Construction of feature selection and efficacy prediction model for transformation therapy of locally advanced pancreatic cancer based on CT, F-FDG PET/CT, DNA mutation, and CA199.

BACKGROUND

Immunotherapy and radiotherapy play crucial roles in the transformation therapy of locally advanced pancreatic cancer; however, the exploration of effective predictive biomarkers has been unsatisfactory. With the rapid development of radiomics, next-generation sequencing, and machine learning, there is hope to identify biomarkers that can predict the efficacy of transformative treatment for locally advanced pancreatic cancer through simple and non-invasive clinical methods. Our study focuses on using computed tomography (CT), positron emission tomography/computed tomography (PET/CT), gene mutations, and baseline carbohydrate antigen 199 (CA199) to identify biomarkers for predicting the efficacy of transformative treatment.

METHODS

We retrospectively collected data from 70 patients with locally advanced pancreatic cancer who had undergone a biopsy for pathological diagnosis. These patients had complete baseline enhanced CT images and baseline CA199 results. Among them, 65 patients had efficacy evaluation results after 4 treatment cycles, 54 patients had complete baseline PET/CT images, 51 patients had complete DNA mutation detection results, and 34 patients had both complete PET/CT images and DNA mutation detection results. Additionally, 47 patients had complete available CT images at baseline, after 2 treatment cycles, and after 4 treatment cycles. We extracted radiomic features from the original lesion-enhanced CT images (including baseline and subsequent follow-up CT scans), radiomic features from baseline 18F-fluoro-2-deoxy-2-D-glucose (F-FDG) PET, and patient-specific features related to abdominal and visceral fat. We used short-term and long-term treatment efficacy as the prediction outcomes and performed statistical and machine learning-based feature selection and COX regression analysis to identify potentially predictive features. Subsequently, we separately or in combination modeled the CT features, PET features, baseline CA199, and gene mutation data to construct efficacy prediction models. Finally, we investigated the mixed effects model of the dynamic changes in CT features at baseline, after 2 treatment cycles, and after 4 treatment cycles on the prediction of short-term treatment efficacy.

RESULTS

We found that a combination of CT radiomic features, including F1_ gray level co-occurrence matrix (GLCM), F2_gray level run length matrix (GLRLM), F5_neighboring gray tone difference matrix (NGTDM), and F6_Shape, PET radiomic features such as visceral adipose tissue (VAT), tumor-to-liver ratio (T/L), standardized uptake value mean (SUVmean), and GLCM, as well as baseline CA199, can be used to predict short-term treatment efficacy. Baseline CA199, GLCM, IntensityDirect, Shape, and PET/CT features are independent factors for long-term treatment efficacy. In constructing the short-term treatment efficacy prediction model, ensemble learning methods such as adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and RandomForest performed the best. However, in terms of model interpretability, decision tree methods provide the most intuitive display of the predictive details of the model. For the time series data of patients' baseline CT, CT after 2 treatment cycles, and CT after 4 treatment cycles, long short-term memory (LSTM) modeling yielded better predictive models.

CONCLUSION

A multimodal combination of radiomics, DNA mutations, and baseline CA199 can predict the efficacy of transformative treatment in locally advanced pancreatic cancer. Various feature selection methods and multimodal fusion approaches contribute to guiding personalized and precise treatment for pancreatic cancer.