Artificial intelligence for sepsis detection has moved from research papers to bedside implementation faster than almost any other clinical AI application. By 2026, the majority of U.S. hospitals running Epic, Cerner, or Philips-integrated EHR systems have some form of machine learning–based sepsis or deterioration alert active. But the implementation gap is real: nurses receive these alerts without always understanding what generated them or what the score actually means clinically.
This article covers the three most widely deployed AI sepsis tools, what inputs drive their scores, what the evidence says about their accuracy, and — most practically — how to use them at bedside without either dismissing alerts or treating every alert as a crisis.
The EDI is built into the Epic EHR and is the most widely encountered AI deterioration tool in U.S. hospitals. It generates a score from 0 to 100 every time the Epic database updates, which can be every few minutes when active monitoring and charting are occurring.
The EDI is a gradient boosting machine learning model trained on millions of Epic patient records. It was not designed to predict sepsis specifically — it predicts general clinical deterioration, defined as events like unexpected ICU transfer, respiratory failure requiring intubation, or cardiac arrest within the next 12 hours.
Key inputs that drive the EDI score:
Alert thresholds are set at the institutional level. A score of 60-65 is a common threshold for a soft alert (visible in the patient list without a push notification); 70+ is often a hard alert that fires a notification. Thresholds below 60 reliably generate alert fatigue without corresponding clinical benefit.
The Rothman Index (RI) is a separate commercial product from PeraHealth (now part of Philips) that runs either standalone or alongside Epic. Unlike the EDI, which is point-in-time, the RI displays as a continuous graph over the patient's entire stay — showing the trajectory of patient wellness rather than a single number.
The RI scores from +100 (healthy) to -91 (most ill). It weights nursing flowsheet documentation heavily — including nursing assessments, vital signs, and explicitly the nursing assessment of patient comfort and appearance. The model was validated specifically for detecting sepsis, cardiac events, and respiratory failure.
The Rothman Index's key differentiator is the trajectory display. A patient at -30 and declining from -15 over 6 hours is clinically more concerning than a patient who has been stable at -35 for 24 hours. The AI gives you both the number and the direction.
Cerner's Sepsis Advisor (also called the Sepsis Sniffer in some hospital implementations) is the Epic EDI equivalent for hospitals running Oracle Cerner (now Oracle Health). It uses a rules-based + machine learning hybrid model that evaluates SIRS criteria, qSOFA components, and organ dysfunction markers against the 3-hour and 6-hour sepsis bundle requirements.
Unlike the EDI, which predicts deterioration, the Cerner Sepsis Advisor is specifically calibrated to the Surviving Sepsis Campaign bundle criteria and generates alerts keyed to bundle timing rather than general deterioration prediction.
The peer-reviewed literature on AI sepsis detection is more nuanced than the vendor messaging.
What the evidence supports:
What the evidence does not support:
When an EDI or Rothman alert fires on your patient, the most effective response is a rapid structured assessment rather than either immediate escalation or dismissal. Consider this sequence:
1. Verify the inputs. What is actually driving the score? If the EDI jumped because a lactate just resulted, go look at the lactate. If the RI dropped because your last vital sign entry showed an oxygen saturation of 91% but the patient is actually at 96%, fix the documentation error. AI models don't know when data is wrong — they process what's in the chart.
2. Assess at bedside. The "I'm just not right" conversation — trust it. An AI score of 65 combined with a patient who looks worse to you than they did two hours ago is a two-signal alarm that warrants earlier action than either signal alone.
3. Check the trend, not just the number. Is this score higher or lower than one hour ago? Has it been rising for three hours or is this a new jump from baseline? Trajectory is clinical information.
4. Close the loop. If you assess, find a benign cause (pain driving tachycardia, patient anxious, BP cuff artifact), and don't escalate — document why. This documentation is what separates effective use of AI tools from alarm fatigue. If you escalate, document that too and track whether the intervention matched the AI prediction.
AI sepsis detection tools show the strongest sensitivity in patient populations where early signs are subtle or where clinical attention is frequently divided:
| Population | Why AI Detection Adds Value |
|---|---|
| Elderly patients (65+) | Classic sepsis signs (fever, tachycardia) often absent; altered mental status and hypotension are late findings. AI detects subtle vital sign drift earlier. |
| Immunocompromised (oncology, transplant) | Baseline vital signs and labs are atypical. AI trained on similar patients adjusts for baseline deviation rather than normal values. |
| Post-operative patients | Pain and inflammation confound bedside assessment. AI tracks unexpected deviations from expected post-op trajectory. |
| High nurse-to-patient ratio units (step-down, med-surg) | Less frequent bedside assessment. AI provides continuous monitoring between nursing contact. |
Nursing documentation directly influences every AI sepsis model in current clinical use. The models are not passively reading monitor data — they are reading the entire flowsheet, including every nursing entry.
The most impactful nursing documentation variables for these models:
The practical implication: accurate, timely nursing documentation is not just a legal and quality requirement — it is the data that feeds the AI that watches your patients while you're in another room.
AI score low, patient looks bad: Trust yourself. Call the provider. The sensitivity of these models for very early sepsis — before any measurable vital sign deviation — is limited. Your gestalt is detecting something the algorithm can't measure yet.
AI score high, patient looks fine: Investigate the inputs rather than dismissing the alert. Look for a documentation error, a spurious lab value, a monitor artifact. If the inputs are clean and the score is elevated, treat it as a clinical question that needs answering — even if the answer is "the patient is okay but let me watch them more closely this hour."
Both signals high: Act sooner rather than later. This is what the tool is designed for.
See also: AI tools ICU nurses are actually using | Epic AI features for nurses | Clinical decision support in the ICU
Get The ICU Notebook Newsletter
Clinical tools and career insights for ICU nurses. One email per week, no fluff.
Yes, send it freeNo spam. Unsubscribe any time.