A machine learning approach for predicting radiation-induced hypothyroidism in patients with nasopharyngeal carcinoma undergoing tomotherapy

PatientsThis examine has been permitted by the ethics committee of Fujian Cancer Hospital (ethics quantity: YKT2020-011-01) and all patients offered written knowledgeable consent previous to enrollment in the examine. All strategies have been carried out in accordance with the Declaration of Helsinki in addition to related pointers and rules. The examine included 219 patients identified with nasopharyngeal carcinoma who underwent tomotherapy at Fujian Cancer Hospital from January 2017 to December 2020. The following medical data was collected: affected person age, gender, TNM stage, and the pre-treatment thyrotropin-stimulating hormone (TSH) ranges. The patients have been randomly divided right into a coaching set (n = 175) and a check set (n = 44) in an 8:2 ratio. There have been no statistically important variations in medical traits between the 2 teams of patients.Inclusion standards: (1) Patients with newly identified nasopharyngeal carcinoma confirmed by pathology; (2) Patients in want of radical radiotherapy; (3) Patients aged between 18 and 70 years; (4) Patients with regular thyroid perform and no underlying thyroid-related ailments; (5) Patients with a PS (ECOG standards) rating of 0 to 1.Exclusion standards: (1) Patients who underwent earlier head and neck radiotherapy or thyroid surgical procedure; (2) Patients with earlier malignant tumors; (3) Patients with extreme cardiovascular ailments or different underlying circumstances which will have an effect on the usual remedy of nasopharyngeal carcinoma. (4) Patients who didn’t have an entire follow-up end result for thyroid perform evaluation.Thyroid perform assessmentBefore remedy, ranges of whole triiodothyronine (TT3), whole thyroxine (TT4), free triiodothyronine (FT3), FT4, TSH, thyroglobulin antibody (TGAb), thyroid peroxidase antibody (TPOAb), and thyroglobulin (TG) are measured utilizing chemiluminescence evaluation to exclude underlying thyroid-related ailments. Thyroid hormone ranges, equivalent to TT3, TT4, FT3, and FT4, play a essential position in the regulation of metabolism and power ranges. TSH is secreted by the pituitary gland to supervise the thyroid’s hormone secretion. TGAb and TPOAb are linked to autoimmune thyroid issues equivalent to Hashimoto’s and Graves’ illness. TG is a protein essential for thyroid hormone synthesis. During the follow-up earlier than, on the finish of, and after radiotherapy, ranges of FT4, FT3, and TSH are decided utilizing chemiluminescence evaluation.Follow-up visits encompass a mixture of outpatient examinations and phone communication. Within the primary two years after finishing remedy, follow-up visits happen each three months. From 3 to five years after remedy, month-to-month follow-up visits are carried out. During every follow-up go to, ranges of FT4, FT3, and TSH are checked. The final follow-up in this examine came about in June 2023, and the first analysis indicator was main hypothyroidism (HT), outlined as a serum TSH stage better than 5.6 μU/mL with or with no lower in FT4 levels29. The incidence time of RIHT was outlined because the time interval between the top of radiotherapy and the primary recorded irregular TSH stage.Image acquisition, contouring, and radiation dose calculationBased on the eighth version of the UICC/AJCC staging system30, patients in stage I bear healing radiotherapy, stage II patients obtain mixed chemotherapy and radiotherapy, and stage III–IVB patients bear mixture remedy, all of which embrace radiation remedy. Positioning CT scans are acquired utilizing the Philips Brilliance Big Bore CT. Patients are positioned supine and immobilized utilizing thermoplastic masks and customised foam. The tube voltage is about at 120 kV, X-ray tube present is 225 mA, CT scan slice thickness is 3 mm, and the scan decision is 512 × 512 pixels.In accordance with the rules RTOG0225 and RTOG0615, skilled radiation oncologists with over 5 years of expertise delineate GTV, CTV, and PTV goal areas. The organs in danger, such because the thyroid, are individually delineated by two junior radiation oncologists, every with a minimal of two years of expertise, and the ultimate delineation is verified by senior radiation oncologists. The prescribed radiotherapy doses have been as follows: GTV: 70–72.6 Gy/31–33 fractions, CTV1: 62–62.7 Gy/31–33 fractions, and for CTV2: 54.4–56.2 Gy/31–33 fractions. The dose limitation of organs in danger (OARs) listed in supplementary materials Table S1.All patients bear intensity-modulated radiation remedy utilizing the Accuray TomoHD helical tomotherapy system (Accuray Inc., Madison, Wisconsin) for remedy planning. The radiation power used is 6 MV, the dose fee is 850 MU/min, and the dose calculation algorithm employed is the convolution/superposition (C/S) algorithm throughout the tumor remedy planning system. The voxel spatial decision for dose calculation is 0.273 × 0.273 × 0.3 cm3. The main goal of the remedy plan is to ship enough and constant dose to the planning goal quantity (PTV) whereas minimizing radiation publicity to OARs.Radiomics and dosiomics options extractionFeature extraction is the method of calculating a lot of particular parameters from a area of curiosity (ROI). This examine makes use of the Pyradiomics open-source bundle, which relies on the Python 3.7 platform, to extract radiomic and dosiomics options. The extracted radiomic options are categorized into three teams: first-order statistical options, form options, and texture options. First-order statistical options point out adjustments in symmetry, uniformity, and native depth distribution throughout the measured ROI area. Shape options present quantitative descriptions of the three-dimensional measurement and morphological data of the ROI area. Texture options mirror the spatial association of grayscale values throughout the ROI area. For detailed descriptions of every function sort, please discuss with the official Pyradiomics documentation31. A whole of 107 radiomic options and 107 dosiomics options have been extracted from every affected person, respectively.Features choice and mannequin constructing of radiomics and dosiomicsThe Kolmogorov–Smirnov check is utilized to evaluate the normality of function distribution. When the computed p-value of the check statistic is beneath the predefined significance threshold (usually 0.05), the null speculation is rejected, signifying a departure from regular distribution. In distinction, if the p-value exceeds the importance stage, the null speculation is upheld, indicating conformity to a traditional distribution. Then, we utilized the Student’s t-test to investigate options that comply with a traditional distribution, and the Mann–Whitney U check to investigate options that don’t comply with a traditional distribution. The null speculation for the Student’s t-test states no important distinction between the technique of the 2 teams in contrast for a particular function, whereas the null speculation for the Mann–Whitney U check asserts no important distinction between the distributions of the 2 teams in contrast for a selected function. When the p-value from the Student’s t-test or Mann–Whitney U check exceeds 0.05 for two units of options, there isn’t a important distinction between them, therefore one function should be randomly eliminated. Additionally, we used the Spearman rank correlation coefficient to measure the correlation between extremely correlation options. If the correlation coefficient between any two options exceeded 0.9, we retained solely one of many options. To protect the descriptive means of the options, we carried out a grasping recursive function elimination technique, eliminating the function with the best correlation in every iteration. Lastly, we employed the Least Absolute Shrinkage and Selection Operator (LASSO) regression mannequin to assemble options in the dataset. By adjusting the regularization weight lambda, the LASSO regression mannequin units all regression coefficients to zero, ensuing in many coefficients of irrelevant options changing into zero. To decide the optimum lambda worth that minimizes the standardized error and achieves the bottom cross-validation error, we carried out tenfold cross-validation. The remaining non-zero coefficient options have been mixed with radiomics options for the regression mannequin. By linearly combining the remaining options and their mannequin coefficient weights, we obtained a radiomics rating for every affected person. We carried out the LASSO regression mannequin utilizing the Python scikit-learn library. After Lasso function screening, we enter the ultimate options into the XGBoost machine learning fashions and so forth for threat mannequin building. The elementary idea of the XGBoost mannequin is to combine quite a few weak classifiers (resolution bushes) to type a sturdy classifier. Each resolution tree is educated on the residuals of the previous tree, progressively diminishing the residuals by iteratively optimizing the loss perform. Simultaneously, the mannequin mitigates the danger of overfitting by managing the bushes’ complexity and implementing regularization phrases. In this examine, the XGBoost is fed with various function sorts as enter, and it produces the chance worth of RIHT incidence as output. Here, we undertake fivefold cross-verification to acquire the ultimate rad signature.Clinical and DVH mannequin constructingThe examine incorporates age, gender, T stage (1, 2, 3, 4), N stage (0, 1, 2), TNM stage (1, 2, 3, 4), and the pre-treatment TSH worth as clinically related options. The DVH options encompass the max dose (Dmax), the min dose (Dmin), the imply dose (Dmean), i cubic centimeter of the utmost dose (Dicc, with i starting from 1 to 10 at an interval of 1), and share of quantity that has acquired at the very least j Gy radiation (VjGy with j starting from 5 to 65 at an interval of 5), ensuing in a complete of 23 thyroid quantity options. The building technique of medical signatures intently resembles that of rad signatures. Initially, baseline statistics (the p-value from the Student’s t-test or Mann–Whitney U check) have been used to pick the options for developing the medical traits. Furthermore, the identical machine learning mannequin was utilized in the development technique of medical and DVH signatures. To guarantee equity in comparability, fivefold cross-validation and a hard and fast experimental queue have been adopted.Combined modelAccording to Fig. 1, the medical options, DVH parameters, radiomics options, and dosiomics options have been built-in as inputs to the composite mannequin. Additionally, priceless options have been chosen utilizing Student’s t-test or Mann–Whitney U check and LASSO regression, after which inputted into the XGBoost machine learning mannequin for thyroid toxicity prediction. Similarly, fivefold cross-validation and a hard and fast experimental queue have been adopted.Figure 1The workflow of the mixed mannequin building.Statistical evaluationIn order to evaluate the diagnostic efficiency, we carried out checks in an experimental cohort. In order to look at the equivalence of affected person attributes between totally different cohorts, we employed unbiased t-tests to investigate usually distributed information and utilized the Mann–Whitney U check to symbolize non-normally distributed information utilizing medians (interquartile ranges). For categorical variables, we used the Chi-square check for evaluation. Additionally, we evaluated the predictive efficiency of the three fashions utilizing receiver working attribute (ROC) curves, the place we calculated the world beneath the ROC curve (AUC), in addition to the trade-off between sensitivity and specificity on the most Youden index. Furthermore, we evaluated the efficiency of those three fashions in the coaching and testing cohorts and assessed the medical utility of the radiomics-clinical mannequin utilizing resolution curve evaluation (DCA).Statistical analyses have been carried out utilizing SPSS (model 21.0; IBM Corp.) and the “One-key AI” platform (https://www.medai.icu), which relies on Pytorch 1.8.0. Statistical significance was outlined as a two-sided p-value ≤ 0.05.


Recommended For You