Improving preliminary clinical diagnosis accuracy through knowledge filtering techniques in consultation dialogues

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Patient symptom descriptions are often vague, leading to inaccurate preliminary clinical diagnoses.
• Existing methods struggle with the ambiguity and complexity of medical symptom reporting.
• There is a need for advanced computational models to improve diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a deep learning model, the knowledgeable diagnostic transformer (KDT), for NLP-based preliminary clinical diagnoses.
• To enhance the accuracy and reliability of preliminary diagnoses derived from patient symptom descriptions.
• To address the challenge of vague or inaccurate patient-reported symptoms in clinical settings.

Main Methods:

• The KDT model utilizes a bipartite medical knowledge graph (bMKG) to extract symptom-disease relation triples.
• A knowledge inclusion-exclusion approach (KIA) is employed to filter out undesirable triples and mitigate knowledge noise.
• Token embedding techniques are combined with the transformer architecture for disease prediction.

Main Results:

• A large-scale Mandarin medical diagnosis question-answering dataset (MDQA) with 2.6 million entries was created for training.
• The KDT model was also trained on the NIH MedQuAD English dataset.
• The KDT achieved a remarkable 99% accuracy across various evaluation metrics, outperforming baseline transformer models.

Conclusions:

• The KDT model demonstrates significant effectiveness in enhancing diagnostic precision for preliminary clinical diagnoses.
• The study highlights the potential of knowledge-based AI and NLP to revolutionize medical diagnostics.
• The KDT offers a pathway to more accurate diagnoses, reduced misdiagnoses, and improved patient outcomes.