Effects of AI and Explainable AI in Myocardial Infarction Detection

How to cite

Li, Ziqing (Violet); Cathy; Bauer, Kevin; Li, Xitong. 2025. "Effects of AI and Explainable AI in Myocardial Infarction Detection." AEA RCT Registry. AEARCTR-0016742. https://doi.org/10.1257/rct.16742-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Ziqing (Violet) Li* , Cathy , Kevin Bauer , Xitong Li

* Primary investigator

Registry ID
AEARCTR-0016742
DOI
https://doi.org/10.1257/rct.16742-1.0
First registered
Last updated
Status
In development
Trial dates
2025-10-08 – 2025-12-10
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/16742

Abstract

This study examines the influence of artificial intelligence (AI) and explainable AI (XAI) on physicians’ diagnostic decision-making in the context of myocardial infarction (MI) detection based on ECG signals. In a two-phase experiment, physicians and medical students first diagnose a set of ECG cases without AI support (baseline). They then diagnose additional cases with predictions from a state-of-the-art AI system, with random assignment to receive either standard AI predictions (control) or AI predictions with explanations (XAI treatment). The primary outcome is diagnostic accuracy and alignment with AI advice. Secondary outcomes include more performance-related measures, workflow efficiency and self-reported well-being. We also examine potential heterogeneous treatment effects and potential mechanisms, such as changes in risk perceptions and responsibility allocation, using survey measures. This design allows us to compare diagnostic performance with and without AI, as well as identify the incremental effect of XAI over standard AI support.

Intervention

This study employs a mixed-design experimental approach, incorporating both between-subjects and within-subjects factors to evaluate the intervention.

The intervention is the provision of AI-based decision support during the diagnostic task in Round 2.

Between-Subjects Factor: Participants are randomly assigned to one of two intervention conditions:
- AI Group: Receives the AI’s predicted probability of MI.
- XAI Group: Receives the AI’s predicted probability plus an explainable visualization (Grad-CAM highlights).

Within-Subjects Factor: All participants complete the diagnostic task without any decision support in Round 1 (baseline) and then again with their assigned decision support in Round 2 (treatment). This allows for a direct comparison of each participant's performance against their own baseline, measuring the change attributable to the intervention.

Experimental design

We conduct a mixed-design randomized controlled experiment to study how AI and explainable AI (XAI) affect physicians’ and medical students’ diagnostic decision-making. All participants first complete a baseline round diagnosing ECG cases without AI support. They are then randomized into one of two groups and complete a second round with either AI predictions (control) or AI predictions with explanations (XAI treatment).

The design allows for both within-subject comparisons (baseline vs. with AI support) and between-subject comparisons (AI vs. XAI). Primary outcomes are diagnostic accuracy and alignment with AI advice; secondary outcomes include more performance-related measures, efficiency and well-being measures. Survey measures capture potential mechanisms such as perceived risk, responsibility, and confidence in AI.

Design details

Randomization method
Randomization done in qualtrics
Randomization unit
individual
Clusters
0
Observations
approx. 130
Arms
approx. 50 per treatment
Primary outcomes
1. Diagnostic accuracy 2. Alignment with AI advice
Secondary outcomes
1. Performance-related measures: False positive rate. False negative rate. Precision. Recall. AUC(Area Under Curve). 2. Efficiency-related measures: Decision to refer for additional tests (binary). Decision to consult another clinician (binary). Time taken for each diagnosis 3. Physician well-being: Burnout, measured via the abbreviated Maslach Burnout Inventory (numeric score).

Cited by

No citation data available for this registration yet.

Keywords