README.md

---
license: apache-2.0
language:
- zh
tags:
- medical
- perioperative
- complications
- lora
- adapter
- clinical-ai
datasets:
- perioperative-complications
pipeline_tag: text-classification
paper_url: https://doi.org/10.1038/s41746-025-02139-3
paper_title: "Enhancing Privacy-Preserving Deployable Large Language Models for Perioperative Complication Detection: A Targeted Strategy with LoRA Fine-tuning"
repository: https://github.com/gscfwid/PeriComp
---

# PeriComp: Perioperative Complication Detection LoRA Adaptors

> **🎉 This work has been accepted by npj Digital Medicine**

![PeriComp Performance](figure6b.png)
*Figure: Performance comparison of fine-tuned models across different sizes*

## 🩺 Model Overview

**PeriComp** is a collection of specialized LoRA (Low-Rank Adaptation) adaptors designed for **perioperative complication detection** from clinical narratives. These adaptors enhance smaller open-source language models to achieve expert-level performance in identifying and grading 22 distinct perioperative complications based on European Perioperative Clinical Outcome (EPCO) definitions.

### 🎯 Key Features

- **Expert-level Performance**: Matches or exceeds human clinician accuracy
- **Multi-scale Detection**: Simultaneous identification and severity grading (mild/moderate/severe)
- **Comprehensive Coverage**: 22 distinct perioperative complications
- **Resource Efficient**: Optimized for deployment on standard clinical infrastructure
- **Privacy Preserving**: Fully deployable on-premises without data transmission

## 📊 Model Collection

This collection includes five optimized LoRA adaptors:

| Model | Base Model | Parameters | F1 Score | Use Case |
|-------|------------|------------|----------|----------|
| **PeriComp-4B** | Qwen3-4B | 4B | 0.55 | Resource-constrained environments |
| **PeriComp-8B** | Qwen3-8B | 8B | 0.61 | Balanced performance/efficiency |
| **PeriComp-14B** | Qwen3-14B | 14B | 0.65 | High-performance deployment |
| **PeriComp-32B** | Qwen3-32B | 32B | 0.68 | Maximum accuracy requirements |
| **PeriComp-QwQ-32B** | QwQ-32B | 32B | 0.70 | Reasoning-enhanced performance |

## 🔬 Research Background

Perioperative complications affect millions of patients globally, with traditional manual detection suffering from:
- **27% under-reporting rate** in clinical registries
- **High variability** in expert performance across institutions
- **Cognitive load limitations** with complex documentation

Our research, published in **npj Digital Medicine** (DOI: [10.1038/s41746-025-02139-3](https://doi.org/10.1038/s41746-025-02139-3)), demonstrates that targeted task decomposition combined with LoRA fine-tuning enables smaller models to achieve expert-level diagnostic capabilities while maintaining practical deployability.

![Strict Performance Evaluation](figure7.png)
*Figure: Strict performance evaluation requiring exact complication type and severity matching*

## 🚀 Quick Start

### Installation

```bash
pip install transformers peft torch
```

### Basic Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and tokenizer
model_name = "Qwen/Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
base_model = AutoModelForCausalLM.from_pretrained(model_name)

# Load PeriComp adaptor
adaptor_name = "gscfwid/Qwen3-8B-PeriComp"
model = PeftModel.from_pretrained(base_model, adaptor_name)

# Prepare clinical input
clinical_text = """
# Objective
The objective is to identify postoperative complications from patient data in medical records, mimicking the diagnostic expertise of a senior surgeon.

# Diagnostic Criteria
The diagnostic criteria for the 22 postoperative complications are as follows:

{the diagnostic criteria for the 22 postoperative complications}

# Guidelines of Output structure

The output format is specified as:
{defined the output structure}

# Data of medical records

- {General Information (De-identified)}
- {Postoperative Medical Record}
- {Abnormal Test Results}
- {Examination Results}
"""

# Prompt preparation format details can be found in the example files:
# - comprehensive_prompts.json for QwQ 32B adapter
# - targeted_prompts.json for Qwen 3 adapters
# Note: Models are trained on Chinese clinical texts; performance on other languages is not validated

# Generate complication assessment
inputs = tokenizer(clinical_text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
```

## 🔧 Technical Details

### Training Methodology

- **Base Architecture**: Qwen3 series and QwQ-32B
- **Fine-tuning Method**: LoRA (Low-Rank Adaptation)
- **Training Data**: 146 complex surgical cases
- **Validation**: Dual-center external validation (52 cases)
- **Task Strategy**: Targeted decomposition approach

### LoRA Configuration

```python
lora_config = {
    "lora_rank": 16,
    "lora_alpha": 32,
    "learning_rate": 1e-4,
    "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"]
}
```

### 💻 Code and Data Access

- **GitHub Repository**: [gscfwid/PeriComp](https://github.com/gscfwid/PeriComp)
- **Complete Implementation**: Training scripts, evaluation code, and data processing pipelines
- **Prompt Templates**: Each model includes optimized prompt files:
  - `comprehensive_prompts.json`: For QwQ-32B adapter (comprehensive approach)
  - `targeted_prompts.json`: For Qwen3 adapters (targeted strategy)
- **Clinical Data**: Available upon reasonable request through institutional collaboration with appropriate ethical approval

## 📋 Supported Complications

The models detect and grade 22 perioperative complications based on European Perioperative Clinical Outcome (EPCO) definitions¹:

1. **Cardiovascular**: Myocardial injury, cardiac arrhythmias
2. **Respiratory**: Pneumonia, respiratory failure  
3. **Renal**: Acute kidney injury
4. **Gastrointestinal**: Paralytic ileus, anastomotic leakage
5. **Infectious**: Surgical site infections, sepsis
6. **Neurological**: Delirium, stroke
7. **Hematological**: Bleeding, thromboembolism
8. **And more...**

Each complication is graded as:
- **Mild**: Minor intervention required
- **Moderate**: Significant medical management
- **Severe**: Life-threatening, intensive intervention

---
¹ Jammer, I. et al. Standards for definitions and use of outcome measures for clinical effectiveness research in perioperative medicine: European Perioperative Clinical Outcome (EPCO) definitions: a statement from the ESA-ESICM joint taskforce on perioperative outcome measures. *Eur J Anaesthesiol* 32, 88-105 (2015). DOI: 10.1097/EJA.0000000000000118

## 🏥 Clinical Applications

### Primary Use Cases

- **Automated Screening**: Continuous 24/7 complication monitoring
- **Quality Assurance**: Systematic complication registry validation  
- **Clinical Decision Support**: "Second opinion" for complex cases
- **Research**: Standardized outcome assessment for clinical studies

### Deployment Scenarios

- **Resource-limited Settings**: Use PeriComp-4B/8B models
- **Standard Clinical Environment**: PeriComp-14B recommended
- **High-accuracy Requirements**: PeriComp-32B for maximum performance
- **Reasoning-enhanced Tasks**: PeriComp-QwQ-32B for complex diagnostic reasoning

## ⚠️ Important Considerations

### Clinical Validation Required

⚠️ **These models are research tools and require clinical validation before use in patient care**

### Limitations

- Training on Chinese medical records (generalizability considerations)
- Performance depends on documentation quality and completeness
- Not a replacement for clinical judgment

### Best Practices

- Use as **screening tool** with clinical oversight
- Validate outputs against clinical judgment
- Consider local adaptation for specific institutional practices

### Data Access

⚠️ **Clinical datasets are not publicly available due to patient privacy protection**

**Data Request Process**:
- Clinical datasets can be requested from corresponding authors for legitimate research purposes
- Requests must include detailed research protocol and intended use
- Institutional ethical approval is required before data sharing
- Data sharing agreements must comply with local privacy regulations
- Contact: gaoshw5@mail.sysu.edu.cn for data access inquiries

## 📚 Citation

If you use PeriComp in your research, please cite:

```bibtex
@article{gao2025pericomp,
  title={Enhancing Privacy-Preserving Deployable Large Language Models for Perioperative Complication Detection: A Targeted Strategy with LoRA Fine-tuning},
  author={Gao, Shaowei and Zhao, Xu and Chen, Lihui and Yu, Junrong and Tian, Shuning and Zhou, Huaqiang and Chen, Jingru and Long, Sizhe and He, Qiulan and Feng, Xia},
  journal={npj Digital Medicine},
  
  year={2025},
  doi={10.1038/s41746-025-02139-3},
  url={https://doi.org/10.1038/s41746-025-02139-3},
  
}
```

**Paper**: [Enhancing Privacy-Preserving Deployable Large Language Models for Perioperative Complication Detection: A Targeted Strategy with LoRA Fine-tuning](https://doi.org/10.1038/s41746-025-02139-3)

**Code**: [GitHub Repository - gscfwid/PeriComp](https://github.com/gscfwid/PeriComp)

## 📧 Contact & Support

For questions, issues, or collaboration opportunities:

- **Research Team**: Department of Anesthesiology, First Affiliated Hospital of Sun Yat-sen University, Guangzhou, China
- **Technical Issues**: gaoshw5@mail.sysus.edu.cn
- **Clinical Data Requests**: gaoshw5@mail.sysus.edu.cn (requires ethical approval and institutional collaboration)
- **Clinical Applications**: Perioperative Complications Detection
- **Code Repository**: [GitHub Issues](https://github.com/gscfwid/PeriComp/issues) for implementation questions

## 📄 License

This work is licensed under Apache License 2.0. See LICENSE for details.

---

*PeriComp: Advancing perioperative patient safety through AI-powered complication detection*