Development and Internal Validation of a Patient‐Level Diagnostic Model Integrating Folate Receptor‐Positive Circulating Tumor Cells and Peripheral CD8 Count for Malignant‐Versus‐Benign Pathology Discrimination in a Pathology‐Enriched Pulmonary Nodule Cohort

ABSTRACT

Background

In pathology‐enriched referred hospital cohorts, stable radiologic‐pathologic linkage is often unavailable, so patient‐level malignant pathology may be a more reliable endpoint than lesion‐specific malignancy. We aimed to develop and internally validate a blood‐based model for patient‐level malignant‐versus‐benign pathology discrimination.

Methods

We retrospectively reviewed 283 consecutive eligible source records from 22 February 2023 to 12 May 2025. After excluding unclassifiable pathology, lesions > 30 mm when verifiable, and obvious extrapulmonary primary tumors, we defined a strict primary cohort. A prespecified complete‐case logistic regression model including FR + CTC, peripheral CD8 count, age, and sex was developed and internally validated with bootstrap resampling.

Results

The strict primary cohort included 234 patients (166 malignant, 68 benign). In the main model (n = 232), FR + CTC was associated with higher odds of malignant pathology (OR 1.31 per 1 FU/3 mL, 95% CI: 1.14–1.50; p < 0.001), whereas peripheral CD8 count was inversely associated (OR 0.50 per 100 cells/μL, 95% CI: 0.41–0.61; p < 0.001). Apparent and optimism‐corrected AUCs were 0.877 and 0.867, respectively. Adding an adjudicated patient‐level CT vascular sign yielded limited incremental value.

Conclusion

In this retrospective single‐center, pathology‐enriched cohort, FR + CTC and peripheral CD8 provided useful internally validated discrimination for patient‐level malignant pathology. The model requires external validation and setting‐appropriate recalibration before broader clinical use.

This graphical abstract summarizes a retrospective pathology‐enriched pulmonary nodule cohort in which a blood‐based patient‐level model integrated FR + CTC, peripheral CD8 count, age, and sex to discriminate malignant from benign pathology. The model achieved apparent and optimism‐corrected AUCs of 0.877 and 0.867, but requires external validation and recalibration before clinical use.

Abbreviations

AUC

area under the receiver operating characteristic curve

CI

confidence interval

CT

computed tomography

EDTA

ethylenediaminetetraacetic acid

FR + CTC

folate receptor‐positive circulating tumor cells

FU/3 mL

folate units per 3 mL of blood

IQR

interquartile range

LR

likelihood ratio

OR

odds ratio

PCR

polymerase chain reaction

PROBAST

Prediction model Risk Of Bias ASsessment Tool

TRIPOD

Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis

1. Background

Pulmonary nodules are frequently encountered in respiratory practice, and their evaluation should be guided by clinical context and the estimated probability of malignancy [1]. The Mayo model was developed for radiologically indeterminate solitary pulmonary nodules in clinical practice [2]. The VA model was likewise derived in a clinical solitary pulmonary nodule population [3]. The Brock model was developed from computed tomography screening cohorts [4]. The Herder approach extends lesion‐level risk estimation by incorporating fluorodeoxyglucose positron emission tomography information into indeterminate pulmonary nodule assessment [5]. Most established calculators are therefore lesion‐oriented tools rather than patient‐level pathology models [6, 7].

In referred hospital cohorts that have already entered invasive work‐up pathways, case mix and pretest probability can differ substantially from those of screening or mixed incidental‐detection populations [8]. In retrospective datasets that include multiple nodules, stable one‐to‐one linkage between a specific imaged lesion and the final pathology result may not be consistently recoverable from routine records. Under those conditions, patient‐level discrimination between malignant and benign pathology during the diagnostic episode is a more defensible analytic endpoint than retrospective assignment of malignancy to a uniquely linked index lesion. Patient‐level discrimination in such pathology‐enriched referred cohorts has been studied less often than conventional lesion‐level malignancy prediction.

Blood‐based markers may provide complementary information in this setting. Circulating tumor cells are biologically plausible because they capture a tumor‐derived signal sampled from peripheral blood [9]. Among circulating tumor cell approaches, folate receptor‐positive circulating tumor cells (FR + CTC) have shown diagnostic potential in non‐small cell lung cancer [10]. In indeterminate pulmonary nodules, FR + CTC has also been associated with malignancy probability and with preoperative assessment of invasiveness [11]. Peripheral T‐cell phenotyping may provide different information by reflecting systemic immune context rather than lesion morphology alone [12].

In this retrospective single‐center study, we aimed to develop and internally validate a model integrating FR + CTC and peripheral CD8 count for patient‐level discrimination between malignant and benign pathology in a pathology‐enriched referred hospital pulmonary nodule cohort. The study was intentionally framed as a patient‐level diagnostic model rather than as a lesion‐specific or index‐nodule malignancy calculator.

2. Methods

2.1. Study Design and Data

This retrospective single‐center study was conducted at Henan Provincial Chest Hospital. The analytical dataset consisted of 283 consecutive eligible source records collected between 22 February 2023 and 12 May 2025, and the unit of analysis was the patient. Eligible source records were pulmonary‐nodule diagnostic records in which routine‐care source metadata contained one record per hospital admission identifier and included preoperative FR + CTC assay results, lymphocyte‐subset testing, CT descriptors, and final pathology text. The dataset therefore should not be interpreted as an unselected hospital‐wide cohort of all patients with pulmonary nodules during the study period. In patients with multiple pulmonary nodules, the retrospective records did not permit stable one‐to‐one linkage between pathology and a uniquely corresponding imaged lesion in all cases. The Ethics Committee of Henan Provincial Chest Hospital approved the retrospective analysis and waived informed consent (approval no. 2025‐09‐02) for the use of de‐identified existing clinical records and laboratory results obtained during routine care; no prospective research‐specific sampling was performed for this study.

2.2. FR + CTC Assay and Lymphocyte‐Subset Measurements

FR + CTC was measured in 3 mL of ethylenediaminetetraacetic acid (EDTA)‐anticoagulated peripheral whole blood using a folate receptor cell‐detection kit based on immunomagnetic negative enrichment and ligand‐targeted polymerase chain reaction (PCR), and was reported as FU/3 mL. Lymphocyte‐subset testing was performed as part of the same routine preoperative clinical work‐up. No additional blood samples were collected solely for this retrospective research analysis. Recorded FR + CTC values were analyzed on their original scale in the primary model. The literature‐linked 8.7 FU/3 mL value was treated as an operational classification cutoff rather than as evidence that sub‐threshold recorded values were analytically zero or uninterpretable; because this retrospective dataset did not include repeated‐assay or cross‐center stability data around that cutoff, threshold‐based and piecewise alternatives were added as supplementary analyses rather than being treated as replacements for the prespecified main model.

2.3. Outcome Definition and Analytic Cohorts

The primary endpoint was pathology‐defined malignant versus benign status at the patient‐level. Because lesion‐by‐lesion linkage was not consistently available in patients with multiple nodules, the endpoint was defined at the patient‐level rather than for a uniquely linked index‐nodule. Detailed pathology adjudication rules are provided in (Additional file 1: Table S3).

Cohort formation proceeded in two steps. First, a pathologically classifiable cohort was defined after exclusion of 16 consecutive records with missing, non‐diagnostic, cytology‐only negative, borderline, or suspicious but non‐definitive pathology entries. Second, a strict primary cohort was defined by excluding classifiable records with a maximum recorded diameter > 30 mm and obvious extrapulmonary primary lesions; nine classifiable cases with missing size data were retained because > 30 mm status could not be determined from the recorded size field. Accordingly, the strict primary cohort should be understood as a nodule‐focused referred diagnostic cohort defined by the available recorded size information rather than a set in which the size criterion was verifiable for every retained case. A sensitivity analysis excluding all size‐missing strict‐cohort cases was added.

2.4. Predictor Derivation

Candidate predictors included age, sex, smoking history, FR + CTC, CD4, CD8, the CD4/CD8 ratio, maximum recorded diameter, nodule count, CT density category, and selected CT features. Smoking history was coded as ever versus never, and maximum diameter was defined as the largest recorded diameter among the patient's pulmonary nodules. Detailed variable definitions, assay information, and coding are provided in (Additional file 1: Table S2).

CT images were independently reviewed by two thoracic radiologists with 5 years of chest‐imaging experience, each blinded to the other's assessments and to the clinical and pathological data. Pleural sign, vascular sign, smooth margin, spiculation, and lobulation were coded as present if at least one pulmonary nodule showed the corresponding feature on final adjudicated review.

CT density was handled as a patient‐level aggregated descriptor. In single‐nodule patients, the recorded density was used directly; in multiple‐nodule patients, a prespecified priority rule was applied (solid > part‐solid > ground‐glass > mass/other). The mass/other category was a harmonized patient‐level residual label for raw density descriptions that did not map cleanly to solid, part‐solid, or ground‐glass and should not be interpreted as synonymous with a confirmed lesion > 30 mm. Raw‐to‐harmonized mapping rules are summarized in (Additional file 1: Table S12).

Interobserver agreement for CT feature assessment was good (Cohen's kappa = 0.891). Discrepant readings were resolved by blinded adjudication from a senior thoracic radiologist with more than 15 years of chest‐imaging experience, and the adjudicated interpretation was used for CT coding.

2.5. Statistical Analysis

Continuous‐variable distributions were assessed using the Shapiro–Wilk test. Because all major continuous variables deviated from normality in the strict cohort (Additional file 1: Table S14), continuous variables are summarized as median (interquartile range [IQR]) and categorical variables as counts and percentages. Group comparisons used the Mann–Whitney U test, chi‐square test, or Fisher exact test, as appropriate; univariable logistic regression was used to estimate crude associations with malignant pathology.

The prespecified main multivariable model included FR + CTC, CD8, age, and sex. This structure was chosen to keep the model parsimonious in a modest development dataset, to retain one tumor‐associated blood marker and one immune‐context marker, and to adjust for basic demographic factors. The CT density category was evaluated in a supplementary robustness model rather than forced into the primary biomarker model, and an extended model additionally included vascular sign, the adjudicated CT feature with the most stable univariable association.

CD4, CD8, and the CD4/CD8 ratio were not entered simultaneously because they represent closely related immune measures and because CD8 was the prespecified immune‐marker in the main model. Formal correlation and variance‐inflation‐factor (VIF) diagnostics were added to assess this decision (Additional file 1: Table S15). FR + CTC was analyzed per 1 FU/3 mL, CD8 per 100 cells/μL, and age per 10 years; models were fitted on complete cases.

Exploratory spline‐based likelihood‐ratio tests assessed linearity for FR + CTC, CD8, and age (Additional file 1: Table S8). FR + CTC parameterization analyses compared continuous FR + CTC, dichotomized FR + CTC using 8.7 FU/3 mL, and threshold‐aware piecewise or hybrid alternatives (Additional file 1: Table S16). Model performance was summarized by apparent and optimism‐corrected area under the receiver operating characteristic curve (AUC), Brier score, and bootstrap mean calibration slope using 1000 bootstrap resamples. Variable‐missingness, normality, correlation structure, and collinearity diagnostics are shown in Additional file 1: Figures S1 and S2 and Tables S14 and S15, and bootstrap validation summaries are provided in (Additional file 1: Table S6). Smoothed calibration plots were generated for the main and extended models. Prespecified sensitivity analyses repeated model fitting and 5‐fold cross‐validated AUC estimation in all pathologically classifiable cases, the single‐nodule subset, the patient‐level non‐solid subset, and the strict cohort after excluding all size‐missing cases (Additional file 1: Table S7). Clinical/CT baseline models were compared with biomarker‐augmented models on the same‐complete‐case‐subset required for each biomarker‐augmented benchmark and were used only as exploratory internal comparisons to contextualize incremental performance (Additional file 1: Table S17). Model reporting was aligned with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement [13]. Risk‐of‐bias considerations were judged against core domains of the Prediction model Risk Of Bias ASsessment Tool (PROBAST) [14]. Bootstrap internal validation was preferred over random split‐sample validation in this modest development dataset [15]. The available sample size was treated as a constraint on model complexity rather than as a rationale for broad data‐driven expansion [16]. Further technical details and supplementary analyses are provided in Additional file 1.

3. Results

3.1. Cohort Formation

Of 283 consecutive eligible source records with pulmonary‐nodule diagnostic documentation, routine preoperative FR + CTC and lymphocyte‐subset results, CT descriptors, and final pathology text, 267 were pathologically classifiable, including 196 malignant and 71 benign cases. After exclusion of 30 classifiable records with a recorded maximum diameter > 30 mm and 3 obvious extrapulmonary primary lesions, the strict primary cohort comprised 234 patients (166 malignant, 68 benign). Nine classifiable cases with missing size information were retained because > 30 mm status could not be determined; excluding these nine cases left 225 strict‐cohort patients (163 malignant, 62 benign) for sensitivity analysis. The cohort derivation is shown in Figure 1.

FIGURE 1.

Case flow and cohort derivation.

3.2. Baseline Characteristics

Baseline characteristics of the strict primary cohort are shown in Table 1. Shapiro–Wilk tests supported use of median and IQR summaries for continuous variables (Additional file 1: Table S14). FR + CTC was higher in patients with malignant pathology than in the benign group as a whole (median 11.8 vs. 8.8 FU/3 mL, p < 0.001), whereas peripheral CD8 was lower (median 398.0 vs. 687.5 cells/μL, p < 0.001). The distributions of FR + CTC and peripheral CD8 are shown in Figure 2. CD4 was also lower and the CD4/CD8 ratio higher in the malignant group. Among CT descriptors, pleural sign and vascular sign were more frequent in malignant cases, whereas smooth margin was more frequent in benign cases.

TABLE 1.

Baseline characteristics of the strict primary cohort.

Variables	Malignant (n = 166)	Benign (n = 68)	p
Age, years	60.0 (53.0–67.8)	59.0 (52.8–68.0)	0.625
Maximum diameter, mm	15.0 (10.0–20.0)	13.0 (9.0–17.0)	0.255
FR + CTC, FU/3 mL	11.8 (9.8–14.0)	8.8 (7.5–10.6)	< 0.001
CD4, cells/μL	645.5 (568.2–689.0)	757.0 (661.2–890.0)	< 0.001
CD8, cells/μL	398.0 (342.8–523.8)	687.5 (532.2–923.2)	< 0.001
CD4/CD8 ratio	1.6 (1.2–1.8)	1.0 (0.8–1.6)	< 0.001
Male sex	76/166 (45.8%)	32/68 (47.1%)	0.859
Ever smoking	42/166 (25.3%)	23/68 (33.8%)	0.186
Multiple nodules	39/165 (23.6%)	24/67 (35.8%)	0.059
Pleural sign	49/166 (29.5%)	11/68 (16.2%)	0.034
Vascular sign	59/166 (35.5%)	13/68 (19.1%)	0.013
Smooth margin	8/166 (4.8%)	14/68 (20.6%)	< 0.001
Spiculation	25/166 (15.1%)	6/68 (8.8%)	0.201
Lobulation	28/166 (16.9%)	9/68 (13.2%)	0.489
CT density category			0.019
Solid	40/166 (24.1%)	30/68 (44.1%)
Part‐solid	70/166 (42.2%)	18/68 (26.5%)
Ground‐glass	51/166 (30.7%)	19/68 (27.9%)
Mass/other	4/166 (2.4%)	1/68 (1.5%)

Note: Values are median (IQR) or n/N (%). Denominators vary because of missing data: maximum diameter was available in 225 patients, nodule count in 232, and CT density category in 233. p values are from the Mann–Whitney U test or χ ²/Fisher exact test, as appropriate. Binary CT signs were coded positive if present in at least one pulmonary nodule on final adjudicated review. In patients with multiple nodules, the final CT density category was assigned using the rule solid > part‐solid > ground‐glass > mass/other.

FIGURE 2.

Distribution of FR + CTC and peripheral CD8 in the strict primary cohort.

The patient‐level CT density category also differed between groups (overall p = 0.019): among 233 patients with non‐missing data, 70 were classified as solid, 88 as part‐solid, 70 as ground‐glass, and 5 as mass/other.

Benign pathology labels were reviewed descriptively to contextualize the CD8 finding. In the strict cohort, 63 of 68 benign cases were classified as active inflammatory or infectious lesions; subgroup‐specific CD8 summaries are shown in Additional file 1: Tables S10 and S11 and were used for descriptive interpretation only.

3.3. Univariable Associations

Spearman correlation analysis showed that the strongest dependency was between CD8 and the CD4/CD8 ratio (rho = −0.841, p < 0.001), followed by more moderate correlations between CD4 and CD8 (rho = 0.382, p < 0.001), FR + CTC and CD8 (rho = −0.295, p < 0.001), and FR + CTC and CD4 (rho = −0.277, p < 0.001; Additional file 1: Table S15 and Figure S2). Univariable logistic regression results are summarized in Table 2. FR + CTC and the CD4/CD8 ratio were positively associated with malignant pathology, whereas CD4 and CD8 were inversely associated. Vascular sign and pleural sign were positively associated and smooth margin inversely associated. The CT density category was associated with malignant pathology overall (likelihood‐ratio p = 0.021), whereas age, sex, smoking history, maximum diameter, spiculation, and lobulation were not significant within the strict cohort.

TABLE 2.

Univariable logistic regression for malignant pathology in the strict primary cohort.

Variables	n	OR (95% CI)	p
Age (per 10 years)	234	1.08 (0.85–1.37)	0.542
Male sex	234	0.95 (0.54–1.67)	0.859
Ever smoking	234	0.66 (0.36–1.22)	0.188
Maximum diameter (per 5 mm)	225	1.13 (0.89–1.43)	0.310
FR + CTC (per 1 FU/3 mL)	232	1.32 (1.18–1.47)	< 0.001
CD4 (per 100 cells/μL)	234	0.53 (0.42–0.67)	< 0.001
CD8 (per 100 cells/μL)	234	0.49 (0.40–0.59)	< 0.001
CD4/CD8 ratio (per 1‐unit increase)	234	4.86 (2.44–9.69)	< 0.001
Multiple nodules	232	0.55 (0.30–1.03)	0.060
Pleural sign	234	2.17 (1.05–4.49)	0.037
Vascular sign	234	2.33 (1.18–4.62)	0.015
Smooth margin	234	0.20 (0.08–0.49)	< 0.001
Spiculation	234	1.83 (0.72–4.69)	0.207
Lobulation	234	1.33 (0.59–2.99)	0.491
CT density category, overall LR test	233	—	0.021
Part‐solid versus solid	233	2.92 (1.45–5.88)	0.003
Ground‐glass versus solid	233	2.01 (0.99–4.09)	0.053
Mass/other versus solid	233	3.00 (0.32–28.25)	0.337

Note: n indicates the non‐missing sample size for that analysis. CT variables are patient‐level aggregated descriptors rather than lesion‐specific annotations.

3.4. Multivariable Models

In the prespecified main multivariable model (Table 3A), FR + CTC and CD8 remained independently associated with malignant pathology after adjustment for age and sex. FR + CTC was associated with higher odds of malignant pathology (OR 1.31 per 1 FU/3 mL, 95% CI: 1.14–1.50; p < 0.001), whereas CD8 was inversely associated (OR 0.50 per 100 cells/μL, 95% CI: 0.41–0.61; p < 0.001). Age and sex were not significant. Formal VIF diagnostics did not indicate problematic collinearity in the main model (VIF range, 1.056–1.126; Additional file 1: Table S15).

TABLE 3A.

Multivariable logistic regression: main model.

Variables	OR (95% CI)	p
FR + CTC (per 1 FU/3 mL)	1.31 (1.14–1.50)	< 0.001
CD8 (per 100 cells/μL)	0.50 (0.41–0.61)	< 0.001
Age (per 10 years)	0.93 (0.68–1.28)	0.655
Male sex	0.53 (0.25–1.13)	0.098

Note: Main‐model predictors were FR + CTC, CD8, age, and sex. Full coefficients and the intercept are provided in (Additional file 1: Table S4).

In the CT‐density‐adjusted robustness model (Additional file 1 Table S13), the associations of FR + CTC and CD8 were materially unchanged. In exploratory FR + CTC parameterization analyses, the 8.7 FU/3 mL operational cutoff representation performed similarly to the prespecified continuous primary specification when added to CD8, age, and sex (5‐fold cross‐validated AUC, 0.872 vs. 0.861), so this small internal difference was treated as a sensitivity finding rather than as evidence to replace the prespecified main model; a simple piecewise linear knot at 8.7 FU/3 mL did not improve fit over the continuous specification (likelihood‐ratio p = 0.871; Additional file 1: Table S16). In the extended model (Table 3B), vascular sign remained independently associated with malignant pathology (OR 3.30, 95% CI: 1.33–8.24; p = 0.010), but discrimination improved only modestly.

TABLE 3B.

Multivariable logistic regression: extended model.

Variables	OR (95% CI)	p
FR + CTC (per 1 FU/3 mL)	1.31 (1.14–1.51)	< 0.001
CD8 (per 100 cells/μL)	0.49 (0.40–0.61)	< 0.001
Age (per 10 years)	1.01 (0.73–1.40)	0.966
Male sex	0.54 (0.25–1.19)	0.125
Vascular sign	3.30 (1.33–8.24)	0.010

Note: The extended model added vascular sign to the main model. Vascular sign was coded as a patient‐level aggregated CT descriptor, positive if present in at least one pulmonary nodule on final adjudicated review.

3.5. Model Performance and Sensitivity Analyses

Model performance in the strict primary cohort is summarized in Table 4. In the complete‐case modeling sample (n = 232; 164 malignant events), the main model had an apparent AUC of 0.877 (95% CI: 0.818–0.912), an optimism‐corrected AUC of 0.867, an apparent Brier score of 0.121, a corrected Brier score of 0.128, and a bootstrap mean calibration slope of 0.932. The corresponding values for the extended model were 0.887 (95% CI: 0.833–0.921), 0.875, 0.117, 0.126, and 0.914. Single‐marker benchmarks showed that CD8 alone had higher apparent and cross‐validated discrimination than FR + CTC alone in this cohort, but this comparison is scale‐ and comparator‐dependent and was not treated as proof of biomarker superiority (Additional file 1: Table S16). Clinical/CT baseline models had lower cross‐validated discrimination than biomarker‐augmented models in same‐complete‐case‐subset exploratory internal benchmarks, but these comparisons should not be read as external validation of incremental clinical utility (Additional file 1: Table S17). Receiver operating characteristic curves for FR + CTC alone, the main model, and the extended model are shown in Figure 3, and calibration plots for the main and extended models are shown in Figure 4. Development‐cohort separation of predicted probabilities by pathology group and bootstrap distributions of apparent AUC are shown in Additional file 1: Figures S3 and S4, respectively.

TABLE 4.

Model performance in the strict primary cohort.

Model	n	Apparent AUC (95% CI)	Apparent brier	Optimism‐corrected AUC	Corrected brier	Bootstrap mean calibration slope (rounded)
Main model: FR + CTC + CD8 + age + sex	232	0.877 (0.818–0.912)	0.121	0.867	0.128	0.932
Extended model: main model + vascular sign	232	0.887 (0.833–0.921)	0.117	0.875	0.126	0.914

Note: n indicates the complete‐case modeling sample. The calibration slope is the bootstrap mean calibration slope, shown to three decimals; it should not be interpreted on its own as evidence of perfect calibration.

FIGURE 3.

Receiver operating characteristic curves for FR + CTC alone, the main model, and the extended model in the complete‐case strict‐cohort sample (n = 232).

FIGURE 4.

Calibration of the main and extended models in the complete‐case strict‐cohort sample (n = 232). Smoothed curves show calibration across predicted risk; points show grouped observed‐versus‐predicted values.

Threshold‐dependent classification measures at the Youden threshold are provided in Additional file 1: Table S5 as descriptive development‐cohort summaries rather than recommended clinical operating points. Illustrative intercept‐only recalibration values for different target prevalences are provided in (Additional file 1: Table S9).

Sensitivity analyses were supportive across prespecified analytic sets. The main and extended models showed similar 5‐fold cross‐validated AUCs in the strict cohort, all pathologically classifiable cases, the single‐nodule subset, the patient‐level non‐solid subset, and the strict cohort after excluding all size‐missing cases. In the size‐missing exclusion analysis, 5‐fold cross‐validated AUCs were 0.874 for the main model and 0.875 for the extended model. Detailed results are provided in (Additional file 1: Table S7 and Figure S5).

4. Discussion

This study addressed a diagnostic problem not fully captured by conventional pulmonary nodule frameworks: patient‐level discrimination between malignant and benign pathology during a referred hospital diagnostic episode in a pathology‐enriched cohort. Within this framework, FR + CTC and peripheral CD8 provided most of the discriminatory information, whereas the incremental contribution of aggregated CT descriptors was limited. These findings suggest that, in a referred cohort already selected for invasive work‐up, blood‐based markers may help refine patient‐level malignant‐versus‐benign pathology discrimination. They do not justify using the model as a lesion‐specific or index‐nodule malignancy calculator.

Comparison with published nodule calculators requires attention to the derivation setting. In screening cohorts, Brock‐based models have shown strong discrimination under screening‐prevalence conditions [17]. In clinical validation studies, however, performance has been more variable [18]. In a large incidental‐detection cohort, existing probability models overestimated cancer risk despite acceptable discrimination [19]. Recent work continues to refine pulmonary nodule triage using data‐driven risk‐stratification systems [20]. Multimodal biomarker models have also improved non‐invasive classification in external‐validation settings [21]. External validations have repeatedly shown that model performance is population dependent rather than universal [22]. Meta‐analytic evidence on Brock‐based tools and deep‐learning approaches likewise indicates that transportability depends on case mix, implementation context, and calibration behavior [23, 24].

The FR + CTC findings are directionally consistent with earlier FR + CTC literature, although comparability is incomplete. Early studies showed that FR + CTC measured by ligand‐targeted polymerase chain reaction could discriminate lung cancer from benign lung disease with moderate‐to‐good accuracy [25]. A recent clinical series also reported that FR + CTC was associated with lung‐cancer diagnosis and with several clinicopathological characteristics [26]. More recent single‐center models in solitary pulmonary nodules have combined circulating‐tumor‐cell information with other blood‐based variables to improve preoperative classification [27]. Another hematology‐based model built around circulating tumor cells also achieved useful discrimination in solitary pulmonary nodules [28]. A 2024 pulmonary‐nodule study reported that FR + CTC alone had moderate diagnostic accuracy and performed better in combined models than as a stand‐alone marker [29]. Systematic reviews have generally concluded that FR + CTC‐based approaches are promising but heterogeneous with respect to population spectrum, thresholds, and assay implementation [30, 31]. In the present internal dataset, the conventional 8.7 FU/3 mL operational threshold‐based representation performed similarly, and numerically slightly higher by cross‐validated AUC, than the continuous primary representation when combined with CD8, age, and sex. We therefore treated 8.7 FU/3 mL as a clinical classification cutoff, not as proof that recorded sub‐threshold values are analytically zero or uninterpretable. Because this retrospective dataset lacked repeated measurements and cross‐center assay data near the cutoff, the results cannot show that near‐threshold continuous variation is more transportable than a threshold‐based representation. This exploratory finding supports presenting threshold‐based results alongside the prespecified continuous model; it does not establish that either FR + CTC representation is more transportable without external validation. Broader prospective CTC screening work suggests that a circulating‐cell signal can complement imaging in early lung‐cancer detection, but that evidence does not directly validate the patient‐level referred‐cohort use case addressed here [32].

The inverse association between peripheral CD8 and malignant pathology should be interpreted more cautiously than the FR + CTC signal. Peripheral immune‐marker studies in lung cancer and pulmonary nodules are heterogeneous with respect to assay platform, whether they analyze absolute counts or proportions, and what they use as the comparison group. Recent peripheral immune‐profiling work reported diagnostic information at the level of multifeature circulating T‐cell patterns rather than a single universally specific marker [12]. Another study found that the proportion of peripheral KLRG1‐positive CD8‐positive T cells was higher in non‐small‐cell lung cancer than in benign pulmonary nodules and healthy controls [33]. A 2025 single‐cell peripheral immune study further suggested that blood immune signatures can help distinguish invasive from non‐invasive small pulmonary nodules [34]. These studies are informative but are not analytically interchangeable with the absolute peripheral CD8 count used here. In our cohort, the benign comparator group was dominated by inflammatory and infectious pathology. Peripheral CD8 is therefore better interpreted as a setting‐dependent marker of systemic immune context than as a tumor‐specific mechanistic signature.

The same setting dependence also helps explain the limited added value of CT morphology in the extended model. Because CT information was intentionally reduced to patient‐level aggregated descriptors, some lesion‐level resolution was inevitably lost. The modest gain from vascular sign should therefore be interpreted as a consequence of the analytic design, not as evidence that CT morphology is unimportant in conventional nodule assessment.

Several methodological features strengthen the interpretability of the present study. First, the endpoint matched what the retrospective data could support reliably: patient‐level malignant versus benign pathology rather than unstable retrospective assignment of pathology to a unique lesion in all cases. Second, pathology adjudication rules, cohort construction, and predictor coding were prespecified and transparent. Third, CT features were assessed by blinded dual readers with adjudication and good interobserver agreement, which reduced predictor‐review bias. Fourth, the multivariable model remained deliberately parsimonious, limiting unnecessary complexity in a modest single‐center dataset. Finally, internal validation used bootstrap resampling rather than an arbitrary split‐sample strategy [15].

The limitations are substantial and should guide interpretation of the model. This was a retrospective, single‐center, pathology‐enriched referred cohort with a high malignant fraction; therefore, absolute risk estimates, predictive values, and threshold‐dependent operating characteristics are setting‐dependent. Although the source records were consecutive eligible diagnostic records during the study period, the cohort was not an unselected hospital‐wide sample of all patients with pulmonary nodules: inclusion required available routine clinical FR + CTC and lymphocyte‐subset testing, CT descriptors, and pathological classification. Spectrum bias and verification bias therefore remain important concerns. Calibration should not be inferred from discrimination alone [35, 36]. If the model is assessed in a new referral population, intercept recalibration might be insufficient if predictor effects or case spectrum shift; coefficient updating might also be required [37]. The modest sample size also argues against aggressive model expansion beyond the prespecified parsimonious structure [16]. Because benign pathology was heterogeneous and heavily weighted toward inflammatory and infectious disease, the observed contribution of peripheral CD8 is likely to be comparator‐dependent. In addition, nine retained strict‐cohort cases lacked size data, so the > 30 mm exclusion criterion could not be verified directly for every included patient, although sensitivity analysis excluding those cases gave similar discrimination; likewise, the harmonized mass/other CT‐density category represented residual patient‐level density wording rather than a confirmed large‐mass phenotype. Most importantly, these data are insufficient to support clinical transport without independent external validation.

Clinically, this model should be interpreted narrowly: it is a patient‐level malignant‐versus‐benign pathology discrimination tool for pathology‐enriched referred hospital populations, not a universal pulmonary nodule malignancy model and not a lesion‐specific or index‐nodule risk score. Its potential role is to refine risk discussion during a diagnostic episode among patients already selected for specialist evaluation, rather than to replace established lesion‐level frameworks used in screening or incidental nodule management. The present study does not establish safety for deferring biopsy, surgery, or other invasive procedures on the basis of model output alone. External validation in independent referred cohorts, followed by setting‐appropriate recalibration and formal assessment of clinical consequences, is required before any broader implementation can be justified.

5. Conclusion

In this retrospective single‐center, pathology‐enriched referred diagnostic cohort centered on pulmonary nodules, a parsimonious patient‐level model integrating FR + CTC and peripheral CD8 showed good internally validated discrimination between malignant and benign pathology. The findings support FR + CTC and CD8 as informative patient‐level signals in this specific referred setting, whereas adding patient‐level aggregated CT descriptors yielded limited incremental improvement. These data do not establish lesion‐specific malignancy prediction, superiority of continuous FR + CTC over threshold‐based representation, or safety for changing invasive‐management decisions without external validation. Independent validation and, where needed, recalibration are required before broader clinical use.

Author Contributions

Hongjian Su: investigation, writing – review and editing. Weidong Zhang: investigation, writing – review and editing. Chunyan Zhang: conceptualization, methodology, data curation, formal analysis, funding acquisition, writing – original draft. Xiaoke Sun: investigation, writing – review and editing. Li Gong: investigation, data curation, writing – review and editing. Xiaojie Ma: investigation, writing – review and editing. Fang Sun: supervision, validation, writing – review and editing, methodology. Yaqian Liu: data curation, writing – review and editing.

Funding

This study was supported by the 2024 Henan Provincial Medical Science and Technology Research Program Joint Co‐construction Project (Grant No. LHGJ20240142). The funder had no role in the design of the study; the collection, analysis, or interpretation of data; the writing of the manuscript; or the decision to submit the work for publication.

Ethics Statement

This retrospective study was approved by the Ethics Committee of Henan Provincial Chest Hospital (approval no. 2025‐09‐02) and was conducted in accordance with the Declaration of Helsinki and relevant institutional requirements. The approval and waiver covered the retrospective use of de‐identified existing clinical records and laboratory data. FR + CTC and lymphocyte‐subset measurements were obtained as routine clinical tests during the diagnostic work‐up; no additional blood samples were collected solely for this retrospective research analysis. The requirement for informed consent was waived because of the retrospective design and use of de‐identified existing data.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Missing‐data profile of candidate variables. Each lollipop shows the percentage of missing observations across all 283 patients. Maximum diameter had the highest proportion of missing values, although overall missingness was low.

Figure S2: Spearman correlation matrix in the strict primary cohort. The strongest dependency was between CD8 and the CD4/CD8 ratio, supporting the decision not to include highly related immune variables in the same primary multivariable model.

Figure S3: Separation of predicted probabilities by pathology group. Both models separated malignant and benign pathology groups in the development‐cohort, with only a modest incremental shift for the extended model.

Figure S4: Bootstrap distributions of apparent AUC. Across bootstrap resamples, the apparent AUC distribution of the extended model was shifted slightly to the right of that of the main model, although overlap remained substantial.

Figure S5: Sensitivity‐analysis discrimination summary. Horizontal connectors link the 5‐fold cross‐validated AUCs of the main and extended models within each prespecified analytic set.

Table S1: Case flow and cohort derivation.

Table S2: Assay details, variable definitions, and coding.

Table S3: Pathology adjudication rules and handling categories.

Table S4: Full logistic regression coefficients and prediction equations.

Table S5: Apparent threshold‐dependent classification metrics at the Youden threshold.

Table S6: Bootstrap internal validation summary.

Table S7: Sensitivity analyses.

Table S8: Exploratory checks of linearity on the log‐odds scale.

Table S9: Example intercept‐only recalibration of the main and extended models under hypothetical target malignant prevalence.

Table S10: Pathology‐label mapping of benign lesions for interpretation of peripheral CD8.

Table S11: Descriptive distribution of peripheral CD8 according to benign‐pathology subgroup in the strict cohort.

Table S12: Raw‐to‐harmonized mapping of CT‐density labels and patient‐level derivation rule.

Table S13: CT‐density‐adjusted robustness model.

Table S14: Shapiro–Wilk normality checks for continuous variables in the strict primary cohort.

Table S15: Spearman correlation and variance‐inflation‐factor diagnostics.

Table S16: FR + CTC parameterization and marker‐benchmark analyses.

Table S17: Clinical/CT baseline and biomarker‐augmented model benchmarking.

Data Availability Statement

The de‐identified analytic dataset and the analysis code that support the findings of this study are available from the corresponding author on reasonable request.