Making sepsis detection transparent, so clinicians can trust the alerts
Sepsis is the third leading cause of death in U.S. hospitals, contributing to at least 350,000 deaths annually. Globally, sepsis accounts for nearly 1 in 5 deaths. One in three patients who die in a hospital had sepsis during their stay.
Early detection saves lives, but existing tools fail to give clinicians the confidence they need to act on alerts.
Trust. Clinicians can't see the underlying data that the model uses to make predictions. Since they are unable to see the full picture, they do not have full confidence in the alerts.
False negatives vs. false positives. Sepsis prediction models are excessively tuned to reduce false negatives as missing a patient with sepsis has much more severe consequences. However, ignoring false positives causes an increased number of alerts and, subsequently, clinicians are overwhelmed leading to less trust in the alerts.
A clinician-facing sepsis prediction and decision support system. It shows the data behind every alert—vitals, procedures, and risk trends—so clinicians can see exactly what the model sees and decide whether to act.
Built on MIMIC-IV, a de-identified ICU electronic health record database with 90,000+ patients from Beth Israel Deaconess Medical Center, which includes:
Predicting sepsis onset in ICU patients using 6-hour sliding windows of clinical data, analyzing trends rather than isolated snapshots.
Three clinician-facing interfaces designed to surface the right information at the right level of detail.
Shows all ICU patients ranked by sepsis risk. Charge nurses use it to prioritize which patients need attention first.
Shows vitals, procedures, and risk trends for one patient. Bedside nurses use it to understand what changed and why.
Shows the risk score alongside the features driving it. Clinicians use it to decide whether to trust and act on an alert.
For charge nurses and unit leads. Displays all ICU patients in a single ranked list by sepsis risk, with a simulation clock that advances hour by hour.
For bedside nurses and attending physicians. Shows vitals trends, procedure events, and how the sepsis risk score changes over time — everything the model sees.
For clinicians evaluating an alert. Shows the hourly risk score alongside the specific features driving it, plus similar patients for clinical context.
How patient data moves through the system every hour
Hourly vitals, labs, and procedures from the ICU
6-hour sliding window of clinical data
Random Forest scores features into hourly sepsis risk
Risk, trends, and context across three views
Features and predictions saved per patient per hour
Built on Django's Model-View-Template pattern — loosely coupled layers that evolve independently
Service Logic
Parse upstream data, manipulate features, make predictions, fetch results, update template
Presentation
Interface
ORM / Data Layer
Data object management
MIMIC-IV Database
Actual Data