A transformer-MLP model with SHAP analysis for predicting postoperative survival in Intrahepatic Cholangiocarcinoma

A transformer-MLP model with SHAP analysis for predicting postoperative survival in Intrahepatic Cholangiocarcinoma | 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 A transformer-MLP model with SHAP analysis for predicting postoperative survival in Intrahepatic Cholangiocarcinoma Yuhan Zhou, Yichen Wang, Haonan Liu, Yiting Dou, Youwei Wu, Xutian Wang, and 20 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9513815/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Intrahepatic cholangiocarcinoma (ICC) carries a dismal postoperative prognosis, with high recurrence rates and inconsistent adjuvant therapy benefits. Current prognostic models are limited by high cost, single‑modal data, or poor accessibility. We developed and externally validated a Transformer‑MLP deep learning model for 2‑year overall survival prediction in 499 resected ICC patients from 8 Chinese centers. A differentiated feature selection strategy was applied: 17 features for traditional machine learning and all 31 routine clinicopathological features for deep learning. The Transformer‑MLP model achieved the best performance in the external test set (AUC = 0.8131, C‑index = 0.7462), with favorable calibration and net clinical benefit. SHAP analysis highlighted adjuvant chemotherapy, AJCC stage, and lymph node metastasis as key predictors. A free online calculator ( https://icc‑prediction.streamlit.app/ ) was built for clinical use. This accessible, interpretable model provides robust risk stratification to guide personalized adjuvant therapy, complementing existing omics‑ or radiomics‑based ICC prognostic tools. Health sciences/Biomarkers Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Intrahepatic cholangiocarcinoma (ICC) multicenter retrospective study adjuvant therapy Deep Learning Transformer-MLP Model SHapley Additive exPlanations (SHAP) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction Intrahepatic cholangiocarcinoma (ICC), the second most common primary liver malignancy, poses a substantial global health burden, with rising incidence and persistently poor prognosis. [1] Surgical resection remains the cornerstone of potentially curative treatment; however, this approach is severely limited by a high recurrence rate—over 50% of patients experience disease recurrence within two years postoperatively—resulting in a dismal 5-year overall survival rate of only 30-40%. [2] There is an urgent need for effective adjuvant therapies targeting residual micrometastatic disease to improve long-term outcomes. Currently, the field of adjuvant therapy for resected ICC lacks consensus, with considerable controversy and a paucity of clear, evidence-based clinical guidelines. [3] A robust prognostic model is a critical decision-support tool in ICC management by enabling precise risk stratification. [4] Such models facilitate the classification of patients into distinct risk groups, a cornerstone of personalized therapy: they help identify high-risk patients most likely to benefit from adjuvant treatment, while sparing low-risk patients unnecessary toxicity. Furthermore, prognostic models optimize the design of more efficient clinical trials via precise patient enrichment, ensuring novel adjuvant strategies are tested in populations with the highest potential for response. [5] By integrating multi-dimensional data—including clinical, pathological, and surgical variables—comprehensive models transcend single-modality assessments to deliver holistic, accurate risk evaluations, which are essential for refining individualized postoperative management in ICC. [6] Prognostic modeling for ICC is a rapidly evolving field. Traditional nomograms mainly rely on clinicopathological features, whereas bioinformatics models leverage genomic data. The advancement of artificial intelligence (AI)—especially machine learning (ML) for predicting short-term surgical outcomes and deep learning (DL) for analyzing complex histopathological and radiological data—marks a major breakthrough. [7] State-of-the-art multi-omics integration approaches further combine genomic, transcriptomic, and radiomic data to enhance predictive performance. [8] However, significant limitations persist in current models: many focus on a single data modality, lack rigorous external validation, and exhibit insufficient interpretability for direct clinical translation. [9] This study was designed to address these gaps by developing and validating a novel Transformer-MLP deep learning model for predicting postoperative survival in ICC patients. A key methodological innovation was the adoption of a differentiated feature-selection strategy tailored to the unique characteristics of each model type. For traditional machine learning algorithms, a refined set of 17 key features was used to maximize interpretability and computational efficiency. In contrast, the deep learning model incorporated the full set of 31 available features, leveraging its inherent capacity to automatically weight features and identify complex non-linear interactions in high-dimensional data. The Transformer-MLP model was trained on a multi-center development cohort and validated on an independent external test set. It also demonstrated good calibration and yielded a positive net clinical benefit in decision curve analysis, underscoring its potential clinical utility. 2. Materials and Methods 2.1. Data Sources and Study Ethics This multicenter, retrospective cohort study was conducted using data from eight tertiary hospitals in China. We identified patients who underwent curative-intent radical liver resection between January 2010 and January 2022 across five cities (Xi'an, Chengdu, Shenyang, Yulin, and Baoji) and had a postoperative pathological confirmation of ICC. After applying the inclusion and exclusion criteria and excluding cases lost to follow-up, a final cohort of 499 patients was included for analysis. The study population comprised 288 (57.7%) male and 211 (42.3%) female patients, with an age range of 26 to 84 years. A detailed flowchart of patient selection is presented in figure S1. This study was approved by the Ethics Committee of The First Affiliated Hospital of Xi’an Jiaotong University (Approval No.: XJTU1AF2025LSYY-393) on 13 March 2025. All procedures conducted in this study were performed in strict accordance with the ethical standards of the institutional research committee, as well as the 1964 Helsinki Declaration and its subsequent amendments. Given the retrospective nature of this study, written informed consent from individual patients was formally waived by the aforementioned ethics committee. 2.2. Data Preprocessing Surgical, Adjuvant Chemotherapy and follow-up Protocols see in method S1. We initially evaluated 31 clinical and pathological features across several domains: patient demographics (sex, age, BMI), comorbidities (hepatitis, cirrhosis, hepatobiliary and cardiopulmonary diseases), surgical details (surgical approach, lymph node dissection, extent of resection, intraoperative blood loss), laboratory values (PLR, NLR, AST, ALT, GGT, TBIL, ALB, PT), and tumor characteristics (AFP, CEA, CA19-9, differentiation grade, microvascular invasion, perineural invasion, capsular invasion, tumor number, tumor size, lymph node metastasis, AJCC stage). Collection of Clinicopathological Data see in method S2. To ensure data quality and improve model stability and predictive performance, we performed systematic pre-processing on the raw clinical data. This included the transformation of numerical variables, encoding of categorical variables, and preparation of the data for survival analysis. Specific data preprocessing process see in method S3. 2.3. Dataset Partitioning and Validation Strategy To comprehensively evaluate the model's predictive stability (internal consistency) and clinical generalizability (external validity), we employed a two-stage validation framework consisting of internal five-fold cross-validation and performance testing on a completely independent external set. The internal development cohort comprised 328 patients from the First Affiliated Hospital of Xi'an Jiaotong University. We performed stratified 5-fold cross-validation on this cohort to ensure reliable model development and initial validation. Specific Steps of Five-Fold Cross-Validation see in method S4. All models were developed through this process. After determining the optimal model architecture and hyperparameters via cross-validation, the final model's generalization capability was rigorously assessed on a completely independent external test set. This set consisted of 171 patients from the remaining 7 participating centers, which were not involved in any stage of internal development. 2.4 Feature selection This study employed a systematic two-stage feature selection approach to identify the most predictive subset from an initial set of 31 clinicopathological features. 2.4.1 Feature importance ranking based on multi-method fusion To robustly identify the most predictive features from an initial set of 31 clinicopathological variables, we implemented a multi-method fusion strategy for feature importance ranking. Three complementary feature selection techniques were employed: univariate F-value analysis, random forest-based importance, and LASSO regression (figure S2-4). To synthesize the results from these diverse methodologies, the importance scores from each method were first normalized to a [0, 1] scale. A weighted fusion scheme was then applied (F-value: 33%, random forest: 33%, LASSO: 34%) to calculate a composite importance score for each feature. This multi-method fusion strategy enhances the robustness of the ranking by leveraging the strengths of each approach .Fusion results demonstrated the stability of these features' prognostic value regardless of the underlying selection logic, and that the selected features provide complementary rather than redundant information for predicting ICC survival outcomes (figure 1). 2.4.2 Feature Number Optimization via Cross-Validation To determine the optimal number of features (k*) across different model architectures, a systematic screening of feature subset sizes was performed. The procedure see in method S5. Systematic assessment of five traditional machine learning models across 1–31 feature subsets revealed distinct performance patterns and optimal feature set sizes by algorithm (table S2 and figure S5ab). Overall, model performance stabilized substantially after 8 features; expanding to 17 features yielded minimal average AUC improvement (<0.03), indicating a clear saturation point with diminishing marginal gains. Comprehensive feature inclusion experiments on six deep learning models revealed a complex non-monotonic relationship between model performance and input feature count. Unlike traditional machine learning models, which often plateau beyond a certain feature threshold, deep learning models exhibited more nuanced fluctuating performance (table S2 and figure S5cd). In stark contrast to traditional ML models (typically saturated beyond 8 features), this highlights their inherent advantage in harnessing high-dimensional data—attributed to integrated feature selection and interaction learning mechanisms (enhanced by feature/paradigm fusion). This adaptability reduces information loss from preemptive feature filtering. 2.5 Model Development & Comparison To evaluate machine learning approaches for predicting 2-year postoperative OS, this study developed and compared 11 prediction models. The model set included five traditional algorithms—logistic regression, support vector machine (SVM), LightGBM, random forest, and XGBoost—chosen to cover a gradient of interpretability and algorithmic complexity, alongside six deep learning architectures: multilayer perceptron (MLP), Transformer, recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), and a novel Transformer-MLP model. Covering diverse representational learning strategies (from feedforward hierarchies to dynamic gating and attention-based feature integration), these models enabled comprehensive assessment of deep learning’s utility in prognostic prediction. Their theoretical rationale see in method S6. 2.6. Design of Transformer-MLP Model This study proposes a composite network architecture that first employs a multi-head attention mechanism to comprehensively capture features, followed by a MLP that performs nonlinear mapping using the extracted high-dimensional features to achieve precise prediction of 2-year survival prognosis. The specific workflow is illustrated in figure S6. After mapping clinical pathological information into a high-dimensional space, the model feeds it into a Transformer layer for more comprehensive and refined feature extraction. Subsequently, a MLP head with 32 neurons outputs a predicted probability value (0-1 range) representing the patient's risk of death within two years. This value is then converted into a survival status classification prediction (0/1) using a 0.5 threshold. In this Transformer-MLP model, the correlation among different clinical indicators is effectively captured by the Transformer model through its attention mechanism. Subsequently, the MLP acts as a dedicated classifier, performing nonlinear inference on the features to output the final risk score. 2.7 SHAP Model Interpretation Analysis To enhance the clinical interpretability of the prognostic model, we performed SHapley Additive exPlanations (SHAP) analysis to quantify and rank the contribution of each clinical and histological variable to model outputs. Rooted in cooperative game theory, SHAP values offer a unified measure of feature importance by fairly attributing the difference between an instance’s prediction and the dataset’s average prediction to each feature, enabling both local (per-instance) and global (dataset-wide) interpretability for identifying key survival-predictive factors. Analysis was conducted on the final trained model with the optimized feature set, using PermutationExplainer to estimate SHAP values—this model-agnostic method approximates Shapley values by permuting features and measuring prediction changes, ensuring reliability for the model’s structure. For global interpretation, we generated SHAP summary plots (bar and scatter plots): bar plots show mean absolute SHAP values (overall feature importance), while scatter plots illustrate SHAP value distributions across samples, revealing impact direction (positive/negative on predictions) and potential non-linear feature-output relationships. For individual patient-level local interpretation, decision plots visualized how feature contributions accumulate to shift predictions from the base value to the final output for specific cases. SHAP analysis confirmed that top features (by mean absolute SHAP value) aligned with established clinical prognostic factors for ICC, validating the model’s clinical relevance and biological plausibility. This framework clarified the model’s decision-making process, provided a rigorous data-driven basis for feature selection, and enhanced model robustness. By identifying key prognostic variables and their directional survival impacts, SHAP analysis delivers transparent, actionable insights to support individualized treatment strategy development. 2.8 Statistical Analysis Statistical analyses were performed using Python 3.9.6 with key libraries including scikit-learn, scikit-survival, and the SHAP package. The analytic cohort included a total of 31 features, comprising 17,964 data points. To compare baseline demographic and clinical characteristics between the model development set and the independent test set, the Mann‑Whitney U test was employed for continuous variables, and the Chi‑square test was applied for categorical variables. Model performance was comprehensively evaluated on the test set. The precision of the performance metrics was estimated using 95% confidence intervals obtained from 1,000 bootstrap resamples. Calibration curves were plotted to visualize the agreement between the predicted probabilities and the actual observed 2‑year overall survival (OS). Survival outcomes across different risk strata, as defined by the model, were compared using Kaplan‑Meier curves with the log‑rank test. Decision curve analysis (DCA) further quantified the net clinical benefit of model-based risk stratification across various decision thresholds. To enhance the interpretability of the final Transformer‑MLP model, SHAP were applied to quantify the contribution of each feature to individual predictions. A two - tailed p - value of less than 0.05 was considered statistically significant for all analyses. 3. Results 3.1 Baseline Characteristics A total of 499 patients meeting the inclusion and exclusion criteria were included in the analysis. Their baseline characteristics are presented in table S1. The vast majority of variables showed no significant differences between the training and test sets (p > 0.05), confirming their comparability. 3.2 Final Feature Set Determination For traditional ML models, the top 17 composite-score features were selected, as performance plateaued beyond 17. This subset balances optimal performance with better training efficiency, inference speed, and interpretability—critical for clinical deployment in resource-constrained or high-transparency settings. In contrast, all 31 features were used for deep learning models, aligning with their data-driven paradigm. Architectures like Transformer-MLP, with self-attention and multi-layer non-linear transformations, inherently perform automatic feature selection/weighting during end-to-end training. The full feature set enables autonomous identification of predictive signals and complex interactions, supported by experiments showing superior overall (highest average AUC) and single-model performance (Transformer-MLP AUC = 0.8131). This differential strategy highlights that optimal feature sets are algorithm-dependent, enabling flexible deployment: traditional models with 17 core features suit scenarios prioritizing speed, low computational cost, and interpretability, while deep learning models with all 31 features maximize predictive accuracy with adequate computational resources. 3.3 Model Performance Evaluation The performance of 11 prediction models on training, validation, and test sets (summarized in tables S3-4) was evaluated using C-index, AUC, sensitivity, specificity, PPV, NPV, and Brier score. Several models (XGBoost, LightGBM [AUC=1.0000]; Random Forest [AUC=0.9770]; SVM [AUC=0.9642]; Transformer-MLP [AUC=0.9627]) exhibited excellent discrimination on the training set, though their generalization performance differed significantly on validation and independent test sets. Notably, the Transformer-MLP model achieved the best overall performance on the independent test set (AUC=0.8131, C-index=0.7462, 95% CI: 0.6898-0.7950), with high sensitivity (0.8361) and excellent NPV (0.8667) that effectively identified most event (e.g., death)-experienced patients and enabled reliable low-risk stratification. Other deep learning models (Transformer [AUC=0.7578], MLP [AUC=0.7317], LSTM [AUC=0.7073]) also generally outperformed traditional machine learning models on the test set, which conversely showed significant performance degradation indicative of overfitting—for example, XGBoost's AUC dropped from 1.0000 (training) to 0.6121 (test), Random Forest had a test-set AUC of 0.6346 but extremely low sensitivity (0.1148) (failing to identify most high-risk patients), and Logistic Regression (AUC=0.6380) and SVM (AUC=0.6387) exhibited limited test-set performance. The Transformer-MLP model's calibration was assessed via calibration curves (figure 2ab for initial training and test sets) and Brier score; to improve test-set calibration, Platt Scaling and Isotonic Regression (trained on the validation set) were applied. Figure 2c compares the original and calibrated models' test-set calibration (with Brier score as a quantitative metric, where lower values indicate better calibration), and the curve shows good agreement between predicted probabilities and observed risk. Additionally, decision curve analysis (figure 3) confirmed the model's higher net clinical benefit across most clinically relevant threshold probabilities. In summary, the Transformer-MLP deep learning model (trained on all 31 features) demonstrated superior, robust predictive performance, good calibration, and clinical utility on the independent multicenter test set, making it a reliable tool for predicting postoperative survival in ICC. 3.4 ROC Curves and Kaplan-Meier Survival Analysis The 5-fold cross-validation performance of Transformer-MLP model see in table S5. The corresponding ROC curve is shown in figure S7. Receiver operating characteristic curves for all 11 models are presented in figure 4. The Transformer-MLP model's test set AUC is highlighted in figure 4c. The high AUC value confirms its excellent discriminative ability. Furthermore, Kaplan-Meier survival curves stratified by the model's risk prediction (high vs. low) showed significant separation (figure 5). The statistical significance of this difference, confirmed by the log-rank test, underscores the model's utility for clinical risk stratification (p < 0.01). 3.5 Feature Importance Analysis SHAP (SHapley Additive exPlanations) was employed to analyze the feature importance of the Transformer-MLP model. Figure 6a ranks postoperative features by mean absolute SHAP value, reflecting their relative predictive importance. The SHAP summary plot (figure 6b) displays the distribution of individual patient SHAP values per feature, with color indicating feature values (blue: low, red: high), clarifying the magnitude and direction of each feature's impact on model predictions. The waterfall chart (figure 6c) details how individual feature values shift the model's baseline prediction toward higher or lower death risk. 3.6 Online Predictor To facilitate clinical application, we developed a publicly accessible online prognostic tool (https://icc-prediction.streamlit.app/). Incorporating all 11 developed models, it delivers individualized risk assessments, generates patient-specific survival probability curves and overall risk classification, and features a user-friendly interface for real-time ICC prognosis evaluation (figure S8). 4. Discussion Postoperative management of intrahepatic cholangiocarcinoma (ICC) after curative resection remains challenging, with tumor recurrence standing as a major clinical obstacle. Over half of patients experience tumor relapse even after complete surgical resection, indicating that surgical intervention alone is insufficient to improve long-term survival.[10] The high postoperative recurrence rate highlights the necessity of exploring reasonable adjuvant treatment strategies for ICC patients. In current clinical practice, adjuvant chemotherapy is widely applied for resected ICC, while debates still exist regarding its overall survival benefit, appropriate candidate selection, and clinical application value.[11] The present prognostic model may assist in precise risk stratification and provide reference for individualized adjuvant treatment decision-making for resected ICC patients. The recommendation for adjuvant chemotherapy is largely based on the aggressive biological characteristics of ICC. Key pathological variables included in our prediction model, such as lymph node metastasis, elevated preoperative CA19-9 level, advanced AJCC stage and poor tumor differentiation, are recognized markers associated with potential micrometastasis and poor prognosis.[12] Eliminating invisible microscopic lesions is considered the core objective of adjuvant systemic therapy, which is recommended by current clinical guidelines for high-risk biliary tract cancer patients.[1] Relevant clinical trials, such as the BILCAP study, have demonstrated the potential survival benefit of capecitabine in resected biliary tract malignancies, including ICC.[13] Nevertheless, the clinical value of adjuvant chemotherapy for ICC remains controversial. Current evidence focusing on ICC-specific cohorts has not yielded consistent and definitive conclusions regarding its long-term survival improvement.[14] Traditional risk evaluation relying on single clinicopathological indicators lacks sufficient accuracy, which may lead to inappropriate clinical decisions, including unnecessary toxic exposure for low-risk patients and insufficient intervention for high-risk individuals. This study integrates multiple conventional clinical risk factors, including adjuvant chemotherapy status, preoperative CEA, liver cirrhosis and systemic inflammatory indicators such as PLR, to construct a comprehensive predictive model. This multi-factor combination framework can better reflect the complex interactions between different clinical variables and generate continuous individualized survival risk evaluation, instead of simple binary high- or low-risk grouping.[15] This comprehensive risk assessment framework may help identify patients who are more likely to obtain clinical benefits from adjuvant chemotherapy and improve the balance between treatment efficacy and safety. This predictive model is established based on readily available routine clinical indicators, showing good accessibility and practical feasibility in clinical scenarios. Emerging prognostic tools, such as radiomics, omics and molecular detection assays, have shown certain predictive potential, but their wide application is limited by specialized equipment requirements, unified detection standards and relatively high economic costs.[16] In comparison, the current model only incorporates general perioperative data, including demographic characteristics, imaging features, intraoperative records and routine histopathological results, which are universally accessible in most medical centers. Combined with advanced analytical algorithms, this model achieves acceptable predictive performance without additional expensive examinations, supporting its potential external application in different medical institutions. The current findings may offer reference for perioperative clinical decision-making in resected ICC. The model provides a data-driven basis for screening suitable candidates for postoperative adjuvant therapy, which is of certain significance for optimizing selective intervention, especially considering the heterogeneous therapeutic effects of adjuvant chemotherapy in unselected patient cohorts.[17] To facilitate clinical transformation, a web-based risk calculator was developed based on the established model. This online tool allows clinicians to input routine clinical parameters to obtain individualized survival prediction results and risk stratification information, which can assist in real-time prognostic evaluation and doctor-patient shared decision-making for postoperative management and adjuvant therapy arrangement. Several limitations of this study should be fully acknowledged. First, this is a retrospective non-randomized study, and inherent selection bias and information bias cannot be completely excluded during data collection and analysis. Second, although we adopted a multi-center design and included an independent external validation cohort to verify model performance, the overall sample size remains relatively limited, and larger prospective multicenter cohorts are still needed to further confirm the long-term generalizability and clinical applicability of the model. Third, only conventional clinicopathological indicators were included in the current study, and some emerging prognostic biomarkers were not incorporated. Molecular indicators such as ctDNA detection and gene mutation profiling (e.g., FGFR2/IDH1) have shown promising value in evaluating minimal residual disease and tumor recurrence risk in ICC.[18,19] Integrating these novel biomarkers into future models may further improve predictive accuracy and better conform to the concept of precision oncology. In addition, slight differences in treatment protocols and follow-up management existed across participating centers, which may introduce potential confounding factors and affect the stability of outcome analysis. In summary, postoperative management for resected ICC requires comprehensive consideration of the potential benefits and existing controversies of adjuvant chemotherapy. Based on conventional clinicopathological variables, this study constructed an individualized prognostic prediction model using comprehensive analytical methods, which may serve as a supplementary tool for clinical risk stratification and help screen potential candidates for adjuvant chemotherapy. Further large-scale prospective studies and external validation are required to confirm its clinical efficacy. Subsequent research can focus on combining molecular biomarkers to optimize the prediction system, so as to formulate more precise, risk-adapted postoperative treatment strategies and promote individualized clinical management for ICC patients after curative resection. Declarations Data availability The datasets and source code used and generated during the current study are included as supplementary material of this manuscript. Due to patient privacy and ethical restrictions, raw clinical data cannot be publicly shared. All relevant analytical code and processed anonymized data are available in the supplementary files for editorial and peer review purposes. Author contributions statement Conceptualization: Xin Zheng, Yang Yang Data curation: Yuhan Zhou, Yichen Wang, Yiting Dou, Haonan Liu, Yitao Liu, Xiaoyu Li, Xiurui Zhao, Tianli Liu, Jiayi Zhang, Yujing Zhang, Yufang Liu, Linwei Yang, Yan Yan, Peiyuan Meng Formal analysis, validation and visualization: Yuhan Zhou, Yichen Wang, Xiurui Zhao Investigation: Yuhan Zhou, Youwei Wu, Xutian Wang, Dong Wang, Wei Peng, Li Yu, Hengchao Yu, Gang Wang, Tao Li, Shangbo Jin, Zhendong Jiao Methodology: Yang Yang, Yichen Wang Project administration: Yuhan Zhou Software: Yichen Wang Supervision: Youwei Wu, Xutian Wang, Dong Wang, Wei Peng, Li Yu, Hengchao Yu Writing – original draft: Yuhan Zhou, Yichen Wang Writing – review & editing: Xin Zheng, Yang Yang Additional Information Assistance with the study: none. Financial support and sponsorship: none. Conflicts of interest: none. Presentation: none. Funding : The author received No Funding for this work. References Banales JM, Burroughs AK, Marin J, et al. Cholangiocarcinoma[J]. Lancet, 2021, 397(10292): 2185-2200. Bridgewater J, Galle PR, Khan SA, et al. Guidelines for the diagnosis and management of intrahepatic cholangiocarcinoma[J]. J Hepatol, 2021, 74(1): 215-245. Banales JM, Burroughs AK, Marin J, et al. Cholangiocarcinoma[J]. Lancet, 2021, 397(10292): 2185-2200. Zhou J, Zhang Y, Liu Y, et al. Development and validation of a nomogram for predicting overall survival in patients with resected intrahepatic cholangiocarcinoma[J]. BMC Cancer, 2020, 20(1): 1064. Wang L, Chen X, Li J, et al. Prognostic models for intrahepatic cholangiocarcinoma: A systematic review and meta-analysis[J]. Front Oncol, 2022, 12: 927634. Zhou S, Zhang H, Li M, et al. Pathway-aware multimodal transformer (PAMT): Integrating pathological image and gene expression for interpretable cancer survival analysis[J]. IEEE Trans Pattern Anal Mach Intell, 2025. Chen Y, Li Y, Zhang Z, et al. Multi-omics integration improves prognostic prediction in intrahepatic cholangiocarcinoma[J]. Comput Biol Chem, 2023, 102: 107689. Jiang Y, Wang H, Zhang L, et al. Limitations of current prognostic models for intrahepatic cholangiocarcinoma: A critical review[J]. World J Gastroenterol, 2022, 28(24): 2685-2702. Zhang Q, Li X, Wang Y, et al. Comprehensive prognostic models integrating clinical and pathological factors for intrahepatic cholangiocarcinoma[J]. J Clin Transl Hepatol, 2023, 11(3): 456-465. Gore ME, Goodman K, Alberts SR, et al. Capecitabine as adjuvant therapy for biliary tract cancer: Results from the BILCAP trial[J]. N Engl J Med, 2021, 384(11): 1031-1042. Marin J, Macarulla T, Lamarca A, et al. Adjuvant therapy in biliary tract cancer: An update[J]. Crit Rev Oncol Hematol, 2023, 183: 103785. Choi JS, Kim HS, Lee JM, et al. Prognostic factors for recurrence and survival after curative resection of intrahepatic cholangiocarcinoma[J]. J Clin Oncol, 2020, 38(15): 1688-1698. Gore ME, Goodman K, Alberts SR, et al. Capecitabine as adjuvant therapy for biliary tract cancer: Results from the BILCAP trial[J]. N Engl J Med, 2021, 384(11): 1031-1042. Yeo YH, Cho YK, Kim BH, et al. Meta-analysis of adjuvant chemotherapy for resected intrahepatic cholangiocarcinoma[J]. Br J Surg, 2020, 107(8): 995-1004. Xu Y, Zhang L, Wang H, et al. A novel integrated prognostic model for individualized survival prediction in patients with resected intrahepatic cholangiocarcinoma[J]. J Hepatol, 2023, 78(3): 567-578. Huang G, Li Y, Zhang H, et al. Limitations of radiomics and multi-omics approaches in clinical prognostic prediction for intrahepatic cholangiocarcinoma[J]. Radiology, 2022, 305(2): 389-400. Chen X, Zhang J, Li H, et al. Heterogeneity in adjuvant therapy trials for intrahepatic cholangiocarcinoma: A systematic review[J]. J Hepatol, 2022, 76(4): 952-961. Cohen JD, Tie J, Wang D, et al. Circulating tumor DNA as a marker of minimal residual disease in biliary tract cancers[J]. N Engl J Med, 2022, 386(13): 1214-1224. Li N, He Y, Zhang S, et al. Circulating tumor DNA-based minimal residual disease detection predicts recurrence in intrahepatic cholangiocarcinoma[J]. Nat Med, 2023, 29(4): 789-798. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterial.pdf SupplementaryCodeandData.zip Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 30 Apr, 2026 Editor assigned by journal 30 Apr, 2026 Editor invited by journal 30 Apr, 2026 Submission checks completed at journal 29 Apr, 2026 First submitted to journal 29 Apr, 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-9513815","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":636310676,"identity":"bf824bf1-ae2d-486b-b3ed-2a4d803bf64b","order_by":0,"name":"Yuhan Zhou","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yuhan","middleName":"","lastName":"Zhou","suffix":""},{"id":636310677,"identity":"f277ab6d-a28e-43a6-8077-5cad1dc43dd6","order_by":1,"name":"Yichen Wang","email":"","orcid":"","institution":"Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yichen","middleName":"","lastName":"Wang","suffix":""},{"id":636310678,"identity":"b01abad6-cb53-4f50-b539-91914885fd0c","order_by":2,"name":"Haonan Liu","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Haonan","middleName":"","lastName":"Liu","suffix":""},{"id":636310679,"identity":"98105ca4-6589-4eb0-9fd3-b2407d62fdc1","order_by":3,"name":"Yiting Dou","email":"","orcid":"","institution":"Sheng Jing Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yiting","middleName":"","lastName":"Dou","suffix":""},{"id":636310680,"identity":"dee31e71-43e3-4e49-9030-a1dfb5bb99c3","order_by":4,"name":"Youwei Wu","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Youwei","middleName":"","lastName":"Wu","suffix":""},{"id":636310681,"identity":"a453500b-b191-4ff4-956d-a68475c4810d","order_by":5,"name":"Xutian Wang","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Xutian","middleName":"","lastName":"Wang","suffix":""},{"id":636310682,"identity":"d99c7c9a-89ed-4767-b911-852a7d07e8e3","order_by":6,"name":"Yitao Liu","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yitao","middleName":"","lastName":"Liu","suffix":""},{"id":636310683,"identity":"5e922499-af8c-406d-8428-7468ba093a5d","order_by":7,"name":"Xiaoyu Li","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoyu","middleName":"","lastName":"Li","suffix":""},{"id":636310684,"identity":"708ca3b6-3ce0-49d8-90ce-2337adc2a6e9","order_by":8,"name":"Xiurui Zhao","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Xiurui","middleName":"","lastName":"Zhao","suffix":""},{"id":636310685,"identity":"a82a80b0-ac0a-4c6d-9589-0e5db47bf9c5","order_by":9,"name":"Tianli Liu","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Tianli","middleName":"","lastName":"Liu","suffix":""},{"id":636310686,"identity":"01dd9566-3b80-459a-a579-1ce67d2567b4","order_by":10,"name":"Jiayi Zhang","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Jiayi","middleName":"","lastName":"Zhang","suffix":""},{"id":636310687,"identity":"3c1de23f-83cb-4b2a-a554-0e2896eaf88c","order_by":11,"name":"Yujing Zhang","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yujing","middleName":"","lastName":"Zhang","suffix":""},{"id":636310688,"identity":"d6ed92ab-3430-46b7-ba0b-3dbaeb1dd50d","order_by":12,"name":"Yufang Liu","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yufang","middleName":"","lastName":"Liu","suffix":""},{"id":636310689,"identity":"999a0e89-af98-4f67-9ef6-6db0c98eeea5","order_by":13,"name":"Yan Yan","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Yan","suffix":""},{"id":636310690,"identity":"88ae3cf7-33a9-4376-a4ea-38000ec461f1","order_by":14,"name":"Linwei Yang","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Linwei","middleName":"","lastName":"Yang","suffix":""},{"id":636310691,"identity":"13edfdc3-2ea1-46d6-a3be-ff1f18d81a53","order_by":15,"name":"Peiyuan Meng","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Peiyuan","middleName":"","lastName":"Meng","suffix":""},{"id":636310692,"identity":"dd0cf060-5d65-4a7c-a136-aebf7f152bfa","order_by":16,"name":"Wei Peng","email":"","orcid":"","institution":"West China Hospital of Sichuan University","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Peng","suffix":""},{"id":636310693,"identity":"f000636b-e765-48ef-bc35-b3891782f554","order_by":17,"name":"Dong Wang","email":"","orcid":"","institution":"Tang Du Hospital","correspondingAuthor":false,"prefix":"","firstName":"Dong","middleName":"","lastName":"Wang","suffix":""},{"id":636310694,"identity":"2a48222c-67d5-49ca-9850-863b9eeaca53","order_by":18,"name":"Li Yu","email":"","orcid":"","institution":"Sheng Jing Hospital","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Yu","suffix":""},{"id":636310695,"identity":"adc5eab8-8e1f-4ea1-b748-4f522b90b3dc","order_by":19,"name":"Hengchao Yu","email":"","orcid":"","institution":"Xijing Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hengchao","middleName":"","lastName":"Yu","suffix":""},{"id":636310696,"identity":"026201d9-f44d-4888-bdd4-d3e8e4ef88f3","order_by":20,"name":"Gang Wang","email":"","orcid":"","institution":"Baoji Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Gang","middleName":"","lastName":"Wang","suffix":""},{"id":636310697,"identity":"dc0983eb-a8ae-46f0-941c-55ab8072ed60","order_by":21,"name":"Tao Li","email":"","orcid":"","institution":"Yulin Hospital of The First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Li","suffix":""},{"id":636310698,"identity":"6aa5fc61-d3a4-4890-95c4-176b3fe3d507","order_by":22,"name":"Shangbo Jin","email":"","orcid":"","institution":"Baoji People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shangbo","middleName":"","lastName":"Jin","suffix":""},{"id":636310699,"identity":"8ef5c129-1ab0-44eb-a592-41b3f428a385","order_by":23,"name":"Zhendong Jiao","email":"","orcid":"","institution":"Baoji People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Zhendong","middleName":"","lastName":"Jiao","suffix":""},{"id":636310700,"identity":"9036e834-d5ef-40aa-bc64-d9d8d3152a4e","order_by":24,"name":"Yang Yang","email":"","orcid":"","institution":"Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Yang","suffix":""},{"id":636310701,"identity":"71f14aed-4920-4ae2-9015-4b0fa4ca4d20","order_by":25,"name":"Xin Zheng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/0lEQVRIiWNgGAWjYDACCSjJz8CQwMBgc4ABTBOlRbKBIbGBIY14LQwMBgcYGInTIj+7+djDr20WecY3Ep4/5km4w8DPnmPA8HMHbi2Mc46lG8uckSg2O3MgsZkn4RmDZM8bA8beM7i1MEvkmElLVEgkbjvekNjM++Mwg8GNHANmxjbcWtgk8r9JSxhIJG5uZgDZcpjBnpAWHokcNskPQFs2sDdAtBhIENAiIZFmJs1wRiJxBtAvM+ckHOaROPOs4GAvHi3yM5KfSf5sq0vsn5GT8OFNwmE5/vbkjQ9+4tECDgIeiBsTwCSIOIBfAzCgf4ApdoIKR8EoGAWjYIQCALh1UkK0HC8DAAAAAElFTkSuQmCC","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":true,"prefix":"","firstName":"Xin","middleName":"","lastName":"Zheng","suffix":""}],"badges":[],"createdAt":"2026-04-24 07:38:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9513815/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9513815/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108973678,"identity":"4f3cd9ae-025a-4aef-8530-3574686a0af8","added_by":"auto","created_at":"2026-05-11 10:43:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":123690,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMultidimensional analysis of key prognostic features for intrahepatic cholangiocarcinoma (ICC) postoperative survival. \u003c/strong\u003e(a) Composite importance of the top 15 features: This visualization presents the composite importance scores of the top 15 discriminative features, which were calculated by a weighted fusion strategy (F-value: 33%, random forest: 33%, LASSO: 34%) after normalizing the importance scores of each feature selection method to a [0, 1] scale, reflecting the relative prognostic value of each feature. (b) Radar chart of the top 8 features: As a typical multivariate comparison tool, this radar chart integrates 8 key features, with each radial axis corresponding to one feature and the closed polyline reflecting the overall performance profile of the top 8 features, intuitively showing the comprehensive characteristics of the core prognostic factors. (c) Correlation heatmap of the top 15 features: The color scale on the right of the heatmap quantifies Pearson correlation coefficients (range: -1 to 1), where values close to 1, -1, and 0 indicate strong positive, strong negative, and weak/no linear correlation, respectively. Most pairwise coefficients are \u0026lt; 0.3, with no strong correlation (|r| ≥ 0.7), indicating that the selected top 15 features provide complementary rather than redundant information for predicting ICC survival outcomes.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/0332b2e99858b6e53793e8b6.png"},{"id":108973792,"identity":"cb49ef1b-62f1-4d7d-b39c-f6235f5b552e","added_by":"auto","created_at":"2026-05-11 10:44:03","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":117780,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eArchitecture Design of Transformer-MLP Model: \u003c/strong\u003eThe proposed Transformer-MLP architecture processes\u0026nbsp;31\u003cem\u003e \u003c/em\u003eclinical features by projecting them into a 64-dimensional embedding space (d\u003csub\u003emodel\u003c/sub\u003e=64). The core encoder consists of\u0026nbsp;2 Transformer blocks, each utilizing\u0026nbsp;4-head\u0026nbsp;self-attention and a feed-forward network with a hidden dimension of\u0026nbsp;128, stabilized by layer normalization and residual connections. The resulting sequence is aggregated via global average pooling into a 64-dimensional vector. Subsequently, this vector feeds into an\u0026nbsp;MLP classification head\u0026nbsp;containing a 32-unit\u0026nbsp;dense layer with ReLU activation and Dropout for feature fusion, culminating in a Sigmoid output layer for survival probability prediction. MLP: Multilayer Perceptron.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/b5c8b5670960ca3d27743d12.png"},{"id":108973785,"identity":"1f741207-76be-4468-9806-0369f0719529","added_by":"auto","created_at":"2026-05-11 10:44:00","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":168318,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCalibration curves of the Transformer-MLP model:\u003c/strong\u003e (a) Calibration curves of the Transformer-MLP model on the training \u0026amp; validation sets, where the x-axis represents the predicted probabilities of the model and the y-axis denotes the actual observed frequencies of the target event; (b) Calibration curve of the Transformer-MLP model on the test set, following the same axis definitions as (a); (c) Calibration curves of the Transformer-MLP model on the test set after applying Platt Scaling and Isotonic Regression calibration methods. MLP: Multilayer Perceptron.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/9514bfebf751352569bd34f9.png"},{"id":108973667,"identity":"3f23b4cc-7f94-4e2f-b83b-389eddf082f0","added_by":"auto","created_at":"2026-05-11 10:43:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":200171,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClinical decision curves of all models:\u003c/strong\u003e. (a) Clinical decision curves of five traditional machine learning models, where the x-axis represents the threshold probability and the y-axis denotes the net benefit of the model; (b) Clinical decision curves of six deep learning models, following the same axis definitions as (a). MLP: Multilayer Perceptron.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/0e0143a32199c550e1476c16.png"},{"id":108973687,"identity":"799ab1f7-c050-4e37-98ee-659dc3e32cef","added_by":"auto","created_at":"2026-05-11 10:43:54","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":129083,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e5-Fold Cross-Validation ROC for Transformer-MLP Model: \u003c/strong\u003eThis figure presents the ROC curves of the Transformer-MLP model across five independent folds of cross-validation. MLP: Multilayer Perceptron, ROC: Receiver Operating Characteristic.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/b811eaf12519943a37d739d0.png"},{"id":108973680,"identity":"570e6a16-5753-405c-9e38-c30b5b2b3340","added_by":"auto","created_at":"2026-05-11 10:43:49","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":122742,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReceiver Operating Characteristic Curves of all models:\u003c/strong\u003e (a) ROC curves of all 11 predictive models on the training \u0026amp; validation sets; (b) ROC curves of 5 traditional machine learning models on the independent test set; (c) ROC curves of 6 deep learning models on the independent test set. ROC: Receiver Operating Characteristic.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/317eb978a585442258758549.png"},{"id":108973793,"identity":"7ef1ef48-eddb-4478-bd3f-c16570fd7227","added_by":"auto","created_at":"2026-05-11 10:44:03","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":5713,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure 7.\u003c/strong\u003e \u003cstrong\u003eKaplan-Meier survival curves of overall survival stratified by the Transformer-MLP model:\u003c/strong\u003e(a) KM curves for high-risk and low-risk groups in the training\u0026amp;validation sets; (b) KM curves for high-risk and low-risk groups in the test set. KM: Kaplan-Meier, OS: overall survival, MLP: Multilayer Perceptron.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/05f6d33b4f6df21272488806.png"},{"id":108973664,"identity":"2c8925ca-257f-49b1-ac70-127047865c1b","added_by":"auto","created_at":"2026-05-11 10:43:40","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":5713,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure 8. SHapley Additive exPlanations plots for interpreting the Transformer-MLP model: \u003c/strong\u003e(a) SHAP bar plot: ranking the relative importance of all predictive features; (b) SHAP scatter plot: illustrating the relationship between key feature values and their corresponding SHAP values; (c) SHAP waterfall plot: showing the contribution of individual features to the model’s prediction for a single sample. SHAP: SHapley Additive exPlanations, AC: adjuvant chemotherapy, CEA: Carcinoembryonic Antigen, AJCC: American Joint Committee on Cancer ,CA19-9: Carbohydrate Antigen 19-9, PLR: platelet-to-lymphocyte ratio.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/c3e326c4a99bea15bf37a04b.png"},{"id":108973922,"identity":"bdc011a8-9ea2-48ea-a85e-cf93e00bb036","added_by":"auto","created_at":"2026-05-11 10:44:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":951926,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/376a0d59-8e1a-462a-a0c0-976c6a86118e.pdf"},{"id":108973679,"identity":"3e716ab2-b4cc-4eab-8ec7-706d8274f909","added_by":"auto","created_at":"2026-05-11 10:43:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":3440709,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/9ed3dfb5c541edd4983f1757.pdf"},{"id":108973784,"identity":"3d9c83fa-c9b0-47ff-b12d-784556e4f800","added_by":"auto","created_at":"2026-05-11 10:44:00","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":20169947,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryCodeandData.zip","url":"https://assets-eu.researchsquare.com/files/rs-9513815/v1/46a7c2bc473f0000e1b2c192.zip"}],"financialInterests":"No competing interests reported.","formattedTitle":"A transformer-MLP model with SHAP analysis for predicting postoperative survival in Intrahepatic Cholangiocarcinoma","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIntrahepatic cholangiocarcinoma (ICC), the second most common primary liver malignancy, poses a substantial global health burden, with rising incidence and persistently poor prognosis.\u003csup\u003e[1]\u003c/sup\u003e Surgical resection remains the cornerstone of potentially curative treatment; however, this approach is severely limited by a high recurrence rate\u0026mdash;over 50% of patients experience disease recurrence within two years postoperatively\u0026mdash;resulting in a dismal 5-year overall survival rate of only 30-40%.\u003csup\u003e[2]\u003c/sup\u003e There is an urgent need for effective adjuvant therapies targeting residual micrometastatic disease to improve long-term outcomes. Currently, the field of adjuvant therapy for resected ICC lacks consensus, with considerable controversy and a paucity of clear, evidence-based clinical guidelines.\u003csup\u003e[3]\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eA robust prognostic model is a critical decision-support tool in ICC management by enabling precise risk stratification.\u003csup\u003e[4]\u003c/sup\u003e Such models facilitate the classification of patients into distinct risk groups, a cornerstone of personalized therapy: they help identify high-risk patients most likely to benefit from adjuvant treatment, while sparing low-risk patients unnecessary toxicity. Furthermore, prognostic models optimize the design of more efficient clinical trials via precise patient enrichment, ensuring novel adjuvant strategies are tested in populations with the highest potential for response.\u003csup\u003e[5]\u003c/sup\u003e By integrating multi-dimensional data\u0026mdash;including clinical, pathological, and surgical variables\u0026mdash;comprehensive models transcend single-modality assessments to deliver holistic, accurate risk evaluations, which are essential for refining individualized postoperative management in ICC.\u003csup\u003e[6]\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003ePrognostic modeling for ICC is a rapidly evolving field. Traditional nomograms mainly rely on clinicopathological features, whereas bioinformatics models leverage genomic data. The advancement of artificial intelligence (AI)\u0026mdash;especially machine learning (ML) for predicting short-term surgical outcomes and deep learning (DL) for analyzing complex histopathological and radiological data\u0026mdash;marks a major breakthrough.\u003csup\u003e[7]\u003c/sup\u003e State-of-the-art multi-omics integration approaches further combine genomic, transcriptomic, and radiomic data to enhance predictive performance.\u003csup\u003e[8]\u003c/sup\u003e However, significant limitations persist in current models: many focus on a single data modality, lack rigorous external validation, and exhibit insufficient interpretability for direct clinical translation.\u003csup\u003e[9]\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eThis study was designed to address these gaps by developing and validating a novel Transformer-MLP deep learning model for predicting postoperative survival in ICC patients. A key methodological innovation was the adoption of a differentiated feature-selection strategy tailored to the unique characteristics of each model type. For traditional machine learning algorithms, a refined set of 17 key features was used to maximize interpretability and computational efficiency. In contrast, the deep learning model incorporated the full set of 31 available features, leveraging its inherent capacity to automatically weight features and identify complex non-linear interactions in high-dimensional data. The Transformer-MLP model was trained on a multi-center development cohort and validated on an independent external test set. It also demonstrated good calibration and yielded a positive net clinical benefit in decision curve analysis, underscoring its potential clinical utility.\u0026nbsp;\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003e\u003cstrong\u003e2.1. Data Sources and Study Ethics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis multicenter, retrospective cohort study was conducted using data from eight tertiary hospitals in China. We identified patients who underwent curative-intent radical liver resection between January 2010 and January 2022 across five cities (Xi\u0026apos;an, Chengdu, Shenyang, Yulin, and Baoji) and had a postoperative pathological confirmation of ICC. After applying the inclusion and exclusion criteria and excluding cases lost to follow-up, a final cohort of 499 patients was included for analysis. The study population comprised 288 (57.7%) male and 211 (42.3%) female patients, with an age range of 26 to 84 years. A detailed flowchart of patient selection is presented in figure S1.\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of The First Affiliated Hospital of Xi\u0026rsquo;an Jiaotong University (Approval No.: XJTU1AF2025LSYY-393) on 13 March 2025. All procedures conducted in this study were performed in strict accordance with the ethical standards of the institutional research committee, as well as the 1964 Helsinki Declaration and its subsequent amendments. Given the retrospective nature of this study, written informed consent from individual patients was formally waived by the aforementioned ethics committee.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2. Data Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSurgical, Adjuvant Chemotherapy and follow-up Protocols see in method S1. We initially evaluated 31 clinical and pathological features across several domains: patient demographics (sex, age, BMI), comorbidities (hepatitis, cirrhosis, hepatobiliary and cardiopulmonary diseases), surgical details (surgical approach, lymph node dissection, extent of resection, intraoperative blood loss), laboratory values (PLR, NLR, AST, ALT, GGT, TBIL, ALB, PT), and tumor characteristics (AFP, CEA, CA19-9, differentiation grade, microvascular invasion, perineural invasion, capsular invasion, tumor number, tumor size, lymph node metastasis, AJCC stage). Collection of Clinicopathological Data see in method S2.\u003c/p\u003e\n\u003cp\u003eTo ensure data quality and improve model stability and predictive performance, we performed systematic pre-processing on the raw clinical data. This included the transformation of numerical variables, encoding of categorical variables, and preparation of the data for survival analysis. Specific data preprocessing process see in method S3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.3. Dataset Partitioning and Validation Strategy\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo comprehensively evaluate the model\u0026apos;s predictive stability (internal consistency) and clinical generalizability (external validity), we employed a two-stage validation framework consisting of internal five-fold cross-validation and performance testing on a completely independent external set.\u003c/p\u003e\n\u003cp\u003eThe internal development cohort comprised 328 patients from the First Affiliated Hospital of Xi\u0026apos;an Jiaotong University. We performed stratified 5-fold cross-validation on this cohort to ensure reliable model development and initial validation. Specific Steps of Five-Fold Cross-Validation see in method S4. All models were developed through this process.\u003c/p\u003e\n\u003cp\u003eAfter determining the optimal model architecture and hyperparameters via cross-validation, the final model\u0026apos;s generalization capability was rigorously assessed on a completely independent external test set. This set consisted of 171 patients from the remaining 7 participating centers, which were not involved in any stage of internal development.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4 Feature selection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study employed a systematic two-stage feature selection approach to identify the most predictive subset from an initial set of 31 clinicopathological features.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4.1 Feature importance ranking based on multi-method fusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo robustly identify the most predictive features from an initial set of 31 clinicopathological variables, we implemented a multi-method fusion strategy for feature importance ranking. Three complementary feature selection techniques were employed: univariate F-value analysis, random forest-based importance, and LASSO regression (figure S2-4).\u003c/p\u003e\n\u003cp\u003eTo synthesize the results from these diverse methodologies, the importance scores from each method were first normalized to a [0, 1] scale. A weighted fusion scheme was then applied (F-value: 33%, random forest: 33%, LASSO: 34%) to calculate a composite importance score for each feature. This multi-method fusion strategy enhances the robustness of the ranking by leveraging the strengths of each approach .Fusion results demonstrated the stability of these features\u0026apos; prognostic value regardless of the underlying selection logic, and that the selected features provide complementary rather than redundant information for predicting ICC survival outcomes (figure 1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4.2 Feature Number Optimization via Cross-Validation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo determine the optimal number of features (k*) across different model architectures, a systematic screening of feature subset sizes was performed. The procedure see in method S5.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSystematic assessment of five traditional machine learning models across 1\u0026ndash;31 feature subsets revealed distinct performance patterns and optimal feature set sizes by algorithm (table S2 and figure S5ab). Overall, model performance stabilized substantially after 8 features; expanding to 17 features yielded minimal average AUC improvement (\u0026lt;0.03), indicating a clear saturation point with diminishing marginal gains.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eComprehensive feature inclusion experiments on six deep learning models revealed a complex non-monotonic relationship between model performance and input feature count. Unlike traditional machine learning models, which often plateau beyond a certain feature threshold, deep learning models exhibited more nuanced fluctuating performance (table S2 and figure S5cd). In stark contrast to traditional ML models (typically saturated beyond 8 features), this highlights their inherent advantage in harnessing high-dimensional data\u0026mdash;attributed to integrated feature selection and interaction learning mechanisms (enhanced by feature/paradigm fusion). This adaptability reduces information loss from preemptive feature filtering.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.5 Model Development \u0026amp; Comparison\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate machine learning approaches for predicting 2-year postoperative OS, this study developed and compared 11 prediction models. The model set included five traditional algorithms\u0026mdash;logistic regression, support vector machine (SVM), LightGBM, random forest, and XGBoost\u0026mdash;chosen to cover a gradient of interpretability and algorithmic complexity, alongside six deep learning architectures: multilayer perceptron (MLP), Transformer, recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), and a novel Transformer-MLP model. Covering diverse representational learning strategies (from feedforward hierarchies to dynamic gating and attention-based feature integration), these models enabled comprehensive assessment of deep learning\u0026rsquo;s utility in prognostic prediction. Their theoretical rationale see in method S6.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.6. Design of Transformer-MLP Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study proposes a composite network architecture that first employs a multi-head attention mechanism to comprehensively capture features, followed by a MLP that performs nonlinear mapping using the extracted high-dimensional features to achieve precise prediction of 2-year survival prognosis. The specific workflow is illustrated in figure S6.\u003c/p\u003e\n\u003cp\u003eAfter mapping clinical pathological information into a high-dimensional space, the model feeds it into a Transformer layer for more comprehensive and refined feature extraction. Subsequently, a MLP head with 32 neurons outputs a predicted probability value (0-1 range) representing the patient\u0026apos;s risk of death within two years. This value is then converted into a survival status classification prediction (0/1) using a 0.5 threshold.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn this Transformer-MLP model, the correlation among different clinical indicators is effectively captured by the Transformer model through its attention mechanism. Subsequently, the MLP acts as a dedicated classifier, performing nonlinear inference on the features to output the final risk score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.7 SHAP Model Interpretation Analysis\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo enhance the clinical interpretability of the prognostic model, we performed SHapley Additive exPlanations (SHAP) analysis to quantify and rank the contribution of each clinical and histological variable to model outputs. Rooted in cooperative game theory, SHAP values offer a unified measure of feature importance by fairly attributing the difference between an instance\u0026rsquo;s prediction and the dataset\u0026rsquo;s average prediction to each feature, enabling both local (per-instance) and global (dataset-wide) interpretability for identifying key survival-predictive factors. Analysis was conducted on the final trained model with the optimized feature set, using PermutationExplainer to estimate SHAP values\u0026mdash;this model-agnostic method approximates Shapley values by permuting features and measuring prediction changes, ensuring reliability for the model\u0026rsquo;s structure.\u003c/p\u003e\n\u003cp\u003eFor global interpretation, we generated SHAP summary plots (bar and scatter plots): bar plots show mean absolute SHAP values (overall feature importance), while scatter plots illustrate SHAP value distributions across samples, revealing impact direction (positive/negative on predictions) and potential non-linear feature-output relationships. For individual patient-level local interpretation, decision plots visualized how feature contributions accumulate to shift predictions from the base value to the final output for specific cases.\u003c/p\u003e\n\u003cp\u003eSHAP analysis confirmed that top features (by mean absolute SHAP value) aligned with established clinical prognostic factors for ICC, validating the model\u0026rsquo;s clinical relevance and biological plausibility. This framework clarified the model\u0026rsquo;s decision-making process, provided a rigorous data-driven basis for feature selection, and enhanced model robustness. By identifying key prognostic variables and their directional survival impacts, SHAP analysis delivers transparent, actionable insights to support individualized treatment strategy development.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.8 Statistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStatistical analyses were performed using Python 3.9.6 with key libraries including scikit-learn, scikit-survival, and the SHAP package. The analytic cohort included a total of 31 features, comprising 17,964 data points. To compare baseline demographic and clinical characteristics between the model development set and the independent test set, the Mann‑Whitney U test was employed for continuous variables, and the Chi‑square test was applied for categorical variables.\u003c/p\u003e\n\u003cp\u003eModel performance was comprehensively evaluated on the test set. The precision of the performance metrics was estimated using 95% confidence intervals obtained from 1,000 bootstrap resamples. Calibration curves were plotted to visualize the agreement between the predicted probabilities and the actual observed 2‑year overall survival (OS). Survival outcomes across different risk strata, as defined by the model, were compared using Kaplan‑Meier curves with the log‑rank test. Decision curve analysis (DCA) further quantified the net clinical benefit of model-based risk stratification across various decision thresholds. To enhance the interpretability of the final Transformer‑MLP model, SHAP were applied to quantify the contribution of each feature to individual predictions. A two - tailed p - value of less than 0.05 was considered statistically significant for all analyses.\u003c/p\u003e"},{"header":"3. Results","content":"\u003cp\u003e\u003cstrong\u003e3.1 Baseline Characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 499 patients meeting the inclusion and exclusion criteria were included in the analysis. Their baseline characteristics are presented in table S1. The vast majority of variables showed no significant differences between the training and test sets (p \u0026gt; 0.05), confirming their comparability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 Final Feature Set Determination\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor traditional ML models, the top 17 composite-score features were selected, as performance plateaued beyond 17. This subset balances optimal performance with better training efficiency, inference speed, and interpretability\u0026mdash;critical for clinical deployment in resource-constrained or high-transparency settings.\u003c/p\u003e\n\u003cp\u003eIn contrast, all 31 features were used for deep learning models, aligning with their data-driven paradigm. Architectures like Transformer-MLP, with self-attention and multi-layer non-linear transformations, inherently perform automatic feature selection/weighting during end-to-end training. The full feature set enables autonomous identification of predictive signals and complex interactions, supported by experiments showing superior overall (highest average AUC) and single-model performance (Transformer-MLP AUC = 0.8131).\u003c/p\u003e\n\u003cp\u003eThis differential strategy highlights that optimal feature sets are algorithm-dependent, enabling flexible deployment: traditional models with 17 core features suit scenarios prioritizing speed, low computational cost, and interpretability, while deep learning models with all 31 features maximize predictive accuracy with adequate computational resources.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 Model Performance Evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe performance of 11 prediction models on training, validation, and test sets (summarized in tables S3-4) was evaluated using C-index, AUC, sensitivity, specificity, PPV, NPV, and Brier score. Several models (XGBoost, LightGBM [AUC=1.0000]; Random Forest [AUC=0.9770]; SVM [AUC=0.9642]; Transformer-MLP [AUC=0.9627]) exhibited excellent discrimination on the training set, though their generalization performance differed significantly on validation and independent test sets.\u003c/p\u003e\n\u003cp\u003eNotably, the Transformer-MLP model achieved the best overall performance on the independent test set (AUC=0.8131, C-index=0.7462, 95% CI: 0.6898-0.7950), with high sensitivity (0.8361) and excellent NPV (0.8667) that effectively identified most event (e.g., death)-experienced patients and enabled reliable low-risk stratification. Other deep learning models (Transformer [AUC=0.7578], MLP [AUC=0.7317], LSTM [AUC=0.7073]) also generally outperformed traditional machine learning models on the test set, which conversely showed significant performance degradation indicative of overfitting\u0026mdash;for example, XGBoost\u0026apos;s AUC dropped from 1.0000 (training) to 0.6121 (test), Random Forest had a test-set AUC of 0.6346 but extremely low sensitivity (0.1148) (failing to identify most high-risk patients), and Logistic Regression (AUC=0.6380) and SVM (AUC=0.6387) exhibited limited test-set performance.\u003c/p\u003e\n\u003cp\u003eThe Transformer-MLP model\u0026apos;s calibration was assessed via calibration curves (figure 2ab for initial training and test sets) and Brier score; to improve test-set calibration, Platt Scaling and Isotonic Regression (trained on the validation set) were applied. Figure 2c compares the original and calibrated models\u0026apos; test-set calibration (with Brier score as a quantitative metric, where lower values indicate better calibration), and the curve shows good agreement between predicted probabilities and observed risk. Additionally, decision curve analysis (figure 3) confirmed the model\u0026apos;s higher net clinical benefit across most clinically relevant threshold probabilities.\u003c/p\u003e\n\u003cp\u003eIn summary, the Transformer-MLP deep learning model (trained on all 31 features) demonstrated superior, robust predictive performance, good calibration, and clinical utility on the independent multicenter test set, making it a reliable tool for predicting postoperative survival in ICC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4 ROC Curves and Kaplan-Meier Survival Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe 5-fold cross-validation performance of Transformer-MLP model see in table S5. The corresponding ROC curve is shown in figure S7.\u0026nbsp;Receiver operating characteristic curves for all 11 models are presented in figure 4. The Transformer-MLP model\u0026apos;s test set AUC is highlighted in figure 4c. The high AUC value confirms its excellent discriminative ability. Furthermore, Kaplan-Meier survival curves stratified by the model\u0026apos;s risk prediction (high vs. low) showed significant separation (figure 5). The statistical significance of this difference, confirmed by the log-rank test, underscores the model\u0026apos;s utility for clinical risk stratification (p \u0026lt; 0.01).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.5 Feature Importance Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSHAP (SHapley Additive exPlanations) was employed to analyze the feature importance of the Transformer-MLP model. Figure 6a ranks postoperative features by mean absolute SHAP value, reflecting their relative predictive importance. The SHAP summary plot (figure 6b) displays the distribution of individual patient SHAP values per feature, with color indicating feature values (blue: low, red: high), clarifying the magnitude and direction of each feature\u0026apos;s impact on model predictions. The waterfall chart (figure 6c) details how individual feature values shift the model\u0026apos;s baseline prediction toward higher or lower death risk.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.6 Online Predictor\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo facilitate clinical application, we developed a publicly accessible online prognostic tool (https://icc-prediction.streamlit.app/). Incorporating all 11 developed models, it delivers individualized risk assessments, generates patient-specific survival probability curves and overall risk classification, and features a user-friendly interface for real-time ICC prognosis evaluation (figure S8).\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003ePostoperative management of intrahepatic cholangiocarcinoma (ICC) after curative resection remains challenging, with tumor recurrence standing as a major clinical obstacle. Over half of patients experience tumor relapse even after complete surgical resection, indicating that surgical intervention alone is insufficient to improve long-term survival.[10] The high postoperative recurrence rate highlights the necessity of exploring reasonable adjuvant treatment strategies for ICC patients. In current clinical practice, adjuvant chemotherapy is widely applied for resected ICC, while debates still exist regarding its overall survival benefit, appropriate candidate selection, and clinical application value.[11] The present prognostic model may assist in precise risk stratification and provide reference for individualized adjuvant treatment decision-making for resected ICC patients.\u003c/p\u003e\n\u003cp\u003eThe recommendation for adjuvant chemotherapy is largely based on the aggressive biological characteristics of ICC. Key pathological variables included in our prediction model, such as lymph node metastasis, elevated preoperative CA19-9 level, advanced AJCC stage and poor tumor differentiation, are recognized markers associated with potential micrometastasis and poor prognosis.[12] Eliminating invisible microscopic lesions is considered the core objective of adjuvant systemic therapy, which is recommended by current clinical guidelines for high-risk biliary tract cancer patients.[1] Relevant clinical trials, such as the BILCAP study, have demonstrated the potential survival benefit of capecitabine in resected biliary tract malignancies, including ICC.[13]\u003c/p\u003e\n\u003cp\u003eNevertheless, the clinical value of adjuvant chemotherapy for ICC remains controversial. Current evidence focusing on ICC-specific cohorts has not yielded consistent and definitive conclusions regarding its long-term survival improvement.[14] Traditional risk evaluation relying on single clinicopathological indicators lacks sufficient accuracy, which may lead to inappropriate clinical decisions, including unnecessary toxic exposure for low-risk patients and insufficient intervention for high-risk individuals. This study integrates multiple conventional clinical risk factors, including adjuvant chemotherapy status, preoperative CEA, liver cirrhosis and systemic inflammatory indicators such as PLR, to construct a comprehensive predictive model. This multi-factor combination framework can better reflect the complex interactions between different clinical variables and generate continuous individualized survival risk evaluation, instead of simple binary high- or low-risk grouping.[15] This comprehensive risk assessment framework may help identify patients who are more likely to obtain clinical benefits from adjuvant chemotherapy and improve the balance between treatment efficacy and safety.\u003c/p\u003e\n\u003cp\u003eThis predictive model is established based on readily available routine clinical indicators, showing good accessibility and practical feasibility in clinical scenarios. Emerging prognostic tools, such as radiomics, omics and molecular detection assays, have shown certain predictive potential, but their wide application is limited by specialized equipment requirements, unified detection standards and relatively high economic costs.[16] In comparison, the current model only incorporates general perioperative data, including demographic characteristics, imaging features, intraoperative records and routine histopathological results, which are universally accessible in most medical centers. Combined with advanced analytical algorithms, this model achieves acceptable predictive performance without additional expensive examinations, supporting its potential external application in different medical institutions.\u003c/p\u003e\n\u003cp\u003eThe current findings may offer reference for perioperative clinical decision-making in resected ICC. The model provides a data-driven basis for screening suitable candidates for postoperative adjuvant therapy, which is of certain significance for optimizing selective intervention, especially considering the heterogeneous therapeutic effects of adjuvant chemotherapy in unselected patient cohorts.[17] To facilitate clinical transformation, a web-based risk calculator was developed based on the established model. This online tool allows clinicians to input routine clinical parameters to obtain individualized survival prediction results and risk stratification information, which can assist in real-time prognostic evaluation and doctor-patient shared decision-making for postoperative management and adjuvant therapy arrangement.\u003c/p\u003e\n\u003cp\u003eSeveral limitations of this study should be fully acknowledged. First, this is a retrospective non-randomized study, and inherent selection bias and information bias cannot be completely excluded during data collection and analysis. Second, although we adopted a multi-center design and included an independent external validation cohort to verify model performance, the overall sample size remains relatively limited, and larger prospective multicenter cohorts are still needed to further confirm the long-term generalizability and clinical applicability of the model. Third, only conventional clinicopathological indicators were included in the current study, and some emerging prognostic biomarkers were not incorporated. Molecular indicators such as ctDNA detection and gene mutation profiling (e.g., FGFR2/IDH1) have shown promising value in evaluating minimal residual disease and tumor recurrence risk in ICC.[18,19] Integrating these novel biomarkers into future models may further improve predictive accuracy and better conform to the concept of precision oncology. In addition, slight differences in treatment protocols and follow-up management existed across participating centers, which may introduce potential confounding factors and affect the stability of outcome analysis.\u003c/p\u003e\n\u003cp\u003eIn summary, postoperative management for resected ICC requires comprehensive consideration of the potential benefits and existing controversies of adjuvant chemotherapy. Based on conventional clinicopathological variables, this study constructed an individualized prognostic prediction model using comprehensive analytical methods, which may serve as a supplementary tool for clinical risk stratification and help screen potential candidates for adjuvant chemotherapy. Further large-scale prospective studies and external validation are required to confirm its clinical efficacy. Subsequent research can focus on combining molecular biomarkers to optimize the prediction system, so as to formulate more precise, risk-adapted postoperative treatment strategies and promote individualized clinical management for ICC patients after curative resection.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets and source code used and generated during the current study are included as supplementary material of this manuscript. Due to patient privacy and ethical restrictions, raw clinical data cannot be publicly shared. All relevant analytical code and processed anonymized data are available in the supplementary files for editorial and peer review purposes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConceptualization: Xin Zheng, Yang Yang\u003c/p\u003e\n\u003cp\u003eData curation: Yuhan Zhou, Yichen Wang, Yiting Dou, Haonan Liu, Yitao Liu, Xiaoyu Li, Xiurui Zhao, Tianli Liu, Jiayi Zhang, Yujing Zhang, Yufang Liu, Linwei Yang, Yan Yan, Peiyuan Meng\u003c/p\u003e\n\u003cp\u003eFormal analysis, validation and visualization: Yuhan Zhou, Yichen Wang, Xiurui Zhao\u003c/p\u003e\n\u003cp\u003eInvestigation: Yuhan Zhou, Youwei Wu, Xutian Wang, Dong Wang, Wei Peng, Li Yu, Hengchao Yu, Gang Wang, Tao Li, Shangbo Jin, Zhendong Jiao\u003c/p\u003e\n\u003cp\u003eMethodology: Yang Yang, Yichen Wang\u003c/p\u003e\n\u003cp\u003eProject administration: Yuhan Zhou\u003c/p\u003e\n\u003cp\u003eSoftware: Yichen Wang\u003c/p\u003e\n\u003cp\u003eSupervision: Youwei Wu, Xutian Wang, Dong Wang, Wei Peng, Li Yu, Hengchao Yu\u003c/p\u003e\n\u003cp\u003eWriting \u0026ndash; original draft: Yuhan Zhou, Yichen Wang\u003c/p\u003e\n\u003cp\u003eWriting \u0026ndash; review \u0026amp; editing: Xin Zheng, Yang Yang\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional Information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAssistance with the study: none.\u003c/p\u003e\n\u003cp\u003eFinancial support and sponsorship: none.\u003c/p\u003e\n\u003cp\u003eConflicts of interest: none.\u003c/p\u003e\n\u003cp\u003ePresentation: none.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding :\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe author received No Funding for this work.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBanales JM, Burroughs AK, Marin J, et al. Cholangiocarcinoma[J]. Lancet, 2021, 397(10292): 2185-2200.\u003c/li\u003e\n\u003cli\u003eBridgewater J, Galle PR, Khan SA, et al. Guidelines for the diagnosis and management of intrahepatic cholangiocarcinoma[J]. J Hepatol, 2021, 74(1): 215-245.\u003c/li\u003e\n\u003cli\u003eBanales JM, Burroughs AK, Marin J, et al. Cholangiocarcinoma[J]. Lancet, 2021, 397(10292): 2185-2200.\u003c/li\u003e\n\u003cli\u003eZhou J, Zhang Y, Liu Y, et al. Development and validation of a nomogram for predicting overall survival in patients with resected intrahepatic cholangiocarcinoma[J]. BMC Cancer, 2020, 20(1): 1064.\u003c/li\u003e\n\u003cli\u003eWang L, Chen X, Li J, et al. Prognostic models for intrahepatic cholangiocarcinoma: A systematic review and meta-analysis[J]. Front Oncol, 2022, 12: 927634.\u003c/li\u003e\n\u003cli\u003eZhou S, Zhang H, Li M, et al. Pathway-aware multimodal transformer (PAMT): Integrating pathological image and gene expression for interpretable cancer survival analysis[J]. IEEE Trans Pattern Anal Mach Intell, 2025.\u003c/li\u003e\n\u003cli\u003eChen Y, Li Y, Zhang Z, et al. Multi-omics integration improves prognostic prediction in intrahepatic cholangiocarcinoma[J]. Comput Biol Chem, 2023, 102: 107689.\u003c/li\u003e\n\u003cli\u003eJiang Y, Wang H, Zhang L, et al. Limitations of current prognostic models for intrahepatic cholangiocarcinoma: A critical review[J]. World J Gastroenterol, 2022, 28(24): 2685-2702.\u003c/li\u003e\n\u003cli\u003eZhang Q, Li X, Wang Y, et al. Comprehensive prognostic models integrating clinical and pathological factors for intrahepatic cholangiocarcinoma[J]. J Clin Transl Hepatol, 2023, 11(3): 456-465.\u003c/li\u003e\n\u003cli\u003eGore ME, Goodman K, Alberts SR, et al. Capecitabine as adjuvant therapy for biliary tract cancer: Results from the BILCAP trial[J]. N Engl J Med, 2021, 384(11): 1031-1042.\u003c/li\u003e\n\u003cli\u003eMarin J, Macarulla T, Lamarca A, et al. Adjuvant therapy in biliary tract cancer: An update[J]. Crit Rev Oncol Hematol, 2023, 183: 103785.\u003c/li\u003e\n\u003cli\u003eChoi JS, Kim HS, Lee JM, et al. Prognostic factors for recurrence and survival after curative resection of intrahepatic cholangiocarcinoma[J]. J Clin Oncol, 2020, 38(15): 1688-1698.\u003c/li\u003e\n\u003cli\u003eGore ME, Goodman K, Alberts SR, et al. Capecitabine as adjuvant therapy for biliary tract cancer: Results from the BILCAP trial[J]. N Engl J Med, 2021, 384(11): 1031-1042.\u003c/li\u003e\n\u003cli\u003eYeo YH, Cho YK, Kim BH, et al. Meta-analysis of adjuvant chemotherapy for resected intrahepatic cholangiocarcinoma[J]. Br J Surg, 2020, 107(8): 995-1004.\u003c/li\u003e\n\u003cli\u003eXu Y, Zhang L, Wang H, et al. A novel integrated prognostic model for individualized survival prediction in patients with resected intrahepatic cholangiocarcinoma[J]. J Hepatol, 2023, 78(3): 567-578.\u003c/li\u003e\n\u003cli\u003eHuang G, Li Y, Zhang H, et al. Limitations of radiomics and multi-omics approaches in clinical prognostic prediction for intrahepatic cholangiocarcinoma[J]. Radiology, 2022, 305(2): 389-400.\u003c/li\u003e\n\u003cli\u003eChen X, Zhang J, Li H, et al. Heterogeneity in adjuvant therapy trials for intrahepatic cholangiocarcinoma: A systematic review[J]. J Hepatol, 2022, 76(4): 952-961.\u003c/li\u003e\n\u003cli\u003eCohen JD, Tie J, Wang D, et al. Circulating tumor DNA as a marker of minimal residual disease in biliary tract cancers[J]. N Engl J Med, 2022, 386(13): 1214-1224.\u003c/li\u003e\n\u003cli\u003eLi N, He Y, Zhang S, et al. Circulating tumor DNA-based minimal residual disease detection predicts recurrence in intrahepatic cholangiocarcinoma[J]. Nat Med, 2023, 29(4): 789-798.\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":"Intrahepatic cholangiocarcinoma (ICC), multicenter retrospective study, adjuvant therapy, Deep Learning, Transformer-MLP Model, SHapley Additive exPlanations (SHAP)","lastPublishedDoi":"10.21203/rs.3.rs-9513815/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9513815/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Intrahepatic cholangiocarcinoma (ICC) carries a dismal postoperative prognosis, with high recurrence rates and inconsistent adjuvant therapy benefits. Current prognostic models are limited by high cost, single‑modal data, or poor accessibility. We developed and externally validated a Transformer‑MLP deep learning model for 2‑year overall survival prediction in 499 resected ICC patients from 8 Chinese centers. A differentiated feature selection strategy was applied: 17 features for traditional machine learning and all 31 routine clinicopathological features for deep learning. The Transformer‑MLP model achieved the best performance in the external test set (AUC = 0.8131, C‑index = 0.7462), with favorable calibration and net clinical benefit. SHAP analysis highlighted adjuvant chemotherapy, AJCC stage, and lymph node metastasis as key predictors. A free online calculator (https://icc‑prediction.streamlit.app/) was built for clinical use. This accessible, interpretable model provides robust risk stratification to guide personalized adjuvant therapy, complementing existing omics‑ or radiomics‑based ICC prognostic tools.","manuscriptTitle":"A transformer-MLP model with SHAP analysis for predicting postoperative survival in Intrahepatic Cholangiocarcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-11 10:41:14","doi":"10.21203/rs.3.rs-9513815/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-30T20:36:26+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-30T20:32:08+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-30T11:51:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-30T02:52:15+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-04-30T02:21:58+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":"e852a220-fc08-4a66-a80d-f3e5d92b2f09","owner":[],"postedDate":"May 11th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewersInvited","content":"17","date":"2026-04-30T20:36:26+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-30T20:32:08+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-30T11:51:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-30T02:52:15+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-04-30T02:21:58+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67727527,"name":"Health sciences/Biomarkers"},{"id":67727528,"name":"Biological sciences/Cancer"},{"id":67727529,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":67727530,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-05-11T10:41:18+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-11 10:41:14","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9513815","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9513815","identity":"rs-9513815","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}