Pfizer Elnaz Engagement — Multi-Stakeholder Framework + Answer Architecture

PFIZER — ELNAZ ENGAGEMENT · Multi-Stakeholder Metrics Framework + Answer Architecture · Priority 1–4 Work Queue

Elnaz Alipour, PhD — Sr Director, Medical Analytics Care Gaps & Customer Segmentation · All report to her: Rachael (AI Strategy R&D), Jay Shetty (Customer Data Acquisition), technical team

Built from Round 4 Continuation Brief · 2026-04-24
Companion to Elnaz Intel + Round 4 Cheat Sheet · Read alongside those

★ PRIORITY 1 — Multi-Stakeholder Metrics Framework · Run every Elnaz question through all 5 lenses in sequence

The rule: For any question Elnaz asks about AI model evaluation, testing outcomes, or business impact — answer it in sequence through all five stakeholder lenses without being prompted. She builds models that touch all five. Demonstrating that framing unprompted signals you understand her domain at the org level, not just the technical level.

Pharma / Pfizer Bishop credential: Amgen (FDA GxP GenAI), NewsRx

Care gap closure rate Script lift HCP reach Treatment initiation rate Biomarker testing rate Adherence Persistence Market share by segment

What changes: did the AI output translate to a rep visit that changed a script? Did the targeted HCP actually initiate treatment for the identified patient?

Payer Bishop credential: Evernorth (Cigna PBM) — 147 CMS KPIs

MLR (Medical Loss Ratio) PMPM cost Adherence rates Formulary compliance Prior auth volume Step therapy rates Specialty spend

PBM view: did adherence improve when the care gap closed? Did specialty drug spend move in the right direction for the targeted segment?

Provider Bishop credential: Invistics (300+ hospitals, 4 EHR systems)

HCP prescribing behavior change Readmission rate Utilization Clinical outcomes Epic/Cerner data quality

Did the HCP in the prioritization model actually change what they were prescribing? Was the underlying EHR data that fed the model clean?

Pharmacist Bishop credential: 21yr Registered Nuclear Pharmacist (RPh)

Dispense accuracy DUR (Drug Utilization Review) flags Patient counseling touchpoints Therapy gap ID at point of dispense

Unique lens: the pharmacist sees the patient at the dispensing event. Therapy gaps and adherence failures often surface there first — before they appear in claims data.

Patient Bishop credential: personal 2yr post-surgery experience

Access barriers Adherence friction Care navigation SDOH factors Patient journey friction

A care gap AI model that doesn't account for SDOH barriers will predict closure rates that don't materialize in practice. The patient lens catches that.

PRIORITY 2 — Evernorth/PBM Bridge · Build the explicit link between Bishop's Cigna work and Elnaz's data

The bridge: Elnaz's care gap models consume claims data that flows from PBMs. Bishop ran data architecture at Evernorth (Cigna's PBM). He has worked on both sides of this data handoff — the PBM producing the claims and the pharma analytics team consuming them.

What PBM claims data looks like (Evernorth):

Pharmacy claims: NDC code, fill date, days supply, quantity, prescriber NPI, pharmacy NPI, paid amount, copay, prior auth indicator
Medical claims: diagnosis codes (ICD-10), procedure codes (CPT/HCPCS), facility type, DRG, allowed amount, payer share
Prior auth records: drug name, diagnosis, approval/denial, dates
Specialty pharmacy: hub enrollment, specialty drug fills, adherence flags, case manager notes
Formulary: tier placement, step therapy requirements, coverage exceptions

How PBM data feeds Pfizer care gap models:

Pfizer receives de-identified or aggregated claims from PBMs through data licensing (IQVIA, Symphony, specialty pharmacy hubs)
Gap calculation = patient has confirmed diagnosis (claims) but no fill for the indicated therapy within lookback window
HCP prioritization = prescribers with high-volume patient population in gap status, ranked by propensity to prescribe given the right engagement

What PBM tracks vs. what pharma commercial tracks — same patient, different view:

PBM (Evernorth) sees	Pfizer commercial sees
Every fill across ALL drugs for this patient	Only Pfizer brand fills (direct or reported)
Prior auth approval/denial history	Market share for their therapeutic area
Formulary tier and copay at point of fill	Script lift from rep visit (lagged 30-90 days)
Step therapy and switch events	HCP prescribing behavior change (aggregated)
Real-time fill status (specialty hub)	Patient persistence at 6/12 month mark

Where the data gaps are: PBM data has fill-level accuracy but is de-identified by the time pharma receives it. Pfizer can see a care gap exists (no fill) but often cannot see why — is it access (prior auth denied), cost (copay too high), behavior (patient forgot), or clinical (HCP decided to wait)? The AI model has to infer root cause from proxy signals. That inference is what needs testing.

Say to Elnaz: "At Evernorth I worked on the data architecture that feeds these analytics downstream. I understand that a care gap in Pfizer's model is an inferred state — the PBM knows there was no fill, but not why. Testing the AI outputs means testing whether the model's root cause inference is reliable, not just whether the gap flag is correct."

PRIORITY 3 — Elnaz's Specific Metric Set · Know these cold

Care gap model outputs — what her team tracks:

Diagnosis rate: % of patients in indicated population with confirmed ICD-10 diagnosis on record. Gap = diagnosed but not treated.
Treatment initiation rate: % of diagnosed patients who start on therapy within a defined window. The core gap metric.
Biomarker testing rate: % of eligible patients who received the required diagnostic test before therapy eligibility. Often the upstream bottleneck — no test, no diagnosis, no treatment.

HCP prioritization model outputs:

Propensity score: probability this HCP will write a new Rx for the indicated therapy in next 90 days given targeted engagement
Decile ranking: HCPs segmented 1–10 by priority score — top decile gets rep visits, bottom gets digital only
Script lift: incremental Rx written per rep visit, compared to control cohort with no engagement. The ROI signal.
ROI per rep visit: script lift × brand revenue per script ÷ rep visit cost. Determines call allocation budget

Care gap closure rate by segment: % of identified gaps that were closed (treatment initiated) within the attribution window, sliced by geography, therapeutic area, HCP decile, and payer segment. This is Elnaz's primary KPI for whether the model is working.

HEOR metrics (longer horizon):

QALY: quality-adjusted life year — clinical outcome measure used in payer and policy contexts
Cost-effectiveness: cost per QALY gained for the intervention
Real-world evidence (RWE) outcomes: post-launch effectiveness data from claims and EHR, compared to clinical trial efficacy

PRIORITY 4 — Answer Framework Template · Any Elnaz question → run this sequence

The pattern: Elnaz's models produce outputs that 5 different stakeholders use to make decisions. When she asks how you'd evaluate an AI output, she wants to know if you think about all 5 downstream consumers — not just the technical metric that's easiest to measure.

Example: "How would you test a care gap AI model?"

Pharma: Did script rates change in the targeted HCP segment after model-driven engagement? Compare treatment initiation rate pre/post for the prioritized decile vs. the control group.

Payer: Did PMPM cost move for the targeted patient population? Did adherence rates improve for patients in the treated segment? Did specialty drug spend shift in the right direction?

Provider: Did the HCP actually change prescribing behavior — and was the prescribing change clinically appropriate given patient characteristics? EHR data validates whether the rep visit changed practice.

Patient: Did the patient receive the medication — and did they stay on it? Adherence and persistence at 6 and 12 months. SDOH factors that the model didn't account for show up here as unexplained closure failures.

Pharmacist: Are dispenses matching the recommended therapy? DUR flags at point of fill that signal therapy mismatch or contraindications the care gap model didn't capture.

The closing line: "What I'm looking for across all five is whether the model's output actually moved the patient down the care pathway — or just generated a list that looked actionable. Those are very different test results."

Testing Angle by Metric Type · What failure looks like at each layer

Metric	Model failure signal	Test method
Treatment initiation rate	Rate doesn't move in prioritized HCP segment despite engagement	Compare treated vs. control cohort at 90-day mark; check if HCP propensity scores were calibrated correctly
Biomarker testing rate	Patients flagged as gaps but biomarker test was already ordered and pending	Validate gap flag against pending lab orders in EHR; claims lag means test was done but not yet in data
HCP propensity score	Top-decile HCPs don't respond; bottom-decile HCPs write scripts with no engagement	Lift analysis with holdout set; check if score is predicting behavior or just reflecting past behavior
Care gap closure rate	Closure rate looks high but is driven by patients who would have initiated anyway	Attribution window and counterfactual analysis; incrementality test with randomized holdout
Adherence/persistence	Treatment initiates but drops off at 30/90 days; gap closes but reopens	Longitudinal tracking beyond initiation; flag patients at SDOH risk for early drop-off
Script lift	Positive lift in claims data but no actual patient outcome change	Cross-validate against EHR outcomes data; script was written but not filled, or filled but discontinued

Questions to ask Elnaz — metrics and validation

Attribution window: "What's your current attribution window for measuring care gap closure — and have you tested whether the window is long enough to capture initiation for patients with access barriers that add 30–60 day delays?"
Biomarker upstream: "Is biomarker testing rate being tracked as a separate upstream metric — or is it folded into the treatment initiation gap? Those require different interventions."
SDOH: "Has there been any subgroup analysis on closure rates by SDOH factors — geography, payer type, insurance tier? I'd expect to see differential closure rates that would tell us whether the model is systematically missing patients with access barriers."
Claims lag: "What's the typical claims lag in your data feed, and how are you handling the interpretation window so a patient who just initiated doesn't stay flagged as a gap for 30–45 days post-fill?"
Incrementality: "Is there a holdout group in the HCP prioritization model — so we can separate the natural prescribers from the ones the engagement actually moved?"

The bridge sentence — Evernorth to Pfizer · Use early

"At Evernorth I worked on the data infrastructure that sits upstream of this kind of analytics — the PBM claims, the formulary data, the prior auth records that feed care gap models on the pharma side. I understand that the gap Pfizer sees is an inference from incomplete data — the PBM knows there was no fill, but not why. That distinction matters for how you test the AI's root cause attribution — is it diagnosing the right failure mode, or is it giving a confident answer about something the underlying data can't actually resolve?"