Summarizing Patients' Problems from Hospital Progress Notes Using Pre-trained Sequence-to-Sequence Models.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Healthcare professionals face information overload from daily patient progress notes.
• Accurate summarization of patient problems is crucial for clinical decision support and efficient care.
• Existing methods struggle to effectively abstract and generate clinical documentation for problem lists.

Purpose of the Study:

• To develop and evaluate a novel Natural Language Processing (NLP) task for automatic patient problem list generation from daily progress notes.
• To investigate the efficacy of state-of-the-art sequence-to-sequence transformer architectures, T5 and BART, for this task.
• To enhance model performance through data augmentation and domain adaptation techniques.

Main Methods:

• Utilized progress notes from the Medical Information Mart for Intensive Care (MIMIC)-III database to create a specialized corpus.
• Trained and evaluated T5 and BART transformer models on the clinical notes dataset.
• Implemented domain adaptation pre-training and data augmentation to improve medical vocabulary and knowledge integration.
• Assessed model performance using ROUGE, BERTScore, cosine similarity, and F-score on medical concepts.

Main Results:

• T5 models, particularly with domain adaptive pre-training, demonstrated significant performance improvements over rule-based systems and general pre-trained models.
• The proposed NLP task and domain adaptation strategies show promise for accurate problem summarization.
• Evaluation metrics indicated strong performance in understanding, abstracting, and generating clinical documentation.

Conclusions:

• Domain adaptive pre-training of T5 models offers a promising approach for automated patient problem list generation.
• This NLP methodology can help alleviate information overload and support clinical decision-making in hospital settings.
• Further research in NLP for clinical documentation holds potential for advancing healthcare informatics.