Development and Validation of a Deep Learning Radiomics Model Based on Ultrasound and Clinical Features to Predict Prognosis in Elderly Patients with Advanced Pancreatic Cancer after HIFU Therapy

Background This study aimed to construct an artificial intelligence model based on ultrasound radiomics and deep learning, integrating clinical features to develop a fusion model for individualized prediction of survival outcomes in elderly patients with advanced pancreatic cancer receiving high-intensity focused ultrasound (HIFU) treatment. Methods This retrospective study enrolled elderly patients with advanced pancreatic cancer admitted to Huadong Hospital Affiliated to Fudan University from March 2015 to March 2024, randomly divided into training and validation cohorts in a 7:3 ratio. Patients were categorized into four groups based on treatment modality: HIFU alone, HIFU combined with 125I seed implantation, HIFU combined with chemotherapy, and triple therapy (HIFU +  ¹²⁵ I + chemotherapy). Traditional radiomics features and deep learning features based on the ResNet architecture were extracted from pre-treatment ultrasound images. After rigorous feature selection, clinical, radiomics, deep learning, and multi-modal fusion Cox proportional hazards models were constructed. Predictive performance for overall survival and clinical utility were evaluated comprehensively. Results This study included 250 elderly pancreatic cancer patients. Multivariate analysis identified liver metastasis, TNM stage, body weight, number of HIFU sessions, and treatment regimen as independent prognostic factors (all p  < 0.05). Predictive models were constructed using selected clinical features, traditional radiomics and deep learning features from ultrasound images. The deep learning radiomics model demonstrated the highest C-indices in both the training and validation sets (0.735 and 0.716, respectively), outperforming the clinical model (0.664, 0.633) and the traditional radiomics model (0.658, 0.592). The combined model, integrating clinical and deep learning features, achieved the best predictive performance, with C-indices of 0.748 (training) and 0.739 (validation). Time-dependent ROC analysis further confirmed that the combined model maintained the highest AUC values for 1-year and 1.5-year survival prediction in the validation set (0.819 and 0.884, respectively), significantly enhancing the accuracy and generalizability of survival stratification. Conclusions The ultrasound-based deep learning radiomics model demonstrated favorable performance in predicting the prognosis of elderly pancreatic cancer patients undergoing HIFU treatment, with performance superior to traditional clinical and radiomics indicators. It can facilitate more accurate individualized survival risk stratification, providing a potentially useful tool for precise treatment decision-making in advanced pancreatic cancer.