We aimed to develop a patient-centered deep learning model capable of automatically detecting dental plaque and gingival inflammation from intraoral images. The findings demonstrate the feasibility of applying an AI model to automate the assessment of key indicators of oral health. Predictive diagnostics are increasingly recognized as a cornerstone of cost-effective prevention in chronic diseases [27], and our model contributes to this paradigm by providing a non-invasive, image-based tool for early risk recognition in periodontal health. To our knowledge, this is the first study to evaluate a DL method capable of both segmenting and classifying dental biofilm and gingival inflammation in RGB intraoral images.

The analysis demonstrated that an AI algorithm, within a supervised learning environment, could effectively identify high levels of plaque and gingivitis achieving DSC scores of 94% and 70% for plaque and gingivitis, respectively. Evidence indicates that stratifying patients into different risk categories supports tailored interventions and monitoring schedules [5, 28]. The YOLOv8Seg model outputs can be applied to identify individuals requiring intensified preventive care or more frequent professional follow-up, thereby contributing to a personalised maintenance programme.

The model demonstrated moderate performance in segmenting dental biofilm, gingival inflammation, and tooth. Quantitative analysis revealed an average mAP of 60%, with better results for tooth segmentation (posterior mAP of 77% and anterior mAP of 71%), likely due to the comparatively simpler nature of tooth segmentation. In classification tasks, the plaque index performance stood out, achieving a precision of 98%, and recall of 91%. These findings indicate that the model is highly effective in classifying areas with plaque but faces challenges in detecting signs of gingival inflammation with similar accuracy.

Classes with larger pixel-area representation and clearer anatomical boundaries, as anterior and posterior teeth, achieved the highest mAP values, whereas inflamed gingiva, the smallest and most visually heterogeneous class (Suppl. S1), exhibited lower DSC and IoU values. Importantly, despite lower overall segmentation performance in this class, the gingivitis index retained a high recall (92%), meaning the model rarely fails to detect inflammatory sites. Minimising false negatives enhances early predictive detection of inflammation, while occasional false positives reinforce preventive behaviour by prompting users to inspect and improve hygiene in flagged areas.

The similarity observed between the validation and hold-out test precision–recall curves (Fig. 4) further indicates that the dataset offered sufficient variability for the model to generalise to unseen samples. Despite class imbalance and the difficulty of detecting small-area gingival sites, the close proximity of the curves suggests that the dataset composition was adequate. Although external remains necessary for scalability, the internal consistency supports the representativeness of this dataset for early PPPM-oriented applications in oral health.

The quantification and classification of dental biofilm are part of the daily clinical routine of all oral health professionals, whether for disease risk classification, hygiene guidance, or periodontal maintenance protocols. However, these tasks require the presence of an oral health professional. In the last century, research focused on developing indices that accurately describe oral hygiene. In the early 2000 s, with the popularisation of digital photography, research shifted towards automating dental biofilm evaluation, aiming to bring reproducibility and objectivity to what remains a subjective process in practice. In recent years, the focus has shifted towards developing methodologies that leverage digital images integrated with new technologies such as artificial intelligence. The proposed AI method demonstrated performance comparable to other models [14, 29] and supports the development of targeted prevention strategies, since site-specific identification of biofilm allows more precise guidance on oral hygiene practices, providing to patients an objective visual feedback on areas requiring improved cleaning, the model reinforces behavioural modification and encourages adherence to preventive routines.

Additionally, the model consistently captured the expected biological relationship between plaque accumulation and inflammatory burden. Cases presenting high plaque index almost invariably exhibited higher predicted gingivitis index values, even in lateral views where gingival sites are smaller and more challenging to segment (Fig. 3). This reflects not only the model’s ability to detect visible plaque but also its capacity to represent clinically meaningful patterns. Such behaviour strengthens its applicability for personalised monitoring, risk stratification, and timely preventive action.

The analysis was conducted comprehensively, evaluating all present teeth, whether deciduous or permanent, rather than focusing on individual teeth. Although dental plaque can be detected in isolated tooth images using DL models, some approaches rely on disclosing agents or high-resolution images [13, 30] to enhance detection, which may limit the applicability in self-monitoring contexts. The ability to assess multiple teeth and gingival regions simultaneously using routine intraoral photographs represents an important advancement in predictive modelling, as it enables early identification of sites with higher inflammatory burden before clinical symptoms become pronounced. By providing automated, objective, and reproducible quantification of plaque and gingival inflammation without the need for adjunctive agents, the model supports preventive strategies aimed at reducing biofilm accumulation and interrupting disease progression at an early stage. Moreover, the segmentation outputs can guide personalised recommendations for oral hygiene reinforcement, with potential integration into participatory digital health monitoring frameworks that empower patients to track their own risk profile over time. Together, these elements position the model as a practical tool for operationalising predictive and preventive dental care within a PPPM framework.

Quantitative and qualitative analysis showed that the presented model was able to detect the disease areas, had limited success in capturing complete segmentation masks, which explains why disease segmentation metrics were not higher. Regarding dental biofilm, as with Andrade et al.‘s [14] study, which reported a DSC of 61% with a U-Net network, we encountered challenges in accurately delineating the biofilm boundaries, as also observed in the qualitative analysis (Fig. 4). A YOLOv8Seg detection model using intraoral images achieved a DSC score of 77% but applied stricter exclusion criteria, including plaque disclosing agents and targeting specific teeth [31, 32], factors that may limit its applicability in real-world oral hygiene self-monitoring contexts.

Annotating the 504 images consumed significant research effort, totalling 480 h of work. Besides, AI models typically require larger datasets [33]; a five-fold cross-validation approach was integrated to enhance prediction performance. Although only one examiner annotated the images, potentially introducing rater bias, a single examiner with expertise in periodontics, computer vision, and AI, the use of a single, expert examiner helped enhance the consistency and robustness of the annotations [17].

Most studies in the medical field rely on metrics such as sensitivity and specificity; however, we did not consider these measures. While sensitivity and specificity are considered gold standard metrics in diagnostic studies, AI models require more robust metrics that reflect the agreement between predicted and target classes, and account for potential imbalanced number of real positive and negative instances [24]. Metrics that use True Negative (TN) confusion matrix values were excluded due to the large number of TN predictions and background pixels in the images, which can skew metrics that rely on TN counts. The substantial or undefined number of TN predictions in segmentation and object detection tasks causes TN focused metrics to appear artificially high, despite the model’s true performance. Furthermore, as this model is designed to be used by non-expert users to identify areas of poor hygiene, it is preferable for the model to prioritise FP rather than FN. This contrasts with other medical domains, where a higher number of FN predictions is preferred to prevent overdiagnosis and overtreatment [34].

This study does not aim to replace clinical diagnosis, but rather to empower patients by facilitating early self-detection of plaque and gingival inflammation. Holistic approaches in PPPM, emphasised that AI-based image analysis should not replace comprehensive periodontal assessment. Instead, such tools should be integrated with microbial, behavioural, and socioeconomic data to generate a more complete patient profile, reinforcing their role in predictive diagnostics, targeted prevention, and personalised treatment planning [35]. Our findings align with PPPM principles recently advocated for dentistry, in which patient stratification and tailored maintenance are central to proactive care. Dental care tailored to the person, integrating risk profiles and cost-effective prevention at population level, represents a cornerstone of this paradigm [36]. By quantifying site-specific plaque and gingival status and operationalising threshold-based alerts, our approach provides actionable strata that map directly onto targeted prevention and timely referral, thereby supporting the paradigm shift towards PPPM-oriented periodontal care.

Periodontal diseases are increasingly recognised as part of a broader chronic inflammatory network with systemic implications, including metabolic, cardiovascular, and immune-mediated pathways. Both periodontal disease and dental caries share fundamental behavioural and lifestyle determinants such as oral hygiene behaviour and socioeconomic factors, supporting the interpretation of plaque control as a modifiable risk factor of systemic relevance [37]. Within this framework, AI-derived digital biomarkers extend beyond local oral disease monitoring by enabling early identification of individuals with elevated inflammatory burden. Integrating image-based outputs with microbial, metabolic, behavioural and socioeconomic data [38,39,40] supports multi-level diagnostics and deeper patient phenotyping. This systems-oriented perspective reinforces the role of AI tools in anticipatory care, where early stratification of risk can guide targeted preventive strategies and personalised intervention pathways before clinical deterioration occurs.

Although the model demonstrated capability in detecting poor oral hygiene and inadequate gingival indices, this study had limitations, and further research in this field is needed. Future studies should focus on associating intraoral images with oral diseases without relying on indices developed for clinical practice. Moreover, the use of standardised images in controlled environments (e.g., saliva and lighting conditions) could hinder real-world implementation, despite our focus on developing a lightweight model suitable for smartphone applications. Longitudinal clinical studies assessing the use of smartphone apps for oral hygiene and gingivitis self-monitoring are crucial for improving the performance of these models. It is essential to identify the areas with dental biofilm and gingival inflammation, and more importantly, demonstrate the physical proximity between biofilm and gingivitis signs, reinforcing their potential role in reducing the incidence and progression of plaque-induced gingivitis and periodontitis.

From a broader PPPM perspective, AI-driven intraoral analysis also contributes to patient phenotyping and multi-level risk stratification by capturing early markers of chronic inflammation and linking them to behavioural and structural determinants of oral health. The integration of imaging biomarkers with multi-parametric data, such as microbial profiles, lifestyle indicators, and socioeconomic risk factors, enables a holistic characterisation of patient trajectories rather than isolated disease episodes [5, 36, 38,39,40]. Such a systems-oriented approach recognises periodontal inflammation as part of a wider chronic inflammatory network, reinforcing the need for predictive modelling that identifies high-risk phenotypes before clinical deterioration. The proposed model provides a foundation for this integrative strategy by generating standardised digital biomarkers suitable for longitudinal digital health monitoring, allowing early alerts, dynamic adjustment of preventive plans, and personalised thresholds for intervention.