Machine Learning-Based Multiomics Prediction Model for Radiation Pneumonitis
Lu Zhou, Yue‐Feng Wen, Guoqian Zhang, Linjing Wang, Shuyu Wu, Shuxu Zhang
Abstract
Objective. The study aims to establish and validate an effective CT-based radiation pneumonitis (RP) prediction model using the multiomics method of radiomics and EQD2-based dosiomics. Materials and Methods. The study performed a retrospective analysis on 91 nonsmall cell lung cancer patients who received radiotherapy from 2019 to 2021 in our hospital. The patients with RP grade ≥1 were labeled as 1, and those with RP grade < 1 were labeled as 0. The whole lung excluding clinical target volume (lung-CTV) was used as the region of interest (ROI). The radiomic and dosiomic features were extracted from the lung-CTV area’s image and dose distribution. Besides, the equivalent dose of the 2 Gy fractionated radiation (EQD2) model was used to convert the physical dose to the isoeffect dose, and then, the EQD2-based dosiomic (eqd-dosiomic) features were extracted from the isoeffect dose distribution. Four machine learning (ML) models, including DVH, radiomics combined with DVH (radio + DVH), radiomics combined with dosiomics (radio + dose), and radiomics combined with eqd-dosiomics (radio + eqdose), were established to construct the prediction model via eleven different classifiers. The fivefold cross-validation was used to complete the classification experiment. The area under the curve (AUC) of the receiver operating characteristics (ROC), accuracy, precision, recall, and F1-score were calculated to assess the performance level of the prediction models. Results. Compared with the DVH, radio + DVH, and radio + dose model, the value of the training AUC, accuracy, and F1-score of radio + eqdose was higher, and the difference was statistically significant <a:math xmlns:a="http://www.w3.org/1998/Math/MathML" id="M1"> <a:mfenced open="(" close=")" separators="|"> <a:mrow> <a:mi>p</a:mi> <a:mo><</a:mo> <a:mn>0.05</a:mn> </a:mrow> </a:mfenced> </a:math> . Besides, the average value of the precision and recall of radio + eqdose was higher, but the difference was not statistically significant <f:math xmlns:f="http://www.w3.org/1998/Math/MathML" id="M2"> <f:mfenced open="(" close=")" separators="|"> <f:mrow> <f:mi>p</f:mi> <f:mo>></f:mo> <f:mn>0.05</f:mn> </f:mrow> </f:mfenced> </f:math> . Conclusion. The performance of using the ML-based multiomics method of radiomics and eqd-dosiomics to predict RP is more efficient and effective.