HAC Machine Learning in Healthcare 2 — Questions and Answers
Question 1: Which evaluation metric is most appropriate when assessing a rare-disease diagnostic model where false negatives carry severe clinical consequences?
- Accuracy
- Specificity
- Recall (Sensitivity) (Correct answer)
- F1 Score
Correct answer: Recall (Sensitivity)
Recall (sensitivity) measures the proportion of actual positives correctly identified, making it critical when missing a true case (false negative) is clinically dangerous.
Question 2: In a clinical NLP pipeline, 'named entity recognition' (NER) is primarily used to:
- Translate clinical notes between languages
- Extract structured entities like diagnoses and medications from free text (Correct answer)
- Classify entire documents into ICD-10 categories
- Generate synthetic patient records
Correct answer: Extract structured entities like diagnoses and medications from free text
NER identifies and classifies named entities (e.g., drug names, diagnoses, lab values) within unstructured clinical text.
Question 3: A hospital wants to predict 30-day readmission. The dataset has 5% readmitted patients. Which preprocessing strategy best addresses class imbalance?
- Delete majority-class records until classes are equal
- Apply SMOTE to oversample the minority class (Correct answer)
- Use accuracy as the sole training metric
- Remove all minority-class samples as outliers
Correct answer: Apply SMOTE to oversample the minority class
SMOTE (Synthetic Minority Over-sampling Technique) generates synthetic minority-class examples to balance the dataset without simply duplicating records.
Question 4: Transfer learning in medical imaging typically involves:
- Training a model from scratch on each new imaging modality
- Fine-tuning a model pre-trained on large non-medical image datasets (Correct answer)
- Transferring patient records between hospitals
- Applying logistic regression weights from one hospital to another
Correct answer: Fine-tuning a model pre-trained on large non-medical image datasets
Transfer learning reuses weights from models trained on large datasets (e.g., ImageNet) and fine-tunes them on smaller medical imaging datasets.
Question 5: Which data splitting strategy is most appropriate for a time-series patient vitals prediction model to prevent data leakage?
- Random k-fold cross-validation
- Stratified random split
- Chronological (temporal) train/test split (Correct answer)
- Leave-one-out cross-validation
Correct answer: Chronological (temporal) train/test split
A chronological split ensures the model is trained on past data and tested on future data, mimicking real deployment and preventing leakage of future information.
Question 6: Federated learning is particularly valuable in healthcare because it:
- Centralizes all patient data in one cloud server for faster training
- Allows model training across institutions without sharing raw patient data (Correct answer)
- Eliminates the need for model validation
- Requires only one hospital's data to generalize nationally
Correct answer: Allows model training across institutions without sharing raw patient data
Federated learning trains models locally at each institution and shares only model updates, preserving patient privacy and complying with regulations like HIPAA.
Question 7: Which technique helps clinicians understand WHY a black-box ML model made a specific prediction for an individual patient?
- Principal Component Analysis (PCA)
- SHAP (SHapley Additive exPlanations) (Correct answer)
- K-means clustering
- Gradient descent
Correct answer: SHAP (SHapley Additive exPlanations)
SHAP values assign each feature a contribution score for a specific prediction, providing local interpretability for individual patient decisions.
HAC Machine Learning in Healthcare 2
Which evaluation metric is most appropriate when assessing a rare-disease diagnostic model where false negatives carry severe clinical consequences?