Neural network predicts three-year local recurrence after lung SBRT with higher accuracy than standard models

An interpretable artificial neural network can predict whether a lung tumor will return within three years after stereotactic body radiation therapy, according to a retrospective study conducted at the Cancer Hospital of Shantou University Medical College in China. The model, built from routinely collected clinical and dosimetric variables, achieved an area under the receiver operating characteristic curve of 0.892 for the six-variable version of the model, outperforming both a conventional logistic regression approach and three established radiobiological models in the same cohort. The findings, reported by a team led by Bao-Tian Huang, suggest that a relatively simple machine learning tool could eventually help clinicians identify patients at elevated risk of local failure after this increasingly common form of radiation treatment.

Stereotactic body radiation therapy, or SBRT, delivers very high radiation doses in a small number of fractions, concentrating biologically effective dose on the tumor while sparing nearby healthy tissue. The approach has become a standard option for patients with inoperable primary lung cancer, and studies using biologically effective doses above 100 Gy have reported local control rates exceeding 80 percent at three years. Yet local recurrence still occurs. Two independent studies cited by the authors reported recurrence rates of 13.0 percent and 18.8 percent after SBRT, underscoring the need for reliable ways to stratify patients before treatment begins.

The research team assembled a cohort of patients with primary and metastatic lung tumors treated with SBRT between July 2011 and December 2018. Of 165 tumors in 164 patients, 127 tumors with complete follow-up data were used to develop the three-year local recurrence prediction models. During the three-year follow-up window, 65 tumors, or 51.2 percent, experienced local recurrence, while 62 tumors, or 48.8 percent, remained locally controlled. The authors note that this recurrence rate is higher than in many published SBRT series because a substantial number of metastatic patients in the cohort received lower, palliative doses, such as 20 to 35 Gy delivered in 4 to 7 fractions, which fall below the biologically effective dose of roughly 100 Gy generally associated with strong local control in stage I disease.

The researchers collected an unusually broad set of candidate predictors. Clinical variables included gender, age, Karnofsky performance status, clinical stage, tumor location, histology, and tumor origin, along with blood markers such as lymphocyte and neutrophil counts, platelet counts, the neutrophil-to-lymphocyte ratio, the platelet-to-lymphocyte ratio, and hemoglobin concentration. Imaging and dosimetric variables included gross and planning target volumes, prescribed dose, the prescribed dose expressed as biologically effective dose using the linear-quadratic model with an alpha/beta ratio of 10 Gy, the maximum biologically effective dose within the planning target volume, treatment duration, fractional dose, number of fractions, maximum tumor diameter, equivalent diameter, and Hounsfield unit statistics derived from the planning computed tomography scans. CT images were acquired with slice thickness below 3 millimeters, using either respiratory-gated four-dimensional CT or conventional three-dimensional CT.

Variable selection proceeded in two stages. Univariate logistic regression first flagged parameters associated with three-year local recurrence at a significance level of P less than 0.10, and backward stepwise multivariable logistic regression guided by the Akaike information criterion then identified the optimal subset. This AIC-based strategy allowed variables that did not reach the conventional P less than 0.05 threshold to remain in the model when their inclusion improved overall fit. Multicollinearity was assessed with variance inflation factors, and when maximum diameter and equivalent diameter proved highly collinear, with variance inflation factors of 10.234 and 18.121 respectively, only maximum diameter was retained. The resulting first model used four predictors: clinical stage, planning target volume, lymphocyte count, and maximum tumor diameter. A second model added the maximum biologically effective dose in the planning target volume and tumor histology, two factors previously linked to local control after SBRT.

The neural network itself was a multilayer perceptron trained with backpropagation, implemented in dedicated clinical prediction software, with a weight decay parameter of 0.6 to control model complexity. Internal validation relied on 10-fold cross-validation repeated five times, and performance estimates were refined through 500 bootstrap resampling iterations to generate 95 percent confidence intervals. The second model reached an AUC of 0.892 (95 percent CI, 0.838 to 0.944), significantly higher than the first model’s AUC of 0.791 (95 percent CI, 0.712 to 0.856). The second model also posted a Brier score of 0.172, accuracy of 0.843, sensitivity of 0.815, specificity of 0.871, precision of 0.869, and an F1 score of 0.841. Calibration plots showed general agreement between predicted and observed probabilities, and decision curve analysis demonstrated net benefit across a range of clinically reasonable threshold probabilities.

A key feature of the study is its use of SHapley Additive exPlanations, or SHAP, to open the black box of the neural network. SHAP values, calculated on the logit scale, quantify how much each predictor pushes an individual patient’s prediction above or below the cohort baseline. In a representative patient whose tumor remained controlled, clinical stage, planning target volume, maximum biologically effective dose, and maximum diameter all contributed negative values that shifted the model output from a baseline logit of 0.30 down to a final logit of 0, corresponding to a prediction of local control. In a representative patient whose tumor recurred, the maximum biologically effective dose, planning target volume, and lymphocyte count contributed strongly positive values, shifting the output from the same baseline to a final logit of 1.0 and a prediction of local recurrence. This patient-level transparency is intended to make the model’s reasoning inspectable by clinicians rather than leaving it opaque.

To benchmark the neural network against established approaches, the team compared it with a conventional logistic regression model built on the same data and with three published radiobiological models of three-year local control: the Liu, Klement, and Gucken models. Because those models estimate local control rather than recurrence, recurrence probability was derived as one minus the predicted control probability. The logistic regression model achieved an AUC of 0.764, while the Liu, Klement, and Gucken models reached AUCs of 0.664, 0.622, and 0.660 respectively. Pairwise DeLong tests showed that the six-variable neural network significantly outperformed every comparator, and calibration plots revealed substantial discrepancies between predicted and observed probabilities for all three radiobiological models, whereas the neural network and logistic regression models tracked observed outcomes more closely.

The authors situate their results within a broader landscape of prediction efforts. Prior work includes a logistic regression model for one-year local control with training and validation AUCs of 0.845 and 0.786, and a support vector machine reported by Klement and colleagues with an AUC of 0.789. Compared with radiomics, which requires laborious feature selection and has been shown to suffer from limited repeatability and reproducibility, and compared with deep learning, which demands large computational resources and long training times, the neural network approach uses variables that are easily accessible in routine clinical workflows and trains quickly on modest hardware. The authors argue that this combination of accessibility, efficiency, and demonstrated discrimination makes the method particularly suitable for eventual clinical integration.

The study carries important limitations. It was retrospective and conducted at a single center, which may limit generalizability to broader populations, and multi-institutional validation is required. The sample, though meeting the common rule of thumb of at least ten events per variable, remains small and vulnerable to overfitting. The cohort mixed primary and metastatic lung tumors, creating heterogeneity, and although prior work by Guckenberger and colleagues suggested comparable dose-response relationships for local control between primary lung cancers and lung metastases, the effect of tumor origin on model performance remains uncertain. The authors also emphasize that their intended target population was patients with lung tumors treated with SBRT rather than only those with early-stage primary lung cancer, and that subgroup and sensitivity analyses await larger cohorts. Notably, histology and maximum biologically effective dose were not statistically significant in the multivariable analysis, in contrast to several earlier studies focused on early-stage patients, a discrepancy the authors attribute to differences in patient cohorts. Even so, adding these two variables to the four-variable model raised the AUC from 0.791 to 0.892, and the team calls for multi-institutional studies to clarify their role. External validation in independent patient cohorts remains the essential next step before such a model could inform treatment decisions.

Subject of Research: Prediction of three-year local recurrence after lung stereotactic body radiation therapy using an interpretable artificial neural network

Article Title: Application of an interpretable artificial neural network for 3-year local recurrence prediction in lung cancer patients treated with stereotactic body radiation therapy

Article References: Huang, B.-T., Wang, S.-Y., Lin, P.-X., Luo, L.-M., Wang, Y., Chen, J.-Z., & Chen, C.-Z. (2026). Application of an interpretable artificial neural network for 3-year local recurrence prediction in lung cancer patients treated with stereotactic body radiation therapy. Heliyon, 12(16), Article e45582. https://doi.org/10.1016/j.heliyon.2026.e45582

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45582

Keywords: artificial neural network, stereotactic body radiation therapy, lung cancer, local recurrence, prediction model, SHAP, radiobiological models, machine learning, radiation oncology, biologically effective dose, model interpretability, clinical prediction