Multimodal Pre-operative Feature Fusion for Robust Post-Surgical Complication Prediction in Lung Cancer

Shubham Pandey, Srirangaraj Setlur, Venu Govindaraju, Kenneth Seastedt

Keywords: Multimodal deep learning, Feature Fusion, Computed Tomography (CT), Translational AI, Trustworthy AI

Abstract: Postoperative complications remain a major concern in lung cancer surgery, often leading to increased morbidity, extended hospital stays and often mortality. Existing risk-prediction methods rely on either clinical variables or radiological assessment independently, which limits their sensitivity and generalizability. To address this, we propose TriFuse, a novel multimodal deep learning architecture that integrates pre-operative thoracic CT scans and structured clinical data to predict probability of postoperative complications following lung cancer surgery. In TriFuse, CT scans are encoded via a Vision backbone that extracts volumetric imaging features and structured clinical variables are processed via a dedicated MLP to produce clinical embeddings. Additionally, a separate branch leverages a large language model to generate natural-language remarks summarizing salient patient-level risk factors from the clinical data and maps these into the shared embedding space. A dynamic feature-fusion module then combines embeddings from these branches into a unified patient representation capturing complementary information from imaging, structured data and semantic summaries. A final classifier predicts probability of postoperative complication risk. We evaluated TriFuse on a curated dataset of 3279 lung-cancer patients, who underwent surgery at Roswell Park Comprehensive Cancer Center, Buffalo, NY (with a train/validation/test split of 2721, 279, 279 respectively).

Primary Subject Area: Integration of Imaging and Clinical Data

Secondary Subject Area: Interpretability and Explainable AI

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 349