A multimodal deep learning model for early prediction of pancreatic cancer progression.

e16368 Background: Early and accurate prediction of pancreatic cancer progression is critical for optimizing treatment regimens and improving patient prognosis, particularly in patients who do not exhibit significant remission during chemotherapy. This study aimed to develop a deep learning-based longitudinal disease monitoring model using multimodal data from routine clinical practice to predict whether pancreatic cancer patients will experience disease progression at their next follow-up during chemotherapy. Methods: We developed a multimodal deep learning model to predict disease progression using baseline clinical information and longitudinal CT imaging data from follow-ups. The study cohort included patients with advanced pancreatic ductal adenocarcinoma receiving first-line chemotherapy from two tertiary medical centers. Data from Sun Yat-sen University Cancer Center between 2011 and 2023 were used to construct the training and internal test sets. Specifically, the training set included patients with stable disease (SD) at the first follow-up (243 patients, with follow-up visits ranging from 2 to 8), while the internal test set included patients with partial response (PR) at the first follow-up (41 patients, with follow-up visits ranging from 2 to 4). Additionally, SD and PR cases from the Sun Yat-sen Memorial Hospital between 2015 and 2023 were collected as an external test set (23 patients, with follow-up visits ranging from 2 to 3). The optimal model was selected using five-fold cross-validation on the training set, followed by performance testing on the internal and external test sets. Results: When using only the learned image features, the model achieved an average AUC of 0.68 in five-fold cross-validation on the training set. However, by combining longitudinal imaging data and clinical information, the model's performance on the training set significantly improved, with an average AUC of 0.78, accuracy of 0.78, sensitivity of 0.72, and specificity of 0.81. In the internal test set, the model achieved an average AUC of 0.74, accuracy of 0.72, sensitivity of 0.76, and specificity of 0.71. In the external test set, the performance metrics were AUC 0.71, accuracy 0.75, sensitivity 0.62, and specificity 0.83. Conclusions: This study demonstrates that the proposed deep learning model integrating baseline clinical information and longitudinal CT imaging data can dynamically monitor disease progression in pancreatic cancer patients during chemotherapy. This approach has great potential to assist clinicians in early predicting chemotherapy outcomes and enabling timely interventions to improve patient prognosis .