Development and application of an elastic net logistic regression model to investigate the impact of cardiac substructure dose on radiation-induced pericardial effusion in patients with NSCLC

Figures & data

Figure 1. Overview of the model construction and analysis workflow.

Figure 2. Boxplot of (A) mean and (B) max dose for the whole heart and each cardiac substructure, grouped according to PCE grade (grade 0, n = 71; ≥grade 2, n = 70). Average DVH plots of the study cohort, as grouped according to PCE incidence (blue, grade 0; red ≥ grade 2), for the (C) PTV, (D) whole heart, (E) left atrium, (F) right atrium, (G) left ventricle, and (H) right ventricle are shown. The solid line represents the mean DVH value and the shaded regions indicate the 95% CIs.

Table 2. Final PCE prediction model.

Variable	Mean ± SD	Coefficient (β)
Intercept	–	−1.370
Gender (F)	–	−0.021
RUL tumor location	–	−0.038
Cardiovascular Issues	–	−0.156
Age (yrs)	–	−0.009
WH Mean Dose	20.64 ± 7.66	0.026
WH V55	17.88 ± 8.40	0.931
WH V60	15.28 ± 7.65	2.013
WH V65	11.33 ± 6.88	0.823
WH V70	7.07 ± 6.51	2.016
LA Volume	67.59 ± 23.23	−0.007
LA Mean Dose	23.16 ± 16.07	0.008
LA V5	63.59 ± 33.02	0.473
LA V20	39.48 ± 28.80	0.134
LA V25	35.25 ± 27.57	0.342
LA V30	31.68 ± 26.41	0.288
LA V35	28.67 ± 25.31	0.089
LA V55	18.24 ± 19.96	0.072
LA V60	15.27 ± 17.97	0.373
LA V65	10.76 ± 14.32	0.043
RV Max Dose	34.94 ± 22.73	0.012
Hyperparameter	Datum
Shrinkage term (λ)	0.0803
Mixing term (α)	0.5477
Performance Metric	Datum
AUC	0.879
Accuracy	0.801
Calibration slope	1.352
Calibration intercept	−0.174

The final fitted elastic net logistic regression model was constructed using all 141 patients. SD: Standard deviation; RUL: Right upper lobe; WH: Whole heart; LA: left atrium; RV: right ventricle; AUC: Area under the curve.

Table 1. Results of the elastic net regression modeling analysis.

Evaluation metric	Heart-only model type (A)	Heart +  substructures model type (B)	Iterations w/higher metric values	p Value (A < B)
AUC	0.799	0.820	74% (B > A)	.00000636
Calibration intercept	−0.729	−0.235	78% (B > A)	.000000677
Calibration slope	2.456	1.356	78% (B > A)	.00000265
Accuracy	0.713	0.743	61% (B > A)	.0217

The median performance metric values across all 100 iterations of cross-validation are reported for each model type. The standard heart-only model type (denoted ‘A’), which can only include DVH metrics from the whole heart dose, is compared to the model types that can include DVH metrics from the whole heart and/or cardiac substructures doses (denoted ‘B’). The percentages of cross-validation iterations in which model type B outscored model type A is shown. The p-values listed were computed from the paired Wilcoxon signed-rank test. AUC: area under the curve.

Figure 3. The percent occurrence of covariates from the 100 iterations of the toxicity modeling for the (A,B) heart-only and (C–H) heart plus all substructures model types.