Analyzing Secondary Cancer Risk: A Machine Learning Approach PDF

**Analyzing Secondary Cancer Risk: A Machine Learning Approach** **E. Hatamabadi Farahani^a^, H. Sadeghi^a^, F. Seif ^b,\*^, M. Azad Marzabadi^a^, R. Rezaee^a^** *^a^ Department of Physics, Faculty of Sciences, Arak University, Arak 38156-8-8349, Iran* *^b^ Department of Radiotherapy and Medical...

**Analyzing Secondary Cancer Risk: A Machine Learning Approach** **E. Hatamabadi Farahani^a^, H. Sadeghi^a^, F. Seif ^b,\*^, M. Azad Marzabadi^a^, R. Rezaee^a^** *^a^ Department of Physics, Faculty of Sciences, Arak University, Arak 38156-8-8349, Iran* *^b^ Department of Radiotherapy and Medical Physics, Arak University of Medical Sciences, Arak, Iran* *^\*^Corresponding author email: s.medphy\@gmail.com (Fatemeh Seif)* **Abstract** Addressing the rising cancer rates through timely diagnosis and treatment is crucial. Additionally, cancer survivors need to understand the potential risk of developing secondary cancer (SC), which can be influenced by several factors including treatment modalities, lifestyle choices, and habits such as smoking and alcohol consumption. Machine learning (ML) models have demonstrated their usefulness in forecasting the likelihood of SC risks based on effective doses in the organ, demonstrating their significance in the field of oncology. Linear regression analysis is a widely utilized technique for examining the relationship between predictor variables and continuous responses, particularly in scenarios with limited sample sizes. This study aims to establish a novel relationship using the linear regression models between dose and the risk of SC, comparing different prediction methods for lung, colon, and breast cancer. The results indicate that the risk of SC increases with the effective dose in the organ, with the linear regression model providing coefficients that mirror the radiation sensitivity of the specific organ. **Keywords:** radiation therapy, second cancer risk, machine learning, regression models. **1. Introduction** Today\'s society is facing a significant challenge with the rising number of individuals diagnosed with cancer. The uncontrolled proliferation of malignant cells can result in the development of cancer. However, advancements in medical science have enabled the timely diagnosis and treatment of cancer, to minimize mortality rates associated with the disease. While some cancer survivors may remain disease-free following initial treatment, others may experience non-cancer-related health issues and side effects from the treatment \[1,2\]. A major concern for individuals who have undergone cancer treatment is the possibility of cancer recurrence. All cancer survivors need to be aware of the potential for developing SC following treatment for the initial cancer. This SC is distinct from the primary cancer in terms of its origin and pathology \[3,4\]. It can manifest in the same organ or area of the body as the initial cancer, or in a completely different organ. It is crucial to understand that SC is not indicative of metastasis from the primary cancer \[5\]. Factors such as treatment methods, smoking, alcohol consumption, and overall lifestyle can contribute to the development of SC. The efficacy of the treatment method plays a crucial role in the development of SC \[6,7\]. Radiation therapy, which involves the use of ionizing radiation to induce breaks in DNA, effectively halts the growth of cancer cells and leads to their destruction. By targeting cells that exhibit uncontrolled proliferation, this method results in the eradication of the disease \[8.9\]. However, the relationship between radiation dosage and the likelihood of cancer must be considered, as this treatment approach can potentially give rise to SC. The scattering of radiation during therapy may impact healthy organs in the body, leading to damage. Organs such as the thyroid and breast, which are particularly sensitive to radiation, are at a higher risk of developing SC. Therefore, it is imperative to assess the risk of SC considering these factors \[10,11\]. There are various approaches to assessing the risk of SC. One of the primary methods involves the use of computational models to calculate the excess relative risk (ERR) and the absolute excess risk (EAR). Additionally, nuclear simulator codes are utilized to determine the effective dose in the organ, which is then used to calculate the risk of SC. Another method involves cohort studies, where databases and patient data are used to estimate the risk of SC among individuals who have undergone primary cancer treatment. This method typically involves studying a significant population within a specific region to ensure the accuracy of the results \[12,13\]. Considering the progress of technology and the utilization of artificial intelligence, particularly ML models in various medical fields, the prediction of SC risk is among the valuable applications of ML models in the oncology domain \[14\]. The ML approach is centered on data and its continual enhancement, relying on statistics and probability. However, the outcomes derived from this approach surpass those of statistical methods \[15\]. ML encompasses diverse models, with the accuracy of results contingent upon the specific models employed. Linear regression analysis, a straightforward and widely used technique for assessing relationships between predictor variables and a continuous response, assumes linearity in the relationships between predictor and target variables. This implies that a consistent unit change in one variable corresponds to a consistent unit change in the other variable. Linear regression is often the preferred option for analyses involving small sample sizes, as these models are straightforward to interpret. Based on the findings related to SC risk and possessing knowledge about the effective dose, the correlation between the effective dose in the organ and the risk of SC can be computed. Recently, the integration of ML models, particularly decision trees, into the research methodology has led to the development of a practical framework for predicting the incidence of SC using patient data. This framework facilitates the classification of patients into high-risk and low-risk categories, thus supporting the formulation of personalized treatment strategies and interventions. Furthermore, it highlights several factors influencing the probability of SC, including radiation exposure, patient age, and genetic factors, while also pointing out the shortcomings of existing models in accounting for all pertinent variables \[16\]. Our study involved the utilization of ML models to analyze past research data to calculate the risk of SC. The primary objective of this research is to determine the correlation between dosage and the likelihood of developing SC, with a specific focus on establishing a relationship using a linear regression model. The paper is structured subsequently: Section 2 and its subsections present a comprehensive analysis of the various ML models explored to identify the most suitable ML model for predicting the SC using patient data. Section 3 compares the results produced by the ML models and discusses the best approach for evaluating feature significance and forecasting SC. Lastly, section 4 provides a summary of the findings and draws informed conclusions from the research conducted. **2. Methodology** The research involving human participants received approval from the Ethics Committee at Arak University of Medical Sciences. This study was conducted by the regulations established by the local authorities and institutional standards. This research was conducted to obtain the relationship between the dose and the risk of SC, and its working method includes two steps: **a)** machine learning and **b)** regression model, which are fully explained below. **a) Machine Learning** Predicting the risk of SC using ML methods is an important and active research field in the field of oncology and medical sciences. This method uses training of its algorithms on patient data to check the risk of SC. The ML method consists of several steps, and we schematically present the steps taken by this method to predict the risk of SC in Figure 1. E:\\phd\\nemudar dose-risk\\latex\\SCELSEVER\\fig1.png Figure 1. The figure depicts the overall workflow diagram. As mentioned, the ML method is data-oriented, and you can see in Fig. 1 that the most basic steps of this method are the selection of data to train the algorithm on them. The dataset used in this research is based on work done in the past decades and collected by databases and includes information such as gender, radiation dose, and age of radiation exposure. Due to the significant importance of data in using the ML method and the information that this data provides us, the data sets were selected with high sensitivity. Table 1 shows details about the data used in this research. Table 1(a). Information about secondary Lung cancer data details +-----------------+-----------------+-----------------+-----------------+ | cancer site | Lung | | | +=================+=================+=================+=================+ | publication | Van Leeuwen et | Mattsson et al. | Davis et al. | | | al. (1995) | (1997) \[18\] | | | | \[17\] | | \(1989) \[19\] | +-----------------+-----------------+-----------------+-----------------+ | all case | 1939 | 1216 | 13385 | +-----------------+-----------------+-----------------+-----------------+ | cases/death | 30 | 19 | 69 | +-----------------+-----------------+-----------------+-----------------+ | women in study% | 41 | 100 | 48.7 | +-----------------+-----------------+-----------------+-----------------+ | Age at exposure | \55 | 8-74 | \ 38 | | (year) | | | | +-----------------+-----------------+-----------------+-----------------+ | Follow-up | 1-\>10 | 5-61 | 0- 50 | | (year) | | | | +-----------------+-----------------+-----------------+-----------------+ | Average dose | 7.2 | 0.75 | 0.84 | | (Sv) | | | | +-----------------+-----------------+-----------------+-----------------+ | Dose Range (Sv) | 0-\>21 | 0-8.98 | 0-\>8 | +-----------------+-----------------+-----------------+-----------------+ Table 1(b). Information about secondary Colon cancer data details +-----------------+-----------------+-----------------+-----------------+ | cancer site | Colon | | | +=================+=================+=================+=================+ | publication | Inskip et al. | Darby et al. | Weiss et al. | | | (1990) \[20\] | (1995) \[21\] | | | | | | \(1994) \[22\] | +-----------------+-----------------+-----------------+-----------------+ | all case | 4153 | 2067 | 14109 | +-----------------+-----------------+-----------------+-----------------+ | cases/death | 73 | 47 | 226 | +-----------------+-----------------+-----------------+-----------------+ | women in study% | 100 | 100 | 17.8 | +-----------------+-----------------+-----------------+-----------------+ | Age at | 13 - 88 | 23 - 65 | \55 | | exposure(year) | | | | +-----------------+-----------------+-----------------+-----------------+ | Follow up(year) | 0 - 60 | 5 \_ 49 | 5 - \>35 | +-----------------+-----------------+-----------------+-----------------+ | Average dose | 1.2 | 3.2 | 4.1 | | (Sv) | | | | +-----------------+-----------------+-----------------+-----------------+ | Dose Range (Sv) | \7.85 | +-----------------+-----------------+-----------------+-----------------+ Table 1(c). Information about secondary Breast cancer data details cancer site Breast ----------------------- ---------------------------- ------------------------------- publication Boice et al. (1989) \[23\] Hildreth et al. (1989) \[24\] all case 12040 1201 cases/death 140 34 women in study% 100 100 Age at exposure(year) \75 \40 0 -\>52 Average dose (Sv) 0.31 0.69 Dose Range (Sv) 0 - 0.98 0.01 -7.1 One of the most essential steps in the data mining process, which has a significant impact on the selection of models for prediction and conclusions, is data processing and the relationship between them, which is examined in Fig. 2 After this step, the data is ready for analysis.  Figure 2. Overall workflow diagram This method examines a part of the data set for training and after this step examines the remaining part of the data set for testing and obtaining results. In this research, the training dataset consists of 70% of the data, while the testing dataset comprises the remaining 30%. In this research, to choose the best model for predicting the risk of SC, we have examined four models: decision tree, random forest, bagging, and AdaBoost. One of the tests that helped us choose the best model is the calculation of the AUC and ROC curve, the minimum value of AUC is zero and the maximum value is one, and the closer this number is to one, it means that the model has strong power. Predictability is therefore very satisfactory if it is in the range of (0.8-0.9) and excellent if it is (0.9-1). Decision Tree (DT): One of the best classification algorithms is DT, which has features such as interpretability, analysis, and simplicity. In this method, a tree structure should be used, which has special rules for the collective implementation of classification processes, in the tree structure, there are three important parts internal nodes, branches, and leaf nodes, which respectively indicate the characteristics, values of the characteristics and the classes that exist in the data set. The internal node that produces the output is called a branch and can be the input of another internal node \[25\]. Fig. 3 shows the results of the AUC and ROC curve for the Decision Tree model. E:\\phd\\nemudar dose-risk\\latex\\SCELSEVER\\fig9.png Figure 3. AUC and ROC curve for the Decision Tree model. Bagging: This technique generates final predictions using a random selection of subsets of the data. Breiman introduced the concept of bagging, also referred to as bootstrap aggregation \[26\]. Fig. 4 shows the results of the AUC and ROC curve for the Bagging model.  Figure 4. AUC and ROC curve for the Bagging model AdaBoost: Adaboost was first introduced as a classification algorithm in 1997 by Freund and Schapire. For training, this method first creates a decision tree in which the data has equal weight at each point, then uses the appropriate model to classify the training set. If this model correctly predicts the weight of the data, it remains unchanged, and if this diagnosis is wrong, the weight of the samples is changed, and after creating a balance between the weights (normalization), a new decision tree is created. This process is repeated until the ideal conditions are reached \[27\]. Fig. 5 shows the results of the AUC and ROC curves for the AdaBoost model. E:\\phd\\nemudar dose-risk\\latex\\SCELSEVER\\fig10.png Figure 5. AUC and ROC curve for the AdaBoost model. Random Forest (Rf): In 1995, Hu introduced the RF model with the idea of taking stochastic methods that include sub-decision trees that function as an ensemble learning classification algorithm. Finally, this method has a higher prediction accuracy than the methods that use a single decision tree \[28\]. Fig. 6 shows the results of the AUC and ROC curve for the Random Forest model.  Figure 6. AUC and ROC curve for the Random Forest model **b) Regression Model** Linear regression is a statistical technique used to estimate the linear association between a single response variable and one or more explanatory variables, which are also referred to as dependent and independent variables. When the model involves only one explanatory variable, it is known as simple linear regression \[29\]. This research, for obtaining the relationship between dose and SC risk has been done by linear regression method, so that, first, the risks obtained by ML are placed in one list and the available doses from previous research are placed in another list. After creating these two lists, we begin by creating a linear regression model in our Python program. We classify the values in these lists as independent variables (in this case, SC risk) and dependent variables (dose) and input them into the model to determine the relationship between these variables. The linear regression model works by initially creating a default first-order linear equation. Y= Ax + B (1) After this step, the program considers the slope of the line (A) to be 1 and the distance from the origin (B) to be 0. It then predicts the relationship between risk and dose as a line and calculates the standard deviation for the values. After calculating the standard deviation, the program again predicts a new line equation by changing the slope and width from the origin and calculates the standard deviation for the new values ​​obtained, if the new standard deviation is smaller than before, it means that this new equation It is more optimal and suitable than the previous equation and until the lowest standard deviation is obtained, the program automatically changes the values ​​of A and B and finally predicts the most optimal and most suitable equation for this relationship and finally the line equation can be drawn and compared with other experimental values. We used the values ​​obtained for the risk of secondary breast, colon, and lung cancer by ML model and considering that these values ​​were obtained using the available data from previous research and the average dose was also available in these data. We obtained the relationship between the risk of SC and the dose using the linear regression method, and after drawing the line, we compared this relationship with other studies. **3. Results and Discussion** The findings obtained from the ML method are displayed in Table 2 and are further enhanced by integrating results from other computational and simulation methods to enable a comprehensive comparison. Donovan et al. conducted an experimental study in 2012, utilizing thermoluminescence dosimeters (TLD) to measure the effective dose in the organ using a phantom. This study encompassed various radiotherapy techniques, including whole breast radiotherapy (WBRT), partial breast irradiation (APBI), and simultaneous integrated enhancement (SIB) with two and three-volume models. Subsequently, the risk of SC was determined using computational models (BEIR VII) \[30\]. Mendes et al. utilized the MCNP code to calculate the risk of SC, employing a virtual phantom known as the VW phantom, which represented a woman with 63 organs, a height of 165 cm, and a weight of 98 kg \[13\]. The absorbed dose in each organ after radiotherapy was calculated using the MCNP code, considering a parallel field of 6 Mv as the radiation source. The risk of SC was then determined using the BEIR VII computational model. The linear regression model outcomes concerning the correlation between dosage and the risk of SC have been visually represented through graphs displayed in Figs. 7 to 9, focusing on SC occurrences in the breast, colon, and lung. Table 2(a). Comparison of different methods for predicting the risk of Second Cancer in the lung Second Cancer Publishers Method Average Dose(sv) Second Cancer Risk (%) --------------- ------------------------------ ---------------------- ------------------ ------------------------ Lung Donovan et al. (2012) \[30\] WBRT 0.68 0.12 Donovan et al. (2012) \[30\] APBI 0.70 0.07 Donovan et al. (2012) \[30\] SIB 2 volume 1.90 1.11 Donovan et al. (2012) \[30\] SIB 3 volume FP IMRT 0.27 0.30 Donovan et al. (2012) \[30\] SIB 3 volume IP IMRT 1.80 0.68 Mendes et al. (2017) \[13\] MCNP 0.22 0.38 This work (2024) ML 7.2 0.77 This work (2024) ML 0.75 0.59 This work (2024) ML 0.84 0.44 Table 2(b). Comparison of different methods for predicting the risk of Second Cancer in the colon Second Cancer Publishers Method Average Dose(sv) Second Cancer Risk (%) --------------- ------------------------------ ---------------------- ------------------ ------------------------ Colon Donovan et al. (2012) \[30\] WBRT 0.08 0.06 Donovan et al. (2012) \[30\] APBI 0.07 0.05 Donovan et al. (2012) \[30\] SIB 2 volume 0.15 0.09 Donovan et al. (2012) \[30\] SIB 3 volume FP IMRT 0.21 0.14 Donovan et al. (2012) \[30\] SIB 3 volume IP IMRT 0.11 0.06 Mendes et al. (2017) \[13\] MCNP 0.06 0.03 This work (2024) ML 1.2 0.38 This work (2024) ML 3.2 0.41 This work (2024) ML 4.1 0.41 Table 2(c). Comparison of different methods for predicting the risk of Second Cancer in the breast Second Cancer Publishers Method Average Dose(sv) Second Cancer Risk (%) --------------- ------------------------------ ---------------------- ------------------ ------------------------ Breast Donovan et al. (2012) \[30\] WBRT 0.6 0.6 Donovan et al. (2012) \[30\] APBI 0.19 0.18 Donovan et al. (2012) \[30\] SIB 2 volume 0.74 0.69 Donovan et al. (2012) \[30\] SIB 3 volume FP IMRT 0.43 0.61 Donovan et al. (2012) \[30\] SIB 3 volume IP IMRT 1.17 1.1 Mendes et al. (2017) \[13\] MCNP 0.27 0.55 This work (2024) ML 0.31 0.59 This work (2024) ML 0.69 0.7 E:\\phd\\nemudar dose-risk\\IMG\_20240719\_200554\_760.png Figure 7. Results of linear regression model for secondary breast cancer  Figure 8. Results of linear regression model for secondary lung cancer E:\\phd\\nemudar dose-risk\\IMG\_20240719\_200727\_499.png Figure 9. Results of the linear regression model by Python program for secondary colon cancer Based on the graphs provided, the outcomes of the regression model demonstrate a satisfactory level of concordance with the outcomes of the computational models. It is evident that, in line with expectations, the likelihood of developing SC escalates with the augmentation of the effective dose in the organ. The model recommended by the BEIR committee has outlined the definitions of excess relative risk (ERR) and excess absolute risk (EAR) is articulated as follows: ERR and EAR = βSD exp (γ e∗) ([\${\\frac{a}{60})}\^{\\eta}\$]{.math.inline} , (2) In the equation provided, D represents the dose administered, while βS, γ, and η are parameters specific to excess relative risk (ERR) and excess absolute risk (EAR) for different organs based on sex. The variable e∗ denotes the age at exposure, and a represents the attained age. Focusing solely on the linear component of Equation (2), which pertains to the correlation between dose and risk, the equation can be reformulated as \[30\]. ERR and EAR = βSD. (3) The linear regression model has yielded a coefficient (A) from Eq. (1) that represents the slope of the graphs depicting the relationship between dose and the risk of SC. Utilizing this model allows for the calculation of the βS coefficient for secondary cancer. Furthermore, the slope of the regression line for secondary breast cancer is 0.879, while for secondary lung cancer, it is 0.420. The slope for secondary colon cancer is 0.657, indicating a potential correlation with the radiation sensitivity of the respective organs. **4. Summary and conclusions** It is crucial to tackle the rise in cancer incidences within the community by promptly diagnosing and treating the disease. Furthermore, it underscores the significance of cancer survivors being mindful of the potential risk of developing SC, which may be impacted by a range of factors including treatment modalities, lifestyle decisions, and behaviors such as smoking and alcohol intake. The use of ML models in predicting the risk of SC based on effective doses in the organ is an effective application in the field of oncology. Linear regression analysis is a popular method for measuring the relationship between predictor variables and continuous response and is often the best choice for analyses with small sample sizes. The research aims to establish a new relationship using the linear regression model between the dose and the risk of SC. We compare different methods for predicting the risk of SC in the lung, colon, and breast. The results indicate that the risk of SC increases with the effective dose in the organ, and the linear regression model provides coefficients that are related to the radiation sensitivity of the specific organ. **Statements and declarations** ** Ethics statement** The studies involving humans were approved by the Ethics Committee of Arak University of Medical Sciences (Approval number **IR.ARAKMU.REC.1403.1**58). The research was conducted in compliance with the regulations of the local district and the standards set by the institution. Due to the retrospective nature of the study, the ethics committee or institutional review board decided to exempt the need for written informed consent from the participants or their legal guardians/next of kin. ** CRediT authorship contribution statement:** **Erfan Hatamabadi Farahani:** Methodology, Investigation, Data curation, Writing -- review & editing, Conceptualization, Validation, Software, Methodology**. Hossein Sadeghi:** Methodology, Investigation, Data curation, Writing -- review & editing, Supervision. **Fatemeh Seif:** Investigation, Data curation, Conceptualization, Writing-- review & editing, Supervision. **Mahdi Azad:** Conceptualization, Validation, Software, Methodology. **Reza Rezaee:** Investigation, Data curation, Conceptualization. ** Data Availability:** All data generated or analyzed during this study are included in this published article. ** Conflict of Interest:** Authors state no conflict of interest. ** Funding:** No funding was received to assist with the preparation of this manuscript. **References**: 1\. DeVita VT, Lawrence TS, Rosenberg SA, editors. DeVita, Hellman, and Rosenberg\'s cancer: principles & practice of oncology. Lippincott Williams & Wilkins; 2008. 2\. Feller A, Matthes KL, Bordoni A, Bouchardy C, Bulliard JL, Hermann C, Konzelmann I, Maspoli M, Mousavi M, Rohrmann S, Staehelin K. The relative risk of second primary cancers in Switzerland. InDGEpi-Jahrestagung: German Society of Epidemiology, September 26-28, Bremen, Germany, 2018 2018. Doi: 10.1186/s12885-019-6452-0 3\. Dasu A, Toma-Dasu I. Models for the risk of secondary cancers from radiation therapy. Physica Medica. 2017 Oct 1; 42:232-8. Doi: 10.1016/j.ejmp.2017.02.015 4\. Hall EJ. Intensity-modulated radiation therapy, protons, and the risk of second cancers. International Journal of Radiation Oncology\* Biology\*Physics.2006May1;65(1):1-7. Doi: 10.1016/j.ijrobp.2006.01.027 5\. Davis RH. Production and killing of second cancer precursor cells in radiation therapy: in regard to Hall and Wuu (Int J Radiat Oncol Biol Phys 2003; 56: 83--88). International journal of radiation oncology, biology, physics. 2004 Jul 1;59(3):916. Doi: 10.1016/j.ijrobp.2003.09.076 6\. Mertens AC, Liu Q, Neglia JP, Wasilewski K, Leisenring W, Armstrong GT, Robison LL, Yasui Y. Cause-specific late mortality among 5-year survivors of childhood cancer: the Childhood Cancer Survivor Study. Journal of the National Cancer Institute. 2008 Oct 1;100(19):1368-79. Doi:10.1093/jnci/djn310 7\. Yerramilli D, Xu AJ, Gillespie EF, Shepherd AF, Beal K, Gomez D, Yamada J, Tsai CJ, Yang TJ. Palliative radiation therapy for oncologic emergencies in the setting of COVID-19: approaches to balancing risks and benefits. Advances in Radiation Oncology. 2020 Jul 1;5(4):589-94. Doi: 10.1016/j.adro.2020.04.001 8\. Rades D, Stalpers LJ, Veninga T, Schulte R, Hoskin PJ, Obralic N, Bajrovic A, Rudat V, Schwarz R, Hulshof MC, Poortmans P. Evaluation of five radiation schedules and prognostic factors for metastatic spinal cord compression. Journal of Clinical Oncology. 2005May 20;23(15):3366 [Doi](https://en.wikipedia.org/wiki/Doi_(identifier)):[10.1200/JCO.2005.04.754](https://doi.org/10.1200%2FJCO.2005.04.754) 9\. Mullenders L, Atkinson M, Paretzke H, Sabatier L, Bouffler S. Assessing cancer risks of low-dose radiation. Nature Reviews Cancer. 2009 Aug;9(8):596-604. Doi: 10.1038/nrc2677 10\. Rades D, Panzner A, Rudat V, Karstens JH, Schild SE. Dose escalation of radiotherapy for metastatic spinal cord compression (MSCC) in patients with relatively favorable survival prognosis. Strahlentherapie und Onkologie. 2011 Nov 1;187(11):729. [Doi](https://en.wikipedia.org/wiki/Doi_(identifier)):[10.1007/s00066-011-2266-y](https://doi.org/10.1007%2Fs00066-011-2266-y) 11\. de Gonzalez AB, Gilbert E, Curtis R, Inskip P, Kleinerman R, Morton L, Rajaraman P, Little MP. Second solid cancers after radiation therapy: a systematic review of the epidemiologic studies of the radiation dose-response relationship. International Journal of Radiation Oncology\* Biology\* Physics. 2013 Jun 1;86(2):224-33. Doi:[10.1016/j.ijrobp.2012.09.001](https://doi.org/10.1016/j.ijrobp.2012.09.001) 12\. Doudoo CO, Gyekye PK, Emi-Reynolds G, Adu S, Kpeglo DO, Tagoe SN, Agyiri K. Dose, and secondary cancer-risk estimation of patients undergoing high dose rate intracavitary gynecological brachytherapy. Journal of Medical Imaging and Radiation Sciences. 2023 Jun 1;54(2):335-42. Doi:10.1016/j.jmir.2023.03.031 13\. Mendes BM, Trindade BM, Fonseca TC, Campos dTP. Assessment of radiation-induced secondary cancer risk in the Brazilian population from left-sided breast-3D-CRT using MCNPX. The British J of Radiol. 2017;90(1080):20170187. Doi: 10.1259/bjr.20170187. 14\. Debnath S, Barnaby DP, Coppa K, et al. Machine learning to assist clinical decision-making during the COVID-19 pandemic. Bioelectron Med. 2020;6(1):1-8. Doi:10.1186/s42234-020-00050-8 15\. Syleouni ME, Karavasiloglou N, Manduchi L, et al. Predicting second breast cancer among women with primary breast cancer using machine learning algorithms, a population-based observational study. Int J of Cancer. 2023;153:932--941. Doi: 10.1002/ijc.34568. ### 16. Sadeghi H, Seif F, Hatam-Abadi E, Khanmohammadi S, Nahidinezhad S. Utilizing Patient Data: A Tutorial on Predicting Second Cancer with Machine Learning Models. Cancer Medicine, In press. Doi: 10.1002/cam4.70231 ### 17\. Van Leeuwen FE, Klokman WJ, Stovall M, Hagenbeek A, Van Den Belt-dusebout AW, Noyon R, Boice Jr JD, Burgers JM, Somers R. Roles of radiotherapy and smoking in lung cancer following Hodgkin\'s disease. JNCI: Journal of the National Cancer Institute. 1995 Oct 18;87(20):1530-7. Doi:10.1093/jnci/87.20.1530 18\. Mattsson A, Hall P, Rudén BI, Rutqvist LE. Incidence of primary malignancies other than breast cancer among women treated with radiation therapy for benign breast disease. Radiation research. 1997 Aug 1;148(2):152-60. Doi:10.2307/3579572 19\. Davis FG, Boice Jr JD, Hrubec Z, Monson RR. Cancer mortality in a radiation-exposed cohort of Massachusetts tuberculosis patients. Cancer research. 1989 Nov 1;49(21):6130-6. 20\. Inskip PD, Monson RR, Wagoner JK, Stovall M, Davis FG, Kleinerman RA, Boice Jr JD. Cancer mortality following radium treatment for uterine bleeding. Radiation research. 1990 Sep 1;123(3):331-44. Doi:10.2307/3577741 21\. Darby SC, Reeves G, Key T, Doll R, Stovall M. Mortality in a cohort of women given X‐ray therapy for metropathia haemorrhagica. International journal of cancer. 1995 Mar 15;56(6):793-801. Doi:10.1002/ijc.2910560606 22\. Weiss HA, Darby SC, Doll R. Cancer mortality following X‐ray treatment for ankylosing spondylitis. International Journal of Cancer. 1994 Nov 1;59(3):327-38. Doi:10.1002/ijc.2910590307 23\. Boice Jr JD, Engholm G, Kleinerman RA, Blettner M, Stovall M, Lisco H, Moloney WC, Austin DF, Bosch A, Cookfair DL, Krementz ET. Radiation dose and second cancer risk in patients treated for cancer of the cervix. Radiation research. 1988 Oct 1;116(1):3-55. Doi:10.2307/3577477 24\. Hildreth NG, Shore RE, Dvoretsky PM. The risk of breast cancer after irradiation of the thymus in infancy. New England Journal of Medicine. 1989 Nov 9;321(19):1281-4. Doi:10.1056/NEJM198911093211901 25\. Wu Y, Ke Y, Chen Z, Liang S, Zhao H, Hong H. Application of alternating decision tree with AdaBoost and bagging ensembles for landslide susceptibility mapping. Catena. 2020 Apr 1;187:104396. Doi:10.1016/j.catena.2019.104396 26\. Alshahrani SM, Albaghdadi MF, Yasmin S, Alosaimi ME, Alsalhi A, Algarni M, Felemban BF, Fadhil AA, Mohammed IM. Green processing based on supercritical carbon dioxide for preparation of nanomedicine: model development using machine learning and experimental validation. Case Studies in Thermal Engineering. 2023 Jan 1;41:102620. Doi:10.1016/j.csite.2022.102620 27\. Mosavi A, Sajedi Hosseini F, Choubin B, Goodarzi M, Dineva AA, Rafiei Sardooi E. Ensemble boosting and bagging based machine learning models for groundwater potential prediction. Water Resources Management. 2021 Jan;35:23-37. Doi:10.1007/s11269-020-02704-3 28\. AlSagri H, Ykhlef M. Quantifying feature importance for detecting depression using random forest. International Journal of Advanced Computer Science and Applications. 2020;11(5). Doi:10.14569/ijacsa.2020.0110577 29\. Freedman DA. A simple regression equation has on the right-hand side an intercept and an explanatory variable with a slope coefficient. A multiple regression e right hand side, each with its own slope coefficient. Statistical Models: Theory and Practice. 2009:26. 30\. Donovan EM, James H, Bonora M, Yarnold JR, Evans PM. Second cancer incidence risk estimates using BEIR VII models for standard and complex external beam radiotherapy for early breast cancer. Medical physics. 2012 Oct;39(10):5814-24. Doi:10.1118/1.4748332