A neural network model to predict lung radiation‐induced pneumonitis

Abstract

A feed-forward neural network was investigated to predict the occurrence of lung radiation-induced Grade 2+ pneumonitis. The database consisted of 235 patients with lung cancer treated using radiotherapy, of whom 34 were diagnosed with Grade 2+ pneumonitis at follow-up. The network was constructed using an algorithm that alternately grew and pruned it, starting from the smallest possible network, until a satisfactory solution was found. The weights and biases of the network were computed using the error back-propagation approach. Momentum and variable leaning techniques were used to speed convergence. Using the growing/pruning approach, the network selected features from 66 dose and 27 non-dose variables. During network training, the 235 patients were randomly split into ten groups of approximately equal size. Eight groups were used to train the network, one group was used for early stopping training to prevent overfitting, and the remaining group was used as a test to measure the generalization capability of the network (cross-validation). Using this methodology, each of the ten groups was considered, in turn, as the test group (ten-fold crossvalidation). For the optimized network constructed with input features selected from dose and nondose variables, the area under the receiver operating characteristics (ROC) curve for cross-validated testing was 0.76 (sensitivity: 0.68, specificity: 0.69). For the optimized network constructed with input features selected only from dose variables, the area under the ROC curve for cross-validation was 0.67 (sensitivity: 0.53, specificity: 0.69). The difference between these two areas was statistically significant (p=0.020), indicating that the addition of non-dose features can significantly improve the generalization capability of the network. A network for prospective testing was constructed with input features selected from dose and non-dose variables (all data were used for training). The optimized network architecture consisted of six input nodes (features), four hidden nodes, and one output node. The six input features were: lung volume receiving >16 Gy (V 16 ), generalized equivalent uniform dose (gEUD) for the exponent a=1 (mean lung dose), gEUD for the exponent a=3.5, free expiratory volume in 1 s (FEV 1 ), diffusion capacity of carbon monoxide (DLCO%), and whether or not the patient underwent chemotherapy prior to radiotherapy. The significance of each input feature was individually evaluated by omitting it during network training and gauging its impact by the consequent deterioration in cross-validated ROC area. With the exception of FEV1 and whether or not the patient underwent chemotherapy prior to radiotherapy, all input features were found to be individually significant (p<0.05). The network for prospective testing is publicly available via internet access.

Modelling of Normal Tissue Complication Probabilities (NTCP): Review of Application of Machine Learning in Predicting NTCP

“…In a publication complimentary to their work using neural networks (discussed in the previous section), Chen et al describe using support vector machines to predict pneumonitis [6] on the same dataset reported for ANN [8]. A radial basis kernel function was chosen for the SVM in preference to a sigmoid or polynomial kernel as the increase in free parameters might result in overfitting.…”

“…In 2007, Chen et al [8] reported results for a larger cohort of lung cancer patients from the same institution, Duke University Medical Centre, North Carolina. Radiation-induced pneumonitis (≥ grade 2) was reported in 34 out of 235 patients, all of whom were treated using 3D conformal radiotherapy.…”

The relationship between pulmonary function metrics and radiation-induced lung injury

“…The importance of these clinical factors have led to the development of complex neural network modeling algorithms that incorporate dosimetric as well as clinical parameters in order to most accurately predict RILI risk [9][10][11] , but these models seem too complicated and unwieldy to be clinically useful or practicable for most radiation oncologists. A recent study proposes a new metric, TFS5, which combines the volume of lung spared from receiving more than 5 Gy and the transfer coefficient for carbon monoxide (KCO) [14] .…”

“…A number of predictive models have been posited to ascertain which patients are likely to develop RILI given similar radiation doses and treatment parameters [2][3][4][5][6][7][8][9][10][11][12][13][14] .…”

A genetic algorithm for variable selection in logistic regression analysis of radiotherapy treatment outcomes

“…A more integrated approach of treatment plan evaluation that includes all these factors has recently sparked a growing interest in the radiotherapy community, creating a need for the development of robust data mining and correlation analysis methods. Software tools were developed to address this issue, using, for example, self-organizing maps, 6 support vector machine algorithms, 7 neural networks, 8 or decision trees. 9 Two of these tools, DREES 10 and EUCLID, 11 which are MATLAB-based programs ͑The MathWorks Inc., Natick, MA͒, use a logistic regression model to correlate treatment outcomes with clinical factors.…”