Company Context
You’re interviewing for Nimbus Health, a Series D healthcare SaaS company that sells an FDA-cleared clinical decision support (CDS) platform to 320 hospitals in the US. Nimbus integrates with Epic and Cerner and provides real-time risk scores in the EHR. The company’s ICU module is used by ~48,000 ICU beds and processes ~1.2B vitals/lab events per month.
Nimbus’ leadership wants to launch a new product: an early-warning system to predict sepsis in ICU patients. The goal is to help clinicians intervene earlier (cultures, fluids, antibiotics) while minimizing false alarms and maintaining clinician trust.
User & Market Scenario
Primary users
- Bedside ICU nurses: monitor patients continuously; respond to alerts; escalate to physicians.
- Intensivists / ICU physicians: make treatment decisions; want actionable, explainable signals.
- Rapid Response / Sepsis teams: handle escalations; care about prioritization and workload.
Competitive landscape
- Epic Sepsis Model (ESM) is already present in ~40% of Nimbus hospitals, but adoption is uneven due to perceived false positives and limited local customization.
- Third-party vendors (e.g., Bayesian Health, Dascena-like offerings) compete on “black-box” accuracy, but hospitals increasingly demand interpretability and governance.
What Nimbus knows from research (last 60 days)
Nimbus ran 18 clinician interviews and analyzed retrospective ICU data from 12 partner hospitals:
- Sepsis incidence in ICU admissions: 7–11% depending on case mix.
- Median time from first physiologic deterioration to antibiotics: 3.6 hours.
- Hospitals with faster antibiotics (within 1 hour of suspicion) show ~10–15% relative mortality reduction in comparable cohorts (observational).
- Current rule-based alerts (SIRS/qSOFA variants) generate ~2.4 alerts per ICU patient-day, and nurses report “alarm fatigue” as the #1 reason alerts are ignored.
- Clinicians want alerts that answer: “Why now?” and “What should I do next?”
Problem Statement
Nimbus’ ICU customers are asking for a sepsis predictor that:
- Identifies high-risk patients early enough to change outcomes (not just confirm obvious sepsis).
- Doesn’t overwhelm staff with false alarms.
- Fits within hospital governance: auditability, fairness checks, and clear clinical ownership.
Nimbus has committed to an MVP in 12 weeks for 5 design-partner hospitals.
Your Task (Deliverables)
As the PM for this product, walk through how you would define and design the prediction capability and the MVP experience.
-
Clarify the user need and Jobs-to-be-Done
- Who is the primary user for the alert in the ICU workflow (nurse vs physician vs sepsis team), and what decision are you trying to change?
-
Propose the feature set (data signals) for the model
- What categories of features would you include (vitals, labs, meds, comorbidities, interventions, notes-derived signals if any)?
- Which features are “must-have for MVP” vs “nice-to-have later,” and why?
-
Define the target variable (label) and prediction horizon
- How would you structure the target so it is clinically meaningful and minimizes label leakage?
- What time window(s) would you predict (e.g., sepsis in next 6 hours vs 12 hours), and how would you handle repeated predictions over time?
-
Design the product experience around the model output
- What does the alert look like in the EHR? Who gets notified? What explanations and recommended next steps are shown?
- How do you prevent alert fatigue (thresholding, tiering, snoozing, escalation rules)?
-
Define success metrics and a rollout plan
- What are your leading and lagging indicators? What targets would you set for the MVP?
- How would you validate safety and performance across hospitals with different patient populations?
Constraints
- Timeline: 12 weeks to MVP; 5 partner hospitals; limited ability to change EHR UI beyond embedded panels and secure messaging.
- Data availability: Real-time feeds include vitals, labs, meds, orders, demographics, problem lists, and device data. Clinical notes arrive with 2–6 hour delays and are inconsistently structured.
- Regulatory: Product must remain within Nimbus’ CDS scope (no autonomous treatment decisions). Must support audit logs and model versioning.
- Operational: ICU staffing is strained; any solution that increases pages/alerts by >10% is likely to be rejected.
- Ethical/clinical: False negatives can be catastrophic; false positives create fatigue and unnecessary antibiotics. Hospitals are sensitive to antibiotic stewardship metrics.
Interviewer guidance
Assume you have access to historical ICU data and can partner with a clinical lead at each hospital. You do not need to choose a specific ML algorithm; focus on product thinking, labeling strategy, and how you’d ship something safe and adopted.
