Model for Endometriosis Detection Using Machine Learning Algorithms

BIBLIOTECA This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. Document downloaded from the institutional repository of the University of Alcala: http://ebuah.uah.es/dspace/ This is a posprint version of the following published document: Bautista, A., Tardillo, J., Castillo Sequera, J.L. & Wong, L. 2025, “Model for endometriosis detection using machine learning algorthms”, in 2025 7th International Conference on Software Engineering and Computer Science (CSECS). Available at https://dx.doi.org/10.1109/CSECS64665.2025.11009460 © 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other users, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works for resale or redistribution to servers or lists, or reuse of any copyrighted components of this work in other works. (Article begins on next page) Model for Endometriosis Detection using Machine Learning Algorithms Alexis Bautista Information Systems Engineering Program Universidad Peruana de Ciencias Aplicadas Lima, Peru ‘[email protected] Jahir Tardillo Information Systems Engineering Program Universidad Peruana de Ciencias Aplicadas Lima, Peru [email protected] José Luis Castillo-Sequera Department of Computer Science Universidad de Alcalá Alcalá de Henares, Spain [email protected] Lenis Wong Information Systems Engineering Program Universidad Peruana de Ciencias Aplicadas Lima, Peru [email protected] Abstract—Endometriosis is a chronic disease that affects a considerable percentage of women of reproductive age and is characterized by the presence of endometrial tissue outside the uterine cavity, leading to symptoms such as pelvic pain and dysmenorrhea. The aim of this study is to develop a predictive model for the classification of endometriosis using four Machine Learning algorithms: Random Forest, LASSO, SVM, and Naive Bayes. For this purpose, a dataset from the Global Health Data Exchange was utilized, consisting of 1,000 cases of patients with endometriosis. The methodology included data cleaning and preprocessing, as well as the evaluation of each algorithm's performance using four metrics: precision, recall, F1-Score, and accuracy. The findings revealed tha t the Random Forest algorithm was the most effective in identifying endometriosis, outperforming the other algorithms with a precision of 0.99 for the "endometriosis" class and an overall accuracy of 0.98. Keywords—endometriosis, machine learning, Random Forest, LASSO I. INTRODUCTION Endometriosis, as defined in [16], is the presence of functional endometrial tissue (glands and stroma) outside the uterine cavity. Additionally, it is a chronic, periodically symptomatic, estrogen-dependent disease that affects between 10% and 30% of wome n of reproductive age and older. According to [13], the primary symptom of endometriosis is pelvic pain, which may occur during vaginal bleeding (dysmenorrhea), during sexual intercourse (dyspareunia), or independently of vaginal bleeding (non -menstrual pe lvic pain). Patients may also experience lower back pain or abdominal discomfort. These symptoms can significantly impact a patient’s physical, mental, and social well -being, thereby impairing their quality of life. Endoscopic Submucosal Dissection (ESD) and Comprehensive Sexuality Education (CSE) provide clinically relevant evaluations of endometriosis symptoms and the disease's impact on patients' lives. Contemporary medicine stands at a critical intersection between the growing volume of available medical data and the need for more precise and personalized tools for disease diagnosis, treatment, and management. In this context, Machine Learning (ML) has emerged as a powerful tool with the potential to radically transform medical practice. According to [19], ML is used in the medical field to analyze large datasets, including clinical, genetic, and medical imaging data, with the goal of improving diagnosis, treatment, and disease management. Beyond endometriosis, ML has also proven essential in other medical domains. For instance, [20] highlights its significance in predicting immunotherapy responses in cancer. The integration of Machine Learning (ML) with endometriosis research offers significant potential, as ML's capability to process large datasets and detect underlying patterns positions it as a key tool for overcoming diagnostic challenges in this gynecological condition. By leveraging ML algorithms, researchers can identify patterns, risk factors, and specific traits associated with endometriosis, providing critical insights for treatment strategies [6]. Additionally, as noted in [4] and [18], these technologies improve diagnostic accuracy and outcome predictions, fostering the development of more accessible non -invasive diagnostic methods. However, as highlighted in [2], the lack of reproducibility in studies involving microRNAs remains a limitation, underl ining the necessity for consistent and reliable results. Early detection of endometriosis is vital, as it enables the application of more effective management approaches and enhances patients’ quality of life [9]. While ML demonstrates the potential for more precise and expedited diagnoses compared to traditional methods, it complements other medical domains that often require longer and riskier diagnostic pathways for patients. With the rapid advancement of technology, this research proposes implementing a model for detecting endometriosis by applying four Machine Learning algorithms: Random Forest, Support Vector Machine, LASSO, and Naive Bayes. The implementation follows the CRISP-DM methodology. II. RELATED WORKS In the literature, various ML algorithms have been applied to address diseases with different purposes, such as analyzing tissue samples, identifying genes, and biomarkers. The most used algorithms include RF, LASSO, and SVM (see Table 1). The Random Forest (RF) algorithm has demonstrated remarkable effectiveness in detecting endometriosis, excelling in processing large datasets and improving classification accuracy for biomarkers and genetic patterns associated with the disease. For instanc e, studies such as [15] and [17] achieved AUC values of 0.721 and 0.939, respectively, by employing RF to classify genes linked to cuproptosis and microRNAs, underscoring the algorithm’s ability to analyze complex data. Additionally, research presented in [14] and [8] reported AUC values of 0.8226 and 0.895 when identifying critical biomarkers and analyzing genes associated with senescence. Despite variations in study methodologies, such as ovarian lesion classification in [3], which achieved an AUC of 0.968, RF consistently proves to be a versatile and robust tool in detecting gynecological diseases, particularly endometriosis. The LASSO algorithm has demonstrated significant efficacy in detecting endometriosis and other gynecological conditions, primarily due to its capability to reduce dimensionality and select relevant features. For example, [8] employed LASSO to identify genes associated with endometriosis and senescence, achieving AUC values of 0.822 during training and 0.895 in validation, highlighting its strong classification performance for endometriosis patients. Similarly, [11] utilized LASSO to select eight key genes related to M2 macrophages, achieving an AUC ≥ 0.65, enabling a detailed analysis of disease severity based on gene expression. Furthermore, [18] demonstrated the al gorithm's ability to enhance sensitivity and specificity in a non-invasive diagnostic model for endometriosis, achieving an AUC of 0.80 for the CA125 marker. In addition, [15] applied LASSO to mitigate overfitting in their classification model, achieving AUC values of 0.781 in training and 0.721 in testing, showcasing its versatility across various applications. Notably, [3] utilized LASSO to develop a logistic regression model for predicting ovarian cancer, achieving an AUC of 0.946, underscoring its robustness in discriminating between benign and malignant lesions. The Support Vector Machine (SVM) algorithm has proven to be highly effective in disease classification and detection, particularly in the medical field, due to its capacity to manage complex datasets. For instance, [3] utilized SVM to distinguish between b enign and malignant ovarian lesions, achieving an AUC of 0.821, which indicates moderate discriminatory capability. This application was compared with other algorithms, highlighting SVM's ability to identify patterns within critical clinical and serological data for precise diagnoses. Similarly, [7] applied SVM for the early diagnosis of endometriosis based on self -reported patient data, achieving an AUC of 0.87, which demonstrates strong predictive performance. While [3] emphasized ovarian lesion classification, [7] showcased SVM's adaptability in addressing other clinical challenges such as endometriosis. Both studies emphasize the robustness of SVM in binary classification tasks, although its performance varies depending on the specific medical issue and the characteristics of the dataset being analyzed. TABLA 1. Algorithm Purpose AUC Ref. RF Distinguish endometrial tissue samples 0.85 [8] RF Identify genes based on the samples 0.78 [15] RF Classify the presence or absence of endometriosis 0.94 [17] RF Identify genes associated with endometriosis 0.78 [14] RF Identify key genes related to M2 macrophages in endometriosis 0.65 [11] RF Build predictive models to classify ovarian lesions. 0.96 [3] LASSO Distinguish features among genes expressed with endometriosis. 0.85 [8] LASSO Identify key genes related to M2 macrophages in endometriosis. AUC > 0.65 [11] LASSO Classify patients based on their symptoms and self -reported characteristics. AUC [7] LASSO Classify segmented images for tumor recognition. AUC > 0.85 [12] LASSO Distinguish characteristics among genes expressed with AUC > 0.85 [8] endometriosis. LASSO Identify key genes associated with M2 macrophages in endometriosis. AUC > 0.65 [11] LASSO Improve diagnostic accuracy. AUC > 0.8 [18] LASSO Identify genes that aid in the prediction of endometriosis. AUC > 0.78 [15] LASSO Filter the most relevant predictors among serological biomarkers. AUC > 0.96 [3] SVM Classification model to predict the presence of ovarian lesions. AUC > 0.96 [3] SVM Classify data within the context of medical decision support systems. AUC > 0.88 [1] SVM Identify genes serving as diagnostic markers for pulmonary arterial hypertension. AUC > 0.94 [5] III. PROPOSED MODEL For this reason, none of these studies focus on the identification of patient symptoms. Our proposal emphasizes this aspect because symptoms are key clinical indicators for the early detection of endometriosis, a condition often diagnosed years after its onset due to its variability and complex presentation. By focusing on symptoms reported by patients, we aim to develop a prediction system that not only facilitates faster and non-invasive diagnoses but also provides a practical tool for physicians in clinical settings. This approach allows for a personalize d strategy tailored to the individual characteristics of each patient, thereby improving the accuracy and effectiveness of treatment from early stages. Fig. 1. Conceptual Model of the Proposed Approach. A. Data selection For this study, a dataset extracted from the Global Health Data Exchange (GHDx) site [9] will be used, containing 250 cases of patients with endometriosis from the year 2020. Each case includes 15 features related to symptoms and other clinically relevant variables for diagnosing the disease. B. Procesamiento de Data Data preprocessing techniques are critical to ensuring data quality and enhancing the performance of Machine Learning models. For this study, the data preprocessing phase involved several steps: data cleaning, feature selection, and variable encoding for numerical representation using Python functions like ‘.map’. For instance, to encode the variable "Infertility," which initially contains the values 'Yes' and 'No,' the ‘.map’ function was applied to convert these values into numerical variables such as '1' for 'Yes' and '0' for 'No.' Table II presents the nine features selected. TABLE II. DATASET CHARACTERISTICS ID Characteristics Description F01 Age Patient's age F02 Ethnicity Human community F03 Days of Menstruation Duration of menstruation in days F04 Pain Presence of pain F05 Increased Bleeding Increased menstrual bleeding F06 Prolonged Menstruation Extended menstrual cycle days F07 Infertility Difficulty conceiving F08 Dyspareunia Genital pain during sexual intercourse F09 Dysmenorrhea Lower abdominal pain C. Algorithm Training For this study, the three most used algorithms in the literature were employed: Random Forest, LASSO, and SVM (Support Vector Machine). In Figure 2, the steps for training the Random Forest algorithm to predict endometriosis diagnosis in Python are illustrated The flowchart illustrates the process of using a machine learning model for diagnosing endometriosis. Initially, patient data and reported symptoms are provided as input for analysis. In Step 1, the dataset is divided into training and testing subsets using the train_test_split function, allocating 70% for training and 30% for testing. Step 2 involves training the Random Forest algorithm with the fit function on the training data (X_train and Y_train), enabling it to learn patterns associated with endometri osis symptoms. In Step 3, the trained model predicts outcomes for the test data (X_test), with results stored in predictions. A confusion matrix evaluates the model's performance by comparing correct and incorrect predictions. Step 4 calculates accuracy us ing the accuracy_score function to determine the model's effectiveness in classifying cases. Additionally, the model's reliability is validated using real clinical data (x_real_data) unseen during training. Finally, the output presents the diagnosis, indic ating whether the patient is likely to have endometriosis. This methodology was consistently applied to other algorithms evaluated in the study. Fig. 2. Steps for Training the Random Forest Model in Python. This process was carried out in a Python -based development environment, specifically using PyCharm. Popular libraries such as pandas, scikit -learn, and matplotlib were employed for data manipulation, model construction, and result visualization. In the correlation analysis of the endometriosis dataset features, a correlation matrix was utilized to identify relationships between different variables and the diagnosis of endometriosis. This matrix aids in visualizing the compatibility of each feature with the desired outcome, in this case, the "Endometriosis Diagnosis." It was observed that certain variables, such as F05 (Increased bleeding) and F09 (Dysmenorrhea), show significant correlation with the diagnosis, which is critical for optimizing the p redictive model. This approach streamlines the feature selection process and enhances the model's accuracy by focusing on the most relevant variables for prediction. Fig. 3. Features of the Correlation Matrix and Dataset. D. Evaluation After the classification process, the results are compared with the actual diagnoses from the dataset to generate four key variables: "True Positives" (TP), "False Positives" (FP), "False Negatives" (FN), and "True Negatives" (TN). These variables are essential for calculating metrics that evaluate the model's performance, as explained in Table IV. TABLE IV. DESCRIPTION OF DATASET VARIABLES Variable Description TP Cases where the model correctly predicted that a patient has endometriosis, and she actually has the disease. FP Cases where the model predicted that a patient has endometriosis, but she actually does not have the disease. FN Cases where the model did not predict endometriosis (negative prediction), but the patient does have the disease. TN Cases where the model correctly predicted that a patient does not have endometriosis, and she actually does not have the disease. The analysis includes the following metrics: Precision (Eq. 1), Recall (Eq. 2), F1 Score (Eq. 3), and Accuracy (Eq. 4), which provide a comprehensive view of how the model predicts endometriosis in patients. These metrics are presented in the following equations: 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃 𝑇𝑃 + 𝐹𝑃 (1) 𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃 𝑇𝑃 + 𝐹𝑁 (2) 𝐹1 𝑠𝑐𝑜𝑟𝑒 = 𝑇𝑁 𝑇𝑁 + 𝐹𝑃 (3) 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 (4) IV. RESULTS AND DISCUSSION Figure 4 illustrates the confusion matrices for the three algorithms: RF (Figure 4a), LASSO (Figure 4b), SVM (Figure 4c), and Naive Bayes (Figure 4d). Additionally, Table 5 summarizes the number of correct and incorrect predictions for each trained model. The results of the confusion matrix metrics for the four algorithms applied to endomet riosis detection are as follows: Random Forest achieved the highest precision, with 237 correct predictions of endometriosis and a low error rate (only 2 errors in each category). LASSO and SVM also delivered strong results, although SVM recorded more erro rs (11) in the "No Endometriosis" category. In contrast, Naive Bayes exhibited the highest number of errors across both categories. (a) (b) (c) (d) Fig. 4. Confusion Matrices for RF (a), LASSO (b), SVM (c), and Naive Bayes (d). TABLE V. CONFUSION MATRIX METRICS FOR THE DATASET Algorithm Feature Correct Prediction Incorrect Prediction Total Random Forest 0=’No Endometriosis’ 60 2 62 1=’ Endometriosis’ 237 2 239 LASSO 0=’No Endometriosis’ 60 2 62 1=’ Endometriosis’ 225 14 239 SVM 0=’No Endometriosis’ 51 11 62 1=’ Endometriosis’ 238 1 239 Naive Bayes 0=’No Endometriosis’ 56 6 62 1=’ Endometriosis’ 214 25 239 Figure 5 presents the results of the algorithms based on the area under the curve (AUC) metrics, where a value closer to 1 indicates a more effective classifier. The results reveal that the SVM algorithm (Figure 5c) achieved the highest classification performance with an AUC of 0.99, followed by Random Forest (Figure 5b). (a) (b) (c) (d) Fig. 5. ROC Curve for RF (a), LASSO (b), SVM (c), and Naive Bayes (d). Table VI displays the training metrics for the four algorithms, highlighting that the Random Forest algorithm demonstrated strong performance with high precision, recall, and F1 score for both "no endometriosis" (0) and "endometriosis" (1) cases. This indi cates significant effectiveness in classifying both types of samples. The LASSO algorithm also delivered satisfactory results, showing better precision in classifying the "endometriosis" (1) class. SVM exhibited higher precision for the classification of t he "endometriosis" (1) class. Finally, the Naive Bayes model showed a marked difference in precision between classes, with the "endometriosis" class achieving the highest accuracy. TABLE VI TRAINING METRICS RESULTS FOR THE DATASET Algorithm Feature Precision Recall F1 Score ACC Random Forest 0 0.96 0.96 0.96 0.98 1 0.99 0.98 0.99 LASSO 0 0.81 0.96 0.88 0.96 1 0.94 0.99 0.96 SVM 0 0.98 0.82 0.89 0.96 1 0.95 0.98 0.97 Naive Bayes 0 0.70 0.90 0.78 0.90 1 0.97 0.89 0.89 V. CONCLUSION In this study, a model for the detection of endometriosis was proposed, focusing on symptoms reported by patients and applying four machine learning algorithms: Random Forest, LASSO, SVM, and Naive Bayes. The model was developed using a dataset extracted from the Global Health Data Exchange (GHDx), comprising 1,000 cases of patients with endometriosis. These algorithms were applied and fine-tuned, with 70% of the data used for training and the remaining 30% reserved for evaluation. Key metrics such as precision, recall, accuracy, and F1 -Score were employed to assess the performance of each algorithm, providing a comprehensive analysis of their effectiveness. The results demonstrated that Random Forest was the most effective algorithm, followed by LASSO and SVM, while Naive Bayes exhibited lower performance. Collectively, these findings highlight the feasibility of the proposed approach to enhance early detection of endometriosis and its potential impact on clinical practice. As a future endeavor, it is suggested to develop a tool utilizing the Random Forest algorithm to provide healthcare professionals with a practical and personalized system, transforming clinical practice and delivering significant benefits to a wide range of women of reproductive age. ACKNOWLEDGMENT We extend our gratitude to the Research Department of the Universidad Peruana de Ciencias Aplicadas for their support in the execution of this project. REFERENCES [1] A. Awaysheh, J. Wilcke, F. Elvinger, L. Rees, W. Fan, y K. L. Zimmerman, "Review of medical decision support and machine learning methods," Journal of Veterinary Diagnostic Investigation , vol. 31, no. 1, pp. 1-14, 2019. doi: 10.1177/1040638718801234. [2] A. Leonova, V. E. Turpin, S. K. Agarwal, M. Leonardi, y W. G. Foster, "A critical appraisal of the circulating levels of differentially expressed microRNA in endometriosis," Biology of Reproduction, vol. 105, no. 5, pp. 1075-1085, 2021. doi: 10.1093/biolre/ioab134. [3] B. Jing, Y. Zhang, Y. Liu, Y. Wang, Y. Chen, Y. Zhang, Y. Li, Y. Zhang, Y. Wang, and Y. Liu, "Development of prediction model to estimate future risk of ovarian lesions," Preventive Medicine Reports, vol. 35, p. 102296, 2023. doi: 10.1016/j.pmedr.2023.102296. [4] B. Sivajohan, M. Elgendi, C. Menon, C. Allaire, P. Yong, and M. A. Bedaiwy, "Machine learning algorithms as new screening approach for patients with endometriosis," Sci. Rep., vol. 12, p. 639, 2022. DOI: 10.1038/s41746-022-00638-1 [5] C. Jiang, Y. Jiang, "Lasso Algorithm and Support Vector Machine Strategy to Screen Pulmonary Arterial Hypertension Gene," Scottish Medical Journal , vol. 68, no. 1, pp. 24 -30, 2022. doi: 10.1177/00369330211064645. [6] C. Xie, Z. Yin, and Y. Liu, "Analysis of characteristic genes and ceRNA regulation mechanism of endometriosis based on full transcriptional sequencing," Frontiers in Genetics, vol. 13, p. 902329, 2022. doi: 10.3389/fgene.2022.902329. [7] D. B. Balogh, G. Hudelist, D. Bliznuks, J. Raghothama, C. M. Becker, R. Horace, H. Krentel, A. W. Horne, N. Bourdel, G. Marki, C. Tomassetti, U. B. Kirk, N. Acs, y A. Bokor, "FEMaLe: The use of machine learning for early diagnosis of endometriosis based on patient self-reported data—Study protocol of a multicenter trial," PLoS ONE, vol. 19, no. 5, p. e0300186, 2024. doi: 10.1371/journal.pone.0300186. [8] D. Zou, L. Zhang, Y. Xu, K. Wang, and J. Zhang, "Identification of senescence hub genes and their association with endometriosis," Frontiers in Pharmacology , vol. 14, p. 1259467, 2023. doi: 10.3389/fphar.2023.1259467. [9] F. G. Martire, L. Lazzeri, F. Conway, T. Siciliano, A. Pietropolli, E. Piccione, E. Solima, G. Centini, E. Zupi, and C. Exacoustos, "Adolescence and endometriosis: symptoms, ultrasound signs and early diagnosis," Fertility and Sterility, vol. 114, no. 5, pp. 1049-1057, Nov. 2020. DOI: 10.1016/j.fertnstert.2020.06.041. [10] Global Health Data Exchange, “Global Health Data Exchange,” [Online]. Available: https://ghdx.healthdata.org/. [11] H. Ding, H. Xu, T. Zhang, and C. Shi, "Identification and validation of M2 macrophage -related genes in endometriosis," Heliyon, vol. 9, e22258, 2023. doi: 10.1016/j.heliyon.2023.e22258. [12] H. I. H. Inbarani, A. T. Azar, and J. G., "Leukemia Image Segmentation Using a Hybrid Histogram -Based Soft Covering Rough K -Means Clustering Algorithm," Electronics, vol. 9, no. 188, pp. 1-22, Jan. 2020. doi: 10.3390/electronics9010188. [13] H. Kitchen, C. Seitz, A. Trigg, N. Aldhouse, T. Willgoss, H. Schmitz, A. Gater, C. Gerlinger, and C. Haberland, "Patients’ and clinicians’ perspectives on item importance, scoring, and clinically meaningful differences for the Endometriosis Symptom Diary ( ESD) and Endometriosis Impact Scale (EIS)," Health Qual. Life Outcomes, vol. 19, no. 7, 2021. DOI: 10.1186/s12955 -020-01579-7. [Online]. Available: https://doi.org/10.1186/s12955-020-01579-7 [14] J. She, D. Su, R. Diao, y L. Wang, "A Joint Model of Random Forest and Artificial Neural Network for the Diagnosis of Endometriosis," Frontiers in Genetics , vol. 13, p. 848116, 2022. doi: 10.3389/fgene.2022.848116. [15] J. Wei, B. Huang, W. Liu, Q. Zhang, Y. Nong, W. Wang, X. Liang, Q. Li, Y. Xie, T. Peng, "Identification of cuproptosis-related differentially expressed genes and their clinical significance in endometriosis," Reproductive Sciences , vol. 30, no. 11, pp. 1841 -1853, 2023. doi: 10.1007/s43032-022-01130-7. [16] M. Anantha Satyanarayana, VV Sushma, K Praveen Kumar, V Sudheer Kumar, K Rama Murthy, and T Asha, "Endometriosis: A Clinicopathological Study of 29 Cases," Journal of Clinical and Diagnostic Research, vol. 12, no. 5, pp. EC05-EC09, May 2018. DOI: 10.7860/JCDR/2018/35204.11460. [Online]. Available: https://doi.org/10.7860/JCDR/2018/35204.11460 [17] M. F. Moustafa, A. M. Al-Hendy, A. M. El-Shal, A. A. El-Badawy, A. A. El-Sayed, A. A. El-Sharkawy, A. A. El-Masry, A. A. El-Sayed, and A. A. El-Masry, "Endometriosis diagnosis by microRNA biomarkers," American Journal of Obstetrics & Gynecology , vol. 223, no. 4, pp. 557.e1-557.e12, 2020. doi: 10.1016/j.ajog.2020.06.014. [18] M. Szubert, A. Rycerz, and J. R. Wilczyński, "How to Improve Non - Invasive Diagnosis of Endometriosis with Advanced Statistical Methods," Medicina, vol. 59, no. 3, p. 499, 2023. doi: 10.3390/medicina59030499. [19] R. Khamisy -Farah, L. B. Furstenau, J. D. Kong, J. Wu, and N. L. Bragazzi, "Gynecology Meets Big Data in the Disruptive Innovation Medical Era: State -of-Art and Future Prospects," in International Journal of Environmental Research and Public Health, vol. 18, p. 5058, May 2021. [Online]. Available: https://doi.org/10.3390/ijerph18105058. [20] W. Jin, Q. Yang, H. Chi, K. Wei, P. Zhang, G. Zhao, S. Chen, Z. Xia, and X. Li, "Ensemble deep learning enhanced with self -attention for predicting immunotherapeutic responses to cancers," Front. Immunol., vol. 13, pp. 1025330, Dec. 2022. [Online]. Availab le: https://doi.org/10.3389/fimmu.2022.1025330.