A Deep Learning Model Incorporating Knowledge Representation Vectors and Its Application in Diabetes Prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Deep learning models show promise in disease prediction but lack interpretability and clinical integration.
• Integrating medical expertise is crucial for enhancing the clinical applicability of AI in healthcare.

Purpose of the Study:

• To develop a disease prediction model combining knowledge representation learning and deep learning.
• To improve the interpretability and clinical utility of deep learning for disease prediction.

Main Methods:

• Constructing a relationship graph of physical indicators and test values based on normal ranges.
• Encoding human physical examination data using knowledge representation learning.
• Inputting patient data vectors into a deep learning model with self-attention and CNN for prediction.

Main Results:

• The model achieved 97.18% accuracy and 87.55% recall in diabetes prediction.
• Outperformed traditional machine learning methods like Lasso, Ridge, SVM, Random Forest, and XGBoost.
• Demonstrated a 5.34% increase in recall compared to the best-performing Random Forest model.

Conclusions:

• Integrating medical knowledge into deep learning via knowledge representation learning enhances disease prediction.
• The proposed model shows significant potential for early diabetes detection and diagnostic assistance.
• This approach offers a pathway to more interpretable and clinically relevant AI in medicine.