About the Client

Stanford University’s Department of Emergency Medicine leads innovation and scientific discovery in the field, collaborating across disciplines at Stanford, Silicon Valley, and globally.

The Challenge

Stanford sought to develop predictive models capable of analyzing patient Electronic Health Records (EHRs) to anticipate both the reason and timing of future emergency room visits with a specific focus on predicting cases related to domestic violence.

The dataset included over 300 million anonymized emergency visits from hospitals across the U.S., containing only relevant demographic and clinical information to ensure data privacy and compliance.

Marvik’s Approach

Our team designed and trained transformer-based predictive models using EHR data to forecast patient visits.
We pre-trained the model on large-scale datasets and then fine-tuned it for the specific domestic violence prediction task, achieving transferable performance to other conditions such as respiratory or cardiac diseases, enabling scalability at minimal cost.

To enhance explainability, we incorporated attention layer visualizations, helping clinicians understand which patient history factors influenced each prediction.
This interpretability ensures the model’s insights are not only accurate but also actionable and transparent for healthcare professionals.

The Results & Impact

• Processed and analyzed 300M+ medical records across U.S. hospitals.
• Achieved 85% AUC accuracy in predicting patient visits related to domestic violence.
• Developed a scalable, explainable AI framework ready to adapt to multiple medical prediction tasks.
• Enabled Stanford’s research team to advance toward more proactive and equitable patient care.

Why This Matters

This project illustrates how AI can transform healthcare from reactive to preventive.By anticipating emergencies before they happen, and explaining why, Stanford is redefining how hospitals allocate resources, prevent crises, and deliver safer, more humane care at scale.