Robust Multimodal Fusion for Survival Prediction in Cancer Patients

4181 Background: Pancreatic cancer is an aggressive malignancy with limited therapeutic options and a poor prognosis. Current approaches to prognostication are limited, especially in advanced disease. We explored whether machine learning integrating multi-modal data could predict outcomes in advanced pancreatic cancer. Methods: We developed and evaluated machine learning models predicting disease control rate and one-year survival from the COMPASS trial (NCT02750657). Data modalities included clinical features, histopathology, radiology, RNAseq, and whole-genome sequencing (WGS). After pre-processing, we applied LASSO and XGBoost to each modality and early and late fusion techniques. Hyperparameter tuning and performance assessment were performed using repeated nested cross-validation. The PurIST RNAseq classifier served as a baseline. Area under the curve (AUC) was the primary metric. Results: The cohort included 260 patients (105 female; median age 64 [IQR 58–70]; 141 treated with FOLFIRINOX, 97 with gemcitabine and nab-paclitaxel). 170 (65%) achieved disease control and 168 (65%) survived at least one year. The performance of the machine learning models is shown in the Table. Predictions from the unimodal models had limited correlation with each other (the maximum pairwise correlation averaged across folds was between clinical and histopathology models, 0.21). The late fusion models up-weighted data modalities with stronger unimodal performance. Conclusions: Multiple individual data modalities can predict outcomes in advanced pancreatic cancer, with PurIST serving as a strong baseline. Despite differing predictions across data modalities, multimodal integration did not improve prognostic performance in this cohort. AUC for the PurIST baseline, the top 2 unimodal models, and the best fusion model for each outcome. Outcome Data Modality AUC (95% confidence interval) Disease control PurIST 0.69 (0.69, 0.70) Radiomics 0.75 (0.72, 0.79) RNAseq 0.71 (0.70, 0.72) Fusion (late) 0.71 (0.69, 0.73) One-year survival PurIST 0.63 (0.62, 0.63) DNA mutations 0.64 (0.61, 0.66) RNAseq 0.57 (0.55, 0.60) Fusion (early) 0.61 (0.56, 0.66)

BackgroundTyrosine kinase inhibitors (TKIs) targeting epidermal growth factor receptor (EGFR) are effective first-line treatments for advanced non-small-cell lung cancer (NSCLC) patients with EGFR mutations. However, some patients do not respond well, and some experience rapid progression despite initial benefit. This study aims to develop multimodal models integrating pre-treatment histopathological images and clinical variables to predict EGFR-TKIs therapy response and progression-free survival (PFS) in stage IV NSCLC patients.MethodsThis retrospective study collected hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) and clinical variables, including clinical characteristics and laboratory indicators, from 247 treatment-naïve, stage IV EGFR-mutated NSCLC patients treated with EGFR-TKIs at Beijing Chest Hospital between 2017 and 2020. For the prediction of treatment response (TR) and progression-free survival (PFS) time, both unimodal and multimodal models were developed. Unimodal models were based on whole-slide images (WSIs) or clinical variables alone, while multimodal models integrated both data types. Model performance was evaluated using distinct metrics, including area under the curve (AUC), accuracy, sensitivity, and specificity for TR prediction, and mean absolute error (MAE), mean squared error (MSE), and Kaplan–Meier analysis for PFS time estimation.ResultsThe multimodal model for predicting therapy response (TR-M) achieved AUC, accuracy, sensitivity and specificity of 0.943, 0.841, 0.800 and 1.000 in test set, respectively, significantly outperforming the two unimodal models. The multimodal PFS model (PFS-M) directly estimated PFS time after EGFR-TKIs therapy with a MAE of 2.690 months and a MSE of 12·829 in the test set, significantly outperforming unimodal models and effectively stratifying patients into high and low risk groups (HR: 10.034, 95% CI 3.879–25.956; p < 0.0001). Furthermore, our models revealed that an increased aggregation of lymphocytes within the patient tissues was indicative of a more favorable therapy response, whereas a higher percentage of lymphocytes in the peripheral blood was associated with prolonged PFS.ConclusionThe multimodal models (TR-M and PFS-M) demonstrated high accuracy in predicting the therapy response and directly estimating of PFS following EGFR-TKIs treatment in stage IV NSCLC patients, revealing significant potential for aiding clinical decisions and facilitating personalized treatment strategies.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12967-025-06956-8.

Multi-modal deep learning for automated assembly of periapical radiographs

Introduction: Accurate prediction of the risk of ischemic stroke (IS) is vital for prevention and would be aided by multimodal biomarkers integrating genetic, clinical, and functional data. The role of imaging and EKG based atrial measurements, other than atrial fibrillation (AF), in IS prediction is uncertain and many strokes remain cryptogenic despite extensive work-up. As an exploratory step to improve stroke evaluation by including atrial traits, we developed a novel multimodal deep learning model integrating demographic and clinical variables with atrial phenotypic and genotypic data. Methods: We collected individuals from UK Biobank (UKBB) and defined ischemic stroke (IS) by the UKBB Algorithmically Defined Outcome (ADO). We developed a multimodal multi-layer perceptron with late fusion (MMLP-LF) model to predict whether a subject has IS by integrating five data modalities from UKBB: 1) MRI and EKG derived atrial traits, 2) lead genetic variants (P&lt;5e-8) from GWAS of atrial traits, 3) patient demographics, 4) ICD codes including AF but excluding IS codes and 5) procedure codes related to cardiac valve surgery. We split the samples into 64%-16%-20% train-validation-test sets via stratified random sampling. Models were trained on the training set with 20 rounds of random initialization. The validation set was used for model selection. We used Area Under the Receiver Operating Characteristic Curve (AUROC) for performance evaluation. Shapley additive explanation (SHAP) analysis was conducted for feature importance attribution. Results: Our dataset included 24,582 individuals from the UKBB, among which 100 (0.41 %) had IS. Our full MMLP-LF model including 5 modalities achieved the highest AUROC of 0.85 for predicting IS on the hold-out test set, substantially outperforming the best unimodal branch (AUROC 0.79 with ICD codes, 0.71 with demographic variables). If we excluded atrial phenotypes, the AUROC was still 0.85. If genetic traits were also removed, the AUROC declined to 0.82. SHAP analysis revealed that ICD codes for hypertension and hyperlipidemia are the most influential contributors to the full model followed by GWAS-identified lead genetic variants associated with atrial traits (Figure). Conclusion: Our MMLP-LF model improved IS prediction over unimodal models by integrating multimodal data and identified genetic and clinical drivers predicting IS. This model establishes a new paradigm for integrating multiple modalities to predict IS outcomes.

ABSTRACTBackground and PurposePrognostic stratification in non‐small cell lung cancer (NSCLC) presents considerable challenges due to tumor heterogeneity. Emerging evidence has proposed that adipose tissue may play a prognostic role in oncological outcomes. This study investigates the integration of deep learning (DL)–derived computed tomography (CT) imaging biomarkers with mediastinal adiposity metrics to develop a multimodal prognostic model for postoperative survival prediction in NSCLC patients.MethodsA retrospective cohort of 702 surgically resected NSCLC patients was analyzed. Tumor radiomic features were extracted using a DenseNet121 convolutional neural network architecture, while mediastinal fat area (MFA) was quantified through semiautomated segmentation using ImageJ software. A multimodal survival prediction model was developed through feature‐level fusion of DL‐extracted tumor characteristics and MFA measurements. Model performance was evaluated using Harrell's concordance index (C‐index) and receiver operating characteristic (ROC) analysis. Risk stratification was performed using an optimal threshold derived from training data, with subsequent Kaplan–Meier survival curve comparison between high‐ and low‐risk cohorts.ResultsThe DL‐based tumor model achieved C‐indices of 0.787 (95% CI: 0.742–0.832) for disease‐free survival (DFS) and 0.810 (95% CI: 0.768–0.852) for overall survival (OS) in internal validation. Integration of MFA with DL‐derived tumor features yielded a multimodal model demonstrating enhanced predictive performance, with C‐indices of 0.823 (OS) and 0.803 (DFS). Kaplan–Meier analysis revealed significant survival divergence between risk‐stratified groups (log‐rank p < 0.05).ConclusionThe multimodal fusion of DL‐extracted tumor radiomics and mediastinal adiposity metrics represents a significant advancement in postoperative survival prediction for NSCLC patients, demonstrating superior prognostic capability compared to unimodal approaches.

To investigate the value of a multimodal deep learning (MDL) model based on computed tomography (CT) and magnetic resonance imaging (MRI) for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC). A total of 287 patients with HCC from our institution and 58 patients from another individual institution were included. Among these, 119 patients with only CT data and 116 patients with only MRI data were selected for single-modality deep learning model development, after which select parameters were migrated for MDL model development with transfer learning (TL). In addition, 110 patients with simultaneous CT and MRI data were divided into a training cohort (n = 66) and a validation cohort (n = 44). We input the features extracted from DenseNet121 into an extreme learning machine (ELM) classifier to construct a classification model. The area under the curve (AUC) of the MDL model was 0.844, which was superior to that of the single-phase CT (AUC = 0.706-0.776, P < 0.05), single-sequence MRI (AUC = 0.706-0.717, P < 0.05), single-modality DL model (AUCall-phase CT = 0.722, AUCall-sequence MRI = 0.731; P < 0.05), clinical (AUC = 0.648, P < 0.05), but not to that of the delay phase (DP) and in-phase (IP) MRI and portal venous phase (PVP) CT models. The MDL model achieved better performance than models described above (P < 0.05). When combined with clinical features, the AUC of the MDL model increased from 0.844 to 0.871. A nomogram, combining deep learning signatures (DLS) and clinical indicators for MDL models, demonstrated a greater overall net gain than the MDL models (P < 0.05). The MDL model is a valuable noninvasive technique for preoperatively predicting MVI in HCC.

In general, water level prediction models using deep learning techniques have been developed using time-series water level observation data from upstream water level stations and target water level stations even though many of physical data are necessary to predict water level. The changes of the water level are greatly affected by rainfall in the basin, therefore rainfall information is needed to more accurately predict the water level. In particular, radar data has the advantage of being able to directly acquire the amount of rainfall occurring within a watershed. This study aims to develop the multimodal deep learning model to predict the water level using 2D grid radar rainfall data and 1D time-series water level observation data. This study proposed two multimodal deep learning models which have different structures. Both multimodal deep learning models predict the water level by simultaneously using the observed water level data up to the present time and the radar rainfall data that affects the water level in the future. The first proposed model consists of a deep learning network that links 2D Average Pooling (AvgPool2D), which compresses 2D radar data to 1D data, and Long Short-Term Memory (LSTM), which predicts 1D time series water level data. The second proposed model consists of a deep learning network that predicts water levels by linking Conv2DLSTM and LSTM, which can reflect the characteristics of 2D radar data without deformation.&amp;#160; The two proposed multimodal deep learning models were learned and evaluated in the upper basin of Hantan River. In addition, it was compared with the results of single-modal LSTM using only water level data. There are three water level stations in the study area, and the objective was to predict the water level of the downstream station up to 180 minutes in advance. For learning and verification of the deep learning model, 10-minute water level and radar rainfall data were collected from May 2019 to October 2021. For the radar data used as input, the grid data included in the target watershed were extracted and used among composite radar data with a resolution of 1 km operating by Ministry of Environment. As a result of evaluating each learned deep learning model, two multimodal models had higher prediction accuracy than the single-modal using only water level data. In particular, second proposed model (Conv2dLSTM+LSTM) had better predictive performance than first proposed model (AvgPool2D+LSTM) at the time of the sudden rise in water level due to rainfall.AcknowledgmentsResearch for this paper was carried out under the KICT Research Program (project no. 202200175-001, Development of future-leading technologies solving water crisis against to water disasters affected by climate change) funded by the Ministry of Science and ICT.

One of the main challenges in CBIR systems is to choose discriminative and compact features, among dozens, to represent the images under comparison. Over the years, a great effort has been made to combine multiple features, mainly using early, late, and hierarchical fusion techniques. Unveiling the perfect combination of features is highly domain-specific and dependent on the type of image. Thus, the process of designing a CBIR system for new datasets or domains involves a huge experimentation overhead, leading to multiple fine-tuned CBIR systems. It would be desirable to dynamically find the best combination of CBIR systems without needing to go through such extensive experimentation and without requiring previous domain knowledge. In this paper, we propose ExpertosLF, a model-agnostic interpretable late fusion technique based on online learning with expert advice, which dynamically combines CBIR systems without knowing a priori which ones are the best for a given domain. At each query, ExpertosLF takes advantage of user’s feedback to determine each CBIR contribution in the ensemble for the following queries. ExpertosLF produces an interpretable ensemble that is independent of the dataset and domain. Moreover, ExpertosLF is designed to be modular, and scalable. Experiments on 13 benchmark datasets from the Biomedical, Real, and Sketch domains revealed that: (i) ExpertosLF surpasses the performance of state of the art late-fusion techniques; (ii) it successfully and quickly converges to the performance of the best CBIR sets across domains without any previous domain knowledge (in most cases, fewer than 25 queries need to receive human feedback).

In advertising, identifying the content safety of web pages is a significant concern since advertisers do not want brands to be associated with threatening content. At the same time, publishers would like to maximize the number of web pages on which they can place ads. Thus, a fine balance must be achieved while classifying content safety in order to satisfy both advertisers and publishers. In this paper, we propose a multimodal machine learning framework that fuses visual and textual information from web pages to improve current predictions of content safety. The primary focus is on late fusion, which involves combining final model outputs of separate modalities, such as images and text, to arrive at a single decision. This paper presents a fully automated machine learning framework that performs binary and multilabel classification using late fusion techniques. We also introduce additional work in early fusion, which involves extracting and fusing intermediate features from the two separate models. Our algorithms are applied to data extracted from relevant web pages in the advertising industry. Both of our late and early fusion methods obtain significant improvements over algorithms currently in use.

Multimodal deep learning for predicting the choice of cut parameters in the milling process

Improving cancer patients Overall Survival (OS) prognosis is critical for personalization of treatment using model-identified drivers of cancer progression. Current cancer prognosis models largely rely on clinical and demographic patient characteristics. Adding ‘omics’-based modalities can help improve patient OS prediction and lead to better disease categorization and understanding. We introduce a data driven methodology for combining multi-omics and clinical data, including clinical/demographics, mutations, gene expression, long non-coding and micro-RNA expression, DNA methylation, and proteomics for improving prediction of OS in cancer patients. High dimensionality of ‘omics’ modalities present challenges to combining them into one model. We propose a late stage modalities fusion where we construct a separate data driven model for OS prediction for each modality, later combining individual predictions in a final linear OS prediction model. With a limited number of patients to develop the model, such an approach helps to better protect against overfitting, and allows to account for different degrees of informativeness of modalities by weighting them according to individual success. We introduce a robust machine learning pipeline with rigorous training, testing and evaluation capabilities, and demonstrate its effectiveness on a suit of TCGA data. When comparing early vs. late fusion of omics and clinical modalities for survival prediction using NSCLC TCGA data, we observe the C-index improvement from 0.57±.04 to 0.61±.01. Best individual modality performance was at 0.59±.02 using clinical modality. Dominant modalities in unimodal survival analysis varied between cancers, with clinical, RNA, and miRNA for LUAD, and clinical, RNA, and RPPA for LUSC. Using pan-cancer TCGA data for survival prediction, the best C-index = 0.77 was achieved using multi-omics model, followed by 0.76±.01 for clinical, 0.75±.01 for RNA seq, and 0.73±.01 for RPPA unimodal models. Citation Format: Nikos Nikolaou, Domingo Salazar, Harish RaviPrakash, Miguel Goncalves, Gustavo Alonso Arango Argoty, Nikolay Burlutsky, Natasha Markuzon, Etai Jacob. Improving survival prediction using flexible late fusion machine learning framework for multi-omics data integration. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5395.

Tissue resident memory T cells (TRM) are a specialized subset of long-lived memory T cells that reside in peripheral tissues. However, whether TRM exerts any immunosurveillance role in the tumor immune microenvironment (TIME) and progression of non-small-cell lung cancer (NSCLC), which accounts for 85% of all lung cancers, remains unclear. Our comprehensive analysis of multiple independent single-cell and bulk RNA-seq datasets of patient NSCLC samples generated reliable, unique TRM signatures, through which we could infer the abundance of TRM in NSCLC. We discovered that TRM abundance is consistently positively correlated with CD4+ T helper 1 cells, M1 macrophages, and resting dendritic cells in TIME and significantly impacts the prognosis of NSCLC patients. In addition, TRM signatures are strongly associated with immune checkpoint genes and the prognosis of NSCLC patients, suggesting that TRM signatures are promising prognostic markers for immunotherapy in NSCLC. We then built a machine learning model to predict patient survival based on the TRM signatures and immune related genes. The accuracy of the model was validated by Kaplan-Meyer survival analysis, receiver operating characteristic curves, principal component analysis, and t-distributed random neighbor embedding. We developed a 4-gene risk score that effectively stratified patients into low-risk and high-risk categories. The patients with high-risk scores had significantly lower overall survival than patients with low-risk. The prognostic value of the risk score was independently validated by the Cancer Genome Atlas Program (TCGA) dataset and multiple independent NSCLC patient datasets. Notably, low-risk NSCLC patients with higher TRM infiltration exhibited enhanced T-cell activation, macrophage regulation, and other TIME immune responses related pathways, indicating a more active immune profile benefitting from immunotherapy. Altogether, this study provides valuable insights into the complex interactions between NSCLC TRM and TIME and their impact on patient prognosis, highlighting the importance of TRM in shaping the NSCLC microenvironment. The development of a simplified 4-gene risk score provides a practical prognostic marker for risk stratification. Keywords: Tissue resident memory T cell, non-small-cell lung cancer, prognosis, tumor immune microenvironment, machine learning Citation Format: Aidan Shen, Aliesha Garrett, Junhua Mai, Yangzhi Zhu, Chongming Jiang. Tissue resident memory T cell abundances impact non-small-cell lung cancer immune microenvironment and patient prognosis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7323.

Background. Lung cancer continues to be a significant health concern globally. Due to the heterogeneity of the disease, using innovative strategies for effective management and treatment of patients is extremely important. Purpose – to characterize the mutational profile of a group of non-small cell lung cancer (NSCLC) patients utilizing a next-generation sequencing technique. Materials and Methods. A total of 42 samples that were fixed in formalin and embedded in paraffin (FFPE) were collected from 42 Ukrainian patients diagnosed with lung cancer who had surgery at the Sumy Regional Clinical Oncology Center. DNA was extracted from FFPE samples using the Omega Bio-tek E.Z.N.A.® FFPE DNA Kit (USA) following the manufacturerʼs instructions. Sequencing was performed on the Illumina NextSeq 550Dx platform (USA) using the Illumina NextSeq 550 Mid-Output Kit. The Cancer Genome Atlas Program (TCGA) database (https://portal.gdc.cancer.gov/) was used for a comparative analysis of the prevalence of genomic mutations in a cohort of Ukrainian and Caucasian patients with NSCLC. Statistical analysis was performed using Stata V.18.0 software (StataCorp, Texas, USA; https://www.stata.com; 2024). The paper belongs to the «description of case series» category which is a type of study recognized by evidence based medicine and does not claim statistical significance of the result. Results. Among the 42 NSCLC samples, 11 (26.19%) carried driver mutations such as EGFR (n=2; L858R), KRAS (n=7; G12C, G12D, G12A and A146S), BRAF (n=1; V600E) and translocation EML4(exon6) – ALK (exon20) (n=1; chr2:42503838 – chr2:29447579). All mutations were mutually exclusive. No NRAS, ROS1, RET, MET, ERBB2, and PIK3CA mutation cases were detected. The number of driver mutations in patients who had never smoked was significantly higher than in former or current smokers (p=0.046). No association was found between age, sex, tumor stage, histology of NSCLC, and driver mutations. Conclusions. Molecular genetic profiling using next-generation sequencing revealed driver mutations in 26.19% of patients with radically treated NSCLC. Most mutations are oncogenic and sensitive to tyrosine kinase inhibitors.

BackgroundThe diagnosis of non‐small cell lung cancer (NSCLC) is a clinical issue that requires attention, and more practical and effective biomarkers need to be selected to assist in diagnosis. This study aimed to examine the diagnostic value of serum albumin (ALB), lactate dehydrogenase (LDH), cytokeratin 19 fragments (CYFRA21‐1), and neuron‐specific enolase (NSE) for NSCLC.MethodsThe clinical data of 1048 NSCLC patients and 1125 healthy subjects were extracted from electronic medical records. Receiver operating characteristic (ROC) curve analysis was performed to assess the diagnostic significance of ALB, LDH, CYFRA21‐1, and NSE for NSCLC. The Cancer Genome Atlas (TCGA) data, which included mRNA profiles for ALB and LDH expression, were acquired from TCGA Program. Finally, interactive survival scatter plots and survival analyses for NSCLC patients were evaluated using the Human Protein Atlas and Kaplan–Meier Plotter.ResultsSignificant differences were noted in the levels of ALB and LDH between NSCLC patients and healthy controls. The areas under the ROC curves (AUCs) for ALB and LDH were 0.754 (95% CI: 0.734–0.774) and 0.681 (95% CI: 0.658–0.704), respectively. Moreover, the combination of ALB and LDH raised the AUC to 0.804 (95% CI: 0.785–0.823), and the incorporation of CYFRA21‐1 and NSE further increased the AUC to 0.903 (95% CI: 0.890–0.916). Notably, ALB and LDH might be related to the overall survival of NSCLC patients.ConclusionThis study revealed that ALB and LDH in NSCLC patient serum could improve the diagnostic accuracy of conventional biomarkers for NSCLC.

Pancreatic ductal adenocarcinoma (PDAC) remains a highly lethal disease with limited tools for predicting treatment response or survival. Our prior work, which applied machine learning to the COMPASS trial (NCT02750657), demonstrated that multimodal integration can enhance performance. Here, we evaluate model generalizability on an independent cohort from the PASS-01 clinical trial (NCT04469556). We developed predictive models using data from the COMPASS trial, incorporating clinical variables, histopathology image features, radiology-derived imaging features, RNA sequencing (RNA-seq), and whole-genome sequencing (WGS). Our updated pipeline applied TabPFN, a transformer-based model, using repeated 5-fold cross-validation. Fusion approaches included both early and late modality integration. We focused on predicting disease control rate (DCR). We compared model performance to the PurIST RNA-seq classifier, a strong baseline. The area under the receiver operating characteristic curve (AUC) was the primary metric. We externally validated the performance of models trained on COMPASS in the PASS-01 trial dataset. Among unimodal models, RNA-seq-based predictors achieved the highest AUC at 0.709 (95% CI: 0.595-0.820), significantly outperforming PurIST (p = 0.01). The performance of other unimodal models varied (clinical: 0.680; DNA: 0.527), with no significant difference compared to PurIST. The late fusion model, “MULTIPL”, integrated clinical, RNA, and DNA modalities and achieved the best overall performance at 0.733 (95% CI: 0.613-0.832), significantly outperforming PurIST (p = 0.002). The top 25th percentile of patients based on MULTIPL predicted DCR had significantly better prognosis (median overall survival 13.9 versus 8.6 months, hazard ratio 0.47 (95% CI: 0.28-0.78). The probability of DCR predicted by MULTIPL was correlated with the PurIST predictions of basal and classical transcriptomic subtypes (r = 0.63, p &amp;lt; 0.001), indicating a shared biology, which was further evidenced with SHapley Additive exPlanation interpretability analyses. Nonetheless, MULTIPL captured additional prognostic information, since PurIST was only modestly associated with DCR (AUC 0.55) and not significantly prognostic. Furthermore, the multimodal model was significantly associated with survival within the classical transcriptomic subtype. In conclusion, multimodal models trained on COMPASS data generalized to the PASS-01 trial in external validation. Late fusion of clinical, RNA, and DNA features achieved the best predictive performance for DCR and was also associated with survival outcomes, including within each transcriptomic subtype. In contrast to other models, which typically identify poor prognostic subgroups such as basal-like cancers, our multimodal model for DCR identifies a subset of patients with a more favourable prognosis. Together, these results demonstrate the potential of multimodal machine learning to improve prognostic modeling in advanced pancreatic cancer and guide personalized treatment strategies. Citation Format: Wei Quan, David Henault, Amy Zhang, Gun Ho Jang, Nicholas Light, Zongliang Ji, Anna Dodd, Julie Wilson, Daniel Renouf, Daniel Laheru, Kenneth Yu, Kimberly Perez, Amber Habowski, Grainne M. O'Kane, Steven Gallinger, David Tuveson, Elizabeth Jaffee, Jennifer J. Knox, Rahul G. Krishnan, Sandra Fischer, Masoom A. Haider, Faiyaz Notta, Robert C. Grant. External validation of a multimodal machine learning system to predict outcomes in advanced pancreatic cancer in the PASS-01 trial [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Pancreatic Cancer Research—Emerging Science Driving Transformative Solutions; Boston, MA; 2025 Sep 28-Oct 1; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl_3):Abstract nr B086.