidence interval (CI): 0.736–0.809] in development and 0.753 (95% CI: 0.700–0.805) in validation. In the validation cohort, the calibration intercept was 0.124 (95% CI: −0.194 to 0.442) and slope was 0.986 (95% CI: 0.706–1.267). Model C also produced ordered low-, intermediate-, and high-risk strata with increasing observed toxicity rates.

Conclusions

A pretreatment model combining clinical variables, FRAIL category, NLR, and PNI separated older patients with lung adenocarcinoma into different levels of early severe toxicity risk. The model showed moderate discrimination, acceptable calibration, and directionally consistent sensitivity analyses. It may support closer observation and supportive-care planning, but routine use should await multicenter external validation and prospective evaluation of clinical impact.

Highlight box.

Key findings

• In this cohort of 1,019 older patients with lung adenocarcinoma, 31.0% developed severe treatment-related toxicity during the first two chemotherapy-based treatment cycles.

• Adding Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight category, neutrophil-to-lymphocyte ratio, and prognostic nutritional index to routine clinical variables produced a pretreatment model with moderate discrimination and acceptable calibration when applied to a later same-center temporal validation cohort.

What is known and what is new?

• Treatment tolerance in older adults with lung cancer is heterogeneous; chronological age alone is a weak summary of early severe toxicity risk. Frailty, functional status, and nutritional reserve add clinically relevant context.

• The present work offers a lung adenocarcinoma-focused prediction approach using routinely collected variables. The coefficients and risk thresholds were fixed after model development and then carried forward unchanged to a later same-center temporal validation cohort.

What is the implication, and what should change now?

• Before starting chemotherapy-based treatment, the model can help flag patients for closer follow-up, nutrition-focused optimization, and early supportive-care planning.

• The evidence remains single-center and has only same-center temporal validation; the model should be viewed as a risk-stratification aid for further validation rather than a ready-to-implement clinical decision tool.

Introduction

Background

Lung cancer accounts for a large share of cancer-related mortality worldwide (1), and adenocarcinoma has become its predominant histological subtype (2). An aging population means that more older adults with lung adenocarcinoma are considered for systemic treatment. For these patients, tolerance is shaped not only by stage or regimen intensity, but also by functional status, comorbidity, nutritional reserve, chronic inflammation, and frailty. Current geriatric oncology guidance encourages assessment of vulnerabilities involving function, nutrition, comorbidity, cognition, and social support before systemic cancer therapy (3). A practical pretreatment tool is therefore needed to identify patients susceptible to early severe toxicity without relying on chronological age as a proxy for fitness.

Rationale and knowledge gap

Several geriatric oncology tools have informed this field. The Cancer and Aging Research Group (CARG) score links chemotherapy toxicity with age, tumor characteristics, treatment intensity, laboratory results, and geriatric assessment variables (4). The Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) also combines patient- and treatment-related factors to estimate severe chemotherapy toxicity (5). These tools are valuable, but they were mainly developed in mixed-cancer cohorts and include elements that may be inconsistently recorded in retrospective electronic medical records. The Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight (FRAIL) scale is more concise, and it fits the clinical need for rapid frailty screening (6,7). Frailty in older patients with lung cancer has also been associated with complications and adverse outcomes in a recent systematic review and meta-analysis (8).

Inflammatory and nutritional status may further characterize treatment reserve. The neutrophil-to-lymphocyte ratio (NLR) can be calculated directly from routine blood counts and is often used as a systemic inflammation marker in solid-tumor research (9). The prognostic nutritional index (PNI), derived from albumin and lymphocyte count (10), has prognostic relevance in lung cancer (11). Because both indices are available from standard pretreatment testing, they are suitable for retrospective risk modeling. Combining FRAIL category with NLR and PNI may capture complementary aspects of vulnerability that are not fully reflected by conventional clinical variables.

Objective

On this basis, we built a pretreatment prediction model incorporating clinical variables, FRAIL category, NLR, and PNI to estimate severe treatment-related toxicity during the first two cycles of cytotoxic chemotherapy-based treatment in older patients with lung adenocarcinoma. The model was fitted in an earlier treatment period and evaluated in a later same-center temporal validation cohort. The analysis was intended for prediction and risk stratification, not for causal inference. We present this article in accordance with the TRIPOD+AI and STROBE reporting checklists (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1930/rc) (12-14).

Methods

Study design and study population

We assembled this single-center retrospective cohort from electronic medical records, nursing notes, laboratory reports, and treatment orders recorded at The First Affiliated Hospital of Bengbu Medical University, Bengbu, Anhui, China from September 2015 to September 2025. Patients were eligible if they were at least 60 years old, had pathologically confirmed lung adenocarcinoma, received cytotoxic chemotherapy-based systemic treatment, and had pretreatment clinical and laboratory assessment available for attempted baseline predictor ascertainment; individual predictor values could be missing and were handled as described below. The age threshold of ≥60 years was used because Chinese law defines older persons as citizens aged 60 years and above and because the World Health Organization/United Nations Decade of Healthy Ageing also uses age 60 years and over to define older people (15,16). This threshold was therefore appropriate for a China-based real-world cohort and allowed evaluation of earlier vulnerability, while age was still modeled continuously per 5-year increase. We acknowledge that many geriatric-oncology guidelines use ≥65 years for geriatric assessment, which may affect comparisons with studies using ≥65 or ≥70 years (3). Predefined study-level exclusion criteria were synchronous targeted therapy; end-of-life care, palliative sedation, or a clinical condition preventing observation of the first two treatment cycles; a non-adjudicable primary outcome; or treatment and follow-up mainly outside the study hospital. Missing predictor values alone did not constitute a study-level clinical exclusion; instead, they determined inclusion in the primary complete-case analysis or the multiple imputation by chained equations (MICE) sensitivity analysis.

Before analysis, the first date of chemotherapy-based treatment was used to define the temporal split. Patients treated from September 2015 through December 2021 formed the development cohort, whereas those treated from January 2022 through September 2025 formed the same-center temporal validation cohort. Among 1,106 screened patients, 50 met predefined clinical exclusion criteria (42 receiving synchronous targeted therapy and 8 receiving end-of-life care or palliative sedation before chemotherapy), 11 had primary outcomes that could not be adjudicated, and 26 had adjudicable outcomes but one or more missing predictor values. Thus, 1,019 patients comprised the primary complete-case analytic cohort (679 development and 340 validation). The 26 patients with adjudicable outcomes and missing predictors were additionally retained for the MICE sensitivity analysis, yielding 1,045 patients; the 11 patients with non-adjudicable outcomes were not imputed. Missingness before complete-case selection is summarized in Table S1. The original development-vs.-validation cohort assignment based on treatment period was preserved in every imputed dataset. Because predictor missingness among otherwise eligible patients was limited, the complete-case cohort was retained as the primary analysis for model transparency and reproducibility, whereas MICE was used as a sensitivity analysis to evaluate the potential impact of missing predictor data.

The endpoint was severe treatment-related toxicity during the first two chemotherapy-based treatment cycles. Severe toxicity was defined as any grade ≥3 adverse event by the National Cancer Institute Common Terminology Criteria for Adverse Events (CTCAE), version 5.0 (17). Follow-up began on cycle 1 day 1 and ended immediately before the planned start of cycle 3 according to the regimen-specific treatment schedule. Because cycle length varied across regimens, a single fixed number of days was not imposed. If cycle 3 was delayed, omitted, or treatment was discontinued, toxicity adjudication extended through the corresponding planned cycle-3 start point defined by the original regimen schedule, using the last available toxicity assessment within that two-cycle window. Events were counted only when the treating physician considered them related to the chemotherapy-based systemic treatment. For immuno-chemotherapy combination regimens, treatment-related toxicity included events attributed to either the cytotoxic chemotherapy component or the concomitant immunotherapy component. Events clearly attributable to tumor progression, non-treatment-related infection, trauma, or worsening of a pre-existing condition were not included. Two trained investigators independently reviewed the records and assigned CTCAE version 5.0 grades; disagreements were resolved by a senior physician. For descriptive characterization of the composite endpoint, grade ≥3 treatment-related toxicities were further grouped into hematologic, gastrointestinal, infectious, hepatic/renal/metabolic, immune-related, and other categories. A primary toxicity type was assigned to each event patient, whereas any-type toxicity indicators allowed more than one category per patient. Patients with more than one grade ≥3 toxicity category were flagged as having multiple toxicity categories. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Bengbu Medical University {approval No. Lun Ke Pi Zi [2024] No. 004}, and the requirement for individual informed consent was waived by the Committee because this was a retrospective analysis of de-identified records.

Variable collection and definitions

Baseline information was extracted from electronic records before treatment initiation. Variables included age, sex, body mass index (BMI), Eastern Cooperative Oncology Group (ECOG) performance status, tumor-node-metastasis (TNM) stage, Charlson Comorbidity Index, treatment regimen, FRAIL score and category, complete blood count, and biochemical markers. For analytic consistency across the 2015–2025 study period, TNM stage was harmonized to the eighth edition of the internationally adopted lung cancer TNM staging system (18); where source records had been documented under a different staging edition, available clinical, imaging, and pathologic information was retrospectively mapped to the eighth-edition stage grouping before analysis. The analytic cohort included patients receiving chemotherapy-based systemic treatment for advanced disease as well as patients with earlier-stage disease who subsequently required systemic treatment, including in the setting of postoperative recurrence. Accordingly, TNM stage I–IV categories were retained as a prespecified tumor-stage covariate rather than used as an eligibility restriction. Laboratory results were taken from tests within 7 days before the first treatment; when more than one value was available, the closest pretreatment result was used. NLR was calculated as neutrophil count divided by lymphocyte count (9). PNI was calculated as albumin (g/L) + 5 × lymphocyte count (10⁹/L), after unit conversion consistent with the original Onodera formula (10). The platelet-to-lymphocyte ratio (PLR) was retained as an alternative inflammatory marker in sensitivity analysis (19), and the systemic immune-inflammation index (SII) was used to represent systemic inflammation involving platelets, neutrophils, and lymphocytes (20).

The treatment exposure of interest was cytotoxic chemotherapy-based systemic treatment. For the first two cycles, regimens were classified according to treatment intensity, clinical use, and sample-size distribution as single-agent chemotherapy, doublet chemotherapy, or immuno-chemotherapy combination. Single-agent chemotherapy was the reference category. Doublet chemotherapy represented a higher-intensity cytotoxic regimen. Immuno-chemotherapy combination was retained as a separate group because immune-related adverse events may contribute to the toxicity profile, although treatment remained chemotherapy-based. Patients receiving synchronous targeted therapy were excluded. These regimen groups were included to adjust risk prediction for treatment pattern rather than to compare individual drug toxicities.

ECOG performance status was used as the measure of pretreatment function and entered as an ordinal continuous variable; the odds ratio (OR) therefore describes the change in toxicity risk per 1-point increase (21). Comorbidity burden was measured with the Charlson Comorbidity Index and modeled as an ordinal continuous variable (22). Frailty was assessed with the FRAIL scale, which covers fatigue, resistance, ambulation, illnesses, and weight loss (6). FRAIL components were retrospectively reconstructed from structured nursing assessment, admission history, performance-status documentation, weight records, and physician chart review before treatment initiation. When a component could not be reliably reconstructed, the FRAIL score was treated as missing. Total scores categorized patients as robust (0 points), pre-frail (1–2 points), or frail (≥3 points). ECOG performance status, Charlson Comorbidity Index, and FRAIL category were all retained because they describe related but distinct dimensions of pretreatment reserve.

Predictor selection, model development, and validation

Because this retrospective cohort included all consecutive eligible patients treated during the prespecified study period, no formal a priori sample-size calculation was performed. The development cohort contained 679 patients with 202 severe-toxicity events and was used for model fitting. Predictors were specified before model fitting, guided by clinical relevance, geriatric oncology evidence, availability in routine records, and the number of outcome events. We avoided screening predictors by univariable P value, because data-driven selection can weaken interpretability and introduce bias in prediction modeling (12,13,23). Model A was the clinical base model and included age, sex, BMI, ECOG performance status, TNM stage, Charlson Comorbidity Index, and treatment regimen. These variables represented demographic factors, functional status, tumor burden, comorbidity, and treatment intensity. Variables used mainly for description, as well as variables overlapping strongly with the components of NLR or PNI, were not used in the primary model. Model B added FRAIL category to Model A. Model C further added NLR and PNI to Model B and was defined in advance as the final primary model. The full Model C equation and coefficients are shown in Table S2.

Age entered the model per 5-year increase, and BMI per 1 kg/m² increase. TNM stage, treatment regimen, and FRAIL category were treated as categorical predictors. NLR and PNI were standardized with the mean and standard deviation (SD) from the development cohort, and the same standardization parameters were used for the validation cohort. All models were fitted only in the development cohort. After the coefficients, NLR/PNI standardization parameters, and Model C risk thresholds were set, they were applied unchanged to the same-center temporal validation cohort. In the validation data, variables were not reselected, the model was not refitted, and risk thresholds were not recalibrated. In the development cohort, Model C predicted probabilities were split into tertiles to form low-, intermediate-, and high-risk groups; identical cut points were applied in the validation cohort. These groups describe observed risk gradients and are not validated intervention thresholds.

Statistical analysis

Statistical analyses and figure generation were performed in Python 3.11, using statsmodels for regression and MICE procedures, scikit-learn for receiver operating characteristic (ROC) analyses and related performance metrics, and Matplotlib for figure generation. Continuous variables are reported as mean ± SD or median (interquartile range), according to distribution; categorical variables are reported as counts and percentages. For two-group comparisons between the development and validation cohorts (Table 1), continuous variables were compared using independent-samples t tests, and categorical variables were compared using Pearson Chi-squared tests; Yates continuity correction was applied to 2×2 tables. For comparisons across the three FRAIL categories (Table 2), continuous variables were compared using the Kruskal-Wallis test and categorical variables using Pearson Chi-squared tests. For Table 3, overall associations between FRAIL category and grade ≥3 treatment-related toxicity or highest toxicity grade were evaluated using Pearson Chi-squared tests. Ordered trends across FRAIL categories were evaluated using the Cochran-Armitage trend test for the binary grade ≥3 endpoint and a linear-by-linear association test for the ordered highest-toxicity-grade categories. Baseline differences between the development and validation cohorts were also described using absolute standardized mean differences. Model A, Model B, and Model C were estimated with multivariable logistic regression in the development cohort. Model A included age, sex, BMI, ECOG performance status, TNM stage, Charlson Comorbidity Index, and treatment regimen. FRAIL category was added in Model B. Standardized NLR and PNI were then added in Model C, the prespecified final model.

Table 1. Baseline characteristics in the development and same-center temporal validation cohorts.

Variable	Overall	Development cohort	Validation cohort	P value	Absolute SMD
Age, years	69.87±5.20	69.82±5.32	69.96±4.98	0.69	0.026
BMI, kg/m2	22.69±2.85	22.79±2.87	22.48±2.79	0.11	0.106
Charlson Comorbidity Index	0.74±0.87	0.75±0.86	0.74±0.88	0.86	0.011
ECOG performance status	0.77±0.80	0.77±0.78	0.77±0.85	0.96	0.003
Gender
Male	701 (68.8)	477 (70.3)	224 (65.9)	0.18	0.094
Female	318 (31.2)	202 (29.7)	116 (34.1)	–	0.094
TNM
Stage I	163 (16.0)	124 (18.3)	39 (11.5)	0.04	0.186
Stage II	98 (9.6)	66 (9.7)	32 (9.4)	–	0.010
Stage III	173 (17.0)	110 (16.2)	63 (18.5)	–	0.062
Stage IV	585 (57.4)	379 (55.8)	206 (60.6)	–	0.096
Treatment regimen
Single-agent chemotherapy	534 (52.4)	353 (52.0)	181 (53.2)	0.64	0.025
Doublet chemotherapy	335 (32.9)	221 (32.5)	114 (33.5)	–	0.021
Immuno-chemotherapy combination	150 (14.7)	105 (15.5)	45 (13.2)	–	0.063
FRAIL classification
Robust	319 (31.3)	219 (32.3)	100 (29.4)	0.65	0.061
Pre-frail	379 (37.2)	249 (36.7)	130 (38.2)	–	0.032
Frail	321 (31.5)	211 (31.1)	110 (32.4)	–	0.027

Data are shown as mean ± standard deviation or n (%), as appropriate. P values compare the development cohort with the same-center temporal validation cohort. Continuous variables were compared using independent samples t-tests. Categorical variables were compared using Pearson Chi-squared tests, with Yates continuity correction for 2×2 tables. For categorical variables, the P value is provided on the first row of the variable. BMI, body mass index; ECOG, Eastern Cooperative Oncology Group; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; SMD, standardized mean difference; TNM, tumor-node-metastasis.

Table 2. Baseline characteristics and inflammatory-nutritional markers by FRAIL category.

Variable	Overall	Robust	Pre-frail	Frail	P value
Age, years	71.0 (66.0–73.0)	68.0 (64.0–71.0)	70.0 (66.0–73.0)	72.0 (70.0–75.0)	<0.001
BMI, kg/m2	22.5 (20.8–24.5)	24.2 (22.8–25.9)	22.6 (20.6–24.6)	20.8 (19.4–22.4)	<0.001
NLR	2.9 (2.2–3.5)	2.6 (2.1–3.3)	2.9 (2.2–3.4)	3.1 (2.5–3.7)	<0.001
PNI	48.6 (44.7–52.4)	51.0 (46.4–54.4)	48.5 (44.8–52.7)	47.0 (43.2–50.1)	<0.001
CRP, mg/L	6.5 (2.5–15.7)	3.6 (1.7–11.1)	6.0 (2.5–15.8)	9.5 (5.1–15.8)	<0.001
PLR	157.2 (120.7–214.1)	149.0 (116.9–197.6)	162.0 (127.2–217.6)	162.5 (122.1–223.1)	0.02
SII	664.0 (488.3–894.1)	586.8 (448.4–783.0)	643.9 (493.3–860.7)	784.2 (569.9–1,031.0)	<0.001
Albumin, g/L	41.0 (37.7–43.9)	43.1 (40.0–45.5)	41.4 (38.1–43.9)	39.1 (36.4–41.5)	<0.001
Hemoglobin, g/L	123.0 (112.1–135.0)	127.1 (116.3–136.0)	121.0 (111.6–132.9)	121.0 (110.4–135.5)	<0.001
Platelet count, ×109/L	243.8 (199.0–292.0)	228.7 (192.6–282.1)	243.0 (200.9–289.5)	261.4 (206.2–299.1)	<0.001
Charlson Comorbidity Index 0	500 (49.1)	198 (62.1)	190 (50.1)	112 (34.9)	<0.001
Charlson Comorbidity Index 1	326 (32.0)	121 (37.9)	95 (25.1)	110 (34.3)	–
Charlson Comorbidity Index 2	146 (14.3)	0 (0.0)	73 (19.3)	73 (22.7)	–
Charlson Comorbidity Index ≥3	47 (4.6)	0 (0.0)	21 (5.5)	26 (8.1)	–
ECOG performance status 0	445 (43.7)	233 (73.0)	137 (36.1)	75 (23.4)	<0.001
ECOG performance status 1	386 (37.9)	67 (21.0)	182 (48.0)	137 (42.7)	–
ECOG performance status 2	163 (16.0)	16 (5.0)	54 (14.2)	93 (29.0)	–
ECOG performance status ≥3	25 (2.5)	3 (0.9)	6 (1.6)	16 (5.0)	–
TNM stage I	163 (16.0)	67 (21.0)	63 (16.6)	33 (10.3)	0.004
TNM stage II	98 (9.6)	30 (9.4)	32 (8.4)	36 (11.2)	–
TNM stage III	173 (17.0)	40 (12.5)	74 (19.5)	59 (18.4)	–
TNM stage IV	585 (57.4)	182 (57.1)	210 (55.4)	193 (60.1)	–
Single-agent chemotherapy	534 (52.4)	177 (55.5)	208 (54.9)	149 (46.4)	0.13
Doublet chemotherapy	335 (32.9)	101 (31.7)	117 (30.9)	117 (36.4)	–
Immuno-chemotherapy combination	150 (14.7)	41 (12.9)	54 (14.2)	55 (17.1)	–
Male	701 (68.8)	203 (63.6)	266 (70.2)	232 (72.3)	0.047
Female	318 (31.2)	116 (36.4)	113 (29.8)	89 (27.7)	–

Continuous variables are expressed as median (interquartile range), and categorical variables as n (%). P values compare FRAIL categories; continuous variables were compared using the Kruskal-Wallis test and categorical variables using Pearson Chi-squared tests. For categorical variables, the P value appears on the first row of the variable. BMI, body mass index; CRP, C-reactive protein; ECOG, Eastern Cooperative Oncology Group; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; PNI, prognostic nutritional index; SII, systemic immune-inflammation index; TNM, tumor-node-metastasis.

Table 3. Severe treatment-related toxicity by FRAIL category.

Variable	Overall	Robust	Pre-frail	Frail	P value	P for trend
Grade ≥3 treatment-related toxicity	316 (31.0)	43 (13.5)	108 (28.5)	165 (51.4)	<0.001	<0.001
Highest toxicity grade	–	–	–	–	<0.001	<0.001
Grade 0–2	703 (69.0)	276 (86.5)	271 (71.5)	156 (48.6)	–	–
Grade 3	219 (21.5)	28 (8.8)	80 (21.1)	111 (34.6)	–	–
Grade 4	97 (9.5)	15 (4.7)	28 (7.4)	54 (16.8)	–	–

Data are presented as n (%). Overall P values were obtained using Pearson Chi-squared tests. P for trend was assessed using the Cochran-Armitage trend test for grade ≥3 treatment-related toxicity and a linear-by-linear association test for the ordered highest-toxicity-grade categories. No grade 5 treatment-related toxicity occurred during the prespecified two-cycle observation window. FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight.

For the MICE sensitivity analysis, missing base-variable values were imputed using the statsmodels MICEData/MICE implementation (statsmodels 0.14.6). Continuous variables with missingness (including BMI, neutrophil count, lymphocyte count, and albumin) were updated using linear conditional models followed by predictive mean matching (PMM). Ordinal variables with missingness (including ECOG performance status and the FRAIL total score) were imputed on their observed integer support using PMM-based donor draws. The fully observed primary outcome was included as a conditioning variable to preserve associations with predictor missingness but was never itself imputed. Fully observed baseline variables, including age, sex, TNM stage, Charlson Comorbidity Index, treatment regimen, and treatment year, were included as conditioning variables where appropriate; the prespecified development-vs.-validation cohort assignment was preserved and was not imputed. After completion of each imputed dataset, derived variables were recalculated deterministically from their completed components: FRAIL category from the FRAIL total score, NLR from neutrophil and lymphocyte counts, PNI from albumin and lymphocyte count, PLR from platelet and lymphocyte counts, and SII from platelet, neutrophil, and lymphocyte counts. A 10-cycle burn-in preceded 20 retained imputations, and the default MICE skip interval (n_skip =3) was used between successive retained analysis samples. Convergence was assessed from the stability of imputed-variable summary statistics and chained-equation history/trace diagnostics across iterations. For comparability with the primary analysis, NLR and PNI in all imputed datasets were standardized using the locked complete-case development-cohort constants (NLR mean 2.9091, SD 0.9628; PNI mean 48.8724, SD 4.9444), rather than re-estimating standardization parameters within each imputed dataset. In each imputed dataset, Model C was refitted in the development cohort and evaluated in the corresponding same-center temporal validation cohort. Regression coefficients and standard errors were combined across datasets using Rubin rules, whereas AUC, Brier score, and calibration metrics were summarized descriptively as the mean and range across the 20 imputed datasets.

The fitted regression coefficients, NLR/PNI standardization parameters, and Model C risk-stratification cut points were transferred to the same-center temporal validation cohort without modification. Discrimination was expressed as the area under the receiver operating characteristic curve (AUC) with 95% confidence interval (CI). AUC 95% CIs were estimated by non-parametric bootstrap resampling within each cohort using 1,000 resamples. Overall prediction error was measured by the Brier score (24). Calibration was assessed with plots and by estimating the calibration intercept and slope from the locked development-model linear predictor (25); 95% CIs for the calibration intercept and slope were Wald intervals derived from the corresponding logistic calibration regressions. Decision curve analysis quantified net benefit across threshold probabilities from 0.05 to 0.50 (26,27). Low-, intermediate-, and high-risk groups were based on tertiles of Model C predicted probability in the development cohort and were then applied unchanged to the validation cohort.

Additional details on missing data handling, multiple-imputation procedures, treatment-era and regimen sensitivity analyses, and toxicity-spectrum classification are provided in Appendix 1. Additional supplementary analyses were performed to address model transparency, treatment heterogeneity, endpoint composition, missingness, and internal optimism. First, model formulas, variable coding, standardization constants, and Model C risk cut-offs were summarized for reproducibility (Table S2). Second, the incremental value of NLR was evaluated by comparing Model B, Model B + PNI, Model B + NLR, and Model C using Akaike information criterion (AIC), a development-cohort model-fit measure that penalizes model complexity, likelihood ratio tests, AUC, Brier score, and calibration slope (Table S3). Third, treatment heterogeneity was examined by summarizing treatment-era and regimen distributions, evaluating Model C performance after excluding immuno-chemotherapy patients, describing observed toxicity rates across regimen categories, and adding treatment year as an additional adjustment term (Tables S4,S5). Fourth, the toxicity spectrum of the composite endpoint was summarized descriptively (Table S6). Fifth, a MICE sensitivity analysis was performed in the eligible missing-data sensitivity-analysis cohort. Model C performance was summarized across 20 completed datasets, and pooled coefficient estimates were summarized using Rubin rules (Table S7).

Bootstrap optimism correction was performed in the development cohort using 1,000 resamples. In each resample, final Model C was refitted and evaluated in both the bootstrap sample and the original development cohort. Mean optimism was then used to obtain optimism-corrected development performance estimates. Smoothed calibration plots for final Model C were generated using decile-group observed risks and a non-parametric smoother for the development and same-center temporal validation cohorts. Detailed model-wise grouped calibration plots for Models A–C are provided in Figure S1. Bootstrap optimism-corrected performance estimates are shown in Table S8.

All tests were two-sided, and P<0.05 was considered statistically significant.

Results

Patient screening and incidence of severe treatment-related toxicity

A total of 1,106 older patients with lung adenocarcinoma were screened. Fifty met predefined clinical exclusion criteria (42 receiving synchronous targeted therapy and 8 receiving end-of-life care or palliative sedation before chemotherapy), 11 had non-adjudicable primary outcomes, and 26 had adjudicable outcomes but missing predictor values. The primary complete-case analytic cohort therefore included 1,019 patients, with 679 in the development cohort and 340 in the same-center temporal validation cohort. The 26 patients with missing predictors were retained only for the MICE sensitivity analysis, yielding a MICE cohort of 1,045 patients. Severe treatment-related toxicity occurred in 316 patients in the complete-case cohort (31.0%): 202 events in the development cohort (29.8%) and 114 in the validation cohort (33.5%). Figure 1 shows the selection process.

Figure 1.

Study flowchart. Patient screening, exclusions, MICE eligibility, and allocation to the development and same-center temporal validation cohorts. Severe toxicity was defined as CTCAE version 5.0 grade ≥3 treatment-related toxicity within the first two chemotherapy-based cycles. CTCAE, Common Terminology Criteria for Adverse Events; MICE, multiple imputation by chained equations.

Comparison of baseline characteristics between the development and validation cohorts

Baseline profiles were generally similar between the two cohorts. Age, BMI, Charlson Comorbidity Index, ECOG performance status, sex, treatment regimen, and FRAIL category did not differ significantly, and most absolute standardized mean differences were small. TNM stage distribution differed (P=0.04), with more stage IV disease in the validation cohort than in the development cohort (60.6% vs. 55.8%) and fewer stage I cases (11.5% vs. 18.3%). Thus, the validation cohort contained a slightly higher proportion of advanced-stage disease, while the overall baseline pattern remained comparable (Table 1).

Baseline characteristics and inflammatory-nutritional markers according to FRAIL category

Patients in higher FRAIL categories were generally older and had lower BMI and nutrition-related values. Median age was 68, 70, and 72 years in the robust, pre-frail, and frail groups, respectively; the corresponding median BMI values were 24.2, 22.6, and 20.8 kg/m². These differences were statistically significant. Markers of inflammation increased across frailty severity: median NLR rose from 2.6 in robust patients to 3.1 in frail patients, median C-reactive protein from 3.6 to 9.5 mg/L, and median SII from 586.8 to 784.2. PNI, albumin, and hemoglobin showed the opposite pattern. Overall, the frail group had a heavier inflammatory burden and poorer nutritional-immune reserve (Table 2).

Severe treatment-related toxicity according to FRAIL category

The occurrence of severe treatment-related toxicity increased across FRAIL categories. Events were recorded in 43 robust patients (13.5%), 108 pre-frail patients (28.5%), and 165 frail patients (51.4%). The overall group comparison and the trend test were both significant (both P<0.001). Highest toxicity grade showed a similar pattern: grade 4 toxicity occurred in 4.7% of robust patients, 7.4% of pre-frail patients, and 16.8% of frail patients. No grade 5 treatment-related toxicity occurred during the prespecified two-cycle observation window. Thus, FRAIL category distinguished both the frequency of severe toxicity and the burden of higher-grade toxicity (Table 3).

Multivariable model development

Model C included all Model B variables and added standardized NLR and PNI. In this final model, higher ECOG performance status, TNM stage IV disease, and frailty indicated greater toxicity risk. A 1-point increase in ECOG performance status was associated with higher risk (OR, 1.62; 95% CI: 1.26–2.08; P<0.001). Compared with TNM stage I, stage IV disease was associated with higher risk (OR, 2.18; 95% CI: 1.23–3.87; P=0.008). Frail patients had higher risk than robust patients (OR, 3.30; 95% CI: 1.79–6.06; P<0.001). PNI showed an inverse association, with each 1-standard-deviation increase corresponding to lower risk (OR, 0.71; 95% CI: 0.56–0.89; P=0.004). NLR did not retain an independent association after adjustment (OR, 1.03; 95% CI: 0.84–1.26; P=0.80). Female sex was associated with lower risk (OR, 0.59; 95% CI: 0.38–0.91; P=0.02). Age, BMI, Charlson Comorbidity Index, and treatment regimen were not independently associated after adjustment (Table 4).

Table 4. Multivariable logistic regression for severe treatment-related toxicity in final Model C.

Variable	Adjusted OR (95% CI)	P value
Age, per 5-year increase	1.03 (0.85–1.25)	0.74
Female vs. male	0.59 (0.38–0.91)	0.02
BMI, per 1 kg/m2 increase	1.03 (0.95–1.10)	0.51
Charlson Comorbidity Index, per 1-point increase	1.14 (0.91–1.43)	0.26
ECOG performance status, per 1-point increase	1.62 (1.26–2.08)	<0.001
TNM stage II vs. stage I	1.74 (0.79–3.87)	0.17
TNM stage III vs. stage I	1.73 (0.85–3.50)	0.13
TNM stage IV vs. stage I	2.18 (1.23–3.87)	0.008
Doublet chemotherapy vs. single-agent chemotherapy	0.85 (0.56–1.28)	0.43
Immuno-chemotherapy combination vs. single-agent chemotherapy	0.60 (0.34–1.04)	0.07
Pre-frail vs. robust	1.35 (0.80–2.29)	0.27
Frail vs. robust	3.30 (1.79–6.06)	<0.001
NLR, per 1-SD increase	1.03 (0.84–1.26)	0.80
PNI, per 1-SD increase	0.71 (0.56–0.89)	0.004

ORs were estimated from the multivariable logistic regression model fitted in the development cohort. Reference categories are male sex, TNM stage I, single-agent chemotherapy, and robust FRAIL category. NLR and PNI were standardized with the development cohort mean and SD, so ORs correspond to a 1-SD increase. Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. BMI, body mass index; CI, confidence interval; ECOG, Eastern Cooperative Oncology Group; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; OR, odds ratio; PNI, prognostic nutritional index; SD, standard deviation; TNM, tumor-node-metastasis.

Same-center temporal validation and model performance

Discrimination and overall prediction error were evaluated in both cohorts. In the development cohort, the AUCs for Model A, Model B, and Model C were 0.725, 0.758, and 0.770. In the validation cohort, the corresponding values were 0.741, 0.749, and 0.753. Adding FRAIL category and then NLR/PNI increased AUC in a stepwise manner, although the gain was modest in the validation cohort. Brier scores changed from 0.184 for Model A to 0.172 for Model C in the development cohort and from 0.188 to 0.186 in the validation cohort. ROC curves are shown in Figure 2, and full performance metrics are summarized in Table 5.

Figure 2.

ROC curves for Models A–C in the development cohort (A) and same-center temporal validation cohort (B). Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. AUC, area under the receiver operating characteristic curve; CI, confidence interval; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index; ROC, receiver operating characteristic.

Table 5. Model performance.

Model (cohort)	AUC (95% CI)	Brier score	Calibration intercept (95% CI)	Calibration slope (95% CI)
A (development cohort)	0.725 (0.684–0.765)	0.184	0.000 (−0.250 to 0.250)	1.000 (0.769–1.231)
A (validation cohort)	0.741 (0.685–0.794)	0.188	0.260 (−0.091 to 0.612)	1.152 (0.818–1.486)
B (development cohort)	0.758 (0.718–0.798)	0.175	0.000 (−0.235 to 0.235)	1.000 (0.797–1.203)
B (validation cohort)	0.749 (0.692–0.802)	0.188	0.174 (−0.155 to 0.502)	1.017 (0.727–1.306)
C (development cohort)	0.770 (0.736–0.809)	0.172	0.000 (−0.231 to 0.231)	1.000 (0.803–1.197)
C (validation cohort)	0.753 (0.700–0.805)	0.186	0.124 (−0.194 to 0.442)	0.986 (0.706–1.267)

AUC (95% CIs) were estimated using 1,000 non-parametric bootstrap resamples within each cohort. Calibration intercept and slope (95% CIs) are Wald intervals from the corresponding logistic calibration regressions. Calibration intercepts and slopes of development cohort describe apparent calibration in the fitting data and should not be read as perfect external performance. Calibration metrics of validation cohort were estimated from the locked development-model linear predictor without refitting, variable reselection, or recalibration of standardization parameters or risk thresholds. Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. AUC, area under the receiver operating characteristic curve; CI, confidence interval; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index.

Calibration assessment showed acceptable agreement between predicted and observed risks. For Model C in the same-center temporal validation cohort, the calibration intercept was 0.124 (95% CI: −0.194 to 0.442) and the calibration slope was 0.986 (95% CI: 0.706 to 1.267). The development cohort intercept of 0 and slope of 1 reflect apparent calibration in the fitting dataset and should not be interpreted as external validation. Figure 3 presents the smoothed calibration plots for final Model C, and detailed model-wise grouped calibration plots for Models A–C are provided in Figure S1.

Figure 3.

Smoothed calibration plots for final Model C. Decile-group observed risks and non-parametric smoothed calibration curves are shown for the development cohort (A) and same-center temporal validation cohort (B). The dashed diagonal line indicates ideal calibration. Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index.

Clinical utility and risk stratification

Across the prespecified threshold probability range of 0.05 to 0.50, decision curve analysis showed that model-guided intensified supportive care generally provided net benefit similar to or greater than the treat-all and treat-none strategies. Model C maintained relatively stable net benefit in both cohorts, supporting its role as a pretreatment risk-stratification aid rather than an automatic treatment decision rule (Figure 4).

Figure 4.

Decision curve analysis. Net benefit of Models A–C across threshold probabilities from 0.05 to 0.50 in the development cohort (A) and same-center temporal validation cohort (B). Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index.

Risk categories were created from tertiles of Model C predicted probability in the development cohort. Observed severe toxicity rates were 8.0%, 28.3%, and 52.9% in the low-, intermediate-, and high-risk groups, respectively. Applying the same cut points to the validation cohort yielded rates of 12.0%, 25.7%, and 57.5%. The ordering persisted in patients treated later, indicating meaningful separation of risk (Figure 5).

Figure 5.

Risk stratification according to Model C. Observed toxicity rates and mean predicted probabilities are shown across low-, intermediate-, and high-risk groups in the development and same-center temporal validation cohorts. Error bars indicate 95% CIs for observed event rates. Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. CI, confidence interval; E, event; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index.

Supplementary and sensitivity analyses

Missingness before complete-case selection is summarized in Table S1. Among 1,106 screened patients, 50 met predefined exclusion criteria for which treatment regimen and primary outcome were recorded as not applicable, 11 had non-adjudicable primary outcomes and were not imputed, and 26 had adjudicable outcomes with missing predictor values and were included with the 1,019 complete cases in the 1,045-patient MICE sensitivity-analysis cohort. In the MICE sensitivity analysis, Model C showed directionally consistent performance compared with the complete-case analysis. Validation AUC decreased mildly from 0.753 to a mean of 0.743 across the 20 imputed datasets, and validation Brier score increased mildly from 0.186 to 0.193 (Table S7).

The incremental contribution of NLR is shown in Table S3. Adding PNI to Model B improved model fit, whereas adding NLR alone produced only limited improvement. After PNI had been included, the addition of NLR in Model C did not materially improve model fit or validation performance. Therefore, NLR was retained as part of the prespecified inflammatory-nutritional marker block, but its independent incremental value appeared limited after FRAIL category and PNI were considered.

Treatment-era and regimen distributions are summarized in Table S4. Immuno-chemotherapy combination was not recorded in 2015–2017 and appeared after 2018, without implying a monotonic increase across treatment eras. Treatment-heterogeneity sensitivity analyses showed no meaningful deterioration in Model C performance after excluding immuno-chemotherapy patients or after adding treatment year as an additional adjustment term (Table S5). Observed regimen-stratified toxicity rates were descriptive and were not interpreted as comparative safety estimates across treatment regimens.

The distribution of grade ≥3 treatment-related toxicity types is shown in Table S6. Among 316 event patients, hematologic toxicity was the most common primary toxicity type [186 (58.9%)], followed by gastrointestinal toxicity [68 (21.5%)] and infectious toxicity [33 (10.4%)]. Because some patients had more than one grade ≥3 toxicity category, any-type toxicity counts exceeded mutually exclusive primary-toxicity counts.

Bootstrap optimism-corrected estimates showed mild optimism in apparent development performance. The optimism-corrected development AUC was 0.747 and the optimism-corrected Brier score was 0.180, compared with an AUC of 0.753 and Brier score of 0.186 in the same-center temporal validation cohort (Table S8). Detailed model-wise grouped calibration plots for Models A–C are shown in Figure S1. Adjusted coefficients from the primary final Model C are summarized visually in Figure 6.

Figure 6.

Forest plot of final Model C. Adjusted ORs and 95% CIs are shown for predictors included in the final multivariable logistic regression model. Model A contained clinical variables, Model B added FRAIL category, and Model C additionally incorporated the NLR and PNI. BMI, body mass index; CI, confidence interval; ECOG, Eastern Cooperative Oncology Group; FRAIL, Fatigue, Resistance, Ambulation, Illnesses, and Loss of weight; NLR, neutrophil-to-lymphocyte ratio; OR, odds ratio; PNI, prognostic nutritional index; SD, standard deviation; TNM, tumor-node-metastasis.

Discussion

Key findings

In this study, we developed a pretreatment model for early severe treatment-related toxicity in older patients with lung adenocarcinoma receiving chemotherapy-based systemic treatment and evaluated the locked model in a later same-center temporal validation cohort. Grade ≥3 treatment-related toxicity occurred in 31.0% of the complete-case cohort during the first two treatment cycles. Toxicity increased markedly across FRAIL categories, from 13.5% in robust patients to 28.5% in pre-frail patients and 51.4% in frail patients. In the final model, frailty, poorer ECOG performance status, stage IV disease, and lower PNI were associated with greater toxicity risk, whereas chronological age was not independently informative. Model C retained moderate discrimination and acceptable calibration in the later cohort, and the low-, intermediate-, and high-risk groups remained clearly separated. These findings support a multidimensional assessment of treatment tolerance rather than reliance on age or regimen category alone.

Strengths and limitations

Several methodological features strengthen the study. Predictors were specified on clinical grounds rather than selected by univariable P values, and the sequential modeling strategy separated the information contributed by conventional clinical variables, frailty, and inflammatory-nutritional markers. The final coefficients, standardization constants, and risk cut points were locked after development and transferred unchanged to the later same-center temporal validation cohort. Model evaluation included discrimination, calibration, Brier score, risk-group separation, and decision-curve analysis, consistent with recommendations that prediction models should not be assessed by discrimination alone (12,13,24-27). Bootstrap optimism correction, detailed calibration assessment, multiple-imputation sensitivity analysis, treatment-era analyses, and characterization of the toxicity spectrum further evaluated potential sources of instability. These features improve transparency and internal credibility, although they do not substitute for independent external validation.

Several limitations should be acknowledged. First, this was a retrospective single-center study. The later-period cohort provides temporal validation under evolving clinical practice but does not establish geographic, institutional, or population transportability; model performance can change when case mix, measurement procedures, treatment patterns, and clinical settings differ (28,29). Multicenter external validation is therefore the most important next step. Second, FRAIL components were retrospectively reconstructed from routine documentation, and attribution of treatment-related toxicity relied on chart review and retrospective adjudication; despite predefined criteria and independent review, measurement and classification error cannot be excluded. Third, the primary analysis used complete cases. Although missingness was limited and the MICE analysis showed broadly consistent results, selection bias related to incomplete documentation remains possible. Fourth, the ≥60-year eligibility threshold reflects the clinical and administrative definition of older persons used in the study setting and allowed assessment of earlier vulnerability, whereas many geriatric-oncology studies and guidelines use ≥65 or ≥70 years; this may affect direct comparison and generalizability (3). Fifth, treatment and supportive-care practices evolved over 2015–2025. Sensitivity analyses were reassuring, but residual calendar-period and treatment-selection effects cannot be eliminated. Sixth, the severe-toxicity endpoint combined hematologic, gastrointestinal, infectious, metabolic, immune-related, and other events that may have different determinants; the study was not powered to develop stable subtype-specific models. Seventh, individual drug combinations, dose intensity, subsequent dose modifications, and detailed supportive-care interventions were not modeled separately. Finally, the risk groups were defined by development-cohort tertiles and have not been prospectively linked to specific clinical actions, and the model predicts toxicity only during the first two chemotherapy-based cycles. Independent external validation, evaluation of clinically meaningful decision thresholds, and prospective assessment of whether model-guided supportive care reduces severe toxicity, unplanned hospitalization, or treatment interruption are required before implementation.

Comparison with similar research

Our findings are consistent with the broader geriatric-oncology literature. The American Society of Clinical Oncology (ASCO) guideline recommends assessment of age-related vulnerabilities, including physical function, comorbidity, cognition, nutrition, and social support, before systemic cancer therapy because routine oncology assessment may not identify clinically relevant deficits (3). The CARG study showed that chemotherapy toxicity in older adults reflects treatment characteristics, laboratory measures, functional limitations, and geriatric-assessment variables rather than age alone (4), while the CRASH score similarly integrates patient- and treatment-related factors (5). A direct score-level comparison with CARG or CRASH was not feasible in the present retrospective dataset because several required components were not consistently available in routine records and could not be reconstructed without substantial missingness or misclassification; consequently, we did not calculate incomplete or locally modified versions of these scores. More recent work indicates that severe treatment-related toxicity prediction remains challenging: a systematic review found that available models generally showed moderate discrimination and limited external validation, while ACTTOP and contemporary validation studies illustrate continuing efforts to improve prediction across solid-tumor populations (28,30,31). A 2026 systematic review and meta-analysis also supports the risk-stratification value of the CARG framework while highlighting heterogeneity across populations and settings (32). Our model was therefore not intended to replace CARG or CRASH, but to address a narrower lung adenocarcinoma-specific question using pretreatment variables that are routinely available in clinical records.

Explanations of findings

Frailty was one of the most informative components of the model. Compared with robust patients, frail patients had substantially higher odds of severe toxicity, and the increase in toxicity from robust to pre-frail and frail categories was accompanied by a parallel increase in grade 4 events. This pattern is clinically plausible because frailty reflects reduced physiologic reserve across multiple domains rather than impairment of a single organ system. Frailty has been associated with adverse outcomes in older patients with lung cancer, and a recent systematic review and meta-analysis further supports its clinical relevance in this population (8). Importantly, FRAIL category remained informative despite the simultaneous inclusion of ECOG performance status and Charlson Comorbidity Index. These measures overlap but are not interchangeable: ECOG primarily reflects functional performance, the Charlson index quantifies chronic disease burden, and FRAIL additionally captures fatigue, mobility limitation, illness burden, and weight loss. Their combined contribution suggests that patients with similar age or ECOG status may still have materially different treatment tolerance when frailty is considered.

PNI provided a complementary dimension of vulnerability. Each 1-SD increase in PNI was associated with lower odds of severe toxicity in the final model. Because PNI combines serum albumin and lymphocyte count, it reflects aspects of nutritional status and immune reserve (10). Previous studies have linked lower PNI with adverse outcomes in lung cancer (11), although most have focused on prognosis rather than short-term treatment toxicity. Our results suggest that nutritional-immune depletion may also identify patients with reduced tolerance to early systemic therapy. This interpretation is consistent with the lower BMI, albumin, and PNI observed in patients with greater frailty. Stage IV disease was also associated with higher toxicity risk than stage I disease, potentially reflecting greater disease burden, systemic inflammation, symptoms, or reduced physiologic reserve. This coefficient should, however, be interpreted as predictive rather than causal, particularly because the cohort included patients with earlier-stage disease who subsequently required systemic treatment after recurrence. The association between female sex and lower toxicity risk likewise requires independent replication before biological interpretation.

The findings for NLR require more cautious interpretation. NLR increased across FRAIL categories and is widely used as a systemic inflammation marker with prognostic value in solid tumors (9), but standardized NLR was not independently associated with early severe toxicity after adjustment for frailty, PNI, ECOG performance status, tumor stage, and other clinical variables. The incremental analysis helps explain this result: adding PNI to Model B improved model fit, whereas adding NLR alone produced limited improvement; after PNI had been included, adding NLR did not materially improve model fit or validation performance. A biomarker may therefore be associated with vulnerability at the population level without providing substantial independent predictive information once correlated variables are included. NLR can be influenced by tumor burden, infection, medication exposure, stress responses, and lymphocyte depletion, and some of this information may already be captured by FRAIL category, PNI, ECOG status, and disease stage. NLR was retained because the inflammatory-nutritional marker block was prespecified, but the present data do not support describing it as a strong independent predictor. Whether a more parsimonious model without NLR performs similarly should be examined in independent datasets rather than determined by post hoc deletion in the current cohort.

The same-center temporal-validation results provide an important perspective on model performance. Model C achieved an AUC of 0.770 in the development cohort and 0.753 in the later same-center temporal validation cohort. More importantly, when the original coefficients and standardization parameters were applied without refitting, the validation calibration intercept was 0.124 and the calibration slope was 0.986. Bootstrap correction yielded an optimism-corrected development AUC of 0.747, close to the same-center temporal-validation AUC, suggesting only modest apparent optimism. The MICE sensitivity analysis produced a slightly lower mean validation AUC of 0.743 and a slightly higher Brier score of 0.193, but the principal coefficient estimates remained directionally consistent. Thus, the model showed reasonable within-center stability, but its discrimination should still be described as moderate rather than high. This is consistent with contemporary evidence showing that moderate discrimination is common among severe-toxicity prediction models and that satisfactory development performance does not establish transportability to other settings (28).

The sequential comparison of Models A, B, and C also clarifies where the additional predictive information arose. In the development cohort, adding FRAIL category to the conventional clinical model increased AUC from 0.725 to 0.758, while subsequent addition of NLR and PNI increased it to 0.770. In the same-center temporal validation cohort, the corresponding AUCs were 0.741, 0.749, and 0.753. The improvement was therefore directionally consistent but modest, particularly in validation. These data do not support claiming a large gain in discrimination from inflammatory biomarkers alone. Rather, the value of the expanded model lies in integrating conventional oncology variables with frailty and nutritional reserve while maintaining calibration. The risk-stratification analysis provides a clinically intuitive illustration: when development-derived cut points were applied unchanged, observed toxicity rates in the validation cohort were 12.0%, 25.7%, and 57.5% in the low-, intermediate-, and high-risk groups, respectively. Because these cut points were based on tertiles rather than prospectively established clinical consequences, they should be regarded as descriptive risk categories rather than validated intervention thresholds.

Treatment heterogeneity also warrants consideration because the study covered a decade during which systemic therapy for lung adenocarcinoma changed substantially. Immuno-chemotherapy was not recorded in 2015–2017 and appeared after 2018, but its proportion did not increase monotonically across subsequent periods. Excluding patients who received immuno-chemotherapy or additionally adjusting for treatment year did not materially change Model C performance. These analyses reduce, but do not eliminate, concern that temporal changes in treatment composition explain the validation results. Regimen coefficients should not be interpreted as comparative safety estimates. In particular, the absence of higher adjusted toxicity risk for doublet chemotherapy or immuno-chemotherapy does not indicate that these regimens are intrinsically safer than single-agent chemotherapy. Treatment selection in routine practice is influenced by performance status, organ function, disease characteristics, physician judgment, and anticipated tolerance; patients selected for single-agent treatment may therefore have been more vulnerable at baseline. Regimen variables in this study should consequently be viewed as adjustment terms within a prediction model rather than causal exposure effects.

The toxicity spectrum further supports the clinical relevance of the early observation window. Hematologic events accounted for most primary grade ≥3 toxicities, followed by gastrointestinal and infectious events, indicating that early surveillance should pay particular attention to cytopenia, infection, and treatment-related nutritional deterioration. No treatment-related grade 5 event occurred during the prespecified two-cycle window. This finding should not be interpreted as evidence that fatal toxicity is absent in older patients. The short observation window, cohort size, clinical selection of patients considered suitable for systemic treatment, and contemporary supportive-care practices all reduce the opportunity to observe rare fatal events. The endpoint was intentionally designed to capture early severe toxicity that could disrupt planned treatment rather than cumulative or late toxicity; the model should therefore not be extrapolated to predict events occurring after later treatment cycles.

Implications and actions needed

From a clinical perspective, the model may be most useful as a pretreatment triage aid rather than a treatment-selection algorithm. The ASCO guideline recommends that geriatric vulnerabilities identified before systemic therapy be linked to appropriate management strategies (3). Randomized evidence from GAP70+ showed that providing geriatric-assessment information together with management recommendations can reduce serious treatment toxicity in older adults with advanced cancer (33), and the advanced-lung-cancer subgroup analysis similarly supported improved treatment tolerability with geriatric-assessment-guided care (34). In this context, patients classified as being at higher predicted risk could be considered for closer early laboratory monitoring, proactive nutritional assessment, surveillance for cytopenia and infection, medication review, and frailty-directed supportive interventions. Risk estimates may also support shared decision-making about expected treatment burden. However, the model should not mandate empirical dose reduction, exclude patients from effective therapy, or automatically favor one regimen over another. Decision-curve analysis suggested potential net benefit across the examined threshold range, but this does not demonstrate that use of the model itself improves outcomes; prospective impact studies are required before routine implementation.

Conclusions

For older patients with lung adenocarcinoma, we developed a pretreatment model for early severe treatment-related toxicity using clinical variables, FRAIL category, NLR, and PNI, and evaluated it in a later same-center temporal validation cohort. FRAIL-defined frailty and PNI added useful information to the final model, whereas NLR provided limited independent incremental value after other vulnerability markers were considered. The model showed moderate discrimination, acceptable calibration, and directionally consistent findings in supplementary sensitivity analyses. At this stage, Model C should be considered a risk-stratification aid for further validation rather than a ready-to-implement clinical decision tool. External multicenter validation and prospective impact evaluation are needed before routine clinical use.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Bengbu Medical University {approval No. Lun Ke Pi Zi [2024] No. 004}, and the requirement for individual informed consent was waived by the Committee because this was a retrospective analysis of de-identified records.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1930/dss

jtd-18-9-1020-dss.pdf ^{(72.6KB, pdf)}

DOI: 10.21037/jtd-2026-1930