Child caries risk dynamic early warning method based on multi-modal data fusion

Abstract

This invention proposes a dynamic early warning method for childhood dental caries risk based on multimodal data fusion, relating to the field of image processing technology. The method includes the following steps: Step 1, acquiring oral examination images, saliva biochemical test indicators, and behavioral record data of the target child; Step 2, constructing a dual-branch Kolmogorov-Arnold network comprising tooth position branches, global branches, and a fusion layer. The tooth position branches receive tooth position-level topological feature vectors, the global branches receive child-level global feature vectors, and the fusion layer outputs a predicted dental caries risk value; Step 3, dividing the tooth position samples in the calibration subset into positive and negative calibration groups according to their actual labels, and determining the conformal prediction thresholds for positive and negative groups respectively. This invention achieves a complete closed loop from multimodal data acquisition to dynamic hierarchical early warning, providing reliable technical support for the early identification and precise intervention of childhood dental caries.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a dynamic early warning method for childhood dental caries risk based on multimodal data fusion. Background Technology

[0002] Dental caries is one of the most common chronic oral diseases in children worldwide. Epidemiological data shows that the prevalence of dental caries in children under 5 years old is close to 50%. Primary tooth caries not only causes irreversible damage to the hard tissues of teeth but can also affect children's chewing function, nutrient intake, maxillofacial development, and even mental health. Because the progression of primary tooth caries is significantly faster than that of permanent tooth caries—from early enamel demineralization to deep dentin caries often taking only a few months—early identification of high-risk children and intervention before lesions form are crucial for reducing the burden of dental caries in children.

[0003] In recent years, machine learning technology has been increasingly explored in the field of dental caries risk prediction. Existing studies have employed algorithms such as random forests, gradient boosting decision trees, support vector machines, and logistic regression to construct caries prediction models based on questionnaire data and clinical examination indicators, outperforming traditional assessment tools in terms of accuracy, sensitivity, and specificity. Furthermore, deep learning methods have been introduced into the field of dental caries image analysis, with convolutional neural networks demonstrating high accuracy in detecting caries areas in dental X-rays and intraoral photographs. Some studies have further proposed incorporating multi-dimensional information such as salivary biochemical indicators, microbiome data, and genetic markers into the prediction model to improve the accuracy of risk assessment.

[0004] However, existing machine learning-based methods for predicting dental caries risk still have the following shortcomings. First, when using oral examination images, existing methods mainly extract visual features such as grayscale, texture, and local morphology through convolutional neural networks, failing to characterize the spatial distribution patterns of caries and demineralized areas on the tooth surface at the topological level. This includes global topological features such as the number of lesion areas, their dispersion, the distance relationships between areas, and the scale information of boundary depressions—information that is crucial for assessing caries progression trends but has long been neglected. Second, when fusing multimodal data, existing methods typically employ simple feature concatenation or weighted voting, lacking fine-grained modeling of the nonlinear interactions between different modalities, particularly lacking a mechanism for effectively coupling local susceptibility at a single tooth position with the child's overall oral environment. Third, the output of existing methods is usually a point estimate probability value or a single classification label, lacking quantification and statistical calibration of prediction uncertainty. This fails to clearly inform clinicians of the reliability of the prediction results when data is insufficient or the model's judgment is ambiguous, resulting in a lack of a safety margin at the probabilistic level for clinical decision-making. Fourth, most existing methods rely on static risk assessment based on single-point cross-sectional data, lacking a dynamic tracking and status update mechanism for risk levels across review points. They cannot adaptively adjust intervention strategies and review frequencies according to risk change trends, making it difficult to achieve true dynamic early warning and vertical management. Summary of the Invention

[0005] The main objective of this invention is to provide a dynamic early warning method for childhood dental caries risk based on multimodal data fusion. It realizes a complete closed loop from multimodal data acquisition, topological feature extraction, nonlinear risk mapping, statistical calibration prediction to dynamic hierarchical early warning, providing reliable technical support for the early identification and precise intervention of childhood dental caries.

[0006] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0007] A dynamic early warning method for childhood dental caries risk based on multimodal data fusion, integrated into the storage media controller layer, includes the following steps:

[0008] Step 1: Collect oral examination images, salivary biochemical test indicators, and behavioral record data of the target children; perform segmentation on the oral examination images of each deciduous tooth and extract the lesion boundary sampling point set; perform persistent homology analysis on the lesion boundary sampling point set to obtain a persistent pair set; convert the persistent pair set into a tooth position-level topological feature vector; normalize the salivary biochemical test indicators and behavioral record data and then concatenate them to form a child-level global feature vector.

[0009] Step 2: Construct a dual-branch Kolmogorov-Arnold network containing tooth position branches, global branches, and a fusion layer. The tooth position branches receive tooth position-level topological feature vectors, the global branches receive child-level global feature vectors, and the fusion layer outputs caries risk prediction values. Divide the children's sample dataset into training subsets and calibration subsets and train the dual-branch Kolmogorov-Arnold network.

[0010] Step 3: Divide the tooth position samples in the calibration subset into positive calibration group and negative calibration group according to the actual labeling, and determine the positive conformity prediction threshold and the negative conformity prediction threshold respectively; generate a prediction label set for each deciduous tooth of the target child based on the positive conformity prediction threshold and the negative conformity prediction threshold, and determine the risk level of the current examination based on the prediction label set of all deciduous teeth; repeat steps 1 to 3 in subsequent re-examinations and compare the risk level of the current re-examination with the risk level of the previous examination to perform dynamic early warning status updates.

[0011] Furthermore, the oral examination images are images of all deciduous teeth taken by an intraoral camera; the salivary biochemical test indicators include at least the salivary pH value, salivary buffering capacity value, and Streptococcus mutans concentration value; the behavioral record data includes at least the daily frequency of sugary food intake, the frequency of fluoride use, and the frequency of nighttime feeding or snacks.

[0012] Furthermore, the segmentation performed on the oral examination image in step 1 is semantic segmentation, which is used to extract binary mask images of carious and demineralized areas. The process of extracting the lesion boundary sampling point set from the binary mask image includes: performing morphological opening operations on the binary mask image to remove isolated noise connected components with an area smaller than a preset pixel number threshold, performing morphological closing operations to fill small cavities inside the lesion area to obtain a smooth binary mask image; extracting the boundary pixel coordinates of all foreground areas in the smooth binary mask image, and downsampling along the boundary contour at equal arc length intervals to retain a fixed number of sampling points for each deciduous tooth, forming a lesion boundary sampling point set.

[0013] Furthermore, the persistent homology analysis process includes: taking each sampling point in the lesion boundary sampling point set as the center, setting a radius threshold that starts from 0 and gradually increases with a fixed step size; at each radius threshold, connecting an edge between two sampling points whose Euclidean distance does not exceed the current radius threshold; constructing a Vietoris-Rips simple complex from all sampling points, connecting edges, and the triangular facets they enclose; tracking the birth and death values ​​of 0-dimensional topological features (i.e., connected components) and 1-dimensional topological features (i.e., ring structures); and pairing all birth and death values ​​to form a persistent pairing set.

[0014] Furthermore, the process of converting the persistent pair set into a tooth position-level topological feature vector includes: establishing a two-dimensional grid, with the horizontal axis covering the range of birth values ​​and the vertical axis covering the range of duration, i.e., the range of the difference between the death value and the birth value; mapping each persistent pair to the corresponding coordinate position in the two-dimensional grid and placing a Gaussian kernel function with a standard deviation fixed to the side length of the grid cell at the corresponding position; accumulating the Gaussian kernel function values ​​corresponding to all persistent pairs on each grid cell to obtain the pixel value of each grid cell, forming a persistent image; and expanding the persistent image of the 0-dimensional topological feature and the persistent image of the 1-dimensional topological feature row by row and concatenating them end to end to form a tooth position-level topological feature vector.

[0015] Furthermore, the tooth position branch includes an input layer, a first hidden layer, and a tooth position branch output layer, while the global branch includes an input layer, a first hidden layer, and a global branch output layer. The fusion layer receives the fusion vector formed by concatenating the output of the tooth position branch output layer and the output of the global branch output layer, and includes a hidden layer and one output node. The output value of the output node is mapped to the 0 to 1 interval by the Sigmoid function and then used as the caries risk prediction value.

[0016] Furthermore, in the dual-branch Kolmogorov-Arnold network, the B-spline activation function configured on each connecting edge between adjacent nodes is a third-order B-spline activation function. Each third-order B-spline activation function is defined by multiple control points uniformly distributed along the input value domain of the connecting edge, and the ordinate value of each control point is a learnable parameter. When the output value of the previous layer node is input along the connecting edge, the third-order B-spline activation function on the connecting edge performs a third-order B-spline interpolation operation on the input value and outputs a scalar value. The next layer node sums the scalar values ​​output by all the incoming edges pointing to itself as its own output value.

[0017] Furthermore, when dividing the children's sample dataset into training and calibration subsets, the division is performed on an individual child basis, with all tooth position samples of the same child being grouped into the same subset. The training subset is used to update the control point ordinate values ​​of the B-spline activation function on all connected edges in the dual-branch Kolmogorov-Arnold network using the backpropagation algorithm until the loss converges.

[0018] Furthermore, the process of determining the positive and negative conformity-preserving prediction thresholds includes: for each sample in the positive calibration group, subtracting the predicted caries risk value from 1 to obtain the positive inconsistency score; for each sample in the negative calibration group, using the predicted caries risk value as the negative inconsistency score; arranging all positive inconsistency scores in ascending order and taking the percentile value corresponding to the preset confidence level as the positive conformity-preserving prediction threshold; and arranging all negative inconsistency scores in ascending order and taking the percentile value corresponding to the preset confidence level as the negative conformity-preserving prediction threshold. The process of generating a prediction label set for each primary tooth of the target child includes: subtracting the predicted caries risk value of each primary tooth from 1 to obtain the positive hypothesis inconsistency score; when the positive hypothesis inconsistency score does not exceed the positive conformity-preserving prediction threshold, the caries occurrence label is included in the prediction label set; and using the predicted caries risk value of each primary tooth as the negative hypothesis inconsistency score; when the negative hypothesis inconsistency score does not exceed the negative conformity-preserving prediction threshold, the oral health label is included in the prediction label set.

[0019] Furthermore, the rules for determining the risk level of a given examination based on the predicted label set of all deciduous teeth are as follows: when the predicted label set of at least one deciduous tooth contains only the caries occurrence label, the risk level of the examination is high; when the predicted label set of all deciduous teeth contains oral health labels and at least one deciduous tooth contains both caries occurrence and oral health labels, the risk level of the examination is medium; when the predicted label set of all deciduous teeth contains only oral health labels, the risk level of the examination is low. The rules for updating the dynamic warning status are as follows: when the risk level of the current follow-up examination increases compared to the previous examination, a consultation reminder and enhanced fluoride intervention process are triggered, and the follow-up interval is shortened; when the risk level of the current follow-up examination is the same as the previous examination, the current intervention plan and follow-up interval are maintained; when the risk level of the current follow-up examination decreases compared to the previous examination and two consecutive follow-up examinations maintain the decreased risk level, the follow-up interval is extended and the intervention intensity is reduced.

[0020] The present invention has the following beneficial effects: The present invention uses the set of lesion boundary sampling points of caries and demineralized areas in oral examination images as input for persistent homology analysis. By tracking the birth and death evolution process of connected components and ring structures in Vietoris-Rips simple complexes under different radius thresholds, the tooth position-level topological feature vector encoding the topological structure of the spatial distribution of lesions is extracted. This vector can characterize the number, dispersion, inter-regional distance relationship, and boundary concavity scale of lesion areas on the same tooth surface, etc. These information cannot be captured by traditional image features based on area, gray level, or texture, enabling the method to distinguish lesion distribution patterns with drastically different clinical significance. This invention employs a dual-branch Kolmogorov-Arnold network to transform local topological features at the tooth position level and global biochemical and behavioral features at the child level through independent branches, before interacting in the fusion layer. This avoids the numerical suppression of low-dimensional global features by high-dimensional topological features, enabling effective coupling of risk information at two different semantic levels in an equivalence representation space. Simultaneously, the independently configured B-spline activation function on each connection edge can adaptively learn the optimal nonlinear transformation shape on each pathway, offering greater flexibility in function approximation compared to traditional multilayer perceptrons with fixed activation functions. This invention introduces a label-conditional conformal prediction framework, calibrating inconsistency score thresholds based on positive and negative labels. This generates a set of predicted labels with statistically guaranteed coverage for each deciduous tooth, rather than a single hard-decision label. The method can explicitly express the model's confidence in each tooth position's determination by the number of labels in the predicted label set. When the model is not sufficiently confident, the predicted label set contains two labels simultaneously, alerting clinicians that the tooth position requires further attention, effectively reducing the risk of missed diagnoses due to model overconfidence. Furthermore, this invention achieves dynamic early warning status updates across time points by repeatedly executing the complete collection, prediction, and calibration process in subsequent re-examinations and comparing the risk level of the current re-examination with that of the previous examination. When the risk increases, enhanced intervention is triggered immediately, and when the risk decreases, intervention can only be relaxed after multiple consecutive confirmations. This asymmetric status transfer strategy aligns with the clinical principle of early detection and early intervention in childhood caries prevention, avoiding the risk of caries recurrence caused by premature removal of protective measures due to occasional improvement. Attached Figure Description

[0021] Figure 1 A schematic diagram of the Vietoris-Rips simple complex evolution process under different radius thresholds in the persistent homology analysis provided in this embodiment of the invention; Figure 2 A schematic diagram of the overall architecture of a dual-branch Kolmogorov-Arnold network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the longitudinal evolution of the risk level of a target child and the dynamic early warning status update process during six consecutive examinations, as provided in an embodiment of the present invention. Detailed Implementation

[0022] A dynamic early warning method for childhood dental caries risk based on multimodal data fusion, integrated into the storage media controller layer, includes the following steps:

[0023] Step 1: Collect oral examination images, salivary biochemical test indicators, and behavioral record data of the target children; perform segmentation on the oral examination images of each deciduous tooth and extract the lesion boundary sampling point set; perform persistent homology analysis on the lesion boundary sampling point set to obtain a persistent pair set; convert the persistent pair set into a tooth position-level topological feature vector; normalize the salivary biochemical test indicators and behavioral record data and then concatenate them to form a child-level global feature vector.

[0024] Step 2: Construct a dual-branch Kolmogorov-Arnold network containing tooth position branches, global branches, and a fusion layer. The tooth position branches receive tooth position-level topological feature vectors, the global branches receive child-level global feature vectors, and the fusion layer outputs caries risk prediction values. Divide the children's sample dataset into training subsets and calibration subsets and train the dual-branch Kolmogorov-Arnold network.

[0025] Step 3: Divide the tooth position samples in the calibration subset into positive calibration group and negative calibration group according to the actual labeling, and determine the positive conformity prediction threshold and the negative conformity prediction threshold respectively; generate a prediction label set for each deciduous tooth of the target child based on the positive conformity prediction threshold and the negative conformity prediction threshold, and determine the risk level of the current examination based on the prediction label set of all deciduous teeth; repeat steps 1 to 3 in subsequent re-examinations and compare the risk level of the current re-examination with the risk level of the previous examination to perform dynamic early warning status updates.

[0026] In a specific implementation scenario, the target children are individuals aged 3 to 12 years in the primary dentition or mixed dentition stage. The first step is to collect data from three modalities. Oral examination images are obtained using an intraoral camera under standardized lighting conditions. Images are taken of the surface of all primary teeth in the target child's mouth, with at least three views (buccal, lingual, and occlusal) for each tooth. The image resolution is no less than 1920 x 1080 pixels, and the color space is RGB3 channels. During the imaging process, a working distance of 10 to 15 mm is maintained between the intraoral camera and the tooth surface. A white LED ring light with a color temperature of 5500K is used to eliminate shadow interference. For younger children with lower cooperation, a panoramic intraoral radiography mode can be used to acquire the entire dental arch image at once, and then cropped into individual images of each primary tooth according to tooth position. Saliva biochemical indicators are collected under non-stimulatory conditions. The target child has not eaten, drunk, or brushed their teeth for at least one hour before collection. At least 1 ml of whole saliva sample is collected using the passive drip method. After the collected saliva samples are sent to the testing device, the saliva pH, saliva buffering capacity, and Streptococcus mutans concentration are measured. Saliva pH is measured using a miniature glass electrode pH meter with an accuracy of 0.01 pH units. Saliva buffering capacity is measured by adding 0.01 mol / L hydrochloric acid solution to the saliva sample and recording the volume of acid required for the pH to drop to a preset endpoint. Streptococcus mutans concentration is measured using a plate count method or real-time quantitative polymerase chain reaction (qPCR), with units of colony-forming units per milliliter. In optional embodiments, the saliva biochemical indicators may further include additional indicators such as lactate concentration, calcium ion concentration, and immunoglobulin A level to enrich the information dimensions of the biochemical characteristics. Behavioral record data is collected through a standardized electronic questionnaire completed by parents, including at least three items: daily frequency of sugary food intake, frequency of fluoride use, and frequency of nighttime feedings or snacks. The daily sugar intake frequency record tracks the average number of times the target child consumes foods or drinks containing free sugars per day over the past 7 days; the fluoride use frequency record tracks the number of times the target child brushes their teeth with fluoride toothpaste per week and whether they receive regular professional fluoride varnish treatments; and the nighttime feeding or snack frequency record tracks whether the target child bottle-feeds or eats after falling asleep at night and the frequency of such behavior per week. In optional implementations, the behavioral record data can also be expanded to include dimensions such as brushing duration, fluoride content in household drinking water, and use of oral hygiene aids.

[0027] After collecting data from three modalities, semantic segmentation was performed on the oral examination images of each deciduous tooth. The goal of semantic segmentation is to assign each pixel in the image to one of three labels: carious area, demineralized area, or healthy tooth tissue area. In this embodiment, a pre-trained deep convolutional neural network is used as the segmentation model. The input is an RGB tooth surface image of a single deciduous tooth, and the output is a pixel-by-pixel label map with the same spatial resolution as the input image. Carious areas typically appear as dark-colored cavity-like structures or irregularly edged brown-black patches in the image, while demineralized areas appear as opaque chalky white spots or stripes on the tooth surface. The segmentation model separates these two types of diseased areas from the background of healthy tooth tissue, forming a binary mask image, in which pixels of carious and demineralized areas are labeled with a foreground value of 1, and pixels of healthy tooth tissue are labeled with a background value of 0. In an optional implementation, the segmentation model can be any one of the U-Net architecture, DeepLab architecture, or SegmentAnything architecture. The training data is a dataset of images of deciduous teeth surface annotated pixel by pixel by dental professionals, with no less than 2,000 annotated images.

[0028] The binary mask output from semantic segmentation may contain small pseudo-lesion regions caused by image noise, reflection artifacts, or saliva occlusion, as well as holes within lesion regions. This noise can interfere with the accuracy of subsequent topological feature extraction. Therefore, a two-step cleaning operation—morphological opening and morphological closing—is performed on the binary mask sequentially. The morphological opening operation first performs erosion followed by dilation, removing isolated noisy connected components with an area smaller than a preset pixel threshold. In this embodiment, the preset pixel threshold is set to 50 pixels, meaning isolated foreground regions with an area less than 50 pixels are identified as noise and removed from the mask. The morphological closing operation first performs dilation followed by erosion, filling in the small holes within lesion regions, making the boundaries of the foreground region more continuous and smooth. Both morphological operations use a 5x5 pixel circular structuring element. After this processing, a smooth binary mask is obtained, eliminating boundary noise and internal holes in the foreground region. In an optional implementation, the preset pixel threshold can be adjusted according to the actual resolution of the oral examination image. In a high-definition image with a resolution of 3840 x 2160 pixels, the threshold can be increased to 200 pixels, and the size of the structural element can also be adjusted accordingly to 7 x 7 pixels or 9 x 9 pixels.

[0029] Next, the lesion boundary sampling point set is extracted from the smoothed binary mask image. Specifically, the entire foreground region in the smoothed binary mask image is traversed, and for each connected foreground region, a contour tracing algorithm is used to extract the coordinates of all pixels on its outer boundary, resulting in one or more closed boundary contour curves. Since the lesion extent varies significantly among different deciduous teeth, directly using all boundary pixels as input for subsequent persistent coherence analysis would lead to excessively large differences in the point set size between different tooth positions, affecting the comparability of topological features. Therefore, downsampling is performed along each boundary contour at equal arc length intervals, so that each deciduous tooth ultimately retains a fixed number of sampling points. In this embodiment, 128 sampling points are retained for each deciduous tooth. The equal arc length downsampling is implemented as follows: first, the total arc length of the boundary contour is calculated, then the total arc length is divided into 128 segments, and the nearest neighbor boundary pixel is taken as the sampling point at the beginning of each arc segment. When there is no foreground region in the smooth binary mask image of a primary tooth, meaning no caries or demineralization lesions are detected at that tooth position, all 128 sampling points at that tooth position are set as the origin. This means that the persistence time of all topological features generated by subsequent persistent cohomology analysis is 0, and the corresponding tooth-level topological feature vector is a zero vector, representing a healthy, lesion-free topological state. After the above processing, each primary tooth corresponds to a set of lesion boundary sampling points containing 128 two-dimensional coordinate points. In optional implementations, the fixed number of sampling points can be set to 64, 256, or 512. A larger number of sampling points preserves more boundary details but also increases the computational cost.

[0030] Persistent homology analysis is performed on the lesion boundary sampling point set for each deciduous tooth. Persistent homology is a topological data analysis method whose core idea is to capture the shape features of the data by observing the connection structure of the point set at different scales. The specific execution process is as follows: Taking each of the 128 sampling points in the lesion boundary sampling point set as the center, a radius threshold is defined, starting from 0 and gradually increasing with a fixed step size. In this embodiment, the fixed step size is set to 0.5 pixels. Under each radius threshold, the Euclidean distance between any two sampling points is calculated. When the Euclidean distance between two sampling points does not exceed the current radius threshold, an edge is connected between the two sampling points. Further, when there are connecting edges between every pair of 3 sampling points, these 3 sampling points and their 3 connecting edges form a triangular facet. The Vietoris-Rips simplex under the current radius threshold is constructed by all sampling points, all connecting edges established under the current radius threshold, and all triangular facets formed. Vietoris-Rips simple complexes are composite topological objects determined purely by the distance relationships between points, independent of the absolute position of the sampling points in the plane. Therefore, they have rotation and translation invariance, a property that makes the topological features unaffected by small changes in the angle of the tooth surface image.

[0031] As the radius threshold gradually increases from 0, the connectivity in the Vietoris-Rips simple complex becomes denser, and its topology evolves accordingly. Persistent homology analysis tracks the birth and death changes of 0-dimensional and 1-dimensional topological features during this evolution. The 0-dimensional topological features are connected components, reflecting the number of independent clusters in the point set. When the radius threshold is 0, the 128 sampling points are isolated, forming 128 independent connected components. As the radius threshold increases, closer sampling points establish connections first, and the connected components containing two sampling points merge into one. The radius threshold corresponding to the first appearance of each connected component is recorded as the birth value, and the radius threshold corresponding to its disappearance due to merging is recorded as the death value. The difference between the death value and the birth value is called the duration. Connected components with longer durations correspond to larger spatial intervals on the lesion boundary, clinically reflecting the distance relationship between multiple scattered lesions on the same tooth surface; connected components with extremely short durations usually correspond to small gaps between closely adjacent sampling points on the boundary contour, belonging to shape noise. The 1D topological feature is a ring structure, reflecting the closed loops formed in the connection relationships of point sets. During the process of gradually establishing connecting edges between sampling points on the boundary contour, a ring structure is generated when a connecting edge forms a closed loop. As the radius threshold continues to increase, the area enclosed by the closed loop is gradually filled by triangular facets; the ring structure disappears when the enclosed area is completely filled. The birth and death values ​​of the ring structure are recorded in the same way. The duration of the ring structure reflects the scale of the concave or pore-like morphological features on the lesion boundary; rings with longer durations correspond to larger boundary concavities, which clinically often indicate that the caries lesion has formed a visible cavity structure.

[0032] The birth and death values ​​of all 0-dimensional and 1-dimensional topological features are paired to form a persistent pairing set. Each element in the persistent pairing set is an ordered pair containing a birth value and a death value. Taking a primary tooth as an example, assuming that persistent homology analysis identifies 128 0-dimensional topological features (because there are 128 independent connected components initially) and several 1-dimensional topological features, then the persistent pairing set contains 128 0-dimensional pairs and the corresponding number of 1-dimensional pairs.

[0033] refer to Figure 1 , Figure 1 It contains 4 sub-images, which correspond to the radius thresholds in order from left to right and top to bottom. The simple complex state is selected using four increasing values: 0, 0.55, 1.1, and 2.8. The four sub-images share the same set of lesion boundary sampling points. These sampling points are distributed in three spatial clusters on the two-dimensional plane. The left cluster contains 18 sampling points arranged roughly in an arc, forming a large annular region. The right cluster contains 12 sampling points, forming a smaller annular region. The lower middle cluster contains only 3 sampling points, which are relatively close together. The radius threshold is used in the first sub-image. All 33 sampling points are isolated, with no connecting edges between them. In this case, the Vietoris-Rips simplex has 33 independent connected components and 0 loops. Each sampling point is marked with a solid circle. In the second subgraph, the radius threshold is increased to... Within the same cluster, adjacent sampling points with a distance not exceeding 0.55 are connected by edges. Sampling points within the three clusters are individually connected, but the clusters themselves are not yet connected. The neighborhood of a single sampling point is marked with a dashed circle in the figure to illustrate the geometric meaning of the connection determination. At this point, the number of connected components decreases from 33 to a few, and some closed loops begin to appear but are not yet filled by triangular faces, thus forming a ring structure. In the third sub-figure, the radius threshold continues to increase to... The connections within each cluster become denser, and the closed loop regions enclosed by the connecting edges begin to be gradually filled by triangular facets. Smaller ring structures disappear as the triangular facets fill the area, while larger ring structures remain. The neighborhood is marked with dashed circles in the figure, and this area is significantly larger than the second subfigure. The triangular facets are marked with light-colored areas. In the fourth subfigure, the radius threshold increases to... All sampling points have been connected into a single connected component. Almost all regions enclosed by closed loops have been filled with triangular facets, and the loop structures have disappeared. The entire simplicial complex appears as a highly dense connected entity. Four sub-figures, from left to right and top to bottom, visually illustrate the topological evolution of the Vietoris-Rips simplicial complex from completely discrete to partially connected and then globally connected as the radius threshold gradually increases from 0. The number of 0-dimensional topological features, i.e., the number of connected components, decreases monotonically, while the number of 1-dimensional topological features, i.e., loop structures, appears first and then disappears. The title of each sub-figure indicates the number of connected components and the number of loop structures at the current radius threshold. Persistent homology analysis captures the multi-scale topological features of the lesion boundary sampling point set by recording the birth and death values ​​of each connected component and each loop structure during this evolutionary process.

[0034] The purpose of converting the persistent pair set into a persistent image is to encode the discrete, variable-number topological pair information into a fixed-dimensional vector representation so that the subsequent network input interface can process it uniformly. The conversion process is as follows: Separate image processing is performed on the persistent pair sets for 0-dimensional and 1-dimensional topological features. Taking 0-dimensional topological features as an example, a 2D grid with a resolution of 20 rows by 20 columns is created. The horizontal axis covers the range from the minimum to the maximum value of the birth value in all 0-persistent persistent pairs, and the vertical axis covers the range from the minimum to the maximum value of the difference between the survival time (i.e., the death value and the birth value) in all 0-persistent persistent pairs. For each persistent pair in the persistent pair set, its coordinates on the horizontal axis are determined based on its birth value, and its coordinates on the vertical axis are determined based on its survival time. A 2D Gaussian kernel function is placed at these coordinate positions. The standard deviation of the Gaussian kernel function is fixed to the side length of the grid cell, which is the smaller of the total range on the horizontal axis divided by 20 or the total range on the vertical axis divided by 20. The reason for using a Gaussian kernel with a fixed standard deviation instead of counting directly within the grid cells is that the smoothing diffusion effect of the Gaussian kernel allows persistent pairs with small perturbations in coordinate position to still contribute to adjacent grid cells, thus giving the persistent image stability against small geometric perturbations in the input point set. The Gaussian kernel function values ​​corresponding to all persistent pairs are accumulated over each grid cell to obtain the pixel value of each grid cell, forming a persistent image of 20 rows by 20 columns. The same operation is performed on the 1D topological features to obtain another persistent image of 20 rows by 20 columns. In an optional implementation, the resolution of the 2D grid can be set to 10 rows by 10 columns, 25 rows by 25 columns, or 30 rows by 30 columns. Higher resolution provides a more detailed characterization of the distribution of topological features on the birth-duration plane, but the vector dimension also increases accordingly.

[0035] The persistent image of 0-dimensional topological features is unfolded row by row into a 1-dimensional vector from row 1 to row 20, resulting in a 400-dimensional sub-vector. Similarly, the persistent image of 1-dimensional topological features is unfolded row by row into a 400-dimensional sub-vector. The two sub-vectors are concatenated end-to-end to form an 800-dimensional tooth-position-level topological feature vector. This complete process from semantic segmentation to persistent image unfolding is performed on all deciduous teeth, resulting in an 800-dimensional tooth-position-level topological feature vector for each tooth. The tooth-position-level topological feature vector encodes the spatial distribution topology of caries and demineralization lesions on the surface of each deciduous tooth, including the number of lesion regions, their dispersion, the distance relationships between regions, and the scale information of boundary depressions. This information cannot be expressed by traditional image features based on area, grayscale, or texture.

[0036] Max-min normalization was performed on each value in the saliva biochemical test indicators and behavioral record data. Max-min normalization maps the raw value of each indicator to the interval between 0 and 1. The mapping method is as follows: for a given indicator, the maximum and minimum values ​​of all samples in the children's sample dataset for that indicator are taken. The raw value of that indicator for each sample is then subtracted from the minimum value, and the result is divided by the difference between the maximum and minimum values ​​to obtain the normalized value. Specifically, the normalization operation is expressed as follows: ,in This refers to the original collected value of a certain indicator. The minimum value of this indicator for all samples in the children's sample dataset. This represents the maximum value of this indicator for all samples in the children's sample dataset. These are the normalized values. This processing aims to eliminate scale inconsistencies caused by differences in dimensions and orders of magnitude between different indicators, ensuring that the salivary pH value (typical range 5.5 to 8.0) and the Streptococcus mutans concentration value (typical range...) are normalized. to Indicators with significant differences in magnitude (colony-forming units per milliliter) receive equal initial influence in subsequent networks. Six scalar values—normalized salivary pH, salivary buffering capacity, Streptococcus mutans concentration, daily frequency of sugary food intake, fluoride usage frequency, and nighttime feeding or snack frequency—are sequentially concatenated to form a 6-dimensional child-level global feature vector. In implementations where additional indicators such as lactate concentration, calcium ion concentration, and immunoglobulin A level are included in the salivary biochemical testing, the dimension of the child-level global feature vector increases to 9 dimensions or higher. The child-level global feature vector represents the overall condition of the target child's oral environment and daily habits. This information applies to all deciduous teeth of the same child, complementing the local morphological information of a single deciduous tooth represented by the tooth position-level topological feature vector. In an optional implementation, the normalization method can be replaced with z-score standardization, which involves subtracting the sample mean from each indicator and dividing by the sample standard deviation. The normalized values ​​no longer strictly fall within the 0-1 range but still eliminate dimensional differences.

[0037] After extracting the tooth position-level topological feature vectors and constructing the child-level global feature vector in step 1, the process moves to the construction and training phase of the dual-branch Kolmogorov-Arnold network. The reason for using a dual-branch architecture instead of directly concatenating the tooth position-level topological feature vectors and the child-level global feature vectors into a single network is that these two types of feature vectors differ fundamentally in semantics and data characteristics. The tooth position-level topological feature vector is an 800-dimensional high-dimensional vector, encoding the spatial topological structure information of lesions on the surface of a single deciduous tooth; the tooth position-level topological feature vectors are different for different deciduous teeth. The child-level global feature vector is a 6-dimensional low-dimensional vector, encoding the overall condition of the target child's oral biochemical environment and behavioral habits; all deciduous teeth of the same child share the same child-level global feature vector. If directly concatenated into an 806-dimensional vector and input into a single network, the high-dimensional topological features would numerically suppress the low-dimensional global features, making it difficult for the network to fully learn the risk information contained in the global features. The dual-branch architecture allows two types of features to undergo nonlinear transformations and dimensionality compression through independent network branches, and then interact with similar dimensions in the fusion layer. This enables effective coupling of information from two different semantic levels in an equivalent representation space. The core design difference between the dual-branch Kolmogorov-Arnold network and the traditional multilayer perceptron lies in this: the traditional multilayer perceptron applies a fixed activation function (such as ReLU or Sigmoid) to each node, and the network's learnability depends entirely on the linear connection weights between nodes; while the Kolmogorov-Arnold network places the learnable nonlinear transformations on the connections rather than on the nodes. Each connection edge is configured with an independent B-spline activation function, and the nodes themselves only perform summation operations without applying any fixed activation function. This design originates from the mathematical structure revealed by the Kolmogorov-Arnold representation theorem: any multivariate continuous function can be decomposed into a finite number of nested compositions and summations of univariate continuous functions. By treating the B-spline activation function on each edge as a learnable approximator of a univariate function, the entire network approximates the complex mapping relationship between multivariate inputs and scalar outputs through layer-by-layer edge activation and node summation. The B-spline function on each edge can adaptively learn the optimal nonlinear transformation shape between the input and output values ​​on that path. Different edges can learn completely different transformation curves, which gives the network extremely high flexibility in function approximation, making it particularly suitable for scenarios such as childhood caries risk prediction where the mapping relationship between input features and risk is complex and may be highly nonlinear.

[0038] The specific network structure of the tooth position branch is as follows: The input layer has 800 nodes, each receiving one of the 800 component values ​​from the tooth position-level topological feature vector. The first hidden layer has 64 nodes. The tooth position branch output layer has 16 nodes. There are a total of [number missing] nodes between the input layer and the first hidden layer. There are 1 connecting edge, and a total of 1 connecting edge between the first hidden layer and the tooth position branch output layer. There are 52,224 connection edges in total for the tooth position branch. The specific network structure of the global branch is as follows: The number of nodes in the input layer is equal to the dimension of the child-level global feature vector. In this embodiment, the child-level global feature vector is 6-dimensional, so the global branch input layer has 6 nodes. The first hidden layer has 16 nodes. The global branch output layer has 8 nodes. There are a total of There are 10 connecting edges between the first hidden layer and the global branch output layer. The global branches contain a total of 224 connecting edges. The number of layers and nodes in the global branches is significantly less than that in the tooth position branches. This matches the characteristics of the child-level global feature vector, which has a lower dimension and relatively concentrated information content, thus avoiding overfitting caused by applying excessive network capacity to the low-dimensional input. In an optional implementation, when the child-level global feature vector is expanded to include additional indicators such as lactate concentration, calcium ion concentration, and immunoglobulin A level, increasing the dimension to 9 dimensions or higher, the number of nodes in the input layer of the global branches increases accordingly, and the number of nodes in the first hidden layer can be adjusted to 32.

[0039] The fusion layer receives the 16-dimensional output from the tooth position branch output layer and the 8-dimensional output from the global branch output layer, concatenating them end-to-end to form a 24-dimensional fusion vector. The fusion layer contains one hidden layer and one output node. The hidden layer has 8 nodes, and there are [number] nodes between the 24-dimensional fusion vector and the hidden layer. There are 10 connecting edges, and a total of 100 connecting edges between the hidden layer and the output node. The fusion layer contains a total of 200 connecting edges. The fusion layer's role is to learn the interactive mapping relationship between tooth-position level topological information and child-level global information. For example, if the topological features of a deciduous tooth show multiple scattered demineralized areas on its surface (corresponding to multiple peaks of high persistence in the 0-persistence image region of the tooth-position level topological feature vector), and the target child has a high frequency of daily sugar intake, the fusion layer should output a higher risk value than a simple superposition of the two individual factors, reflecting the synergistic amplification effect between local morphological susceptibility and global behavioral risk. The output value of the output node is mapped to the 0-1 interval by the Sigmoid function and used as the caries risk prediction value. The expression for the Sigmoid function is... ,in This is the original output scalar value obtained after performing edge activation and summing with the node. It is a natural constant. This represents the mapped predicted risk value for dental caries. The closer the value is to 1, the higher the likelihood that the tooth will develop caries during the future observation period. The closer the value is to 0, the higher the likelihood that the tooth position will maintain oral health.

[0040] In the dual-branch Kolmogorov-Arnold network, all B-spline activation functions configured on the connecting edges are 3rd-order B-spline activation functions. A 3rd-order B-spline is a piecewise cubic polynomial curve that maintains the continuity of its second derivative at the junctions of adjacent polynomial segments, thus exhibiting smoothness and avoiding the non-differentiability discontinuities at zeros present in the traditional ReLU function. Each 3rd-order B-spline activation function is defined by eight control points uniformly distributed along the input range of the connecting edge. The x-coordinates of the control points are evenly spaced within the input range of the connecting edge, and the upper and lower bounds of the input range are automatically determined based on the range of input values ​​on that connecting edge using training data and fixed before training begins. The y-coordinates of each of the eight control points are unique learnable parameters for that connecting edge, randomly initialized before training using a Gaussian distribution with a mean of 0 and a standard deviation of 0.1. The eight control points, through their x-coordinates and y-coordinates, collectively define the shape of a 3rd-order B-spline curve. When the output value of a node in the current layer is input along a connecting edge, the 3rd-order B-spline activation function on the connecting edge performs 3rd-order B-spline interpolation on the input value, using the ordinates of the 8 control points as interpolation constraints. The interpolation process is as follows: based on the position of the input value on the horizontal axis, determine which two adjacent control points it falls within the interval. Use the 3rd-order B-spline basis function to perform a weighted combination of the ordinates of the 4 control points near that interval, resulting in a scalar output value. The 3rd-order B-spline basis function is defined by a recursive relation, starting from the 0th-order basis function (piecewise constant function) and iterating sequentially to the 3rd order. Each recursion is essentially a linear interpolation and interval expansion operation of the previous-order basis function. The node in the next layer sums the scalar values ​​output by all incoming edges pointing to itself; the sum is its own output value. Unlike traditional multilayer perceptrons, which use a fixed activation function after summation, nodes in Kolmogorov-Arnold networks perform only pure summation; the responsibility for nonlinear transformation is entirely handled by the B-spline activation functions on the incoming edges. In optional implementations, the B-spline order can be 2nd or 4th, and the number of control points can be 6, 10, or 12. Higher order results in a smoother curve but higher computational cost; more control points provide a more flexible curve shape but require more training data to prevent overfitting.

[0041] refer to Figure 2 , Figure 2The network topology connections of the input layer, tooth position branch, global branch, splicing operation, and fusion layer are shown from left to right. The upper half of the network is the tooth position branch, and the lower half is the global branch. The two branches merge on the right side and enter the fusion layer. The leftmost end of the tooth position branch is marked with an input arrow, next to which is labeled an 800-dimensional tooth position-level topological feature vector, indicating that the tooth position branch receives the 800-dimensional tooth position-level topological feature vector generated in step 1 as input. The tooth position branch has three columns of nodes arranged from left to right. The first column is the input layer containing 800 nodes (shown as 6 circular nodes with ellipses in the figure), the second column is the first hidden layer containing 64 nodes (shown as 5 nodes in the figure), and the third column is the output layer of the tooth position branch containing 16 nodes (shown as 4 nodes in the figure). Each pair of nodes in adjacent columns is connected by a thin line to represent a fully connected edge, and each connected edge is configured with an independent 3rd-order B-spline activation function. The input layer of the tooth position branch is labeled with 800 nodes, the first hidden layer with 64 nodes, and the output layer with 16 nodes. The global branch has a similar structure but is smaller, with the leftmost input arrow labeled as the child-level global feature vector. The global branch also has three columns of nodes arranged from left to right: the first column is the input layer with 6 nodes, the second column is the first hidden layer with 16 nodes, and the third column is the global branch output layer with 8 nodes. Each layer of the global branch is labeled with its name and number of nodes. The 16-dimensional output of the tooth position branch and the 8-dimensional output of the global branch are concatenated at the position labeled "24-dimensional concatenation," forming a 24-dimensional fusion vector. This fusion vector enters the fusion layer, which contains one hidden layer (8 nodes, shown as 3 nodes in the diagram) and one output node. Output nodes are marked with large, padded circles to distinguish them from hollow nodes. Below each output node is a label indicating that it outputs a caries risk prediction value after Sigmoid mapping. The diagram also includes annotated arrows pointing to connecting edges in the tooth position branches, with annotations indicating that each connecting edge is configured with an independent 3rd-order B-spline activation function, emphasizing the core design that distinguishes this network from traditional multilayer perceptrons. The entire diagram clearly shows the complete data flow path of the 800-dimensional tooth position-level topological feature vector and the child-level global feature vector, which undergo nonlinear transformations and dimensionality compression in independent branches, interact in the fusion layer with similar dimensions, and finally output a caries risk prediction value between 0 and 1.

[0042] In this embodiment, the total number of learnable parameters for the entire dual-branch Kolmogorov-Arnold network is calculated as follows: 52,224 connecting edges for the tooth branch, 224 connecting edges for the global branch, and 200 connecting edges for the fusion layer, totaling 52,648 connecting edges. The ordinate values ​​of the 8 control points on each connecting edge are used as learnable parameters. Therefore, the entire network has a total of... One learnable parameter.

[0043] The dataset of children's samples labeled with caries occurrence was divided into a training subset and a calibration subset. The labeling process was conducted by dental professionals based on clinical examination results and imaging evidence for each primary tooth. Teeth with actual caries were labeled positively, while teeth maintaining oral health were labeled negatively. The partitioning was performed on an individual child basis, with all tooth samples from the same child grouped into the same subset. Different tooth positions from the same child were not allowed to be scattered across the training and calibration subsets. This constraint is necessary because different tooth positions from the same child share the same child-level global feature vector, and the pathological states of different tooth positions in the same oral environment exhibit strong statistical correlations. If some tooth positions from the same child were allowed to enter the training subset and others the calibration subset, the child-specific global information learned during training would leak into the calibration subset through the shared child-level global feature vector. This would lead to overly optimistic threshold estimations in subsequent conformal prediction calibration steps, ultimately causing the coverage guarantee of the predicted label set to fail. In this embodiment, the ratio of the training subset to the calibration subset is 7:3. That is, 70% of the children participating in the study, along with all their tooth position samples, are randomly selected and assigned to the training subset, while the remaining 30% of the children, along with all their tooth position samples, are assigned to the calibration subset. In an optional embodiment, the ratio can be adjusted to 8:2 or 6:4. The specific ratio depends on the total number of individuals in the children's sample dataset. When the total number of individuals is small, the ratio of the calibration subset can be appropriately increased to ensure the stability of the threshold estimation.

[0044] The two-branch Kolmogorov-Arnold network was trained using a training subset. The training objective was to make the network's output caries risk predictions reflect the actual probability of caries development in each primary tooth as accurately as possible. The loss function used was binary cross-entropy loss, expressed as follows: ,in This represents the total number of tooth position samples in the training subset. For the first The actual labeled value of each tooth position sample (positive label is 1, negative label is 0). For the two-branch Kolmogorov-Arnold network to the first The predicted caries risk value output for each tooth position sample. The value of the binary cross-entropy loss is smaller when the predicted caries risk for teeth with actual caries is closer to 1 and the predicted caries risk for teeth with actual oral health is closer to 0. During training, the control point ordinates of all connected edges are updated using the backpropagation algorithm. The backpropagation algorithm starts from the output node, calculates the gradient of the loss function with respect to the original output value of the output node, and then backtracks layer by layer along the network structure towards the input, using the chain rule to calculate the partial derivative of the loss function with respect to the ordinate of each control point on each connected edge. The control point ordinates are updated in the negative direction of the partial derivative, and the update step size is controlled by the optimizer. In this embodiment, the Adam optimizer is used, with an initial learning rate of 0.001, a batch size of 32 tooth samples, and a maximum training epoch of 200 rounds. After each training epoch (i.e., traversing all tooth samples in the training subset once), the current binary cross-entropy loss value is calculated on the training subset. When the loss value decreases by no more than 0.0001 over 20 consecutive training cycles, the loss is considered to have converged and training is terminated. In an optional implementation, the optimizer can be replaced with stochastic gradient descent with momentum or the AdaGrad optimizer, and the initial learning rate can be gradually decayed from 0.01 to 0.00001 using a cosine annealing scheduling strategy.

[0045] During training, L2 regularization can be introduced to suppress excessively large values ​​of the control point ordinates, thereby mitigating the risk of overfitting. L2 regularization adds a penalty term to the binary cross-entropy loss. ,in The total number of ordinate values ​​for all control points in the entire network is 421,184. For the first The ordinate values ​​of each control point The regularization strength coefficient is set to 0.0001 in this embodiment. The penalty term continuously applies a constraint close to 0 to the ordinate value of the control point during training, preventing excessive oscillations in the B-spline curve and ensuring that the learned nonlinear transformation maintains a relatively smooth extrapolation behavior even outside the coverage of the training data. In optional implementations, total variation regularization can be applied to the first derivative of the B-spline curve to further constrain the complexity of the curve shape, or an early stopping strategy can be used to monitor loss changes on an independent validation set to determine the timing of training termination.

[0046] After training, the ordinate values ​​of the control points of the 3rd-order B-spline activation functions on all 52,648 connecting edges in the dual-branch Kolmogorov-Arnold network are fixed. The network can receive the 800-dimensional tooth position-level topological feature vector of any deciduous tooth and the corresponding child-level global feature vector. After the layer-by-layer edge activation and node summation operations of the tooth position branch, global branch and fusion layer, the caries risk prediction value of that tooth position is output.

[0047] After the dual-branch Kolmogorov-Arnold network is trained in step 2, the ordinate values ​​of the control points of the B-spline activation functions on all connected edges of the network are fixed, enabling it to output a caries risk prediction value for any tooth position sample. However, this caries risk prediction value is itself a point estimate and cannot directly answer the most crucial question in clinical decision-making: for a specific deciduous tooth, is it reliable enough to determine whether it "will develop caries" or "will remain healthy" based solely on this point estimate? Traditionally, a fixed threshold (e.g., 0.5) is set, with a prediction value above the threshold considered positive and below the threshold considered negative. However, this approach lacks any statistical coverage guarantee, potentially leading to numerous misjudgments in actual deployment without quantifying the probability boundary of these misjudgments. The conformal prediction framework provides a method to wrap a statistical calibration mechanism around an arbitrary point prediction model. Its output is no longer a single label but a set of predicted labels, and this set of predicted labels has a theoretically guaranteed coverage when the data commutativity condition is met—that is, the probability that the true label is included in the set of predicted labels is not lower than a pre-set confidence level. For the clinical application scenario of early warning of childhood dental caries risk, this probabilistic set prediction is safer than the traditional hard threshold judgment: when the model is not confident enough in judging a certain deciduous tooth, the prediction label set will include both the caries occurrence label and the oral health label, clearly informing the clinician that the judgment result of the tooth position is uncertain and needs further attention; rather than forcibly giving a single conclusion that may be wrong like the hard threshold judgment.

[0048] The specific execution process for conformal prediction calibration is as follows: First, all tooth position samples in the calibration subset are input one by one into the trained two-branch Kolmogorov-Arnold network to obtain the caries risk prediction value for each tooth position sample in the calibration subset. In this embodiment, it is assumed that the calibration subset contains tooth position samples of a total of 2400 deciduous teeth from 150 children. The tooth position samples in the calibration subset are divided into a positive calibration group and a negative calibration group according to their actual labels. The positive calibration group contains all tooth position samples that actually have caries, and the negative calibration group contains all tooth position samples that actually maintain oral health. It is assumed that the positive calibration group contains 360 tooth position samples and the negative calibration group contains 2040 tooth position samples. The reason for grouping by label instead of mixing all samples is that children's caries datasets usually exhibit class imbalance, with the number of tooth positions that actually have caries being far less than the number of tooth positions that maintain health. Under these unbalanced conditions, if a single threshold is calculated using all calibration samples, the resulting threshold will be dominated by the majority of negative samples, potentially leading to a severe under-coverage of the positive class. In other words, teeth that are actually prone to caries are more likely to be missed from the predicted label set, which contradicts the clinical requirement of early warning methods to "over-report rather than under-report." Calibration by label grouping ensures that both positive and negative classes independently meet the coverage requirements of the preset confidence level.

[0049] For each sample in the positive calibration group, a positive inconsistency score was calculated. The positive inconsistency score was calculated by subtracting the predicted caries risk value for that sample from 1. This can be described mathematically as follows: ,in The first in the positive calibration group Positive nonconformity score of each sample This represents the caries risk prediction value output by the two-branch Kolmogorov-Arnold network for this sample. The intuitive meaning of this definition is: if a tooth position actually has caries, the higher its caries risk prediction value (the closer to 1), the smaller 1 minus the prediction value, the lower the positive inconsistency score, indicating that the model's prediction is more consistent with the actual situation; conversely, if the model only gives a low prediction value (e.g., 0.3) for a tooth position actually having caries, the positive inconsistency score is 0.7, indicating that the model's prediction deviates significantly from the actual situation. For a specific numerical example: assuming the caries risk prediction value for a tooth position sample in the positive calibration group is 0.85, then its positive inconsistency score is... The other positive sample had a caries risk predictive value of 0.42, therefore its positive inconsistency score was... .

[0050] For each sample in the negative calibration group, a negative inconsistency score was calculated. The negative inconsistency score was calculated by directly using the sample's predicted caries risk value as the negative inconsistency score. This can be described mathematically as follows: ,in The first in the negative calibration group Negative inconsistency score of each sample This is the predicted caries risk value for this sample. The intuitive meaning of this definition is symmetrical to the positive side: if a tooth position that actually maintains oral health has a lower predicted caries risk value (closer to 0), then the negative inconsistency score is lower, and the model prediction is more consistent with the actual situation; if the model incorrectly gives a higher predicted value (such as 0.7) for a healthy tooth position, then the negative inconsistency score is 0.7, which means that the model judgment deviates significantly from the actual situation.

[0051] Next, conformal prediction thresholds were determined for both the positive and negative calibration groups. The positive inconsistency scores of all 360 samples in the positive calibration group were arranged in ascending order to form an ordered positive inconsistency score sequence. A preset confidence level was set. ,in This represents the allowable probability of missing coverage. In this embodiment, the preset confidence level is set to 0.9, i.e. This means that, in at least 90% of cases, the true label is expected to be covered by the predicted label set. In an ordered positive inconsistency score sequence, the first... The scores corresponding to each position are used as the positive class conformal prediction threshold, where The total number of samples in the positive calibration group is 360. This indicates the floor function (rounding up). Substitute the values ​​into the calculation. That is, the 325th value among the 360 ​​positive inconsistency scores arranged in ascending order is taken as the positive class conformal prediction threshold, denoted as . Assuming that the calculation in this embodiment yields... Similarly, the negative inconsistency scores of all 2040 samples in the negative calibration group are sorted from smallest to largest, and the score of the next highest score is taken. The scores corresponding to each position are used as the conformal prediction threshold for the negative class, where The total number of samples in the negative calibration group is 2040. Substitute the values ​​into the calculation. That is, the 1837th value is taken as the negative class conformal prediction threshold, denoted as Assuming that the calculation in this embodiment yields... In an optional implementation, the preset confidence level can be adjusted to 0.85 or 0.95 according to clinical needs. A higher preset confidence level results in more lenient positive and negative conformity-preserving prediction thresholds, a higher proportion of teeth in the prediction label set containing both labels, a tendency to report more uncertain results, a lower rate of missed diagnoses but a higher rate of over-warnings; conversely, a lower preset confidence level results in more single-label judgments, higher decision-making efficiency but a correspondingly increased risk of missed diagnoses.

[0052] After determining the positive and negative conformity-preserving prediction thresholds, a set of predicted labels can be generated for each primary tooth of the target child. For a specific primary tooth of the target child, the caries risk prediction value for that tooth position is first obtained by following the complete process of steps 1 and 2. Then, two independent label inclusion decisions are performed. The first decision is for the caries occurrence label: the positive hypothesis inconsistency score for that tooth position is calculated by subtracting the caries risk prediction value for that tooth position from 1. When the positive hypothesis inconsistency score does not exceed the positive conformity-preserving prediction threshold, the caries occurrence label is included in the predicted label set. Mathematically, this can be described as: calculating... ,in This is the predicted risk value for caries at this tooth position. The score for inconsistency between the positive and negative hypotheses; judgment If the caries risk prediction for that tooth position is true, the caries occurrence label is included in the predicted label set. The second determination is for the oral health label: the predicted caries risk for that tooth position is used as the negative hypothesis inconsistency score for that tooth position. If the negative hypothesis inconsistency score does not exceed the negative conformity prediction threshold, the oral health label is included in the predicted label set. Mathematically, this can be described as: calculating... ,judge If the condition is met, the oral health label will be included in the predicted label set.

[0053] Based on the two determinations above, the predicted label set for each deciduous tooth may present three scenarios. A specific numerical example illustrates the formation mechanism of these three scenarios. Assume a positive class conformal prediction threshold. Negative class conformal prediction threshold Scenario 1: If the predicted risk of caries for a certain primary tooth is 0.82, then the positive hypothesis inconsistency score is... The caries occurrence label is included in the predicted label set because the negative hypothesis inconsistency score is 0.82, which is above 0.35. Therefore, the oral health label is not included in the predicted label set. Ultimately, the predicted label set for this tooth position only includes the caries occurrence label. This means the model has a high degree of confidence in predicting caries at this tooth position. Scenario 2: The predicted caries risk of a primary tooth is 0.48, then the positive hypothesis inconsistency score is... If the caries risk predictive value for a primary tooth is no more than 0.62, the caries occurrence label is included; if the negative hypothesis inconsistency score is 0.48, exceeding 0.35, the oral health label is not included. The predicted label set for that tooth position still only includes the caries occurrence label. Scenario 3: If the caries risk predictive value for a primary tooth is 0.25, then the positive hypothesis inconsistency score is... If the caries risk score exceeds 0.62, the caries hazard label is not included; if the negative hypothesis inconsistency score is 0.25 but not exceeding 0.35, the oral health label is included. The predicted label set for this tooth position only includes the oral health label. There is also an intermediate case: if the predicted caries risk for a primary tooth is 0.55, then the positive hypothesis inconsistency score is... A caries risk score not exceeding 0.62 is considered for inclusion in the caries hazard label; a negative hypothesis inconsistency score of 0.55 or higher than 0.35 excludes the oral health label. However, if the caries risk predictive value of another primary tooth is 0.30, then the positive hypothesis inconsistency score is [not specified in the original text]. If the inconsistency score exceeds 0.62, the caries occurrence label is not included; if the negative hypothesis inconsistency score is 0.30 but not exceeding 0.35, the oral health label is included. When the predicted value falls within the range that ensures both inconsistency scores do not exceed their respective thresholds, both labels are included simultaneously. The predicted label set contains both the caries occurrence label and the oral health label, which is precisely how the model expresses uncertainty. Specifically, the condition for simultaneous inclusion of both labels is... and Established simultaneously, that is and Substituting the values ​​into this embodiment yields... and Since 0.38 is greater than 0.35, this interval does not exist in this embodiment, therefore the predicted label set for each deciduous tooth contains exactly one label. However, in the embodiment where the preset confidence level is set to 0.95, both thresholds become more lenient. and The value of increases, making Decrease As the range increases, the predicted value range that includes both labels simultaneously appears and expands, at which point the predicted label set for some tooth positions will contain both labels.

[0054] After generating predictive label sets for all primary teeth of the target child, the risk level of the current examination is determined based on the predictive label sets of all primary teeth. The risk level determination follows these rules: If at least one primary tooth of the target child has a predictive label set containing only the caries occurrence label, the risk level of the current examination is high-risk. A high-risk level means that the model determines with high confidence that at least one primary tooth will develop caries during the observation period, requiring immediate intensive intervention. If all predictive label sets of the target child's primary teeth contain oral health labels (i.e., no tooth's predictive label set contains only the caries occurrence label), and at least one primary tooth has a predictive label set containing both caries occurrence and oral health labels, the risk level of the current examination is medium-risk. A medium-risk level means that the model considers the risk of caries in some tooth positions to be uncertain; although there is no high-confidence positive determination, there is a significant potential risk, requiring increased monitoring frequency. If all predictive label sets of the target child's primary teeth contain only oral health labels, the risk level of the current examination is low-risk. A low-risk level means that the model has a high degree of confidence in the judgment that all teeth are healthy, and oral health management can be carried out at the usual pace.

[0055] Let's illustrate this with a complete case study. Assume the target child has 20 primary teeth. After completing steps 1 and 2, the predicted risk values ​​for caries in these 20 primary teeth are 0.05, 0.08, 0.12, 0.06, 0.03, 0.75, 0.88, 0.11, 0.04, 0.09, 0.22, 0.15, 0.07, 0.06, 0.03, 0.65, 0.10, 0.08, 0.04, and 0.91, respectively. In this embodiment... , Under the given conditions, the positive hypothesis inconsistency score for the 6th deciduous tooth (predicted value 0.75) was 0.25, not exceeding 0.62, therefore the caries occurrence label was included; the negative hypothesis inconsistency score was 0.75, exceeding 0.35, therefore the oral health label was not included, and the predicted label set only included the caries occurrence label. The cases for the 7th deciduous tooth (predicted value 0.88) and the 20th deciduous tooth (predicted value 0.91) were similar, with the predicted label set only including the caries occurrence label. For the 16th deciduous tooth (predicted value 0.65), the positive hypothesis inconsistency score was 0.35, not exceeding 0.62, therefore the caries occurrence label was included; the negative hypothesis inconsistency score was 0.65, exceeding 0.35, therefore the oral health label was not included, and the predicted label set only included the caries occurrence label. The predicted values ​​for the remaining 16 deciduous teeth were all below 0.38. The positive hypothesis inconsistency scores all exceeded 0.62, therefore the caries occurrence label was not included. Meanwhile, the predictive values ​​all did not exceed 0.35, therefore the oral health label was included. The predicted label sets all contained only the oral health label. Because at least one deciduous tooth (6th, 7th, 16th, and 20th teeth) had a predicted label set containing only the caries occurrence label, the risk level for that examination was determined to be high risk.

[0056] In subsequent follow-up examinations, steps 1 to 3 are repeated to obtain the risk level for each follow-up examination. At each follow-up examination, oral examination images, salivary biochemical test indicators, and behavioral record data of the target child are re-collected. Tooth position-level topological feature vectors are re-extracted for each deciduous tooth, and a new child-level global feature vector is constructed. All tooth position samples are input into a pre-trained two-branch Kolmogorov-Arnold network to obtain the caries risk prediction value for each tooth position. A prediction label set is generated for each deciduous tooth using the established positive and negative conformity-preserving prediction thresholds. The risk level for each follow-up examination is determined based on the prediction label set of all deciduous teeth. It should be noted that the two-branch Kolmogorov-Arnold network, as well as the positive and negative conformity-preserving prediction thresholds used in each follow-up examination, use the results already trained and calibrated in steps 2 and 3, and are not retrained or recalibrated for each follow-up examination. In an optional implementation, when the accumulated amount of review data reaches the preset update conditions (such as adding complete labeled data of more than 300 children), the training process in step 2 and the calibration process in step 3 can be re-executed using the expanded children's sample dataset to update the parameters and conformal prediction threshold of the dual-branch Kolmogorov-Arnold network.

[0057] Dynamic early warning status updates are implemented by comparing the risk level of the current follow-up examination with that of the previous examination. The core idea of ​​dynamic early warning status updates is to make intervention decisions not only based on the risk level of a single examination, but also to comprehensively consider the changing trend of the risk level across follow-up examination points to adjust the intervention strategy and the frequency of follow-up examinations. The results of a single examination may be affected by accidental factors, such as the target child consuming a large amount of sugary food a few days before the examination, causing a temporary deterioration in salivary indicators, or food debris on the surface of a demineralized tooth being mistakenly identified as a demineralized area. These accidental factors may cause the single risk level to deviate from the child's true long-term risk trajectory. By comparing the direction of risk level changes between two adjacent examinations, it is possible to distinguish between the true risk trend change and accidental fluctuations.

[0058] refer to Figure 3 , Figure 3The horizontal axis represents the examination time points, labeled from left to right as the initial diagnosis, first follow-up examination, second follow-up examination, third follow-up examination, fourth follow-up examination, and fifth follow-up examination, totaling six time points. The vertical axis represents the risk level, divided into three discrete levels from bottom to top, labeled as low risk, medium risk, and high risk. A stepped broken line connects the risk levels corresponding to the six examination time points in the graph, with a solid dot marking the risk level value for each examination time point. The longitudinal evolution trajectory of the target child's risk level is as follows: low risk at the initial diagnosis, rising to medium risk at the first follow-up examination, continuing to rise to high risk at the second follow-up examination, remaining high risk at the third follow-up examination, falling to medium risk at the fourth follow-up examination, and remaining medium risk at the fifth follow-up examination. Above the broken line segment between adjacent examinations, upward arrows, equal signs, or downward arrows indicate the direction of risk level change. Each examination time point is accompanied by boxed annotations indicating the corresponding dynamic early warning status update decision. The first follow-up examination, marked as an increase in risk level and a shortened follow-up interval, indicates that the shortened interval was triggered when the risk level rose from low to medium. The second follow-up examination, marked as an increase in risk level and triggering a doctor's visit reminder and enhanced fluoride intervention, indicates that the highest level of intervention response was triggered when the risk level rose from medium to high. The third follow-up examination, marked as the same risk level and maintaining the current intervention plan and follow-up interval, indicates that the existing intervention strategy remains unchanged when both examinations are at high risk. The fourth follow-up examination, marked as a decrease in risk level but still under observation (as it is the first decrease), indicates that although the risk level has decreased from high to medium, the confirmation condition of maintaining a decrease for two consecutive examinations has not yet been met, therefore the intervention is not relaxed. The fifth follow-up examination, marked as a decrease in risk level for two consecutive examinations, therefore the follow-up interval is extended and the intervention intensity is reduced, indicates that both the fourth and fifth follow-ups maintained a medium risk level, thus meeting the confirmation condition, and adjustments were made to extend the follow-up interval and reduce the intervention intensity. The entire diagram visually illustrates the core logic of the dynamic early warning status update mechanism: when the risk level rises, an enhanced response is immediately triggered; when the risk level remains the same, the status quo is maintained; and when the risk level falls, two consecutive confirmations are required before intervention can be relaxed.

[0059] The dynamic early warning status update follows these rules. When the risk level of the current follow-up examination increases compared to the previous examination (i.e., from low risk to medium risk, from low risk to high risk, or from medium risk to high risk), a visit reminder and enhanced fluoride intervention process are triggered, and the follow-up interval is shortened. The visit reminder sends an immediate notification to the parents of the target child, recommending that they take the target child to the dentist for a professional examination within 2 weeks. The enhanced fluoride intervention process includes recommending that parents increase the frequency of fluoride toothpaste use to twice daily, recommending professional fluoride varnish application, and recommending the use of fluoride mouthwash. Shortening the follow-up interval means reducing the planned interval for the next follow-up examination from the current interval to half of the original planned interval. For example, if the original planned follow-up was in 6 months, it will be shortened to 3 months. Triggering a response immediately when the risk increases, rather than waiting for confirmation, reflects the principle of "early detection and early intervention" in childhood caries prevention, because the progression of caries in primary teeth is significantly faster than that in permanent teeth, and delayed intervention may lead to rapid expansion of caries lesions.

[0060] If the risk level of the follow-up examination is the same as that of the previous examination, the current intervention plan and follow-up interval will be maintained. This means that regardless of whether the current risk level is high, medium, or low, as long as the risk level does not change between two examinations, the existing intervention intensity and follow-up frequency will not be adjusted. If the previous examination triggered an enhanced intervention, the enhanced intervention plan will be maintained for this examination; if the previous examination was under routine management, routine management will be maintained for this examination.

[0061] When the risk level of a follow-up examination decreases compared to the previous examination (i.e., from high risk to medium risk, from high risk to low risk, or from medium risk to low risk), the additional condition of maintaining the decreased risk level in two consecutive follow-up examinations is also required before extending the follow-up interval and reducing the intervention intensity. The reason for establishing a two-time confirmation mechanism, rather than immediately relaxing intervention upon the first decrease in risk level, is to eliminate false risk decreases caused by accidental improvement. For example, if a child's saliva indicators and behavioral scores temporarily improve during the first follow-up examination due to parents' short-term strict control of sugary diet, causing the risk level to decrease from high risk to medium risk, but if the high-sugar diet is subsequently resumed, the risk level may rise back to high risk in the next follow-up examination. The two-time confirmation mechanism ensures that the intervention strategy is only adjusted after the downward trend in risk level is repeatedly verified, avoiding missing the critical window for caries control due to premature relaxation of intervention. Extending the follow-up interval means adjusting the planned time for the next follow-up visit to 1.5 times the current interval. For example, if the current interval is 3 months, it will be extended to 4.5 months (rounded up to 5 months). Reducing the intensity of intervention means restoring the enhanced fluoride intervention process to routine oral hygiene guidance, or changing emergency appointment reminders to regularly sending oral health education content.

[0062] The following example illustrates the evolution of a target child's risk level across five consecutive checkups. The first checkup (initial record keeping) indicates a low-risk level, and oral health education content is delivered according to the standard 6-month cycle. The second checkup indicates a medium-risk level, an increase from the first, triggering a visit reminder and intensified fluoride intervention, with the follow-up interval shortened from 6 months to 3 months. The third checkup indicates a medium-risk level, the same as the second, maintaining the current intensive intervention plan and 3-month follow-up interval. The fourth checkup indicates a low-risk level, a decrease from the third, but since this is the first decrease in risk level, the confirmation condition of two consecutive checkups is not yet met, so the intervention strategy and follow-up interval remain unchanged, continuing with the 3-month follow-up interval and intensive intervention plan. The risk level of the 5th check-up remained low, the same as the 4th check-up. The 4th and 5th consecutive check-ups maintained the reduced low risk level, meeting the condition for two consecutive confirmations. The interval between check-ups was extended to 5 months, and the intervention plan was changed from the enhanced fluoride intervention process to the delivery of oral health education content according to the preset regular cycle.

[0063] In an optional implementation, the number of consecutive confirmations can be adjusted to 3 or 1, depending on the rigor of clinical management. For areas with a high incidence of dental caries or high-risk populations, the number of consecutive confirmations can be increased to 3 to further reduce the risk of relapse due to premature relaxation of intervention. For areas with a low incidence of dental caries and limited oral health management resources, the number of consecutive confirmations can be set to 1, meaning that the intervention downgrade is triggered upon the first decrease in caries incidence, thus reducing unnecessary occupation of medical resources. Furthermore, in an optional implementation, the adjustment coefficient for the follow-up interval can also be flexibly set according to specific scenarios; when shortened, it can be adjusted to one-third of the original planned interval instead of one-half, and when extended, it can be adjusted to twice the original planned interval instead of 1.5 times.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic early warning method for childhood dental caries risk based on multimodal data fusion, characterized in that, Includes the following steps: Step 1: Collect oral examination images, salivary biochemical test indicators, and behavioral record data of the target children; perform segmentation on the oral examination images of each deciduous tooth and extract the lesion boundary sampling point set; perform persistent homology analysis on the lesion boundary sampling point set to obtain a persistent pair set; convert the persistent pair set into a tooth position-level topological feature vector; normalize the salivary biochemical test indicators and behavioral record data and then concatenate them to form a child-level global feature vector. Step 2: Construct a dual-branch Kolmogorov-Arnold network containing tooth position branches, global branches, and a fusion layer. The tooth position branches receive tooth position-level topological feature vectors, the global branches receive child-level global feature vectors, and the fusion layer outputs caries risk prediction values. The children's sample dataset was divided into a training subset and a calibration subset, and a dual-branch Kolmogorov-Arnold network was trained. Step 3: Divide the tooth position samples in the calibration subset into positive calibration group and negative calibration group according to the actual labeling, and determine the positive conformity prediction threshold and the negative conformity prediction threshold respectively; generate a prediction label set for each deciduous tooth of the target child based on the positive conformity prediction threshold and the negative conformity prediction threshold, and determine the risk level of the current examination based on the prediction label set of all deciduous teeth; repeat steps 1 to 3 in subsequent re-examinations and compare the risk level of the current re-examination with the risk level of the previous examination to perform dynamic early warning status updates.

2. The method as described in claim 1, characterized in that, The oral examination images are images of all deciduous teeth taken by an intraoral camera; the salivary biochemical test indicators include at least the salivary pH value, salivary buffering capacity value, and Streptococcus mutans concentration value; the behavioral record data includes at least the daily frequency of sugary food intake, the frequency of fluoride use, and the frequency of nighttime feeding or snacks.

3. The method as described in claim 1, characterized in that, The segmentation performed on the oral examination image in step 1 is semantic segmentation, which is used to extract binary mask images of caries and demineralized areas; The process of extracting the lesion boundary sampling point set from the binary mask image includes: performing morphological opening operations on the binary mask image to remove isolated noise connected components with an area smaller than a preset pixel number threshold, performing morphological closing operations to fill small holes inside the lesion area to obtain a smooth binary mask image; extracting the coordinates of the boundary pixel points of all foreground regions in the smooth binary mask image, and downsampling along the boundary contour at equal arc length intervals to retain a fixed number of sampling points for each deciduous tooth, thus forming the lesion boundary sampling point set.

4. The method as described in claim 1, characterized in that, The process of persistent homology analysis includes: taking each sampling point in the lesion boundary sampling point set as the center, setting a radius threshold that starts from 0 and gradually increases with a fixed step size; under each radius threshold, connecting an edge between two sampling points whose Euclidean distance does not exceed the current radius threshold; constructing a Vietoris-Rips simple complex from all sampling points, connecting edges and the triangular facets they enclose; tracking the birth and death values ​​of 0-dimensional topological features (i.e., connected components) and 1-dimensional topological features (i.e., ring structures); and pairing all birth and death values ​​to form a persistent pairing set.

5. The method as described in claim 4, characterized in that, The process of converting a set of persistent pairs into a tooth position-level topological feature vector includes: establishing a two-dimensional grid, with the horizontal axis covering the range of birth values ​​and the vertical axis covering the range of duration, i.e., the range of the difference between the death value and the birth value; mapping each persistent pair to the corresponding coordinate position in the two-dimensional grid and placing a Gaussian kernel function with a standard deviation fixed to the side length of the grid cell at the corresponding position; accumulating the Gaussian kernel function values ​​corresponding to all persistent pairs on each grid cell to obtain the pixel value of each grid cell, forming a persistent image; and expanding the persistent image of the 0-dimensional topological feature and the persistent image of the 1-dimensional topological feature row by row and concatenating them end to end to form a tooth position-level topological feature vector.

6. The method as described in claim 1, characterized in that, The tooth position branch includes an input layer, a first hidden layer, and a tooth position branch output layer. The global branch includes an input layer, a first hidden layer, and a global branch output layer. The fusion layer receives the fusion vector formed by concatenating the output of the tooth position branch output layer and the output of the global branch output layer, and includes a hidden layer and one output node. The output value of the output node is mapped to the 0 to 1 interval by the Sigmoid function and used as the caries risk prediction value.

7. The method as described in claim 1, characterized in that, In a dual-branch Kolmogorov-Arnold network, the B-spline activation function configured on each connecting edge between adjacent nodes is a 3rd-order B-spline activation function. Each 3rd-order B-spline activation function is defined by multiple control points uniformly distributed along the input value domain of the connecting edge, and the ordinate value of each control point is a learnable parameter. When the output value of the previous layer node is input along the connecting edge, the 3rd-order B-spline activation function on the connecting edge performs a 3rd-order B-spline interpolation operation on the input value and outputs a scalar value. The next layer node sums the scalar values ​​output by all the incoming edges pointing to itself as its own output value.

8. The method as described in claim 1, characterized in that, When dividing the children's sample dataset into training and calibration subsets, the division is performed on an individual child basis, with all tooth position samples of the same child being grouped into the same subset. The training subset is used to update the control point ordinates of the B-spline activation functions on all connected edges in the dual-branch Kolmogorov-Arnold network using the backpropagation algorithm until the loss converges.

9. The method as described in claim 1, characterized in that, The process of determining the positive and negative conformity-preserving prediction thresholds includes: for each sample in the positive calibration group, subtracting the predicted caries risk value from 1 to obtain the positive inconsistency score; for each sample in the negative calibration group, using the predicted caries risk value as the negative inconsistency score; arranging all positive inconsistency scores in ascending order and taking the percentile value corresponding to the preset confidence level as the positive conformity-preserving prediction threshold; and arranging all negative inconsistency scores in ascending order and taking the percentile value corresponding to the preset confidence level as the negative conformity-preserving prediction threshold. The process of generating a prediction label set for each primary tooth of the target child includes: subtracting the predicted caries risk value of each primary tooth from 1 to obtain the positive hypothesis inconsistency score; when the positive hypothesis inconsistency score does not exceed the positive conformity-preserving prediction threshold, the caries occurrence label is included in the prediction label set; and using the predicted caries risk value of each primary tooth as the negative hypothesis inconsistency score; when the negative hypothesis inconsistency score does not exceed the negative conformity-preserving prediction threshold, the oral health label is included in the prediction label set.

10. The method as described in claim 1, characterized in that, The rule for determining the risk level of a given examination based on the predicted label set of all primary teeth is as follows: when the predicted label set of at least one primary tooth contains only caries-related labels, the risk level of that examination is high. When the predictive label set of all deciduous teeth contains oral health labels and there is at least one deciduous tooth whose predictive label set contains both caries occurrence and oral health labels, the risk level of that examination is medium risk. When the predictive label set for all deciduous teeth contains only oral health labels, the risk level of the current examination is low. The rules for updating the dynamic warning status are as follows: when the risk level of the current follow-up examination increases compared to the risk level of the previous examination, a visit reminder and enhanced fluoride intervention process are triggered, and the follow-up interval is shortened; when the risk level of the current follow-up examination is the same as the risk level of the previous examination, the current intervention plan and follow-up interval are maintained; when the risk level of the current follow-up examination decreases compared to the risk level of the previous examination and the risk level is maintained for two consecutive follow-ups, the follow-up interval is extended and the intervention intensity is reduced.

Images