سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / اولین همایش بین المللی تحول خدمات سلامت: تلفیق مراقبت، درمان، بازتوانی و امیدبخشی

Accuracy and Efficiency of the RoboDent Smart Robot in Diagnosing Intraoral Lesions and Primary Patient Screening: A Prospective, Blinded, Proof-of-Concept Diagnostic Accuracy Study

نویسندگان :

Parham Farzam¹ Amirmansour Shirani² Danoush Dehghanian³ Hossein Turk⁴

1- Department of Diagnostic Sciences, School of Dentistry, Islamic Azad University, Isfahan (Khorasgan) Branch, Isfahan, Iran 2- Department of Oral and Maxillofacial Pathology, School of Dentistry, Islamic Azad University, Isfahan (Khorasgan) Branch, Isfahan, Iran 3- Department of Oral Health and Community Dentistry, School of Dentistry, Islamic Azad University, Isfahan (Khorasgan) Branch, Isfahan, Iran 4- دانشگاه آزاد نجف آباد

کلمات کلیدی :

Artificial intelligence،Diagnostic accuracy،Intraoral screening،Dental caries،Deep learning،Oral mucosal lesions،Proof-of-concept study

چکیده :

Background: Oral diseases remain among the most prevalent noncommunicable conditions worldwide, yet access to timely screening is limited outside clinical settings. RoboDent is a domestically developed, artificial-intelligence-powered kiosk that combines an intraoral camera with a convolutional neural network pipeline (ResNet50 and EfficientNet backbones) to screen for intraoral lesions without requiring a dentist to be present. Its diagnostic accuracy against a specialist reference standard has not previously been evaluated in a real clinical population. Objective: To determine the accuracy and efficiency of RoboDent in detecting dental caries, periodontal/mucosal soft-tissue lesions, and other intraoral pathology, relative to a panel of oral medicine specialists, in a prospective Proof-of-Concept (PoC) study. Methods: In this prospective, single-centre, blinded diagnostic accuracy PoC study, adults attending the Dental Faculty of Islamic Azad University, Isfahan (Khorasgan) were screened sequentially by RoboDent, by two calibrated dental students, and by a reference panel of faculty oral-medicine specialists (index test and reference standard interpreted independently and blinded to each other). Sample size (n = 200) was calculated using the Buderer method for an expected sensitivity of 90%, specificity of 87%, an assumed lesion prevalence of 20%, 95% confidence, and an absolute margin of error of 10 percentage points, inflated for 10% anticipated non-completion. Patient-level diagnostic accuracy (sensitivity, specificity, accuracy, predictive values, Cohen's kappa, area under the receiver-operating-characteristic curve [AUC]) and lesion-level multiclass performance (precision, recall, F1-score, macro-averaged across caries, periodontal/hard-tissue, and mucosal/soft-tissue categories) were computed. Screening time and a 5-point Likert user-experience score were compared across the three methods. Results: Simulated, internally consistent placeholder results (not real patient data) were generated to populate this PoC template. At the patient level (reference-standard lesion prevalence 20.0%, 40/200), RoboDent achieved a sensitivity of 90.0% (95% CI 76.9-96.0), specificity of 88.1% (95% CI 82.2-92.3), overall accuracy of 88.5% (95% CI 83.3-92.2), positive predictive value of 65.5%, negative predictive value of 97.2%, Cohen's kappa of 0.68 against the reference standard, and an AUC of 0.93 (95% CI 0.89-0.96). At the lesion level, macro-averaged precision was 87.9%, recall 89.6%, and F1-score 88.7%. RoboDent completed screening in a mean of 2.5 (SD 0.6) minutes, versus 9.0 (SD 2.1) minutes for dental students and 14.0 (SD 3.0) minutes for the specialist reference examination (all pairwise comparisons p < .001). Mean user-experience score was 4.2/5 (SD 0.6), and the image-quality failure rate was 5.0%. Conclusion: In this simulated PoC template, the RoboDent system shows the pattern of diagnostic performance and time-efficiency that would be needed to support its use as an autonomous first-line intraoral screening and triage tool. These placeholder findings illustrate the analysis plan only; a prospective clinical study with real, specialist-verified data, independent of the device-development team, is required before any accuracy claim can be made.