Imagine you're reviewing a patient's radiographs at the end of a busy clinic. The AI software marks the images and reports no evidence of caries or significant periodontal bone loss. Would you feel reassured enough to move on, or would you still question its assessment?
As artificial intelligence continues to make its way into dental practice, one question remains at the forefront: how reliable are these systems in real-world clinical settings?
A recent pilot study sought to answer exactly that by evaluating two US Food and Drug Administration (FDA)-cleared AI-based clinical decision support systems for detecting dental caries and clinically relevant periodontal bone loss.
Unlike many previous AI studies that rely solely on expert opinion as the reference standard, the researchers used a more robust approach. They retrospectively analysed records of 90 patients treated at the Veterans Affairs Greater Los Angeles Health Care System, comparing AI findings with longitudinal clinical data collected over 6–12 months. Ground truth was established through a structured validation process involving calibrated examiners and adjudication.
The study focused on two important measures—specificity and negative predictive value (NPV)—to determine how reliably the AI systems could identify patients who were free of disease.
For caries detection, one AI system achieved a specificity of 80.17% with an NPV of 96.92%, while the second demonstrated a specificity of 83.76% and an NPV of 97.36%.
For clinically relevant periodontal bone loss, specificity exceeded 93% for both systems, with NPVs ranging from 75.20% to 80.64%.
These findings suggest that when the AI identified no evidence of disease, it was usually correct. In other words, the systems performed particularly well at ruling out disease rather than confirming its presence.
Free for members
Keep reading — it's free
The rest of this article is open to DentalReach members. Create a free account, or sign in, and every article on the site opens in full.