Academic medical center - early deterioration prediction

The challenge

COVID-19 patients can worsen quickly. A leading academic medical center needed signals based on routine electronic health record activity to surface timely risks and prioritize attention without overwhelming staff. The project focused on learning from the hospital's own historical admissions and on making the rationale behind each prediction transparent to clinicians, who needed to trust and document any model-derived risk indication.

What we built

We built production-grade deep learning models with two variants: an admission model that scores risk using information available at admission, and a hybrid tracking model that updates risk as new clinical events arrive during the stay. The system provides both local explanations (key factors for an individual prediction) and global explanations (model behavior summaries for clinicians and quality teams). Delivered as a production-grade service so scores are consistently available to clinical staff alongside the explanatory views.

What changed