Over reliance on artificial intelligence: An emerging risk in the care of arrhythmia

Dear Editor,

Artificial intelligence (AI) is rapidly changing cardiology by interpreting patterns in clinical data that often escape human perception. Machine learning based algorithms are starting to change the way electrocardiograms (ECGs) are interpreted and used for predicting cardiac risk, which can help us create new possibilities for the early detection and prevention.^[1,2] However, one emerging issue with this transformation is that this increasing AI based routine interpretation may increase the risk of over reliance and weaken the fundamentals of trainees in ECG. This issue is becoming increasingly relevant in both education and clinical practice.

The landmark foundational study by Hannun et al. showed a deep neural network trained on single lead ECG recordings accurately classified 12 common cardiac rhythms with an average area under the curve of 0.97 and has outperformed an average cardiologist level performance.^[1]

Similarly, newer systems, which are trained on multi-lead ECGs and wearable cardiovascular monitoring devices, have demonstrated comparable accuracy in everyday settings, which have benefited from continuous and passive arrhythmia screening.^[2] In the same fashion, an AI enabled ECG risk score developed in Japan successfully identified individuals at high risk for occult atrial fibrillation even when their ECGs showed normal sinus rhythm.^[3] This reflects

AI’s ability to detect subtle P wave and atrial conduction changes that are easily overlooked by trainees.^[4] This is both an advantage and a challenge because it may encourage students to defer to the algorithm instead of attempting careful waveform analysis themselves.

Despite these advances, AI in cardiology is not without challenges. While these tools can help support personalized management, their rising accuracy also increases the temptation among the trainees to accept algorithmic output as definitive rather than supportive. AI generated summaries can reduce the motivation to practice independent ECG interpretation, and this happens especially when algorithms appear authoritative and definitive.^[5,6] A recent systematic review of physician and trainee attitudes toward clinical AI further highlights widespread concern regarding over reliance on algorithmic outputs, particularly in diagnostic tasks that traditionally rely on pattern recognition and experiential learning.^[7]

The reliability of any algorithm ultimately depends on the quality and diversity of the data on which it is trained. Most AI models are trained on tertiary center populations, and thus, they may perform less accurately in rural, minority, or underrepresented groups whose ECG patterns and comorbidity profiles were not adequately captured in the original datasets.^[8,9] This issue becomes especially important in training, as students may not recognize when AI interpretations are less reliable for certain demographic groups and may start to trust outputs that should be questioned. Another key apprehension is the “black-box” manner in which many AI systems operate. Clinicians and patients often lack clarity on how algorithms reach their conclusions, which can limit trust and hinder clinical validation.^[7] For trainees, the inability to understand which ECG features influenced an automated diagnosis makes it difficult to compare the algorithm’s reasoning with their own, which reinforces dependence on these tools instead of promoting learning. These concerns are consistent with broader analyses of AI driven clinical decision support systems, which have identified reduced opportunities for active learning and skill acquisition as key mechanisms underlying AI induced deskilling in medical education.^[6]

Given the increasing risk of over reliance on AI tools, clear recommendations are needed. Medical education should focus on fostering strong ECG interpretation skills before introducing AI tools in early training. AI systems should highlight the ECG features that inform each prediction, which will help clinicians maintain an active engagement in interpretation, rather than directly accepting final results. Stronger regulatory oversight is also essential to define when and how AI outputs can be used in clinical decision making, which will help ensure that automated interpretations support and not replace the physician’s accountability for final decisions. Furthermore, patient privacy must be considered of utmost importance when using sensitive ECG and health data for AI development. Regularly checking and validating these algorithms can help prevent their use in settings where they are likely to underperform. With thoughtful integration, AI can strengthen arrhythmia care while preserving the essential human reasoning skills that trainees must develop to become competent and responsible clinicians.