Inspiration
What it doeIt’s 2:00 AM in a crowded emergency department (ED). A febrile two-year-old waits with a high heart rate. Vital signs are an hour old, no one is assigned to reassess, and the parent says, "Something is wrong. My child is getting worse." Sepsis caused about 49 million cases and 11 million deaths worldwide in 2017, with over 40% of cases in children under five. However, current alerts are failing. In a study of ~167,000 pediatric ED visits, the most widely used alert caught only 36% of cases with a dismal 3% positive predictive value (PPV). Clinicians told us an alert seen 40 times a day is simply ignored, and unowned tasks are passed "like a hot potato." We realized the solution isn't just another algorithm; it’s about connecting age-aware risk, parental concern, and a closed-loop response with a named owner. ⚙️ What it doesBeyond the Alert is a context-aware sepsis alert prioritization system for pediatric EDs. It does not diagnose or initiate treatment; instead, it answers one crucial question: Should a clinician see this child sooner? The system operates in three tiers:Red Flag: Existing mandatory rules that cannot be overridden and trigger immediate emergency pathways. Watch: Activated by parent or nurse concern, regardless of the model score, prompting a reassessment within 30 minutes. Alert (Top Model Tier): Interrupts the assigned physician, requiring acknowledgement within 15 minutes and documented action within 30 minutes. If unanswered, the system escalates to the charge nurse. 🛠️ How we built itWe designed a phased data strategy and AI architecture:Data Layers: We utilized three layers of data. Layer 1 uses public PECARN data (5.83 million ED encounters) for profiling. Layer 2 targets de-identified ED extraction from the Waterloo Region Health Network (WRHN). Layer 3 uses synthetic data to stress-test rare deterioration and missing data. AI & Decision Design: The model predicts the risk of newly meeting Phoenix sepsis criteria within 24 hours after triage. We employed regularised logistic regression, with gradient boosting as a comparator, using only triage-time information (e.g., age- and temperature-adjusted vital signs, triage acuity, and missingness indicators). To prevent data leakage, later lab results define outcomes only. ⚠️ Challenges we ran intoMissing and Uneven Data: Vital signs are often taken just once at triage, with nothing monitored in the waiting room. Furthermore, blood pressure is taken in only 35% of children under two. Missing values had to be flagged, never assumed normal. The Rarity of Sepsis: Sepsis occurs in about 1 in 880 visits. Because it is so rare, false alarms are mathematically certain. This required us to design strict safety protocols where only the top tier interrupts workflow. Sample Size Constraints: With 15 inputs, we need about 300 cases (roughly 265,000 visits) to train effectively, meaning we must train across multiple hospitals or years. 🎯 Accomplishments that we're proud ofWe established strict mathematical and clinical targets to ensure our model reduces alert fatigue while maintaining safety. Using LaTeX formatting for our quantitative goals, we proposed:Target Alert Burden: $\le 5$ model-generated interrupting alerts per 1,000 pediatric ED visits. Target Precision: $\ge 10\%$ PPV (a significant improvement over the existing 3% baseline). Target Sensitivity: $\ge 45\%$ sensitivity. Most importantly, we integrated caregiver concern as a formal, non-overridable input, recognizing that parent concern is independently associated with ICU admission (6.9% vs 1.8%). 🚀 What's next for PEACE / Ctrl+Shift+HealOur Go/No-Go plan involves a phased evaluation. Phase 1 requires no new hardware, relying entirely on existing recorded data. Phase 2 introduces continuous monitoring (pulse oximeters, parent check cards every 30 minutes) exclusively for children on the "watch list". We will execute three steps: test on past records, run silently for three months, and finally go live. We advance only what is shown to improve clinical action safely. s
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Context Aware Sepsis Alert Prioritization for Pediatric
Built With
- clinical
- decision
- emergency
- healthcare
- medtech
- pediatrics
- sepsis
- support
