Food adaptation method and system for dysphagia patients based on FOIS classification

Abstract

The application discloses a food adaptation method and system for dysphagia patients based on FOIS classification, relates to the field of intelligent diet management, collects dysphagia function evaluation data, clinical data and the risk perception level of patients for hidden aspiration, calculates the baseline aspiration risk probability based on a multi-factor logistic regression model, and evaluates the FOIS classification; physiological signals in the swallowing process are collected through a wearable sensor, and a characteristic vector is extracted. The application realizes dynamic quantification of aspiration risk of dysphagia patients and personalized food adaptation, combines multi-dimensional baseline data and real-time physiological signals, continuously corrects the aspiration risk probability through a Bayesian dynamic updating model, and improves the real-time performance and accuracy of risk judgment. The adaptive adjustment of the risk threshold is realized based on the Thompson sampling strategy of reinforcement learning, individual differences of different patients can be accurately matched, and the aspiration risk avoidance and the nutritional intake demand of patients can be effectively balanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent diet management technology, and in particular to a method and system for food adaptation for patients with dysphagia based on FOIS classification. Background Technology

[0002] Dysphagia is a common complication of stroke patients. More than half of stroke patients will experience dysphagia, which can easily lead to occult or overt aspiration. Occult aspiration is difficult to identify in the early stages because it has no obvious clinical symptoms. Repeated occurrences can lead to serious complications such as aspiration pneumonia and chronic bronchitis, which significantly increase the risk of death and seriously affect the prognosis and quality of life of stroke patients.

[0003] Currently, dietary management for patients with dysphagia largely relies on clinical swallowing imaging and other examinations to assess swallowing ability, and then provides basic dietary recommendations based on the FOIS classification. However, these assessments are mostly static baseline assessments, unable to capture real-time changes in the patient's physiological state during swallowing, making it difficult to dynamically adjust dietary adaptation strategies. Furthermore, the risk of aspiration in patients is influenced by multiple factors, including swallowing function, comorbidities, risk perception, and food texture. Existing methods have not achieved dynamic risk quantification that integrates multiple factors, and the detection and adaptation of food texture largely depend on the subjective judgment of medical staff or caregivers, lacking standardized and intelligent detection methods, making it difficult to accurately implement dietary adaptation requirements in non-medical environments.

[0004] Furthermore, existing dietary adaptation programs often have fixed risk thresholds that cannot be adaptively adjusted based on the patient's actual eating outcomes. This makes it difficult to match individual differences among patients and dynamic risk changes during eating, easily leading to over- or under-adaptation of the diet. This not only affects the patient's nutritional intake but also fails to effectively avoid the risk of aspiration. Simultaneously, the dietary management data for patients with dysphagia lacks systematic recording and feedback, hindering continuous model optimization and remote monitoring by healthcare professionals, thus limiting the accuracy and convenience of clinical applications. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for food adaptation in patients with dysphagia based on the FOIS classification. The technical solution adopted is as follows:

[0006] Food adaptation methods for patients with dysphagia based on FOIS classification include the following steps:

[0007] Step 1: Collect the patient's swallowing function assessment data, clinical data, and the patient's perceived risk level of silent aspiration. Calculate the baseline aspiration risk probability based on a multivariate logistic regression model. And assess the FOIS classification;

[0008] Step 2: Collect physiological signals during swallowing using wearable sensors and extract feature vectors. ;

[0009] Step 3: Use a Bayesian dynamic update model based on the currently observed feature vectors. Risk probability at the previous moment Calculate the real-time probability of aspiration risk at the current moment. ;

[0010] Step 4: Set three risk thresholds: low, medium, and high, based on the real-time probability of accidental aspiration. By comparing with risk thresholds, the range of IDDSI levels of foods that patients are allowed to eat is dynamically adjusted. The selection of risk thresholds adopts the Thompson sampling strategy in reinforcement learning, and the threshold distribution is updated according to the eating outcome.

[0011] Step 5: Detect food characteristics using multiple sensors, map them to IDDSI levels, and generate dietary recommendations based on the allowable range.

[0012] Optionally, in step 1, the baseline aspiration risk probability The calculation formula is:

[0013] ;

[0014] in The i-th baseline risk factor includes pharyngeal function score, number of comorbidities, number of strokes, patient's perceived risk level of silent aspiration, and food texture level; The regression coefficients corresponding to the i-th baseline risk factor are obtained through maximum likelihood estimation using historical data. Here, n represents the regression coefficient of the intercept term in the model, and n is the total number of baseline risk factors. It is a natural constant.

[0015] Optionally, in step 3, the real-time probability of accidental aspiration. The Bayesian dynamic update formula is:

[0016] ;

[0017] in It is a feature vector extracted from real-time physiological signals. It's an aspiration state. It is in a state of no accidental aspiration. and These are the likelihood functions of the feature in the false aspiration state and the non-false aspiration state, respectively, which are estimated from the training data.

[0018] Optionally, the likelihood function can be a multivariate Gaussian distribution:

[0019] ; ;

[0020] and These are the mean vectors and covariance matrices of the eigenvectors for the aspiration class and the non-aspiration class, respectively, which are estimated using historical data.

[0021] Optionally, in the adaptive food adaptation threshold adjustment, the threshold selection adopts the Thompson sampling strategy, and each threshold... It follows a Beta distribution, and the parameters α and β are updated according to the outcome after each feeding:

[0022] , ;

[0023] in Success is defined as no aspiration event occurs after feeding; otherwise, it is considered a failure. Before each feeding, thresholds are sampled from the current Beta distribution corresponding to each threshold for risk classification.

[0024] Optionally, the permissible food IDDSI level range is based on the real-time probability of aspiration risk. Determined by comparison with the sampling threshold:

[0025] like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit;

[0026] like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit minus one level;

[0027] like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit minus two levels;

[0028] like It prohibits eating and sounds an alarm;

[0029] The FOIS basic recommendation upper limit is predefined. It is a low-risk threshold. It is a medium-risk threshold. It is a high-risk threshold.

[0030] Optionally, in food property detection, a portable detector integrating a viscosity sensor, a pressure sensor, and a near-infrared spectroscopy module can be used to detect the flowability, hardness, viscosity, and particle size parameters of food in real time, and input them into a pre-trained deep neural network to output the corresponding IDSI level.

[0031] Optionally, after the dietary recommendations are generated, the system can output the current food compatibility results and adjustment suggestions to the user through voice broadcast, screen display, or indicator lights.

[0032] Optionally, real-time aspiration risk probability data, feeding records, and outcomes can be uploaded to the cloud or hospital information system for remote monitoring and intervention by medical staff, and for continuous optimization of the model.

[0033] The FOIS-based food adaptation system for patients with dysphagia is used to implement a food adaptation method for patients with dysphagia based on the FOIS classification. The system includes a baseline assessment module, a physiological signal monitoring module, a dynamic risk assessment module, an adaptive threshold decision module, a food detection module, an adaptation judgment and interaction module, and a data management and update module.

[0034] The baseline assessment module is used to collect multidimensional patient data and calculate baseline risk probability based on a multivariate logistic regression model. And assess the FOIS classification;

[0035] The physiological signal monitoring module includes wearable sensors for collecting physiological signals during swallowing and extracting feature vectors. ;

[0036] The dynamic risk assessment module deploys a Bayesian dynamically updated model based on feature vectors. Risk probability at the previous moment Calculate the real-time probability of accidental aspiration at the current moment. ;

[0037] Adaptive threshold decision module: Stores threshold distribution parameters, uses Thompson sampling to select the threshold in real time, and based on... The allowable range of food IDDSI levels is determined by comparison with a threshold.

[0038] The food detection module integrates multiple sensors to detect food characteristics and map them to IDDSI levels;

[0039] The adaptation judgment and interaction module generates dietary suggestions based on the allowed range and the food's IDDSI level, and outputs them through the human-computer interaction interface.

[0040] The data management and update module records the data and outcome of each feeding, updates the threshold distribution parameters, and uploads them to the cloud for model optimization.

[0041] In summary, the present invention has at least one of the following beneficial technical effects:

[0042] This invention provides a method and system for food adaptation for patients with dysphagia based on FOIS classification. It achieves dynamic quantification and personalized food adaptation of aspiration risk in these patients. By combining multi-dimensional baseline data and real-time physiological signals, and continuously correcting the aspiration risk probability through a Bayesian dynamic update model, it overcomes the limitations of traditional static assessments and improves the real-time performance and accuracy of risk judgment. A reinforcement learning-based Thompson sampling strategy enables adaptive adjustment of the risk threshold, updating the threshold distribution based on the patient's actual eating outcomes. This accurately matches individual differences among patients, avoiding adaptation bias caused by fixed thresholds, and effectively balancing aspiration risk avoidance with the patient's nutritional intake needs.

[0043] By integrating multiple sensors for food detection, the system achieves standardized and intelligent detection of food properties and precise mapping of IDDSI levels, replacing traditional subjective judgment methods and improving the operability of food matching in non-medical environments. The system can dynamically adjust the range of IDDSI levels for permissible foods based on real-time aspiration risk, generate personalized dietary recommendations, and provide output through various forms of human-computer interaction. It is easy to operate and convenient for patients and caregivers to follow.

[0044] Meanwhile, the data management and cloud upload functions enable the systematic recording of eating data and risk data, and remote monitoring by medical staff. This provides data support for continuous model optimization and clinical intervention, effectively reduces the incidence of aspiration in patients with dysphagia, improves the accuracy and intelligence of dietary management, and provides reliable technical support for the long-term dietary management of patients with dysphagia after stroke. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the food adaptation method for patients with dysphagia based on the FOIS classification of the present invention.

[0046] Figure 2 This is a schematic diagram of the first page of the personalized dietary adaptation report for patients with dysphagia, according to a specific embodiment of the present invention.

[0047] Figure 3 This is a schematic diagram on the second page of the personalized dietary adaptation report for patients with dysphagia, according to a specific embodiment of the present invention.

[0048] Figure 4 This is a schematic diagram on the third page of the personalized dietary adaptation report for patients with dysphagia, which is a specific embodiment of the present invention. Detailed Implementation

[0049] The following is in conjunction with the accompanying drawings. Figures 1-4 The present invention will be described in further detail below.

[0050] This invention discloses a method and system for food adaptation for patients with dysphagia based on the FOIS classification.

[0051] Example 1

[0052] Food adaptation methods for patients with dysphagia based on FOIS classification include the following steps:

[0053] Step 1: Collect the patient's swallowing function assessment data, clinical data, and the patient's perceived risk level of silent aspiration. Calculate the baseline aspiration risk probability based on a multivariate logistic regression model. And assess the FOIS classification;

[0054] Step 2: Collect physiological signals during swallowing using wearable sensors and extract feature vectors. ;

[0055] Step 3: Use a Bayesian dynamic update model based on the currently observed feature vectors. Risk probability at the previous moment Calculate the real-time probability of aspiration risk at the current moment. ;

[0056] Step 4: Set three risk thresholds: low, medium, and high, based on the real-time probability of accidental aspiration. By comparing with risk thresholds, the range of IDDSI levels of foods that patients are allowed to eat is dynamically adjusted. The selection of risk thresholds adopts the Thompson sampling strategy in reinforcement learning, and the threshold distribution is updated according to the eating outcome.

[0057] Step 5: Detect food characteristics using multiple sensors, map them to IDDSI levels, and generate dietary recommendations based on the allowable range.

[0058] By adopting the above technical solutions, this method focuses on the dynamic assessment of aspiration risk in patients with dysphagia. It combines the FOIS grading of swallowing ability assessment standards with the IDDSI food texture grading standards. Through multi-dimensional data fusion, dynamic risk modeling, and adaptive threshold decision-making, it achieves precise adaptation of food textures for patients with dysphagia. The core principle revolves around five progressively advancing stages: baseline risk quantification, real-time physiological signal perception, dynamic risk updating, adaptive threshold adjustment, and intelligent detection and adaptation of food textures. Each stage supports the others to form a closed-loop personalized dietary adaptation system.

[0059] Based on a multi-dimensional analysis of risk factors from biopsychosocial dimensions, this study integrates patient swallowing function assessment data, clinical data, and psychosocial factors. Utilizing the statistical analysis capabilities of a multivariate logistic regression model, the influence weights of each risk factor are quantitatively fitted to calculate the probability of aspiration risk reflecting the patient's baseline condition, thus achieving accurate assessment of the patient's static baseline aspiration risk. Simultaneously, based on the FOIS grading system and combined with objective examination results such as pharyngography, the patient's oral intake capacity and degree of swallowing disorder are graded, clarifying the patient's baseline food adaptation ability and providing an initial reference for subsequent dynamic dietary adaptation.

[0060] By utilizing the non-invasive sensing characteristics of wearable sensors, physiological signals during the patient's swallowing process are continuously collected, capturing changes in physiological characteristics corresponding to swallowing behavior. Through signal processing and feature engineering, feature vectors with discriminability and relevance are extracted from the original physiological signals, transforming the dynamic swallowing physiological process into quantifiable digital features. This enables precise perception of the patient's real-time swallowing status and provides real and continuous physiological data support for the dynamic updating of aspiration risk.

[0061] Using Bayesian probabilistic inference as its core, this method takes the aspiration risk probability from the previous moment as the prior probability and combines it with the physiological feature vector extracted during the current swallowing process. By calculating the likelihood probability of the feature vector under both aspiration and no-aspiration states, the prior probability is dynamically corrected and updated to obtain a real-time aspiration risk probability reflecting the patient's current swallowing state. This process transforms aspiration risk from a static baseline to a dynamic real-time scenario, accurately capturing risk changes at different moments during the patient's swallowing process and ensuring that the risk assessment results are synchronized with the patient's actual swallowing state.

[0062] Three risk thresholds (low, medium, and high) are set as the basis for determining the risk of food adaptation. The Thompson sampling strategy in reinforcement learning is used to dynamically select and update the thresholds. A Beta distribution is used to model the probability distribution of each risk threshold. The actual outcome after each patient's meal is used as a feedback signal. If no aspiration occurs after eating, the distribution parameters are updated to increase the probability of selecting that threshold; if aspiration occurs, the parameters are updated in reverse, continuously optimizing the threshold distribution to better match the individual patient characteristics. Based on the comparison between the real-time aspiration risk probability and the sampled thresholds, the IDDSI level range of foods allowed for patient consumption is dynamically adjusted, achieving a precise match between the risk thresholds and the individual patient's aspiration risk characteristics, making the food adaptation strategy more personalized.

[0063] This system integrates multiple types of sensors to perform multi-dimensional detection of the physical properties of food, accurately acquiring key parameters such as fluidity, hardness, viscosity, and particle size. Based on standardized mapping rules, the detected physical parameters are converted into IDDSI (Independent Dynamic Discharge Index) levels, achieving standardized and digital judgment of food properties. Using dynamically adjusted IDDSI level allowable ranges as a basis, the system matches and analyzes the detected food IDDSI levels against these ranges, providing adaptation suggestions for foods within the range and adjustment prompts for foods outside the range. This achieves precise matching of food properties with the patient's real-time aspiration risk and swallowing ability, mitigating aspiration risks from the food end.

[0064] Example 2

[0065] In step 1, the baseline aspiration risk probability The calculation formula is:

[0066] ;

[0067] in The i-th baseline risk factor includes pharyngeal function score, number of comorbidities, number of strokes, patient's perceived risk level of silent aspiration, and food texture level; The regression coefficients corresponding to the i-th baseline risk factor are obtained through maximum likelihood estimation using historical data. Here, n represents the regression coefficient of the intercept term in the model, and n is the total number of baseline risk factors. It is a natural constant.

[0068] Example 3

[0069] In step 3, the probability of real-time accidental aspiration. The Bayesian dynamic update formula is:

[0070] ;

[0071] in It is a feature vector extracted from real-time physiological signals. It's an aspiration state. It is in a state of no accidental aspiration. and These are the likelihood functions of the feature in the false aspiration state and the non-false aspiration state, respectively, which are estimated from the training data.

[0072] Example 4

[0073] The likelihood function is a multivariate Gaussian distribution:

[0074] ; ;

[0075] and These are the mean vectors and covariance matrices of the eigenvectors for the aspiration class and the non-aspiration class, respectively, which are estimated using historical data.

[0076] By adopting the above technical solution, with the quantitative calculation and dynamic updating of aspiration risk in stroke patients with dysphagia as the core, the baseline aspiration risk is determined through multi-factor statistical modeling, and the risk is dynamically corrected by relying on Bayesian probabilistic inference combined with real-time physiological signals. By fitting the probability distribution of physiological characteristics with multivariate Gaussian distribution, the aspiration risk assessment is extended from static baseline to real-time dynamic, so as to achieve accurate and continuous judgment of the patient's aspiration risk.

[0077] Based on five core baseline risk factors identified in clinical studies—pharyngeal function score, number of comorbidities, number of strokes, patient's perceived risk level of latent aspiration, and food texture level—a logistic regression model was used to fit historical clinical data. Maximum likelihood estimation was employed to determine the intercept term and regression coefficients corresponding to each risk factor, thus establishing the weight of each factor's influence on aspiration risk. Utilizing the sigmoid mapping property of logistic regression, the linear combination results of the multi-dimensional risk factors were mapped to a probability range of 0 to 1, calculating the baseline aspiration risk probability reflecting the patient's basic physiological, clinical, and psychosocial status. This standardized and quantified assessment of the patient's initial aspiration risk provides a priori probabilistic basis for subsequent real-time updates to the aspiration risk probability.

[0078] Using the aspiration risk probability from the previous moment as the prior probability, and combining it with feature vectors extracted from swallowing physiological signals collected by wearable sensors, the prior probability is dynamically corrected and updated by calculating the likelihood values ​​of the feature vectors under both aspiration and no-aspiration states. The real-time aspiration risk probability is calculated by using the product of the likelihood value of the feature vector under the aspiration state and the prior probability as the numerator, and the sum of the products of the likelihood values ​​under both aspiration and no-aspiration states and the prior probability as the denominator. This ratio is then used to calculate the real-time aspiration risk probability. This process incorporates the dynamic changes in real-time physiological signals into the risk calculation, allowing the risk assessment results to reflect the changes in the patient's physiological state during swallowing in real time, achieving continuous and dynamic updates to the aspiration risk.

[0079] For the multidimensional physiological feature vectors extracted during swallowing, a multivariate Gaussian distribution is used to fit their probability distributions under two states: aspiration and no aspiration. Using historical training data, the mean vector and covariance matrix of the feature vectors under aspiration and no aspiration states are estimated. These two multivariate Gaussian distributions characterize the distribution patterns and correlations between the physiological feature vectors under the two states. Substituting the real-time extracted feature vectors into the probability density functions of the two multivariate Gaussian distributions, the likelihood values ​​of the feature vectors under the corresponding states are calculated. This provides a precise probability calculation basis for Bayesian dynamic updates, making the correlation analysis between physiological features and aspiration states more closely reflect the actual distribution characteristics of the data and improving the accuracy of real-time aspiration risk probability calculation.

[0080] Example 5

[0081] In the adaptive food matching threshold adjustment, the threshold selection adopts the Thompson sampling strategy, and each threshold... It follows a Beta distribution, and the parameters α and β are updated according to the outcome after each feeding:

[0082] , ;

[0083] in Success is defined as no aspiration event occurs after feeding; otherwise, it is considered a failure. Before each feeding, thresholds are sampled from the current Beta distribution corresponding to each threshold for risk classification.

[0084] By employing the above technical solution, for the three food adaptation risk thresholds (low, medium, and high), the probability distribution characteristics of each threshold are modeled using a Beta distribution. This transforms the selection of the threshold into a random sampling process based on probability distribution. The Beta distribution can flexibly characterize the probability distribution characteristics within the 0-1 interval and can continuously adjust its shape based on actual feedback data to adapt to individual differences in aspiration risk among different patients. This provides a probabilistic basis for the dynamic selection of the threshold, ensuring that the determination of the threshold is no longer a fixed value but rather a probability distribution result based on the actual situation of the patient.

[0085] Before each meal, the risk threshold for that meal is obtained by random sampling based on the current Beta distribution parameters of each risk threshold. The Thompson sampling strategy takes into account both the exploration and utilization of probability distributions. It selects a threshold with a high probability of fit based on the distribution characteristics formed by existing meal outcomes, and also explores other thresholds within the distribution through random sampling. This avoids the rigidity of fit caused by fixed thresholds, allowing the threshold selection to continuously adapt to the dynamic changes in the patient's swallowing state while conforming to the patient's existing risk characteristics, thus improving the flexibility and adaptability of threshold selection.

[0086] The actual outcome of each patient's meal is used as a feedback signal to drive the dynamic update of the corresponding threshold Beta distribution parameters. If no aspiration occurs after eating, it is considered a success, and the Alpha parameter of the corresponding threshold is cumulatively updated; if aspiration occurs, it is considered a failure, and the Beta parameter of the corresponding threshold is updated based on the failure outcome. This process allows the shape of the Beta distribution to be continuously corrected according to the patient's actual aspiration outcomes. Successful outcomes strengthen the selection probability of the corresponding threshold, while failed outcomes reduce its selection probability. This continuously optimizes the threshold probability distribution towards a direction that better matches the individual patient's aspiration risk characteristics, making subsequent threshold sampling results more closely reflect the patient's actual swallowing ability and aspiration risk status.

[0087] Overall, this technical solution integrates probabilistic distribution modeling and reinforcement learning sampling strategies. It uses Beta distribution to probabilistically characterize risk thresholds, relies on Thompson sampling to dynamically select thresholds, and uses the actual aspiration outcomes of patients as feedback to continuously update distribution parameters, forming a closed-loop feedback adjustment between threshold selection and actual outcomes. This allows the risk threshold for food matching to dynamically change with individual patient characteristics and swallowing status, achieving personalized risk grading and food matching, effectively balancing aspiration risk avoidance and patient dietary needs.

[0088] Example 6

[0089] Permissible food IDSSI level range based on real-time aspiration risk probability Determined by comparison with the sampling threshold:

[0090] like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit;

[0091] like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit minus one level;

[0092] like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit minus two levels;

[0093] like It prohibits eating and sounds an alarm;

[0094] The FOIS basic recommendation upper limit is predefined. It is a low-risk threshold. It is a medium-risk threshold. It is a high-risk threshold.

[0095] By adopting the above technical solution, based on the assessment results of the patient's swallowing ability according to the FOIS grading system, the upper limit of the patient's basic food IDDSI level is determined in advance. This upper limit is a basic dietary adaptation standard formulated in combination with the patient's swallowing function and oral intake ability, providing a fixed initial reference benchmark for subsequent adjustment of the diet range according to the real-time aspiration risk probability, and ensuring that the initial direction of dietary adaptation matches the patient's basic swallowing ability.

[0096] By setting three dynamic sampling thresholds (low, medium, and high), the real-time aspiration risk probability of patients is divided into four gradient intervals. By comparing the real-time aspiration risk probability with each threshold level, the current aspiration risk level of the patient is accurately classified. Different risk intervals correspond to different risk levels during swallowing, with the risk level gradually increasing as the real-time probability rises. This achieves quantitative stratification of aspiration risk, providing a clear basis for adjusting the range of dietary risk levels.

[0097] A linkage adjustment rule was established between the real-time aspiration risk probability level and the allowable food IDDSI level range, with a negative correlation between the risk level and the upper limit of the allowable food level. When the patient's real-time aspiration risk probability is at a low level, the upper limit of the food level recommended by the FOIS is maintained to ensure the patient's dietary choices. As the risk level gradually increases, the upper limit of the allowable food level is lowered sequentially to reduce the possibility of aspiration during swallowing by reducing the complexity of food texture. When the risk reaches the high-risk threshold, food intake is directly prohibited to completely avoid the risk of serious aspiration from the dietary perspective, achieving a gradient adjustment of food adaptation from lenient to strict.

[0098] When the real-time aspiration risk probability reaches or exceeds the high-risk threshold, an emergency warning signal is used to trigger the prohibition of eating and issue an alarm. This promptly reminds patients and caregivers that the current swallowing risk is extremely high and that safe eating is not possible. Intervention measures must be taken to improve the swallowing status before the feasibility of eating can be reassessed. Through immediate risk warnings, serious aspiration consequences caused by eating in a high-risk state can be avoided, thus ensuring the patient's swallowing safety.

[0099] Example 7

[0100] In food property testing, a portable detector integrating a viscosity sensor, a pressure sensor, and a near-infrared spectroscopy module is used to detect the flowability, hardness, viscosity, and particle size parameters of food in real time. These parameters are then input into a pre-trained deep neural network, which outputs the corresponding IDSI level.

[0101] By adopting the above technical solution, the portable detector integrates a viscosity sensor, a pressure sensor, and a near-infrared spectroscopy module. Each sensor performs its specific function and works collaboratively to achieve comprehensive real-time detection of key physical properties of food. The viscosity sensor directly captures the stickiness and flowability characteristics of food, the pressure sensor senses the hardness and softness of food through contact detection, and the near-infrared spectroscopy module analyzes the particle size and internal texture distribution of food using spectral reflectance characteristics. Multiple sensors collect physical characteristic data of food from different dimensions, forming a complete set of food property parameters, providing comprehensive and objective raw data support for subsequent grading.

[0102] The raw data collected by each sensor undergoes signal processing and feature extraction to remove redundant and interfering information, retaining core characteristic parameters that characterize food properties. This process integrates scattered, single-dimensional detection data into structured, multi-dimensional feature vectors. This transformation from raw sensor signals to effective feature parameters allows the detection data to be recognized and analyzed by deep neural networks. Furthermore, feature integration allows characteristics such as food fluidity, hardness, viscosity, and particle size to form an organic whole, better aligning with the comprehensive assessment requirements of IDDSI (Integrated Device Index) for food properties.

[0103] The integrated multidimensional feature vectors of food are input into a pre-trained deep neural network. This network, trained on a large number of samples labeled with food physical parameters and corresponding IDDSI levels, has learned the intrinsic correlation between food physical characteristics and IDDSI levels. Through feature learning and nonlinear mapping across multiple layers of neurons, the network performs layer-by-layer analysis and classification calculations on the input feature vectors, automatically matching the corresponding IDDSI levels and outputting the results. The pre-trained deep neural network can capture the complex correspondence between food physical characteristics and IDDSI levels, overcoming the limitations of traditional manual judgment or simple threshold division, and improving the accuracy and consistency of mapping food traits to IDDSI levels.

[0104] Example 8

[0105] After dietary recommendations are generated, the system outputs the current food compatibility results and adjustment suggestions to the user through voice broadcast, screen display, or indicator lights.

[0106] Example 9

[0107] Real-time aspiration risk probability data, feeding records, and outcomes are uploaded to the cloud or hospital information system for remote monitoring and intervention by medical staff, and are used for continuous model optimization.

[0108] Example 10

[0109] The FOIS-based food adaptation system for patients with dysphagia is used to implement a food adaptation method for patients with dysphagia based on the FOIS classification. The system includes a baseline assessment module, a physiological signal monitoring module, a dynamic risk assessment module, an adaptive threshold decision module, a food detection module, an adaptation judgment and interaction module, and a data management and update module.

[0110] The baseline assessment module is used to collect multidimensional patient data and calculate baseline risk probability based on a multivariate logistic regression model. And assess the FOIS classification;

[0111] The physiological signal monitoring module includes wearable sensors for collecting physiological signals during swallowing and extracting feature vectors. ;

[0112] The dynamic risk assessment module deploys a Bayesian dynamically updated model based on feature vectors. Risk probability at the previous moment Calculate the real-time probability of accidental aspiration at the current moment. ;

[0113] Adaptive threshold decision module: Stores threshold distribution parameters, uses Thompson sampling to select the threshold in real time, and based on... The allowable range of food IDDSI levels is determined by comparison with a threshold.

[0114] The food detection module integrates multiple sensors to detect food characteristics and map them to IDDSI levels;

[0115] The adaptation judgment and interaction module generates dietary suggestions based on the allowed range and the food's IDDSI level, and outputs them through the human-computer interaction interface.

[0116] The data management and update module records the data and outcome of each feeding, updates the threshold distribution parameters, and uploads them to the cloud for model optimization.

[0117] The following specific embodiments illustrate the implementation principle of the present invention:

[0118] This study focuses on a 65-year-old patient with dysphagia following ischemic stroke. The patient had a history of two strokes and two underlying diseases: hypertension and diabetes. The patient was assessed as having moderate dysphagia by veno-femoral stenosis (VFSS) and a moderate perceived risk of latent aspiration. The clinical plan is to conduct a food adaptation intervention based on the FOIS classification. The adaptation method and system of this technical protocol will be applied throughout the process to achieve dynamic assessment of aspiration risk and personalized dietary adaptation.

[0119] System Deployment:

[0120] This embodiment employs a FOIS-based food adaptation system for patients with dysphagia, comprising a baseline assessment module, a physiological signal monitoring module, a dynamic risk assessment module, an adaptive threshold decision module, a food detection module, an adaptation judgment and interaction module, and a data management and update module. The physiological signal monitoring module uses a wrist-worn wearable sensor to collect physiological signals such as pharyngeal electromyography and swallowing acceleration during swallowing. The food detection module is a portable multi-sensor detector integrating a viscosity sensor, a pressure sensor, and a near-infrared spectroscopy module. The adaptation judgment and interaction module is equipped with voice broadcasting, an LCD screen display, and red, yellow, and green indicator lights. The data management and update module enables local data storage and synchronization with the hospital information system in the cloud.

[0121] Adaptation and implementation steps:

[0122] Step 1, Baseline Risk Calculation and FOIS Classification:

[0123] The system's baseline assessment module collected multidimensional data from the patient, including a pharyngeal function score of 65, two comorbidities, two strokes, and a patient's perceived risk level of 3 for latent aspiration. Combined with the baseline levels of common clinical food characteristics, the above data were input into a multivariate logistic regression model, and the patient's baseline aspiration risk probability P0 was calculated to be 0.45.

[0124] In conjunction with the patient's VFSS test results, and based on the FOIS grading system, the patient's FOIS grade was assessed as level 6, and the upper limit of the FOIS basic recommended food IDDSI level was determined to be level 6 (soft food).

[0125] Step 2, Swallowing physiological signal acquisition and feature vector extraction:

[0126] The system equips patients with wearable sensors on their wrists to continuously collect physiological signals such as pharyngeal electromyography and swallowing acceleration during their attempts to eat. The system then performs noise reduction and filtering on the raw physiological signals to extract feature vectors that characterize the swallowing state. This provides data support for real-time calculation of accidental aspiration risk.

[0127] Step 3, Real-time dynamic calculation of the probability of accidental aspiration:

[0128] The system dynamic risk assessment module calls the Bayesian dynamic update model to update the baseline aspiration risk probability obtained in step 1. As the initial moment Combined with the real-time feature vector extracted in step 2 The likelihood values ​​of the feature vectors under aspiration and no-aspiration states are calculated, and the prior probabilities are dynamically adjusted to obtain the real-time aspiration risk probability of the patient at the current moment. It is 0.52.

[0129] Step 4, Risk threshold sampling and IDDSI level range adjustment:

[0130] The system's adaptive threshold decision module samples the Beta distribution parameters of low, medium, and high risk thresholds in real time. This sampling yields the low-risk threshold. The threshold for medium risk is 0.40. 0.50, high-risk threshold It is 0.70.

[0131] Real-time probability of accidental aspiration =0.52 compared with the sampling threshold, which is consistent with... ≤ < Based on the criteria for judgment, the system dynamically adjusts the range of permissible food IDSI levels for this patient to the upper limit of the FOIS basic recommendation minus two levels, that is, the permissible IDSI level ≤ 4 (extremely viscous liquid).

[0132] Step 5, Food trait detection and IDDSI level mapping:

[0133] The patient and caregiver prepare three foods: steamed egg custard, thick porridge, and warm water. The portable food detection module is placed into each food in turn. The detector works in concert with multiple sensors to detect the fluidity, hardness, viscosity, and particle size parameters of the three foods. The detected multidimensional parameters are integrated into a feature vector and then input into a pre-trained deep neural network. After the model calculates, it outputs the corresponding IDDSI level for the three foods: steamed egg custard is level 6, thick porridge is level 5, and warm water is level 1.

[0134] Step 6, Dietary suitability assessment and result output:

[0135] The system's compatibility judgment and interaction module matches and analyzes the IDSI level of each food with the allowable range (≤4 levels) determined in step 4. The compatibility results are displayed on the LCD screen, and voice announcements are made simultaneously. In addition, red, yellow and green indicator lights provide prompts: Warm water has an IDSI level of 1, which meets the compatibility requirements, so the green light is on and the voice announcement says "Food is compatible and can be eaten normally"; Thick porridge has an IDSI level of 5 and steamed egg custard has an IDSI level of 6, both of which exceed the compatibility range, so the yellow light is on and the voice announcement says "Food is not compatible, it is recommended to adjust the texture of the food".

[0136] Step 7, Feeding Outcome Recording and Model Optimization:

[0137] The patient chose to consume warm water and did not experience aspiration after eating. The caregiver recorded the feeding outcome through the system. The system's data management and update module locally stored the real-time aspiration risk probability data, food detection data, and feeding outcome information, and simultaneously uploaded it to the hospital's information system cloud for remote viewing by medical staff. At the same time, based on the "successful feeding" outcome, the system updated the Beta distribution parameters of the risk threshold, providing a more accurate basis for subsequent threshold sampling that better reflects the patient's individual characteristics, thus enabling continuous model optimization.

[0138] Step 8, Verification of high-risk early warning scenarios:

[0139] When the patient attempted to eat soft bread while in poor condition, the system detected a real-time aspiration risk probability of Pt=0.75, exceeding the high-risk threshold τh=0.70. The system immediately triggered a high-risk warning, with a red light flashing continuously and a voice message repeatedly announcing "The current aspiration risk is extremely high, eating is prohibited." At the same time, the high-risk warning information was pushed to the hospital information system terminal of medical staff, facilitating timely intervention by medical staff.

[0140] The attached diagram in this embodiment is a personalized dietary adaptation report for patients with dysphagia based on the FOIS classification. It is mainly divided into five areas: patient basic information area, risk assessment area, food testing area, adaptation result area, and dietary recommendation area. The content and layout of each area are described below:

[0141] The patient basic information area includes basic data such as patient name, age, stroke history, comorbidities, FOIS classification, and baseline aspiration risk probability.

[0142] The risk assessment area includes the real-time probability of aspiration risk, the low, medium and high risk thresholds for this sample, and the allowable range of food IDDSI levels, displayed in a combination of numerical values ​​and textual annotations. It is also accompanied by a risk level bar chart to intuitively present the current risk range.

[0143] The food testing area is in tabular form, which includes the name of the food to be tested, the flowability, hardness, viscosity, particle size testing parameters of each food, and the mapped IDDSI level.

[0144] The matching results area corresponds to the food to be tested in the food testing area. The food is displayed in the form of food name combined with matching label, corresponding to "matched" and "unmatched" respectively. At the same time, it is indicated whether each food exceeds the allowable IDSSI level range.

[0145] The dietary advice section is in text format, including recommendations for foods that can be eaten normally, suggestions for adjusting the texture of foods that are not suitable, and precautions for eating. It also includes simple diagrams of food texture adjustment, such as "liquids can be thickened and solids can be ground into a paste" and other intuitive illustrations.

[0146] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for food adaptation in patients with dysphagia based on FOIS classification, characterized in that, Includes the following steps: Step 1: Collect the patient's swallowing function assessment data, clinical data, and the patient's perceived risk level of silent aspiration. Calculate the baseline aspiration risk probability based on a multivariate logistic regression model. And assess the FOIS classification; Step 2: Collect physiological signals during swallowing using wearable sensors and extract feature vectors. ; Step 3: Use a Bayesian dynamic update model based on the currently observed feature vectors. Risk probability at the previous moment Calculate the real-time probability of aspiration risk at the current moment. ; Step 4: Set three risk thresholds: low, medium, and high, based on the real-time probability of accidental aspiration. By comparing with risk thresholds, the range of IDDSI levels of foods that patients are allowed to eat is dynamically adjusted. The selection of risk thresholds adopts the Thompson sampling strategy in reinforcement learning, and the threshold distribution is updated according to the eating outcome. Step 5: Detect food characteristics using multiple sensors, map them to IDDSI levels, and generate dietary recommendations based on the allowable range.

2. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 1, characterized in that, In step 1, the baseline aspiration risk probability The calculation formula is: ； in The i-th baseline risk factor includes pharyngeal function score, number of comorbidities, number of strokes, patient's perceived risk level of silent aspiration, and food texture level; The regression coefficients corresponding to the i-th baseline risk factor are obtained through maximum likelihood estimation using historical data. Here, n represents the regression coefficient of the intercept term in the model, and n is the total number of baseline risk factors. It is a natural constant.

3. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 2, characterized in that, In step 3, the probability of real-time accidental aspiration. The Bayesian dynamic update formula is: ； in It is a feature vector extracted from real-time physiological signals. It's an aspiration state. It is in a state of no accidental aspiration. and These are the likelihood functions of the feature in the false aspiration state and the non-false aspiration state, respectively, which are estimated from the training data.

4. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 3, characterized in that, The likelihood function is a multivariate Gaussian distribution: ； ； and These are the mean vectors and covariance matrices of the eigenvectors for the aspiration class and the non-aspiration class, respectively, which are estimated using historical data.

5. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 4, characterized in that, In the adaptive food matching threshold adjustment, the threshold selection adopts the Thompson sampling strategy, and each threshold... It follows a Beta distribution, and the parameters α and β are updated according to the outcome after each feeding: ， ； in For a successful event to be considered, it is defined as no aspiration event occurring after feeding; otherwise, it is considered a failure. Before each feeding, thresholds are sampled from the current Beta distribution corresponding to each threshold for risk classification.

6. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 5, characterized in that, Permissible food IDSSI level range based on real-time aspiration risk probability Determined by comparison with the sampling threshold: like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit; like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit minus one level; like It allows IDSSI levels to be less than or equal to the FOIS basic recommendation upper limit minus two levels; like It prohibits eating and sounds an alarm; The FOIS basic recommendation upper limit is predefined. It is a low-risk threshold. It is a medium-risk threshold. It is a high-risk threshold.

7. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 6, characterized in that, In food property testing, a portable detector integrating a viscosity sensor, a pressure sensor, and a near-infrared spectroscopy module is used to detect the flowability, hardness, viscosity, and particle size parameters of food in real time. These parameters are then input into a pre-trained deep neural network, which outputs the corresponding IDSI level.

8. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 7, characterized in that, After dietary recommendations are generated, the system outputs the current food compatibility results and adjustment suggestions to the user through voice broadcast, screen display, or indicator lights.

9. The method for food adaptation for patients with dysphagia based on FOIS classification according to claim 8, characterized in that, Real-time aspiration risk probability data, feeding records, and outcomes are uploaded to the cloud or hospital information system for remote monitoring and intervention by medical staff, and are used for continuous model optimization.

10. A food adaptation system for patients with dysphagia based on FOIS classification, characterized in that, The system for implementing the FOIS-based food adaptation method for patients with dysphagia as described in claim 9 includes a baseline assessment module, a physiological signal monitoring module, a dynamic risk assessment module, an adaptive threshold decision module, a food detection module, an adaptation judgment and interaction module, and a data management and update module. The baseline assessment module is used to collect multidimensional patient data and calculate baseline risk probability based on a multivariate logistic regression model. And assess the FOIS classification; The physiological signal monitoring module includes wearable sensors for collecting physiological signals during swallowing and extracting feature vectors. ; The dynamic risk assessment module deploys a Bayesian dynamically updated model based on feature vectors. Risk probability at the previous moment Calculate the real-time probability of accidental aspiration at the current moment. ; Adaptive threshold decision module: Stores threshold distribution parameters, uses Thompson sampling to select the threshold in real time, and based on... The allowable range of food IDDSI levels is determined by comparison with a threshold. The food detection module integrates multiple sensors to detect food characteristics and map them to IDDSI levels; The adaptation judgment and interaction module generates dietary suggestions based on the allowed range and the food's IDDSI level, and outputs them through the human-computer interaction interface. The data management and update module records the data and outcome of each feeding, updates the threshold distribution parameters, and uploads them to the cloud for model optimization.

Images