Prediction of Postoperative ICU Stay in Oral Cancer Using Dynamic Ensemble of LLMs & Small ML Models

Sunday, March 22, 2026

12:15 PM - 1:15 PM Central Time

Location: Connections Central - RST 03

Junpeng Tang, N/A (he/him/his)

CHN

Disclosure(s): No relevant financial relationship(s) to disclose.
SL

Siqi Liu, N/A

Doctor
Sun Yat-sen Memorial Hospital，Sun Yat-sen University

Disclosure information not submitted.
JZ

Jinzhao Zhang, n/a

Doctor
Sun Yat-sen Memorial Hospital，Sun Yat-sen University

Disclosure information not submitted.
QY

Qiyu Yang, N/A

Professor
School of Automation, Guangdong University of Technology

Disclosure information not submitted.
MW

Mike Wang, N/A

Master
School of Automation, Guangdong University of Technology

Disclosure information not submitted.
ZH

Zhengwei Huang, N/A

Master
School of Automation, Guangdong University of Technology

Disclosure information not submitted.
JT

Jiaheng Tian, N/A

Master
School of Automation, Guangdong University of Technology

Disclosure information not submitted.
JL

Jun Lu, N/A

Professor
School of Automation, Guangdong University of Technology

Disclosure information not submitted.
ZY

Zhengfei Yang, n/a

Professor
Sun Yat-sen Memorial Hospital，Sun Yat-sen University

Disclosure information not submitted.

Introduction: CU length of stay (LOS) prediction faces challenges including heterogeneity in multimodal data and static parameter constraints in conventional artificial intelligence (AI) ensemble models.This study aims to develop EPPO (Ensemble Prediction System Integrating Large Language Models for Postoperative ICU Length-of-Stay in Oral Cancer Patients) – a dynamic fusion framework that synergizes large language models (LLMs) with small machine learning models to enhances both prediction accuracy and clinical utility for postoperative ICU LOS in oral cancer patients.

Methods: Using multimodal data from 412 oral cancer patients at Sun Yat-sen Memorial Hospital, EPPO combined unstructured clinical narratives (processed by GPT-4o, Gemini and Claude) and structured variables (e.g., vital signs and laboratory results).The framework employed adaptive weighting of predictions from LLMs and AMFormer (Adversarial Mining Transformer). Key mechanisms include: (i) Gaussian distance normalization to quantify domain relevance between training and testing logits; (ii) entropy-based confidence calibration; (iii) dynamic fusion of textual and tabular predictions using distribution shift compensation. Performance was evaluated via the Area Under the Receiver Operating Characteristic Curve (AUROCC) and Area Under the Precision-Recall Curve (AUPRC) across 24-, 48- and 72-hour ICU discharge thresholds.

Results: EPPO achieved superior predictive performance with mean AUROCC values of 0.617 (24 h), 0.707 (48 h), and 0.769 (72 h), which exceeded those of standalone LLMs and small machine learning models (XGBoost, logistic regression, and AMFormer). The fused model reduced overfitting risks in small cohorts (n=412) and improved precision–recall metrics versus single-model approaches. Calibration curves showed stable alignment with observed outcomes.

Conclusions: EPPO established a clinically robust paradigm for multimodal fusion in ICU prediction tasks and addressed critical limitations of small sample sizes and data heterogeneity. By dynamically balancing LLM-derived semantic insights and structured physiological patterns, the framework enables earlier risk stratification, potentially reducing ICU bed occupancy through optimized discharge planning.