CSC Technology & Digital Applications 4 — Questions and Answers
Question 1: Artificial intelligence algorithms applied to a 12-lead ECG obtained in normal sinus rhythm have demonstrated the ability to detect:
- Subclinical or paroxysmal atrial fibrillation not present on the current tracing (Correct answer)
- Coronary artery calcium score from electrical signal patterns
- Real-time myocardial perfusion defects during ischemia
- Pulmonary artery systolic pressure with accuracy equivalent to echocardiography
Correct answer: Subclinical or paroxysmal atrial fibrillation not present on the current tracing
AI-ECG models can identify a 'fingerprint' of subclinical atrial cardiopathy in sinus rhythm that predicts current or future paroxysmal AF.
Question 2: In AI-driven clinical decision support for chest pain evaluation, which input combination provides the most accurate early risk stratification?
- Patient-reported pain severity score alone
- Integration of high-sensitivity troponin, ECG findings, and structured clinical history (Correct answer)
- Physician gestalt estimate based on clinical experience
- Population-level mortality rates for the patient's demographic group
Correct answer: Integration of high-sensitivity troponin, ECG findings, and structured clinical history
Validated tools such as the HEART score integrate troponin, ECG, and clinical history into a calibrated risk prediction that outperforms individual data elements alone.
Question 3: Which machine learning architecture is most widely used for automated segmentation of cardiac chambers from echocardiographic and cardiac MRI images?
- Convolutional neural networks (CNNs) (Correct answer)
- Logistic regression with L2 regularization
- Random forest with decision tree ensembles
- K-nearest neighbor classification
Correct answer: Convolutional neural networks (CNNs)
CNNs excel at image-based tasks by learning hierarchical spatial features, making them the dominant architecture for cardiac image segmentation.
Question 4: In the context of multicenter AI research in cardiology, 'federated learning' is best described as:
- Merging raw patient datasets from all institutions into one central server for model training
- Training AI models locally at each institution and sharing only model parameters, not patient data (Correct answer)
- A federal regulatory framework governing how cardiac AI must be validated
- A multi-center randomized trial design for evaluating AI diagnostic tools
Correct answer: Training AI models locally at each institution and sharing only model parameters, not patient data
Federated learning trains models locally and aggregates only weight updates, allowing collaborative AI development without transferring protected health information.
Question 5: At the population scale, natural language processing (NLP) of electronic health record clinical notes is most valuable for identifying patients with:
- Structured billing codes already present in administrative databases
- Undiagnosed atrial fibrillation mentioned in free-text notes but not coded (Correct answer)
- Laboratory values outside the normal reference range
- Imaging findings already stored in structured DICOM metadata
Correct answer: Undiagnosed atrial fibrillation mentioned in free-text notes but not coded
NLP extracts clinically relevant diagnoses and risk factors from unstructured text, capturing AF diagnoses and symptoms that were never assigned a billing code.
Question 6: A deep learning model trained on non-contrast cardiac CT data can predict which outcome that is NOT directly visible on the images?
- Coronary artery calcium Agatston score
- Five-year risk of major adverse cardiovascular events (MACE) (Correct answer)
- Left ventricular ejection fraction from cine sequences
- Degree of coronary artery luminal stenosis
Correct answer: Five-year risk of major adverse cardiovascular events (MACE)
AI models extract subtle imaging biomarkers (e.g., pericardial fat, myocardial texture) that encode long-term prognostic information beyond what human readers directly measure.
Question 7: In digital phenotyping for heart failure management, continuous passive data collected from smartphones and wearables is primarily used to:
- Completely replace scheduled clinic visits and in-person assessments
- Detect subtle changes in activity, sleep, and heart rate that may signal impending decompensation (Correct answer)
- Calculate serum BNP levels from wearable biometric signals
- Monitor direct oral anticoagulant drug levels in AF patients
Correct answer: Detect subtle changes in activity, sleep, and heart rate that may signal impending decompensation
Passive wearable signals such as declining step count, altered sleep patterns, and reduced HRV can identify early decompensation days before overt symptoms appear.
CSC Technology & Digital Applications 4
Artificial intelligence algorithms applied to a 12-lead ECG obtained in normal sinus rhythm have demonstrated the ability to detect: