A method and public platform for dynamically predicting the risk of osteopenia in psoriasis patients

Abstract

The method of the embodiment of the present application proposes a method and a public platform for dynamically predicting the risk of osteopenia of psoriasis patients. First, high-quality data are screened from the historical medical records, bone density data and biomarker data of the patients, and multi-modal feature extraction is performed. The high-order interaction of the features of each mode is captured through a graph attention network to realize multi-modal feature fusion. Then, a deep attention neural network is constructed to realize the probabilistic prediction of the risk of osteopenia of psoriasis patients. Next, the individual prediction results are visualized and explained by using the model attention weight and SHAP value, thereby enhancing the clinical usability. Subsequently, the network is compressed and optimized through model distillation, pruning and quantization, thereby realizing the efficient deployment of the model on the edge computing device. Finally, the model parameters are continuously optimized in combination with the actual clinical results and the feedback of doctors. The present application effectively improves the accuracy of risk prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical model processing, and particularly relates to a method for dynamically predicting the risk of osteopenia of a psoriasis patient and a public platform. BACKGROUND

[0002] In recent years, research shows that as a systemic inflammatory disease, the risk of osteopenia and osteoporosis of a psoriasis patient is significantly higher than that of ordinary people, and the disease course is complex and the mechanism is diverse, so a single clinical indicator cannot accurately predict individual risk. How to realize early and accurate identification and risk intervention of osteopenia of a psoriasis patient is a hot and difficult point in the fields of rheumatism and immunity, bone metabolism and dermatology.

[0003] Traditional osteopenia risk assessment tools are mostly based on a single mode, ignoring the high-dimensional interaction information of multi-source data such as patient medical records, biomarkers, drug history and lifestyle, and have limited prediction accuracy and generalization ability. In addition, most traditional models use linear or shallow structures such as logistic regression and decision tree, which cannot effectively capture complex nonlinear and cross-modal interaction relationships, and lack automatic explanation ability for key risk factors of individual patients, which cannot meet the needs of modern precision medicine.

[0004] At present, there is still a lack of intelligent platform systems for the whole process, closed loop, multi-modal, interpretable and edge deployable risk of osteopenia of psoriasis patients at home and abroad. An innovative technical solution covering data collection and cleaning, multi-modal feature fusion, deep prediction, model explanation, compression deployment and feedback optimization is urgently needed to realize intelligent, personalized and real-time prediction and clinical auxiliary decision-making for the risk of osteopenia of psoriasis patients, and to improve the early screening, early prevention and precision intervention ability. The establishment of such a technical platform not only helps to improve the risk prediction level of single disease, but also provides an innovative paradigm with promotional value for the application of artificial intelligence in chronic disease management and precision medicine. SUMMARY

[0005] Therefore, an MRI image processing method for a cochlear implant patient is provided in the embodiments of the application.

[0006] The method for dynamically predicting the risk of osteopenia of a psoriasis patient provided in the embodiments of the application comprises the following steps.

[0007] S1, high-quality data is screened from each psoriasis patient record, wherein the psoriasis patient record comprises historical medical records, bone density data and biomarker data;

[0008] S2, multi-modal feature extraction is performed on each high-quality data, and multi-modal fusion is performed after high-order feature interaction relationships are captured through a graph attention network;

[0009] S3. Construct a deep attention neural network to predict the risk of osteopenia in psoriasis patients, and output the predicted probability of osteopenia risk through the deep attention neural network.

[0010] S4. Establish a model based on model attention weights and SHAP values ​​to visualize individual prediction results;

[0011] S5. Distill, prune and quantize the deep attention neural network, and compress it through an edge computing model;

[0012] S6. Collect actual clinical results and doctor feedback data to optimize the parameters of the deep attention neural network.

[0013] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S1 specifically includes:

[0014] S11. Extract the feature vector X from the historical medical records, bone density data, and biomarker data. i =[X ehr ,X bmd ,X bio ];in,

[0015] Natural language processing is performed on historical medical records, and medical text embedding models are used to extract medical record text features X. ehr ;

[0016] Extract bone mineral density measurement indicators: lumbar spine bone mineral density (L-BMD), femoral neck bone mineral density (F-BMD), T-score, and Z-score, to construct the bone mineral density feature vector X. bmd =[L-BMD,F-BMD,T-score,Z-score];

[0017] Biomarker indicators were extracted: inflammatory factors IL-6, TNF-α, CRP and bone metabolism indicators OC, CTX, P1NP, to construct a bone mineral density feature vector X. bmd =[IL-6,TNF-α,CRP,OC,CTX,P1NP];

[0018] S12. Calculate the data reconstruction error using an autoencoder model, determine the anomaly scoring threshold, and perform anomaly data filtering; whereby...

[0019] If the reconstruction error AS i If X > T, then the data X i Data deemed outliers are filtered out; the outlier scoring threshold T = μ(AS) + α·σ(AS), and the reconstruction error. h i =f enc (X i ), hi For the encoded representation of data, f enc and f dec Here are the encoding and decoding functions of the autoencoder, α is the anomaly threshold adjustment parameter, and μ(AS) and σ(AS) are the mean and standard deviation of the anomaly score.

[0020] S13. Calculate the integrity score and consistency score of the data after initial screening. Obtain high-quality data scores through a multi-factor comprehensive scoring model and screen out high-quality data.

[0021] The multi-factor comprehensive scoring model is expressed by the formula: Integrity score is expressed as The consistency score is expressed as n missing n represents the number of missing features. total n represents the total number of features. inconsistent Let n be the number of features that violate the rules. checks For the total number of rule checks, AS max ω1, ω2, and ω3 represent the maximum abnormal score and are weighting coefficients.

[0022] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S2 specifically includes:

[0023] S21. Construct a multimodal feature heterogeneity map based on medical record text features, bone density features, and biomarker features; among which...

[0024] Construct a node set: V = V text ∪V bmd ∪V bio V text V is a set of feature nodes for medical record text. bmd V is a set of bone mineral density index nodes. bio A set of biomarker indicator nodes;

[0025] Construct the edge set: E = {(v i ,r k ,v j )|v i ,v j ∈V,r k ∈R}, which includes same-modal edges, cross-modal edges, and edge types;

[0026] Construct a set of relation types: R = {r1, r2, ..., v} k} should at least include co-occurrence, statistical correlation, pathway regulation, clinical guidance, pharmacodynamics, and intramodal similarity;

[0027] Define the modal feature heterogeneity graph as G = (V, E, R);

[0028] S22, learning the high-order feature interaction relationship between nodes based on the graph attention network to obtain node embedding features; wherein,

[0029] The feature updating formula of the graph attention network layer is: is the feature representation of node i at the lth layer, W (l) is the weight matrix learned at the lth layer, σ(·) is an activation function, and N(i) is a neighbor node set of node i;

[0030] The attention weight formula is represented as: a (l) and W (l) are model trainable parameters, is the contribution of node j to the feature update of node i;

[0031] S23, calculating the feature fusion weight between modalities through cross-modal attention to output the fused multi-modal feature representation; wherein,

[0032] The fused multi-modal feature is h fusion = Concat(z text , z bmd , z bio , Att text←bmd , Att text←bio , Att bmd←bio ); each modality obtains an overall representation through pooling: The attention calculation formula of modality m to modality n is: W Q , W K and W V are learnable weight matrices, and d is the feature dimension.

[0033] According to any possible implementation of the above-mentioned aspect, further provided is an implementation, and the S3 specifically comprises:

[0034] S31, constructing a deep attention neural network, the network comprising at least two nonlinear hidden layers, and introducing a feature attention weighting mechanism in the hidden layer output feature representation;

[0035] S32, outputting a continuous probability prediction value of the risk of osteopenia in psoriasis patients through the deep attention neural network;

[0036] S33, generating an individualized osteopenia risk prediction report for psoriasis patients according to the predicted probability value; the report comprises an osteopenia risk prediction probability, an impact feature of the osteopenia risk and its contribution degree, and an individualized prevention or treatment strategy recommended based on the prediction result.

[0037] Aspects and any possible implementation manners described above, further provide an implementation manner, the S31 specifically comprises:

[0038] S311, mapping the input multi-modal fusion feature representation to a first hidden layer to obtain an intermediate representation:

[0039] h (1) =σ(W (1) h fusion +b (1) );

[0040] S312, mapping the intermediate representation to a second hidden layer to obtain a higher-order feature representation:

[0041] h (2) =σ(W (2) h (1) +b (2) );

[0042] S313, calculating an attention weight vector for the feature representation of the second hidden layer through an attention mechanism:

[0043] α=softmax(W attn h (2) +b attn );

[0044] S314, element-wise weighting the feature representation of the second hidden layer using the attention weight vector to obtain a weighted feature representation:

[0045] h attn =αe h (2)

[0046] S315, mapping to a bone loss risk prediction probability value through an output layer:

[0047]

[0048] Wherein, W (1) , W (2) and W (3) are weight matrices of each layer, b (1) , b (2) and b (3) are bias terms corresponding to the layer, σ(·) is a nonlinear activation function, W attn and b attn are attention weights and biases, h (1) and h (2) are hidden layer feature representations;

[0049] The cross-entropy loss function of the deep attention neural network is represented as:

[0050]

[0051] wherein y i denotes the true label of the i-th sample, denotes the predicted osteoporosis risk probability value of the i-th sample, N is the total number of training samples, and θ is a set of model trainable parameters containing all weights and bias terms.

[0052] According to the aspect and any possible implementation manner as described above, an implementation manner is further provided, and the S4 specifically comprises:

[0053] S41, extracting attention weights of each input feature of a patient sample in a prediction process;

[0054] S42, using a Shapley additive explanation algorithm to calculate a Shapley value of each feature for the current input feature and the prediction result of the patient sample, and a calculation formula is as follows:

[0055]

[0056] wherein, is the Shapley value of the i-th feature for the current sample x, N is the total feature set, M is the total number of features, S is any feature subset not containing i, and f x is the model prediction output for a specific feature set, and S∪{i} is the union set of the subset S and the feature i;

[0057] S43, drawing a feature importance bar chart with the Shapley value as the horizontal axis to display the quantitative influence and positive or negative direction of the main features on individual prediction; and constructing a one-dimensional heat map based on the feature-level attention vector α, arranging the feature names on the axis, and using color depth to represent the size of the corresponding α value;

[0058] S44, integrating the graphics generated by S41-S43 and the specified numerical values into an interactive interface for display.

[0059] According to the aspect and any possible implementation manner as described above, an implementation manner is further provided, and the S5 specifically comprises:

[0060] S51, migrating the knowledge of a higher-performance original teacher model to a student model with lighter structure through knowledge distillation of the trained deep attention neural network;

[0061] S52, pruning the weight parameters, neurons, feature channels or branches in the neural network according to the weight absolute value, feature attention weight or structure importance score, and removing the parts below the set threshold; and a pruning formula is as follows:

[0062]

[0063] wherein, θ j is a pruning parameter, and Threshold is a threshold value;

[0064] S53, converting the model weight and activation value of the pruned model from a high-precision floating point number to a low bit width, using a linear or dynamic quantization method:

[0065]

[0066] wherein, ω is the original weight, ω q is the quantized weight, ω min and ω max are the weight intervals, and b is the quantization bit number;

[0067] S54, exporting the compressed model processed by distillation, pruning and quantization to a format supported by the edge computing device, and deploying it to the specified edge platform.

[0068] Aspects and any possible implementation modes described above, further provide an implementation mode, wherein the S51 specifically comprises:

[0069] S511, taking the soft label probability distribution generated by the teacher model as the student model training target, and defining the distillation loss function as:

[0070] L distill = KL(softmax(z T / T), softmax(z S / T));

[0071] wherein, z T is the output logits of the teacher model, z S is the output logits of the student model, T is the distillation temperature coefficient, and KL is the Kullback-Leibler divergence;

[0072] S512, weighting and fusing the distillation loss and the cross-entropy loss of the student model on the real label to form the total loss function:

[0073] L total = λ1L hard + λ2L distill ;

[0074] wherein, λ1 and λ2 are weighting coefficients;

[0075] S513, training the student model by back propagation and gradient descent algorithm.

[0076] Aspects and any possible implementation modes described above, further provide an implementation mode, wherein the S6 specifically comprises:

[0077] S61, collect the clinical actual result and the doctor feedback data; wherein the clinical actual result is expressed by a binary label The feedback data given by the doctor according to the model prediction result is expressed as F j ={(f i ,v i )},j=1,2,...,J,f i is the feature that the doctor thinks the model underestimates or overestimates, v i is the correction value of the corresponding feature;

[0078] S62, construct a joint incremental loss function based on the historical training data loss term, the clinical actual result loss term and the doctor feedback data loss term:

[0079]

[0080] Wherein θ is the set of parameters to be optimized of the model, λ real and λ fb are loss term weight coefficients, y i is the true label in the original training data, is the prediction probability of the model to the data, is the collected actual clinical diagnosis label, N is the sample quantity of the original training data set, N r is the sample quantity of the new feedback clinical data, J is ω j is the weight of the feature f i in the current model, is the weight correction value of the expected feature f i of the doctor feedback, L base (θ) is the loss of the historical training data, L real (θ) is the loss of the clinical actual result, L fb (θ) is the loss of the doctor feedback data;

[0081] S63, update the parameters based on the joint incremental loss function by gradient descent method.

[0082] In another aspect, the embodiment of the present application provides a public platform for dynamically predicting the risk of osteopenia of psoriasis patients, which comprises:

[0083] a central server, a medical terminal and a user interaction device;

[0084] The central server comprises:

[0085] a screening module for screening high-quality data from each psoriasis patient record, wherein the psoriasis patient record comprises historical medical records, bone density data and biomarker data;

[0086] An extraction module is configured to perform multi-modal feature extraction on each of the high-quality data, and perform multi-modal fusion after capturing high-order feature interaction relationships through a graph attention network.

[0087] A prediction module is configured to construct a deep attention neural network for predicting the risk of osteopenia in psoriasis patients, and output a prediction probability of the risk of osteopenia through the deep attention neural network.

[0088] A visualization module is configured to visualize the prediction results of the model on individuals based on model attention weights and SHAP values.

[0089] A compression module is configured to perform distillation, pruning and quantization processing on the deep attention neural network, and perform compression through an edge computing model.

[0090] An optimization module is configured to collect actual clinical results and doctor feedback data, and optimize the parameters of the deep attention neural network.

[0091] One of the above technical solutions has the following beneficial effects:

[0092] (1) By innovatively integrating medical record texts, bone density, biomarkers and other multi-modal data, and using graph neural networks and cross-modal attention mechanisms, the complex high-order relationships between features can be fully explored, and more accurate and personalized risk probability assessments can be given for each patient, which is superior to traditional single modal or linear models.

[0093] (2) The technology combines the intrinsic attention weights of the model with the SHAP-based feature attribution method, clearly showing the main influencing features and their contribution, providing traceable and verifiable decision-making basis for doctors and patients, and enhancing the trustworthiness and usability of the artificial intelligence system.

[0094] (3) Through knowledge distillation, pruning and quantization, the model compression technology significantly reduces the number of model parameters and the computational complexity, enabling high-performance prediction models to run efficiently on edge computing platforms such as hospital local servers, mobile terminals or wearable devices, meeting the real-time and energy requirements in clinical practice.

[0095] (4) The platform can continuously collect real clinical results and doctor feedback, use incremental learning and expert guidance to dynamically optimize and correct the model parameters, so that the prediction system always remains highly consistent with the latest clinical practice, improving the long-term stability and adaptability of the model, and greatly improving the early identification and intervention level of osteopenia and other complications in high-risk populations such as psoriasis. BRIEF DESCRIPTION OF DRAWINGS

[0096] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings described below are only some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0097] Figure 1 A flowchart of a method for dynamically predicting the risk of osteopenia in psoriasis patients provided by an embodiment of the present application;

[0098] Figure 2 A schematic block diagram of a public platform central server provided by an embodiment of the present application;

[0099] Figure 3 A structural schematic diagram of a public platform provided by an embodiment of the present application.

DETAILED DESCRIPTION

[0100] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with specific embodiments and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0101] Please refer to Figure 1 which is a flowchart of a method for dynamically predicting the risk of osteopenia in psoriasis patients provided by an embodiment of the present application. As shown in the figure, the method includes the following steps:

[0102] S1, high-quality data is selected from each psoriasis patient record, the psoriasis patient record including historical medical records, bone density data, and biomarker data;

[0103] S2, multi-modal feature extraction is performed on each of the high-quality data, and multi-modal fusion is performed after capturing high-order feature interaction relationship through a graph attention network;

[0104] S3, a deep attention neural network for predicting the risk of osteopenia in psoriasis patients is constructed, and a prediction probability of the risk of osteopenia is output through the deep attention neural network;

[0105] S4, a model based on model attention weight and SHAP value is established to visually display the individual prediction results;

[0106] S5, the deep attention neural network is distilled, pruned and quantized, and compression is performed through an edge computing model;

[0107] S6, collect actual clinical results and doctor feedback data, optimize the parameters of the deep attention neural network.

[0108] The purpose of S1 is to accurately extract feature vectors from the historical medical records, bone density data and biomarker data of psoriasis patients, automatically screen high-quality data through an innovative anomaly detection and data quality scoring model, and provide a solid data foundation for subsequent multi-modal fusion prediction. S1 specifically includes:

[0109] S11, extracting feature vectors X from the historical medical records, bone density data, and biomarker data i = [X ehr , X bmd , X bio ]; wherein,

[0110] Natural language processing is performed on the historical medical records, and medical text embedding model is used to extract medical record text features X ehr ;

[0111] Extracting bone density measurement indicators: lumbar spine bone density L-BMD, femoral neck bone density F-BMD, T-score and Z-score, to form bone density feature vector X bmd = [L-BMD, F-BMD, T-score, Z-score];

[0112] Extracting biomarker indicators: inflammatory factors IL-6, TNF-α, CRP, and bone metabolism indicators OC, CTX, P1NP, to form bone density feature vector X bmd = [IL-6, TNF-α, CRP, OC, CTX, P1NP];

[0113] Bone density measurement indicators include lumbar spine bone density (L-BMD), femoral neck bone density (F-BMD), T-score and Z-score. Biomarker indicators include inflammatory factors (IL-6, TNF-α, CRP) and bone metabolism indicators (OC, CTX, P1NP).

[0114] S12, calculating data reconstruction error with autoencoder model, determining anomaly score threshold and performing abnormal data screening; wherein,

[0115] If the reconstruction error AS i > T, the data X i is determined to be abnormal data and is screened out; the anomaly score threshold T = μ(AS) + α·σ(AS), the reconstruction error h i = f enc (X i ), h i is the encoded representation of the data, fenc and f dec are the encoding and decoding functions of the autoencoder, a is the anomaly threshold adjustment parameter, and μ(AS) and σ(AS) are the mean and standard deviation of the anomaly score;

[0116] The autoencoder model can be trained in an unsupervised manner, and the optimization objective is to minimize the reconstruction error of the normal data in the training set.

[0117] S13, calculate the integrity score and consistency score of the data after preliminary screening, obtain a high-quality data score through a multi-factor comprehensive scoring model, and screen out high-quality data; wherein,

[0118] The multi-factor comprehensive scoring model is expressed by the formula: The integrity score is expressed as The consistency score is expressed as n missing is the number of missing features, n total is the total number of features, n inconsistent is the number of features that violate the rules, n checks is the total number of rule checks, AS max is the maximum anomaly score, and ω1, ω2 and ω3 are weight coefficients.

[0119] S2 accurately captures the high-order nonlinear interaction relationship between the multi-modal data (medical record text, bone density, and biomarker) of psoriasis patients through an innovative graph neural network (GNN) and cross-modal attention mechanism, and realizes more accurate and personalized risk prediction. S2 specifically includes:

[0120] S21, a multi-modal feature heterogeneous graph is constructed for medical record text features, bone density features, and biomarker features; wherein,

[0121] A node set is constructed: V = V text ∪V bmd ∪V bio , V text is a medical record text feature node set, V bmd is a bone density index node set, and V bio is a biomarker index node set;

[0122] An edge set is constructed: E = {(v i ,r k ,v j )|v i ,v j ∈V,r k ∈R}, including intra-modal edges, cross-modal edges and edge types;

[0123] A relationship type set is constructed: R = {r1, r2,..., v kco-occurrence, statistical correlation, pathway regulation, clinical guidance, drug effect, intra-modality similarity;

[0124] The modality-specific heterogeneous graph is defined as G=(V, E, R);

[0125] The same modality edges include Text-Text, BMD-BMD and Bio-Bio, such as ("Psoriatic arthritis", "Methotrexate", co-occurrence), ("T-score", "Z-score", strong correlation) and ("IL-6", "TNF-alpha", inflammation co-regulation), etc. Cross-modality edges include Text-BMD, Text-Bio and BMD-Bio, such as ("Psoriasis", "L-BMD", guide detection), ("Chronic inflammation", "IL-6", common pathway) and ("Z-score", "Osteocalcin", metabolic correlation), etc. The relationship types include co-occurrence, statistical correlation, pathway regulation, clinical guidance, cause / effect / drug effect, intra-modality similarity, etc.

[0126] S22, learning the high-order feature interaction relationship between nodes based on a graph attention network, to obtain node embedding features; wherein,

[0127] The feature update formula of the graph attention network layer is: is the feature representation of node i at the l-th layer, W (l) is the weight matrix learned at the l-th layer, σ(·) is an activation function, and N(i) is the neighbor node set of node i;

[0128] The attention weight formula is represented as: a (l) and W (l) are model trainable parameters, is the contribution degree of node j to the feature update of node i;

[0129] S23, calculating the feature fusion weight between modalities through cross-modality attention, and outputting the fused multi-modal feature representation; wherein,

[0130] The fused multi-modal feature is h fusion = Concat(z text , z bmd , z bio , Att text←bmd , Att text←bio , Att bmd←bio ); Each modality obtains an overall representation through pooling: The attention calculation formula of modality m to modality n is: W Q, W K and W V are learnable weight matrices, d is the feature dimension.

[0131] Specifically, S3 comprises:

[0132] S31, constructing a deep attention neural network, the network comprising at least two layers of nonlinear hidden layers,

[0133] and introducing a feature attention weighting mechanism in the hidden layer output feature representation;

[0134] S31 specifically comprises:

[0135] S311, mapping the input multi-modal fusion feature representation to the first hidden layer to obtain an intermediate representation:

[0136] h (1) = σ(W (1) h fusion + b (1) );

[0137] S312, mapping the intermediate representation to the second hidden layer to obtain a higher-order feature representation:

[0138] h (2) = σ(W (2) h (1) + b (2) );

[0139] S313, calculating an attention weight vector for the feature representation of the second hidden layer through an attention mechanism:

[0140] α = softmax(W attn h (2) + b attn );

[0141] S314, element-wise weighting the second hidden layer feature representation using the attention weight vector to obtain a weighted feature representation:

[0142] h attn = αe h (2)

[0143] S315, mapping to bone loss risk prediction probability value through the output layer:

[0144]

[0145] where W (1) , W (2) and W (3) are weight matrices of each layer, b (1) , b (2) and b (3)for the bias term of the corresponding layer, σ(·) is a nonlinear activation function, W attn and b attn are attention weights and bias, h (1) and h (2) are hidden layer feature representations;

[0146] The cross-entropy loss function of the deep attention neural network is represented as:

[0147]

[0148] where y i represents the true label of the i-th sample, represents the predicted osteopenia risk probability value of the i-th sample, N is the total number of training samples, and θ is a set of model trainable parameters containing all weights and bias terms.

[0149] S32, outputting a continuous probability prediction value of the osteopenia risk of the psoriasis patient through the deep attention neural network;

[0150] S33, generating an individualized osteopenia risk prediction report of the psoriasis patient according to the predicted probability value; the report contains an osteopenia risk prediction probability, an impact feature of the osteopenia risk and its contribution degree, and an individualized prevention or treatment strategy recommended based on the prediction result.

[0151] Specifically, S4 includes:

[0152] S41, extracting the attention weight of each input feature of the patient sample in the prediction process;

[0153] S42, using the Shapley additive explanation algorithm to calculate the Shapley value of each feature for the current input feature and the prediction result of the patient sample, and the calculation formula is:

[0154]

[0155] wherein, is the Shapley value of the i-th feature for the current sample x, N is the total feature set, M is the total number of features, S is any feature subset not containing i, f x is the model prediction output for a specific feature set, and S∪{i} is the union set of the subset S and the feature i;

[0156] S43, drawing a feature importance bar chart with the Shapley value as the horizontal axis to display the quantitative influence and positive or negative direction of the main features on individual prediction; and constructing a one-dimensional heat map based on the feature-level attention vector α, arranging the feature names on the axis, and using color depth to represent the size of the corresponding α value;

[0157] S44, integrate the graphics and the specified numerical value generated by S41-S43 into an interactive interface for display.

[0158] S5 specifically comprises:

[0159] S51, through knowledge distillation, migrate the knowledge of the original teacher model with higher performance to the student model with lighter structure.

[0160] S51 specifically comprises:

[0161] S511, take the soft label probability distribution generated by the teacher model as the training target of the student model, and define the distillation loss function as:

[0162] L distill = KL(softmax(z T / T), softmax(z S / T));

[0163] Wherein, z T is the output logits of the teacher model, z S is the output logits of the student model, T is the distillation temperature coefficient, and KL is the Kullback-Leibler divergence.

[0164] S512, weight and fuse the distillation loss and the cross-entropy loss of the student model to the real label to form the total loss function:

[0165] L total = λ1L hard + λ2L distill ;

[0166] Wherein, λ1 and λ2 are weighting coefficients.

[0167] S513, train the student model through back propagation and gradient descent algorithm.

[0168] S52, according to the weight absolute value, the feature attention weight or the structure importance score, cut the weight parameters, neurons, feature channels or branches in the neural network, and cut the parts below the set threshold; The pruning formula is:

[0169]

[0170] Wherein, θ j is the pruning parameter, and Threshold is the threshold.

[0171] S53, convert the model weight and activation value of the pruned model from high-precision floating point number to low bit width, and use linear or dynamic quantization method:

[0172]

[0173] wherein ω is the original weight, ω q is the quantized weight, ω min and ω max are the weight intervals, and b is the number of quantization bits;

[0174] S54, export the compressed model processed by distillation, pruning and quantization into a format supported by the edge computing device, and deploy it to the specified edge platform.

[0175] S6 specifically includes:

[0176] S61, collect clinical actual results and doctor feedback data; wherein the clinical actual results are represented by binary labels as The feedback data given by the doctor according to the model prediction result is represented as F j ={(f i ,v i )},j=1,2,...,J,f i is the feature that the doctor thinks the model underestimates or overestimates, v i is the correction value of the corresponding feature;

[0177] S62, construct a joint incremental loss function based on the historical training data loss term, the clinical actual result loss term and the doctor feedback data loss term:

[0178]

[0179] wherein θ is the set of parameters to be optimized in the model, λ real and λ fb are loss term weight coefficients, y i is the true label in the original training data, is the prediction probability of the model on the data, is the collected actual clinical diagnosis label, N is the sample size of the original training data set, N r is the sample size of the new feedback clinical data, J is, ω j is the weight of the feature f i in the current model, is the weight correction value of the feature f i expected by the doctor feedback, L base (θ) is the loss of historical training data, L real (θ) is the loss of clinical actual results, L fb (θ) is the loss of doctor feedback data;

[0180] S63, update the parameters based on the joint incremental loss function by gradient descent method.

[0181] Through the above steps, the application achieves the following technical effects:

[0182] (1) By innovatively integrating medical record text, bone density, biomarker and other multi-modal data, and using graph neural network and cross-modal attention mechanism, the complex high-order relationships between features can be fully mined, and more accurate and personalized risk probability assessment can be given to each patient, which is better than traditional single modal or linear model.

[0183] (2) The technology combines the intrinsic attention weight of the model with the SHAP-based feature attribution method, clearly shows the main influencing features and their contribution, provides traceable and verifiable decision basis for doctors and patients, and enhances the trustworthiness and usability of the artificial intelligence system.

[0184] (3) Through model compression technologies such as knowledge distillation, pruning and quantization, the number of model parameters and the computational complexity are significantly reduced, so that the high-performance prediction model can run efficiently on edge computing platforms such as hospital local servers, mobile terminals or wearable devices, meeting the real-time and energy consumption requirements in clinical actual scenarios.

[0185] (4) The platform can continuously collect real clinical results of patients and feedback from doctors, use incremental learning and expert guidance to realize dynamic optimization and self-correction of model parameters, so that the prediction system always keeps high consistency with the latest clinical practice, improves the long-term stability and adaptability of the model, and greatly improves the early identification and intervention level of osteopenia and other complications in high-risk groups such as psoriasis.

[0186] The embodiment of the application further provides a device embodiment for implementing each step and method in the above method embodiment.

[0187] Please refer to Figure 2 , which is a public platform center server provided by the embodiment of the application. Figure 2 The center server comprises:

[0188] The screening module 210 is used for screening high-quality data from each psoriasis patient record, and the psoriasis patient record comprises historical medical records, bone density data and biomarker data.

[0189] The extraction module 220 is used for multi-modal feature extraction of each high-quality data, and multi-modal fusion is performed after capturing high-order feature interaction relationship through a graph attention network.

[0190] The prediction module 230 is used for constructing a deep attention neural network for risk of osteopenia in psoriasis patients, and outputs a prediction probability of risk of osteopenia through the deep attention neural network.

[0191] Visualization module 240 is used to visualize the individual prediction results based on the model attention weights and SHAP values.

[0192] Compression module 250 is used to perform distillation, pruning and quantization on deep attention neural networks, and to compress them through an edge computing model;

[0193] The optimization module 260 is used to collect actual clinical results and doctor feedback data to optimize the parameters of the deep attention neural network.

[0194] Because each unit module in this embodiment can execute Figure 1 For the methods shown, parts not described in detail in this embodiment, please refer to the [examples / descriptions]. Figure 1 Related explanations.

[0195] Please refer to Figure 3 This is a schematic diagram of the structure of the common platform system provided in the embodiments of the present invention.

[0196] include Figure 2 The diagram shows the central server, medical terminals and user interaction devices, as well as edge inference and data acquisition devices. Medical terminals and user interaction devices include, but are not limited to, doctor workstations, mobile medical tablets, and patient self-service inquiry terminals. Edge inference and data acquisition devices include, but are not limited to, departmental edge AI boxes and bedside all-in-one machines.

[0197] At the hardware level, the central server may include a processor, and optionally an internal bus, network interface, and memory. The memory may include RAM, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the central server may also include other hardware required for other business operations.

[0198] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be categorized into address bus, data bus, control bus, etc.

[0199] The memory is configured to store a program. Specifically, the program can include program code including computer operation instructions. The memory can include an internal memory and a nonvolatile memory, and provide instructions and data for the processor.

[0200] The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as hardware code processing executed by a processor, or executed by a combination of hardware and software modules in the processor. The software modules can be located in the internal memory, the flash memory, the read-only memory, the programmable read-only memory, the electrically programmable read-only memory, the register, or other mature storage mediums in the art. The memory is located in the storage medium, and the processor reads information in the memory and combines the hardware to implement the steps of the above method.

[0201] The systems, apparatuses, modules or units illustrated by the above embodiments can be specifically implemented by a computer chip or entity, or by a product with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0202] For the convenience of description, the above apparatus is described as various units or modules in function respectively when described. Of course, the functions of each unit or module can be implemented in the same or multiple software and / or hardware when implementing the present application.

[0203] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0204] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device for implementing the functions described in the flowcharts and / or block diagrams. Figure 1 Each flow or multiple flows and / or blocks Figure 1an apparatus to perform each function

[0205] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a Figure 1 one or more processes and / or blocks Figure 1 an apparatus to perform each function

[0206] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing processes Figure 1 one or more processes and / or blocks Figure 1 an apparatus to perform each function

[0207] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0208] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the computer stores information about an operating system or application programs, e.g., when the computing device is active. The memory is an example of computer readable media.

[0209] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile discs (DVDs) or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0210] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0211] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0212] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0213] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0214] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for dynamically predicting the risk of osteopenia in psoriasis patients, characterized in that, The method includes: S1. Select high-quality data from each psoriasis patient record, including historical medical records, bone density data, and biomarker data; S2. Multimodal feature extraction is performed on the high-quality data, and multimodal fusion is performed after capturing high-order feature interaction relationships through a graph attention network. S3. Construct a deep attention neural network to predict the risk of osteopenia in psoriasis patients, and output the predicted probability of osteopenia risk through the deep attention neural network. S4. Establish a model based on model attention weights and SHAP values ​​to visualize individual prediction results; S5. Distill, prune and quantize the deep attention neural network, and compress it through an edge computing model; S6. Collect actual clinical results and doctor feedback data to optimize the parameters of the deep attention neural network.

2. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 1, characterized in that, S1 specifically includes: S11. Extract feature vectors from the historical medical records, bone density data, and biomarker data. ;in, Natural language processing was performed on historical medical records, and medical text embedding models were used to extract textual features from the medical records. ; Extract bone mineral density measurement indicators: lumbar spine bone mineral density (L-BMD), femoral neck bone mineral density (F-BMD), T-score, and Z-score, to construct a bone mineral density feature vector. ; Biomarker indicators were extracted: inflammatory factors IL-6, TNF-α, CRP and bone metabolism indicators OC, CTX, P1NP, to construct a bone mineral density feature vector. ; S12. Calculate the data reconstruction error using an autoencoder model, determine the anomaly scoring threshold, and perform anomaly data filtering; whereby... If reconstruction error Then the data Data identified as outliers is filtered out; anomaly scoring threshold. Reconstruction error , , , For the encoding representation of data, and These are the encoding and decoding functions for an autoencoder. This is an abnormal threshold adjustment parameter. and The mean and standard deviation of the outlier scores; S13. Calculate the integrity score and consistency score of the data after initial screening. Obtain high-quality data scores through a multi-factor comprehensive scoring model and screen out high-quality data. The multi-factor comprehensive scoring model is expressed by the formula: Integrity score is expressed as The consistency score is expressed as , This represents the number of missing features. For the total number of features, For the number of features that violate the rules, The total number of checks for the rule, The maximum abnormal score. , and These are the weighting coefficients.

3. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 1, characterized in that, S2 specifically includes: S21. Construct a multimodal feature heterogeneous map based on medical record text features, bone density features, and biomarker features; among which... Build a node set: , This is a set of feature nodes for medical record text. For the bone mineral density index node set, A set of biomarker indicator nodes; Constructing the edge set: It includes same-modal edges, cross-modal edges, and edge types; Construct a set of relation types: It should include at least co-occurrence, statistical correlation, pathway regulation, clinical guidance, pharmacodynamics, and intramodal similarity; Define modal feature heterogeneity graph as ; S22. Based on graph attention networks, learn the high-order feature interaction relationships between nodes to obtain node embedding features; whereby... The formula for updating features in a graph attention network layer is: , Let i be the feature representation of node i in the l-th layer. Let be the weight matrix learned in the l-th layer. For activation function, Let i be the set of neighboring nodes of node i; The attention weight formula is expressed as: , and These are the trainable parameters of the model. Let j be the contribution of node j to the feature update of node i. S23. Calculate the feature fusion weights between modalities through cross-modal attention, and output the fused multimodal feature representation; where... The fused multimodal features are Each modality is represented as a whole through pooling. The formula for calculating the attention of mode m to mode n is as follows: ; , and For learnable weight matrix, For feature dimensions.

4. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 3, characterized in that, S3 specifically includes: S31. Construct a deep attention neural network, wherein the network contains at least two nonlinear hidden layers, and a feature attention weighting mechanism is introduced into the feature representation output by the hidden layers; S32. Output a continuous probability prediction value of the risk of osteopenia in psoriasis patients through the deep attention neural network; S33. Based on the predicted probability value, generate an individualized osteopenia risk prediction report for psoriasis patients; the report includes the predicted probability of osteopenia risk, the impact characteristics and contribution of osteopenia risk, and recommendations for individualized prevention or treatment strategies based on the prediction results.

5. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 4, characterized in that, S31 specifically includes: S311. Map the input multimodal fusion feature representation to the first hidden layer to obtain the intermediate representation: ； S312. Map the intermediate representation to the second hidden layer to obtain a higher-order feature representation: ； S313. The attention weight vector is calculated for the feature representation of the second hidden layer using an attention mechanism: ； S314. Element-wise weighting of the second hidden layer feature representation is performed using the attention weight vector to obtain the weighted feature representation: ； S315. Mapping the output layer to the predicted probability value of osteopenia risk: ； in, , and Weight matrices for each layer , and For the bias term of the corresponding layer, It is a non-linear activation function. and For attention weights and biases, and This represents the hidden layer features; The cross-entropy loss function of a deep attention neural network is expressed as: ； in, This represents the true label of the i-th sample. This represents the predicted probability value of bone loss for the i-th sample. The total number of training samples, This is the set of trainable parameters for the model, which includes all weights and biases.

6. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 4, characterized in that, S4 specifically includes: S41. Extract the attention weights of each input feature of the patient sample during the prediction process; S42. Using the Shapley additive interpretation algorithm, calculate the Shapley value for each feature of the current input features and the prediction result for the patient sample. The calculation formula is as follows: ； in, Let be the Shapley value of the i-th feature for the current sample x. For the complete feature set, For the total number of features, For any feature subset that does not contain i, To predict the output of a model for a specific feature set, Let S be the union of subset S and feature i; S43. A feature importance bar chart plotted with Shapley values ​​on the horizontal axis shows the quantitative impact of major features on individual predictions and their positive or negative directions; and a one-dimensional heatmap constructed based on the feature-level attention vector α, with feature names arranged on the axis and color intensity representing the magnitude of the corresponding α value. S44. Integrate the graphics and specified values ​​generated in S41-S43 into the interactive interface for display.

7. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 1, characterized in that, S5 specifically includes: S51. The trained deep attention neural network is transferred from the high-performance original teacher model to the more lightweight student model through knowledge distillation. S52. Prune the weight parameters, neurons, feature channels, or branches in the neural network based on the absolute value of the weights, feature attention weights, or structural importance scores, removing portions below a set threshold; the pruning formula is: ； in, The parameters to be pruned are as follows: For threshold; S53. Convert the model weights and activation values ​​of the pruned model from high-precision floating-point numbers to low-bit width using linear or dynamic quantization methods: ； in, The original weights, To quantify weights, and For the weighted interval, For the number of quantization bits; S54. Export the compressed model, which has undergone distillation, pruning, and quantization, to a format supported by edge computing devices and deploy it to the specified edge platform.

8. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 7, characterized in that, S51 specifically includes: S511. Using the soft label probability distribution generated by the teacher model as the training objective of the student model, define the distillation loss function as: ； in, The output logits of the teacher model, The output logits of the student model, This is the distillation temperature coefficient. The Kullback-Leibler divergence; S512. Weighted fusion of the distillation loss and the student model's cross-entropy loss on the true labels to form the total loss function: ； in, and These are weighting coefficients; S513. Train the student model using backpropagation and gradient descent algorithms.

9. The method for dynamically predicting the risk of osteopenia in psoriasis patients according to claim 1, characterized in that, S6 specifically includes: S61. Collect actual clinical outcomes and physician feedback data; whereby actual clinical outcomes are represented using binary labels. The feedback data given by the doctor based on the model's prediction results is expressed as follows: , For features that doctors believe the model underestimates or overestimates, This is the correction value for the corresponding feature; S62. Construct a joint incremental loss function based on historical training data loss terms, actual clinical outcome loss terms, and physician feedback data loss terms: ； in, The set of parameters to be optimized for the model. and The weighting coefficient for the loss term. These are the true labels from the original training data. The model predicts the probability of the data. For the collection of actual clinical diagnostic labels, The sample size of the original training dataset. For the sample size of the new feedback clinical data, for, Features in the current model The weight, Characteristics of feedback expectations The weight adjustment value, Due to the loss of historical training data, For the loss of actual clinical outcomes, Report data loss to doctors; S63. Update parameters using gradient descent based on the joint incremental loss function.

10. A public platform for dynamically predicting the risk of osteopenia in psoriasis patients using the method of any one of claims 1-9, characterized in that, The public platform includes: Central server, medical terminals and user interaction devices; The central server includes: The filtering module is used to filter high-quality data from each psoriasis patient record, which includes historical medical records, bone density data, and biomarker data. The extraction module is used to extract multimodal features from the high-quality data, and then perform multimodal fusion after capturing high-order feature interaction relationships through a graph attention network. The prediction module is used to construct a deep attention neural network for the risk of osteopenia in psoriasis patients, and outputs the predicted probability of osteopenia risk through the deep attention neural network. The visualization module is used to visualize the individual prediction results of a model based on model attention weights and SHAP values. The compression module is used to perform distillation, pruning, and quantization on deep attention neural networks, and to compress them through an edge computing model; The optimization module is used to collect actual clinical results and doctor feedback data to optimize the parameters of the deep attention neural network.

Images