Company Interview Questions · Healthcare & Pharma

Novartis, GE HealthCare, Optum & IQVIA: Data Science

❓ 6 questions🏢 Novartis & GE HealthCare & Optum & IQVIA🆕 Recent reports

Novartis, GE HealthCare, Optum & IQVIA: Data Science Interview Experiences (2025–26). Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145📝 Assignments
About this interview

Compiled from public candidate reports: Novartis Data Scientist (New Delhi, August 2025, plus two January 2025 reports), GE HealthCare Data Scientist (Haifa, December 2025; a referral offer in March 2025; Paris), Optum Senior Data Analyst (Noida and others) and IQVIA Data Analyst (New Delhi, Bengaluru, 2025–26). The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Novartis GE HealthCare Optum IQVIA
Role Data Scientist Data Scientist Senior Data Analyst Data Analyst
Rounds 2 technical + HR HM background (1 h) → technical with 2 team members (1 h); or a project presentation + technical Written test (tech basics + SQL, above 65% to pass) → interview Guesstimate + SQL → HR; or an online assessment
Questions Resume projects; how ML algorithms work in practice; making a project plan Per-class evaluation in multi-class problems; backprop and gradient descent; how you designed the training data Explain SQL joins with examples Average daily income of an 8-lane toll plaza; Python and SQL basics; why IQVIA
Outcome Not selected (3) 1 offer (referral), 2 not selected Mixed 1 offer, 3 not selected

01Rounds at a glance

02Questions and model answers

Q1
How do you evaluate performance on each class in a multi-class problem?
  • Confusion matrix (N×N), then per-class precision, recall, F1 (one-vs-rest).
  • Aggregate: macro-average (each class counts equally, which matters for rare diseases), weighted (by support), micro (overall, equal to accuracy for single-label problems).
  • Per-class ROC-AUC / PR-AUC, plus calibration per class.
  • In medical imaging, report sensitivity and specificity per condition with confidence intervals; clinically important misses get extra weight.
Q2
Explain backpropagation.
  • Forward pass: compute the activations layer by layer and the loss.
  • Backward pass: apply the chain rule from the loss backwards. For each layer, ∂L/∂W = (∂L/∂output) × (∂output/∂W), and the gradient passes to the previous layer via ∂L/∂input.
  • Frameworks build a computation graph and do this automatically (autograd). The optimiser then updates the weights.
  • Issues: vanishing or exploding gradients, solved with good initialisation, ReLU-family activations, normalisation, residual connections and gradient clipping.
Q3
How did you design your training dataset (GE HealthCare)?
  • Representativeness: patients from several sites, scanners and vendors, demographics, disease severities. Avoid single-hospital bias.
  • Splits by patient (never by image) to prevent leakage; an external site as the test set.
  • Labels: expert annotation with an inter-rater agreement check (Cohen's κ); adjudicate disagreements.
  • Imbalance: stratify; oversample rare findings carefully; augmentation that is realistic for the modality.
  • Documentation: a datasheet covering source, inclusion criteria and known gaps, needed for regulatory submission.
Q4
IQVIA guesstimate: average daily income of an 8-lane toll plaza.

Answer (structure):

  • 8 lanes, so about 4 per direction. Peak throughput with FASTag is about 300 vehicles per lane per hour; the average over 24 hours is about 40% of peak, giving about 120 vehicles per lane per hour.
  • 8 × 120 × 24 ≈ 23,000 vehicles a day.
  • Mix: 70% cars at ₹100, 20% buses and trucks at ₹300, 10% heavy multi-axle at ₹500 → average ≈ 70 + 60 + 50 = ₹180.
  • Daily income ≈ 23,000 × ₹180 ≈ ₹41 lakh. Sanity-check against highway traffic and state your assumptions.
Q5
Explain SQL join types with a healthcare example.
  • patients(patient_id), claims(patient_id, amount):
  • INNER JOIN: patients with at least one claim.
  • LEFT JOIN: all patients, with NULL claim fields for those without claims (to find patients with no claims, add WHERE c.patient_id IS NULL).
  • RIGHT JOIN: all claims, even orphan ones (a data-quality check).
  • FULL OUTER JOIN: everything from both tables, for reconciliation.
Q6
Novartis: how would you plan a data-science project from scratch?
  1. Define: the business or clinical question, the stakeholders, the success metric, the constraints (privacy, timeline).
  2. Data: sources, access approvals, a quality assessment.
  3. Approach: baseline → model plan, with an evaluation design (including bias and fairness checks).
  4. Milestones: discovery (2 weeks), proof of concept (4 weeks), validation, deployment and monitoring. Assign owners and risks.
  5. Governance: documentation, review gates, reproducibility. - Use a simple plan (timeline + RACI) and agree checkpoints with stakeholders.

03How to prepare

Sources

Created by Ajeetkumar