MedAI-DX: An AI-Powered Real-Time Clinical Decision Support System for Resource-Constrained Healthcare Environments

Abstract: Rural and peri-urban healthcare facilities in developing nations face a critical physician shortage, with a single clinician often managing 80–120 patients daily under severe diagnostic resource constraints. Existing clinical decision support systems are predominantly cloud-only, English-exclusive, unimodal, or cost-prohibitive for deployment at primary health centres. This paper presents MedAI-DX, a multimodal, real-time AI-powered clinical decision support platform designed specifically for resource-constrained environments. The proposed system integrates a fine-tuned clinical Natural Language Processing (NLP) engine based on BioMistral-7B for multilingual symptom extraction and ICD-11 mapping, an EfficientNet-B4 computer vision module trained on NIH ChestX-ray14, ISIC 2020, and APTOS 2019 datasets for diagnostic image classification with Grad-CAM explainability, and a weighted evidence fusion risk stratification engine producing Green/Amber/Red triage classifications with structured referral recommendations. The system achieves a clinical entity F1 score of 0.84, image classification AUC-ROC of 0.87 across four disease categories, and an end-to-end latency under 2.5 seconds on standard cloud infrastructure. The full-stack deployment — React frontend, FastAPI backend, PostgreSQL with pgvector — is validated through a live interactive demonstration accessible via web browser. This work contributes a comprehensive, ethically grounded framework for AI-augmented clinical decision-making in multilingual, low-resource settings.

Keywords: Clinical decision support, multimodal AI, natural language processing, computer vision, risk stratification, ICD-11, Grad-CAM, EfficientNet, BioMistral, FastAPI, pgvector, healthcare AI, triage, multilingual NLP.