CVS Health·Data Scientist·Technical Phone Screen·Senior

Senior

May 2026

Summary

CVS Health data scientist interview with a single massive ML design question covering missing data, class imbalance, metrics, and logistic regression improvements all at once. The depth expected was pretty intense for what I assumed would be a standard technical screen.

Questions Asked (1)

You're building a binary classifier to predict 30-day hospital readmission using claims and EHR data, with ~30% missing values and only 6% positive class rate. Walk through your full plan: how do you diagnose and handle missing data, choose and justify classification metrics, deal with class imbalance, and improve logistic regression beyond the baseline, all while preventing leakage?

Technical Trade-offsProduct Analytics & MetricsData Modeling

Author's notes

This was basically four questions stapled together and I did not realize that until I was already three minutes into the missing data section.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer as a pipeline: first diagnose missingness and handle it with appropriate imputation, then select metrics that reflect the clinical and business costs of errors, address class imbalance with resampling or weighting, and iteratively improve logistic regression using feature engineering and regularization—all while enforcing strict train-test separation and temporal validation to prevent leakage.

Pro tip: Emphasize that missingness itself can be informative—create missingness indicators and check if they correlate with readmission. Also, use a temporal split (e.g., train on earlier years, test on later) to mimic real deployment and avoid leakage from future data.

1. Diagnose and Handle Missing Data

Analyze missingness patterns (MCAR, MAR, MNAR) and create missingness indicators. Impute using domain-informed methods (e.g., median for labs, 'missing' category for categoricals) and consider multiple imputation if appropriate.

2. Choose Evaluation Metrics

Select metrics that align with clinical and business goals: recall (sensitivity) to catch high-risk patients, precision to avoid alert fatigue, and AUC-ROC/PR for ranking. Use F1 or cost-sensitive metrics if costs are known.

3. Address Class Imbalance

Apply techniques like class weighting, oversampling (SMOTE), or undersampling, but only on training data. Evaluate impact via cross-validation and consider ensemble methods if imbalance persists.

4. Improve Logistic Regression Beyond Baseline

Engineer features (e.g., interactions, aggregations, temporal trends), apply regularization (L1/L2), and tune hyperparameters. Consider splines or binning for non-linear relationships.

5. Prevent Leakage

Use a temporal split (train on past, test on future) and ensure all preprocessing (imputation, scaling, resampling) is fit only on training data. Avoid using future information or target-derived features.

Key Points to Mention

Missingness indicators and pattern analysis (MCAR/MAR/MNAR)
Metric selection: recall, precision, AUC-PR, and cost-sensitive evaluation
Class imbalance techniques: class weights, SMOTE, and evaluation on original distribution
Feature engineering: interactions, aggregations, and temporal features
Regularization and hyperparameter tuning for logistic regression
Temporal validation and strict train-test separation to prevent leakage

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.