Accurate and automated assignment of ICD-9 codes from clinical narratives is essential for healthcare analytics, and clinical decision support. However, the task remains complex due to the high dimensionality of the label space, variability in clinical language, and the multi-label nature of diagnostic documentation. This paper presents a novel deep learning framework for multi-label ICD-9 code prediction from discharge summaries using a fine-tuned Bidirectional and Auto-Regressive Transformers(BART). Targeting real-world clinical documentation scenarios, the proposed approach models unstructured medical narratives using a sequence-to-sequence architecture that captures both global context and fine-grained semantic cues. Each discharge summary is tokenized and encoded using BART’s bidirectional encoder and autoregressive decoder, enabling robust multi-label classification across the top 20 most frequent ICD-9 codes. Evaluation on the MIMIC-III dataset demonstrates strong performance, achieving a F1 score of 71%, while also showing balanced precision-recall behavior across high and low-frequency codes. ROC-AUC confirms the model’s discriminative capability across imbalanced labels. These results validate the model’s capability to handle complex, multi-label classification scenarios common in clinical documentation.