Processing of clinical notes for efficient diagnosis with feedback attention-based BiLSTM.

Area of Science:

• Intelligent Medicine
• Clinical Informatics
• Artificial Intelligence in Healthcare

Background:

• Clinical records are valuable for predicting patient health but are underutilized due to complexity and sparsity.
• Effective analysis of clinical notes can significantly enhance diagnostic accuracy and patient care.
• Existing methods struggle to fully leverage the rich information within unstructured clinical data.

Purpose of the Study:

• To propose a novel Feedback Attention-based Bidirectional Long Short-Term Memory (FABiLSTM) model for enhanced disease prediction using clinical records.
• To address the challenges of complexity, high dimensionality, and sparsity in clinical notes.
• To improve the accuracy and efficiency of clinical decision-making through advanced machine learning.

Main Methods:

• Utilized PubMedBERT for filtering irrelevant information from clinical records.
• Enhanced word embeddings with global vector representations and K-means clustering for numerical data.
• Employed Term Frequency-Inverse Document Frequency (TF-IDF) analysis and a billiards-inspired optimization algorithm.
• Developed a Feedback Attention-based Bidirectional Long Short-Term Memory (FABiLSTM) architecture.

Main Results:

• The FABiLSTM model demonstrated high performance on the MIMIC-III dataset.
• Achieved an accuracy of 98.52%, precision of 98%, F1 score of 98.2%, and recall of 98.2%.
• The model effectively captured complex information for accurate disease prediction, outperforming current practices.

Conclusions:

• The proposed FABiLSTM model offers a significant advancement in utilizing clinical records for disease prediction.
• This approach enhances diagnostic capabilities and supports better patient care outcomes.
• The findings highlight the potential of sophisticated AI models in intelligent medicine for clinical decision support.