A clinically actionable and explainable real-time risk assessment framework for stroke-associated pneumonia

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Predictive Analytics

Background:

• Current healthcare is reactive, missing opportunities for early disease detection.
• Early detection improves care quality and reduces costs.
• Clinically actionable predictions are crucial for proactive healthcare.

Purpose of the Study:

• To develop and validate a real-time predictive model for early disease detection.
• To utilize electronic health records (EHRs) for proactive patient monitoring.
• To enhance the interpretability of AI-driven clinical predictions.

Main Methods:

• Developed a transformer-encoder-based model for analyzing heterogeneous EHR data.
• Applied the model to predict stroke-associated pneumonia (SAP).
• Utilized Integrated Gradient for model interpretability.

Main Results:

• The model demonstrated accuracy and stability on an independent test set.
• Achieved 98.6% detection rate for SAP patients with timely alerts.
• Alerts preceded physician diagnoses by an average of 2.71 days.
• Integrated Gradient highlighted clinically relevant historical patient events.

Conclusions:

• AI models can enable proactive healthcare through accurate, real-time predictions.
• Interpretable AI enhances clinical trust and utility in predictive models.
• This approach shows promise for improving early detection of conditions like SAP.