Towards stroke prediction using electronic health records

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Stroke is a leading cause of death in Japan, necessitating proactive healthcare strategies.
• Predicting stroke diagnosis allows for timely interventions and improved patient outcomes.

Purpose of the Study:

• To develop and evaluate a predictive model for stroke diagnosis within one year using electronic health records.
• To improve the accuracy of stroke prediction models by addressing class imbalance issues.

Main Methods:

• Utilized approximately 8000 electronic health records from Tsuyama Jifukai Tsuyama Chuo Hospital.
• Transformed non-homogeneous temporal data for neural network input and designed custom regularization terms for the cross-entropy loss function.
• Evaluated model performance using Receiver Operating Characteristic (ROC) analysis.

Main Results:

• A dual-input neural network topology achieved an area under the ROC curve of 0.669.
• Custom regularization terms positively impacted the training process compared to standard cross-entropy loss.
• The developed models demonstrated effectiveness in handling imbalanced stroke case data.

Conclusions:

• The described techniques are viable for stroke prediction.
• The developed models serve as a foundation for a national clinical decision support system.
• This approach supports proactive healthcare measures for stroke prevention.