Machine Learning in Healthcare Flashcards

Recurrent Neural Networks (RNNs) and LSTMs are particularly suited for which type of clinical data?

Answer: Sequential time-series data such as ICU vital signs

LSTMs handle sequential dependencies in temporal data (e.g., ICU vitals over hours), maintaining memory of past states through gating mechanisms.

An ML model predicts a 90% probability of a patient developing diabetes within 5 years. According to responsible AI principles, the clinician should:

Answer: Use the prediction as one input alongside clinical judgment and patient context

Responsible clinical AI treats model outputs as decision support, not autonomous decisions; clinicians integrate predictions with full patient context.

Which approach is used to combine predictions from multiple ML models to improve overall predictive performance in healthcare?

Answer: Ensemble methods (e.g., stacking, bagging)

Ensemble methods aggregate predictions from multiple diverse models, reducing variance and bias compared to any single model.

A healthcare organization wants to validate an ML model prospectively. This means:

Answer: Deploying the model in a real clinical environment and evaluating future outcomes

Prospective validation involves deploying the model in real clinical workflows and measuring its performance on new patients going forward.

In the context of healthcare ML, 'feature engineering' refers to:

Answer: Transforming raw clinical data into informative input variables for a model

Feature engineering transforms raw data (e.g., converting lab timestamps into time-since-last-result) into meaningful variables that improve model learning.

Which FDA regulatory pathway is most relevant for AI/ML-based software that functions as a medical device (SaMD)?

Answer: 510(k) premarket notification or De Novo classification

The FDA regulates AI/ML SaMD through the 510(k) pathway (demonstrating substantial equivalence) or De Novo (for novel low-to-moderate risk devices).

What is the primary advantage of using a random forest over a single decision tree in clinical prediction tasks?

Answer: Random forests reduce overfitting by averaging predictions across many decorrelated trees

By averaging predictions from many trees trained on random data/feature subsets, random forests reduce variance and overfitting compared to a single decision tree.