Bridging the Version Gap: Multi-version Training Improves ICD Code Prediction, Especially for Rare Codes

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Health & Wellbeing — Artificial Intelligence & Machine Learning, Medical Devices & Health Technology · Depth: Expert, medium

Summary

A study presented at BioNLP 2026 by Jinghui Liu and Anthony Nguyen introduces a multi-version training approach to enhance International Classification of Diseases (ICD) code prediction, particularly for rare codes. Traditional models are often optimized for single ICD versions, despite the continuous evolution of these systems and the "long-tail problem" where rare codes are difficult to predict accurately. The researchers investigated training version-independent models by integrating data from different ICD versions. Specifically, they incorporated ICD-9 data into the training of a modified label-wise attention model designed for ICD-10 prediction. This method yielded a 27% increase in micro F1 score for 18K rare ICD-10 codes compared to training solely on ICD-10 data. Furthermore, the multi-version training substantially improved macro metrics for 8K frequent ICD-10 codes while utilizing fewer model parameters.

Key takeaway

For Machine Learning Engineers developing clinical coding automation, you should consider multi-version training to enhance model robustness and accuracy. Integrating older ICD version data, such as ICD-9, into current ICD-10 prediction models can significantly boost performance for rare codes by 27% and improve overall macro metrics. This approach allows you to build more implementable models with fewer parameters, effectively addressing the long-tail problem inherent in medical coding.

Key insights

Combining data from different ICD versions significantly improves rare code prediction and overall model performance.

Principles

• ICD version mismatch can be overcome with multi-version training.
• Multi-version training addresses the long-tail problem in clinical coding.
• Fewer model parameters can achieve better performance with diverse data.

Method

A modified label-wise attention model is trained for ICD-10 prediction by incorporating ICD-9 data, creating a version-independent model.

In practice

• Integrate historical ICD data (e.g., ICD-9) into current version training (e.g., ICD-10).
• Prioritize multi-version training for improving rare medical code accuracy.

Topics

• ICD Code Prediction
• Multi-version Training
• Rare Code Performance
• Clinical Coding Automation
• Label-wise Attention Model
• Long-tail Problem

Best for: NLP Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Severity-Aware Curriculum Learning with Multi-Model Response Selection for Medical Text Generation — Severity-aware curriculum learning with multi-model response selection significantly improves medical text generation quality and relevance.
• Systematic Evaluation of the Quality of Synthetic Clinical Notes Rephrased by LLMs at Million-Note Scale — LLM-rephrased clinical notes retain coarse-grained utility but lose fine-grained detail, with chunking offering mitigation.
• How We Built Zeta2: Training an Edit Prediction Model in Production — Ben Kunkle, Zed — Training a specialized code edit prediction model in production leverages frontier model distillation and filtered user-settled data for optimal examples.
• Multi-LCB: Extending LiveCodeBench to Multiple Programming Languages — Multi-LCB extends code generation benchmarks beyond Python to expose LLM multilingual performance disparities and overfitting.
• Contrastive Training with LLM-generated Near-Misses for Robust Code-Switching Speech Recognition — Contrastive training with LLM-generated near-misses improves code-switching ASR by targeting critical regions.
• On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance — LLM annotation reliability hinges on definition alignment, not just text memorization, as prompt-based corrections are largely ineffective.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.