Multi-modal help risk assessment method based on order-preserving calibration

Abstract

The invention relates to the technical field of artificial intelligence assisted medical treatment, and discloses a multi-modal help risk assessment method based on order-preserving calibration, and the method comprises the steps: mapping the multi-modal data of a to-be-tested sample to a joint risk measurement manifold space, and carrying out the fusion; then, a clinical semantic anchor point sequence subjected to order-preserving constraint is used as a reference system, and a modal conflict vector is calculated to quantify semantic inconsistency between modals; perturbation is applied to the anchor points based on the conflict intensity to generate an anti-fact anchor point cloud, and the pairwise dominant probability of the to-be-tested sample relative to the virtual anchor points is calculated through a constructed differential partial sequence discriminator; and finally, performing statistical aggregation to obtain a partial order dominance degree sequence, and defining a dynamic risk confidence interval. According to the method, through explicit modeling of modal conflicts and construction of dynamic calibration boundaries, the problem of assessment uncertainty caused by heterogeneous data is solved, risk underestimation is effectively avoided, and an auxiliary diagnosis basis with high robustness and interpretability is provided for clinical decision making.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted medical technology, specifically to a multimodal help-seeking risk assessment method based on order-preserving calibration. Background Technology

[0002] With the deepening application of artificial intelligence technology in the field of mental health, help-seeking risk assessment based on multimodal data has become an important development direction for intelligent assisted diagnosis. Compared with single-modal analysis, multimodal fusion technology can capture an individual's behavioral characteristics from multiple dimensions such as vision, hearing, and language, and theoretically can provide a more comprehensive and objective psychological state profile, thereby assisting professionals in suicide risk screening or crisis intervention level determination.

[0003] However, existing multimodal risk assessment techniques, when processing heterogeneous data, typically employ feature cascading or attention-based fusion strategies, whose core assumption is often based on the consistency of semantic expression across modalities. In real clinical scenarios, clients often exhibit significant modal semantic conflicts. For example, a patient with "smiling depression" may display a positive facial expression, but their speech content or vocal features reveal extreme pessimism and despair. Traditional fusion algorithms often treat this inconsistency between modalities as noise or error to be eliminated, attempting to obtain a unified feature representation through weighted averaging or forced alignment. This approach not only loses the highly diagnostic pathological feature of "semantic inconsistency" but also easily leads to feature cancellation when faced with conflicting information, resulting in mediocre or erroneous predictions that fail to accurately reflect the client's true level of psychological crisis.

[0004] Furthermore, most existing risk assessment models model it as a standard classification or regression problem, with outputs typically being discrete risk category labels or single risk scores. This point-like deterministic output ignores the inherent ordinal relationships of risk levels and makes it difficult to quantify the uncertainty of the model when faced with complex inputs. When the modal conflicts of the input data are severe, existing models still tend to give a single prediction with high confidence. This "blindly confident" judgment mechanism lacks the ability to dynamically define risk boundaries and cannot provide clinicians with a reference range for the reliability of diagnostic results, easily leading to missed diagnoses or misdiagnoses. It is difficult to meet the stringent requirements of robustness and interpretability of assessment results in medical scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multimodal help-seeking risk assessment method based on order-preserving calibration. This method solves the problems of existing psychological risk assessment methods lacking an effective conflict quantification mechanism when faced with inconsistencies in multimodal semantic expressions, and being unable to provide dynamic and reliable risk confidence intervals based on the uncertainty of the assessment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multimodal help-seeking risk assessment method based on order-preserving calibration, the method comprising the following steps:

[0007] S1: Obtain the multimodal data of the sample to be tested, map it to the joint risk metric manifold space, and fuse it based on the attention mechanism to obtain the target fusion feature representation;

[0008] S2: Call the preset clinical semantic anchor sequence, which consists of several discrete anchors corresponding to different risk levels, and each anchor satisfies the inherent partial order constraint relationship in the joint risk metric manifold space.

[0009] S3: Calculate the deviation between each modal feature and the target fused feature representation, and quantify it to obtain the modal conflict vector and conflict intensity scalar to characterize the semantic inconsistency between modalities;

[0010] S4: Perturb each anchor in the clinical semantic anchor sequence based on the modal conflict vector and conflict intensity scalar to generate a corresponding counterfactual anchor cloud, and use a differential partial order discriminator to calculate the pairwise advantage probability of the target fusion feature representation relative to each virtual anchor in the counterfactual anchor cloud;

[0011] S5: Statistically aggregate the pairwise dominance probabilities to obtain the partial order dominance sequence for each risk level, thereby defining the dynamic risk confidence interval and outputting the evaluation results.

[0012] Preferably, the target fusion feature representation obtained in S1 based on the attention mechanism specifically includes:

[0013] The attention energy value of each modality is obtained by calculating the dot product of the vector of each modality feature after linear transformation and nonlinear activation with a learnable context vector.

[0014] The attention energy value is normalized to obtain the attention weight, and the attention weight is used to perform a weighted summation of the modal features to generate the target fusion feature representation located in the joint risk metric manifold space;

[0015] The learnable context vector is a parameter that is randomly initialized and jointly optimized during the model training phase.

[0016] Preferably, the process of constructing and constraining the pre-set clinical semantic anchor sequence in S2 specifically includes:

[0017] Calculate the centroid of the fusion features of all samples under the same risk category in the historical confirmed data, and define the centroid as the clinical semantic anchor point of that risk level;

[0018] During the model training phase, the clinical semantic anchor sequence is constrained by the order-preserving loss function, which forces that the difference between the projection values ​​of two adjacent anchors in the direction of the principal risk component is greater than a preset positive boundary threshold, so as to ensure that the clinical semantic anchor sequence has topological order.

[0019] Preferably, the quantization of the modal conflict vector and conflict intensity scalar in S3 specifically includes:

[0020] Calculate the vector difference between the feature representation of each independent modality and the target fused feature representation to obtain the deviation vector of each modality;

[0021] The attention weights calculated in S1 are used to perform a weighted summation of the bias vectors of all modalities to obtain the modal conflict vector indicating the direction of semantic conflict.

[0022] Calculate the L2 norm of the modal conflict vector and multiply it by a normalization factor set based on the statistical distribution of the training set to obtain the conflict intensity scalar used to control the perturbation range.

[0023] Preferably, the generation of the corresponding counterfactual anchor cloud in step S4 specifically includes:

[0024] For each anchor in the clinical semantic anchor sequence, a compound perturbation is applied to generate multiple virtual anchors, which constitute a counterfactual anchor cloud centered on the anchor.

[0025] The composite perturbation includes directional conflict perturbation and random noise perturbation. The directional conflict perturbation is to scale the modal conflict vector according to the conflict intensity scalar and then superimpose it onto the original anchor point. The random noise perturbation is to superimpose a Gaussian random vector whose variance is controlled by the conflict intensity scalar.

[0026] Preferably, in step S4, the calculation of the pairwise advantage probability using a differential partial order discriminator specifically includes:

[0027] A neural network-based binary classifier is constructed as the differential partial order discriminator, whose input is the difference vector of two feature vectors and whose output is a probability value indicating the relative risk level.

[0028] The target fusion feature representation is vector-differentially divided with the virtual anchor points sampled from the counterfactual anchor cloud, and the generated difference feature vector is input into the difference partial order discriminator.

[0029] The risk level of the target fusion feature representation is higher than that of the virtual anchor point after processing by the fully connected layer of the differential partial order discriminator and the Sigmoid activation function, i.e., the pairwise advantage probability.

[0030] Preferably, the partial order dominance sequence obtained by statistical aggregation in S5 specifically includes:

[0031] For each preset risk level, the pairwise advantage probability is calculated by extracting all virtual anchors contained in the corresponding counterfactual anchor cloud and fusing them with the target feature representation.

[0032] Calculate the arithmetic mean of all pairwise dominance probabilities under this risk level to obtain the partial order dominance of this risk level;

[0033] The partial order dominance values ​​of all risk levels are aggregated to generate the partial order dominance sequence, which shows a decreasing trend as the risk level increases.

[0034] Preferably, defining the dynamic risk confidence interval in S5 specifically includes:

[0035] Set high confidence thresholds and low confidence thresholds;

[0036] In the partial order dominance sequence, the highest risk level whose partial order dominance value is maintained above the high confidence threshold is determined as the lower bound anchor point by reverse retrieval.

[0037] In the partial order dominance sequence, a forward search is performed, and the risk level at which the partial order dominance value first falls below the low confidence threshold is determined as the upper bound anchor point.

[0038] The continuous range of levels covered by the lower bound anchor point to the upper bound anchor point is defined as the risk confidence interval.

[0039] Preferably, the output evaluation result in S5 specifically includes:

[0040] The span of the risk confidence interval is calculated as a diagnostic uncertainty indicator;

[0041] If the span value is within a preset low threshold range, the representative level within the risk confidence interval is output as the diagnosis result.

[0042] If the span value exceeds the preset high threshold, the sample to be tested is marked as a high-concern object, the risk confidence interval is output and the modal conflict vector that causes uncertainty is displayed to trigger the manual review process.

[0043] The multimodal help-seeking risk assessment system based on order-preserving calibration, and the multimodal help-seeking risk assessment method based on order-preserving calibration according to any one of claims 1-9, include:

[0044] The multimodal feature joint mapping module is used to acquire the multimodal data of the sample to be tested, map it to the joint risk metric manifold space, and fuse it based on the attention mechanism to obtain the target fused feature representation;

[0045] A clinical semantic anchor construction module is used to provide a pre-set clinical semantic anchor sequence, which is subject to an inherent partial order constraint to maintain the monotonicity of risk level in feature space;

[0046] The modal conflict vector quantization module is used to calculate the deviation between each modal feature and the target fused feature representation, and generate modal conflict vectors and conflict intensity scalars;

[0047] The counterfactual anchor perturbation generation and discrimination module is used to generate a counterfactual anchor cloud based on conflict features and calculate the pairwise advantage probability of the target fusion features relative to the virtual anchors through a differential partial order discriminator.

[0048] The risk calibration boundary determination module is used to determine the lower and upper bound anchor points of the risk confidence interval based on the partially ordered dominance sequence after statistical aggregation, and output the evaluation results according to the interval span.

[0049] This invention provides a multimodal help-seeking risk assessment method based on order-preserving calibration. It has the following beneficial effects:

[0050] 1. This invention calculates modal conflict vectors and generates counterfactual anchor cloud containing directional perturbations, explicitly transforming the semantic inconsistency between multimodalities into the geometric diffusion range in the feature space. It can simulate the potential drift path of risk features under extreme conflict, effectively avoiding the risk underestimation caused by mutual cancellation of information in traditional feature fusion, and significantly improving the evaluation robustness of the model under heterogeneous data.

[0051] 2. This invention establishes a dynamic definition mechanism for upper and lower bounds of risk confidence by constructing a differential partial order discriminator and calculating the partial order dominance sequence. This mechanism abandons the traditional single-point prediction mode and instead adaptively outputs a calibrated risk interval based on the modal conflict intensity. While ensuring a high-confidence basic assessment, the span of the interval intuitively reflects the degree of uncertainty of the current diagnosis, providing clinicians with safer and more interpretable decision support.

[0052] 3. This invention constructs a standardized risk reference system with strict topological order by pre-setting a sequence of clinical semantic anchor points constrained by the order loss in the joint risk metric manifold space. This ensures that the risk assessment of the test sample is based on the relative partial order relationship with the standard clinical representation, rather than a simple absolute distance measurement. This guarantees the logical monotonic consistency of the assessment results and strictly conforms to the clinical pathological law of the increasing severity of mental illnesses. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2This is a system architecture diagram of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example:

[0057] See attached document Figure 1 This invention provides a multimodal help-seeking risk assessment method based on order-preserving calibration, comprising the following steps:

[0058] S1: Perform multimodal feature extraction and joint manifold space mapping operations to obtain the text, audio and video modal features of the object to be evaluated, map them to a unified joint risk metric manifold space and generate a fused feature representation;

[0059] S2: Construct a discrete sequence of clinical semantic anchor points and define standardized reference anchor points that cover different risk levels and satisfy inherent partial order relations in the joint risk measure manifold space;

[0060] S3: Perform modal conflict vector quantization and perturbation intensity calculation operations, calculate the degree of discreteness of each modal feature relative to the fused feature representation, and generate modal conflict vectors and conflict intensity scalars representing the direction of semantic inconsistency;

[0061] S4: Generate counterfactual anchor cloud and perform partial order relation discrimination. Use modal conflict vectors to perturb clinical semantic anchors to generate virtual anchor cloud, and calculate the partial order dominance of the test sample relative to the virtual anchor cloud of each risk level.

[0062] S5: Determine the risk calibration boundary and output the evaluation results. Identify the flip critical region based on the partial order dominance sequence, and determine the lower and upper bounds of risk confidence to generate the risk calibration interval.

[0063] See attached document Figure 2 This invention provides a multimodal help-seeking risk assessment system based on order-preserving calibration, including: a multimodal feature joint mapping module, a clinical semantic anchor construction module, a modal conflict vector quantization module, a counterfactual anchor perturbation generation and discrimination module, and a risk calibration boundary determination module.

[0064] In the multimodal feature extraction stage, the system receives text modal data, audio modal data, and video modal data of the object to be evaluated, respectively. Based on the physical characteristics of different modal data, corresponding deep neural network models are used for independent encoding to obtain the original high-dimensional feature representations of each modality. For text modal data, a pre-trained language model is used for semantic feature extraction.

[0065] Specifically, the text input from the person seeking help is serialized and then fed into the BERT or RoBERTa model. The model's deep Transformer encoder captures contextual dependencies and sentiment semantic information within the text, extracting the original feature vector of the text modality, denoted as... The text feature vector It includes a deep semantic representation of the input text. For the specific structure and parameter settings of the BERT model, those skilled in the art can refer to relevant literature, which will not be elaborated here. For audio modal data, a speech model based on convolutional neural networks or self-attention mechanisms is used for acoustic feature extraction.

[0066] Specifically, the original speech waveform or spectrogram is input into the Wav2Vec2.0 or VGGish model. Local acoustic features are extracted through convolutional layers, and combined with temporal modeling layers to capture non-linguistic emotional cues such as intonation, speech rate, and pauses, thus extracting the original feature vector of the audio modality, denoted as . The audio feature vector It can characterize the speaker's emotional state and prosodic features.

[0067] For video modal data, a spatiotemporal feature extraction network is used for facial and body behavior analysis. The model simultaneously performs convolutional operations on the spatial and temporal dimensions of the video, capturing facial micro-expression changes and body movement sequences, extracting the original feature vector of the video modality, denoted as . The video feature vector It aggregates dynamic behavioral features from the visual channel.

[0068] The encoding processes for each of the above modalities are executed independently and in parallel, each outputting text features. Audio features and video features These original feature vectors reside in their respective independent feature spaces, with different dimensional distributions and physical properties, providing the basic input for subsequent heterogeneous space mapping.

[0069] To eliminate the heterogeneity between the original features of different modalities, enabling them to be compared and fused within the same dimensional space, it is necessary to construct a joint risk metric manifold space. This space is a unified, low-dimensional embedding space whose metric structure reflects the risk similarity between samples. To achieve this goal, the system processes the original feature vectors of each modality. (in Nonlinear projection transformations are performed on text, audio, and video modalities, respectively. Specifically, an independent projection head is set up for each modality. This projection head can consist of one or more fully connected layers with a non-linear activation function. Its function is to map the high-dimensional, heterogeneous original feature vectors to a dimensionless vector. In the joint risk measure manifold space.

[0070] The projection process can be represented as:

[0071] ;

[0072] in, It is modal In the joint risk measurement manifold space Feature representation in the text. Through this step, the original text features are... Audio features and video features They were transformed into forms with the same dimension and similar distribution characteristics. The parameters of this transformation function are optimized through end-to-end learning during model training to ensure that the distance in the joint space can effectively measure the semantic association between different samples or different modalities.

[0073] After obtaining the feature representations of each modality in the joint manifold space, these features need to be fused to form a global representation that can represent the overall risk status of the sample under test. Since different modalities contribute differently to risk assessment, simple averaging fusion cannot effectively capture key risk clues. Therefore, an attention-based feature aggregation method is adopted to adaptively learn the importance weights of each modality.

[0074] In practice, the modal features are first calculated. Attention energy value. Features are analyzed using a learnable parameter matrix and bias terms. A linear transformation is performed, followed by nonlinear activation using the hyperbolic tangent function, and then a dot product operation is performed with a learnable context vector to obtain the energy value in scalar form.

[0075] Next, the energy values ​​of each modality are normalized using the Softmax function to obtain the attention weights for each modality. The weight value ranges from 0 to 1, and the sum of the weights of all modalities is 1. The larger the weight, the more important the modality is in the current risk assessment.

[0076] Finally, based on the calculated attention weights For each modal feature Perform weighted summation to generate a fusion feature representation of the test sample. This fusion feature Also located in the joint risk metric manifold space In, and the dimension is It integrates information from text, audio, and video modalities, and highlights the contributions of key modalities, providing discriminative input for subsequent conflict detection and anchor point comparison.

[0077] To establish a standardized reference system for risk assessment, it is necessary to consider the joint risk measurement manifold space. A set of discrete clinical semantic anchor sequences is pre-constructed. These anchors do not depend on the current test sample, but are generated based on historical expert diagnosis data, representing the standard form of each risk level in the feature space.

[0078] Specifically, assuming the historical training dataset contains Each risk category is labeled differently. For each specific risk category... (in We collect the fused feature vectors of all confirmed samples in this category. Then, we calculate the centroid of these feature vectors using a statistical averaging method, and define this centroid as the first... Clinical semantic anchors of risk level .

[0079] The calculation process is as follows: Traverse all samples in the training set; if the true label of a sample is... Then, the corresponding fusion features are added to the accumulator. Finally, the sum is divided by the total number of samples in that class to obtain the prototype center, i.e., the anchor point, of that class. Anchor point sequences generated in this way It can stably characterize the distribution center of each risk level in the feature space, providing a benchmark for subsequent partial order comparisons. These anchor points are fixed parameters during the system inference phase and do not need to be recalculated.

[0080] To ensure that the sequence of clinical semantic anchor points accurately reflects the increasing relationship of risk levels, rather than merely discrete category centers, the system imposes a strict inherent partial order constraint on the anchor points within the joint risk metric manifold space. This constraint mechanism requires a set of anchor points... It satisfies the mathematical property of monotonicity, that is, for any two anchor points in the sequence... and If its index satisfies Then their corresponding risk metrics in the manifold space must satisfy... .

[0081] This partial order relationship is achieved by introducing an order-preserving loss during the model training phase. In practice, the system does not enforce the absolute coordinates of the anchor points in space, but rather constrains their relative positional relationships. For each pair of adjacent anchor points... and The system sets a positive boundary threshold, which forces... The projection value along the principal component of risk is at least greater than The threshold is large.

[0082] Through this mechanism, the joint risk metric manifold space is constructed as a topologically ordered structure. Risk level determination no longer depends on the absolute distance between a sample and a point, but rather on the sample's relative position within this ordered sequence of anchor points.

[0083] In multimodal help-seeking scenarios, the inconsistency in semantic representation between modalities is the main source of evaluation uncertainty. This method does not treat this inconsistency as noise that needs to be eliminated, but rather quantifies it into a measurable technical feature, namely the modal conflict vector, to guide the subsequent dynamic calibration process.

[0084] The calculation of modal conflict vectors in the joint risk measure manifold space The process is as follows: Specifically, the system calculates the feature representation of each independent mode. (in Fusion feature representation of the sample The deviation vector between them. This deviation vector It indicates the direction and distance of a single modal feature from the multimodal semantic center.

[0085] To obtain a comprehensive vector that characterizes the overall conflict direction, the system performs a weighted sum of the deviation vectors of all modes. The weights here are the attention weights for each mode calculated previously. In this way, modalities with higher contribution and greater credibility have a greater impact on the definition of conflict direction.

[0086] The final generated modal conflict vector Its direction points to the direction in which semantic conflicts are most significant in the feature space. Physically, it represents the main direction in which the feature vector needs to move in order to resolve the multimodal semantic contradictions of the current sample. This vector is the core driving force for the subsequent generation of the counterfactual anchor cloud, providing a foundation for quantifying uncertainty.

[0087] To quantify the severity of semantic inconsistencies between modalities and determine the spread of subsequent counterfactual anchor clouds in the manifold space, a scalar index, namely the conflict intensity scalar, needs to be extracted from the modal conflict vector. In practice, the modal conflict vector is calculated. The L2 norm is used. A larger L2 norm value indicates higher dispersion of modal features around the fusion center, and more severe semantic conflicts. To keep this intensity value within a controllable and statistically significant range, a normalization factor is introduced. Scaling is applied to it. Normalization factor. Typically, the statistical distribution of the conflict norms of all samples in the training set is predefined.

[0088] The calculation formula is:

[0089] ;

[0090] in, This represents the L2 norm operation of a vector. The calculated result... It is a non-negative real number, which directly determines the coverage radius of the subsequently generated virtual anchor point cloud. When When the value approaches 0, it indicates a high degree of semantic consistency across modalities, and the virtual anchor cloud will shrink to the vicinity of the original anchor points, leading to more accurate evaluation results; when... When the value is large, it indicates severe modal conflict, and the virtual anchor cloud will spread significantly, resulting in a wider risk range for the assessment results, thus accurately reflecting the uncertainty of the current diagnosis.

[0091] After determining the direction and intensity of modal conflict, one of the core steps of this invention is to generate a counterfactual anchor cloud. This process transforms static, discrete clinical semantic anchors into a dynamic, continuous virtual anchor distribution whose shape is directly driven by the modal conflict characteristics of the current sample to be tested.

[0092] In practice, for each fixed clinical semantic anchor point The system applies a composite perturbation to it. This perturbation consists of two main parts: directional conflict perturbation and random noise perturbation.

[0093] Directed conflict perturbation is the process of applying modal conflict vectors... According to the intensity of conflict After scaling, add it to the original anchor point. Above. This operation shifts the original anchor point along the direction where the semantic conflict of the current sample is most significant, simulating the worst-case drift that the standard form of this risk level might undergo in the context of the current conflict.

[0094] Random noise perturbation, based on the aforementioned directional offset, adds a random vector sampled from a Gaussian distribution. The variance of this random vector is also determined by the conflict intensity. Control. Its role is to form a distribution cloud around the main axis of directional offset, covering other possible, secondary semantic change directions, thereby exploring the uncertainty space more comprehensively.

[0095] By repeating this process N times, the system generates a value for each original anchor point. N corresponding virtual anchor points were generated. These virtual anchor points together constitute a network. The cloud of counterfactual anchors centers on a specific modal conflict. When the modal conflict is small, the cloud contracts tightly; when the modal conflict is large, the cloud is significantly stretched and diffused along the conflict direction. This generated counterfactual anchor cloud explicitly encodes the semantic uncertainty of the sample under test into the geometric distribution of the reference frame, providing a foundation for subsequent robust partial order discrimination.

[0096] To quantify the relative risk levels between the test sample and each virtual anchor point in the counterfactual anchor cloud, a specialized differential partial order discriminant needs to be constructed. This discriminant aims to determine the relative order of risk levels between two feature points in the joint risk metric manifold space, rather than their absolute distance in computational geometry.

[0097] In its implementation, the discriminator employs a neural network-based binary classification architecture. The system performs vector difference operations between the fused feature representation of the test sample and a virtual anchor sampled from the counterfactual anchor cloud, generating a difference feature vector. This difference feature vector preserves the relative position and orientation differences of the test sample relative to the virtual anchor in each feature dimension.

[0098] Subsequently, the differential feature vector is input into a multilayer perceptron network containing hidden layers. After linear transformation and mapping by a nonlinear activation function in a fully connected layer, the network outputs a scalar probability value between 0 and 1 through the sigmoid activation function at the end.

[0099] The output value is defined as the pairwise advantage probability, used to characterize the confidence level that the risk severity of the test sample is higher than that of a specific virtual anchor. If the output value approaches 1, it indicates that the risk of the test sample is significantly higher than that of the virtual anchor; if the output value approaches 0, it indicates that the risk of the test sample is significantly lower than that of the virtual anchor; if the output value is around 0.5, it indicates that the two are in a state of ambiguity or closeness in terms of risk measurement. The network weight parameters of the discriminator are optimized during the offline training phase of the system by using supervised learning with paired samples having known partial order relations to ensure that it can accurately identify the dominant direction representing increasing risk in the manifold space.

[0100] After obtaining the pairwise odds between the sample to be tested and a single virtual anchor point, statistical aggregation is needed to transform the micro-level pairwise comparison results into a macro-level metric for a specific risk level, namely, the partial order dominance. This metric aims to eliminate random fluctuations caused by a single perturbation sample, thereby establishing the relative strength of the sample to be tested relative to a certain level of risk standard reference in a statistical sense.

[0101] In practice, for each preset risk level The system extracts all the counterfactual anchor points contained in its corresponding counterfactual anchor cloud. Each virtual anchor point is used. The fused features of the sample to be tested are paired with each of these virtual anchor points, and each pair of feature vectors is input into the differential partial order discriminator. This process will generate... Each pairwise odds value represents a set of probability distributions for that risk level. Then, statistical aggregation is performed on this probability set. Typically, the arithmetic mean is calculated by summing all probability values ​​in the set and dividing by the total number of virtual anchors, thus yielding the probability distribution of the test sample relative to the first digit. Partial order dominance of level risk This advantage This is a scalar value between 0 and 1, and its physical meaning represents the actual risk level of the sample being tested, after comprehensively considering various potential semantic drifts and uncertainties caused by modal conflicts, which is higher than the average confidence level of the standard reference system for this level. The above calculation steps are repeated for all defined risk levels to finally generate a set containing... This is a partial order dominance sequence of numerical values. Ideally, this sequence exhibits a monotonically decreasing trend: for extremely low risk levels, the dominance approaches 1; as the reference level increases, the dominance gradually decreases; for extremely high risk levels, the dominance approaches 0. This sequence not only identifies the risk position of the sample, but the steepness or gentleness of its numerical decrease also directly reflects the range of uncertainty in the assessment results, providing a quantitative basis for subsequently determining the dynamic risk boundary.

[0102] After obtaining the partial order dominance sequence for each risk level, the system identifies the relative position of the test sample within the risk reference system based on the numerical trend of this sequence. Since the clinical semantic anchors are strictly ordered from low to high risk, ideally, the calculated partial order dominance sequence exhibits a monotonically decreasing characteristic, meaning the value gradually transitions from near 1 to near 0. The critical region in this numerical decline process is the partial order reversal boundary, marking the logical boundary where the risk level of the test sample changes from significantly higher than the anchor point to significantly lower than the anchor point.

[0103] To accurately pinpoint this boundary, the system sets a central decision threshold. This threshold represents a balance between the risk of the sample under test and the risk of the current anchor point. The identification logic traverses the dominance sequence to find the specific location where the value crosses this central decision threshold. Specifically, the system locates the risk level index of the last dominance greater than or equal to the central decision threshold, and the risk level index of the first dominance less than the central decision threshold. These two adjacent or overlapping index positions constitute the core benchmark for partial order reversal.

[0104] If the dominance sequence exhibits non-monotonic fluctuations locally, the system employs smoothing or maximum likelihood estimation strategies to determine the descent crossover point that best aligns with the overall trend. Identifying this flip boundary is not merely for obtaining a single classification result, but rather for determining the central anchor position of the risk interval, providing an arithmetic benchmark for subsequently generating dynamic confidence intervals based on uncertainty information. This process maps continuous probability assessments back to a discrete clinical risk grading framework while preserving the relative positional information of the assessments.

[0105] This invention goes beyond simply outputting a single risk level prediction value; instead, it constructs a dynamic risk confidence interval encompassing a range of possible risks. This interval is defined directly based on the numerical decay characteristics of the partial order dominance sequence, mapping the statistical features of the probability distribution to specific clinical risk level boundaries.

[0106] In practice, the system introduces a pair of statistically significant cutoff thresholds: a high-confidence threshold and a low-confidence threshold. The high-confidence threshold is typically set to a higher probability value to ensure a minimum level of risk that actually exists in the sample; the low-confidence threshold is typically set to a lower probability value to ensure a maximum level of risk that is highly unlikely to be exceeded. These two thresholds are not fixed and can be adjusted according to specific clinical requirements for controlling missed diagnoses or misdiagnoses.

[0107] To determine the lower bound of risk confidence, the system performs a reverse search in the partial order dominance sequence to find the highest risk level index whose value still remains above the high confidence threshold. The risk level corresponding to this index is marked as the lower bound anchor point. Its physical meaning represents that, after fully considering the uncertainty disturbances caused by modal conflicts, the sample to be tested at least has the risk characteristics of this level, which is a relatively conservative assessment baseline.

[0108] To determine the upper bound of risk confidence, the system searches for the risk level index in the sequence where the value first falls below the low confidence threshold. The risk level corresponding to this index is marked as the upper bound anchor point. Its physical meaning is that even in the worst-case scenario caused by modal conflict, the risk characteristics of the sample under test are unlikely to exceed this level, which belongs to the safety upper limit of the assessment.

[0109] The continuous range of levels covered by the lower and upper bound anchor points constitutes the final risk confidence interval. This mechanism establishes a dynamic relationship between interval width and modal conflict intensity: when there is a high degree of semantic consistency between modalities, the dominance sequence decreases sharply in a step-like manner, causing the upper and lower bounds to coincide or be adjacent, forming a narrow, precise interval; conversely, when modal conflict is severe, the dominance sequence decreases gradually and smoothly, leading to a significant separation between the upper and lower bounds, forming a wide, ambiguous interval. This dynamic definition method can intuitively alert clinicians to the potential ambiguity in the current diagnosis, avoiding the erroneous judgments made by traditional models when faced with heterogeneous data.

[0110] Based on the risk confidence lower and upper bound anchor points determined in the preceding steps, the system constructs a closed set containing these two endpoints and all intermediate risk levels in between, thereby generating the final calibration risk interval. This interval is no longer a single-point prediction, but a comprehensive description of the psychological state of the current test sample, and its span directly reflects the model's confidence level when facing multimodal semantic conflicts.

[0111] To apply this interval to actual triage decisions, the system further calculates the span value of the interval, which is the difference between the upper and lower bound anchor indices. This span value is output as a diagnostic uncertainty indicator. Based on the magnitude of this indicator, the system executes differentiated subsequent processing logic: if the span value is zero or within a preset low threshold range, it indicates high semantic consistency among modalities, and the system outputs a specific level within the interval as the basis for diagnosis, and classifies the sample into a standardized automatic grading queue; if the span value exceeds a preset high threshold, it indicates significant semantic gaps between modalities, and the system marks the sample as a "high concern" object and triggers a manual review process.

[0112] In this review process, the system not only outputs a broad risk range but also calls upon the previously calculated modal conflict vector to visually demonstrate to reviewers the specific sources of conflict leading to uncertainty, thereby achieving interpretable assisted diagnosis. This range-based output mechanism effectively avoids forced model fitting under heterogeneous data input, ensuring the safety and reliability of mental health screening.

Claims

1. A method for multi-modal helpdesk risk assessment based on order preserving calibration, characterized in that, The method comprises the following steps: S1: obtaining multi-modal data of a sample to be tested, mapping the multi-modal data to a joint risk metric manifold space, and obtaining a target fusion feature representation based on an attention mechanism; S2: calling a preset clinical semantic anchor sequence, wherein the clinical semantic anchor sequence comprises a plurality of discrete anchors corresponding to different risk levels, and each anchor satisfies an inherent partial order constraint relationship in the joint risk metric manifold space; S3: calculating deviations between each modality feature and the target fusion feature representation, quantifying a modality conflict vector and a conflict intensity scalar to represent semantic inconsistency between modalities; S4: perturbing each anchor in the clinical semantic anchor sequence based on the modality conflict vector and the conflict intensity scalar, generating a corresponding counterfactual anchor cloud, and calculating a pair-wise dominance probability of the target fusion feature representation relative to each virtual anchor in the counterfactual anchor cloud by using a differential partial order discriminator; S5: statistically aggregating the pair-wise dominance probability to obtain a partial order dominance sequence for each risk level, and defining a dynamic risk confidence interval and outputting an evaluation result based on the partial order dominance sequence.

2. The method of claim 1, wherein, In S1, the target fusion feature representation is obtained based on the attention mechanism, specifically including: calculating the dot product of each modality feature after linear transformation and nonlinear activation and a learnable context vector to obtain an attention energy value of each modality; normalizing the attention energy value to obtain an attention weight, and performing weighted summation on each modality feature by using the attention weight to generate the target fusion feature representation in the joint risk metric manifold space; wherein the learnable context vector is a parameter that is randomly initialized and jointly optimized during model training.

3. The method of claim 1, wherein, In S2, the construction and constraint process of the preset clinical semantic anchor sequence specifically includes: calculating the centroid of all sample fusion features in the same risk category in historical confirmed data, and defining the centroid as the clinical semantic anchor of the risk level; in the model training stage, a preservation order sorting loss function is used to constrain the clinical semantic anchor sequence, and the difference between the projection values of two adjacent anchors in the risk principal component direction is forced to be greater than a preset positive boundary threshold, so as to ensure that the clinical semantic anchor sequence has topological order.

4. The method of claim 1, wherein, In S3, the modality conflict vector and the conflict intensity scalar are quantified, specifically including: calculating the vector difference between the feature representation of each independent modality and the target fusion feature representation to obtain a deviation vector of each modality; performing weighted summation on the deviation vectors of all modalities by using the attention weight calculated in S1 to obtain the modality conflict vector indicating the semantic conflict direction; calculating the L2 norm of the modality conflict vector, and multiplying the L2 norm by a normalization factor set based on the statistical distribution of the training set to obtain the conflict intensity scalar for controlling the perturbation range.

5. The method of claim 1, wherein, In S4, the corresponding counterfactual anchor cloud is generated, specifically including: for each anchor in the clinical semantic anchor sequence, a composite perturbation is applied to generate a plurality of virtual anchors, and the plurality of virtual anchors constitute a counterfactual anchor cloud centered on the anchor; The complex perturbation includes a directional conflict perturbation and a random noise perturbation, wherein the directional conflict perturbation is superimposed to the original anchor point after the modal conflict vector is scaled according to the conflict intensity scalar, and the random noise perturbation is superimposed to a Gaussian random vector with a variance controlled by the conflict intensity scalar.

6. The method of claim 1, wherein, The S4 includes calculating a pair-wise advantage probability by using a differential partial order discriminator, and specifically includes: A neural network-based binary classifier is constructed as the differential partial order discriminator, an input of which is a differential vector of two feature vectors, and an output of which is a probability value indicating a relative risk level; The target fusion feature representation is vector-differentiated with a virtual anchor point sampled from the counterfactual anchor point cloud, and a differential feature vector generated is input into the differential partial order discriminator; The differential partial order discriminator is processed through a fully connected layer and a Sigmoid activation function, and outputs a confidence degree that the risk level of the target fusion feature representation is higher than that of the virtual anchor point, i.e., the pair-wise advantage probability.

7. The method of claim 1, wherein, The S5 includes statistically aggregating a partial order advantage degree sequence, and specifically includes: For each preset risk level, a pair-wise advantage probability calculated from all virtual anchor points in a corresponding counterfactual anchor point cloud and the target fusion feature representation is extracted; An arithmetic mean of all pair-wise advantage probabilities at the risk level is calculated to obtain a partial order advantage degree of the risk level; All partial order advantage degrees of the risk levels are summarized to generate the partial order advantage degree sequence showing a downward trend in value with the increase of the risk level.

8. The method of claim 1, wherein, The S5 includes defining a dynamic risk confidence interval, and specifically includes: High and low confidence thresholds are set; The highest risk level at which the partial order advantage degree value is maintained above the high confidence threshold is determined as a lower bound anchor point through reverse retrieval in the partial order advantage degree sequence; The risk level at which the partial order advantage degree value first falls below the low confidence threshold is determined as an upper bound anchor point through forward retrieval in the partial order advantage degree sequence; A continuous level range covered from the lower bound anchor point to the upper bound anchor point is determined as the risk confidence interval.

9. The method of claim 1, wherein, The S5 includes outputting an evaluation result, and specifically includes: A span value of the risk confidence interval is calculated as a diagnostic uncertainty indicator; If the span value is within a preset low threshold range, a representative level within the risk confidence interval is output as a confirmed diagnosis result; If the span value exceeds a preset high threshold, the sample under test is marked as a high-attention object, and the risk confidence interval and the modal conflict vector causing uncertainty are output to trigger an artificial review process.

10. A multi-modal help-seeking risk assessment system based on order preserving calibration according to any one of claims 1-9, characterized in that, The method includes: a multi-modal feature joint mapping module, configured to obtain multi-modal data of a sample under test, map the multi-modal data to a joint risk measurement manifold space, and fuse the multi-modal data based on an attention mechanism to obtain a target fusion feature representation; a clinical semantic anchor point construction module, configured to provide a preset clinical semantic anchor point sequence, the anchor point sequence being constrained by an inherent partial order to maintain monotonicity of risk levels in a feature space; a modal conflict vector quantification module, configured to calculate deviations of modal features from the target fusion feature representation, and generate a modal conflict vector and a conflict intensity scalar; The counter-factual anchor point generation and discrimination module is configured to generate a counter-factual anchor point cloud based on the conflict features, and calculate a pair-wise advantage probability of the target fusion features relative to the virtual anchor point through a differential partial order discriminator; The risk calibration boundary determination module is configured to determine a lower anchor point and an upper anchor point of a risk confidence interval based on the statistically aggregated partial order advantage degree sequence, and output an evaluation result according to the interval span.

Images