Company Context

You’re the PM for OncoGuide, a Series D healthcare AI company that sells a clinical decision support product to 120 US oncology clinics (mix of academic centers and community practices). OncoGuide integrates into Epic and Cerner via FHIR and is used by ~3,500 oncologists and ~90,000 active cancer patients/year across partner sites.

OncoGuide’s revenue model is per-site annual contracts ($250K–$1.2M ARR depending on volume) plus a smaller per-test fee when the product triggers molecular testing workflows. The company is competing with Tempus, Foundation Medicine, and several EHR-native analytics tools. Your differentiation today is workflow integration and fast turnaround on real-world evidence insights.

The CEO wants to expand into a high-stakes module: predicting patient response to immunotherapy (primarily PD-1/PD-L1 and CTLA-4 inhibitors) to help oncologists decide whether to prescribe immunotherapy, combine it with chemo, or pursue alternative regimens.

User / Market Scenario

Immunotherapy is clinically transformative for some patients but expensive and risky:

• Typical drug cost: $10K–$20K/month (payer-negotiated varies)
• Serious immune-related adverse events (irAEs): ~10–20% depending on regimen
• Response rates vary widely by tumor type and biomarkers

Personas (based on your research)

Current Product & Data Reality

OncoGuide already ingests:

• Structured EHR data (diagnosis, stage, meds, labs, vitals)
• Pathology reports (semi-structured)
• Some genomics (NGS panels) for ~35% of patients
• Imaging metadata (not raw images) for ~20% of patients

However, data quality is uneven:

• PD-L1 status missing in ~28% of NSCLC cases in community sites
• ECOG performance status missing in ~40% of sites (often in free text)
• Response outcomes are inconsistently captured; RECIST is common in academic centers but rare in community practices

Problem / Opportunity

Clinics are asking for help answering: “Is this patient likely to respond to immunotherapy, and what’s the expected benefit vs risk?”

Your internal analysis across partner sites suggests:

• Immunotherapy is used in ~52% of eligible advanced solid tumor cases
• Of those treated, ~35% show meaningful response (proxy: progression-free survival beyond a threshold; varies by tumor)
• ~18% experience moderate-to-severe irAEs requiring steroids, hospitalization, or discontinuation

Commercially, the sales team reports that 30% of renewal conversations mention “need stronger treatment selection support,” and two large health systems have threatened churn unless OncoGuide can show measurable impact on outcomes and cost.

Your Task (Deliverables)

Design the product (not just the ML model) that delivers immunotherapy response prediction in a way clinicians will trust and adopt.

1. Define the product vision and target users. Which persona(s) do you build for first and why?
2. Clarify the use cases and workflow integration points. Where in the clinical workflow does this appear (ordering, tumor board, prior auth, follow-up)?
3. Specify the “prediction” output as a product experience. What do you show (risk score, responder categories, confidence intervals, explanations, recommended next steps, missing-data prompts)?
4. Prioritize an MVP scope using a structured method (e.g., RICE or Kano). Include what you will explicitly not do in v1.
5. Define success metrics and an evaluation plan that balances adoption, clinical impact, and safety (including how you’ll monitor harm).

Constraints

• Timeline: MVP must launch to 10 pilot sites in 12 weeks.
• Team: 1 PM (you), 6 engineers, 2 data scientists, 1 clinical informaticist, 0 dedicated designer (shared).
• Regulatory: You must assume this may be considered SaMD depending on claims; Legal requires conservative positioning (“decision support,” not “recommendation”) for MVP.
• Explainability: At least one explanation mechanism is required for clinician trust and auditability.
• Equity: The model must be monitored for performance disparities across race/ethnicity, sex, age, and site type.
• Data limitations: You cannot rely on raw imaging in MVP; only metadata and reports.

What We’re Looking For

We’re evaluating your ability to translate a complex ML capability into a safe, adoptable product: crisp problem framing, user-centered design, prioritization under constraints, and measurable success criteria with clear trade-offs.