← Back to Home

AI helps novices capture interpretable heart ultrasound images, but quality gaps remain in community use

Eight studies show that real-time AI guidance can help nonexpert operators obtain images that meet quality standards, but results diverge markedly outside settings with more comprehensive training and supervision. Whether capturing these images translates into reliable diagnosis and patient benefit remains to be verified.

By SURL BioNews

The first hurdle in obtaining clear heart ultrasound images often lies with the hand holding the probe. Even when equipment reaches the patient, the operator still needs to find the right position and angle. A meta-analysis published on November 16 in BMC Cardiovascular Disorders suggests that real-time AI guidance could help people without formal training in cardiac ultrasound obtain interpretable images. Whether those results extend to routine community services, however, depends on additional support measures.

The analysis combined eight studies involving human patients, covering 74 novice or nonexpert operators, five AI systems, and 2,365 observations. These tools provide feedback on probe positioning and image quality during scanning to help operators acquire images. The primary analysis estimated that 96.3% met the quality thresholds predefined by each study, with a 95% confidence interval of 85.5% to 99.3%. Differences between studies were substantial, so this figure cannot be directly treated as the expected success rate for any new service.

The gap was more pronounced when deployment approaches were examined separately. Six studies conducted under controlled conditions or with clearly defined procedures had a pooled image quality pass rate of 98.0%. In two decentralized community studies, the rate was 63.7%, with a wide 95% confidence interval of 20.6% to 92.6%. Reports on the same findings by Bioengineer.org and Unite.AI also highlighted this gap: feasibility in a carefully organized research setting does not mean the same quality will automatically be maintained as services expand.

However, this comparison was an exploratory analysis conducted after the primary analysis had been completed. The community studies differed from the other studies in supervision, training, operator backgrounds, and patient selection, so the results cannot be attributed to the setting alone. With only two community studies, 63.7% is better viewed as a description of the current evidence than as a fixed ceiling for community use.

Definitions of “meeting quality standards” also varied. Five studies used independent experts to assess images, while the other three used whether software could successfully generate a left ventricular ejection fraction value as their criterion. This value measures the heart’s pumping function, but successfully producing a number alone does not establish that all clinical questions can be reliably assessed. Moreover, the 2,365 observations did not all come from different patients: one study counted repeat scans, while another counted individual images. This distinction must be retained when interpreting the scale of the data.

The analysis supports the feasibility of AI-assisted image acquisition by novices. Because almost all included studies used single-arm designs, it cannot yet clearly answer how much operators improve compared with working without AI assistance. Images meeting study-defined thresholds do not establish diagnostic accuracy, and the current data have not demonstrated that this guidance improves subsequent treatment decisions or patient health outcomes.

To make cardiac testing services available in more places, the next step must go beyond whether software can guide a probe to establishing how training, expert support, and image quality checks work together. The researchers identified these as priorities for wider deployment and called for comparative studies linking image acquisition quality, diagnostic accuracy, and clinical benefit. AI has shown the potential to help more people capture usable images. Whether that potential can become a consistent capacity to deliver care still requires evidence more closely reflecting routine services.

References

  1. BMC Cardiovascular Disorders
  2. Bioengineer.org
  3. Unite.AI