Machine Learning in Healthcare Flashcards

Which method is commonly used to handle missing laboratory values in an EHR-based ML dataset?

Answer: Multiple imputation or mean/median imputation with missingness indicators

Multiple imputation or statistical imputation combined with a missingness indicator flag preserves sample size while signaling to the model that data was absent.

A natural language processing model assigns ICD-10 codes to clinical discharge summaries automatically. The primary quality metric for this task is:

Answer: Micro-averaged F1 score across all ICD codes

Micro-averaged F1 score accounts for class imbalance across hundreds of ICD codes by aggregating true/false positives globally before computing precision and recall.

Which architectural innovation allows Transformer-based models (e.g., ClinicalBERT) to capture long-range dependencies in clinical notes?

Answer: Self-attention mechanisms that relate every token to every other token

Self-attention allows each word to attend to all other words in the document simultaneously, capturing relationships regardless of distance in the text.

A predictive model for patient deterioration is deployed in an ICU. Over six months, its performance declines. This is most likely due to:

Answer: Concept drift as clinical practices, patient populations, or documentation patterns change

Concept drift occurs when the relationship between input features and the target outcome changes over time due to shifts in patient population, treatment protocols, or coding practices.

When using logistic regression to predict hospital-acquired infection, the output of the sigmoid function represents:

Answer: The probability that a patient will develop an infection

The sigmoid function maps the linear combination of features to a value between 0 and 1, which is interpreted as the predicted probability of the positive outcome.

For a healthcare ML model to comply with HIPAA, which of the following is required before using patient data for model training?

Answer: Data must be de-identified per HIPAA Safe Harbor or Expert Determination standards, or a valid authorization/waiver obtained

HIPAA permits using PHI for research/model training if data is properly de-identified or if a valid patient authorization, waiver, or data use agreement is in place.

Which technique is used to visually inspect which image regions most influenced a CNN's prediction on a chest X-ray?

Answer: Gradient-weighted Class Activation Mapping (Grad-CAM)

Grad-CAM produces a heatmap highlighting the image pixels that contributed most to a CNN's prediction, aiding radiologist review of model reasoning.