An adaptive ordering method and system

Abstract

The application discloses a kind of self-adapting ordering method and system.Collect the electronic medical record of patient brain disease and antiepileptic drug prescription, extract brain nutrition constraint characteristics such as the upper limit of MCT, the lower limit of B6, and drug-food conflict, emotional trigger risk and other interaction characteristics, splice into multidimensional brain-food feature vector, calculate meal antiepileptic adaptation degree score;When compliance menu is empty, constraint relaxation is automatically executed based on conflict type;In touchless interface, extract lateralization index, θ / β power ratio and ERD characteristics by fNIRS and electroencephalogram collection amygdala blood oxygen and electroencephalogram signal, identify and confirm intention by abnormal detection model, generate double-factor confirmation instruction combined with face authentication.The application quantifies drug interaction and nutrition constraint jointly, realizes self-relaxation under over-constraint and non-sense confirmation in high-noise environment, improves the compliance and interactive convenience of antiepileptic diet management.

Description

Technical Field

[0001] This invention relates to the field of adaptive ordering technology, specifically to an adaptive ordering method and system. Background Technology

[0002] Currently, intelligent ordering systems are being gradually adopted in hospitals, airports, and nursing homes. However, significant technological gaps remain in addressing the specific dietary management needs of patients with epilepsy and other brain diseases. In clinical practice, antiepileptic treatment often requires meeting the strict nutritional parameters of the ketogenic diet (such as the fat-to-protein / carbohydrate ratio, upper limit of medium-chain triglycerides, and lower limit of vitamin B6) as well as the interactions between antiepileptic drugs and specific foods (e.g., the absolute contraindication between sodium valproate and grapefruit, and the synergistic sedative effect of cholecystokinin and caffeine). However, existing systems can only screen a few macro-nutritional indicators (such as total calories and fat) and cannot jointly quantify and assess multiple constraints (drug-food conflicts, nutrient restrictions, and the risk of emotion-food triggers). This forces patients and their families to rely on manual verification, which is both time-consuming and prone to errors. Incorrect; more challenging is that when multiple constraints overlap, the system often falls into an over-constraint predicament of "no food options available." Existing technologies can only simply indicate no results or require users to manually relax the conditions, lacking an intelligent mechanism to automatically relax constraints based on clinical safety boundaries. Furthermore, in high-noise or high-cognitive-load environments such as airport waiting areas and operating room waiting areas, traditional touchscreens or voice interactions are not only severely affected by noise but also require explicit confirmation actions from users, which is extremely unfriendly to epilepsy patients who may have pre-existing cognitive impairment. Existing contactless gesture recognition also relies on deliberate actions and cannot achieve truly covert intent perception. Therefore, how to incorporate antiepileptic drug-food interactions and ketogenic nutritional constraints into a unified automated detection framework, how to achieve clinically acceptable constraint relaxation when over-constrained, and how to complete two-factor authentication of intent confirmation through non-invasive brain signal acquisition in high-noise environments have become urgent technical challenges to be solved in this field. Summary of the Invention

[0003] The technical problems to be solved by this invention are: how to incorporate antiepileptic drug-food interactions and ketogenic nutritional constraints into a unified automated detection framework, how to achieve clinically acceptable constraint relaxation when there is over-constraint, and how to complete two-factor authentication of intent confirmation through non-invasive brain signal acquisition in a high-noise environment.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] An adaptive ordering method includes the following steps:

[0006] Step 1: Antiepileptic Dietary Conflict Detection Steps: In response to a user's food ordering request, collect the user's electronic medical record summary information on brain diseases and antiepileptic drug prescription information; extract brain nutritional constraint features based on the brain disease electronic medical record summary information, which includes at least one of the following: ketogenic diet ratio requirements, upper limit of medium-chain triglycerides, and lower limit of daily vitamin B6 intake; extract AEDs-food interaction features based on the antiepileptic drug prescription information, which includes conflict relationship features between at least one antiepileptic drug and prohibited or reduced-efficacy foods, as well as emotion-related information obtained from the amygdala-hippocampal circuit functional status assessment. Food-triggered risk level; concatenating the brain nutrition constraint features with the AEDs-food interaction features to form a multidimensional brain-food feature vector; obtaining a meal ordering candidate set, which includes at least one candidate meal's brain nutrition data and meal ingredient composition data; using the vector dot product of the multidimensional brain-food feature vector and the meal's brain nutrition data as the base score for anti-epileptic fit, using the vector dot product of the multidimensional brain-food feature vector and the meal's ingredient composition data as the AEDs-food conflict penalty score, and subtracting the AEDs-food conflict penalty score from the base score for anti-epileptic fit to generate an anti-epileptic fit score for each candidate meal;

[0007] Step 2: Compliant Menu Generation Steps: Traverse the candidate menu set and select candidate menus with anti-epileptic fit scores higher than a first preset threshold to construct a compliant candidate subset; In response to the empty set detection result of the compliant candidate subset, look up the pre-generated constraint relaxation strategy mapping table based on the conflict type index and extract the corresponding constraint relaxation strategy; Perform local screening and repair operations on each candidate menu in the candidate menu set according to the constraint relaxation strategy until a non-empty compliant candidate subset that meets all the brain nutrition constraint features and the AEDs-food interaction features is generated;

[0008] Step 3: Amygdala-linked concealed confirmation step: In response to the selection of any candidate meal from the compliant candidate subset, the user's blood oxygen dynamic signals in the prefrontal cortex and bilateral amygdala regions are collected via functional near-infrared spectroscopy in a contactless interactive interface, while the user's EEG signals are also collected; the amygdala activation lateralization index, EEG theta / β power ratio, and event-related desynchronization features are extracted; the extracted features are input into a pre-trained anomaly detection model to generate a user confirmation intent score; in response to the confirmation intent score being lower than a second preset threshold, a first authentication factor is generated; simultaneously, a second authentication factor is generated based on the user's iris or facial features; the first authentication factor and the second authentication factor are logically ANDed to generate a confirmation completion instruction.

[0009] Furthermore, in the above adaptive ordering method, the antiepileptic fit score is calculated according to the following formula:

[0010]

[0011] in, Indicates the first Antiepileptic suitability score of each candidate meal;

[0012] It is a multidimensional brain-food feature vector that includes the brain nutritional constraint features and the AEDs-food interaction features;

[0013] For the first The brain nutrition data vector of each candidate meal, its dimension and The components that are identical and correspond to brain nutritional constraint features are standardized using Z-score.

[0014] For the first The conflict vector of ingredient composition for each candidate dish, its dimension and The same value, and the component corresponding to the AEDs-food interaction feature is 0 or 1, indicating whether there is a corresponding conflicting food or emotional risk triggering factor.

[0015] This is the preset penalty coefficient for drug-food conflicts.

[0016] Furthermore, in the above adaptive ordering method, the constraint relaxation strategy mapping table stores the mapping relationship between conflict types and constraint relaxation operations. The constraint relaxation operation is selected from at least one of the following strategies: reducing the ketogenic diet ratio requirement to the original value minus 0.8, increasing the upper limit of medium-chain triglycerides to 1.3 times the original value, reducing the lower limit of daily vitamin B6 intake to 0.8 times the original value, and relaxing the conflict determination threshold in the AEDs-food interaction feature.

[0017] Furthermore, in the above adaptive ordering method, the amygdala activation lateralization index is calculated according to the following formula:

[0018]

[0019] in, Indicates the amygdala activation lateralization index. and These represent the amplitudes of blood oxygen concentration changes in the scalp regions corresponding to the right and left amygdala, as measured by functional near-infrared spectroscopy.

[0020] The EEG The power ratio is calculated using the following formula:

[0021]

[0022] in, Indicates brainwave Power ratio, and They represent the central preamble region. Electrode position at frequency band and Power spectral density of the frequency band;

[0023] The event-related desynchronization features are calculated according to the following formula:

[0024]

[0025] in, Indicates event-related desynchronization characteristics. This represents the average amplitude of the μ-band EEG when the user is looking at the selected food item. This represents the reference amplitude when the user is at rest with their eyes closed.

[0026] Furthermore, in the above adaptive ordering method, the amplitude of the blood oxygen concentration change is calculated using a modified Beer-Lambert law:

[0027]

[0028] in, This indicates changes in oxyhemoglobin concentration. The difference in extinction coefficients. This is the differential path length factor. The distance between the light source and the detector. Baseline light intensity, To measure light intensity.

[0029] Furthermore, in the above adaptive ordering method, the anomaly detection model is a semi-supervised anomaly detection model built on a single-class support vector machine; the pre-trained anomaly detection model is trained using a standard brain-body dynamic feature sample set of multiple users in a normal confirmation intention state, where the normal confirmation intention state is defined as the feature pattern when a user clearly makes a food selection in a conscious state without epileptic aura or anxiety interference.

[0030] Furthermore, in the above adaptive ordering method, the second preset threshold is determined by maximizing the F1 score on the validation set, and the first preset threshold is determined by the maximum point of the Youden index in the ROC curve analysis.

[0031] The present invention also relates to an adaptive ordering system, comprising:

[0032] The data acquisition unit is used to collect the user's electronic medical record summary information on brain diseases and anti-epileptic drug prescription information in response to the user's order request;

[0033] The feature extraction unit is used to extract brain nutritional constraint features based on the brain disease electronic medical record summary information, extract AEDs-food interaction features based on the antiepileptic drug prescription information, and concatenate the brain nutritional constraint features with the AEDs-food interaction features to form a multidimensional brain-food feature vector.

[0034] The menu generation engine is used to obtain a set of ordering candidates, generate an anti-epileptic fit score for each candidate, traverse the set of ordering candidates to filter candidates with anti-epileptic fit scores higher than a first preset threshold to construct a compliant candidate subset, respond to the empty set detection result of the compliant candidate subset, look up a pre-generated constraint relaxation strategy mapping table based on the conflict type index, extract the corresponding constraint relaxation strategy, and perform local filtering and repair operations on each candidate in the set of ordering candidates according to the constraint relaxation strategy until a non-empty compliant candidate subset is generated;

[0035] The amygdala-linked concealment confirmation module is used to collect dynamic brain signals from users through functional near-infrared spectroscopy and electroencephalography (EEG) in a contactless interactive interface, extracting the amygdala activation lateralization index and EEG data. The power ratio and event-related desynchronization features are used to input the extracted features into a pre-trained anomaly detection model to generate a user confirmation intent score. In response to the confirmation intent score being lower than a second preset threshold, a first authentication factor is generated, and a second authentication factor is generated based on the user's iris or facial features. The first authentication factor and the second authentication factor are then logically ANDed to generate a confirmation completion instruction.

[0036] In the above adaptive ordering system, the amygdala-linked concealed confirmation module includes:

[0037] The brain signal acquisition submodule includes a functional near-infrared spectroscopy probe array and a dry electrode EEG cap. The probe array covers the scalp region corresponding to the prefrontal lobe and bilateral temporal lobes and amygdala. The dry electrode EEG cap includes at least Fz, C3, C4 electrodes and bilateral mastoid reference electrodes.

[0038] The signal processing submodule is used to calculate the amygdala activation lateralization index, EEG theta / β power ratio, and event-related desynchronization features;

[0039] The anomaly detection submodule is used to input the above features into a pre-trained anomaly detection model, generate a confirmation intent score, and generate a first authentication factor in response to the score being lower than a second preset threshold.

[0040] The fusion authentication submodule is used to perform a logical AND operation between the first authentication factor and a second authentication factor based on facial features or iris to generate a confirmation completion instruction.

[0041] The beneficial effects of this invention are as follows: First, it is the first to incorporate the interaction between antiepileptic drugs and food and the nutrient constraints of the ketogenic diet into a unified vector dot product scoring framework. By introducing a penalty coefficient λ to dynamically balance the efficacy and safety of the formula, it solves the technical problem that existing nutritional meal systems, which only classify diseases coarsely, cannot handle refined multi-source constraints and dynamic drug and food contraindications, and achieves a processing speed of seconds. Second, addressing the over-constraint contradiction of "no meal options" caused by excessively strong medical constraints, it creatively introduces a constraint relaxation strategy mapping table triggered by conflict types. This ensures that feasible meal plans always exist while guaranteeing patient safety. Existing ordering systems can only indicate "no matching items" or require users to manually relax the conditions, failing to achieve this autonomous, precise, and traceable soft constraint repair. Third, it breaks through the mindset of traditional interaction modes by incorporating pupil dynamic indicators (lateralization index, etc.) originally used in cognitive psychology and driver fatigue monitoring. The power ratio and event-related desynchronization were transferred across boundaries to the recognition of order confirmation intent. Combined with contactless acquisition of fNIRS and EEG, covert two-factor authentication without any explicit gestures or voice was achieved in high-noise environments. This not only eliminated physical contact and additional operational burden, but also significantly outperformed existing solutions for voice or gesture recognition. Fourth, the constraint relaxation strategy not only solved the usability problem, but also reduced the frequency of epileptic seizures caused by poor patient compliance in practical applications. The covert confirmation mechanism completed identity authentication and intent authentication at the same time, without any active cooperation from the user, which is especially suitable for special scenarios such as the epileptic prodromal period or high-noise environments such as airports. Attached Figure Description

[0042] Figure 1 This is a flowchart of the adaptive ordering method of the present invention;

[0043] Figure 2 This is a block diagram of the adaptive ordering system of the present invention. Detailed Implementation

[0044] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0045] Example 1

[0046] An adaptive meal ordering method for anti-epileptic diets for patients with brain diseases;

[0047] This embodiment uses Xiaoming, a child suffering from epilepsy, and his mother in a ward of an epilepsy specialist hospital as an example to illustrate the specific implementation of the present invention.

[0048] Step 1: Antiepileptic diet conflict detection procedure;

[0049] This step corresponds to the "anti-epileptic diet conflict detection step" in claim 1, and is used to calculate the anti-epileptic suitability score of each candidate meal based on the user's electronic medical record of brain disease and anti-epileptic drug prescription.

[0050] Step 1.1: Data Acquisition;

[0051] The child's mother clicked the "Start Ordering" button on the contactless ordering terminal next to the bed. In response to the order request, the system performed the following data collection operations:

[0052] Collection of summary information from electronic medical records for brain diseases: The system connects to the hospital's HIS system via the HL7 / FHIR interface to obtain the structured electronic medical record of the patient, Xiaoming. The extracted content includes: disease diagnosis (G40.309 drug-resistant focal epilepsy), ketogenic diet prescription (classic 4:1 ketogenic diet, starting on the 5th day of current hospitalization), upper limit of daily intake of medium-chain triglycerides (MCT) (30g, because the child's weight is greater than 30kg), and lower limit of daily intake of vitamin B6 (2mg, because sodium valproate can interfere with B6 metabolism).

[0053] Information on antiepileptic drug prescriptions was collected: the current dosages of sodium valproate (20 mg / kg) and crobazal (0.5 mg / kg) were extracted from the medical order system. Simultaneously, known drug-food interaction data were obtained from a drug-food conflict knowledge base, including: absolute contraindication between sodium valproate and grapefruit juice (conflict weight 1.0); strong recommendation to avoid sodium valproate and high-tyramine foods (such as aged cheese and fermented soy products) (conflict weight 0.7); and recommended limit on crobazal and excessive caffeine intake (conflict weight 0.4). Furthermore, using Xiaoming's previous EEG and functional magnetic resonance imaging (fMRI) data, the functional status of the amygdala-hippocampal circuit was assessed: resting-state fMRI was used to calculate the functional connectivity strength between the bilateral amygdala and hippocampus / prefrontal cortex. Connectivity strength below the normal reference value (mean - 1.5 standard deviation) was defined as high risk, resulting in an emotion-food trigger risk level of "high" (quantified as 0.8, range 0-1).

[0054] Step 1.2: Feature extraction and multidimensional brain-food feature vector generation;

[0055] Based on the information collected above, the feature extraction unit extracts the following features:

[0056] Characteristics of brain nutritional constraints: ketogenic diet ratio requirements (Dimensionless ratio); Upper limit of medium-chain triglycerides Lower limit of daily intake of vitamin B6 To eliminate the influence of dimensions, these three features are standardized using Z-scores. Standardization parameters... (mean) and (Standard deviation) was obtained from statistical analysis of corresponding characteristic data of 1000 epilepsy patients stored in the hospital information system over the past two years: , ; , ; , Standardized formula .

[0057] AEDs - Food Interaction Features: The original 0-1 values ​​are preserved without Z-score standardization to ensure the positive bias of the penalty terms. Specifically, this includes: Grapefruit Conflict Weights. High-tyramine food conflict weight Caffeine conflict weight Emotion-Food Trigger Risk Level .

[0058] The above features are concatenated to form a 7-dimensional multidimensional brain-food feature vector. :

[0059]

[0060] The first three components are the standardized components. , , The last four components are the original weight values.

[0061] Step 1.3: Obtaining the candidate set of dishes and vectorizing the dishes;

[0062] The system retrieves a candidate set of lunches for the day from the cafeteria's food database, containing 15 candidate dishes. Each dish has its own nutritional data (raw values) and a conflict flag for its ingredient composition. The nutritional data uses the same data as the user's characteristics. and Standardization is performed to obtain standardized nutritional vectors for the meals. (7 dimensions, the last four dimensions are 0). Conflict vector of food ingredient composition. It is 7-dimensional, with the first three dimensions corresponding to brain nutrition-related conflict markers (set to 0 in this embodiment), and the last four dimensions corresponding to [ , , , ], which can take the value 0 or 1.

[0063] Let's take two typical dishes as examples:

[0064] Meal A: "MCT-stir-fried broccoli with egg and cheese pancake". Original nutritional data: Ketogenic ratio MCT content Vitamin B6 content After standardization: , , Therefore .

[0065] Item B: "Grapefruit Chicken Salad with Caffeine Bar". Original Nutritional Data: MCT content Vitamin B6 content After standardization: , , Therefore Conflict vector: Contains grapefruit ( High-tyramine foods = 0, caffeine levels are excessive ( Emotional risk trigger ( ), so .

[0066] Step 1.4: Calculation of antiepileptic fit score;

[0067] Calculate the antiepileptic fit score for each candidate meal:

[0068]

[0069] in, (Penalty coefficient, obtained through F1 optimization using historical data from 50 children). Specific calculation:

[0070] Meal A:

[0071]

[0072]

[0073]

[0074] Item B:

[0075]

[0076]

[0077]

[0078] The first preset threshold is 0.5 (based on ROC curve analysis using 200 positive and 200 negative samples, at the maximum Youden index). Dish A has a score of 1.81875 > 0.5, while Dish B has a score of -10.859375 < 0.5. The system includes all dishes with scores higher than 0.5 in the compliance candidate subset.

[0079] Step 2: Steps to generate a compliant menu;

[0080] This step is used to trigger the constraint relaxation mechanism when the first screening result is empty, ensuring that a non-empty compliant candidate subset is generated.

[0081] Step 2.1: Initial screening results:

[0082] After reviewing 15 candidate dishes, the system initially selected 7 dishes with a rating higher than 0.5 to form a compliant candidate subset. Under normal meal service conditions, this subset is not empty, and the system directly displays it to the user for selection.

[0083] Step 2.2: Empty set detection and constraint relaxation:

[0084] In another extreme case (e.g., the proportion of ketogenic foods served in the kitchen on the same day is generally low), if the system detects that the compliant candidate subset is empty (the empty set detection result is true), then the constraint relaxation mechanism is triggered.

[0085] The system maintains a conflict type index table, automatically identifying the conflict type based on the current cause of the empty set. For example, by analyzing the rating components of all rejected dishes, it discovers the conflict types of all dishes. All were below the threshold, resulting in a penalty of 0, primarily due to the ketogenic ratio component (the standardized ketogenic ratio of all candidate meals). If the conflict type is "TYPE_KETO_LOW", then the conflict type is determined to be "TYPE_KETO_LOW".

[0086] The system pre-generates a constraint relaxation policy mapping table, which stores the correspondence between conflict types and relaxation operations. For "TYPE_KETO_LOW", the corresponding relaxation operation is: reduce the ketogenic ratio requirement from the current value to... (That is, from 4.0 to 3.2). This relaxation level is based on a clinical study (NEJM 2018) showing that a 3:1 ketogenic ratio is still effective for most children.

[0087] The system performs a relaxation operation: update The value is 3.2; recalculate the standardized value. and update The first component is -0.375, and the remaining components remain unchanged. Then, a new... Recalculate all candidate dishes The system then re-filters the samples. After relaxation, the system finds at least one meal (e.g., a meal with a ketogenic ratio of 3.5:1 has a positive score higher than 0.5), generating a non-empty compliant candidate subset. The system displays the samples sorted by score and simultaneously sends a notification to the nutritionist's terminal: "The ketogenic ratio requirement has been automatically reduced from 4:1 to 3.2:1. Please confirm whether the child's condition allows it."

[0088] Step 3: Amygdala linkage concealment confirmation step;

[0089] This step is used to collect brain signals through a contactless interactive interface after the user selects a meal, confirm the user's intention, and generate a confirmation command.

[0090] Step 3.1: Food Selection and Contactless Interaction:

[0091] After Xiaoming's mother selected dish A from the compliant candidate subset, the system entered the hidden confirmation phase. The contactless interface integrates aerial projection and floating touch technology, requiring no additional buttons or voice prompts.

[0092] Step 3.2: Brain signal acquisition and preprocessing;

[0093] The system acquires dynamic brain signals from the user using functional near-infrared spectroscopy (fNIRS) and electroencephalography (EEG). Xiaoming's mother wore a disposable dry electrode EEG cap (electrode positions: Fz, C3, C4, with reference electrodes placed bilaterally on the mastoid process), with a sampling rate of 250 Hz and a bandpass filter of 0.5-50 Hz. Simultaneously, the fNIRS probe array covered the bilateral temporal lobe regions: probes were placed at T3 and T4 positions (corresponding to the scalp projections of the left and right amygdala), with a light source-detector spacing of 30 mm. The device used a dual-wavelength (690 nm and 830 nm) continuous wave system with a sampling rate of 10 Hz. The system instructed the user to naturally gaze at the selected food icon projected in the air for approximately 1.5 seconds, during which signals were automatically acquired. The system simultaneously acquired a 30-second resting, closed-eye EEG as a baseline (acquired before ordering began).

[0094] EEG preprocessing: Independent component analysis (ICA) was used to remove electrooculography (EOG) artifacts, identifying and removing components related to blinking and eye movement; a bandpass filter (0.5-50 Hz) and a notch filter (50 Hz) were used; time periods with amplitudes exceeding ±100 μV were marked as artifacts and removed (removal rate not exceeding 10%).

[0095] fNIRS blood oxygenation calculation: Changes in oxyhemoglobin concentration are calculated using a modified Beer-Lambert law. :

[0096]

[0097] in, (Difference in extinction coefficient) (Differential path length factor) (Light source-detector spacing) Baseline light intensity (average value 500ms before fixation). To measure light intensity.

[0098] Step 3.3: The signal processing submodule calculates the following three key features in real time:

[0099] amygdala activation lateralization index ,in and The values ​​represent the amplitude of blood oxygen concentration changes (peak value minus baseline) measured by the fNIRS channels on the right side (T4) and left side (T3), respectively. In this example... , , .

[0100] EEG Power ratio :

[0101]

[0102] in, and They are respectively Electrode in Frequency band (4-8 Hz) and Power spectral density in the frequency band (13-30 Hz) (Welch method, FFT point count 512). In this example... , ,have to .

[0103] Event-related desynchronization features :

[0104]

[0105] During fixation Average amplitude of EEG in frequency band (8-12 Hz) (mean values ​​of C3 and C4). This is the reference amplitude at rest with eyes closed. In this example... , ,have to .

[0106] Step 3.4: Pre-trained anomaly detection model and intent verification determination;

[0107] The above three features are combined into a feature vector. The pre-trained anomaly detection model employs a semi-supervised model based on a one-class support vector machine (One-Class SVM). Training process: 35 healthy adults were recruited, and samples confirming intent (positive class, 1200 samples) and non-confirming intent (negative class, 1000 samples) were collected in a simulated food ordering environment. The same features were extracted, and the model was trained only on the positive class using Python scikit-learn's OneClassSVM (RBF kernel, nu=0.05, gamma=0.1). The model outputs the decision function value. Mapped to confirmation intention score Anomaly detection score The second preset threshold is set to 0.7 (determined by maximizing the F1 score on the validation set).

[0108] In this example, the input get ,but , .because The model determines that the user is in a normal state of confirmation intent and generates the first authentication factor. .

[0109] Step 3.5: Generate two-factor authentication and confirmation completion instructions;

[0110] Simultaneously, a visible light camera captures the user's facial image and compares it with a pre-stored facial template of the patient's caregiver. A 512-dimensional feature vector is extracted using a ResNet50-based convolutional neural network, and cosine similarity is calculated with a threshold of 0.85. If a match is successful, a second authentication factor is generated. .

[0111] The hidden confirmation module executes the following logic and operations: The system generates a confirmation instruction, automatically submits the order, and prompts, "Order confirmed, meal will be delivered to the ward within 20 minutes." The entire process requires no explicit gestures or voice input from the user.

[0112] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An adaptive ordering method, characterized in that, Includes the following steps: Step 1: Antiepileptic diet conflict detection steps: In response to the user's order request, collect the user's electronic medical record summary information on brain diseases and antiepileptic drug prescription information; Brain nutrition constraint features are extracted based on the electronic medical record summary information of the brain disease. The brain nutrition constraint features include at least one of the following: ketogenic diet ratio requirements, upper limit of medium chain triglycerides, and lower limit of daily vitamin B6 intake. AEDs-food interaction features are extracted based on the prescription information of the antiepileptic drugs. The AEDs-food interaction features include the conflict relationship between at least one antiepileptic drug and prohibited or reduced-efficacy foods, as well as the emotion-food trigger risk level obtained by the functional status assessment of the amygdala-hippocampal circuit. The brain nutrition constraint features are concatenated with the AEDs-food interaction features to form a multidimensional brain-food feature vector; a meal ordering candidate set is obtained, which includes the brain nutrition data and food ingredient data of at least one candidate meal; the vector dot product of the multidimensional brain-food feature vector and the brain nutrition data of the meal is used as the basic score for anti-epileptic fit, and the vector dot product of the multidimensional brain-food feature vector and the food ingredient data of the meal is used as the AEDs-food conflict penalty score; the anti-epileptic fit score is generated by subtracting the AEDs-food conflict penalty score from the basic score for anti-epileptic fit. Step 2: Compliant Menu Generation Steps: Traverse the set of ordering candidates, filter the candidate dishes whose anti-epileptic fit scores are higher than the first preset threshold to construct a compliant candidate subset; in response to the empty set detection result of the compliant candidate subset, look up the pre-generated constraint relaxation strategy mapping table based on the conflict type index, and extract the corresponding constraint relaxation strategy; For each candidate meal in the ordering candidate set, a local screening and repair operation is performed according to the constraint relaxation strategy until a non-empty compliant candidate subset that meets all the brain nutrition constraint features and the AEDs-food interaction features is generated. Step 3: Amygdala-linked concealed confirmation step: In response to the selection of any candidate meal from the compliant candidate subset, the user's blood oxygen dynamic signals in the prefrontal cortex and bilateral amygdala regions are collected via functional near-infrared spectroscopy in a contactless interactive interface, while the user's EEG signals are also collected; the amygdala activation lateralization index, EEG theta / β power ratio, and event-related desynchronization features are extracted; the extracted features are input into a pre-trained anomaly detection model to generate a user confirmation intent score; in response to the confirmation intent score being lower than a second preset threshold, a first authentication factor is generated; simultaneously, a second authentication factor is generated based on the user's iris or facial features; the first authentication factor and the second authentication factor are logically ANDed to generate a confirmation completion instruction.

2. The adaptive ordering method according to claim 1, characterized in that, The antiepileptic suitability score is calculated according to the following formula: in, Indicates the first Antiepileptic suitability score of each candidate meal; It is a multidimensional brain-food feature vector that includes the brain nutrition constraint features and the AEDs-food interaction features; For the first The brain nutrition data vector of each candidate meal, its dimension and The components that are identical and correspond to brain nutritional constraint features are standardized using Z-score. For the first The conflict vector of ingredient composition for each candidate dish, its dimension and The same value, and the component corresponding to the AEDs-food interaction feature is 0 or 1, indicating whether there is a corresponding conflicting food or emotional risk triggering factor. This is the preset penalty coefficient for drug-food conflicts.

3. The adaptive ordering method according to claim 1, characterized in that, The constraint relaxation strategy mapping table stores the mapping relationship between conflict types and constraint relaxation operations. The constraint relaxation operations are selected from at least one of the following strategies: reducing the ketogenic diet ratio requirement to the original value minus 0.8, increasing the upper limit of medium chain triglycerides to 1.3 times the original value, reducing the lower limit of daily vitamin B6 intake to 0.8 times the original value, and relaxing the conflict determination threshold in the AEDs-food interaction characteristics.

4. The adaptive ordering method according to claim 1, characterized in that, The amygdala activation lateralization index is calculated according to the following formula: in, Indicates the amygdala activation lateralization index. and These represent the amplitudes of blood oxygen concentration changes in the scalp regions corresponding to the right and left amygdala, as measured by functional near-infrared spectroscopy. The EEG The power ratio is calculated using the following formula: in, Indicates brainwave Power ratio, and They represent the central preamble region. Electrode position at frequency band and Power spectral density of the frequency band; The event-related desynchronization features are calculated according to the following formula: in, Indicates event-related desynchronization characteristics. This represents the average amplitude of the μ-band EEG when the user is looking at the selected food item. This represents the reference amplitude when the user is at rest with their eyes closed.

5. The adaptive ordering method according to claim 1, characterized in that, The amplitude of the blood oxygen concentration change was calculated using a modified Beer-Lambert law: in, This indicates changes in oxyhemoglobin concentration. The difference in extinction coefficients. This is the differential path length factor. The distance between the light source and the detector. Baseline light intensity, To measure light intensity.

6. The adaptive ordering method according to claim 1, characterized in that, The anomaly detection model is a semi-supervised anomaly detection model based on a single-class support vector machine. The pre-trained anomaly detection model is trained using a standard brain-body dynamic feature sample set of multiple users in a normal confirmation intention state. The normal confirmation intention state is defined as the feature pattern when a user clearly makes a meal selection in a conscious state without epileptic aura or anxiety interference.

7. The adaptive ordering method according to claim 1, characterized in that, The second preset threshold is determined by maximizing the F1 score on the validation set, and the first preset threshold is determined by the maximum point of the Youden index in the ROC curve analysis.

8. An adaptive ordering system, characterized in that, include: The data acquisition unit is used to collect the user's electronic medical record summary information on brain diseases and anti-epileptic drug prescription information in response to the user's order request; The feature extraction unit is used to extract brain nutritional constraint features based on the brain disease electronic medical record summary information, extract AEDs-food interaction features based on the antiepileptic drug prescription information, and concatenate the brain nutritional constraint features with the AEDs-food interaction features to form a multidimensional brain-food feature vector. The menu generation engine is used to obtain a set of ordering candidates, generate an anti-epileptic fit score for each candidate, traverse the set of ordering candidates to filter candidates with anti-epileptic fit scores higher than a first preset threshold to construct a compliant candidate subset, respond to the empty set detection result of the compliant candidate subset, look up a pre-generated constraint relaxation strategy mapping table based on the conflict type index, extract the corresponding constraint relaxation strategy, and perform local filtering and repair operations on each candidate in the set of ordering candidates according to the constraint relaxation strategy until a non-empty compliant candidate subset is generated; The amygdala-linked concealment confirmation module is used to collect dynamic brain signals from users through functional near-infrared spectroscopy and electroencephalography (EEG) in a contactless interactive interface, extracting the amygdala activation lateralization index and EEG data. The power ratio and event-related desynchronization features are used to input the extracted features into a pre-trained anomaly detection model to generate a user confirmation intent score. In response to the confirmation intent score being lower than a second preset threshold, a first authentication factor is generated. At the same time, a second authentication factor is generated based on the user's iris or facial features. The first authentication factor and the second authentication factor are then logically ANDed to generate a confirmation completion instruction.

9. The adaptive ordering system according to claim 8, characterized in that, The amygdala-linked concealment confirmation module includes: The brain signal acquisition submodule includes a functional near-infrared spectroscopy probe array and a dry electrode EEG cap. The probe array covers the scalp region corresponding to the prefrontal lobe and bilateral temporal lobes and amygdala. The dry electrode EEG cap includes at least Fz, C3, C4 electrodes and bilateral mastoid reference electrodes. The signal processing submodule is used to calculate the amygdala activation lateralization index, EEG theta / β power ratio, and event-related desynchronization features; The anomaly detection submodule is used to input the above features into a pre-trained anomaly detection model, generate a confirmation intent score, and generate a first authentication factor in response to the score being lower than a second preset threshold. The fusion authentication submodule is used to perform a logical AND operation between the first authentication factor and a second authentication factor based on facial features or iris to generate a confirmation completion instruction.

Images