結合 LLM 與 LSTM 之時序缺失值填補混合架構於 MIMIC 資料庫之敗血症早期預測

Abstract

敗血症為加護病房中死亡率極高之急重症，早期預測對提升病患存活率具有關鍵影響。ICU 臨床時間序列資料普遍存在高缺失率與非隨機缺失（MNAR）問題，現有缺失值填補方法多依賴數值時序建模，難以整合臨床語義知識，使得缺失機制本身所隱含之臨床資訊無法被有效利用。 本研究提出結合長短期記憶網路（LSTM）與大型語言模型（Llama 3 8B）之混合式架構，採用雙分支設計進行缺失值填補與敗血症早期預測。第一階段訓練缺失值填補模型，LSTM 數值分支負責建模時間序列動態，LLM 語義分支透過模板式文字轉換捕捉臨床語義知識，並以門控融合機制整合兩者之互補資訊。第二階段則以融合表示作為輸入，訓練下游分類器進行敗血症早期預測。 基於 MIMIC-IV 資料庫之實驗結果顯示，本研究方法在多數實驗設定下，於填補與分類效能上皆優於 7 種基準方法，且優勢隨預測時距增加而擴大。在發病前 11 小時之早期預測情境下，AUROC 達 0.801，較次佳方法 SAITS（0.771）高出 0.030，顯示本架構在資料極度稀疏的挑戰性情境下，仍具備穩健的早期預警潛力。本研究將語義資訊引入 ICU 時序缺失值填補，不僅能提升填補品質，亦能改善下游預測效能，為未來結合大型語言模型與臨床時序資料之研究提供研究基礎。

Sepsis is a life-threatening condition in the intensive care unit (ICU) with extremely high mortality rates, and early prediction plays a critical role in improving patient survival. ICU clinical time-series data are characterized by high missing rates and Missing Not at Random (MNAR) mechanisms. Existing imputation methods predominantly rely on numerical time-series modeling, limiting their ability to incorporate clinical semantic knowledge and effectively leverage the clinical information embedded in missingness patterns. This study proposes a hybrid framework combining Long Short-Term Memory (LSTM) and a large language model (Llama 3 8B), adopting a dual-branch design for missing value imputation and early sepsis prediction. In the first stage, an imputation model is trained with a dual-branch design, where the LSTM numerical branch models temporal dynamics and the LLM semantic branch captures clinical semantic knowledge through template-based text serialization. A gated fusion mechanism integrates the complementary information from both branches. In the second stage, the fused representation serves as input to train a downstream classifier for early sepsis prediction. Experimental results on the MIMIC-IV database demonstrate that the proposed method outperforms seven baseline methods in most experimental settings for both imputation performance and classification performance, with the performance advantage increasing as the prediction horizon extends. At 11 hours before sepsis onset, the proposed framework achieves an AUROC of 0.801, surpassing the second-best method SAITS (0.771) by 0.030, demonstrating robust early warning capability even under severely sparse data conditions. This study incorporates semantic information into ICU time-series missing value imputation, demonstrating that it not only improves imputation quality but also enhances downstream prediction performance, providing a reference foundation for future research combining large language models with clinical time-series data.