-
-
Early sepsis detection: from live FHIR data and AI reasoning to real-time ICU alerts and mobile notifications.
Sepsis Early Detection:
AI-Powered Clinical Intelligence
Tagline:
Bridging the "Golden Hour" gap with FHIR-integrated AI and real-time ICU alerts.
Inspiration
Sepsis is a race against the clock. Mortality increases by approximately $8\%$ for every hour that treatment is delayed. In high-pressure ICU environments, clinicians are often drowning in data but starving for actionable insights. We were inspired to build a bridge between raw EHR data and the "Golden Hour" of clinical intervention—using AI not just to alert, but to explain.
What it does
Our platform is an end-to-end clinical decision support tool:
1.Ingests real-time patient data via HL7 FHIR standards.
2.Analyzes complex physiological trends using Gemini 1.5 Pro, calculating SOFA/qSOFA scores with an AUROC of 0.85+.
3.Explains the risk through an automated SBAR (Situation, Background, Assessment, Recommendation) report.
4.Notifies the care team via a high-priority mobile push notification, allowing them to initiate life-saving sepsis bundles instantly.
How we built it
The Brain: We used Google AI Studio to engineer prompts for Gemini 1.5 Pro. We leveraged its long context window to analyze hours of historical patient vitals and lab trends (MAP, Lactate, Respiratory Rate).
The Interface: A responsive React & TypeScript dashboard designed for high-stress clinical environments.
The Data: Integrated Synthetic FHIR data to demonstrate "plug-and-play" interoperability with modern EHRs like Epic and Cerner.
The Mobile Action: A storyboarded notification system that brings the "Intelligence" directly to the clinician's pocket.
We built a full-stack clinical intelligence pipeline:
The Data Layer: We utilized HL7 FHIR standards to ensure interoperability. The system ingests synthetic patient resources (Observations, Patient, and DiagnosticReports).
The Reasoning Engine: Developed in Google AI Studio, we leveraged Gemini 1.5 Pro. The model processes multi-modal inputs, including trended lab data and mean arterial pressure (MAP), where:
$$MAP \approx \frac{1}{3} \text{Systolic BP} + \frac{2}{3} \text{Diastolic BP}$$
The Frontend: A high-performance React & TypeScript dashboard that translates AI reasoning into a standard medical SBAR (Situation, Background, Assessment, Recommendation) format.
Technical Deep Dive: The Gemini Reasoning Engine
The "Intelligence" of our system goes beyond simple pattern matching. We designed a multi-stage prompt architecture within Google AI Studio to turn noisy FHIR data into a clinical SBAR report.
- Long-Context Trend Analysis
While traditional algorithms often look at a single snapshot in time, we utilize Gemini’s long-context window to ingest the last 24 hours of a patient's physiological trajectory. This allows the model to calculate the rate of change ($\Delta$) for critical markers:
$$\text{Risk} \propto \frac{\Delta \text{Lactate}}{\Delta t} + \frac{1}{\Delta \text{MAP}}$$
- Chain-of-Thought (CoT) Clinical Reasoning
We implemented a Chain-of-Thought prompting strategy. Instead of asking for a "Sepsis: Yes/No" answer, the model is instructed to follow the Sepsis-3 Guidelines logic:
Step 1: Evaluate the qSOFA criteria (Respiratory Rate $\ge 22$, Altered Mentation, Systolic BP $\le 100$).
Step 2: Cross-reference with lab results (Creatinine, Bilirubin, Platelets) to calculate the full SOFA Score.
Step 3: Assess for Septic Shock indicators (requirement for vasopressors or Lactate $> 2$ mmol/L).
- Handling Multi-Modal Input
One of the most powerful features of our Gemini implementation is its vision-to-data capability. If a FHIR server is unavailable, the model can parse a photo of a bedside monitor (OCR + Visual Reasoning) to extract vitals:
Input: Raw Image/Stream from a Bionet or Philips Monitor.
Output: Structured JSON containing $HR$, $SpO_2$, and $RR$.
- The Prompt "Guardrails"
To ensure safety, we utilized System Instructions that force the model to cite its sources. For every alert generated, the model must map the risk back to a specific observation ID from the FHIR bundle:
"The alert was triggered because Observation/102 (Lactate) showed a $25\%$ increase over the last 4 hours, exceeding the threshold of $2.0$ mmol/L."
Challenges we ran into
The primary challenge was Signal vs. Noise. Clinical data is notoriously "messy." A single low blood pressure reading might just be a sensor error. We had to engineer prompts that look for temporal patterns rather than isolated data points. We achieved this by utilizing Gemini's long context window to evaluate the last 6–12 hours of a patient's trajectory, maintaining an AUROC of:
$$AUROC \approx 0.85$$
This ensures the tool provides high sensitivity while minimizing the "alert fatigue" that plagues modern hospitals.
Accomplishments that we're proud of
We successfully moved from a raw data "fetch" to a fully reasoned clinical SBAR report in under 5 seconds. Achieving a balance between a high-tech AI backend and a simple, "one-tap" clinical UI is something we are incredibly proud of.
What we learned
We learned that in healthcare, Explainability is as important as Accuracy. A doctor won't initiate a sepsis bundle just because a "black box" said so. By forcing our AI to output its reasoning via specific risk drivers (e.g., "Rising Lactate + Decreasing $GCS$"), we built a tool that clinicians can actually trust.
What's next for Sepsis Early Detection
Future Roadmap
Dynamic Weighting: Implementing Bayesian updates to the risk model as more data flows in.
Computer Vision: Integrating the ability to "see" and digitize vitals from legacy bedside monitors using Gemini's vision capabilities.
Mobile-First Deployment: Refining the React Native push-notification system for instant bedside response.
Log in or sign up for Devpost to join the conversation.