HAC Healthcare Analytics Predictive Analytics & Trends 3 — Questions and Answers
Question 1: Which approach allows multiple hospitals to collaboratively train a predictive model without sharing raw patient data, preserving privacy?
- Transfer learning
- Federated learning (Correct answer)
- Ensemble learning
- Active learning
Correct answer: Federated learning
Federated learning trains models across decentralized data sources by sharing model updates rather than raw data, protecting patient privacy.
Question 2: A hospital's sepsis prediction algorithm flags 80% of actual sepsis cases but also generates many false alarms. Which metric directly measures the proportion of flagged cases that are true sepsis?
- Recall
- Precision (Correct answer)
- AUC-ROC
- Specificity
Correct answer: Precision
Precision measures the proportion of positive predictions that are actually correct, directly reflecting the false alarm rate.
Question 3: Predictive models in healthcare increasingly incorporate social determinants of health (SDOH). Which of the following is an example of an SDOH variable?
- Serum creatinine level
- Housing instability (Correct answer)
- Heart rate variability
- Hemoglobin A1c
Correct answer: Housing instability
Housing instability is a social determinant of health, reflecting non-clinical factors that influence health outcomes and are increasingly used in predictive models.
Question 4: When building a 30-day mortality prediction model, including the discharge disposition variable (e.g., 'discharged to hospice') causes near-perfect accuracy. This is best described as:
- Model overfitting
- Label leakage (Correct answer)
- Multicollinearity
- Survivorship bias
Correct answer: Label leakage
Label leakage occurs when features that are direct consequences of the outcome are included in training data, artificially inflating model performance.
Question 5: Which statistical method is commonly used to handle missing data in healthcare predictive models by estimating plausible values based on observed data?
- One-hot encoding
- Multiple imputation (Correct answer)
- Min-max normalization
- Dimensionality reduction
Correct answer: Multiple imputation
Multiple imputation fills missing values by creating several plausible datasets based on observed data distributions, reducing bias from complete-case analysis.
Question 6: Which trend describes the use of wearable device data streams to continuously update patient risk scores in real time?
- Batch predictive modeling
- Continuous or streaming analytics (Correct answer)
- Retrospective analysis
- Descriptive dashboarding
Correct answer: Continuous or streaming analytics
Streaming analytics processes continuous data from wearables or IoT sensors in real time, allowing dynamic and up-to-date risk stratification.
Question 7: A health system uses a gradient boosting model to predict patient falls. Clinicians distrust the model because they cannot understand its decisions. The best solution is to:
- Replace it with logistic regression only
- Apply explainability tools such as LIME or SHAP (Correct answer)
- Increase model complexity further
- Remove all non-clinical features
Correct answer: Apply explainability tools such as LIME or SHAP
Explainability tools like LIME and SHAP provide local and global interpretations of complex models, helping clinicians understand and trust predictions.
HAC Healthcare Analytics Predictive Analytics & Trends 3
Which approach allows multiple hospitals to collaboratively train a predictive model without sharing raw patient data, preserving privacy?