Deep Neural Networks For Reliable Detection Of Ai-Generated Text In Healthcare Systems

Abstract

Trust serves as a cornerstone of healthcare, fundamentally shaping patient confidence in medical advice and influencing clinical decision-making. The growing integration of artificial intelligence (AI) into medical consultations, however, has introduced concerns related to reliability, bias, and transparency. AI systems rely heavily on large, high-quality datasets for training, and deficiencies such as incomplete, imbalanced, or noisy data can lead to flawed or biased outputs. These limitations may compromise patient care and erode trust in healthcare services. To address this challenge, we propose an innovative framework for distinguishing AI-generated responses from those authored by human physicians within health consultation contexts. Central to this approach is MedXNet, an efficient hybrid neural network architecture designed to capture both contextual and structural patterns in medical text. The model employs word-level tokenization to transform textual data into numerical sequences, facilitating effective vocabulary management and fixed-length input processing while preserving semantic information. MedXNet integrates Bidirectional Long Short-Term Memory (BiLSTM), Transformer, and Convolutional Neural Network (CNN) layers to extract global dependencies, contextual relationships, and local linguistic features within consultation responses. A fully connected dense layer with dropout regularization and softmax activation is used for final classification. We evaluated MedXNet against several baseline deep learning architectures, including LSTM, Bi-LSTM, GRU, and 1D-CNN models, using three datasets of increasing complexity. Experimental results demonstrate the robustness and generalizability of the proposed architecture across diverse consultation scenarios.