Learnijoy
ResearchAI ResearchAI Engineering & DevTools

LLM Embeddings Improve Multimodal ICD-10 Prediction

Chengyuan Liu, Xinyue Zhang, Yao Li, Guanting Chen· June 30, 2026 View original

Summary

Researchers demonstrated that frozen medical LLM representations can serve as a shared embedding space for multimodal primary diagnosis category prediction, outperforming baselines by integrating both clinical narratives and structured EHR data. This approach allows efficient reuse of clinical representations across modalities and datasets.

This study explores the use of frozen medical large language model (LLM) representations as a unified embedding space for predicting primary ICD-10 diagnosis categories. The objective was to overcome the limitations of existing automated coding systems that struggle to integrate diagnostic signals from both clinical narratives and structured electronic health record (EHR) variables. Researchers constructed a cohort from MIMIC-IV, focusing on the ten most frequent primary ICD-10 codes, consolidated into seven categories. They serialized structured variables into narratives and combined them with discharge notes. Using a frozen MedFound-Llama3-8B-finetuned backbone, hidden states were extracted from transformer layers to train linear probes for structured-only, unstructured-only, and combined inputs. The combined probe achieved the best performance on MIMIC-IV, significantly outperforming single-modality probes and established baselines. The study also found that diagnostic information became more linearly separable in deeper layers and that a compact adapter enabled effective cross-dataset transfer to MIMIC-III. This method offers a practical way to unify and adapt clinical representations across different data types and datasets.

Why it matters

This advancement can significantly improve the accuracy and efficiency of automated medical coding, leading to better reimbursement, more reliable research data, and enhanced population health surveillance.

How to implement this in your domain

  1. 1Evaluate integrating LLM-based multimodal embedding techniques into existing clinical coding systems.
  2. 2Pilot the use of this approach for automated primary diagnosis prediction in a specific healthcare setting.
  3. 3Collaborate with AI researchers to adapt and fine-tune medical LLMs for specific institutional EHR data.

Who benefits

HealthcareHealth InsuranceMedical ResearchPharmaceuticals

Key takeaways

  • Frozen medical LLM embeddings can unify structured and narrative EHR data.
  • Multimodal probing significantly improves primary diagnosis prediction accuracy.
  • Diagnostic information becomes more separable in deeper LLM layers.
  • This approach enables efficient transfer of clinical representations across datasets.

Original post by Chengyuan Liu, Xinyue Zhang, Yao Li, Guanting Chen

"arXiv:2606.28798v1 Announce Type: new Abstract: Objective: ICD codes are central to reimbursement, research, and population health surveillance, yet automated coding systems often struggle to integrate diagnostic signals from both clinical narratives and structured electronic hea…"

View on X

Originally posted by Chengyuan Liu, Xinyue Zhang, Yao Li, Guanting Chen on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

BaRA Improves LoRA Fine-Tuning with Adaptive Rank Allocation

Researchers introduce BaRA, a Bayesian Adaptive Rank Allocation framework for parameter-efficient fine-tuning, which dynamically adjusts adaptation capacity based on context. This method enhances predictive performance, robustness, and uncertainty calibration compared to standard LoRA and other Bayesian LoRA variants.

Zhibin Duan, Yuhong Wang, Jiahong Fu, Zongsheng Yue, Bo Chen, Zongben XuJun 30, 2026
AI ResearchAI Engineering & DevTools

New Preconditioner Improves Deep Network Training Stability and Performance

Researchers introduce Dead-Direction Conditioners (DDC), a novel preconditioning method that leverages gauge-equivariant optimization to prevent deep network training from drifting along symmetry orbits. This technique improves model stability, reduces overfitting, and enhances performance in language and vision models.

Tejas Pradeep ShirodkarJun 30, 2026
AI ResearchAI Engineering & DevTools

SMDA Traces Training Data Influence on LLM Behavioral Policies

Researchers introduce Symbolic Mechanistic Data Attribution (SMDA), a framework that attributes specific training examples to the interpretable symbolic policies governing an LLM's high-level behavior. SMDA offers a fine-grained diagnostic tool to understand how training data shapes model decisions, revealing safety gaps and unintended influences.

Reza Habibi, Darian Lee, Magy Seif El-NasrJun 30, 2026