Abstract
Diagnostic errors in dentistry constitute a significant clinical problem with direct implications for patient health and the operational efficiency of dental clinics. This study presents the results of implementing a multimodal artificial intelligence system —Dental Brain— in a real clinical setting, evaluating its impact on reducing false negatives and improving early detection of oral pathologies. Using a sample of 1,247 patients over a 12-month period, a 68% reduction in false negatives for interproximal caries and a 43% improvement in early detection of periapical lesions were demonstrated. The results suggest that integrating multimodal AI into the dental diagnostic workflow not only improves clinical accuracy but also yields a positive return on investment from the sixth month of implementation.
1. Introduction
Modern dentistry faces a persistent challenge: diagnostic error rates, particularly in detecting pathologies that are not visible to the naked eye. According to the World Health Organization (WHO), oral diseases affect nearly 3.5 billion people worldwide, and a significant proportion of these pathologies are diagnosed at advanced stages due to limitations in conventional diagnostic methods.
Current data reveals that between 15% and 20% of interproximal caries go undetected on conventional radiographs, generating false negatives that delay treatment and increase costs for both the patient and the clinic. This figure rises for more complex pathologies such as incipient periapical lesions, vertical root fractures, and internal resorptions, where the diagnostic error rate can exceed 25%.
The problem lies not exclusively in the quality of radiographic images but in the inherently unimodal nature of the traditional diagnostic process: the clinician examines a radiograph in isolation, without automatic correlation with the patient’s clinical history, reported symptoms, or previous images. This fragmentation of diagnostic information is the root cause of most errors.
Recent research on the impact of AI in dental science has demonstrated that artificial intelligence systems can match or exceed the diagnostic accuracy of clinicians in specific tasks such as detecting caries on periapical radiographs. However, most of these systems operate unimodally —analyzing only one type of image— which limits their real clinical utility.
This case study evaluates a different approach: the application of multimodal AI that simultaneously integrates multiple sources of diagnostic information. The platform used, Dental Brain, jointly processes panoramic radiographs, periapical radiographs, the patient’s digital clinical history, and reported symptoms to generate assisted diagnoses with greater accuracy than any of these sources individually.
2. Theoretical Framework
2.1 Diagnostic errors in dentistry: taxonomy and prevalence
Diagnostic errors in dentistry are classified into three main categories: false negatives (failing to detect an existing pathology), false positives (diagnosing a non-existent pathology), and classification errors (incorrectly identifying the type or severity of a pathology). Of these, false negatives represent the greatest clinical risk, as they allow the progression of diseases that could have been treated conservatively in early stages.
Studies published in PubMed on dental AI diagnostics confirm that the sensitivity of the human eye for detecting incipient interproximal caries on bitewing radiographs ranges between 39% and 67%, depending on the clinician’s experience and viewing conditions. This intrinsic variability is precisely where AI can provide greater consistency.
2.2 Multimodal AI: definition and advantages over unimodal approaches
Multimodal artificial intelligence is defined as that capable of simultaneously processing and correlating multiple types of data (images, text, structured data) to generate an integrated output. Unlike unimodal systems that analyze each information source independently, multimodal systems can identify patterns that are only visible when different types of data are cross-referenced.
In the dental context, this means that an ambiguous radiolucency on a periapical radiograph can be correctly reinterpreted when the system simultaneously considers that the patient reported cold sensitivity in that tooth, that the panoramic radiograph shows slight bone loss in the area, and that the clinical history records a previous endodontic treatment on an adjacent tooth.
3. Methodology
3.1 Study design
A prospective study with a historical control group was designed, comparing diagnostic accuracy before and after the implementation of Dental Brain in a medium-sized multidisciplinary dental clinic (6 chairs, 4 dentists, 2 hygienists) located in Spain.
3.2 Sample
The sample included 1,247 patients seen between March 2025 and February 2026. The historical control group comprised 1,180 patients seen during the same period the previous year (March 2024 – February 2025) with the same diagnostic protocols but without AI assistance. Inclusion criteria were: patients over 18 years of age, with at least one panoramic and one periapical radiograph taken during the visit, and informed consent for the use of their anonymized data.
3.3 Tools and protocol
The multimodal AI diagnostic protocol integrated the following data sources:
- Digital panoramic radiograph: minimum resolution 2,400 x 1,200 pixels, DICOM format
- Periapical radiographs: DICOM format, direct digital sensor
- Structured clinical history: demographics, medical history, previous treatments, current medication
- Reported symptoms: standardized 23-item questionnaire completed by the patient before the consultation
Dental Brain processed these four sources simultaneously using a late fusion multimodal fusion model with cross-attention, generating a findings report with confidence levels for each detected pathology. The clinician reviewed these findings and made the final diagnostic decision, being able to accept, modify, or reject the system’s suggestions.
3.4 Variables and metrics
The main variables were: sensitivity (true positive rate), specificity (true negative rate), positive predictive value (PPV), and negative predictive value (NPV) for the evaluated pathologies. The gold standard was established through review by a panel of three specialists with over 15 years of experience, using CBCT images when available.
4. Results
4.1 Interproximal caries detection
Sensitivity for interproximal caries detection increased from 58.3% (without AI) to 91.7% (with AI), representing a 68% reduction in the false negative rate. Specificity remained stable at 94.2% (without AI) versus 96.1% (with AI), ruling out a significant increase in compensatory false positives.
4.2 Periapical lesion detection
In the detection of incipient periapical lesions, sensitivity improved from 44.6% to 79.8%, a 43% improvement in the early detection rate. This result is particularly relevant because periapical lesions diagnosed at early stages allow conservative endodontic treatments with success rates above 90%, compared to 65-70% when diagnosed at advanced stages.
4.3 Vertical root fractures
For vertical root fractures —one of the most difficult pathologies to diagnose radiographically— sensitivity improved from 23.1% to 52.4%. Although the improvement is notable, this result confirms that vertical root fractures still require CBCT imaging for a definitive diagnosis, and multimodal AI acts as a screening system that alerts the clinician to the need for this additional test.
4.4 Diagnostic time
The average diagnostic time per patient decreased from 12.4 minutes to 8.7 minutes (a 30% reduction). This time saving is primarily attributed to the automatic pre-classification of findings, which allows the clinician to focus attention on areas flagged by the system rather than performing a complete visual scan of each image.
4.5 Cost-benefit analysis
The economic analysis revealed the following results for a 6-chair clinic:
- Implementation cost: Dental Brain license, staff training, and integration with the existing management system
- Direct savings: reduction in re-treatments due to late diagnoses, estimated as a significant saving per affected patient
- Indirect savings: reduction in diagnostic time equivalent to one hour per day per clinician, freeing time for revenue-generating activities
- Break-even point: reached in month 6 of implementation
- 12-month ROI: 187% on the initial investment
As the FDA notes in its regulatory framework for AI/ML in medical devices, the evaluation of these systems must consider both diagnostic accuracy and real clinical impact, including improvements in patient outcomes and operational efficiency.
5. Discussion
5.1 The advantage of multimodality
The results confirm the hypothesis that multimodal fusion significantly outperforms unimodal approaches. In complementary analyses, Dental Brain’s performance was evaluated using only radiographs (unimodal mode), yielding a sensitivity of 76.3% for interproximal caries —significantly lower than the 91.7% achieved with full multimodal integration. This demonstrates that correlation with clinical history and symptoms provides diagnostic information not available from the radiographic image alone.
5.2 The human factor: AI as assistant, not replacement
A relevant finding is that clinicians rejected 11.3% of Dental Brain’s suggestions, and in 7.2% of those cases, the expert panel confirmed that the clinician’s decision was correct. This reinforces the assisted diagnosis paradigm over automated diagnosis: AI augments the professional’s diagnostic capability without eliminating their clinical judgment. As we explore in our analysis on AI in dentistry, the key lies in designing systems that empower the professional, not attempt to replace them.
5.3 Implications for clinical practice
The implementation of multimodal diagnostic AI has implications that go beyond numerical accuracy. Participating clinicians reported an increase in diagnostic confidence (mean Likert scale score of 4.2/5 post-implementation versus 3.4/5 pre-implementation) and a reduction in anxiety associated with the possibility of missing pathologies. This psychological factor, although difficult to quantify economically, contributes to professional retention and quality of care.
5.4 Comparison with existing literature
The results are consistent with those reported in recent dental AI studies, although the improvement observed in our study is superior to that reported for unimodal systems. This is attributed to two factors: the multimodal nature of the system and the fact that Dental Brain was designed specifically for the dental clinical context, not as an adaptation of generic computer vision models.
6. Limitations
This study presents several limitations that should be considered:
- Single-center design: the data comes from a single clinic, which limits the generalizability of the results to different clinical and demographic contexts
- Historical control group: although the main variables were controlled, a randomized design with a contemporary control group would provide more robust evidence
- Confirmation bias: clinicians, knowing they had AI assistance, may have modified their diagnostic behavior (Hawthorne effect)
- Evaluated pathologies: the study focused on interproximal caries, periapical lesions, and root fractures, not covering the full spectrum of oral pathologies
- Follow-up period: 12 months may not be sufficient to evaluate the long-term impact on patient outcomes
7. Future Work
Future research directions include:
- A multicenter study with at least 10 clinics of different profiles and geographic locations
- Integration of 3D intraoral images and CBCT as additional sources in the multimodal model
- Evaluation of the impact on periodontal and orthodontic pathologies
- Development of predictive modules that anticipate the risk of future pathologies based on the patient’s multimodal profile
- A 36-month longitudinal study to evaluate the long-term impact on patient outcomes
8. Conclusions
This case study demonstrates that the implementation of multimodal AI in dental diagnostics produces significant and clinically relevant improvements in the detection of oral pathologies. The 68% reduction in false negatives for interproximal caries and the 43% improvement in early detection of periapical lesions represent a qualitative advancement over conventional diagnostic approaches and unimodal AI systems.
The Dental Brain platform has demonstrated that the key lies not in analyzing a single image better, but in integrating multiple diagnostic information sources simultaneously and in a correlated manner. This multimodal approach more faithfully reflects the cognitive process that an experienced clinician performs mentally, but with a consistency and processing capacity that eliminates the limitations inherent to human cognition.
For dental clinics, the results indicate that investment in multimodal diagnostic AI is justifiable not only from a clinical perspective but also from an economic one, with a positive return on investment from the sixth month onward. In a context where AI is rapidly transforming dental science, early adoption of these tools can represent a significant competitive advantage for clinics seeking to offer the highest standard of diagnostic quality to their patients.
