Peking University First Hospital Team Publishes World's First LLM for Acute Kidney Injury Risk Prediction and Attribution

Background

Acute kidney injury (AKI) is a common critical condition among hospitalized patients, with high incidence and mortality rates. Currently there is no specific treatment available, making early identification of high-risk patients key to improving prognosis.

Existing risk prediction models face two major challenges in clinical deployment: high false positive rates ranging from 70% to 94% that can lead to alarm fatigue among medical staff, and their black-box characteristics that do not output explainable or executable clinical factors.

What was achieved

A team from the Renal Division of Peking University First Hospital led by Professor Li Yang and Associate Professor Xizi Zheng developed what the hospital news describes as the world's first large language model framework for acute kidney injury risk prediction and attribution. Professor Li Yang is a professor, chief physician, doctoral supervisor, vice president of the hospital, deputy director of the Peking University Institute of Nephrology, and director of the Peking University Department of Nephrology. Associate Professor Xizi Zheng is an associate professor, deputy chief physician, and secretary of the Peking University Department of Nephrology.

The paper titled "Large language model driven multicenter prediction and explainable risk attribution of acute kidney injury" was published in Nature Communications. The study was a multicenter retrospective cohort analysis using electronic medical record data from 140,637 inpatient admissions across four geographically diverse Chinese hospitals of different levels.

The team used Qwen2.5-7B as the base model. After domain-specific pre-training and fine-tuning on a 4.26-billion-token nephrology corpus, they created AKI-PM for prediction and AKI-RAM for risk attribution. AKI-PM achieved an AUC of 0.95, accuracy of 0.91 and positive predictive value of 0.68 in internal validation, maintaining AUC of 0.92-0.96 and PPV of 0.69-0.74 in external validation after few-shot adaptation. AKI-RAM was evaluated by six nephrologists on 200 cases, receiving high scores (Likert scale 4.18-4.88) with moderate to good inter-rater reliability. The team plans prospective validation in real clinical scenarios.

How it works

Large language models are a form of artificial intelligence trained on massive amounts of text to understand context, process information and generate natural language outputs. The researchers constructed a specialized kidney-focused corpus (AKICorpus) of 4.26 billion tokens drawn from evidence-based medicine guidelines on AKI, kidney-related PubMed publications, clinical notes from public databases including MIMIC-IV and eICU, and proprietary electronic health record notes from 47,750 admissions at Peking University First Hospital. This corpus was used for continued pre-training of the Qwen2.5-7B model to embed nephrology domain knowledge, followed by supervised fine-tuning on patient data and expert alignment.

AKI-PM analyzes real-time data from a patient's electronic medical record to dynamically predict the risk of AKI occurring within the next 24 hours, with fast inference time supporting potential real-time application. For patients predicted to be at high risk, AKI-RAM generates structured, explainable outputs that distinguish modifiable clinical risk factors (those amenable to intervention such as medication adjustments or fluid management) from non-modifiable ones and provides personalized, clinically actionable intervention recommendations in language understandable to clinicians.

What it means for international patients

Peking University First Hospital's Renal Division is a national leader in kidney disease care and research. International patients seeking nephrology care at the hospital can access services through its Special Needs Clinic, which provides priority access to specialists. Some physicians are able to communicate in English, and the hospital accepts certain international insurance (typically requiring upfront payment and subsequent reimbursement claims).

The AKI models described represent a research advance based on retrospective data. Prospective real-world clinical validation is planned but current routine clinical availability is not reported. International patients should contact the hospital directly to confirm eligibility, availability of related services, language support and any integration of such predictive tools for their individual case.

About the hospital

Peking University First Hospital is a large comprehensive tertiary Grade A hospital located in Beijing. Founded in 1915, it is affiliated with Peking University and integrates medical services, teaching, research and preventive medicine. It was designated as a Central Government Designated Hospital and serves as a base for national health initiatives.

The hospital's Renal Division established nephrology as a specialty in China in the 1950s and is recognized as a leading center for kidney disease, home to the Key Laboratory of Renal Disease of the Ministry of Health and contributing to national clinical guidelines. It maintains multiple national key clinical specialties, conducts high volumes of complex care including kidney biopsies, and operates special needs services for patients including those from overseas.