A Nutritional Decision-Making Method and System Based on Multi-Task Learning (Integrating Diagnosis, Treatment, and Assessment)

Abstract

This invention relates to a multi-task learning-based integrated nutritional decision-making method and system for examination, diagnosis, treatment, and evaluation, belonging to the interdisciplinary field of clinical medicine and artificial intelligence. This method aims to address the problems of isolated tasks, limited data utilization, and lack of quantitative recommendation and prognostic prediction capabilities in existing nutritional decision-making technologies. The technical solution involves constructing a serial multi-task learning neural network model, sequentially including a feature extraction module, a nutritional diagnosis module, a nutritional treatment recommendation module, and a prognostic and efficacy evaluation module. A joint loss function that dynamically balances the weights of each task and reuses prognostic labels is designed for training. This invention achieves integrated decision-making throughout the entire process of examination, diagnosis, treatment, and prognosis. It can integrate multimodal data to achieve accurate feature extraction, provide personalized quantitative nutritional recommendations, and prospectively predict efficacy to guide clinical intervention, significantly improving the efficiency and accuracy of nutritional decision-making.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of clinical medicine and artificial intelligence, and relates to a nutritional decision-making method and system based on multi-task learning that integrates diagnosis, treatment and evaluation. Background Technology

[0002] Nutrition-related problems, including malnutrition, cachexia, sarcopenia, and asthenia, have become a significant global public health burden and a major challenge for clinical diagnosis and treatment. Poor nutritional status significantly affects the effectiveness and safety of disease treatment, increases the consumption of medical resources, prolongs hospital stays, reduces patients' quality of life, and leads to a series of other adverse health consequences. Therefore, in clinical practice, it is crucial to actively examine, assess, diagnose, intervene in, and evaluate the efficacy and prognosis of patients' nutritional problems.

[0003] In recent years, relevant international diagnostic guidelines, such as those from the Global Leadership Initiative on Malnutrition (GLIM) and the European Society for Clinical Nutrition and Metabolism (ESPEN), have provided methodological references for clinical practice. Simultaneously, domestic and international research teams have developed decision-making models for specific issues such as malnutrition diagnosis. However, these existing technical frameworks typically have the following shortcomings: First, they usually treat nutritional diagnosis, treatment recommendations, and prognostic assessment as independent processes, requiring clinicians to make judgments based on different criteria, lacking an integrated decision-making tool that can integrate the entire process of examination, diagnosis, treatment, and prognosis. Second, existing technologies mainly rely on static indicators such as weight and body mass index, failing to effectively integrate static tabular data at patient admission with dynamic time-series data, and failing to fully mine the deeper information in multimodal and heterogeneous data. Third, recommendations for nutritional therapy are mostly based on experience or static guidelines, which typically provide rough suggestions for a specific population, lacking the ability to dynamically predict and accurately recommend nutrient intake such as energy, protein, and fat based on individual patient characteristics. Finally, existing technologies struggle to predict patients’ responses to nutritional therapy and the future outcome of their nutritional status, and cannot simultaneously use prognostic information to improve treatment strategies. They also cannot proactively identify target populations that may benefit from intensive therapy or recover poorly to guide precise interventions. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a nutritional decision-making method and system based on multi-task learning that integrates diagnosis, treatment and evaluation.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multi-task learning-based integrated nutritional decision-making method for diagnosis, treatment, and assessment includes the following steps: Establish a population dataset that includes cross-sectional data, time-series data, nutritional diagnostic labels, nutritional therapy recommendation labels, and efficacy and prognostic evaluation labels; The population dataset is preprocessed and divided into training and testing sets; A multi-task learning neural network model is constructed, comprising a feature extraction module, a nutritional diagnosis module, a nutritional therapy recommendation module, and a prognosis and efficacy evaluation module connected in series. The feature extraction module is used to fuse the cross-sectional data and the time-series data. The nutritional diagnosis module takes the fused features output by the feature extraction module as input and outputs hidden layer information. The nutritional therapy recommendation module takes the hidden layer information output by the nutritional diagnosis module and the fused features output by the feature extraction module as input and outputs hidden layer information. The prognosis and efficacy evaluation module takes the hidden layer information output by the nutritional therapy recommendation module, the hidden layer information output by the nutritional diagnosis module, and the fused features output by the feature extraction module as input. Design a joint loss function, which is used to dynamically balance the loss weights of the nutrition diagnosis module, the nutrition therapy recommendation module, and the prognosis and efficacy evaluation module, and to perform label-aware adjustment on the loss weight of the prognosis and efficacy evaluation module based on the batch label mean. The multi-task learning neural network model is trained and optimized using the training set and the joint loss function. The trained and optimized model is selected and evaluated using the test set.

[0006] Furthermore, the feature extraction module includes a tabular data processing submodule, a time series data processing submodule, and a feature fusion submodule. The tabular data processing submodule uses a fully connected layer to process the cross-sectional data, the time series data processing submodule uses a bidirectional long short-term memory network (BiLSTM) or a Transformer module to process the time series data, and the feature fusion submodule adds or concatenates the outputs of the tabular data processing submodule and the time series data processing submodule.

[0007] Furthermore, the joint loss function is:

[0008] in, For the binary classification cross-entropy loss of the nutritional diagnostic task, Mean squared error loss for the nutritional therapy recommendation task. The mean squared error loss for prognostic and efficacy evaluation tasks. , , These are the corresponding dynamic balance weights, and the dynamic balance weights are... The calculation method is as follows:

[0009] in For the task k The exponential moving average of historical losses; The label-perceived adjustment of the loss weights in the prognosis and efficacy evaluation module based on the batch label mean refers to adjusting the loss weights based on the mean of the true prognosis labels in the current batch. Adjustment :

[0010] in These are the initial weights before adjustment.

[0011] Furthermore, the joint loss function is:

[0012] in, For the binary classification cross-entropy loss of the nutritional diagnostic task, Mean squared error loss for the nutritional therapy recommendation task. The negatively biased log-likelihood loss of the Cox proportional hazards for the prognostic and efficacy evaluation task. , , These are the corresponding dynamic balance weights, and the dynamic balance weights are... The calculation method is as follows:

[0013] in For the task k The exponential moving average of historical losses; The label-aware adjustment of the loss weights in the prognosis and efficacy evaluation module based on the batch label mean refers to adjusting the event incidence rate in the current batch. Adjustment :

[0014] in These are the initial weights before adjustment.

[0015] Furthermore, the population dataset is constructed for specific disease subgroups such as those with chronic kidney disease, cancer, severe illness, or old age.

[0016] A multi-task learning-based integrated nutritional decision-making system for diagnosis, treatment, and evaluation, the system comprising: The data acquisition module is used to collect model training data and individual prediction data; The data analysis module, connected to the data acquisition module, includes a prediction model modeling submodule and an execution prediction submodule. The prediction model modeling submodule is used to execute the method described above to construct a multi-task learning neural network model, and the execution prediction submodule is used to predict the input data using the multi-task learning neural network model. The results output module, connected to the data analysis module, is used to output the model performance evaluation index of the prediction model modeling submodule or the individual integrated nutrition decision report of the execution prediction submodule.

[0017] Furthermore, the data acquisition module includes a model training data input submodule and an individual prediction data input submodule. The model training data input submodule is used to collect training data from the electronic medical data system, and the individual prediction data input submodule is used to receive model independent variable information input by the user.

[0018] Furthermore, the execution prediction submodule in the data analysis module is used to obtain the prediction category and prediction probability corresponding to each subtask of nutritional diagnosis, nutritional intervention treatment plan, and prognostic evaluation.

[0019] Furthermore, the result output module includes a batch modeling result output submodule and an individual prediction result output submodule. The batch modeling result output submodule is used to output model performance evaluation indicators and visualization charts, and the individual prediction result output submodule is used to output an integrated report that includes nutritional diagnosis, quantitative nutritional treatment plan and prognostic risk warning.

[0020] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method.

[0021] The beneficial effects of this invention are as follows: (1) This invention integrates the four key aspects of clinical nutrition decision-making—examination, diagnosis, treatment, and prognosis—into a unified multi-task learning model for the first time, achieving end-to-end integrated intelligent support from multimodal data input to multidimensional decision output. This overcomes the drawbacks of existing technologies where each aspect is isolated and requires separate judgment by clinicians, greatly improving the coherence and overall efficiency of clinical decision-making.

[0022] (2) This invention can effectively integrate patients' static tabular data and dynamic time series data. By introducing advanced neural network structures such as bidirectional long short-term memory networks and attention mechanisms, it can automatically learn and extract deep and comprehensive phenotypic features related to nutritional status. This solves the problem that existing methods rely on a single static indicator and cannot make full use of heterogeneous information, thereby uncovering complex associations and patterns that are difficult to discover using traditional methods.

[0023] (3) The nutritional therapy recommendation module proposed in this invention can quantify and dynamically predict and recommend the intake targets of nutrients such as energy, protein, and fat based on the individual characteristics of the patient, including their specific diagnostic results and nutritional status. This breaks through the limitations of existing technologies that rely on general guidelines or experience and lack personalized and precise recommendation capabilities, and provides a reliable tool for realizing individualized nutritional therapy.

[0024] (4) This invention innovatively reuses label information from prognostic evaluation tasks through a label-aware adjustment mechanism in the joint loss function. This mechanism enables the model to dynamically adjust its learning focus based on the overall prognostic level of the patient population, achieving not only effective prediction of efficacy and prognostic outcomes but also proactively identifying high-risk groups. This provides a strong basis for formulating and optimizing precise intervention strategies, forming a closed loop of diagnosis and treatment.

[0025] (5) The multi-task learning framework constructed in this invention has high flexibility and scalability. Its modular design allows for flexible selection of feature extraction networks according to different data types, and the diagnosis, treatment, and prognosis sub-tasks can be customized, expanded, or reduced according to actual clinical needs. The framework has good extrapolation capabilities and can be adapted to other clinical decision-making scenarios with similar data collection, situation assessment, intervention decision-making, and effect evaluation processes.

[0026] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the model architecture of the present invention; Figure 2 This is a schematic diagram of the system architecture of the present invention; Figure 3The training and validation loss curves of the model as a whole and the learning curves of each prediction subtask are shown in Embodiment 1 of the present invention. Figure 4 This refers to the changes in model loss for each subtask and the changes in weights of each subtask in the joint loss function in Embodiment 1 of the present invention. Figure 5 This is the performance evaluation result of the classification model for the diagnostic task in Embodiment 1 of the present invention; Figure 6 This is the performance evaluation result of the regression model for treatment recommendation and prognosis tasks in Embodiment 1 of the present invention; Figure 7 The training and validation loss curves of the model as a whole and the learning curves of each prediction subtask are shown in Embodiment 2 of the present invention. Figure 8 This refers to the changes in model loss for each subtask and the changes in weights of each subtask in the joint loss function in Embodiment 2 of the present invention. Figure 9 This is the performance evaluation result of the classification model for the diagnostic task in Embodiment 2 of the present invention; Figure 10 This is the performance evaluation result of the regression model for the treatment recommendation task in Embodiment 2 of the present invention; Figure 11 This is the survival curve for the prognostic task in Embodiment 2 of the present invention. Detailed Implementation

[0028] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0030] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0031] In a first aspect, the present invention provides a method for constructing an integrated nutritional decision-making artificial intelligence model based on multi-task learning, comprising the following steps: 1.1 Establishing a population dataset Inpatients were selected as the modeling population, and variable information was collected through public databases, electronic medical record systems, and manual data entry. Modeling information includes, but is not limited to: 1) Cross-sectional data: static clinical information such as age, gender, height, disease diagnosis, past medical history, and metabolomics test results.

[0032] 2) Time series data: dynamic changes in weight, body mass index, laboratory test indicators (such as albumin and prealbumin), dietary intake records, etc. at multiple time points.

[0033] 3) Task Two (Diagnosis) Labels: Based on standards such as the Global Leaders' Malnutrition Initiative and the Nutrition Risk Screening 2002 Scale, classify and code the nutritional diagnosis results (e.g., malnutrition, normal nutritional status). Alternatively, form continuous values ​​based on nutritional status scores.

[0034] 4) Task 3 (Treatment) Labels: Record the patient's actual intake of nutrients such as energy (kcal / kg body weight / day), protein (g / kg body weight / day), and fat (g / kg body weight / day) as regression task labels. Alternatively, categorize intake levels into insufficient intake, adequate intake, and excessive intake based on quantiles of continuous nutrient intake values ​​or clinical recommendations.

[0035] 5) Task Four (Prognostic) Labels: Use future efficacy evaluation indicators or clinical outcomes as classification or regression labels. For example, discretize the percentage change in weight after one month into multiple levels (e.g., 0-9, a total of ten levels) as a classification label for prognostic assessment. Alternatively, use the patient's risk of death as a continuous label.

[0036] 1.2 Data Preprocessing The sample dataset established in step 1 is preprocessed, including but not limited to imputing missing data using multiple imputation, deleting outliers, excluding samples with missing key modeling information, performing one-hot encoding on multi-class data, standardizing continuous variables, deleting variables with the same values ​​for all samples, and generating new variables.

[0037] 1.3 Splitting the training and test sets The preprocessed sample dataset is randomly divided into a training set and a test set in a certain ratio (e.g., 7:3). The training set is used for model training, hyperparameter tuning, and model selection, while the test set is used to evaluate the model's generalization ability.

[0038] 1.4 Design and Construction of Multi-Task Learning Neural Network Model Architecture To construct a multi-task learning neural network model, see [link / reference]. Figure 1 The model contains four cascaded task modules: 1) Task 1: Feature Extraction Module Tabular data processing submodule: Uses neural networks to process input tabular data (such as a single fully connected layer).

[0039] Time series data processing submodule: Uses neural network structures capable of processing time series data, such as bidirectional long short-term memory networks or the Transformer module.

[0040] Feature fusion submodule: This module fuses the processed tabular features and time series features, for example, by adding dimensions.

[0041] 2) Task Two: Nutritional Diagnosis Module Taking the fused features extracted from Task 1 as input, this module processes information through a combination of one or more neural network components, including attention modules, fully connected layers, activation functions, and layer normalization, to output a classification result for nutritional diagnosis. This module also outputs hidden layer information for use in subsequent tasks.

[0042] 3) Task 3: Nutritional Therapy Recommendation Module Taking the hidden layer information from Task 2 and the fused features from Task 1 as input, this module processes the information through a combination of one or more neural network components, such as attention modules, fully connected layers, activation functions, and layer normalization, to output recommended results for energy, protein, and fat intake targets. This task can be designed as a regression task (predicting specific values) or a classification task (predicting intake levels). Simultaneously, this module also outputs hidden layer information for use in subsequent tasks.

[0043] 4) Task Four: Prognosis and Treatment Evaluation Module Taking the hidden layer information output from Task 3 as input, and combining the hidden layer information output from Task 2 with the fusion features from Task 1, the information is processed through a combination of neural network components, such as one or more attention modules, fully connected layers, activation functions and layer normalization, to output the prediction results of the prognostic evaluation index.

[0044] 5) Module optimization components: In each task module, add neural network components such as residual connections to enhance the model's learning ability and training stability.

[0045] 1.5 Design of Joint Loss Function The design incorporates a joint loss function that dynamically balances the importance of each task. Its core ideas include: a dynamic balancing mechanism, where the weight of a sub-task is automatically increased when its loss value suddenly increases, guiding the model to focus on the sub-task that is currently difficult to learn; and label-aware adjustment, where Task 4 (prognosis and efficacy evaluation) can be implemented as continuous prognosis prediction (prog) or survival analysis (surv) according to clinical needs, both being different implementations within the same task module. For Task 4 (efficacy and prognosis evaluation), its loss weight is dynamically adjusted based on the mean true label value of samples in each batch (representing the overall prognosis level). When the label prediction tends towards a worse outcome (e.g., most patients experience severe weight loss), its weight is automatically reduced; conversely, its weight is increased, enabling the reuse of label information. For common label types in classification, regression, and survival prediction sub-tasks, this invention proposes two methods for implementing the joint loss function: 1) Let the loss for Task 2 (nutritional diagnosis, binary classification task) in the current training batch be... (Binary classification cross-entropy); The loss for Task 3 (protein recommendation, regression task) is... (Mean squared error); The loss for Task 4 (prognostic evaluation) is (Mean squared error, regression task). Define historical loss (exponential moving average):

[0046] in This is the smoothing coefficient. The dynamic balance weight calculation method is as follows:

[0047] k,j {diag, prot, prog}, where diag is an abbreviation for diagnosis, corresponding to Task 2: Nutritional Diagnosis Module, whose function is to classify and diagnose the nutritional status of patients (e.g., determine whether malnutrition exists). prot is an abbreviation for protein, corresponding to Task 3: Nutritional Therapy Recommendation Module, whose function is to recommend individualized protein intake targets (in grams per kilogram of body weight per day). In Example 1, this task is implemented using a regression model to predict specific intake values. prog is an abbreviation for prognosis, corresponding to Task 4: Prognosis and Efficacy Evaluation Module in the document. This task predicts a continuous efficacy indicator, namely the "weight change rate" one month after the patient's discharge. Therefore, the task set includes: diagnosis, protein recommendation, and weight prognosis.

[0048] Label perception regulation (Task 4):

[0049] in This represents the average of the true prognostic labels in the current batch.

[0050] Final weight normalization:

[0051] (In the actual code, the calculation is performed first) Then adjust Task 4, and finally renormalize, which is simplified here to represent the final weights. ) Joint loss function:

[0052] Specific forms of loss for each mission:

[0053]

[0054]

[0055] in For genuine diagnostic labels, To predict probabilities; To predict protein intake values, The actual value; To predict prognostic values, The actual value; This refers to the batch size.

[0056] This loss function achieves: dynamic balancing, which automatically adjusts the weights based on the relative magnitude of the losses for each task, allowing the model to focus on the sub-tasks that are currently difficult to train; and label-aware reuse, which uses the mean information of prognostic labels to adjust the weights of task four, thereby strengthening the learning of prognostic prediction in individuals with mild symptoms.

[0057] 2) Let the loss of Task 2 (nutritional diagnosis, binary classification task) in the current batch be... (Binary classification cross-entropy); The loss for Task 3 (protein recommendation, regression task) is... (Mean squared error); The loss for Task 4 (survival prognosis, survival prediction task) is (Cox proportional risk negative skewed log-likelihood loss). Define historical loss (exponential moving average):

[0058] in This is the smoothing coefficient. ,j diag,prot,surv

[0059] The dynamic balance weight calculation method is as follows:

[0060] ,j diag,prot,surv Surv is an abbreviation for (survival), corresponding to Task 4: Prognosis and Efficacy Evaluation module. In Example 2, it is specifically implemented as a survival analysis task, predicting the patient's survival risk using a Cox proportional hazards model. Therefore, the task set includes: diagnosis, protein recommendation, and survival prognosis.

[0061] Label perception regulation (Task 4):

[0062] in This represents the proportion of events (deaths) that occurred in the current batch.

[0063] Final weight normalization:

[0064] Joint loss function:

[0065] Specific forms of loss for each mission:

[0066]

[0067]

[0068] in For genuine diagnostic labels, To predict probabilities; To predict protein intake values, The actual value; The risk score predicted by the model. For the observation period, For event indication; The batch size is used. This loss function achieves: dynamic balancing, which automatically adjusts the weights based on the relative magnitude of the losses for each task, allowing the model to focus on the subtasks that are currently difficult to train; and label-aware adjustment, which uses event occurrence information to adjust the weights of task four, strengthening survival prediction learning in batches with high event rates.

[0069] 1.6 Model Training and Optimization Using the training set data, the model built in step 2 is trained end-to-end using the backpropagation algorithm and an optimizer (such as Adam). During training, the total loss is calculated using the joint loss function designed in step 3. Based on the performance on the validation set, hyperparameters such as the learning rate and batch size are adjusted, and strategies such as early stopping are employed to prevent overfitting.

[0070] 1.7 Model Selection and Evaluation On the test set, the model with the minimum validation loss is selected, and the trained model is comprehensively evaluated. Evaluation metrics include: 1) Diagnostic tasks: For regression tasks, root mean square error, coefficient of determination, etc. are used. For classification tasks, area under the curve, precision-recall curve, confusion matrix, etc. are used.

[0071] 2) Treatment tasks: For regression tasks, root mean square error, coefficient of determination, etc. are used. For classification tasks, area under the curve, precision-recall curve, confusion matrix, etc. are used.

[0072] 3) Prognostic evaluation tasks: For regression tasks, root mean square error, coefficient of determination, etc. are used; for classification tasks, area under the curve, precision-recall curve, confusion matrix, etc. are used; for survival time data, C-index, time-dependent C-index, survival curve, etc. are used.

[0073] Preferably, in step 1, a dedicated model can be constructed for a specific disease subgroup (such as chronic kidney disease, tumors, severe illness, or the elderly) to improve the model's relevance and accuracy.

[0074] Preferably, in the feature fusion submodule of step 2, an attention mechanism can be used to automatically learn the importance weights of different modal features for the final decision.

[0075] Preferably, in the design of the joint loss function in step 3, a temperature parameter can be introduced to control the sensitivity of label perception adjustment and improve the flexibility of the model.

[0076] Preferably, in the feature extraction module of Task 1 in step 4, a padding and masking mechanism is used for data with different sequence lengths.

[0077] Preferably, in the feature fusion module of Task 1 in step 4, a cross-attention mechanism is used for feature fusion.

[0078] Preferably, in step 4, multiple sets of processes from Task 1 to Task 4 are constructed simultaneously to learn multiple decision-making scenarios in parallel and achieve multidisciplinary assisted decision-making.

[0079] Preferably, in step 4, some tasks from task 2 to task 4 are skipped for modeling, thereby reducing model complexity.

[0080] Secondly, this invention provides an integrated nutritional decision-making system for diagnosis, treatment, and evaluation based on multi-task learning, see [link to relevant documentation]. Figure 2 The system mainly includes the following modules: 2.1 Data Acquisition Module: Model training data input submodule: Used to collect training data required for modeling from electronic medical data systems, local data sources or other data sources, including but not limited to age, gender, height, body mass index, current weight, historical weight at multiple time points, limb skeletal muscle index at multiple time points, basal metabolic rate at multiple time points, malnutrition diagnosis results, cachexia diagnosis results, nutritional risk screening results, sarcopenia diagnosis results, nutritional intervention treatment plan (type, route, dosage, brand, time), and clinical outcomes (death, postoperative complications, weight) as input data.

[0081] Individual Prediction Data Input Submodule: Receives single or batch model independent variable information input by the user as input data.

[0082] 2.2 Data Analysis Module Predictive model modeling submodule: Using the data obtained by the input module 1), perform the modeling steps 1 to 7 in the first aspect to obtain the performance evaluation index of the multi-task model and each sub-task model.

[0083] The prediction submodule performs predictions based on the multi-task model obtained by the data analysis and processing module 1) and the input data acquired by the data acquisition module. It obtains the prediction categories and / or probabilities corresponding to each sub-task, including nutritional diagnosis, nutritional intervention treatment plan, and prognostic evaluation.

[0084] 2.3 Result Output Module Batch modeling result output submodule: Based on the modeling results of data analysis and processing module 1 in step 1, output model performance evaluation indicators and visualization charts (such as subject operating characteristic curves and confusion matrices).

[0085] Individual prediction result output submodule: Based on the calculation results of the data analysis module 2, output an integrated "examination-diagnosis-treatment-evaluation" report for the individual, including nutritional diagnosis, quantitative nutritional treatment plan and prognostic risk warning.

[0086] Example 1: In this embodiment, a method for constructing an integrated nutritional decision-making artificial intelligence model based on multi-task learning for diagnosis, treatment, and evaluation is implemented. See [link to relevant documentation]. Figure 2 It includes the following steps: 3.1 Establishing a population dataset We selected 1000 hospitalized patients and collected individual baseline cross-sectional data including age (years, continuous), sex (binary), body mass index (BMI, continuous), serum albumin (continuous), and C-reactive protein (continuous). We also collected individual historical time-series data, including weight, BMI, and skeletal muscle index (ASMI) at admission, three months prior to admission, one month prior to admission, and at admission. Additionally, we collected labeling data for Tasks 2 through 4: Task 2 nutritional diagnosis was a binary label (presence or absence of malnutrition based on the Global Leaders' Malnutrition Initiative); Task 3 nutritional therapy recommendations were daily protein intake (g / kg body weight / day, continuous); and Task 4 efficacy prognostic evaluation was the rate of weight change one month after discharge (positive for weight gain, negative for weight loss, continuous).

[0087] 3.2 Data Preprocessing The collected sample datasets were preprocessed, and continuous variables were standardized using z-scores. Specifically, each feature dimension of the time series data was standardized separately.

[0088] 3.3 Splitting the training and testing sets The preprocessed sample dataset from step 2 is randomly divided into a training set and a test set in a 7:3 ratio.

[0089] 3.4 Design and Construction of Multi-Task Learning Neural Network Model Architecture Based on the training set partitioned in step 3, a cascaded multi-task deep learning model is constructed in the Python 3.9.11 environment. This model comprises four sequentially executed task modules, with the overall architecture as follows: Figure 1 As shown. Wherein: 1) Task 1: Feature Extraction Module The cross-sectional data processing submodule employs a three-layer fully connected network, with each layer having an output dimension of 128 and using ReLU activation. Each layer is followed by a LayerNorm normalization layer, and the input dimension is 5. The time series data processing submodule uses a bidirectional Long Short-Term Memory (BiLSTM) network with an input dimension of 3, a hidden layer dimension of 64 (128 in total), two layers, and dropout=0.1. The sequence length is fixed at 3, and the output of the last time step is taken as the global feature of the time series. The feature fusion submodule concatenates the cross-sectional feature vector (128-dimensional) with the time series feature vector (128-dimensional), and then passes it through a fully connected network (256 inputs, 128 outputs) and layer normalization to obtain the fused feature vector.

[0090] 2) Task Two: Nutritional Diagnosis Module Using the fused features output from Task 1 as input, the feature weights are first calculated through an attention module (single-layer fully connected + Tanh activation) to obtain weighted features. Then, the features pass through two fully connected layers (128→64→32), each followed by ReLU, layer normalization, and dropout (0.2). The final layer is a binary classification output layer, using the Sigmoid activation function to output the probability p_diag of malnutrition. Simultaneously, the output (32-dimensional) of the second fully connected layer is passed as the hidden state h2 to the downstream task.

[0091] 3) Task 3: Nutritional Therapy Recommendation Module The input consists of the hidden state h2 (32-dimensional) from Task 2 and the fused feature fused (128-dimensional) from Task 1, concatenated to obtain a 160-dimensional vector. This vector passes through an attention layer and a fully connected layer (160→128→64) similar to those used in Task 2, and finally outputs the protein uptake target value prot_pred (a continuous value) through a linear regression layer. Simultaneously, the output of the penultimate layer (64-dimensional) is used as the hidden state h3 and passed to Task 4.

[0092] 4) Task Four: Therapeutic Effect and Prognostic Evaluation Module The input consists of the hidden state h3 (64-dimensional) from Task 3, the hidden state h2 (32-dimensional) from Task 3, and the fused feature (128-dimensional) from Task 1. These three are concatenated to obtain a 224-dimensional vector. This vector then passes through an attention layer and a fully connected layer (224→128→64), and finally outputs the prognostic index prog_pred (weight change rate) through a linear regression layer.

[0093] All modules employ residual connections and layer normalization to mitigate gradient vanishing and accelerate training.

[0094] 3.5 Joint Loss Function: Use the first type of joint loss function described in step 5 of the technical solution.

[0095] 3.6 Model Training and Optimization 1) Hardware environment: NVIDIA GPU RTX 4060 (CUDA available). This example uses GPU acceleration.

[0096] 2) Optimizer: Adam, initial learning rate 0.001, weight decay 0.00001.

[0097] 3) Loss Function: The custom dynamic balancing joint loss function described in step 5 is used. Task 2 uses Binary Cross-Entropy Loss (BCE Loss), while Tasks 3 and 4 use Mean Squared Error Loss (MSE Loss). Loss values ​​for all three tasks are calculated for each batch, and historical losses are maintained using Exponential Moving Average (EMA). The larger the ratio of the current batch loss to the historical loss, the higher the weight of that task is automatically increased, thus guiding the model to focus on sub-tasks that are currently difficult to learn. Label-aware adjustment (Task 4): The mean value μ_prog of the prognostic labels in the current batch is calculated. If μ_prog > -0.2 (i.e., most patients do not experience severe weight loss), the weight of Task 4 is increased by a factor of 1.2; if μ_prog ≤ -0.2 (most patients experience severe weight loss), the weight of Task 4 is decreased to a factor of 0.8. This mechanism enables the reuse of prognostic label information, allowing the model to focus more on prognostic prediction in mild cases.

[0098] 4) Training rounds: 50 epochs, with the total loss calculated on the validation set after each epoch.

[0099] 3.7 Model Selection and Prediction 1) Select and save the model parameters that minimize the validation loss.

[0100] 2) Prediction: Based on the best model obtained in step 1), quickly predict the test set samples.

[0101] 3.8 Results of the Examples 1) Evaluation methods: For the classification task (Task 2), plot the area under the receiver operating characteristic (AUC) curve and its 95% confidence interval (Bootstrap method, 1000 resampling), precision-recall curve and area under the curve (PR-AUC), and confusion matrix. For the regression tasks (Task 3 and Task 4), plot scatter plots of predicted and true values, calculate root mean square error (RMSE), coefficient of determination (R²), and plot residual plots and residual distribution histograms.

[0102] This embodiment verifies the feasibility and superiority of the described cascaded multi-task deep learning system in nutrition decision-making under conditions of simplified features (5-dimensional cross-sectional and 3-dimensional time series with 3 time points) and small sample size (1000 cases). The overall training and validation loss curves of the model, as well as the learning curves of each prediction subtask, are shown below. Figure 3 As the number of training epochs increases, the overall loss of the model gradually decreases, while the learning curves of each subtask gradually increase, indicating that the model has learned the knowledge from the data. The changes in model loss for each subtask and the changes in the weights of each subtask in the joint loss function are shown in [reference needed]. Figure 4 As the number of training rounds increases, the loss of each subtask of the model gradually decreases, while the weights of each subtask in the loss function are dynamically adjusted. The performance evaluation results of the classification model for the diagnostic task (Task 2) show... (See...) Figure 5 The model's predictive performance is good (area under the curve > 0.95). Performance evaluation results of the regression model for treatment recommendations and prognostic tasks (Tasks 3 and 4) show [details omitted]. Figure 6 The model has good predictive performance (coefficient of determination > 0.8).

[0103] This embodiment demonstrates that the model can simultaneously perform three tasks: nutritional diagnosis, protein intake target recommendation, and weight change rate prediction, with all indicators achieving excellent levels. The dynamic balance loss function and label reuse strategy effectively improve the stability of multi-task learning and the accuracy of prognostic prediction. This model can be further deployed as a clinical decision support tool, providing patients with an integrated nutritional management plan from assessment to intervention to prognostic prediction.

[0104] Example 2: This embodiment is largely the same as Embodiment 1. See [link / reference] Figure 1 However, the training sample size, the learning objective for Task 4, and the design of the loss function were changed. A method for constructing an integrated diagnostic, treatment, and assessment nutrition decision-making artificial intelligence model based on multi-task learning includes the following steps: 4.1 Establishing a population dataset We selected 1500 hospitalized patients and collected individual baseline cross-sectional data including age (years, continuous), sex (binary), body mass index (BMI, continuous), serum albumin (continuous), and C-reactive protein (continuous). We also collected individual historical time-series data, including weight, BMI, and skeletal muscle index (ASMI) at admission, three months prior to admission, one month prior to admission, and at admission. Additionally, we collected labeling data for Tasks 2 through 4: Task 2 nutritional diagnosis was a binary label (presence or absence of malnutrition based on the Global Leaders' Malnutrition Initiative); Task 3 nutritional therapy recommendations were daily protein intake (g / kg body weight / day, continuous); and Task 4 survival prognosis prediction was survival time (days, continuous) and event indicators (binary, 1 for death / event occurrence, 0 for censoring).

[0105] 4.2 Data Preprocessing The collected sample datasets were preprocessed, and continuous variables were standardized using z-scores. Specifically, each feature dimension of the time series data was standardized separately.

[0106] 4.3 Splitting the training and testing sets: The preprocessed sample dataset from step 2 is randomly divided into a training set and a test set in a 7:3 ratio, and stratified sampling is performed based on the labels of Task 2 to ensure consistent category distribution.

[0107] 4.4 Multi-task learning neural network model architecture design and construction Based on the training set partitioned in step 3, a cascaded multi-task deep learning model is constructed in the Python 3.9.11 environment. This model comprises four sequentially executed task modules, with the overall architecture as follows: Figure 2 As shown. Wherein: 1) Task 1: Feature Extraction Module The cross-sectional data processing submodule employs a three-layer fully connected network, with each layer having an output dimension of 128 and using ReLU activation. Each layer is followed by a LayerNorm normalization layer, and the input dimension is 5. The time series data processing submodule uses a bidirectional Long Short-Term Memory (BiLSTM) network with an input dimension of 3, a hidden layer dimension of 64 (128 in total), two layers, and dropout=0.1. The sequence length is fixed at 3, and the output of the last time step is taken as the global feature of the time series. The feature fusion submodule concatenates the cross-sectional feature vector (128-dimensional) with the time series feature vector (128-dimensional), and then passes it through a fully connected network (256 inputs, 128 outputs) and layer normalization to obtain the fused feature vector.

[0108] 2) Task Two: Nutritional Diagnosis Module Using the fused features output from Task 1 as input, the feature weights are first calculated through an attention module (single-layer fully connected + Tanh activation) to obtain weighted features. Then, the features pass through two fully connected layers (128→64→32), each followed by ReLU, layer normalization, and dropout (0.2). The final layer is a binary classification output layer, using the Sigmoid activation function to output the probability p_diag of malnutrition. Simultaneously, the output (32-dimensional) of the second fully connected layer is passed as the hidden state h2 to the downstream task.

[0109] 3) Task 3: Nutritional Therapy Recommendation Module The input consists of the hidden state h2 (32-dimensional) from Task 2 and the fused feature fused (128-dimensional) from Task 1, concatenated to obtain a 160-dimensional vector. This vector passes through an attention layer and a fully connected layer (160→128→64) similar to those used in Task 2, and finally outputs the protein uptake target value prot_pred (a continuous value) through a linear regression layer. Simultaneously, the output of the penultimate layer (64-dimensional) is used as the hidden state h3 and passed to Task 4.

[0110] 4) Task Four: Survival Prognosis Prediction Module This module extends traditional prognostic assessment into a survival analysis task. The input consists of the hidden state h3 (64-dimensional) from Task 3, the hidden state h2 (32-dimensional) from Task 2, and the fused feature (128-dimensional) from Task 1. These three are concatenated to obtain a 224-dimensional vector. This vector then passes through an attention layer and a fully connected layer (224→128→64), and finally outputs a risk score (a continuous value without an activation function) directly through a linear layer. A higher risk score indicates a higher risk of death and a shorter expected survival time.

[0111] All modules employ residual connections and layer normalization to mitigate gradient vanishing and accelerate training.

[0112] 4.5 Joint Loss Function: Use the second type of joint loss function described in step 5 of the technical solution.

[0113] 4.6 Model Training and Optimization 1) Hardware environment: NVIDIA GPU RTX 4060 (CUDA available). This example uses GPU acceleration.

[0114] 2) Optimizer: Adam, initial learning rate 0.001, weight decay 0.00001.

[0115] 3) Loss Function: The custom dynamic balancing joint loss function described in step 5 is used. Task 2 uses Binary Cross-Entropy Loss (BCE Loss), Task 3 uses Mean Squared Error Loss (MSE Loss), and Task 4 uses Cox Loss (Negative Biased Log-Likelihood). Loss values ​​for all three tasks are calculated for each batch, and historical losses are maintained using Exponential Moving Average (EMA). The larger the ratio of the current batch loss to the historical loss, the higher the weight of that task is automatically increased, thus guiding the model to focus on sub-tasks that are currently difficult to learn. Label-Aware Adjustment (Task 4): The event occurrence rate (event_rate) in the current batch is calculated. If event_rate > 0.5 (i.e., high-risk batch), the weight of Task 4 is increased by 1.2 times; if event_rate ≤ 0.5 (low-risk batch), the weight of Task 4 is decreased to 0.8 times. This mechanism enables the reuse of survival label information, allowing the model to focus more on survival prediction in high-risk populations.

[0116] 4) Training epochs: 50 epochs. The total loss is calculated on the validation set after each epoch.

[0117] 4.7 Model Selection and Prediction 1) Select and save the model parameters that minimize the validation loss.

[0118] 2) Prediction: Based on the best model obtained in step 1), quickly predict the validation set samples and output the probability of malnutrition, the recommended protein value, and the survival risk score.

[0119] 4.8 Results of the Examples 1) Evaluation Methods: For the classification task (Task 2), plot the area under the receiver operating characteristic (AUC) curve and its 95% confidence interval (Bootstrap method, 1000 resampling), precision-recall curve and area under the curve (PR-AUC), and confusion matrix. For the regression task (Task 3), plot a scatter plot of predicted and true values, calculate the root mean square error (RMSE) and coefficient of determination (R²), and plot the residual plot and residual distribution histogram. For the survival prediction task (Task 4), calculate the consistency index (C-index) and plot Kaplan-Meier survival curves stratified by the median predicted risk.

[0120] This embodiment verifies the feasibility and superiority of the described cascaded multi-task deep learning system in nutrition decision-making under conditions of simplified features (5-dimensional cross-sectional and 3-dimensional time series with 3 time points) and a moderate sample size (1500 cases). The overall training and validation loss curves of the model, as well as the learning curves for each prediction subtask, are shown below. Figure 7As the number of training epochs increases, the overall model loss gradually decreases, while the performance metrics of each subtask (diagnosis AUC, protein R², survival C-index) gradually increase or stabilize, indicating that the model has effectively learned the knowledge from the data. The changes in model loss for each subtask and the changes in the weights of each subtask in the joint loss function are shown in [see...]. Figure 8 As the number of training rounds increases, the loss of each subtask of the model gradually decreases, while the weights of each subtask in the loss function are dynamically adjusted. The performance evaluation results of the classification model for the diagnostic task (Task 2) show... (See...) Figure 9 The model's predictive performance is good (area under the curve > 0.85). The performance evaluation results of the regression model for the treatment recommendation task (Task 3) show [see details]. Figure 10 The model exhibits good predictive performance (coefficient of determination > 0.80). Evaluation results for the survival prediction task (Task 4) show that the model effectively distinguishes between high- and low-risk groups (survival C-index > 0.75), and the Kaplan-Meier curves demonstrate a significant separation in survival probabilities between the high-risk and low-risk groups. (See...) Figure 11 .

[0121] The above is another embodiment of the present invention. This embodiment shows that the model can simultaneously complete three tasks: nutritional diagnosis, protein intake target recommendation, and survival rate prediction, and all indicators reach excellent levels. This further illustrates that the cascaded multi-task learning architecture proposed in this invention has good flexibility and can be adapted to different application scenarios by replacing sub-task modules.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A diagnosis-treatment evaluation integrated nutrition decision-making method based on multi-task learning, characterized in that: Includes the following steps: Establish a population dataset that includes cross-sectional data, time-series data, nutritional diagnostic labels, nutritional therapy recommendation labels, and efficacy and prognostic evaluation labels; The population dataset is preprocessed and divided into training and testing sets; A multi-task learning neural network model is constructed, comprising a feature extraction module, a nutritional diagnosis module, a nutritional therapy recommendation module, and a prognosis and efficacy evaluation module connected in series. The feature extraction module is used to fuse the cross-sectional data and the time-series data. The nutritional diagnosis module takes the fused features output by the feature extraction module as input and outputs hidden layer information. The nutritional therapy recommendation module takes the hidden layer information output by the nutritional diagnosis module and the fused features output by the feature extraction module as input and outputs hidden layer information. The prognosis and efficacy evaluation module takes the hidden layer information output by the nutritional therapy recommendation module, the hidden layer information output by the nutritional diagnosis module, and the fused features output by the feature extraction module as input. Design a joint loss function, which is used to dynamically balance the loss weights of the nutrition diagnosis module, the nutrition therapy recommendation module, and the prognosis and efficacy evaluation module, and to perform label-aware adjustment on the loss weight of the prognosis and efficacy evaluation module based on the batch label mean. The joint loss function is: wherein, is the binary cross-entropy loss for the nutritional diagnosis task, is the mean squared error loss for the nutritional therapy recommendation task, is the mean squared error loss for the prognosis and efficacy evaluation task, , , are the corresponding dynamic balance weights, respectively, the calculation method of the dynamic balance weights is: in For the task k The exponential moving average of historical losses; The label-perceived adjustment of the loss weights in the prognosis and efficacy evaluation module based on the batch label mean refers to adjusting the loss weights based on the mean of the true prognosis labels in the current batch. Adjustment : in The initial weights before adjustment; The multi-task learning neural network model is trained and optimized using the training set and the joint loss function. The trained and optimized model is selected and evaluated using the test set.

2. The integrated nutritional decision-making method based on multi-task learning for diagnosis, treatment, and evaluation according to claim 1, characterized in that: The feature extraction module includes a tabular data processing submodule, a time series data processing submodule, and a feature fusion submodule. The tabular data processing submodule uses a fully connected layer to process the cross-sectional data. The time series data processing submodule uses a bidirectional long short-term memory network (BiLSTM) or a Transformer module to process the time series data. The feature fusion submodule adds or concatenates the outputs of the tabular data processing submodule and the time series data processing submodule.

3. The integrated nutritional decision-making method based on multi-task learning for diagnosis, treatment, and evaluation according to claim 1, characterized in that: The joint loss function is replaced with: in, For the binary classification cross-entropy loss of the nutritional diagnostic task, Mean squared error loss for the nutritional therapy recommendation task. The negatively biased log-likelihood loss of the Cox proportional hazards for the prognostic and efficacy evaluation task. , , These are the corresponding dynamic balance weights, and the dynamic balance weights are... The calculation method is as follows: in For the task k The exponential moving average of historical losses; The label-aware adjustment of the loss weights in the prognosis and efficacy evaluation module based on the batch label mean refers to adjusting the event incidence rate in the current batch. Adjustment : in These are the initial weights before adjustment.

4. The integrated nutritional decision-making method based on multi-task learning for diagnosis, treatment, and evaluation according to claim 1, characterized in that: The population dataset was constructed for specific disease subgroups of people with chronic kidney disease, cancer, severe illness, or old age.

5. A nutritional decision-making system integrating diagnosis, treatment, and evaluation based on multi-task learning, characterized in that: The system includes: The data acquisition module is used to collect model training data and individual prediction data; The data analysis module, connected to the data acquisition module, includes a prediction model modeling submodule and an execution prediction submodule. The prediction model modeling submodule is used to execute the method of claim 1 or 2 to construct a multi-task learning neural network model, and the execution prediction submodule is used to predict the input data using the multi-task learning neural network model. The results output module, connected to the data analysis module, is used to output the model performance evaluation index of the prediction model modeling submodule or the individual integrated nutrition decision report of the execution prediction submodule.

6. The integrated nutritional decision-making system for diagnosis, treatment, and evaluation based on multi-task learning according to claim 5, characterized in that: The data acquisition module includes a model training data input submodule and an individual prediction data input submodule. The model training data input submodule is used to collect training data from the electronic medical data system, and the individual prediction data input submodule is used to receive model independent variable information input by the user.

7. The integrated nutritional decision-making system for diagnosis, treatment, and evaluation based on multi-task learning according to claim 5, characterized in that: The execution prediction submodule in the data analysis module is used to obtain the prediction category and prediction probability corresponding to each subtask of nutritional diagnosis, nutritional intervention treatment plan and prognosis evaluation.

8. The integrated nutritional decision-making system for diagnosis, treatment, and evaluation based on multi-task learning according to claim 5, characterized in that: The result output module includes a batch modeling result output submodule and an individual prediction result output submodule. The batch modeling result output submodule is used to output model performance evaluation indicators and visualization charts, and the individual prediction result output submodule is used to output an integrated report that includes nutritional diagnosis, quantitative nutritional treatment plan and prognostic risk warning.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method as described in claim 1 or 2.

Images