Artificial Intelligence-Driven Prediction of post Neoadjuvant Treatment Toxicities and Biomarkers Identifications in Rectal Cancer

Artificial Intelligence-Driven Prediction of post Neoadjuvant Treatment Toxicities and Biomarkers Identifications in Rectal Cancer | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Artificial Intelligence-Driven Prediction of post Neoadjuvant Treatment Toxicities and Biomarkers Identifications in Rectal Cancer Mariem Chouchen, Chamseddine Barki, Christophe Badie, Afsheen Raza, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8967708/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Neoadjuvant chemotherapy (nCT) and chemoradiotherapy (nCRT) are widely used cancer treatments, however, patients’ response and subsequent clinical toxicity vary substantially. Predicting toxicity using artificial intelligence (AI) allows risk-adapted treatment and overcomes statistical models’ limitations. This study describes a novel approach using two complementary AI models to achieve three objectives. Initially, a fine-tuned pre-trained multilingual BERT model (mBERT), was used to analyse radiological reports and identify patients at risk of developing toxicities post-nCT or nCRT for rectal cancer (RC). Then, a MultiLayer Perceptron (MLP) neural network was developed to predict toxicities and identify its key biomarkers. Our results demonstrate that the mBERT model and MLP algorithm achieved strong performance in classifying patients at risk and predicting toxicities following nCT or nCRT for RC (mBERT: precision = 1, recall = 0.94 and F1-score = 0.97; MLP: accuracy = 0.90, mean squared error = 0.06 and mean absolute error = 0.16). The MLP algorithm identified toxicity biomarkers not previously reported in machine learning models. Furthermore, our study recommends using (y)pTNM staging as a biomarker for potential toxicity. In conclusion, this study presents a new AI approach to classify patients at risk and predict toxicities following nCT or nCRT for RC supporting personalized neoadjuvant therapeutic strategies. Health sciences/Biomarkers Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Deep Learning Large Language Model multilingual BERT MultiLayer Perceptron Neoadjuvant Chemoradiotherapy and Chemotherapy Toxicity Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Colorectal cancer is the second leading cause of cancer-related mortality and the third most diagnosed malignant tumor [ 1 ]. Rectal cancer (RC) accounts for 42% of colorectal cancers, and its incidence tends to gradually increase over the following years [ 2 ]. Neoadjuvant chemotherapy (nCT) and chemoradiotherapy (nCRT) are two of the therapeutic modalities for RC, offering an advantage in reducing the risks of mortality and recurrence [ 3 ]. However, the rate of pathological complete response varies widely among treated patients [ 4 ], between 73 to 85% of patients do not achieve a complete pathological response due to diverse risk factors such as tumor histopathological heterogeneity and individual clinical characteristics [ 5 ]. These inherent risk factors can increase the propensity to develop radiation-induced toxicity regardless of treatment duration [ 6 ] and have a significant impact on patients' quality of life. Serval studies [ 7 ] suggest that the combination of surgery and nCT or nCRT can result in early toxicities (such as diarrhea [ 8 ], urinary tract infections [ 9 ], etc.) as well as late toxicity (such as fistula and hemorrhage [ 6 ]). Moreover, additional complications such as extramural vascular invasion and tumor invasion into the mesorectum [ 10 ]. Therefore, it is crucial to identify and analyze patient-specific biomarkers to assess individual sensitivity and guide towards a personalized neoadjuvant treatment protocol for rectal cancer. Indeed, radiotherapy alters the biological characteristics of tumor cells, contributing to radioresistance through radio-induced hypoxia, activation of tumor-associated fibroblasts and macrophages [ 11 ]. In the same context, the presence of the PREX2 gene alters the effectiveness of radiotherapy by promoting DNA replication errors in the tumor cell. It is a biomarker potentially radioresistance [ 12 ]. Similarly, proteins responsible for immunosuppression can be biomarkers of radioresistance (PITPNC1); the activation of these proteins alters the immune function of lymphocytes (CD8 + T), thus promoting the proliferation of tumor cells [ 13 ]. In addition, the status of the circumferential radical margin (CRM) is a powerful predictor of local and distant recurrence as well as survival rate. Patients with a positive CRM (CRM ≤ 1 mm) have a 35% risk of developing distant metastases [ 14 ]. Metastasis and profiling of RC cells can be promoted by the presence of circular RNA from tumor cell (such as circ-0001313), which can contribute to resistance to chemoradiotherapy [ 15 ]. Also, carcinogenesis can be stimulated by oral microbiomes through disruption and translocation of the intestinal wall, leading to intestinal dysbiosis. These microbiomes can compromise the effectiveness of radiotherapy [ 16 ]. Furthermore, recent studies have shown that patients with T4 stages have a significant risk of anastomotic leakage after nCRT, while immunocompromised patients face a significant risk of radical infection [ 17 , 18 , 19 , 20 ]. Nevertheless, research aimed at predicting the toxicity of neoadjuvant treatment for RC based on clinical, biological, therapeutic, pathological, and radiological data remains limited. Most studies focus primarily on dosimetric data, with a few studies addressed patient specific clinical or biological data, and heterogeneity of patient populations which can restrict the ability to develop robust or individualized predictive models [ 21 , 22 ]. Since the widespread use of electronic radiology reporting systems has provided objective and real-time information on post-chemoradiotherapy toxicity in patients with RC [ 23 ]. Several studies have used these reports to support clinical decision-making by relying on artificial intelligence technologies, such as large language models (LLM) and deep learning [ 24 , 25 ]. In this study, we aim to develop a new approach using two complementary artificial intelligent models to achieve three primary objectives. Firstly, we aim to develop pre-trained LLM to identify and classify patients at risk of developing toxicities following a neoadjuvant treatment for RC using radiological reports. This is followed by a deep learning algorithm to specify the toxicities (early or late). This algorithm is based on tabular dataset that’s includes demographic, comorbidities, dental health status, tumor characteristics (pre-therapeutic), neoadjuvant treatment, surgical parameters, biological biomarkers and post-operative staging. Thirdly, our deep learning algorithm serves to identify key biomarkers associated with the development of these toxicities, providing deeper insights into the biological and clinical indicators of post neoadjuvant treatment of RC. 2. Materials and Methods 2.1. Study Population and Data Collection A public database of patients with RC was obtained from the Mendeley database (https://data.mendeley.com/). The present study dataset represents a newly generated database created by merging two previously published public datasets: One dataset, collected from a single-center study, was used for short-term follow-up of treatment-related toxicities, while the other, derived from a multicenter cohort in Ethiopia, provided long-term toxicity and survival data. We included in this study patients who received a nCT or nCRT with or without subsequent total mesorectal excision (TME). For our LLM we analyze radiology reports data, including positron emission tomography-computed tomography (PET-CT) reports in Spanish, which were selected for their high sensitivity in detecting toxicities [26, 27]. This type of examination is highly recommended in the management and prognosis analysis of RC [28]. A total of 199 PET-CT reports were used to identify patients at risk of toxicities following nCT or nCRT. Our study also utilizes a tabular dataset, including data from 54 patients treated between January 2019 to October 2020 were used to predict early toxicities following nCT or nCRT. In addition, data from 160 patients treated between January 2010 to May 2017, were used to predict late toxicities. The prescribed dose was 45 Gy delivered in 25 fractions. Overall, our tabular dataset included 28 columns representing the input features to predict six toxicities (output features) following the neoadjuvant treatment. The tabular dataset for the final shape is provided in an Excel file. The Python programming language was employed for the deep learning algorithm fitting and analysis, and Python with graphics processing units (GPU) support was used for LLM processing. 2.2. Data Preparation LLM: Corpus Annotation We pre-analyzed PET-CT reports to annotate our dataset. The annotation team consisted of two researchers specializing in radiation therapy and oncology and an expert in medical informatics were responsible for the manual annotation of the corpus. To have a quality annotation corpus, we followed the latest version of the annotation recommendations [29]. Since the objective of our LLM is to identifier patient at risk of developing toxicities following nCT or nCRT, a classification approach was used. To this end, the annotation scheme included two entities: T0 for the absence of toxicities, invasion or distant metastasis and T1 for their presence. Deep Learning Algorithm In our study we generated a new dataset by combining two datasets. This process may result in missing values. To address this issue, we use the k-nearest neighbors imputation method (highly recommended for manipulating healthcare databases) [30]. Our imputation was performed with 5 nearest neighbors (k=5). This generates a new dataset containing 28 input features and six outputs (six toxicities). Table 1 provides an overview of the variables used in this study. Table 1. Overview of Variables of Deep Learning Algorithm Variables Description Input features gender: Gender of patient Age : Current age Tumor type: the type of tumor BMI : Body mass index CV diseases : Cardiovascular diseases L diseases : Lung diseases Smoking : Smoking status NDRT : Dental health status NAT : The type of neoadjuvant therapy cTNM : Pre-treatment clinical staging MRI-CRM : Magnetic resonance imaging -based circumferential radical margin ycTNM : Post-treatment clinical staging Tumor level : Tumor level (cm from anal verge) C or F RNA : Circular or fixed RNA from tumor cell Surgical approach : Type of anastomosis Ileostomy : Temporary intestinal diversion anastomosis Level: Level of the anastomosis (cm) CRP : C-reactive protein level (mg/L) Album : Albumin level (g/L) Prealb : Prealbumin level (g/L) Fe : Iron level (umol/L) (y)pTNM : Post-treatment pathological staging Output features pCRM : Circumferential radial margin (mm) LVI : Lymphovascular invasion Perineural spread : Tumor nerve invasion Response in NAT : Therapy response grading (Dworak classification after neoadjuvant therapy) TD : Gastrointestinal symptoms Dist_metastasis : Distant metastasis 2.3. Model Selection For our LLM, we employ a fine-tuned pre-trained multilingual BERT (mBERT) model for binary classification. This model was chosen for robustness and better performance in biomedical text analysis tasks since mBERT was trained on a rich biomedical corpus [31]. Also, it is open-source, publicly available, and has been extensively pre-trained on radiological reports. For deep learning algorithm, we use the MultiLayer Perceptron (MLP) Neural Network algorithm, as it is recommended for predicting toxicity following nCT or nCRT based on tabulated dataset [32, 33, 34, 35]. Our proposed approach is clearly illustrated in Fig. 1. For this algorithm, we fixed the number of epochs at 300, and a MinMaxScaler normalization method was applied to stabilize the fitting process and eliminate the risk of overfitting or underfitting [36]. Finally, our tabular dataset was divided into a 70% sample for training and a 30% sample for testing and validation, which was also used for our mBERT model. The architecture of the MLP algorithm used in this study is presented in the Table 2. Table 2. MLP Algorithm Structure 2.4. Text Pre-Processing, Encoding and Embedding the Clinical Procedures To preserve the semantic information in our mBERT, we followed a systematic pre-processing pipeline. Each text, representing a PET-CT report was compressed into a ZIP file and converted into a Hugging Face dataset to create a homogeneous corpus. The text is tokenized using an AutoTokenizer from the Hugging Face library with a pre-trained mBERT tokenizer. This choice was made due to the robust tokenization capabilities of mBERT, which demonstrates its place in pre-training on biomedical corpora in Spanish [37, 38]. The mBERT tokenizer segments the report text into meaningful subword units and usable numerical representations. To achieve a balance between character-level and word-level tokenization, allowing efficient management of technical terms related to toxicity following nCT or nCRT for RC, a processing pipeline using truncation and padding to fix the length of 128 tokens was applied via the. map() method. This allows for the standardization of the input and optimizes the training process. 2.5. Hyperparameters Optimization To optimize the hyperparameters of the mBERT we manually defined the training arguments using the Hugging Face’s Trainer API. We set a learning rate of 2e-5, the batch and evaluation size equal to 4, a number of training epochs to 3, and the weight decay regularization equal to 0.01. These parameters were chosen to optimize the model and prevent overfitting. In addition, to optimize the hyperparameters of our MLP algorithm, the loss function was the training mean squared error (MSE) and the Adam optimizer was used. Each dense layer was followed by an L2 regularization. Each L2 regularization was followed by a dropout layer with a rate of 0.1(to prevent overfitting and underfitting [39]). A learning rate of 0.001 was used for our algorithm. 2.6. Model Performance Evaluation To assess our mBERT model, we calculated the accuracy (the number of correctly predicted observations), the recall (the number of data predicted positive and which are truly positive), the precision (rate of correct positive predictions), the F1 score (the ability of an algorithm to correctly predict positive variables) [40], and the loss. To evaluate MLP algorithm, we used the mean absolute error (MAE) to estimate the variance between the true variables and the predicted output values. The mean MSE was calculated as the average of the squared difference between the true variables and the predicted output values [41]. As well as to analyze the overfitting and underfitting, we examine the accuracy curves and the loss curves, for both the training and validation datasets. 3. Results 3.1. LLM Performance Evaluation The mBERT model achieved exceptional metrics in extracting relevant data from radiological reports with accuracy of 0.97, F1 score of 0.97, recall of 0.94, precision of 1 and a low loss rate of 0.29 (91.66% were classified as a true negative and 97% were classified as a true positive). To further evaluate the performance of this model, we compare it with three studies from the literature using pre-trained BERT models that aimed to analysis, predict toxicity, metastasis and therapeutic response following neoadjuvant therapy for RC using radiological reports or clinical notes. In each study, we selected the best performing BERT model; PubMedBERT [42], RadBERT [ 43 ] and MetBERT [ 44 ], which were compared based on F1 score, precision and recall on the test set. The mBERT model developed in this study achieved superior performance outperforming PubMedBERT (precision 0.90, recall 0.89, F1-score 0.87), RadBERT (precision 0.92, recall 0.78, F1-score 0.84), and MetBERT (precision 0.82, recall 0.76, F1-score 0.80). 3.2. Patients Our tabular dataset shows that 92.96% of patients have a rectal adenocarcinoma Grade 1, 38.50% of patients with cT4 stage and 10.33% of patients with cN1 stage. Due to the advanced stages of RC, the majority of patients undergo TME, with 31.46% of patients receiving nCRT. Despite this robust therapeutic protocol, our dataset shows that 24.41% of patients have an incomplete therapeutic response, with 16.43% classified as grade 0 per the Dworak classification after neoadjuvant therapy. As a result, 46% of patients exhibit gastrointestinal symptoms such as diarrhea, constipation, rectovaginal fistula or blood in stool. Furthermore, 30.52% of patients develop distant metastases. Specifically, 11.27% show invasion into perirectal adipose tissue ((y)pT3 staging), 9.39% show lymph node invasion ((y)pN1 staging) and 31.92% have lymphovascular and extramural invasion. These findings demonstrate that the therapeutic response was influenced by several factors that are resistant to radiotherapy or chemotherapy. The characteristics of the RC patients are presented in Table 3 . Table 3 Patient characteristics from tabular dataset Features Number of patients (in %) Gender Female: 15.96 Male: 84.04 Age (years) Mean 50 (16–86) Tumor type Mucinous rectal adenocarcinoma: 3.28 Rectal adenocarcinoma, Grade 1: 92.96 Rectal adenocarcinoma, Grade 2: 3.76 Rectal neuroendocrine tumor: 1 BMI; mean (min-max) 27 (3.13–38.6) Cardiovascular diseases No: 73.71 Acute myocardial infarction: 7.04 Acute Mesenteric Ischemia: 19.25 Lung diseases No: 94.84 Yes: 5.16 Smoking status No: 87.32 Yes: 9.39 Ex-smoker: 3.29 Dental health status Non-defective: 77.46 Restored: 21.13 Dental bridge: 0.94 Dentures below, anodontia above :0.47 Type of neoadjuvant therapy nCT: 68.54 nCRT: 31.46 cTNM cT0: 0.47 cT1: 4.69 cT2: 32.39 cT3: 23.94 cT4: 38.51 cN 0: 84.04 cN1: 10.33 cN2: 5.63 cM0 :98.12 cM1:1.88 MRI-CRM Negative: 92.02 Positive: 7.98 ycTNM ycT0: 88.73 ycT2: 2.82 ycT3: 7.51 ycT4: 0.94 ycN0: 96.24 ycN1: 3.76 ycM0: 99.06 ycM1: 0.94 Tumor level from anal verge (cm); mean (min-max) 7,08 (3,50 − 10) Circular or Fixed RNA No : 94.37 RNA circulaire : 4.22 RNA fixed : 1.41 Surgical approach No : 1.41 TME : 98.59 Ileostomy No : 99.06 Yes :0.94 anastomosis Level (cm); mean (min-max) 3,70 (1.-7) C-reactive protein level (mg/L); mean (min-max) 3,49 (0,50 − 23,7) Albumin level (g/L); mean (min-max) 38,21 (25,20–50,4) Prealbumin (g/L); mean (min-max) 0,254 (0,17–0,3) Iron level Fe (umol/L); mean (min-max) 17,02 (2,1–75) (y)pTNM (y)pT 0 : 77.93 (y)pT1: 1.88 (y)pT2: 7.51 (y)pT3: 11.27 (y)pT4: 1.41 (y)pN0: 88.73 (y)pN1: 9.39 (y)pN2: 1.89 (y)pM0: 98.59 (y)pM1: 1.41 pCRM Negative: 97.18 Positive : 2.82 lymphovascular invasion status No: 68.08 Lymphovascular invasion: 27.23 Extramural vascular invasion: 4.69 Perineural spread Negative: 92.96 Positive: 7.04 Tumor Repine Grade (Dworak classification) Grade 0: 16.43 Grade 1: 2.82 Grade 2: 4.23 Grade 3: 0.94 Grade 4: 75.58 Gastrointestinal symptoms No: 53.99 Diarrhea: 14.09 Constipation :8.92 Diarrhea and constipation: 4.69 Blood stool :16.90 Rectovaginal fistula : 1.41 Distant metastasis No: 69.48 Liver: 25.35 Lung: 1.88 Both: 3.29 3.3. Correlation Matrix between Input and Output Features In order to identify biomarkers that increase or decrease consistently depending on the stage of the disease we measured the monotonic relationship between the input and output features of our MLP algorithm, Fig. 2 presents a heat map plot illustrating the Spearman correlation coefficient [45]. This information is critical for feature selection, ensuring that the MLP focuses on variables that have meaningful and interpretable associations with clinical outcomes [45, 46]. Our results reveal a moderate positive correlation ρ = 0.4 between the (y)pN stage and the pCRM. As the (y)pN stage advances, the probability of having a positive pCRM increases. Additionally, there is a moderate positive correlation between the (y)pN stage and lymphovascular invasion status, ρ = 0.5. As the (y)pN stage become more advanced, the probability to develop lymphovascular invasion increases. For perineural spread a moderate negative correlation ρ = -0.41 exists with the type of neoadjuvant therapy. This suggests that the nCT or nCRT have a protective effect against the development of perineural spread. This output also shows a moderate positive correlation with the (y)pT stage, the dental health status, the (y)pN stage, the cardiovascular disease and the tumor type. In addition, the therapeutic response shows a strong negative correlation ρ = -0.88, ρ = -0.84, and ρ = -0.8 respectively with the (y)pT stage, the cardiovascular disease and the dental health status. Similarly, the type of neoadjuvant therapy and the therapeutic response show a strong positive correlation ρ = 0.71. For gastrointestinal symptoms we observe a weak negative correlation with most of the input features practically with the cardiovascular disease and the (y)pT stage. The distant metastases increase with higher cT stage, showing a strong correlation attaining 0.73. 3.4. MLP Performance Evaluation Through the prediction of 6 toxicities following nCT or nCRT for RC, the results indicate that MLP algorithm performs well, achieving an accuracy of 0.90. Regarding the prediction error, the MLP algorithm achieved very low MSE of 0.06 and MAE of 0.16. Figure 3, support these results, which show that there is no evidence of overfitting or underfitting in the MLP algorithm. This is supported by the fact that the validation loss remains comparable to the training loss, indicating a good balance of dataset. 3.5. Key Biomarkers and Features Importance for Post-Treatment Toxicities To identify the most important features for each output of the MLP algorithm, the Mean Absolute Pearson Correlation (MAPC) was calculated, as it demonstrates in Fig. 4 . Our results indicate that for predicting pCRM, the (y)pN stage is the most informative variable, with MAPC value of 0.45. Furthermore, the MRI-CRM, dental health status, (y)pT stage, and (y)cT stage exhibit limited predictive power for this output, with MAPC values ranging from 0.28 to 0.36. In contrast, for the prediction of lymphovascular invasion status, the (y)pN stage is the most informative variable, with MAPC value of 0.51. This strong correlation indicates that this biomarker plays a major role in prediction of lymphovascular invasion. In addition, for predicting the perineural spread, the (y)pT and the (y)pN stages are the most informative variables, with MAPC values of 0.53 and 0.51 respectively. While, the dental health status, tumor type, cardiovascular disease, RNA tumor status, type of neoadjuvant therapy and cN stage show a moderate MAPC with this output (ranging from 0.40 to 0.49). Also, our results indicate that the most influential feature for predicting therapeutic response ranked from highest to lowest MAPC are (y)pT stage, cardiovascular disease, dental health status, type of neoadjuvant therapy, and cN stage, with MAPC values ranging from 0.52 to 0.79. For predicting the presence of gastrointestinal symptoms, the (y)pT stage, cardiovascular disease, and age are the most informative variables, with a moderate MAPC ranging from 0.22 to 0.29. Finally, the cT stage is the strongest feature to predict distant metastases, with MAPC value of 0.60. 4. Discussion This study aimed to develop an mBERT classification model to identify patients at risk of developing toxicities following nCT or nCRT for RC, using PET-CT reports. In addition, an MLP algorithm is developed to specify the type of toxicities (early or late) to which these patients are vulnerable. Based on the MLP algorithm, we identified the key biomarkers responsible for the development of these toxicities. Among the three pre-trained BERT models developed and evaluated by other studies (PubMedBERT [42], RadBERT [ 43 ] and MetBERT [ 44 ]), our mBERT model is the most effective achieving an F1 score and precision of 0.97, 1 respectively. This superior performance can be attributed to specific limitations affecting other models: the low performance observed with PubMedBERT [42] and RadBERT [ 43 ] models is explained by the limited heterogeneity of their datasets, while the training dataset for MetBERT [ 44 ] presents an imbalance in the distribution of classes. Notably, PubMedBERT model trained on 1,524 clinical notes generally outperformed those trained on larger datasets (4,522 radiology reports in the RadBERT model and 6634 in the MetBERT model). This suggests that model performance could benefit from smaller scale training datasets. To address these limitations, our pre-trained BERT model overcomes. First, the multilingual BERT is tested on a large corpus covering 104 languages ​​using the masked language modelling method [ 47 ], effectively overcoming the limitation posed by the Spanish language in our PET-CT reports. In addition, our model is pre-trained on the smallest dataset among the other LLM models evaluated in this study. This approach is supported by numerous studies showing that accuracy improves with smaller training sets of radiological reports (even with 50 cases) and shorter training time [ 25 ]. This method reduces the heterogeneity of these reports (inaccurate or inconsistent reports) distorting the annotation [ 48 ]. In addition, our mBERT model demonstrates its strength in terms of identifying patients at risk of developing toxicities, by reducing false negatives and false positives, thereby outperforming other LLM models. Our mBERT undergoes fine-tuning by adapting an LLM model trained on a large dataset to a specific task. This method allows us to achieve performance exceeding 0.9 even with a limited dataset [ 49 ]. Moreover, this fine-tuning significantly improves our results, leading to increase the performance of our mBERT with 4.61 points on F1 score [ 50 ]. Subsequently, performing intermediate fine-tuning enables the LLM models to integrate specialized knowledge, thus reinforcing the effectiveness of our small dataset [51]. In parallel, our mBERT model utilizes the WordPiece tokenization method, which allows better management of out-of-vocabulary words by using byte pair encoding [ 52 ]. To improve Our mBERT model, we developed a MLP algorithm to predict six types of toxicities following nCT or nCRT for RC using 28 features. Our MLP algorithm outperformed other machine learning approaches, achieving higher accuracy in predicting therapeutic response for RC and metastases (0.90 vs. 0.83 with convolutional neural network and 89,63% with random forest algorithm) [53, 54]. This strong performance can be attributed to its architecture with four hidden layers (128, 64, 32 and 12 neurons) with ReLU activation functions, optimized via stochastic gradient descent (learning rate = 0.001) and L2 regularization [55]. This architecture empowered our MLP algorithm to learn nuanced patterns from the 28 features and toxicity outcomes yielding an impressive MSE of 0.06 and MAE of 0.16 on the test set. By integrating diverse inputs, our MLP algorithm provides radiotherapists and clinicians with data driven insights to personalize nCT or nCRT and better protect patients with RC. For this reason, we analyze the spearman correlation between input and output features as well as the mean absolute pearson correlation to identify the most influential features for each MLP output prediction. Figure 5 , represents the main biomarkers of toxicities identified by the MLP algorithm. Our results show that both biomarkers (y)pN and pCRM reflect tumor aggressiveness and prognosis (rs = 0.45, ρ = 0.4). Patients with high (y)pN stage positivity face a 26% risk of local recurrence after 7 years due to invaded radical margin [ 56 ]. Meanwhile, patients with a pCRM < 1 mm negatively influence the survival and quality of life of patients (the risk of mortality will be multiplied by 7) [ 57 ]. To address this pCRM monitoring is performed by MRI-CRM (rs = 0.37 and ρ = 0.37) allows to determine the probability and identify patients at risk of having a positive pCRM [58]. For lymphovascular invasion prediction, the (y)pN stage is the most informative variable for our MLP algorithm, (rs = 0.51, ρ = 0.5). The risk of developing this type of toxicity increases with the ypTNM staging [ 59 ] and influenced by circular or fixed RNA from tumor cells which affect proliferation, invasion, metastasis, and apoptosis [ 15 ]. For perineural spread prediction, (y)pT and (y)pN, are the most informative variable (rs = 0.53 and 0.51, ρ = 0.51, ρ = 0.49, respectively). Patients with advanced ypT (ypT2) influence perineural spread [60]. In addition, there is a strong relationship between perineural invasion and dental health status (ρ = 0.51, rs = 0.49) due to Fusobacterium nucleatum which promotes tumor cell migration and metastases by generating chronic inflammation [ 61 ]. Thus, the perineural progression is influenced by the histological type of tumor (ρ = 0.46, rs = 0.48) [ 62 ]. Our results demonstrate also that an effective choice of therapeutic approach based on the characteristics of the patient positively influences the therapeutic response (ρ = 0.71, rs = 0.67). This can prevent toxicity following nCT or nCRT [63]. Furthermore, dental health status strongly influences the therapeutic response. The persistence of Fusobacterium nucleatum in cases of locally advanced RC shows a significant risk of recurrence. This type of microbiome is resistant to chemotherapy and to radiotherapy. It promotes carcinogenesis, inflammatory processes and stimulates autophagy [ 64 ]. Furthermore, our results demonstrate that elderly patients with a history of cardiovascular disease are at risk of anastomotic fistula due to atherosclerosis affecting the mesenteric vessels, which compromises the blood supply to the anastomosis area [ 65 ]. Also, chemoradiotherapy is accompanied by diarrhea and constipation, as well as rectovaginal fistula induced by surgery [66]. Moreover, distant metastases increase with higher cT stage (ρ = 0.73 and rs = 0.60). Patients with advanced cT stage have a risk of developing metastases. [ 67 ]. In this context, several studies have improved the potential of genome-wide association studies (GWAS) to analyze and interpret toxicities following radiation therapy for rectal cancer [ 68 ]. GWAS is utilized to identify genetic variants (SNP) [ 68 ]. These SNP are markers of severe cutaneous toxicities of cetuximab (an EGFR inhibitor) [69,70]. Additionally, other researchers have identified genes such as MROH5, which are associated with neutropenia (a drop in white blood cells in patients receiving XELOX) and markers near the SLC26A7 gene that are linked to vomiting caused by oxaliplatin and fluoropyrimidine chemotherapy (including FOLFOX and XELOX) [ 71 , 72 ]. These studies increase the power to detect genetic factors by combining datasets and are crucial for personalized treatments by predicting who might experience severe side effects, allowing for targeted interventions or prophylactic measures. Building on this, our MLP algorithm can identify new biomarkers responsible for toxicity (late or early) that are not commonly used in machine learning models, such as circular or fixed RNA from the tumor cells, dental health status and comorbidities (cardiovascular disease). Since, all these biomarkers influence the neoadjuvant therapeutic response in RC, our study recommends using (y)pTNM staging to assess potential toxicity including lymphovascular invasion, analyze the therapeutic response, look for signs of gastrointestinal or identify a perineural spread. Blinding on these foundations, our study introduces a pioneering approach involving two complementary intelligent models. A pre-trained mBERT model to identify patients at risk of developing toxicities following nCT or nCRT for RC using PET-CT reports. This is followed by MLP algorithms to specify the toxicities (early or late) to which they are vulnerable by identifying up to six different toxicities based on the 28 features. This study is a complementary solution to help radiotherapists and clinicians establish personalized neoadjuvant therapeutic strategies tailored to the characteristics of each patient (comorbidities, histological type and tumor stage). Our study not only helps reduce clinical errors and calibration errors (even if clinicians and radiotherapists are competent) but also improves medical decision-making. It can serve as a basis for a clinical decision support system based on a pre-trained mBERT model and MLP algorithm. Although promising results and relevant clinical information were obtained, this study has some limitations. To confirm the performance of our models, a prospective multicenter study is necessary. Thus, predicting post-treatment toxicity is inherently complex due to many intertwined features. The use of different data sets by our MLP approach mitigates these difficulties but limits their interpretability by its "black box" nature. Thus, our toxicity assessment is limited to the use of cross-sectional data. Furthermore, our toxicity prediction models are not yet authorized by health authorities, and they are not practically applied in the radiotherapy departments. 5. Conclusion Our study presents a comprehensive pipeline that combines patients-risk identification, toxicity prediction and biomarkers identification of each toxicity following nCT or nCRT for RC. It demonstrates the use of a pre-trained mBERT model and an MLP algorithm in the assessment of nCT or nCRT for RC protocols. Our study innovates by its approach and methodology which are based on PET-CT reports, as well as demographic and anthropometric data, comorbidities, dental health status, tumor characteristics (pre-therapeutic), neoadjuvant treatment, surgical parameters, biological biomarkers and post-operative staging. Our models demonstrate their performance in terms of sensitivity, specificity and precision. Thus, their ability to identify new biomarkers for each type of late or early toxicity. This study is a complementary solution to help radiotherapists and clinicians establish personalized neoadjuvant therapeutic strategies for RC tailored to each patient's characteristics. It also serves as a basis for decisions regarding the patient during future oncology consultations. Declarations Acknowledgements : The authors of this study would like to express their sincere gratitude and acknowledge the valuable resources provided by Alexander Ferko for his dataset available on Mendeley Data, and by Hamid Y. Hassen, Mohammed A. Teka, Jemal Beksisa, and Jilcha D. Feyisa for their dataset also available on Mendeley Data. These datasets were very useful for our study. Authors' contributions : Conceptualization, M.C., H.B.R. and C.Bo.; Methodology, original draft preparation, M.C., H.B.R. and C.Bo.; Software development and data processing, M.C., and H.B.R ; Resources, writing, M.C., H.B.R., A.R., A.G. and C.Bo.; Writing and Original draft preparation, M.C., H.B.R., A.R., A.G. and C.Bo.; Supervision, H.B.R, A.R., A.G., C.Ba. Y.M., and C.Bo.; Original draft preparation, Review and editing, M.C., H.B.R., A.R., A.G., C.Ba., Y.M. and C.Bo. All authors have read and agreed to the published version of the manuscript. Availability of data and materials : The data sets created in this study are private, contained within the extracted datasets (Excel tables and ZIP files), and are not publicly available. They can be shared upon reasonable request from the authors. Conflict of interest : The authors of this study declare no conflicts of interest. Funding : This research received no external funding. Ethics Approval and Consent to Participate: This study involved secondary analysis of two previously published and publicly available anonymized datasets obtained from the Mendeley Data repository. The first dataset (short-term follow-up toxicities), derived from a single-center study, received ethical approval from the Jessenius Faculty of Medicine Comenius University Bratislava in Martin, Institutional Review Board (IRB No. IRB00005636; Approval Number EK1/2019). All patients included in that study provided written informed consent prior to participation. The second dataset (long-term toxicity), derived from a multicenter cohort in Ethiopia, received ethical clearance from the Institutional Review Board of Saint Paul's Hospital Millennium Medical College and Addis Ababa University, College of Health Sciences (Ref. No: PM23/92). Written informed consent was obtained from patients or their caretakers before data collection. Patient data were anonymized, and confidentiality was ensured throughout the data collection and processing stages. No new patients were recruited, and no direct interaction with human participants occurred in this study. All analyzed data were fully anonymized prior to access and contained no identifiable information. Therefore, no additional institutional ethical approval was required for this secondary analysis, in accordance with international research standards. All research procedures were conducted in accordance with the ethical principles of the Declaration of Helsinki and relevant guidelines and regulations. Identifying Information: No identifiable personal information, clinical images, or protected health information (HIPAA identifiers) are included in this manuscript or supplementary materials. References Matsuda, T., Fujimoto, A., & Igarashi, Y. (2025). Colorectal Cancer: Epidemiology, Risk Factors, and Public Health Strategies. Digestion.106 (2): 91–99. https://doi.org/10.1159/000543921 Li, L. B., Wang, L. Y., Chen, D. M., Liu, Y. X., Zhang, Y. H., Song, W. X. & Ma, Z. Y. (2023). A systematic analysis of the global and regional burden of colon and rectum cancer and the difference between early-and late-onset CRC from 1990 to 2019. Frontiers in Oncology, 13, 1102673.https://doi.org/10.3389/fonc.2023.1102673 Gandini, A., Sciallero, S., Martelli, V., Pirrone, C., Puglisi, S., Cremante, M., ... & Pastorino, A. (2025). A Comprehensive Approach to Neoadjuvant Treatment of Locally Advanced Rectal Cancer. Cancers, 17(2), 330. https://doi.org/10.3390/cancers17020330 Guo, Y., Guo, Z., Zhang, J., Qian, G., Ji, W., Song, L., ... & Han, Z. (2025). Short-and long-term outcomes of neoadjuvant chemotherapy compared with neoadjuvant chemoradiotherapy for locally advanced rectal cancer: an updated meta-analysis. BMC gastroenterology, 25(1), 87. Qian, C., Yang, S., Chen, Y., Ge, R., Shi, F., Liu, C., ... & Guo, Y. (2025). Predicting pathological response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer with two step feature selection and ensemble learning. Scientific Reports, 15(1), 9936. Mahdavi, H., Dashti, S., & Jafari, S. (2025). Strategies to reduce intestinal toxicity in neoadjuvant management of locally advanced rectal cancer. Discover Oncology , 16 (1), 1886. https://doi.org/10.1007/s12672-025-03526-8 Jung, W. B. (2024). Beyond survival: a comprehensive review of quality of life in rectal cancer patients. Annals of Coloproctology, 40(6), 527. https://doi.org/10.3393/ac.2024.00745.0106 Joye, I., & Haustermans, K. (2014). Early and late toxicity of radiotherapy for rectal cancer. Early gastrointestinal cancers II: rectal cancer , 189-201. Yang, J., Luo, Y., Tian, T., Dong, P., & Fu, Z. (2022). Effects of Neoadjuvant Radiotherapy on Postoperative Complications in Rectal Cancer: A Meta‐Analysis. Journal of Oncology , 2022 (1), 8197701. https://doi.org/10.1155/2022/8197701 10. Tawfik, M. A. M., Gheida, U. E. S., Khedr, R. A., & Khadrah, R. S. (2025). Multiparametric MRI in the assessment of the response of locally advanced rectal cancer to neoadjuvant chemoradiotherapy. Tanta Medical Journal, 53(2), 161-170. Jain, S. M., Nagainallur Ravichandran, S., Murali Kumar, M., Banerjee, A., Sun-Zhang, A., Zhang, H., ... & Pathak, S. (2024). Understanding the molecular mechanism responsible for developing therapeutic radiation-induced radioresistance of rectal cancer and improving the clinical outcomes of radiotherapy-A review. Cancer Biology & Therapy, 25(1), 2317999. Li, M., Xiao, J., Song, S., Han, F., Liu, H., Lin, Y., ... & Liang, L. (2024). PREX2 contributes to radiation resistance by inhibiting radiotherapy-induced tumor immunogenicity via cGAS/STING/IFNs pathway in colorectal cancer. BMC medicine, 22(1), 154. Liang, J., Liao, L., Xie, L., Tang, W., Yu, X., Lu, Y., ... & Tan, Y. (2024). PITPNC1 Suppress CD8+ T cell immune function and promote radioresistance in rectal cancer by modulating FASN/CD155. Journal of Translational Medicine, 22(1), 117. Ma, Y., Ma, D., Xu, X., Li, J., & Guan, Z. (2024). Progress of MRI in predicting the circumferential resection margin of rectal cancer: a narrative review. Asian Journal of Surgery, 47(5), 2122-2131. Hussen, B. M., Abdullah, S. R., Mohammed, A. A., Rasul, M. F., Hussein, A. M., Eslami, S., ... & Taheri, M. (2024). Advanced strategies of targeting circular RNAs as therapeutic approaches in colorectal cancer drug resistance. Pathology-Research and Practice, 260, 155402. Kudra, A., Muszyński, D., Sobocki, B. K., Atzeni, A., Carbone, L., Kaźmierczak-Siedlecka, K., ... & Kalinowski, L. (2023). Insights into oral microbiome and colorectal cancer–on the way of searching new perspectives. Frontiers in Cellular and Infection Microbiology, 13, 1159822. Zeng, D. X., Liu, R. N., Ren, X. K., Zhang, P., Tang, L. H., Tan, L., ... & Xiao, J. W. (2025). Comparison of the efficacy of neoadjuvant chemotherapy and neoadjuvant chemoradiotherapy in locally advanced rectal cancer patients: meta-analysis of randomized controlled trials. International Journal of Surgery , 111 (3), 2686-2696. Tao, W., Cheng, Y., Wang, P., Wen, H., & Xiao, W. (2025). Comparison of LNR-and LODDS-based predictive models for prognosis in non-elderly patients with locally advanced rectal cancer undergoing neoadjuvant therapy. International Journal of Colorectal Disease , 40 (1), 157. https://doi.org/10.1007/s00384-025-04942-6 Yu, X. N., Xu, L. M., Bin, Y. W., Yuan, Y., Tian, S. B., Cai, B., ... & Wang, Z. (2022). Risk factors of anastomotic leakage after anterior resection for rectal cancer patients. Current Medical Science , 42 (6), 1256-1266. https://doi.org/10.1007/s11596-022-2616-2 Ju, H. E., Lee, C. S., Bae, J. H., Lee, H. J., ran Yoon, M., Al-Sawat, A., ... & Han, S. R. (2022). High incidence of late anastomosis leakage in patients for rectal cancer after neoadjuvant chemoradiotherapy: A comparative study. Asian Journal of Surgery , 45 (10), 1832-1842. https://doi.org/10.1016/j.asjsur.2021.10.039 Corbeau, A., Heemsbergen, W. D., Kuipers, S. C., Godart, J., Creutzberg, C. L., Nout, R. A., & de Boer, S. M. (2024). Predictive factors for toxicity after primary chemoradiation for locally advanced cervical cancer: a systematic review. International Journal of Radiation Oncology* Biology* Physics , 119 (1), 127-142. https://doi.org/10.1016/j.ijrobp.2023.11.010 Tohidinezhad, F., Willems, Y., Berbee, M., Van Limbergen, E., Verhaegen, F., Dekker, A., & Traverso, A. (2022). Prediction models for brachytherapy-induced rectal toxicity in patients with locally advanced pelvic cancers: a systematic review. Journal of Contemporary Brachytherapy , 14 (4), 411-422. https://doi.org/10.5114/jcb.2022.119427 Asoglu, O., Tokmak, H., Bakir, B., Demir, G., Ozyar, E., Atalar, B., ... & Guven, K. (2020). The impact of total neo-adjuvant treatment on nonoperative management in patients with locally advanced rectal cancer: The evaluation of 66 cases. European Journal of Surgical Oncology, 46(3), 402-409. Liu, C., Liu, Z., Holmes, J., Zhang, L., Zhang, L., Ding, Y., ... & Liu, W. (2023). Artificial general intelligence for radiation oncology. Meta-radiology, 1(3), 100045. Yan, A., McAuley, J., Lu, X., Du, J., Chang, E. Y., Gentili, A., & Hsu, C. N. (2022). RadBERT: adapting transformer-based language models to radiology. Radiology: Artificial Intelligence, 4(4), e210258. Ma, W., Chen, B., Zhu, F., Yang, C., & Yang, J. (2024). Diagnostic role of F-18 FDG PET/CT in determining preoperative Lymph node status of patients with rectal cancer: a meta-analysis. Abdominal Radiology, 49(6), 2125-2134. Kawamura, M., Kamomae, T., Yanagawa, M., Kamagata, K., Fujita, S., Ueda, D., ... & Naganawa, S. (2024). Revolutionizing radiation therapy: the role of AI in clinical practice. Journal of radiation research, 65(1), 1-9. Dadgar, H., Norouzbeigi, N., Jafari, E., Al-balooshi, B., Al-Ibraheem, A., Haidar, M., ... & Assadi, M. (2024). Exploring the efficacy of FAPI PET/CT in the diagnosis and treatment management of colorectal cancer: a comprehensive literature review and initial experience. Clinical and Translational Imaging, 12(3), 235-252. Barabadi, M. A., Chan, W. Y., Zhu, X., Simpson, A. L., & Do, R. K. (2024, August). Adapting large language models for automatic annotation of radiology reports for metastases detection. In 2024 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE) (pp. 340-345). IEEE. Joel, L. O., Doorsamy, W., & Paul, B. S. (2024). On the performance of imputation techniques for missing values on healthcare datasets. arXiv preprint arXiv:2403.14687. Karabacak, M., & Margetis, K. (2023). Embracing large language models for medical applications: opportunities and challenges. Cureus , 15 (5). Lee, S., Lim, J., Shin, J., Kim, S., & Hwang, H. (2021). Pathologic complete response prediction after neoadjuvant chemoradiation therapy for rectal cancer using radiomics and deep embedding network of MRI. Applied Sciences, 11(20), 9494. Li, J., Zhang, W., Chen, L., Mao, X., Wang, X., Liu, J., ... & Wang, T. (2024). SNPs and blood inflammatory marker featured machine learning for predicting the efficacy of fluorouracil-based chemotherapy in colorectal cancer. Scientific Reports, 14(1), 27700. Lou, X., Zhou, N., Feng, L., Li, Z., Fang, Y., Fan, X., ... & Huang, Y. (2022). Deep learning model for predicting the pathological complete response to neoadjuvant chemoradiotherapy of locally advanced rectal cancer. Frontiers in Oncology, 12, 807264. YAKAR, M., ETİZ, D., BADAK, B., ÇELİK, Ö., KÜTRİ, D., ÖZEN, A., & YILMAZ, E. (2021). Prediction of Response to Neoadjuvant Chemoradiotherapy with Machine Learning in Rectal Cancer: A Pilot Study. TURK ONKOLOJI DERGISI, 36(4). Szmigiel, A., Apel, D. B., Wang, J., Pu, Y., Wojtecki, L., & Pourrahimian, Y. (2024). Enhancing open stope stability prediction in mining engineering: Optimal configuration of an artificial neural network model. Journal of Industrial Safety, 1(1), 100008. Solarte-Pabón, O., Montenegro, O., García-Barragán, A., Torrente, M., Provencio, M., Menasalvas, E., & Robles, V. (2023). Transformers for extracting breast cancer information from Spanish clinical narratives. Artificial Intelligence in Medicine, 143, 102625. Gardazi, N. M., Daud, A., Malik, M. K., Bukhari, A., Alsahfi, T., & Alshemaimri, B. (2025). BERT applications in natural language processing: a review. Artificial Intelligence Review, 58(6), 1-49. Santos, C. F. G. D., & Papa, J. P. (2022). Avoiding overfitting: A survey on regularization methods for convolutional neural networks. ACM Computing Surveys (Csur), 54(10s), 1-25. Kozak, J., Probierz, B., Kania, K., & Juszczuk, P. (2022). Preference-driven classification measure. Entropy, 24(4), 531. Jierula, A., Wang, S., Oh, T. M., & Wang, P. (2021). Study on accuracy metrics for evaluating the predictions of damage locations in deep piles using artificial neural networks with acoustic emission data. Applied Sciences, 11(5), 2314. Chen, S., Guevara, M., Ramirez, N., Murray, A., Warner, J. L., Aerts, H. J., ... & Bitterman, D. S. (2023). Natural language processing to automatically extract the presence and severity of esophagitis in notes of patients undergoing radiotherapy. JCO Clinical Cancer Informatics, 7, e2300048. Tay, S. B., Low, G. H., Wong, G. J. E., Tey, H. J., Leong, F. L., Li, C., ... & Tan, R. S. Y. C. (2024). Use of natural language processing to infer sites of metastatic disease from radiology reports at scale. JCO Clinical Cancer Informatics, 8, e2300122. Liu, K., Kulkarni, O., Witteveen-Lane, M., Chen, B., & Chesla, D. (2022). MetBERT: a generalizable and pre-trained deep learning model for the prediction of metastatic cancer from clinical notes. AMIA Summits on Translational Science Proceedings, 2022, 331. Ryu, J., & Cho, S. B. (2002, January). Towards optimal feature and classifier for gene expression classification of cancer. In AFSS International Conference on Fuzzy Systems (pp. 310-317). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-45631-7_41 Guo, Q. H., Xie, F. C., Zhong, F. M., Wen, W., Zhang, X. R., Yu, X. J., ... & Wang, X. Z. (2024). Application of interpretable machine learning algorithms to predict distant metastasis in ovarian clear cell carcinoma. Cancer Medicine , 13 (7), e7161. https://doi.org/10.1002/cam4.7161 Khan, W., Leem, S., See, K. B., Wong, J. K., Zhang, S., & Fang, R. (2025). A comprehensive survey of foundation models in medicine. IEEE Reviews in Biomedical Engineering. Tejani, A. S., Ng, Y. S., Xi, Y., Fielding, J. R., Browning, T. G., & Rayan, J. C. (2022). Performance of multiple pretrained BERT models to automate and accelerate data annotation for large datasets. Radiology: Artificial Intelligence, 4(4), e220007. Bai, H., Zhang, W., Hou, L., Shang, L., Jin, J., Jiang, X., ... & King, I. (2020). Binarybert: Pushing the limit of bert quantization. arXiv preprint arXiv:2012.15701. Beltagy, I., Lo, K., & Cohan, A. (2019). SciBERT: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676. Kalyan, K. S., Rajasekharan, A., & Sangeetha, S. (2022). AMMU: a survey of transformer-based biomedical pretrained language models. Journal of biomedical informatics, 126, 103982. Mutuvi, S., Boros, E., Doucet, A., Lejeune, G., Jatowt, A., & Odeo, M. (2020, December). Multilingual epidemiological text classification: a comparative study. In COLING, International Conference on Computational Linguistics (pp. 6172-6183). Shi, L., Zhang, Y., Nie, K. E., Sun, X., Niu, T., Yue, N., ... & Su, M. Y. (2019). Machine learning for prediction of chemoradiation therapy response in rectal cancer using pre-treatment and mid-radiation multi-parametric MRI. Magnetic resonance imaging, 61, 33-40. Volovat, C. C., Scripcariu, D. V., Boboc, D., Volovat, S. R., Vasilache, I. A., Ursulescu-Lupascu, C., ... & Scripcariu, V. (2024). Machine Learning-Based Algorithms for Enhanced Prediction of Local Recurrence and Metastasis in Low Rectal Adenocarcinoma Using Imaging, Surgical, and Pathological Data. Diagnostics, 14(6), 625. Salehin, I., & Kang, D. K. (2023). A review on dropout regularization approaches for deep neural networks within the scholarly domain. Electronics, 12(14), 3106. Sorrentino, L., Sileo, A., Daveri, E., Battaglia, L., Guaglio, M., Centonze, G., ... & Cosimelli, M. (2023). Impact of Microscopically Positive (≤ 1 mm) Distal Margins on Disease Recurrence in Rectal Cancer Treated by Neoadjuvant Chemoradiotherapy. Cancers, 15(6), 1828. Calvo, F. A., Tudela, M., Serrano, J., Muñoz-Fernández, M., Peligros, M. I., Garcia-Alfonso, P., & Del Valle, E. (2023). Post-chemoradiation metastatic, persistent and resistant nodes in locally advanced rectal cancer: metrics and their impact on long-term outcome. Cancers, 15(18), 4591. Volovat, C. C., Scripcariu, D. V., Boboc, D., Volovat, S. R., Vasilache, I. A., Lupascu-Ursulescu, C., ... & Scripcariu, V. (2024). Predicting the Feasibility of Curative Resection in Low Rectal Cancer: Insights from a Prospective Observational Study on Preoperative Magnetic Resonance Imaging Accuracy. Medicina, 60(2), 330. Kim, Y. I., Cho, H., Kim, C. W., Park, Y., Kim, J., Ro, J. S., ... & Kim, J. C. (2021). Prognostic impact of extranodal extension in rectal cancer patients undergoing radical resection after preoperative chemoradiotherapy. Clinical colorectal cancer, 20(1), e35-e42. Cui, Y., Song, M., Tie, J., Li, S., Wang, H., Zhang, Y., ... & Wang, W. (2024). Clinicopathological factors predict residual lymph node metastasis in locally advanced rectal cancer with ypT0-2 after neoadjuvant chemoradiotherapy. Journal of Cancer Research and Clinical Oncology, 150(4), 176. Chiscuzzu, F., Crescio, C., Varrucciu, S., Rizzo, D., Sali, M., Delogu, G., & Bussu, F. (2025). Current Evidence on the Relation Between Microbiota and Oral Cancer—The Role of Fusobacterium nucleatum—A Narrative Review. Cancers, 17(2), 171. Que, Y., Wu, R., Li, H., & Lu, J. (2024). A prediction nomogram for perineural invasion in colorectal cancer patients: a retrospective study. BMC surgery, 24(1), 80. Kokaine, L., Gardovskis, A., & Gardovskis, J. (2021). Evaluation and predictive factors of complete response in rectal cancer after neoadjuvant chemoradiation therapy. Medicina, 57(10), 1044. Dong, J., Li, Y., Xiao, H., Zhang, S., Wang, B., Wang, H., ... & Cui, M. (2021). Oral microbiota affects the efficacy and prognosis of radiotherapy for colorectal cancer in mouse models. Cell reports, 37(4). Niu, R., Jiang, Y., Bi, Z., Zhang, H., Mei, X., Bi, J., ... & Liang, J. (2024). The effect of cardiovascular disease on the perioperative period of radical surgery in elderly rectal cancer. BMC gastroenterology, 24(1), 256. Varela, C., & Kim, N. K. (2021). Surgical treatment of low-lying rectal cancer: updates. Annals of coloproctology, 37(6), 395. Lai, I. L., You, J. F., Chern, Y. J., Tsai, W. S., Chiang, J. M., Hsieh, P. S., ... & Hsu, Y. J. (2021). The risk factors of local recurrence and distant metastasis on pT1/T2N0 mid-low rectal cancer after total mesorectal excision. World Journal of Surgical Oncology, 19, 1-7. Kerns, S. L., Stock, R. G., Stone, N. N., Rath, L., Vega, A., Fachal, L., ... & Ostrer, H. (2013). Genome-wide association study identifies a region on chromosome 11q14. 3 associated with late rectal bleeding following radiation therapy for prostate cancer. Radiotherapy and Oncology, 107(3), 372-376. https://doi.org/10.1016/j.radonc.2013.05.001 Benitez, C. M., & Knox, S. J. (2020). Harnessing genome-wide association studies to minimize adverse radiation-induced side effects. Radiation Oncology Journal, 38(4), 226. https://doi.org/10.3857/roj.2020.00556 Fernández-Mateos, J., Seijas-Tamayo, R., Mesía, R., Taberna, M., Borgoñón, M. P., Pérez-Ruiz, E., ... & Cruz-Hernández, J. J. (2016). Epidermal growth factor receptor (EGFR) pathway polymorphisms as predictive markers of cetuximab toxicity in locally advanced head and neck squamous cell carcinoma (HNSCC) in a Spanish population. Oral oncology, 63, 38-43. https://doi.org/10.1016/j.oraloncology.2016.10.006 Watts, K. (2023). Genome-wide analyses to identify biomarkers of toxicity in the treatment of advanced colorectal cancer (Doctoral dissertation, Cardiff University). https://orca.cardiff.ac.uk/id/eprint/163425 Watts, K., Wills, C., Madi, A., Palles, C., Maughan, T. S., Kaplan, R., ... & Cheadle, J. P. (2021). Genome‐wide association studies of toxicity to oxaliplatin and fluoropyrimidine chemotherapy with or without cetuximab in 1800 patients with advanced colorectal cancer. International journal of cancer, 149(9), 1713-1722. https://doi.org/10.1002/ijc.33739 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 31 Mar, 2026 Reviews received at journal 30 Mar, 2026 Reviewers agreed at journal 30 Mar, 2026 Reviews received at journal 14 Mar, 2026 Reviewers agreed at journal 04 Mar, 2026 Reviewers invited by journal 03 Mar, 2026 Editor assigned by journal 03 Mar, 2026 Editor invited by journal 03 Mar, 2026 Submission checks completed at journal 01 Mar, 2026 First submitted to journal 01 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8967708","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":600642750,"identity":"36824e82-d8e2-4012-8810-7bf63d4fef34","order_by":0,"name":"Mariem Chouchen","email":"","orcid":"","institution":"University of Tunis El Manar","correspondingAuthor":false,"prefix":"","firstName":"Mariem","middleName":"","lastName":"Chouchen","suffix":""},{"id":600642751,"identity":"9e0a865f-5229-4e15-986e-fb84d3881f3c","order_by":1,"name":"Chamseddine Barki","email":"","orcid":"","institution":"University of Tunis El Manar","correspondingAuthor":false,"prefix":"","firstName":"Chamseddine","middleName":"","lastName":"Barki","suffix":""},{"id":600642752,"identity":"e127792a-246f-4b16-9492-fbb99c2ae053","order_by":2,"name":"Christophe Badie","email":"","orcid":"","institution":"UK Health Security Agency","correspondingAuthor":false,"prefix":"","firstName":"Christophe","middleName":"","lastName":"Badie","suffix":""},{"id":600642753,"identity":"25a30ba0-bbd6-45b4-b718-5e4f244935a5","order_by":3,"name":"Afsheen Raza","email":"","orcid":"","institution":"Abu Dhabi University","correspondingAuthor":false,"prefix":"","firstName":"Afsheen","middleName":"","lastName":"Raza","suffix":""},{"id":600642754,"identity":"063eed6a-4312-48f8-a6ea-ee0e38f4af83","order_by":4,"name":"Amal Gadalla","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAUlEQVRIiWNgGAWjYHACxgMgooGduYHhAwMzcXogWpgZGxhnkKyFmYcYLfzTDj84wPDLTra/mbHxs02NdWID++EHDD8qcGuRuJ1mcICxL9l4xmHGZumcY+mJDTxpBow9Z/BYczsBqKWHObHhMGODdA7b4cQGhhwGBt423Drkb6d/AGqpT5wPtOW3xT+gFv43DIx//+HWYnA7x+AAw4/DiRsOM7ZJM7YBtUjkMDDzNuDWYng7p+BAYsNx441ALZa9fenGbRLPDA7LHMOtRe52+sYHH/5Uy8473nz4xo9v1rL9/MkPH76pweN9EEhE9iwbAySmCIA/hJWMglEwCkbBCAYAOw9ZlezOW38AAAAASUVORK5CYII=","orcid":"","institution":"Abu Dhabi University","correspondingAuthor":true,"prefix":"","firstName":"Amal","middleName":"","lastName":"Gadalla","suffix":""},{"id":600642755,"identity":"708aa986-9886-4ae7-bb37-b77a87eef0da","order_by":5,"name":"Yasser Maghrbi","email":"","orcid":"","institution":"Gulf University for Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Yasser","middleName":"","lastName":"Maghrbi","suffix":""},{"id":600642756,"identity":"eb87eb33-21e6-4088-802a-e82269b22690","order_by":6,"name":"Hanene Boussi","email":"","orcid":"","institution":"University of Tunis El Manar","correspondingAuthor":false,"prefix":"","firstName":"Hanene","middleName":"","lastName":"Boussi","suffix":""}],"badges":[],"createdAt":"2026-02-25 12:23:36","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8967708/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8967708/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104404663,"identity":"9f1d5f9c-3b40-4309-b31d-5f7f7af625df","added_by":"auto","created_at":"2026-03-11 12:20:46","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":365376,"visible":true,"origin":"","legend":"\u003cp\u003eAI-driven Workflow for Predicting Early and Late Toxicity in Neoadjuvant RC Therapy\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8967708/v1/b259cf09bcc7412f916de3e5.png"},{"id":104204270,"identity":"a2a05d38-a517-4bb1-8f29-98f18b11ad2c","added_by":"auto","created_at":"2026-03-09 06:32:42","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":200985,"visible":true,"origin":"","legend":"\u003cp\u003eSpearman Correlation matrix\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8967708/v1/6c354fade2c9818416f5ed66.png"},{"id":104204272,"identity":"36ad4855-0370-423b-befb-eedcd3e1a06a","added_by":"auto","created_at":"2026-03-09 06:32:43","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":75200,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy and Loss Curve of MLP algorithm\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8967708/v1/8e44319095b7d802b9d5e9aa.png"},{"id":104204268,"identity":"2dad7449-522f-402f-a7be-25c2a9578eb1","added_by":"auto","created_at":"2026-03-09 06:32:42","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":170574,"visible":true,"origin":"","legend":"\u003cp\u003eMean Absolute Pearson Correlation: The Most Important Features for Prediction\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8967708/v1/459b9cdd944043f3a521a1bd.png"},{"id":104779573,"identity":"895d7a95-3654-4d9b-bd90-3598b2fef96a","added_by":"auto","created_at":"2026-03-17 07:42:28","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":444913,"visible":true,"origin":"","legend":"\u003cp\u003eToxicity Biomarkers Map in RC\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8967708/v1/44a4c7fdec9b428ede148016.png"},{"id":104783992,"identity":"381a623b-9d94-419c-a215-92fa663b86dc","added_by":"auto","created_at":"2026-03-17 08:04:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2241082,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8967708/v1/9da89987-ba6a-4f9a-a3cd-3f1110bb0f24.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial Intelligence-Driven Prediction of post Neoadjuvant Treatment Toxicities and Biomarkers Identifications in Rectal Cancer","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eColorectal cancer is the second leading cause of cancer-related mortality and the third most diagnosed malignant tumor [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Rectal cancer (RC) accounts for 42% of colorectal cancers, and its incidence tends to gradually increase over the following years [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Neoadjuvant chemotherapy (nCT) and chemoradiotherapy (nCRT) are two of the therapeutic modalities for RC, offering an advantage in reducing the risks of mortality and recurrence [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. However, the rate of pathological complete response varies widely among treated patients [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], between 73 to 85% of patients do not achieve a complete pathological response due to diverse risk factors such as tumor histopathological heterogeneity and individual clinical characteristics [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. These inherent risk factors can increase the propensity to develop radiation-induced toxicity regardless of treatment duration [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] and have a significant impact on patients' quality of life. Serval studies [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] suggest that the combination of surgery and nCT or nCRT can result in early toxicities (such as diarrhea [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], urinary tract infections [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], etc.) as well as late toxicity (such as fistula and hemorrhage [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]). Moreover, additional complications such as extramural vascular invasion and tumor invasion into the mesorectum [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Therefore, it is crucial to identify and analyze patient-specific biomarkers to assess individual sensitivity and guide towards a personalized neoadjuvant treatment protocol for rectal cancer. Indeed, radiotherapy alters the biological characteristics of tumor cells, contributing to radioresistance through radio-induced hypoxia, activation of tumor-associated fibroblasts and macrophages [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In the same context, the presence of the PREX2 gene alters the effectiveness of radiotherapy by promoting DNA replication errors in the tumor cell. It is a biomarker potentially radioresistance [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Similarly, proteins responsible for immunosuppression can be biomarkers of radioresistance (PITPNC1); the activation of these proteins alters the immune function of lymphocytes (CD8\u0026thinsp;+\u0026thinsp;T), thus promoting the proliferation of tumor cells [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In addition, the status of the circumferential radical margin (CRM) is a powerful predictor of local and distant recurrence as well as survival rate. Patients with a positive CRM (CRM\u0026thinsp;\u0026le;\u0026thinsp;1 mm) have a 35% risk of developing distant metastases [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Metastasis and profiling of RC cells can be promoted by the presence of circular RNA from tumor cell (such as circ-0001313), which can contribute to resistance to chemoradiotherapy [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Also, carcinogenesis can be stimulated by oral microbiomes through disruption and translocation of the intestinal wall, leading to intestinal dysbiosis. These microbiomes can compromise the effectiveness of radiotherapy [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Furthermore, recent studies have shown that patients with T4 stages have a significant risk of anastomotic leakage after nCRT, while immunocompromised patients face a significant risk of radical infection [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Nevertheless, research aimed at predicting the toxicity of neoadjuvant treatment for RC based on clinical, biological, therapeutic, pathological, and radiological data remains limited. Most studies focus primarily on dosimetric data, with a few studies addressed patient specific clinical or biological data, and heterogeneity of patient populations which can restrict the ability to develop robust or individualized predictive models [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Since the widespread use of electronic radiology reporting systems has provided objective and real-time information on post-chemoradiotherapy toxicity in patients with RC [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Several studies have used these reports to support clinical decision-making by relying on artificial intelligence technologies, such as large language models (LLM) and deep learning [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. In this study, we aim to develop a new approach using two complementary artificial intelligent models to achieve three primary objectives. Firstly, we aim to develop pre-trained LLM to identify and classify patients at risk of developing toxicities following a neoadjuvant treatment for RC using radiological reports. This is followed by a deep learning algorithm to specify the toxicities (early or late). This algorithm is based on tabular dataset that\u0026rsquo;s includes demographic, comorbidities, dental health status, tumor characteristics (pre-therapeutic), neoadjuvant treatment, surgical parameters, biological biomarkers and post-operative staging. Thirdly, our deep learning algorithm serves to identify key biomarkers associated with the development of these toxicities, providing deeper insights into the biological and clinical indicators of post neoadjuvant treatment of RC.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003ch2\u003e2.1. Study Population and Data Collection\u003c/h2\u003e\n\u003cp\u003eA public database of patients with RC was obtained from the Mendeley database (https://data.mendeley.com/). The present study dataset represents a newly generated database created by merging two previously published public datasets:\u003c/p\u003e\n\u003cp\u003eOne dataset, collected from a single-center study, was used for short-term follow-up of treatment-related toxicities, while the other, derived from a multicenter cohort in Ethiopia, provided long-term toxicity and survival data. We included in this study patients who received a nCT or nCRT with or without subsequent total mesorectal excision (TME). For our LLM we analyze radiology reports data, including positron emission tomography-computed tomography (PET-CT) reports in Spanish, which were selected for their high sensitivity in detecting toxicities [26, 27]. This type of examination is highly recommended in the management and prognosis analysis of RC [28]. A total of 199 PET-CT reports were used to identify patients at risk of toxicities following nCT or nCRT. Our study also utilizes a tabular dataset, including data from 54 patients treated between January 2019 to October 2020 were used to predict early toxicities following nCT or nCRT. \u0026nbsp; In addition, data from 160 patients treated between January 2010 to May 2017, were used to predict late toxicities. The prescribed dose was 45 Gy delivered in 25 fractions. Overall, our tabular dataset included 28 columns representing the input features to predict six toxicities (output features) following the neoadjuvant treatment. \u0026nbsp;The tabular dataset for the final shape is provided in an Excel file. The Python programming language was employed for the deep learning algorithm fitting and analysis, and Python with graphics processing units (GPU) support was used for LLM processing.\u003c/p\u003e\n\u003ch2\u003e2.2. Data Preparation\u003c/h2\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eLLM: Corpus Annotation\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eWe pre-analyzed PET-CT reports to annotate our dataset. The annotation team consisted of two researchers specializing in radiation therapy and oncology and an expert in medical informatics were responsible for the manual annotation of the corpus. To have a quality annotation corpus, we followed the latest version of the annotation recommendations [29]. Since the objective of our LLM is to identifier patient at risk of developing toxicities following nCT or nCRT, a classification approach was used. To this end, the annotation scheme included two entities: T0 for the absence of toxicities, invasion or distant metastasis and T1 for their presence.\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eDeep Learning Algorithm\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eIn our study we generated a new dataset by combining two datasets. This process may result in missing values. To address this issue, we use the k-nearest neighbors imputation method (highly recommended for manipulating healthcare databases) [30]. Our imputation was performed with 5 nearest neighbors (k=5). This generates a new dataset containing 28 input features and six outputs (six toxicities). Table 1 provides an overview of the variables used in this study.\u003c/p\u003e\n\u003cp\u003eTable 1. Overview of Variables of Deep Learning Algorithm\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"652\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 219px;\"\u003e\n \u003cp\u003eVariables\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"22\" valign=\"top\" style=\"width: 219px;\"\u003e\n \u003cp\u003eInput features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003egender: Gender of patient\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eAge : Current age\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eTumor type: the type of tumor\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eBMI : Body mass index\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eCV diseases\u0026nbsp;: Cardiovascular diseases\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eL\u0026nbsp;diseases\u0026nbsp;: Lung diseases\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eSmoking :\u0026nbsp;Smoking status\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eNDRT :\u0026nbsp;Dental health status\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eNAT\u0026nbsp;: The type of neoadjuvant therapy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003ecTNM\u0026nbsp;: Pre-treatment clinical staging\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eMRI-CRM\u0026nbsp;: Magnetic resonance imaging -based circumferential radical margin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eycTNM\u0026nbsp;: Post-treatment clinical staging\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eTumor level\u0026nbsp;: Tumor level (cm from anal verge)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eC or F RNA\u0026nbsp;: Circular or fixed RNA from tumor cell\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eSurgical approach : Type of anastomosis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eIleostomy : Temporary intestinal diversion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eanastomosis Level: Level of the anastomosis (cm)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eCRP\u0026nbsp;: C-reactive protein level (mg/L)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eAlbum\u0026nbsp;: Albumin level (g/L)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003ePrealb\u0026nbsp;: Prealbumin level (g/L)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eFe\u0026nbsp;: Iron level (umol/L)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003e(y)pTNM\u0026nbsp;: Post-treatment pathological staging\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"6\" valign=\"top\" style=\"width: 219px;\"\u003e\n \u003cp\u003eOutput features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003epCRM\u0026nbsp;: Circumferential radial margin (mm)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eLVI\u0026nbsp;: Lymphovascular invasion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003ePerineural spread\u0026nbsp;: Tumor nerve invasion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eResponse in NAT\u0026nbsp;: Therapy response grading (Dworak classification after neoadjuvant therapy)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eTD :\u0026nbsp;Gastrointestinal symptoms\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 433px;\"\u003e\n \u003cp\u003eDist_metastasis :\u0026nbsp;Distant metastasis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e2.3. Model Selection\u003c/h2\u003e\n\u003cp\u003eFor our LLM, we employ a fine-tuned pre-trained multilingual BERT (mBERT) model for binary classification. This model was chosen for robustness and better performance in biomedical text analysis tasks since mBERT was trained on a rich biomedical corpus [31]. Also, it is open-source, publicly available, and has been extensively pre-trained on radiological reports. \u0026nbsp;For deep learning algorithm, we use the MultiLayer Perceptron (MLP) Neural Network algorithm, as it is recommended for predicting toxicity following nCT or nCRT based on tabulated dataset [32, 33, 34, 35]. Our proposed approach is clearly illustrated in Fig. 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor this algorithm, we fixed the number of epochs at 300, and a MinMaxScaler normalization method was applied to stabilize the fitting process and eliminate the risk of overfitting or underfitting [36]. \u0026nbsp; Finally, our tabular dataset was divided into a 70% sample for training and a 30% sample for testing and validation, which was also used for our mBERT model. The architecture of the MLP algorithm used in this study is presented in the Table 2.\u003c/p\u003e\n\u003cp\u003eTable 2. MLP Algorithm Structure\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"721\" height=\"169\"\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e2.4. Text Pre-Processing, Encoding and Embedding the Clinical Procedures\u003c/h2\u003e\n\u003cp\u003eTo preserve the semantic information in our mBERT, we followed a systematic pre-processing pipeline. Each text, representing a PET-CT report was compressed into a ZIP file and converted into a Hugging Face dataset to create a homogeneous corpus. The text is tokenized using an AutoTokenizer from the Hugging Face library with a pre-trained\u0026nbsp;mBERT tokenizer. This choice was made due to the robust tokenization capabilities of mBERT, which demonstrates its place in pre-training on biomedical corpora in Spanish [37, 38].\u0026nbsp;The mBERT tokenizer segments the report text into meaningful subword units and usable numerical representations. To achieve a balance between character-level and word-level tokenization, allowing efficient management of technical terms related to toxicity following nCT or nCRT for RC, a processing pipeline using truncation and padding to fix the length of 128 tokens was applied via the. map() method. This allows for the standardization of the input and optimizes the training process.\u003c/p\u003e\n\u003ch2\u003e2.5. Hyperparameters Optimization\u003c/h2\u003e\n\u003cp\u003eTo optimize the hyperparameters of the mBERT we manually defined the training arguments using the Hugging Face\u0026rsquo;s Trainer API. We set a learning rate of 2e-5, the batch and evaluation size equal to 4, a number of training epochs to 3, and the weight decay regularization equal to 0.01. These parameters were chosen to optimize the model and prevent overfitting. In addition, to optimize the hyperparameters of our MLP algorithm, the loss function was the training mean squared error (MSE) and the Adam optimizer was used. Each dense layer was followed by an L2 regularization. Each L2 regularization was followed by a dropout layer with a rate of 0.1(to prevent overfitting and underfitting [39]). \u0026nbsp;A learning rate of 0.001 was used for our algorithm. \u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e2.6. Model Performance Evaluation\u003c/h2\u003e\n\u003cp\u003eTo assess our mBERT model, we calculated the accuracy (the number of correctly predicted observations), the recall (the number of data predicted positive and which are truly positive), the precision (rate of correct positive predictions), the F1 score (the ability of an algorithm to correctly predict positive variables) [40], and the loss. To evaluate MLP algorithm, we used the mean absolute error (MAE) to estimate the variance between the true variables and the predicted output values. The mean MSE was calculated as the average of the squared difference between the true variables and the predicted output values [41]. As well as to analyze the overfitting and underfitting, we examine the accuracy curves and the loss curves, for both the training and validation datasets.\u003c/p\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1. LLM Performance Evaluation\u003c/h2\u003e \u003cp\u003eThe mBERT model achieved exceptional metrics in extracting relevant data from radiological reports with accuracy of 0.97, F1 score of 0.97, recall of 0.94, precision of 1 and a low loss rate of 0.29 (91.66% were classified as a true negative and 97% were classified as a true positive). To further evaluate the performance of this model, we compare it with three studies from the literature using pre-trained BERT models that aimed to analysis, predict toxicity, metastasis and therapeutic response following neoadjuvant therapy for RC using radiological reports or clinical notes. In each study, we selected the best performing BERT model; PubMedBERT [42], RadBERT [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] and MetBERT [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e], which were compared based on F1 score, precision and recall on the test set. The mBERT model developed in this study achieved superior performance outperforming PubMedBERT (precision 0.90, recall 0.89, F1-score 0.87), RadBERT (precision 0.92, recall 0.78, F1-score 0.84), and MetBERT (precision 0.82, recall 0.76, F1-score 0.80).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Patients\u003c/h2\u003e \u003cp\u003eOur tabular dataset shows that 92.96% of patients have a rectal adenocarcinoma Grade 1, 38.50% of patients with cT4 stage and 10.33% of patients with cN1 stage. Due to the advanced stages of RC, the majority of patients undergo TME, with 31.46% of patients receiving nCRT. Despite this robust therapeutic protocol, our dataset shows that 24.41% of patients have an incomplete therapeutic response, with 16.43% classified as grade 0 per the Dworak classification after neoadjuvant therapy. As a result, 46% of patients exhibit gastrointestinal symptoms such as diarrhea, constipation, rectovaginal fistula or blood in stool. Furthermore, 30.52% of patients develop distant metastases. Specifically, 11.27% show invasion into perirectal adipose tissue ((y)pT3 staging), 9.39% show lymph node invasion ((y)pN1 staging) and 31.92% have lymphovascular and extramural invasion. These findings demonstrate that the therapeutic response was influenced by several factors that are resistant to radiotherapy or chemotherapy. The characteristics of the RC patients are presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatient characteristics from tabular dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeatures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNumber of patients (in %)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGender\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eFemale: 15.96\u003c/p\u003e \u003cp\u003eMale: 84.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAge (years)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eMean 50 (16\u0026ndash;86)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTumor type\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eMucinous rectal adenocarcinoma: 3.28\u003c/p\u003e \u003cp\u003eRectal adenocarcinoma, Grade 1: 92.96\u003c/p\u003e \u003cp\u003eRectal adenocarcinoma, Grade 2: 3.76\u003c/p\u003e \u003cp\u003eRectal neuroendocrine tumor: 1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eBMI; mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e27 (3.13\u0026ndash;38.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCardiovascular diseases\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo: 73.71\u003c/p\u003e \u003cp\u003eAcute myocardial infarction: 7.04\u003c/p\u003e \u003cp\u003eAcute Mesenteric Ischemia: 19.25\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLung diseases\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo: 94.84\u003c/p\u003e \u003cp\u003eYes: 5.16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSmoking status\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo: 87.32\u003c/p\u003e \u003cp\u003eYes: 9.39\u003c/p\u003e \u003cp\u003eEx-smoker: 3.29\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDental health status\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNon-defective: 77.46\u003c/p\u003e \u003cp\u003eRestored: 21.13\u003c/p\u003e \u003cp\u003eDental bridge: 0.94\u003c/p\u003e \u003cp\u003eDentures below, anodontia above :0.47\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eType of neoadjuvant therapy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003enCT: 68.54\u003c/p\u003e \u003cp\u003enCRT: 31.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ecTNM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ecT0: 0.47\u003c/p\u003e \u003cp\u003ecT1: 4.69\u003c/p\u003e \u003cp\u003ecT2: 32.39\u003c/p\u003e \u003cp\u003ecT3: 23.94\u003c/p\u003e \u003cp\u003ecT4: 38.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecN 0: 84.04\u003c/p\u003e \u003cp\u003ecN1: 10.33\u003c/p\u003e \u003cp\u003ecN2: 5.63\u003c/p\u003e \u003cp\u003ecM0 :98.12\u003c/p\u003e \u003cp\u003ecM1:1.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMRI-CRM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNegative: 92.02\u003c/p\u003e \u003cp\u003ePositive: 7.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eycTNM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eycT0: 88.73\u003c/p\u003e \u003cp\u003eycT2: 2.82\u003c/p\u003e \u003cp\u003eycT3: 7.51\u003c/p\u003e \u003cp\u003eycT4: 0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eycN0: 96.24\u003c/p\u003e \u003cp\u003eycN1: 3.76\u003c/p\u003e \u003cp\u003eycM0: 99.06\u003c/p\u003e \u003cp\u003eycM1: 0.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTumor level from anal verge (cm); mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e7,08 (3,50\u0026thinsp;\u0026minus;\u0026thinsp;10)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCircular or Fixed RNA\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo : 94.37\u003c/p\u003e \u003cp\u003eRNA circulaire : 4.22\u003c/p\u003e \u003cp\u003eRNA fixed : 1.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSurgical approach\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo\u0026nbsp;: 1.41\u003c/p\u003e \u003cp\u003eTME\u0026nbsp;: 98.59\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIleostomy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo\u0026nbsp;: 99.06\u003c/p\u003e \u003cp\u003eYes\u0026nbsp;:0.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eanastomosis Level (cm); mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e3,70 (1.-7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eC-reactive protein level (mg/L); mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e3,49 (0,50\u0026thinsp;\u0026minus;\u0026thinsp;23,7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAlbumin level (g/L); mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e38,21 (25,20\u0026ndash;50,4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrealbumin (g/L); mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e0,254 (0,17\u0026ndash;0,3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIron level Fe (umol/L); mean (min-max)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e17,02 (2,1\u0026ndash;75)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e(y)pTNM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(y)pT 0 : 77.93\u003c/p\u003e \u003cp\u003e(y)pT1: 1.88\u003c/p\u003e \u003cp\u003e(y)pT2: 7.51\u003c/p\u003e \u003cp\u003e(y)pT3: 11.27\u003c/p\u003e \u003cp\u003e(y)pT4: 1.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e(y)pN0: 88.73\u003c/p\u003e \u003cp\u003e(y)pN1: 9.39\u003c/p\u003e \u003cp\u003e(y)pN2: 1.89\u003c/p\u003e \u003cp\u003e(y)pM0: 98.59\u003c/p\u003e \u003cp\u003e(y)pM1: 1.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003epCRM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNegative: 97.18\u003c/p\u003e \u003cp\u003ePositive : 2.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003elymphovascular invasion status\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo: 68.08\u003c/p\u003e \u003cp\u003eLymphovascular invasion: 27.23\u003c/p\u003e \u003cp\u003eExtramural vascular invasion: 4.69\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePerineural spread\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNegative: 92.96\u003c/p\u003e \u003cp\u003ePositive: 7.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTumor Repine Grade (Dworak classification)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eGrade 0: 16.43\u003c/p\u003e \u003cp\u003eGrade 1: 2.82\u003c/p\u003e \u003cp\u003eGrade 2: 4.23\u003c/p\u003e \u003cp\u003eGrade 3: 0.94\u003c/p\u003e \u003cp\u003eGrade 4: 75.58\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGastrointestinal symptoms\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo: 53.99\u003c/p\u003e \u003cp\u003eDiarrhea: 14.09\u003c/p\u003e \u003cp\u003eConstipation\u0026nbsp;:8.92\u003c/p\u003e \u003cp\u003eDiarrhea and constipation: 4.69\u003c/p\u003e \u003cp\u003eBlood stool\u0026nbsp;:16.90\u003c/p\u003e \u003cp\u003eRectovaginal fistula\u0026nbsp;: 1.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDistant metastasis\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eNo: 69.48\u003c/p\u003e \u003cp\u003eLiver: 25.35\u003c/p\u003e \u003cp\u003eLung: 1.88\u003c/p\u003e \u003cp\u003eBoth: 3.29\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Correlation Matrix between Input and Output Features\u003c/h2\u003e \u003cp\u003eIn order to identify biomarkers that increase or decrease consistently depending on the stage of the disease we measured the monotonic relationship between the input and output features of our MLP algorithm, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents a heat map plot illustrating the Spearman correlation coefficient [45]. This information is critical for feature selection, ensuring that the MLP focuses on variables that have meaningful and interpretable associations with clinical outcomes [45, 46].\u003c/p\u003e \u003cp\u003eOur results reveal a moderate positive correlation ρ\u0026thinsp;=\u0026thinsp;0.4 between the (y)pN stage and the pCRM. As the (y)pN stage advances, the probability of having a positive pCRM increases. Additionally, there is a moderate positive correlation between the (y)pN stage and lymphovascular invasion status, ρ\u0026thinsp;=\u0026thinsp;0.5. As the (y)pN stage become more advanced, the probability to develop lymphovascular invasion increases. For perineural spread a moderate negative correlation ρ = -0.41 exists with the type of neoadjuvant therapy. This suggests that the nCT or nCRT have a protective effect against the development of perineural spread. This output also shows a moderate positive correlation with the (y)pT stage, the dental health status, the (y)pN stage, the cardiovascular disease and the tumor type. In addition, the therapeutic response shows a strong negative correlation ρ = -0.88, ρ = -0.84, and ρ = -0.8 respectively with the (y)pT stage, the cardiovascular disease and the dental health status. Similarly, the type of neoadjuvant therapy and the therapeutic response show a strong positive correlation ρ\u0026thinsp;=\u0026thinsp;0.71. For gastrointestinal symptoms we observe a weak negative correlation with most of the input features practically with the cardiovascular disease and the (y)pT stage. The distant metastases increase with higher cT stage, showing a strong correlation attaining 0.73.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4. MLP Performance Evaluation\u003c/h2\u003e \u003cp\u003eThrough the prediction of 6 toxicities following nCT or nCRT for RC, the results indicate that MLP algorithm performs well, achieving an accuracy of 0.90. Regarding the prediction error, the MLP algorithm achieved very low MSE of 0.06 and MAE of 0.16. Figure\u0026nbsp;3, support these results, which show that there is no evidence of overfitting or underfitting in the MLP algorithm. This is supported by the fact that the validation loss remains comparable to the training loss, indicating a good balance of dataset.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.5. Key Biomarkers and Features Importance for Post-Treatment Toxicities\u003c/h2\u003e \u003cp\u003eTo identify the most important features for each output of the MLP algorithm, the Mean Absolute Pearson Correlation (MAPC) was calculated, as it demonstrates in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e. Our results indicate that for predicting pCRM, the (y)pN stage is the most informative variable, with MAPC value of 0.45. Furthermore, the MRI-CRM, dental health status, (y)pT stage, and (y)cT stage exhibit limited predictive power for this output, with MAPC values ranging from 0.28 to 0.36. In contrast, for the prediction of lymphovascular invasion status, the (y)pN stage is the most informative variable, with MAPC value of 0.51. This strong correlation indicates that this biomarker plays a major role in prediction of lymphovascular invasion. In addition, for predicting the perineural spread, the (y)pT and the (y)pN stages are the most informative variables, with MAPC values of 0.53 and 0.51 respectively. While, the dental health status, tumor type, cardiovascular disease, RNA tumor status, type of neoadjuvant therapy and cN stage show a moderate MAPC with this output (ranging from 0.40 to 0.49). Also, our results indicate that the most influential feature for predicting therapeutic response ranked from highest to lowest MAPC are (y)pT stage, cardiovascular disease, dental health status, type of neoadjuvant therapy, and cN stage, with MAPC values ranging from 0.52 to 0.79. For predicting the presence of gastrointestinal symptoms, the (y)pT stage, cardiovascular disease, and age are the most informative variables, with a moderate MAPC ranging from 0.22 to 0.29. Finally, the cT stage is the strongest feature to predict distant metastases, with MAPC value of 0.60.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis study aimed to develop an mBERT classification model to identify patients at risk of developing toxicities following nCT or nCRT for RC, using PET-CT reports. In addition, an MLP algorithm is developed to specify the type of toxicities (early or late) to which these patients are vulnerable. Based on the MLP algorithm, we identified the key biomarkers responsible for the development of these toxicities. Among the three pre-trained BERT models developed and evaluated by other studies (PubMedBERT [42], RadBERT [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] and MetBERT [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]), our mBERT model is the most effective achieving an F1 score and precision of 0.97, 1 respectively. This superior performance can be attributed to specific limitations affecting other models: the low performance observed with PubMedBERT [42] and RadBERT [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] models is explained by the limited heterogeneity of their datasets, while the training dataset for MetBERT [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e] presents an imbalance in the distribution of classes. Notably, PubMedBERT model trained on 1,524 clinical notes generally outperformed those trained on larger datasets (4,522 radiology reports in the RadBERT model and 6634 in the MetBERT model). This suggests that model performance could benefit from smaller scale training datasets. To address these limitations, our pre-trained BERT model overcomes. First, the multilingual BERT is tested on a large corpus covering 104 languages ​​using the masked language modelling method [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e47\u003c/span\u003e], effectively overcoming the limitation posed by the Spanish language in our PET-CT reports. In addition, our model is pre-trained on the smallest dataset among the other LLM models evaluated in this study. This approach is supported by numerous studies showing that accuracy improves with smaller training sets of radiological reports (even with 50 cases) and shorter training time [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. This method reduces the heterogeneity of these reports (inaccurate or inconsistent reports) distorting the annotation [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. In addition, our mBERT model demonstrates its strength in terms of identifying patients at risk of developing toxicities, by reducing false negatives and false positives, thereby outperforming other LLM models. Our mBERT undergoes fine-tuning by adapting an LLM model trained on a large dataset to a specific task. This method allows us to achieve performance exceeding 0.9 even with a limited dataset [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Moreover, this fine-tuning significantly improves our results, leading to increase the performance of our mBERT with 4.61 points on F1 score [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. Subsequently, performing intermediate fine-tuning enables the LLM models to integrate specialized knowledge, thus reinforcing the effectiveness of our small dataset [51]. In parallel, our mBERT model utilizes the WordPiece tokenization method, which allows better management of out-of-vocabulary words by using byte pair encoding [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e52\u003c/span\u003e]. To improve Our mBERT model, we developed a MLP algorithm to predict six types of toxicities following nCT or nCRT for RC using 28 features.\u003c/p\u003e \u003cp\u003eOur MLP algorithm outperformed other machine learning approaches, achieving higher accuracy in predicting therapeutic response for RC and metastases (0.90 vs. 0.83 with convolutional neural network and 89,63% with random forest algorithm) [53, 54]. This strong performance can be attributed to its architecture with four hidden layers (128, 64, 32 and 12 neurons) with ReLU activation functions, optimized via stochastic gradient descent (learning rate\u0026thinsp;=\u0026thinsp;0.001) and L2 regularization [55]. This architecture empowered our MLP algorithm to learn nuanced patterns from the 28 features and toxicity outcomes yielding an impressive MSE of 0.06 and MAE of 0.16 on the test set. By integrating diverse inputs, our MLP algorithm provides radiotherapists and clinicians with data driven insights to personalize nCT or nCRT and better protect patients with RC. For this reason, we analyze the spearman correlation between input and output features as well as the mean absolute pearson correlation to identify the most influential features for each MLP output prediction. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e, represents the main biomarkers of toxicities identified by the MLP algorithm.\u003c/p\u003e \u003cp\u003eOur results show that both biomarkers (y)pN and pCRM reflect tumor aggressiveness and prognosis (rs\u0026thinsp;=\u0026thinsp;0.45, ρ\u0026thinsp;=\u0026thinsp;0.4). Patients with high (y)pN stage positivity face a 26% risk of local recurrence after 7 years due to invaded radical margin [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e56\u003c/span\u003e]. Meanwhile, patients with a pCRM\u0026thinsp;\u0026lt;\u0026thinsp;1 mm negatively influence the survival and quality of life of patients (the risk of mortality will be multiplied by 7) [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e]. To address this pCRM monitoring is performed by MRI-CRM (rs\u0026thinsp;=\u0026thinsp;0.37 and ρ\u0026thinsp;=\u0026thinsp;0.37) allows to determine the probability and identify patients at risk of having a positive pCRM [58]. For lymphovascular invasion prediction, the (y)pN stage is the most informative variable for our MLP algorithm, (rs\u0026thinsp;=\u0026thinsp;0.51, ρ\u0026thinsp;=\u0026thinsp;0.5). The risk of developing this type of toxicity increases with the ypTNM staging [\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e59\u003c/span\u003e] and influenced by circular or fixed RNA from tumor cells which affect proliferation, invasion, metastasis, and apoptosis [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. For perineural spread prediction, (y)pT and (y)pN, are the most informative variable (rs\u0026thinsp;=\u0026thinsp;0.53 and 0.51, ρ\u0026thinsp;=\u0026thinsp;0.51, ρ\u0026thinsp;=\u0026thinsp;0.49, respectively). Patients with advanced ypT (ypT2) influence perineural spread [60]. In addition, there is a strong relationship between perineural invasion and dental health status (ρ\u0026thinsp;=\u0026thinsp;0.51, rs\u0026thinsp;=\u0026thinsp;0.49) due to Fusobacterium nucleatum which promotes tumor cell migration and metastases by generating chronic inflammation [\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e]. Thus, the perineural progression is influenced by the histological type of tumor (ρ\u0026thinsp;=\u0026thinsp;0.46, rs\u0026thinsp;=\u0026thinsp;0.48) [\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e]. Our results demonstrate also that an effective choice of therapeutic approach based on the characteristics of the patient positively influences the therapeutic response (ρ\u0026thinsp;=\u0026thinsp;0.71, rs\u0026thinsp;=\u0026thinsp;0.67). This can prevent toxicity following nCT or nCRT [63]. Furthermore, dental health status strongly influences the therapeutic response. The persistence of Fusobacterium nucleatum in cases of locally advanced RC shows a significant risk of recurrence. This type of microbiome is resistant to chemotherapy and to radiotherapy. It promotes carcinogenesis, inflammatory processes and stimulates autophagy [\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e64\u003c/span\u003e]. Furthermore, our results demonstrate that elderly patients with a history of cardiovascular disease are at risk of anastomotic fistula due to atherosclerosis affecting the mesenteric vessels, which compromises the blood supply to the anastomosis area [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e]. Also, chemoradiotherapy is accompanied by diarrhea and constipation, as well as rectovaginal fistula induced by surgery [66]. Moreover, distant metastases increase with higher cT stage (ρ\u0026thinsp;=\u0026thinsp;0.73 and rs\u0026thinsp;=\u0026thinsp;0.60). Patients with advanced cT stage have a risk of developing metastases. [\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e67\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn this context, several studies have improved the potential of genome-wide association studies (GWAS) to analyze and interpret toxicities following radiation therapy for rectal cancer [\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e]. GWAS is utilized to identify genetic variants (SNP) [\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e]. These SNP are markers of severe cutaneous toxicities of cetuximab (an EGFR inhibitor) [69,70]. Additionally, other researchers have identified genes such as MROH5, which are associated with neutropenia (a drop in white blood cells in patients receiving XELOX) and markers near the SLC26A7 gene that are linked to vomiting caused by oxaliplatin and fluoropyrimidine chemotherapy (including FOLFOX and XELOX) [\u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e71\u003c/span\u003e, \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e72\u003c/span\u003e]. These studies increase the power to detect genetic factors by combining datasets and are crucial for personalized treatments by predicting who might experience severe side effects, allowing for targeted interventions or prophylactic measures.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBuilding on this, our MLP algorithm can identify new biomarkers responsible for toxicity (late or early) that are not commonly used in machine learning models, such as circular or fixed RNA from the tumor cells, dental health status and comorbidities (cardiovascular disease). Since, all these biomarkers influence the neoadjuvant therapeutic response in RC, our study recommends using (y)pTNM staging to assess potential toxicity including lymphovascular invasion, analyze the therapeutic response, look for signs of gastrointestinal or identify a perineural spread. Blinding on these foundations, our study introduces a pioneering approach involving two complementary intelligent models. A pre-trained mBERT model to identify patients at risk of developing toxicities following nCT or nCRT for RC using PET-CT reports. This is followed by MLP algorithms to specify the toxicities (early or late) to which they are vulnerable by identifying up to six different toxicities based on the 28 features. This study is a complementary solution to help radiotherapists and clinicians establish personalized neoadjuvant therapeutic strategies tailored to the characteristics of each patient (comorbidities, histological type and tumor stage). Our study not only helps reduce clinical errors and calibration errors (even if clinicians and radiotherapists are competent) but also improves medical decision-making. It can serve as a basis for a clinical decision support system based on a pre-trained mBERT model and MLP algorithm.\u003c/p\u003e \u003cp\u003eAlthough promising results and relevant clinical information were obtained, this study has some limitations. To confirm the performance of our models, a prospective multicenter study is necessary. Thus, predicting post-treatment toxicity is inherently complex due to many intertwined features. The use of different data sets by our MLP approach mitigates these difficulties but limits their interpretability by its \"black box\" nature. Thus, our toxicity assessment is limited to the use of cross-sectional data. Furthermore, our toxicity prediction models are not yet authorized by health authorities, and they are not practically applied in the radiotherapy departments.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eOur study presents a comprehensive pipeline that combines patients-risk identification, toxicity prediction and biomarkers identification of each toxicity following nCT or nCRT for RC. It demonstrates the use of a pre-trained mBERT model and an MLP algorithm in the assessment of nCT or nCRT for RC protocols. Our study innovates by its approach and methodology which are based on PET-CT reports, as well as demographic and anthropometric data, comorbidities, dental health status, tumor characteristics (pre-therapeutic), neoadjuvant treatment, surgical parameters, biological biomarkers and post-operative staging. Our models demonstrate their performance in terms of sensitivity, specificity and precision. Thus, their ability to identify new biomarkers for each type of late or early toxicity. This study is a complementary solution to help radiotherapists and clinicians establish personalized neoadjuvant therapeutic strategies for RC tailored to each patient's characteristics. It also serves as a basis for decisions regarding the patient during future oncology consultations.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e: The authors of this study would like to express their sincere gratitude and acknowledge the valuable resources provided by Alexander Ferko for his dataset available on Mendeley Data, and by Hamid Y. Hassen, Mohammed A. Teka, Jemal Beksisa, and Jilcha D. Feyisa for their dataset also available on Mendeley Data. These datasets were very useful for our study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e: Conceptualization, M.C., H.B.R. and C.Bo.; Methodology, original draft preparation, M.C., H.B.R. and C.Bo.; Software development and data processing, M.C., and H.B.R ; Resources, writing, M.C., H.B.R., A.R., A.G. and C.Bo.; Writing and Original draft preparation, M.C., H.B.R., A.R., A.G. and C.Bo.; Supervision, H.B.R, A.R., A.G., C.Ba. Y.M., and C.Bo.; Original draft preparation, Review and editing, M.C., H.B.R., A.R., A.G., C.Ba., Y.M. and C.Bo. All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e: The data sets created in this study are private, contained within the extracted datasets (Excel tables and ZIP files), and are not publicly available. They can be shared upon reasonable request from the authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of interest\u003c/strong\u003e: The authors of this study declare no conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e: This research received no external funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate:\u003c/strong\u003e This study involved secondary analysis of two previously published and publicly available anonymized datasets obtained from the Mendeley Data repository. The first dataset (short-term follow-up toxicities), derived from a single-center study, received ethical approval from the Jessenius Faculty of Medicine Comenius University Bratislava in Martin, Institutional Review Board (IRB No. IRB00005636; Approval Number EK1/2019). All patients included in that study provided written informed consent prior to participation. The second dataset (long-term toxicity), derived from a multicenter cohort in Ethiopia, received ethical clearance from the Institutional Review Board of Saint Paul\u0026apos;s Hospital Millennium Medical College and Addis Ababa University, College of Health Sciences (Ref. No: PM23/92). Written informed consent was obtained from patients or their caretakers before data collection. Patient data were anonymized, and confidentiality was ensured throughout the data collection and processing stages. No new patients were recruited, and no direct interaction with human participants occurred in this study. All analyzed data were fully anonymized prior to access and contained no identifiable information. Therefore, no additional institutional ethical approval was required for this secondary analysis, in accordance with international research standards. All research procedures were conducted in accordance with the ethical principles of the Declaration of Helsinki and relevant guidelines and regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIdentifying Information:\u003c/strong\u003e No identifiable personal information, clinical images, or protected health information (HIPAA identifiers) are included in this manuscript or supplementary materials.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMatsuda, T., Fujimoto, A., \u0026amp; Igarashi, Y. (2025). Colorectal Cancer: Epidemiology, Risk Factors, and Public Health Strategies. Digestion.106 (2): 91\u0026ndash;99. https://doi.org/10.1159/000543921\u003c/li\u003e\n\u003cli\u003eLi, L. B., Wang, L. Y., Chen, D. M., Liu, Y. X., Zhang, Y. H., Song, W. X. \u0026amp; Ma, Z. Y. (2023). A systematic analysis of the global and regional burden of colon and rectum cancer and the difference between early-and late-onset CRC from 1990 to 2019. Frontiers in Oncology, 13, 1102673.https://doi.org/10.3389/fonc.2023.1102673\u003c/li\u003e\n\u003cli\u003eGandini, A., Sciallero, S., Martelli, V., Pirrone, C., Puglisi, S., Cremante, M., ... \u0026amp; Pastorino, A. (2025). A Comprehensive Approach to Neoadjuvant Treatment of Locally Advanced Rectal Cancer. Cancers, 17(2), 330. https://doi.org/10.3390/cancers17020330\u003c/li\u003e\n\u003cli\u003eGuo, Y., Guo, Z., Zhang, J., Qian, G., Ji, W., Song, L., ... \u0026amp; Han, Z. (2025). Short-and long-term outcomes of neoadjuvant chemotherapy compared with neoadjuvant chemoradiotherapy for locally advanced rectal cancer: an updated meta-analysis. BMC gastroenterology, 25(1), 87. \u003c/li\u003e\n\u003cli\u003eQian, C., Yang, S., Chen, Y., Ge, R., Shi, F., Liu, C., ... \u0026amp; Guo, Y. (2025). Predicting pathological response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer with two step feature selection and ensemble learning. Scientific Reports, 15(1), 9936. \u003c/li\u003e\n\u003cli\u003eMahdavi, H., Dashti, S., \u0026amp; Jafari, S. (2025). Strategies to reduce intestinal toxicity in neoadjuvant management of locally advanced rectal cancer. \u003cem\u003eDiscover Oncology\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(1), 1886. https://doi.org/10.1007/s12672-025-03526-8\u003c/li\u003e\n\u003cli\u003eJung, W. B. (2024). Beyond survival: a comprehensive review of quality of life in rectal cancer patients. Annals of Coloproctology, 40(6), 527. https://doi.org/10.3393/ac.2024.00745.0106\u003c/li\u003e\n\u003cli\u003eJoye, I., \u0026amp; Haustermans, K. (2014). Early and late toxicity of radiotherapy for rectal cancer. \u003cem\u003eEarly gastrointestinal cancers II: rectal cancer\u003c/em\u003e, 189-201.\u003c/li\u003e\n\u003cli\u003eYang, J., Luo, Y., Tian, T., Dong, P., \u0026amp; Fu, Z. (2022). Effects of Neoadjuvant Radiotherapy on Postoperative Complications in Rectal Cancer: A Meta‐Analysis. \u003cem\u003eJournal of Oncology\u003c/em\u003e, \u003cem\u003e2022\u003c/em\u003e(1), 8197701. https://doi.org/10.1155/2022/8197701 \u003c/li\u003e\n\u003cli\u003e10. Tawfik, M. A. M., Gheida, U. E. S., Khedr, R. A., \u0026amp; Khadrah, R. S. (2025). Multiparametric MRI in the assessment of the response of locally advanced rectal cancer to neoadjuvant chemoradiotherapy. Tanta Medical Journal, 53(2), 161-170.\u003c/li\u003e\n\u003cli\u003eJain, S. M., Nagainallur Ravichandran, S., Murali Kumar, M., Banerjee, A., Sun-Zhang, A., Zhang, H., ... \u0026amp; Pathak, S. (2024). Understanding the molecular mechanism responsible for developing therapeutic radiation-induced radioresistance of rectal cancer and improving the clinical outcomes of radiotherapy-A review. Cancer Biology \u0026amp; Therapy, 25(1), 2317999.\u003c/li\u003e\n\u003cli\u003eLi, M., Xiao, J., Song, S., Han, F., Liu, H., Lin, Y., ... \u0026amp; Liang, L. (2024). PREX2 contributes to radiation resistance by inhibiting radiotherapy-induced tumor immunogenicity via cGAS/STING/IFNs pathway in colorectal cancer. BMC medicine, 22(1), 154.\u003c/li\u003e\n\u003cli\u003eLiang, J., Liao, L., Xie, L., Tang, W., Yu, X., Lu, Y., ... \u0026amp; Tan, Y. (2024). PITPNC1 Suppress CD8+ T cell immune function and promote radioresistance in rectal cancer by modulating FASN/CD155. Journal of Translational Medicine, 22(1), 117.\u003c/li\u003e\n\u003cli\u003eMa, Y., Ma, D., Xu, X., Li, J., \u0026amp; Guan, Z. (2024). Progress of MRI in predicting the circumferential resection margin of rectal cancer: a narrative review. Asian Journal of Surgery, 47(5), 2122-2131.\u003c/li\u003e\n\u003cli\u003eHussen, B. M., Abdullah, S. R., Mohammed, A. A., Rasul, M. F., Hussein, A. M., Eslami, S., ... \u0026amp; Taheri, M. (2024). Advanced strategies of targeting circular RNAs as therapeutic approaches in colorectal cancer drug resistance. Pathology-Research and Practice, 260, 155402. \u003c/li\u003e\n\u003cli\u003eKudra, A., Muszyński, D., Sobocki, B. K., Atzeni, A., Carbone, L., Kaźmierczak-Siedlecka, K., ... \u0026amp; Kalinowski, L. (2023). Insights into oral microbiome and colorectal cancer\u0026ndash;on the way of searching new perspectives. Frontiers in Cellular and Infection Microbiology, 13, 1159822.\u003c/li\u003e\n\u003cli\u003eZeng, D. X., Liu, R. N., Ren, X. K., Zhang, P., Tang, L. H., Tan, L., ... \u0026amp; Xiao, J. W. (2025). Comparison of the efficacy of neoadjuvant chemotherapy and neoadjuvant chemoradiotherapy in locally advanced rectal cancer patients: meta-analysis of randomized controlled trials. \u003cem\u003eInternational Journal of Surgery\u003c/em\u003e, \u003cem\u003e111\u003c/em\u003e(3), 2686-2696.\u003c/li\u003e\n\u003cli\u003eTao, W., Cheng, Y., Wang, P., Wen, H., \u0026amp; Xiao, W. (2025). Comparison of LNR-and LODDS-based predictive models for prognosis in non-elderly patients with locally advanced rectal cancer undergoing neoadjuvant therapy. \u003cem\u003eInternational Journal of Colorectal Disease\u003c/em\u003e, \u003cem\u003e40\u003c/em\u003e(1), 157. https://doi.org/10.1007/s00384-025-04942-6 \u003c/li\u003e\n\u003cli\u003eYu, X. N., Xu, L. M., Bin, Y. W., Yuan, Y., Tian, S. B., Cai, B., ... \u0026amp; Wang, Z. (2022). Risk factors of anastomotic leakage after anterior resection for rectal cancer patients. \u003cem\u003eCurrent Medical Science\u003c/em\u003e, \u003cem\u003e42\u003c/em\u003e(6), 1256-1266. https://doi.org/10.1007/s11596-022-2616-2 \u003c/li\u003e\n\u003cli\u003eJu, H. E., Lee, C. S., Bae, J. H., Lee, H. J., ran Yoon, M., Al-Sawat, A., ... \u0026amp; Han, S. R. (2022). High incidence of late anastomosis leakage in patients for rectal cancer after neoadjuvant chemoradiotherapy: A comparative study. \u003cem\u003eAsian Journal of Surgery\u003c/em\u003e, \u003cem\u003e45\u003c/em\u003e(10), 1832-1842. https://doi.org/10.1016/j.asjsur.2021.10.039\u003c/li\u003e\n\u003cli\u003eCorbeau, A., Heemsbergen, W. D., Kuipers, S. C., Godart, J., Creutzberg, C. L., Nout, R. A., \u0026amp; de Boer, S. M. (2024). Predictive factors for toxicity after primary chemoradiation for locally advanced cervical cancer: a systematic review. \u003cem\u003eInternational Journal of Radiation Oncology* Biology* Physics\u003c/em\u003e, \u003cem\u003e119\u003c/em\u003e(1), 127-142. https://doi.org/10.1016/j.ijrobp.2023.11.010 \u003c/li\u003e\n\u003cli\u003eTohidinezhad, F., Willems, Y., Berbee, M., Van Limbergen, E., Verhaegen, F., Dekker, A., \u0026amp; Traverso, A. (2022). Prediction models for brachytherapy-induced rectal toxicity in patients with locally advanced pelvic cancers: a systematic review. \u003cem\u003eJournal of Contemporary Brachytherapy\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(4), 411-422. https://doi.org/10.5114/jcb.2022.119427\u003c/li\u003e\n\u003cli\u003eAsoglu, O., Tokmak, H., Bakir, B., Demir, G., Ozyar, E., Atalar, B., ... \u0026amp; Guven, K. (2020). The impact of total neo-adjuvant treatment on nonoperative management in patients with locally advanced rectal cancer: The evaluation of 66 cases. European Journal of Surgical Oncology, 46(3), 402-409.\u003c/li\u003e\n\u003cli\u003eLiu, C., Liu, Z., Holmes, J., Zhang, L., Zhang, L., Ding, Y., ... \u0026amp; Liu, W. (2023). Artificial general intelligence for radiation oncology. Meta-radiology, 1(3), 100045.\u003c/li\u003e\n\u003cli\u003eYan, A., McAuley, J., Lu, X., Du, J., Chang, E. Y., Gentili, A., \u0026amp; Hsu, C. N. (2022). RadBERT: adapting transformer-based language models to radiology. Radiology: Artificial Intelligence, 4(4), e210258.\u003c/li\u003e\n\u003cli\u003eMa, W., Chen, B., Zhu, F., Yang, C., \u0026amp; Yang, J. (2024). Diagnostic role of F-18 FDG PET/CT in determining preoperative Lymph node status of patients with rectal cancer: a meta-analysis. Abdominal Radiology, 49(6), 2125-2134.\u003c/li\u003e\n\u003cli\u003eKawamura, M., Kamomae, T., Yanagawa, M., Kamagata, K., Fujita, S., Ueda, D., ... \u0026amp; Naganawa, S. (2024). Revolutionizing radiation therapy: the role of AI in clinical practice. Journal of radiation research, 65(1), 1-9.\u003c/li\u003e\n\u003cli\u003eDadgar, H., Norouzbeigi, N., Jafari, E., Al-balooshi, B., Al-Ibraheem, A., Haidar, M., ... \u0026amp; Assadi, M. (2024). Exploring the efficacy of FAPI PET/CT in the diagnosis and treatment management of colorectal cancer: a comprehensive literature review and initial experience. Clinical and Translational Imaging, 12(3), 235-252.\u003c/li\u003e\n\u003cli\u003eBarabadi, M. A., Chan, W. Y., Zhu, X., Simpson, A. L., \u0026amp; Do, R. K. (2024, August). Adapting large language models for automatic annotation of radiology reports for metastases detection. In 2024 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE) (pp. 340-345). IEEE.\u003c/li\u003e\n\u003cli\u003eJoel, L. O., Doorsamy, W., \u0026amp; Paul, B. S. (2024). On the performance of imputation techniques for missing values on healthcare datasets. arXiv preprint arXiv:2403.14687.\u003c/li\u003e\n\u003cli\u003eKarabacak, M., \u0026amp; Margetis, K. (2023). Embracing large language models for medical applications: opportunities and challenges. \u003cem\u003eCureus\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(5).\u003c/li\u003e\n\u003cli\u003eLee, S., Lim, J., Shin, J., Kim, S., \u0026amp; Hwang, H. (2021). Pathologic complete response prediction after neoadjuvant chemoradiation therapy for rectal cancer using radiomics and deep embedding network of MRI. Applied Sciences, 11(20), 9494.\u003c/li\u003e\n\u003cli\u003eLi, J., Zhang, W., Chen, L., Mao, X., Wang, X., Liu, J., ... \u0026amp; Wang, T. (2024). SNPs and blood inflammatory marker featured machine learning for predicting the efficacy of fluorouracil-based chemotherapy in colorectal cancer. Scientific Reports, 14(1), 27700.\u003c/li\u003e\n\u003cli\u003eLou, X., Zhou, N., Feng, L., Li, Z., Fang, Y., Fan, X., ... \u0026amp; Huang, Y. (2022). Deep learning model for predicting the pathological complete response to neoadjuvant chemoradiotherapy of locally advanced rectal cancer. Frontiers in Oncology, 12, 807264.\u003c/li\u003e\n\u003cli\u003eYAKAR, M., ETİZ, D., BADAK, B., \u0026Ccedil;ELİK, \u0026Ouml;., K\u0026Uuml;TRİ, D., \u0026Ouml;ZEN, A., \u0026amp; YILMAZ, E. (2021). Prediction of Response to Neoadjuvant Chemoradiotherapy with Machine Learning in Rectal Cancer: A Pilot Study. TURK ONKOLOJI DERGISI, 36(4).\u003c/li\u003e\n\u003cli\u003eSzmigiel, A., Apel, D. B., Wang, J., Pu, Y., Wojtecki, L., \u0026amp; Pourrahimian, Y. (2024). Enhancing open stope stability prediction in mining engineering: Optimal configuration of an artificial neural network model. Journal of Industrial Safety, 1(1), 100008.\u003c/li\u003e\n\u003cli\u003eSolarte-Pab\u0026oacute;n, O., Montenegro, O., Garc\u0026iacute;a-Barrag\u0026aacute;n, A., Torrente, M., Provencio, M., Menasalvas, E., \u0026amp; Robles, V. (2023). Transformers for extracting breast cancer information from Spanish clinical narratives. Artificial Intelligence in Medicine, 143, 102625.\u003c/li\u003e\n\u003cli\u003eGardazi, N. M., Daud, A., Malik, M. K., Bukhari, A., Alsahfi, T., \u0026amp; Alshemaimri, B. (2025). BERT applications in natural language processing: a review. Artificial Intelligence Review, 58(6), 1-49.\u003c/li\u003e\n\u003cli\u003eSantos, C. F. G. D., \u0026amp; Papa, J. P. (2022). Avoiding overfitting: A survey on regularization methods for convolutional neural networks. ACM Computing Surveys (Csur), 54(10s), 1-25.\u003c/li\u003e\n\u003cli\u003eKozak, J., Probierz, B., Kania, K., \u0026amp; Juszczuk, P. (2022). Preference-driven classification measure. Entropy, 24(4), 531.\u003c/li\u003e\n\u003cli\u003eJierula, A., Wang, S., Oh, T. M., \u0026amp; Wang, P. (2021). Study on accuracy metrics for evaluating the predictions of damage locations in deep piles using artificial neural networks with acoustic emission data. Applied Sciences, 11(5), 2314.\u003c/li\u003e\n\u003cli\u003eChen, S., Guevara, M., Ramirez, N., Murray, A., Warner, J. L., Aerts, H. J., ... \u0026amp; Bitterman, D. S. (2023). Natural language processing to automatically extract the presence and severity of esophagitis in notes of patients undergoing radiotherapy. JCO Clinical Cancer Informatics, 7, e2300048.\u003c/li\u003e\n\u003cli\u003eTay, S. B., Low, G. H., Wong, G. J. E., Tey, H. J., Leong, F. L., Li, C., ... \u0026amp; Tan, R. S. Y. C. (2024). Use of natural language processing to infer sites of metastatic disease from radiology reports at scale. JCO Clinical Cancer Informatics, 8, e2300122.\u003c/li\u003e\n\u003cli\u003eLiu, K., Kulkarni, O., Witteveen-Lane, M., Chen, B., \u0026amp; Chesla, D. (2022). MetBERT: a generalizable and pre-trained deep learning model for the prediction of metastatic cancer from clinical notes. AMIA Summits on Translational Science Proceedings, 2022, 331.\u003c/li\u003e\n\u003cli\u003eRyu, J., \u0026amp; Cho, S. B. (2002, January). Towards optimal feature and classifier for gene expression classification of cancer. In \u003cem\u003eAFSS International Conference on Fuzzy Systems\u003c/em\u003e (pp. 310-317). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-45631-7_41 \u003c/li\u003e\n\u003cli\u003eGuo, Q. H., Xie, F. C., Zhong, F. M., Wen, W., Zhang, X. R., Yu, X. J., ... \u0026amp; Wang, X. Z. (2024). Application of interpretable machine learning algorithms to predict distant metastasis in ovarian clear cell carcinoma. \u003cem\u003eCancer Medicine\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(7), e7161. https://doi.org/10.1002/cam4.7161\u003c/li\u003e\n\u003cli\u003eKhan, W., Leem, S., See, K. B., Wong, J. K., Zhang, S., \u0026amp; Fang, R. (2025). A comprehensive survey of foundation models in medicine. IEEE Reviews in Biomedical Engineering.\u003c/li\u003e\n\u003cli\u003eTejani, A. S., Ng, Y. S., Xi, Y., Fielding, J. R., Browning, T. G., \u0026amp; Rayan, J. C. (2022). Performance of multiple pretrained BERT models to automate and accelerate data annotation for large datasets. Radiology: Artificial Intelligence, 4(4), e220007. \u003c/li\u003e\n\u003cli\u003eBai, H., Zhang, W., Hou, L., Shang, L., Jin, J., Jiang, X., ... \u0026amp; King, I. (2020). Binarybert: Pushing the limit of bert quantization. arXiv preprint arXiv:2012.15701.\u003c/li\u003e\n\u003cli\u003eBeltagy, I., Lo, K., \u0026amp; Cohan, A. (2019). SciBERT: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676.\u003c/li\u003e\n\u003cli\u003eKalyan, K. S., Rajasekharan, A., \u0026amp; Sangeetha, S. (2022). AMMU: a survey of transformer-based biomedical pretrained language models. Journal of biomedical informatics, 126, 103982.\u003c/li\u003e\n\u003cli\u003eMutuvi, S., Boros, E., Doucet, A., Lejeune, G., Jatowt, A., \u0026amp; Odeo, M. (2020, December). Multilingual epidemiological text classification: a comparative study. In COLING, International Conference on Computational Linguistics (pp. 6172-6183).\u003c/li\u003e\n\u003cli\u003eShi, L., Zhang, Y., Nie, K. E., Sun, X., Niu, T., Yue, N., ... \u0026amp; Su, M. Y. (2019). Machine learning for prediction of chemoradiation therapy response in rectal cancer using pre-treatment and mid-radiation multi-parametric MRI. Magnetic resonance imaging, 61, 33-40.\u003c/li\u003e\n\u003cli\u003eVolovat, C. C., Scripcariu, D. V., Boboc, D., Volovat, S. R., Vasilache, I. A., Ursulescu-Lupascu, C., ... \u0026amp; Scripcariu, V. (2024). Machine Learning-Based Algorithms for Enhanced Prediction of Local Recurrence and Metastasis in Low Rectal Adenocarcinoma Using Imaging, Surgical, and Pathological Data. Diagnostics, 14(6), 625.\u003c/li\u003e\n\u003cli\u003eSalehin, I., \u0026amp; Kang, D. K. (2023). A review on dropout regularization approaches for deep neural networks within the scholarly domain. Electronics, 12(14), 3106.\u003c/li\u003e\n\u003cli\u003eSorrentino, L., Sileo, A., Daveri, E., Battaglia, L., Guaglio, M., Centonze, G., ... \u0026amp; Cosimelli, M. (2023). Impact of Microscopically Positive (\u0026le; 1 mm) Distal Margins on Disease Recurrence in Rectal Cancer Treated by Neoadjuvant Chemoradiotherapy. Cancers, 15(6), 1828.\u003c/li\u003e\n\u003cli\u003eCalvo, F. A., Tudela, M., Serrano, J., Mu\u0026ntilde;oz-Fern\u0026aacute;ndez, M., Peligros, M. I., Garcia-Alfonso, P., \u0026amp; Del Valle, E. (2023). Post-chemoradiation metastatic, persistent and resistant nodes in locally advanced rectal cancer: metrics and their impact on long-term outcome. Cancers, 15(18), 4591.\u003c/li\u003e\n\u003cli\u003eVolovat, C. C., Scripcariu, D. V., Boboc, D., Volovat, S. R., Vasilache, I. A., Lupascu-Ursulescu, C., ... \u0026amp; Scripcariu, V. (2024). Predicting the Feasibility of Curative Resection in Low Rectal Cancer: Insights from a Prospective Observational Study on Preoperative Magnetic Resonance Imaging Accuracy. Medicina, 60(2), 330.\u003c/li\u003e\n\u003cli\u003eKim, Y. I., Cho, H., Kim, C. W., Park, Y., Kim, J., Ro, J. S., ... \u0026amp; Kim, J. C. (2021). Prognostic impact of extranodal extension in rectal cancer patients undergoing radical resection after preoperative chemoradiotherapy. Clinical colorectal cancer, 20(1), e35-e42.\u003c/li\u003e\n\u003cli\u003eCui, Y., Song, M., Tie, J., Li, S., Wang, H., Zhang, Y., ... \u0026amp; Wang, W. (2024). Clinicopathological factors predict residual lymph node metastasis in locally advanced rectal cancer with ypT0-2 after neoadjuvant chemoradiotherapy. Journal of Cancer Research and Clinical Oncology, 150(4), 176.\u003c/li\u003e\n\u003cli\u003eChiscuzzu, F., Crescio, C., Varrucciu, S., Rizzo, D., Sali, M., Delogu, G., \u0026amp; Bussu, F. (2025). Current Evidence on the Relation Between Microbiota and Oral Cancer\u0026mdash;The Role of Fusobacterium nucleatum\u0026mdash;A Narrative Review. Cancers, 17(2), 171.\u003c/li\u003e\n\u003cli\u003eQue, Y., Wu, R., Li, H., \u0026amp; Lu, J. (2024). A prediction nomogram for perineural invasion in colorectal cancer patients: a retrospective study. BMC surgery, 24(1), 80.\u003c/li\u003e\n\u003cli\u003eKokaine, L., Gardovskis, A., \u0026amp; Gardovskis, J. (2021). Evaluation and predictive factors of complete response in rectal cancer after neoadjuvant chemoradiation therapy. Medicina, 57(10), 1044.\u003c/li\u003e\n\u003cli\u003eDong, J., Li, Y., Xiao, H., Zhang, S., Wang, B., Wang, H., ... \u0026amp; Cui, M. (2021). Oral microbiota affects the efficacy and prognosis of radiotherapy for colorectal cancer in mouse models. Cell reports, 37(4).\u003c/li\u003e\n\u003cli\u003eNiu, R., Jiang, Y., Bi, Z., Zhang, H., Mei, X., Bi, J., ... \u0026amp; Liang, J. (2024). The effect of cardiovascular disease on the perioperative period of radical surgery in elderly rectal cancer. BMC gastroenterology, 24(1), 256.\u003c/li\u003e\n\u003cli\u003eVarela, C., \u0026amp; Kim, N. K. (2021). Surgical treatment of low-lying rectal cancer: updates. Annals of coloproctology, 37(6), 395.\u003c/li\u003e\n\u003cli\u003eLai, I. L., You, J. F., Chern, Y. J., Tsai, W. S., Chiang, J. M., Hsieh, P. S., ... \u0026amp; Hsu, Y. J. (2021). The risk factors of local recurrence and distant metastasis on pT1/T2N0 mid-low rectal cancer after total mesorectal excision. World Journal of Surgical Oncology, 19, 1-7.\u003c/li\u003e\n\u003cli\u003eKerns, S. L., Stock, R. G., Stone, N. N., Rath, L., Vega, A., Fachal, L., ... \u0026amp; Ostrer, H. (2013). Genome-wide association study identifies a region on chromosome 11q14. 3 associated with late rectal bleeding following radiation therapy for prostate cancer. Radiotherapy and Oncology, 107(3), 372-376. https://doi.org/10.1016/j.radonc.2013.05.001 \u003c/li\u003e\n\u003cli\u003eBenitez, C. M., \u0026amp; Knox, S. J. (2020). Harnessing genome-wide association studies to minimize adverse radiation-induced side effects. Radiation Oncology Journal, 38(4), 226. https://doi.org/10.3857/roj.2020.00556 \u003c/li\u003e\n\u003cli\u003eFern\u0026aacute;ndez-Mateos, J., Seijas-Tamayo, R., Mes\u0026iacute;a, R., Taberna, M., Borgo\u0026ntilde;\u0026oacute;n, M. P., P\u0026eacute;rez-Ruiz, E., ... \u0026amp; Cruz-Hern\u0026aacute;ndez, J. J. (2016). Epidermal growth factor receptor (EGFR) pathway polymorphisms as predictive markers of cetuximab toxicity in locally advanced head and neck squamous cell carcinoma (HNSCC) in a Spanish population. Oral oncology, 63, 38-43. https://doi.org/10.1016/j.oraloncology.2016.10.006 \u003c/li\u003e\n\u003cli\u003eWatts, K. (2023). Genome-wide analyses to identify biomarkers of toxicity in the treatment of advanced colorectal cancer (Doctoral dissertation, Cardiff University). https://orca.cardiff.ac.uk/id/eprint/163425 \u003c/li\u003e\n\u003cli\u003eWatts, K., Wills, C., Madi, A., Palles, C., Maughan, T. S., Kaplan, R., ... \u0026amp; Cheadle, J. P. (2021). Genome‐wide association studies of toxicity to oxaliplatin and fluoropyrimidine chemotherapy with or without cetuximab in 1800 patients with advanced colorectal cancer. International journal of cancer, 149(9), 1713-1722. https://doi.org/10.1002/ijc.33739 \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Deep Learning, Large Language Model, multilingual BERT, MultiLayer Perceptron, Neoadjuvant Chemoradiotherapy and Chemotherapy, Toxicity","lastPublishedDoi":"10.21203/rs.3.rs-8967708/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8967708/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eNeoadjuvant chemotherapy (nCT) and chemoradiotherapy (nCRT) are widely used cancer treatments, however, patients\u0026rsquo; response and subsequent clinical toxicity vary substantially. Predicting toxicity using artificial intelligence (AI) allows risk-adapted treatment and overcomes statistical models\u0026rsquo; limitations. This study describes a novel approach using two complementary AI models to achieve three objectives. Initially, a fine-tuned pre-trained multilingual BERT model (mBERT), was used to analyse radiological reports and identify patients at risk of developing toxicities post-nCT or nCRT for rectal cancer (RC). Then, a MultiLayer Perceptron (MLP) neural network was developed to predict toxicities and identify its key biomarkers. Our results demonstrate that the mBERT model and MLP algorithm achieved strong performance in classifying patients at risk and predicting toxicities following nCT or nCRT for RC (mBERT: precision\u0026thinsp;=\u0026thinsp;1, recall\u0026thinsp;=\u0026thinsp;0.94 and F1-score\u0026thinsp;=\u0026thinsp;0.97; MLP: accuracy\u0026thinsp;=\u0026thinsp;0.90, mean squared error\u0026thinsp;=\u0026thinsp;0.06 and mean absolute error\u0026thinsp;=\u0026thinsp;0.16). The MLP algorithm identified toxicity biomarkers not previously reported in machine learning models. Furthermore, our study recommends using (y)pTNM staging as a biomarker for potential toxicity. In conclusion, this study presents a new AI approach to classify patients at risk and predict toxicities following nCT or nCRT for RC supporting personalized neoadjuvant therapeutic strategies.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence-Driven Prediction of post Neoadjuvant Treatment Toxicities and Biomarkers Identifications in Rectal Cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-09 06:32:32","doi":"10.21203/rs.3.rs-8967708/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-31T14:26:56+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-31T03:54:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"210675880233107436910933347097124765558","date":"2026-03-31T02:18:55+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-14T04:29:44+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"273632372518366372382668896634644482082","date":"2026-03-04T13:05:06+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-03T11:30:31+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-03T11:24:03+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-03T11:05:50+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-01T19:38:38+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-03-01T19:34:38+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f26c0581-2cf8-4a98-9952-1c8a14ce03cd","owner":[],"postedDate":"March 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":63919356,"name":"Health sciences/Biomarkers"},{"id":63919357,"name":"Biological sciences/Cancer"},{"id":63919358,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":63919359,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-04-29T01:54:40+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-09 06:32:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8967708","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8967708","identity":"rs-8967708","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}