Data Science for a Safety Net Hospital

Challenge

Community-acquired infections (CAI) are among the most common reasons for admission at regional safety-net hospitals, and every case raises a triage question: manage it outpatient, on the floor, or in the ICU? Patients who start on the floor and deteriorate into the ICU a day or two later may signal mis-triage: the wrong level of care, and worse outcomes. The hospital wanted answers, but couldn’t fund a multi-year, multimillion-dollar prospective study.

Approach

Instead of a prospective study, we used data science on data the hospital already had: discharge abstracts and daily itemized billing from its purchasing cooperative. Working with business leaders, clinical SMEs, and IT, we defined the study carefully, agreeing on conditions, definitions, and metrics. From the billing data we could see the actions taken (antibiotics by class/dose, floor vs. ICU location, organ support like ventilation, vasopressors, dialysis), identify CAI and severe sepsis, define mis-triage as ICU admission on day two or three after initial floor care, and characterize patients by care pattern.

Outcome

Mis-triage was common (1 in 8 patients) and drove significantly greater resource use and increased mortality. It was more frequent in older patients, those without complex co-morbidity, and those started on a single rather than multiple antibiotics; notably, smaller hospitals had lower mis-triage rates. We shared the findings across the network so better-triaging hospitals’ practices could spread, and flagged specific cases for clinical review.

Why Trexin

Data they already had, answering a clinical-operations question fast and affordably: the foundation-first mindset, a decade before it became an AI prerequisite.