Radiation Doses to Cardiac Substructures Predict Elevation in High-sensitivity Cardiac Troponin T Levels in Radiation Therapy for Lung Cancer

Abstract

Purpose:

Cardiotoxicity is a major concern for patients undergoing thoracic radiation therapy. This study compared the predictive power of radiation dose-volume histogram (DVH) parameters and radiomic/dosiomic features of the whole heart (WH) and cardiac substructures for elevated circulating high-sensitivity cardiac troponin T (hs-cTnT), a biomarker for early detection of cardiac adverse events.

Methods and Materials:

A retrospective cohort of 160 patients with non-small cell lung cancer (NSCLC) from a completed prospective trial and a prospective cohort of 57 patients with NSCLC enrolled in an ongoing trial were analyzed. The endpoint was hs-cTnT elevation, indicated by an increase of ≥5 ng/L from baseline. An in-house auto-segmentation model delineated 19 cardiac substructures. DVH parameters, radiomic, and dosiomic features were extracted from each patient. A 100-iteration Monte Carlo cross-validation (75%/25% split) was conducted within the retrospective cohort to mitigate random split bias. Logistic regression models using different input combinations were compared, with a model using only clinical factors serving as the baseline. Key predictive features were identified using permutation importance during training. Models were validated by holdout in the prospective cohort to evaluate robustness. Model performance was assessed by the area under the receiver operating characteristic curve (AUROC).

Results:

Incidence of hs-cTnT elevation was 31.9% in the retrospective and 29.8% in the prospective cohort. The substructure DVH model achieved the highest predictive performance in cross-validation (mean AUROC, 0.71; 95% CI, 0.70-0.73) and demonstrated greater robustness in holdout validation compared with WH-based models (AUROC: 0.60 vs ≤0.51). Feature analysis identified the left anterior descending coronary artery V20Gy as the most dominant predictor, with cutoff values ranging from 0.2% to 5% using various indices.

Conclusions:

Cardiac substructure DVH parameters have superior predictive power and robustness over WH variables for predicting hs-cTnT elevation in NSCLC radiation therapy, emphasizing the need to use cardiac substructures for cardiotoxicity risk assessment.

Introduction

Cardiotoxicity has long been recognized as a major complication for patients receiving thoracic radiation therapy, with substantial evidence linking thoracic irradiation to increased cardiac risk.^1–4 Cardiac diseases accounted for 5.3% of total deaths in a disease-specific mortality analysis of more than 270,000 patients with lung cancer, ranking second only to tumor-related mortality.⁵ Moreover, radiation-related cardiovascular mortality may be underestimated because of frequent cardiac-related hospitalizations and numerous sudden deaths without documented cancer recurrence.⁶ After radiation therapy, cardiotoxicity presents significant clinical challenges because it negatively affects both overall survival and quality of life. A study of patients with classic Hodgkin lymphoma indicated that cardiovascular disease has surpassed the primary cancer and other malignancies as the leading cause of mortality.⁷ In breast cancer, relative risks for specific cardiac conditions and cardiac mortality range from 1.29 to 1.97 across several investigations.^8–10 In lung cancer, increased radiation dose is associated with reduced overall survival, with a hazard ratio of 1.38 compared with standard dose treatment.¹¹ Darby et al¹² reported a 7.4% increase in the rates of major coronary events per 1 Gy increase in mean heart dose. However, the mean heart dose fails to account for heterogeneous dose distribution within the heart and may not accurately reflect the actual dose delivered to cardiac substructures.¹³ Emerging evidence suggests that dosimetric parameters of specific cardiac substructures better predict cardiotoxicity than whole heart (WH) dose metrics.^8,14–21 Cardiac substructures such as the left ventricle,^16,18 left atrium,¹⁵ left anterior descending coronary artery (LAD),^14,17,20,21 and heart valves¹⁹ have been implicated in radiation-induced cardiac damage. Efforts to spare these critical substructures during treatment planning may reduce the risk of cardiotoxicity while maintaining high doses to targets for tumor control.

The incidence of cardiac complications in patients with lung cancer can reach up to 33%.²² Various studies have proposed different dose constraints to mitigate cardiac risk.²³ Although factors such as pre-existing heart conditions and smoking status also influence risk, an individualized approach for predicting cardiotoxicity is crucial for personalized treatment planning. Because radiation-induced cardiac damage can develop gradually, several biomarkers have emerged as potential indicators of subclinical cardiotoxicity during chemoradiation therapy (CRT).²⁴ Cardiac troponin T, released into the serum by damaged cardiac myocytes, serves as a specific marker of myocardial injury in routine cardiology practice.²⁵ Elevation in high-sensitivity cardiac troponin T (hs-cTnT) levels has demonstrated predictive value as an early biomarker of radiation- or chemo-induced myocardial damage.^26–30 A relationship with clinical cardiotoxicity has, however, not been fully demonstrated yet.

Historically, the labor-intensive process of manually delineating cardiac substructures posed a significant barrier to large-scale dosimetric analyses and cardiotoxicity prediction studies. However, advances in deep-learning–based algorithms have improved the efficiency and consistency of cardiac substructure delineation, producing clinically acceptable contours suitable for dosimetric studies.^31–36 In this study, we leveraged deep-learning–based auto-segmentation models to facilitate the development of elastic net models for predicting hs-cTnT elevation as a surrogate marker of radiation-induced cardiotoxicity. The models integrated a comprehensive set of inputs, including clinical factors, dose-volume histogram (DVH) parameters, and radiomic and dosiomic features. Dosiomic features are radiomic features extracted from a 3-dimensional dose map. We evaluated the predictive performance of WH-based and cardiac substructure-based variables and assessed model robustness through holdout validation by using data from patients prospectively enrolled in an ongoing clinical trial. This approach holds promise for future integration into personalized treatment planning workflows, enabling more precise cardiotoxicity risk assessment in lung cancer radiation therapy.

Methods and Materials

Patient cohorts

This study was approved by the institutional review board. The retrospective data set was obtained from a completed randomized prospective trial comparing intensity modulated radiation therapy (IMRT) with passive scattering proton therapy (PSPT) in patients with non-small cell lung cancer (NCT009105005).³⁷ Of 225 eligible patients, 160 were selected after excluding those who withdrew consent for blood sample donation or lacked sufficient samples for hs-cTnT level analysis. Those patients received either IMRT or PSPT with prescribed doses ranging from 60 to 74 Gy between 2009 and 2014, combined with concurrent chemotherapy. Clinical factors such as sex (biologically, male or female), age, smoking status, and pre-existing heart diseases were recorded, as detailed in prior publications.^29,37 Another 74 patients were recruited from an ongoing clinical trial (NCT05010109) starting in 2021, with 57 meeting the same inclusion criteria. Those patients received photon-based treatments, including IMRT or volumetric-modulated arc therapy, or proton-based treatments, including PSPT or intensity modulated proton therapy, with prescribed doses ranging from 60 to 72 Gy. For the majority of patients, tumor internal motion was assessed using 4-dimensional computed tomography (CT) to generate an internal planning target volume, while a minority were treated under breath-hold conditions as per the decision of physicians. In intensity modulated proton therapy plans, repainting and robust optimization techniques were employed to mitigate range uncertainties and interplay effects associated with respiratory motion. In the current study, the retrospective data set was used for training and Monte Carlo cross-validation (MCCV) of a prediction model for hs-cTnT elevation, and the prospective data set served as an independent holdout validation set for the hs-cTnT elevation prediction model, as illustrated in Figure 1.

Fig. 1.

CONSORT diagram with schematic representation of the data collection, Monte Carlo cross-validation, and holdout validation processes. Abbreviations: hs-cTnT = high-sensitivity cardiac troponin T; NSCLC = non-small cell lung cancer.

Hs-cTnT measurement

Baseline blood samples were collected within 2 months before CRT for the retrospective data set and within 2 weeks before CRT for the prospective data set. Additional samples were obtained 2 to 3 times during CRT, including during the final week of treatment in both data sets. Serum was separated within 2 hours of collection and stored at −80 °C for testing and long-term preservation. The hs-cTnT concentrations in serum were measured in duplicate with an Elecsys TnT Gen 5 STAT assay (Roche Diagnostics) on a Cobas e411 analyzer. Samples within each data set were analyzed using the same assay. A delta in hs-cTnT level exceeding 5 ng/L has been linked with a significantly increased risk of cardiac adverse events.²⁹ Therefore, in this study, hs-cTnT elevation was defined as a maximum change in hs-cTnT levels exceeding 5 ng/L during CRT and was considered the primary endpoint. Although dichotomizing the hs-cTnT outcome into elevated versus nonelevated may result in the loss of detailed quantitative information, it mitigates the influence of outlier values and reduces bias introduced by extreme hs-cTnT level changes observed in individual patients.

Auto-segmentation

Figure 2 illustrates the overall workflow of this study. An in-house nnU-Net based auto-segmentation model was applied to the planning average 4-dimensional CT images to delineate a total of 19 substructures, including the WH, left atrium, right atrium, left ventricle, right ventricle, ascending aorta, descending aorta, pulmonary artery, pulmonary vein, superior vena cava, inferior vena cava, aortic valve, mitral valve, pulmonary valve, tricuspid valve, LAD, left main coronary artery, left circumflex coronary artery, and right coronary artery.³¹ The model was trained using contours developed according to consensus guidelines informed by the University of Michigan cardiac atlas³⁸ and refined with input from our institutional cardiology experts, as detailed in the appendix of the original publication. This model has been extensively evaluated and has achieved an overall rate of clinically acceptable contours of 94%.

Fig. 2.

Schematic of study workflow. Abbreviations: AUPRC = area under the precision-recall curve; AUROC = area under the receiver operating characteristic curve; CT = computed tomography; CV = cross-validation; DVH = dose-volume histogram; EHR = electronic health record; FPR = false-positive rate; TPR = true-positive rate; VIF = variance inflation factor; WH = whole heart.

DVH parameters

DVH parameters were extracted from clinically delivered radiation therapy treatment plans by using the open-source Python package “PlatiPy.” For patients with multiple treatment plans or significant plan adaptations, accumulated dose maps were generated by registering the re-simulated CT images to the original planning CT images by deformable image registration, conducted in Raystation 12A (RaySearch Laboratories AB). A total of 23 DVH metrics were extracted for each substructure, including minimum dose, mean dose, maximum dose, SD, relative volumes receiving at least x or x Gy (relative biologic effectiveness [RBE]) (including V5Gy and V10-V70Gy in 10-Gy increments), doses received by at least x% of the volume (including D5% and D10%-D70% in 10% increments), and dose received by absolute x cc volumes (including D0.1, D1.0, and D5.0 cc). In total, 437 DVH parameters were extracted for each patient.

Radiomic and dosiomic features

A total of 106 radiomic features were extracted from the planning CT for each cardiac substructure. The same feature categories were extracted from the 3-dimensional dose maps and denoted as dosiomic features. Feature extraction was done with the open-source Python package “pyradiomics.”³⁹ This process resulted in a total of 2014 radiomic features and 2014 dosiomic features per patient.

To enhance reliability and generalizability, robustness analysis was done with the retrospective data set. Unstable features were excluded through an image perturbation approach.⁴⁰ A perturbation chain was applied to each patient to acquire a set of perturbed data including CT images, substructure masks, and dose maps: (1) applying random rotations to CT images, substructure masks, and dose maps to mimic variations in patient positioning; (2) adding random Gaussian noise to CT images to account for differences in imaging systems; (3) performing random volume shrinkage/growth on substructure masks to reflect volume variations in manual delineation; and (4) introducing random deformation of masks to represent contour uncertainties. The same features described earlier were extracted from perturbed CT images, masks, and dose maps. Features with Spearman correlation coefficients of >0.9 from before to after the perturbation were considered robust and were retained for further processing. Detailed procedures for feature extraction and robustness analysis are provided in Appendices E1 and E2.

Univariate analysis and preselection of variables

Clinical factors, DVH parameters, radiomic features, and dosiomic features were compared between patients with and without hs-cTnT elevation during CRT. The Wilcoxon rank-sum test and Pearson χ² test were used for continuous and categorical features, respectively. Given the large number of DVH, radiomic, and dosiomic features relative to the patient sample size, univariate statistical testing was conducted to identify the most meaningful variables for model input. For each substructure, the 2 most significant variables within each feature group (DVH, radiomics, dosiomics) were selected based on the lowest P values from the Wilcoxon rank-sum tests within the retrospective data set. These selected features were used to construct elastic net models in the MCCV of the retrospective data set. A significance level of .00015 was applied, considering Bonferroni correction for multiple comparisons.

All clinical factors were included in the model construction process, regardless of their individual correlation with hs-cTnT elevation, to account for potential intricate interactions with other features. Statistical test results for the prospective data set are presented to provide additional insight into the data set characteristics, but were not incorporated into the preselection process to prevent data leakage. Before being input into the model, categorical variables among clinical factors were encoded by using a OneHot encoder with the open-source package “scikit-learn.”

Model construction

We applied a penalized logistic regression, an elastic net logistic regression, in this study. An MCCV approach was used to randomly divide the patients from the retrospective data set into training and testing subsets with a split of 75%/25% for 100 times. Input variables were grouped as clinical factors, selected DVH parameters, selected robust radiomic features, and selected robust dosimetric features.

To assess the predictive roles of cardiac substructures and variable groups in hs-cTnT elevation, multiple models were constructed for each training–testing split: (1) clinical model included only clinical factors as a baseline; (2) WH DVH model: clinical factors combined with WH-based DVH parameters; (3) substructure (SUB) DVH model: clinical factors combined with substructure-based DVH parameters (including the WH); (4) WH radiomic model: clinical factors combined with WH-based radiomic features; (5) SUB radiomic model: clinical factors combined with substructure-based radiomic features; (6) WH dosiomic model: clinical factors combined with WH-based dosiomic features; (7) SUB dosiomic model: clinical factors combined with substructure-based dosiomic features; (8) WH hybrid model: hybrid inputs of clinical factors and all WH-based variables; and (9) SUB hybrid model: hybrid inputs of clinical factors and all substructure-based variables.

Model training and evaluation

In each iteration, variables within a specific model from the training samples were normalized to a range of 0 to 1 by using a MinMaxScaler. Multicollinearity correction was performed by eliminating features with high collinearity that could negatively influence model performance. The variance inflation factor (VIF) was used as the elimination criterion, calculated as follows:

$${VIF}_i=\frac{1}{1-R_i^2}$$ (1)

where ${\text{R}}_{\text{i}}^2$ represents the unadjusted coefficient of determination for regressing the i-th variable on the remaining variables. Variables with the highest VIFs were iteratively removed until all VIFs were below 5, corresponding to an R²< 0.8, indicating acceptable collinearity levels. Next, recursive feature elimination (RFE) was applied to select the top 10 variables for model input, balancing the number of inputs with the number of training samples. A default random forest classifier was used within the RFE process to leverage Gini importance and reduce overfitting risks in subsequent logistic regression models.

A unique elastic net logistic regression model was constructed for each of the 100 iterations by using the selected 10 features. An inner 5-fold cross-validation was used to optimize hyperparameters, including the regularization strength, L1/L2 penalty ratio, and solver type. Balanced accuracy was used as the validation metric because of data imbalance in the data set. The trained model was then applied to the testing samples by using the corresponding selected features, scaled with the same MinMaxScaler. Evaluation metrics, including sensitivity, specificity, area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve (AUPRC), were calculated based on predicted and ground-truth labels for each iteration. These metrics were aggregated across all iterations for statistical analysis. Model training, testing, data splitting, and hyperparameter tuning were implemented with the open-source package “scikit-learn.”

The MCCV approach increased data set partitioning, thus reducing estimation variability and enabling direct comparison of performance across different models. Because identical training and testing samples were used across iterations, the Wilcoxon signed-rank tests were applied, which allowed pairwise comparisons of performance metrics between models.

Feature importance

To evaluate the predictive significance of input variables and identify superior predictors of hs-cTnT elevation, permutation-based feature importance was calculated during model training in each iteration.⁴¹ Each input variable was randomly permuted 30 times, and the corresponding changes in the training balanced accuracy score were measured as importance. A substantial decrease in the training score after permutation of a specific variable indicated its importance in prediction. Permutation importance values for each variable were recorded across all 100 iterations for further analysis. Variables excluded during multicollinearity correction or RFE were assigned an importance score of zero. For the identified important variable, several common approaches were applied to determine optimal cutoff value regarding hs-cTnT elevation, including Youden index, Euclidean distance, Index of Union, maximum product of sensitivity and specificity, and minimum P value.⁴²

Holdout validation

The prospective patient data set was used as an independent holdout set to evaluate the performance of the different models. This data set was exclusively applied during holdout validation, ensuring no data leakage and minimizing potential bias. To enhance prediction robustness, an ensemble prediction approach was implemented for each model type by using the 100 elastic net logistic regression models trained during the MCCV. The final probability of hs-cTnT elevation was determined by averaging the logits from each model, calculated as follows:

$$P=\frac{1}{1+e^{--\frac{1}{100}{\sum }_{i=1}^{100}\left[{\beta }_{i,0}+{\sum }_{j=1}^{10}\left({\beta }_{i,j}{X}_{i,j}\right)\right]}}$$ (2)

Where X_i,j represents the j-th input variable in the i-th model, β_i,j represents the j-th coefficient in the i-th model, and β_i,0 represents the intercept in the i-th model. A probability threshold of .5 was used to derive the final prediction. Similarly, sensitivity, specificity, AUROC, and AUPRC were calculated based on the ensembled predictions and ground-truth labels.

Results

Patient characteristics: Retrospective versus prospective

Characteristics of the retrospective and prospective cohorts are summarized in Table 1. Chemotherapy details are omitted because all patients received the same regimen. No statistically significant differences were observed between the cohorts in terms of sex, tumor stage, tumor histology, tumor location, pre-existing heart disease, Karnofsky Performance Status score, treatment modality, or hs-cTnT elevation (the primary endpoint). Notably, compared with the retrospective cohort, prospective patients were older (67.8 vs 64.7 years; P = .037), had fewer smokers (active: 12% vs 21%; former: 65% vs 70%; never: 23% vs 9%; P = .013), higher baseline hs-cTnT levels (11.2 vs 7.9 ng/L; P = .0002), and lower prescribed radiation doses (66.7 vs 70.5 Gy; P < .0001).

Table 1.

Clinical characteristics in the retrospective and prospective datasets.

	Value or no. of patients (%)
	Retrospective (n = 160)	Prospective (n = 57)	P value
Age, y, mean (SD)	64.7 (8.9)	67.8 (10.1)	.037
Sex			.302
Female	78 (49)	33 (58)
Male	82 (51)	24 (42)
Race			.280
White	145 (91)	48 (84)
Other	15 (9)	9 (16)
Tumor stage			.070
≤IIIA	74 (46)	22 (39)
≥IIIB	86 (54)	35 (61)
Tumor histology			.927
Adeno	91 (57)	34 (60)
Squamous	52 (32)	17 (30)
Other	17 (11)	6 (10)
Tumor location			.969
Left/mediastinal	65 (41)	24 (42)
Right	95 (59)	33 (58)
Smoking status			.013
Never	14 (9)	13 (23)
Former	112 (70)	37 (65)
Active	34 (21)	7 (12)
Pre-existing heart disease			.456
No	130 (81)	43 (75)
Yes	30 (19)	14 (25)
KPS			1.000
<80	15 (9)	6 (10)
≥80	145 (91)	51 (90)
hs-cTnT baseline, ng/L, mean (SD)	7.9 (9.6)	11.2 (9.5)	<.0001
Prescribed dose, Gy, mean (SD)	70.5 (4.6)	66.7 (3.2)	<.0001
Modality			.454
Photon	98 (61)	31 (54)
Proton	62 (39)	26 (46)
hs-cTnT elevation			.904
No	109 (68)	40 (70)
Yes	51 (32)	17 (30)

Abbreviations: hs-cTnT = high-sensitivity cardiac troponin T; KPS = Karnofsky Performance Status score.

Continuous variables are reported as mean (SD), and categorical variables are reported as frequency (percentage). P values are from the Wilcoxon rank-sum test for continuous values and Pearson χ² test for categorical variables.

Patient characteristics: Photon versus proton

The incidence of hs-cTnT elevation was nonsignificantly lower in patients undergoing proton therapy in the retrospective cohort (37% vs 24%; P = .138) and comparable in the prospective cohort (29% vs 31%; P = 1.000), compared with patients undergoing photon therapy. Among the 160 retrospective patients, no significant differences in baseline characteristics were observed between the 2 treatment groups, except for smoking status (active: 14% vs 32%; former: 76% vs 61%; never: 10% vs 6%; P = .024). In the prospective data set, patients undergoing proton therapy were significantly older (72.5 vs 63.9 years; P = .001) and had a higher prevalence of adenocarcinoma histology (adenocarcinoma: 73% vs 49%; squamous cell carcinoma: 27% vs 32%; other: 0% vs 6%; P = .037). Detailed comparisons of patient characteristics are provided in Appendices E3 and E4. Given the relatively minor differences in hs-cTnT elevation rates and patient characteristics, b patients undergoing both photon and proton therapies were included in the same predictive model, with treatment modality incorporated as an input variable to assess its potential impact on cardiotoxicity prediction.

Robustness analysis and preselection of variables

After the perturbation chain analysis, 265 of 2014 radiomic features and 1131 of 2014 dosiomic features were identified as being robust against image perturbations. Detailed robust feature counts per substructure are provided in Appendix E2. During variable preselection, 2 DVH parameters, 2 radiomic features, and 2 dosiomic features were chosen for each substructure based on the statistical distribution observed in the retrospective data set. A comparative summary of preselected WH-based and LAD-based variables is presented in Table 2. The preselection process was conducted exclusively on the retrospective data set; however, to provide additional context on population differences, prospective data set statistics are also included in Table 2. In the retrospective data set, all WH-based variables demonstrated no significant association with hs-cTnT elevation (P > .001, significance level 0.00015), indicating their limited predictive value. In comparison, retrospective patients with hs-cTnT elevation (n = 51) had significantly higher LAD D1.0 cc (31.2 vs 13.8 Gy; P < .0001) and LAD V20Gy (38.7% vs 13.2%; P < .0001) compared with those without elevation (n = 109). The dosiomic feature LAD glszm_ZoneEntropy (which quantifies the complexity of zone sizes with different dose levels) also significantly distinguished the 2 patient groups. Notably, the selected dosiomic features LAD glszm_ZoneEntropy and LAD firstorder_90Percentile were strongly correlated with LAD V20Gy, with Spearman correlation coefficients of 0.84 and 0.91. A comparison between the retrospective and prospective cohorts revealed that prospective patients received significantly lower WH V20Gy (10.7% vs 21.7%; P < .0001), although no significant differences in LAD-based variables were observed (P > .001). However, both WH- and LAD-based variables showed no significant differences between prospective patients with and without hs-cTnT elevation, highlighting substantial variability within the retrospective and prospective data sets. Comprehensive tables of preselected DVH, radiomic, and dosiomic features for each cardiac substructure are presented in Appendices E5–E7. In addition, a comparison of WH and LAD DVH parameters between photon and proton patients is presented in Appendix E8.

Table 2.

Comparison of whole heart-based and left anterior descending coronary artery-based variables between patients with and without hs-cTnT elevation in the retrospective and prospective cohorts.

Variables	Retrospective (n = 160)				Prospective (n = 57)				P* (Retrospective vs prospective)
	All (n = 160)	Elevated (n = 51)	Nonelevated (n = 109)	P value	All (n = 57)	Elevated (n = 17)	Nonelevated (n = 40)	P value
WH V20Gy, %	21.7 (18.5)	28.0 (21.6)	18.7 (16.0)	.016	10.7 (11.7)	15.0 (14.1)	8.6 (9.4)	.101	<.0001
WH D60.0, Gy	4.2 (6.1)	6.5 (7.8)	3.2 (4.7)	.004	2.0 (3.3)	2.7 (4.4)	1.4 (2.1)	.820	.381
WH M2DDS (radiomic)	136.0 (16.4)	140.6 (14.8)	134.3 (16.6)	.018	133.5 (14.3)	131.7 (15.3)	134.3 (13.8)	.565	.423
WH M2DDC (radiomic)	140.0 (14.4)	141.9 (13.5)	139.3 (14.6)	.168	137.7 (12.7)	134.7 (14.0)	139.0 (11.8)	.346	.399
WH M2DDS (dosiomic)	136.0 (16.4)	140.6 (14.8)	134.3 (16.6)	.018	133.5 (14.3)	131.7 (15.3)	134.3 (13.8)	.565	.423
WH firstorder_TotalEnergy (dosiomic)	3.3 × 108   (2.9 × 108)	4.2 × 108   (3.2 × 108)	3.0 × 108   (2.8 × 108)	.035	1.4 × 108   (1.6 × 108)	1.7 × 108   (1.4 × 108)	1.2 × 108   (1.6 × 108)	.062	<.0001
LAD D1.0 cc, Gy	19.4 (21.2)	31.2 (22.6)	13.8 (18.1)	<.0001	12.4 (16.0)	15.7 (17.5)	11.1 (15.0)	.595	.042
LAD V20Gy, %	21.4 (30.8)	38.7 (36.9)	13.2 (23.5)	<.0001	11.1 (21.9)	16.2 (24.9)	9.0 (20.2)	.553	.014
LAD M2DDC (radiomic)	27.3 (10.6)	31.3 (13.6)	25.8 (8.7)	.036	27.1 (8.6)	32.0 (10.8)	25.0 (6.5)	.023	.510
LAD M2DDS (radiomic)	36.8 (11.9)	39.8 (11.4)	35.7 (11.9)	.036	36.4 (10.4)	37.7 (12.3)	35.9 (9.4)	.951	.969
LAD glszm_ZoneEntropy (dosiomic)	3.3 (1.8)	4.1 (1.4)	3.0 (1.8)	<.0001	2.6 (1.7)	2.7 (1.8)	2.6 (1.7)	.708	.013
LAD firstorder_90Percentile (dosiomic)	22.9 (22.7)	33.8 (23.0)	18.9 (21.3)	<.0001	14.1 (16.9)	17.2 (19.1)	12.7 (15.7)	.754	.012

Abbreviations: glszm = gray level size zone matrix; hs-cTnT = high-sensitivity cardiac troponin T; LAD = left anterior descending coronary artery; M2DDC = maximum 2-dimensional diameter column; M2DDS = maximum 2-dimensional diameter slice; WH = whole heart.

P values were calculated with the Wilcoxon rank-sum test, with a significance level of .00015 after Bonferroni correction for multiple comparisons. P* indicates the overall comparison between retrospective and prospective patients. Numbers were reported as mean (SD). WH firstorder_TotalEnergy (dosiomic) measures the magnitude of dose intensities within WH. LAD glszm_ZoneEntropy (dosiomic) quantifies the complexity of zone sizes with different dose levels.

It is noteworthy that factors commonly considered cardiotoxicity risk indicators, such as baseline hs-cTnT levels (8.1 vs 7.8 ng/L; P = .095) and the presence of pre-existing heart disease (19.6% vs 18.3%; P = 1.000), did not show statistically significant differences between patients with and without hs-cTnT elevation. Nonetheless, all clinical variables were retained as model inputs to account for potential complex interaction effects among predictors.

Model evaluation

The sensitivity, specificity, AUROC, and AUPRC results from MCCV and holdout validation are summarized in Table 3. The baseline clinical model achieved a mean AUROC of 0.66 in MCCV. Incorporating WH-based variables marginally improved the predictive performance of the model (AUROC: WH DVH, 0.67; P = .0003; WH radiomics, 0.67; P = .030; WH dosiomics, 0.66; P = .083). In contrast, incorporating substructure-based DVH parameters and dosiomic features significantly improved the model’s performance, yielding mean AUROCs of 0.71 (P < .0001) for clinical plus substructure DVH parameters and 0.70 (P < .0001) for clinical plus substructure dosiomic features. However, adding substructure-based radiomic features reduced the AUROC to 0.57 (P < .0001). The hybrid model incorporating substructure-based features (SUB hybrid) performed worse than the SUB DVH model (AUROC: 0.68 vs 0.71; P < .0001). In comparison, the WH hybrid model, combining WH-based DVH parameters, radiomic, and dosiomic features, achieved an AUROC of 0.70, comparable to the SUB DVH model (P = .051). Nevertheless, in holdout validation, performance of the clinical baseline model and all WH-based models was limited (AUROC: clinical, 0.45; WH DVH, 0.50; WH radiomics, 0.48; WH dosiomics, 0.48; WH hybrid, 0.51), indicating poor generalizability across data sets. Conversely, the SUB DVH model achieved an AUROC of 0.60, with a sensitivity of 0.47 and specificity of 0.75. The SUB dosiomics model performed worse (AUROC: 0.56), despite high correlations between selected dosiomic features and DVH parameters in the retrospective data set. The SUB radiomics model achieved an AUROC of 0.61; however, its lower MCCV performance suggests that its higher AUROC in holdout validation resulted from data set variability rather than true predictive strength. These trends also applied to AUPRC, which is considered more robust for imbalanced data sets.

Table 3.

Comparison of model performance in Monte Carlo cross-validation (within the retrospective cohort) and in holdout validation (within the prospective cohort).

Models	Retrospective (Monte Carlo cross-validation)				Prospective (holdout validation)
	Sensitivity	Specificity	AUROC	AUPRC	Sensitivity	Specificity	AUROC	AUPRC
Clinical	0.57 (0.55, 0.60)	0.64 (0.62, 0.66)	0.66 (0.64, 0.67)	0.50 (0.48, 0.52)	0.53	0.48	0.45	0.27
WH DVH	0.60 (0.57, 0.63)	0.65 (0.62, 0.67)	0.67 (0.66, 0.69)	0.52 (0.50, 0.54)	0.47	0.50	0.50	0.29
SUB DVH	0.64 (0.61, 0.67)	0.66 (0.64, 0.69)	0.71 (0.70, 0.73)	0.56 (0.54, 0.58)	0.47	0.75	0.60	0.39
WH radiomics	0.60 (0.57, 0.63)	0.63 (0.61, 0.66)	0.67 (0.65, 0.68)	0.51 (0.49, 0.53)	0.53	0.50	0.48	0.28
SUB radiomics	0.50 (0.47, 0.53)	0.59 (0.57, 0.62)	0.57 (0.56, 0.59)	0.45 (0.43, 0.47)	0.94	0.33	0.61	0.35
WH dosiomics	0.59 (0.56, 0.62)	0.65 (0.62, 0.67)	0.66 (0.65, 0.68)	0.51 (0.49, 0.53)	0.35	0.55	0.48	0.28
SUB dosiomics	0.64 (0.61, 0.67)	0.64 (0.62, 0.67)	0.70 (0.69, 0.71)	0.51 (0.50, 0.53)	0.35	0.70	0.56	0.35
WH hybrid	0.62 (0.59, 0.65)	0.64 (0.62, 0.66)	0.70 (0.68, 0.71)	0.55 (0.53, 0.58)	0.47	0.50	0.51	0.31
SUB hybrid	0.61 (0.58, 0.64)	0.64 (0.62, 0.67)	0.68 (0.66, 0.70)	0.52 (0.50, 0.54)	0.41	0.55	0.61	0.40

Abbreviations: AUPRC = area under the precision-recall curve; AUROC = area under the receiver operating characteristic curve; DVH = dose-volume histogram; SUB = substructure; WH = whole heart.

Values in Monte Carlo cross-validation are shown as mean (95% CI). Clinical factors were included in every model.

Feature importance

The permutation importance of input variables across different models in MCCV is summarized in Figure 3. In the clinical model, tumor location emerged as the most influential predictor for predicting hs-cTnT elevation. Similarly, tumor location remained the dominant feature across all WH-based models. WH V20Gy and WH D60.0 exhibited median importance values of 0 in the WH DVH model, indicating limited predictive relevance. In the SUB DVH model, LAD V20Gy was identified as the most critical predictor. Similarly, the SUB dosiomics model highlighted LAD glszm_ZoneEntropy as a key variable, with both features ranking among the top predictors in the SUB hybrid model. Given the superior performance of the SUB DVH model in both MCCV and holdout validation, LAD V20Gy was considered the most dominant predictor among all input variables. Optimal cutoff LAD V20Gy for hs-cTnT elevation was calculated among retrospective patients as 0.3%, 1.9%, 5.0%, 1.1%, and 0.2%, using Youden index, Euclidean distance, Index of Union, maximum product of sensitivity and specificity, and minimum P value approaches, respectively.

Fig. 3.

Importance of input variables in permutations in the (A) clinical model, (B) WH DVH model, (C) substructure (SUB) DVH model, (D) WH radiomics model, (E) SUB radiomics model, (F) WH dosiomics model, (G) SUB dosiomics model, (H) WH hybrid model, and (I) SUB hybrid model. Error bars indicate the variability in feature importance across folds of Monte Carlo cross-validation. “Prescription” refers to the prescribed dose to the target volume, and “modality” indicates the treatment type (photon or proton). Abbreviations: AA = ascending aorta; AVV = aortic valve; CS = cluster shade; DA = descending aorta; DVH = dose-volume histogram; glcm = gray level co-occurrence matrix; GLNU = gray-level nonuniformity; glrlm = gray level run length matrix; glszm = gray level size zone matrix; IVC = inferior vena cava; LAD = left anterior descending coronary artery; LAL = least axis length; LCX = left circumflex coronary artery; LV = left ventricle; M3DD = maximum 3-dimensional diameter; M2DDC = maximum 2-dimensional diameter column; M2DDR = maximum 2-dimensional diameter row; M2DDS = maximum 2-dimensional diameter slice; MVV = mitral valve; ngtdm = neighboring gray tone difference matrix; PA = pulmonary artery; pre_existing HD = pre-existing heart disease; PV = pulmonary vein; PVV = pulmonary valve; RA = right atrium; RCA = right coronary artery; RMS = root mean square; RV = right ventricle; SVC = superior vena cava; TE = total energy; TVV = tricuspid valve; WH = whole heart; ZE = zone entropy.

To reduce the risk of multicollinearity and enhance clinical interpretability, predictive models should favor a smaller set of dominant input features. A simplified model constructed using the top 5 features identified in Figure 3C (LAD V20Gy, right ventricle max, right atrium SD, superior vena cava (SVC) D70, SVC V70) achieved an AUROC of 0.74 in MCCV and 0.60 in holdout validation.

Discussion

In this study, we undertook a comprehensive comparison between WH-based and cardiac substructure-based variables for predicting hs-cTnT elevation. We evaluated the predictive performance of DVH parameters, radiomic, and dosiomic features by using both MCCV and holdout validation. Our findings demonstrate that incorporating substructure-based DVH parameters alongside clinical factors yielded the best performance in MCCV and superior robustness in holdout validation. Current clinical guidelines predominantly use WH constraints, such as mean heart dose < 26 Gy or Heart V30Gy < 46%, as recommended by the Quantitative Analyses of Normal Tissue Effects in the Clinic guidelines.⁴³ However, our findings reveal that WH-based dose metrics are poor predictors of hs-cTnT elevation, likely because of their inability to capture localized dose distributions. Notably, dose metrics specific to the LAD were more strongly associated with hs-cTnT elevation than traditional WH dose metrics, such as WH V20Gy. In our feature analysis, tumor location emerged as the most critical factor in all WH-based models, whereas WH-based variables failed to reflect specific treatment sites. Tumor location likely serves as a surrogate for substructure doses, as evidenced by its significant correlations with specific metrics, such as LAD V20Gy (Spearman r = 0.58) and SVC V70Gy (r = −0.55), and its inferior predictive value. Our results underscore the importance of sparing the LAD during lung cancer treatment planning to minimize the risk of radiation-induced cardiotoxicity.

Identification of heart dose-volume parameters has been inconsistent across different studies,⁴⁴ which could stem from the variability in the ability of the WH DVH parameters to capture substructure-specific doses. Although WH DVH parameters may correlate with substructure doses under fixed tumor locations (eg, breast radiation therapy), this relationship weakens in more heterogeneous treatment scenarios, such as lung and esophageal cancers. To mitigate this limitation, we examined a comprehensive set of 19 cardiac substructures, including chambers, great vessels, valves, and coronary arteries. This approach reduced the risk of confounding effects and enabled the identification of LAD V20Gy as a robust predictor after correction for multicollinearity. In a dosimetric study of patients with breast cancer using a ≥30% increase in hs-cTnT from baseline as the endpoint, LAD V20 and V15Gy were significantly higher among patients with elevated hs-cTnT levels (21% incidence).⁴⁵ Although that study also identified WH V20Gy as a significant discriminator, our results indicate that WH V20Gy is not a reliable predictor of hs-cTnT elevation in patients with lung cancer. This discrepancy may be partly due to the more uniform treatment setup in breast cancer, where WH V20Gy could serve as a surrogate for LAD V20Gy. Our findings further support the importance of LAD sparing, consistent with prior literature.⁴⁶

Proton therapy is generally associated with reduced cardiac dose in patients with lung cancer. In our dosimetric analysis, proton therapy resulted in significantly lower WH doses compared with photon therapy (eg, WH D60.0: 0.7 vs 6.5 Gy in the retrospective cohort; 0.3 vs 3.5 Gy in the prospective cohort; P < .0001 for both, Appendix E8). Among the retrospective patients used for model training, the LAD V20Gy was significantly lower in patients undergoing proton therapy than in patients undergoing photon therapy (16.6% vs 34.3%; P = .0001). In contrast, this difference was not statistically significant in the prospective cohort (9.0% vs 16.2%; P = .553). These findings are consistent with the slightly lower incidence (although nonsignificant) of hs-cTnT elevation observed in patients undergoing proton therapy within the retrospective cohort, and the comparable incidence rates between modalities in the prospective cohort. Overall, treatment modality did not emerge as a significant predictor in the model.

Radiomics has had significant roles in studies of outcome prediction in radiation therapy and has demonstrated potential for enhancing predictive performance.^47,48 Recently, Talebi et al⁴⁹ reported that radiomic features significantly enhanced cardiotoxicity prediction in breast cancer based on a cohort of 83 patients. However, reproducibility and generalizability remain critical challenges because of the high-dimensional feature space, which increases the risk of data overfitting. In this study, we used an image perturbation strategy to identify robust features resilient to simulated variations in image acquisition and contour delineation. As a result, only 13% of radiomic features and 56% of dosiomic features were retained before feature selection and model construction. Despite an improvement in predictive performance by incorporating radiomic and dosiomic features during MCCV (an increase in AUROC from 0.67-0.70), this enhancement did not translate to holdout validation (AUROC: 0.50 vs 0.51). Moreover, although selected dosiomic features were highly correlated with substructure DVH parameters, the SUB dosiomic model exhibited limited robustness in holdout validation. These findings underscore the necessity of validating radiomics-based models across independent data sets to ensure their clinical applicability. Variability in imaging protocols, treatment planning quality, and delivery processes may further complicate the reliable application of prediction models in clinical practice.

In this study, we used a clinically validated deep learning–based auto-segmentation model to delineate cardiac substructures from planning CT images. This model demonstrated high geometric and dosiomic accuracy, offering several advantages for dosimetric studies. Notably, it reduced interobserver variability and improved the consistency of dosimetric metrics. Manual segmentation of cardiac structures, particularly small substructures such as the LAD, is prone to substantial variability.⁵⁰ The use of auto-segmentation not only enhances contour consistency but also improves model performance, as evidenced by recent studies demonstrating superior prediction accuracy using auto-contours compared with manual contours.⁵¹ The adoption of auto-segmentation in both retrospective and prospective patient cohorts also minimized potential deviations caused by the decade-long gap between updates of the contouring guidelines. Moreover, auto-segmentation mitigated the labor-intensive and time-consuming process of manual contouring, facilitating large-scale retrospective dosimetric analyses. Our results hold promise for the application of auto-segmentation models in advancing future large-scale cardiotoxicity research.

One acknowledged limitation of the current study is that it was conducted at a single institution, which may limit the generalizability of the findings. Expanding the analysis to a multi-institutional data set would enhance the robustness and external validity of the results. Nonetheless, the inclusion of newly enrolled prospective patients as a holdout validation cohort improved the reliability of the model evaluation. Another limitation is the relatively small number of patients, which was driven by the requirement for hs-cTnT assays and thereby restricting the number of evaluable cases. To address this constraint, we used MCCV to reduce random biases introduced during data splitting and to enable direct pairwise comparisons between models incorporating different variables.

Our results demonstrated moderate cross-validation AUROC and low holdout validation AUROC, even with the best-performing model, suggesting the presence of unaccounted underlying factors. The elevation of hs-cTnT may not be solely attributable to radiation dose and the clinical variables included here, underscoring the need for broader patient data curation in clinical trials to enhance prediction models. Moreover, hs-cTnT elevation is not always a definitive indicator of underlying cardiac disease.²⁵ Accurately assessing adverse cardiac events requires long-term follow-up of prospectively enrolled patients, which is ongoing. We plan to validate our workflow by using actual cardiac adverse events as outcomes in the future. In this study, a fixed RBE of 1.1 was applied for proton dose calculations. This simplification may introduce uncertainties in outcome analysis and potentially underestimate the cardiovascular risk associated with proton therapy. Further investigation into the tissue-specific RBE of protons, particularly in cardiovascular structures, is warranted to determine whether incorporating variable RBE values could improve model performance and cardiotoxicity risk prediction. Despite these challenges, predicting hs-cTnT elevation enables early intervention to improve treatment plan quality, which ultimately benefits patients. Future research will focus on integrating this predictive model into treatment planning to optimize radiation therapy plans, aiming to reduce cardiotoxicity risk and evaluate the associated clinical benefits.

Conclusions

We demonstrated that cardiac substructure DVH parameters provide superior predictive power and robustness compared with WH-based variables for predicting hs-cTnT elevation in radiation therapy for non-small cell lung cancer. LAD V20Gy emerged as a significant predictor of elevated hs-cTnT levels and perhaps of radiation-induced cardiotoxicity. The optimal cutoff value of LAD V20Gy ranged from 0.2% to 5% using different approaches. This model offers valuable insights for cardiotoxicity risk assessment, facilitating the development of personalized treatment plans aimed at minimizing cardiotoxicity in affected patients.

Supplementary Material

Supplementary material associated with this article can be found in the online version at doi:10.1016/j.ijrobp.2025.10.008.

Data Sharing Statement:

The data can be made available on reasonable request to Zhongxing Liao (zliao@mdanderson.org).