AI Engineer for Medical Data Extraction

Role

AI / Machine Learning Engineer – Large Language Models & Clinical NLP

We are looking for an experienced AI/ML researcher or engineer to support a research project focused on automated clinical data abstraction from pediatric congenital heart surgery records using locally deployed open-weight large language models.

Project

Automated Clinical Abstraction in Pediatric Congenital Heart Surgery Using Open-Weight Large Language Models: Extraction of Diagnoses, Procedures, and Surgical Risk Categories

The project aims to evaluate whether open-weight LLMs running locally can accurately extract structured clinical information from unstructured medical documents, including:

* Operative notes
* Discharge summaries
* Echocardiography reports
* Cardiac catheterisation reports

The system will extract diagnoses, surgical procedures, operative variables, outcomes, and surgical risk categories such as RACHS-1 and STS-EACTS.

Main Responsibilities

The collaborator will be responsible primarily for the technical AI component, including:

* Building an end-to-end clinical NLP/LLM extraction pipeline.
* Deploying and testing open-weight LLMs locally, without sending clinical data to external APIs.
* Running and comparing models of different sizes.
* Implementing models using frameworks such as Hugging Face Transformers and vLLM.
* Working with quantised models using approaches such as AWQ/GGUF where appropriate.
* Designing and optimizing prompts for structured clinical information extraction.
* Comparing:
* Zero-shot prompting
* Few-shot prompting
* Retrieval-based few-shot prompting
* Creating structured JSON output schemas for extracted clinical variables.
* Implementing constrained or structured generation to ensure valid outputs.
* Developing an evidence-span validation system, where every extracted value is linked to supporting text from the original clinical note.
* Implementing a not_documented / abstention mechanism to reduce hallucinations.
* Developing the second-stage reasoning pipeline for assigning:
* RACHS-1 categories
* STS-EACTS mortality categories
* Comparing LLM-generated outputs with clinician-created reference-standard data.
* Calculating model performance including:
* Accuracy
* Precision
* Recall
* F1-score
* Weighted Cohen’s kappa
* Hallucination rate
* Confident error rate
* Performing error analysis to determine whether failures arise from:
* Incorrect information extraction
* Incorrect application of risk-scoring rules
* Ambiguous clinical documentation
* Measuring computational performance including:
* GPU memory requirements
* Inference speed
* Processing time per clinical document
* Approximate computational cost
* Maintaining reproducible code, model versions, prompts, and experiment configurations.
* Supporting interpretation of the results and preparation of the AI/methodology sections of the scientific manuscript.

Required Technical Skills

The ideal candidate should have strong experience in:

* Python
* Machine Learning / Deep Learning
* Large Language Models
* Natural Language Processing
* Hugging Face Transformers
* PyTorch
* Local LLM deployment
* vLLM or similar inference frameworks
* Prompt engineering
* Structured JSON generation
* Model quantisation
* Evaluation of NLP/LLM systems
* Data processing with Pandas / NumPy
* Git / version control

Highly Desirable Experience

Preference will be given to candidates with experience in one or more of the following:

* Clinical NLP
* Medical AI
* Electronic health records
* Information extraction from medical notes
* Biomedical language models
* Open-weight models such as Qwen, Gemma, MedGemma, Phi, or OpenBioLLM
* GPU-based LLM inference
* Research publication in AI, medical informatics, or biomedical NLP

Important

This is not primarily a predictive machine-learning project and is not simply a classification task.

The main technical challenge is to make open-weight LLMs read real-world clinical narrative, extract predefined clinical variables accurately, provide textual evidence for each extracted value, avoid unsupported information, and subsequently assign surgical risk categories.

Experience specifically with LLMs, NLP, information extraction, and evaluation of generative AI systems is therefore much more important than experience only with traditional models such as Random Forest, XGBoost, or standard regression.