AI-driven ultrasound image acquisition in medical student training of echocardiography

Background Sudden cardiac events in young athletes remain a major public health concern, and large-scale cardiac ultrasound is a promising preventive strategy. However, reliance on expert operators remains the principal limitation to widespread implementation. Current training programs depend on extensive supervision by experienced clinicians, making large-scale deployment difficult. Advances in artificial intelligence (AI), particularly real-time visual guidance systems, offer scalable solutions to enhance standardized training in ultrasound image acquisition. This study represents a first step toward establishing a large-scale screening program in which medical students could perform cardiac ultrasounds to detect cardiac abnormalities in athletes. Purpose To determine whether AI-guided echocardiography (HeartFocus by DESKi) improves acquisition of diagnostic-quality echocardiographic images among medical students compared with traditional instruction, and to assess differences between in-person and remote AI-supported training. Methods In this prospective randomized study, 29 preclinical medical students were allocated to one of three groups: in-person instruction with AI-guided acquisition using HeartFocus (Group 1a, n=6), remote instruction with AI guidance (Group 1b, n=8), or traditional expert-led training without AI assistance (Group 2, n=15). Following half day training including practice on 4 standardized patients, each participant performed between 6 to 10 echocardiographic examinations on different standardized patients by acquiring 10 standard cardiac views. For each clip, diagnostic quality was independently assessed by blinded experienced sonographer using American College of Emergency Physicians grading criteria (ACEP ≥3). Success rates were compared between groups using mixed-effects logistic regression accounting for clustering by standardized patient. Results A total of 60, 73, and 139 standardized patients were scanned by Groups 1a, 1b, and 2, respectively. Success rates were 37.9% (226/600) for Group 1a, 27.9% (204/730) for Group 1b, and 10.2% (142/1390) for Group 2. Medical students assisted by AI demonstrated significantly higher odds of successful acquisition compared with Group 2. The odds of success were 5.6-fold higher with in-person training (Group 1a; OR 5.6, 95% CI 4.7–6.6) and 3.6-fold higher with fully remote training (Group 1b; OR 3.6, 95% CI 3.0–4.1). Group 1a achieved higher performance than Group 1b (OR 1.58, 95% CI 1.3–2.0), indicating a graded training effect. Conclusion AI-guided echocardiography, particularly when combined with in-person instruction, significantly improves transthoracic echocardiographic image acquisition performance among novice learners compared with traditional training alone. These findings support the integration of AI-assisted, real-time acquisition support systems into cardiovascular imaging education to enhance technical standardization and training efficiency.