About

I'm a PhD student in Computer Science at the University of Otago, working on AI for healthcare and clinical NLP.

My research focuses on automated medical coding and language-model agents for clinical decision-making. I'm particularly interested in how medical knowledge, structured representations, and agent harness design can make language models more reliable in complex clinical workflows. I also work on post-training methods, including supervised fine-tuning and reinforcement learning, to improve how these systems reason, interact with their environment, and use feedback.

Before starting my PhD, I completed an MSc in Data Science and Analytics at the University of Leeds and worked as a data analyst in industry.

News

  • 2026.09

    AcceptedCoLa-ICD was accepted to Findings of EMNLP 2026. I'll present the work in Budapest this October.

  • 2026.08

    Released the CoLa-ICD preprint on arXiv.

  • 2026.02

    Started serving as Postgraduate Representative for the Department of Computer Science.

  • 2025

    Tutored COSC326: Computational Problem Solving at the University of Otago.

  • 2024.11

    Started my PhD at the University of Otago.

Publications

Findings of EMNLP 2026PosterFirst author

CoLa-ICD: A Knowledge-Enhanced Framework for Long-Tail Automated Medical Coding

Yihang Cheng, Veronica Liesaputra, Andrew Trotman

0.108 vs 0.069+57%

Macro-F1 on MIMIC-IV ICD-10, compared with GKI-ICD

0.204 vs 0.127+61%

Micro-F1 on rare codes seen only 1 to 10 times in training

Paper 
@inproceedings{cheng2026colaicd,
  title     = {CoLa-ICD: A Knowledge-Enhanced Framework for Long-Tail Automated Medical Coding},
  author    = {Cheng, Yihang and Liesaputra, Veronica and Trotman, Andrew},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

Research

Long-tail Automated Medical Coding Published

My work on automated medical coding focuses on rare ICD codes, where limited training examples make conventional models difficult to generalize. CoLa-ICD combines medical terminology knowledge, relationships between ICD codes, and clinical evidence alignment to improve long-tail prediction. The method was evaluated on MIMIC-III and MIMIC-IV and is published in Findings of EMNLP 2026.

LLM Agents for Sequential Clinical Diagnosis Ongoing

I'm developing LLM agents that approach diagnosis as a sequential decision-making problem. The agent interacts with a simulated patient environment, decides which questions or tests to request, maintains a diagnostic state, and updates its differential diagnosis as new evidence becomes available. My current work focuses on agent harness design and post-training, including trajectory generation and reinforcement learning, as well as evaluating diagnostic accuracy, evidence acquisition, interaction efficiency, and consultation cost.

Semantic IDs for Generative Medical Coding Ongoing

I'm exploring structured output representations for generative medical coding. Instead of generating ICD codes directly, language models predict Semantic IDs that allow related diagnoses to share parts of their representation while preserving exact code identity. The project studies how the organization of the output space affects learning, generation, and long-tail medical coding.

Background

Education

2024 – Present
PhD Student in Computer Science
University of Otago, New Zealand

Clinical NLP · Medical Coding · LLM Agents

2021 – 2022
MSc Data Science and AnalyticsDistinction
University of Leeds, UK

Academic & Service

2026 – Present
Postgraduate Representative
University of Otago

Represent postgraduate students and contribute to postgraduate research-community activities.

2025
Tutor · COSC326 Computational Problem Solving
University of Otago

Led tutorials and labs and contributed to assessment and marking for ~70 students.

2026
Student Volunteer
EMNLP 2026
Current
Member
AI in Health Research Network, New Zealand

Industry

2022 – 2024
Data Analyst
Sensors Data, Shenzhen

Worked on product analytics, experimentation, user behaviour analysis, and metric design using Python and SQL.

Selected Activities

2026
EMNLP 2026
Poster · Budapest, Hungary

CoLa-ICD: A Knowledge-Enhanced Framework for Long-Tail Automated Medical Coding

2026
HINZ / AI in Health Pitch Fest
Pitch · Dunedin, New Zealand

How Can AI Support Doctors in the Clinical Workflow?

2022
Oxford Machine Learning Summer School (OxML 2022)
MLx Health & Bio · MLx Finance

Skills and honours

Skills

Research
LLM AgentsPost-trainingReinforcement LearningClinical NLPMedical CodingAgent Evaluation
Tools
PyTorchTransformersPythonLinuxSlurm/HPC

Honours

2021.05
Data Science Scholarship
University of Leeds
2020.05
First-class Scholarship
East China Jiaotong University
2019.05
Second Prize
China University Business Challenge (CUBC)