A deep learning framework for burn patient analytics: predictive modeling, anomaly detection, and image classification

Inaccurate severity assessments have made burn care a critical challenge to public hospitals. Fragmented clinical data and limited access to real-time decision support systems, especially in resource-constrained environments, pose significant challenges to accurate and timely clinical assessment. Current approaches focus on a single task, either burn classification or length of stay prediction, but not providing an integrated approach. This work presents a new artificial intelligence (AI)-driven framework which unifies the predictive analytics of anomaly detection and image based classification of the severity of burns into one real time analytics platform. The framework uniquely integrates multiple advanced AI techniques: an optimized Artificial Neural Network (ANN) for forecasting hospital stay duration (R 2 = 0.82; MAE ≈ 2 days), a Vision Transformer (ViT) for high precision burn severity classification (accuracy = 96.85%; precision = 96.5%; F1-score = 96.7%), Generative Adversarial Networks (GANs) and Autoencoders for detecting irregular treatment patterns (GAN-based detection accuracy = 97%), and Bayesian models for probabilistic outcome prediction based on key clinical parameters. Additionally, clinically relevant patient segments identified via K-Means clustering support the risk stratification, and targeted care strategies. The proposed system will combine these capabilities to addresses the major gaps in burn care for accurate diagnosis, efficient resource distribution, and real time proactive anomaly monitoring. This is demonstrated through extensive testing with real world clinical data and large scale image datasets, representing a unique integrated solution for the management of burns in low resource healthcare settings.