PeriComp: Perioperative Complication Detection LoRA Adaptors

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 as a preprint on medRxiv, demonstrates that targeted task decomposition combined with LoRA fine-tuning enables smaller models to achieve expert-level diagnostic capabilities while maintaining practical deployability.

Figure: Strict performance evaluation requiring exact complication type and severity matching

🚀 Quick Start

Installation

pip install transformers peft torch

Basic Usage

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

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
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:

Cardiovascular: Myocardial injury, cardiac arrhythmias
Respiratory: Pneumonia, respiratory failure
Renal: Acute kidney injury
Gastrointestinal: Paralytic ileus, anastomotic leakage
Infectious: Surgical site infections, sepsis
Neurological: Delirium, stroke
Hematological: Bleeding, thromboembolism
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: [email protected] for data access inquiries

📚 Citation

If you use PeriComp in your research, please cite:

@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={medRxiv},
  pages={2025.06.11.25329235},
  year={2025},
  doi={10.1101/2025.06.11.25329235},
  url={https://doi.org/10.1101/2025.06.11.25329235},
  publisher={Cold Spring Harbor Laboratory Press}
}

Paper: Enhancing Privacy-Preserving Deployable Large Language Models for Perioperative Complication Detection: A Targeted Strategy with LoRA Fine-tuning

Code: GitHub Repository - 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: [email protected]
Clinical Data Requests: [email protected] (requires ethical approval and institutional collaboration)
Clinical Applications: Perioperative Complications Detection
Code Repository: GitHub 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

gscfwid
/

Qwen3-8B-PeriComp