This repository contains the source code and implementation for a comprehensive Artificial Intelligence-assisted Clinical Decision Support System (AI-CDSS) designed for automated and interpretable chronic wound triage. The system integrates deep convolutional feature extractors with generative Large Language Models (LLMs) to bridge the gap between opaque AI predictions and actionable clinical documentation.
Figure 1: Overview of the proposed AI-CDSS workflow, from visual data ingestion to LLM-generated reports.
Accurate triage of chronic wounds (diabetic foot, pressure, venous, and surgical ulcers) is a persistent challenge due to high inter-observer variability and the administrative burden of manual documentation. Our system provides a vision-only workflow that:
- Classifies wound images into six distinct categories using a fine-tuned VGG16 backbone.
- Interprets probabilistic outputs through a "Narrative Interpretability" layer powered by Google Gemini.
- Deploys via a scalable, asynchronous microservices architecture to ensure point-of-care responsiveness.
The AI-CDSS is engineered to decouple computationally intensive deep learning inference from the user interface, ensuring a seamless experience for healthcare professionals.
The system adopts the VGG16 architecture as a stable and transparent feature extractor. To adapt it to wound triage, the ImageNet classifier head was replaced with a custom, highly regularized head consisting of dense layers and dropout to prevent overfitting on specialized medical data.
The platform is orchestrated via Docker containerization and comprises four primary components:
- Frontend: An interactive, mobile-responsive portal built with Streamlit.
- Backend & Database: A Python-based API utilizing PostgreSQL for clinical traceability and longitudinal follow-up.
- Task Queue: Redis and Celery for managing asynchronous inference tasks.
- Generative Worker: Communicates with the vision engine and the LLM API to produce narrative reports.
The database ensures that every generated pre-report is auditable and tied to a specific patient history and model version.
Figure 2: Entity-relationship diagram illustrating the structure for clinical persistence and auditability.
The model development followed a three-phase strategy: Transfer Learning (TL), Extended TL with weight restoration, and Specialist Fine-Tuning (FT) of high-level convolutional blocks.
The system achieved stable convergence through targeted optimization, minimizing the gap between training and validation loss.
| Initial Training (Epochs 1-50) | Extended Training (At Epoch 94) |
|---|---|
 |
 |
| Rapid initial convergence to 72.22% accuracy. | Stabilization at 74.79% accuracy prior to fine-tuning. |
Fine-tuning the block5 convolutional filters transformed the model into a "venous ulcer specialist," increasing its recall for the most prevalent wound etiology.
- Global Accuracy: 75.21%.
- Venous Ulcer Recall: 95.16%.
| Confusion Matrix (Baseline) | Confusion Matrix (Fine-Tuned) | F1-Score Comparison |
|---|---|---|
 |
 |
 |
The interface is designed for point-of-care usability, allowing clinicians to register patients, upload images, and interact with the AI-generated narratives.
Rather than providing a "black-box" classification, the system uses Google Gemini to translate probabilistic distributions and patient metadata into editable clinical drafts. These reports explicitly highlight model uncertainties and suggest differential probabilities to ensure human-in-the-loop safety.
| Patient Registration | LLM Interpretability Interface |
|---|---|
 |
 |
| Ingestion modules allow for structured clinical data alongside high-resolution uploads. | Interactive templates contextualize model uncertainty for human verification. |
The research was conducted through a partnership between authors from the Federal University of Rio Grande do Norte (UFRN) and the Northeast Center for Strategic Technologies (CETENE).
 |
 |
 |
|---|---|---|
| Ernane Ferreira Rocha Junior | Ignacio Sánchez-Gendriz | Luiz Affonso Guedes |
| B.S. in Science and Technology (2023) and Computer Engineering (2026) from UFRN, Brazil . He is currently an M.Sc. student in Electrical and Computer Engineering at UFRN. Research focuses on AI, LLMs, intelligent agents, and software engineering. | Ph.D. in Mechanical Engineering (USP, 2017) and M.Sc. in Automatic Engineering . He is a researcher at MCTI/CETENE specializing in scientific computing. Research focuses on emerging AI technologies, LLMs, applied data science, and signal analysis. | Ph.D. in Automation (Unicamp, 1999) and Full Professor of Computer Engineering at UFRN. He has over 20 years of R&D experience in industrial software. Research focuses on data science, machine learning, and evolving systems for industrial applications. |
If you use this work in your research, please cite our official paper:
@article{rocha2026interpretable,
title={An AI-Enabled Digital Healthcare Workflow: Integrating Deep Learning and Large Language Models for Interpretable Wound Triage},
author={ROCHA JUNIOR, Ernane Ferreira and SÁNCHEZ-GENDRIZ, Ignacio and GUEDES, Luiz Affonso},
journal={},
year={2026},
volume={XX},
number={XX},
doi={XXXXXXX}
}Watch the detailed explanation of the project on YouTube:
