Enterprise AI Analysis

FT-ARM: Fine-Tuned Agentic Reflection Multimodal Language Model for Pressure Ulcer Severity Classification with Reasoning

Pressure Ulcers (PUs) are a prevalent and serious healthcare concern. To guide appropriate treatment, accurate categorization of PU severity into one of four categories (I - IV) is required. However, as severity categories often have subtle and subjective visual distinctions, manual staging is challenging and prone to variability across clinicians, necessitating automated solutions. Prior Al-driven approaches explored Convolutional Neural Networks (CNN) and Vision Transformers (ViT), which yielded promising image classification results but their predictions had limited interpretability. Multimodal large language models (MLLMs), which integrate vision and language understanding, are an emerging paradigm for contextualized and explainable image classification. We present FT-ARM (Fine-Tuned Agentic Reflection Multimodal model), which combines a fine-tuned MLLM with an agentic self-reflection mechanism to classify pressure ulcer image severity classification, while also providing rich rationale (reasoning) and context for its classifications. Inspired by diagnostic reassessments by human clinicians, FT-ARM's self-reflection strategy, performs iterative self-refinement of its initial predictions by reasoning over visual features and encoded clinical knowledge (via natural language understanding of clinical notes) to improve classification accuracy and consistency. In experiments on the publicly available Pressure Injury Image Dataset (PIID), our fine-tuned model-FT-ARM with LLaMA 3.2 90B as backbone-achieved an accuracy of 85% in classifying pressure ulcer stages I-IV, outperforming prior CNN-based models (by +4%). It is also instructive to note that prior work utilizing CNN or ViT models, typically reported model performance in offline evaluations, which would likely degrade in live deployments. In contrast, FT-ARM is designed for and evaluated in a live inference scenario that reflect real-time deployment conditions, enhancing its potential for clinical application. Beyond predictive performance, FT-ARM generates clinically grounded natural language explanations (reasons) for each prediction, offering interpretability aligned with expert reasoning. By combining fine-tuning with reflective reasoning on multimodal inputs, FT-ARM advances the reliability, transparency, and clinical utility of automated wound assessment systems-addressing a critical need for consistent and explainable pressure ulcer staging to support improved patient care.

Keywords: Pressure Ulcers, Deep Learning, Multimodal Large Language Model, Fine-Tuning, Agentic Reflection