Nutrition risk screening method and system

Abstract

The invention relates to the technical field of nutrition risk screening, and discloses a nutrition risk screening method and system, and the method comprises the steps: collecting user body data and historical data, and carrying out the preprocessing; comparing the historical data with the body data to obtain a plurality of first historical deviation values, and determining a current user deviation condition based on the plurality of first historical deviation values; determining a nutrition risk level of the user based on the deviation condition, and inputting the body data into a nutrition risk model for comparison based on the nutrition risk level; and determining severity based on the comparison result, and sending corresponding early warning information based on the severity. The accuracy of nutrition screening is improved.

Description

Technical Field

[0001] This invention relates to the field of nutritional risk screening technology, and more specifically, to a nutritional risk screening method and system. Background Technology

[0002] Epilepsy is a common chronic neurological disorder characterized by abnormally excessive electrical discharges in the brain, leading to recurrent seizures. Although a variety of antiepileptic drugs are available, approximately 30% of patients develop drug-resistant epilepsy, meaning their seizures cannot be effectively controlled even with two or more appropriate medications. For these patients, especially children, seeking non-pharmacological treatment options is crucial.

[0003] In the treatment of drug-resistant epilepsy, the ketogenic diet, as a non-pharmacological therapy with a clear neuromodulatory effect, has established an important therapeutic position. By inducing the body to produce ketone bodies and altering energy metabolism patterns, it offers hope for reduced seizures or even seizure-free periods to a significant number of patients. However, this therapy is inherently an extreme and forced physiological and metabolic intervention, and its treatment process is accompanied by a series of complex and unique nutritional and metabolic risks.

[0004] However, the core idea of ​​existing nutritional screening tools is to "identify and correct nutritional deficiencies and metabolic imbalances," aiming to restore or maintain standard physiological homeostasis in patients. The treatment logic of the ketogenic diet is exactly the opposite; it is a "proactive construction and precise maintenance of a specific metabolic imbalance state to achieve therapeutic goals," namely, persistent physiological ketosis. Therefore, the scoring items of general tools directly lose their evaluative meaning and may even lead to misleading conclusions.

[0005] Therefore, it is necessary to design a nutritional risk screening method and system to address the problems existing in current technologies. Summary of the Invention

[0006] In view of this, the present invention proposes a nutritional risk screening method and system, which aims to solve the problem of low efficiency in current aggregated payment.

[0007] This invention proposes a nutritional risk screening method, comprising: Collect and preprocess user body data and historical data; The historical data and the body data are compared to obtain several first historical deviation values, and the current user deviation is determined based on the several first historical deviation values. Based on the aforementioned deviations, the user's nutritional risk level is determined, and the body data is then input into the nutritional risk model for comparison based on the nutritional risk level. The severity is determined based on the comparison results, and corresponding warning information is sent based on the severity.

[0008] Furthermore, when collecting and preprocessing user body data and historical data, the following steps are included: The physical data include at least blood ketone levels, carbohydrate intake percentage, seizure frequency, blood potassium / magnesium levels, and ketogenic compliance score; Based on the historical data, filter and remove data from abnormal periods; The historical baseline and historical baseline standard deviation are determined based on the preprocessed historical data.

[0009] Furthermore, when comparing the historical data and the body data to obtain several first historical deviation values, the following are included: The relative deviation and absolute deviation are obtained by comparing the historical baseline value and the historical baseline standard deviation with the body data; The relative and absolute deviations are verified based on the normal range of ketogenic adaptation. When the verification is valid, the relevant data is retained. When the verification is invalid, it is marked as normal fluctuation and assigned a value of 0.

[0010] Furthermore, when comparing the historical data and the body data to obtain several first historical deviation values, the method further includes: Based on the historical data, the stability coefficient of ketogenic diet implementation and the correlation coefficient of clinical symptoms were determined; Based on the stability coefficient of the ketogenic diet and the correlation coefficient of the clinical symptoms, the relative deviation and the absolute deviation after verification are verified a second time to obtain several first historical deviation values.

[0011] Furthermore, when determining the current user deviation based on several of the first historical deviation values, the process includes: When the absolute deviation is greater than 30% or the absolute deviation rate exceeds the normal range of ±50%, it is a high-weight indicator and is judged as a serious deviation. When the relative deviation rate is greater than 40% or the absolute deviation rate exceeds the normal range of ±60%, it is a medium-weighted indicator and is judged as a moderate deviation. When the relative deviation rate is greater than 50% or there is a significant abnormality, it is a low-weight indicator and is judged as a slight deviation.

[0012] Furthermore, when determining a user's nutritional risk level based on the aforementioned deviations, the process includes: When only one low-weight indicator deviates slightly, and the high / medium-weight indicators do not deviate, it is judged as a low-risk level; When 2-3 medium / low weighted indicators deviate moderately, or 1 high weighted indicator deviates slightly, it is judged as a medium risk level; When ≥2 high-weight indicators deviate significantly, or ≥4 medium / low-weight indicators deviate moderately, the risk level is determined to be high.

[0013] Furthermore, when inputting the body data into the nutritional risk model for comparison based on the nutritional risk level, the process includes: The historical data is divided into a training set and a test set, and the encoder is pre-trained using the training set. Freeze the encoder parameters and train the main network; Unfreeze the encoder parameters and fine-tune them synchronously with the main network to obtain the nutrition risk model; When the nutritional risk level reaches the medium or high risk level, the body data is input into the nutritional risk model for comparison.

[0014] Furthermore, when determining severity based on the comparison results, this includes: A preliminary severity assessment is made based on the comparison results, and the preliminary severity level is determined based on the probability distribution fusion function; Based on historical data, a boundary region is pre-defined and it is determined whether the comparison result is within the boundary region. When the comparison result is within the boundary region, the comparison result is corrected. When the comparison results are corrected, the corrected comparison results are used as the final grade. If the comparison results are not corrected, the preliminary severity level will be used as the final level.

[0015] Furthermore, when sending corresponding warning information based on the severity level, it includes: The severity levels are categorized into Level 1, Level 2, and Level 3 warnings. When the final level is mild, a Level 1 warning is triggered; when the final level is moderate, a Level 2 warning is triggered; and when the final level is severe, a Level 3 warning is triggered. The confidence level is calculated for the final level, and the confidence level includes high confidence, medium confidence and low confidence. When the confidence level is high, the warning level is determined based on the final level; When the confidence level is medium confidence level, a level 2 warning is triggered; When the confidence level is low, a level 3 warning is triggered.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By constructing a closed-loop screening process based on individualized historical benchmarks and intelligent stratification, the individualized accuracy and early detection capability of risk assessment are improved in the field of clinical nutrition risk management, especially in managing complex therapeutic nutritional interventions such as the ketogenic diet. By comparing the current physical data with the patient's own historical data baseline, a "personalized" health and risk coordinate system is established. This self-comparison-based assessment mechanism can keenly capture early abnormal trends that have not yet exceeded the general clinical scope but have deviated significantly from the patient's personal baseline.

[0017] (2) It optimized the efficiency of clinical resource allocation and reduced unnecessary medical burden. By introducing a hierarchical risk assessment logic, namely, preliminary screening is conducted first through historical deviation calculation, and only those cases reaching the medium-to-high risk level are analyzed in depth using more complex nutritional risk models, thus achieving "intelligent triage" of screening work. This avoids the waste of resources caused by indiscriminately inputting all data into complex model calculations, ensuring that high-value computing resources and clinical attention can be concentrated on patients who are truly facing higher risks. At the same time, it reduces the possibility of unnecessary anxiety and over-examination among low-risk patients due to false alarms from general tools, making the entire screening and intervention process more focused and efficient.

[0018] (3) It enhances the clinical relevance and operability of risk warnings, and improves the timeliness and pertinence of intervention measures. This invention uses an intelligent model to conduct in-depth comparisons and severity assessments of medium- and high-risk cases, and its output results can be more closely integrated with specific clinical scenarios and pathophysiological mechanisms. The generated warning information not only indicates the risk level, but also relates to specific abnormal indicators, deviation directions, and durations, so that subsequent warning information and intervention recommendations are no longer general health education, but rather targeted clinical action guidelines.

[0019] (4) Its output of early warning and intervention results can be fed back to update the patient's historical data baseline. This allows the patient's "health baseline" to dynamically evolve with changes in the treatment stage and the body's adaptation, and the risk assessment standards are continuously individualized accordingly. This dynamism overcomes the drawback of fixed standards being unable to adapt to changes in the patient's disease course, enabling screening to be carried out throughout the entire treatment process, whether in the adaptation period, the stable period, or the period of complication, providing risk assessments that fit the current state. In the long term, this accumulated continuous data also provides valuable resources for understanding the long-term response patterns of individuals to specific nutritional interventions.

[0020] (5) By finely classifying the severity of risks and matching them with differentiated early warning and response mechanisms, the system ensures that appropriate levels of response are taken for situations of different levels of urgency. For potential high-risk complications, the system can trigger a rapid clinical response through high-level early warnings to maximize patient safety; for management deviations of medium and low risk, timely reminders can also help patients and their families fine-tune their behavior and consolidate treatment adherence. This tailored communication and intervention approach helps to establish a more proactive doctor-patient collaborative management relationship, transforming nutritional risk management from a passive crisis management to a proactive, collaborative health maintenance process.

[0021] On the other hand, this application also provides a nutrition risk screening system for applying the above-mentioned nutrition risk screening method, including: The data acquisition module is configured to collect and preprocess user body data and historical data. The feature extraction module is configured to compare the historical data and the body data to obtain several first historical deviation values, and determine the current user deviation based on the several first historical deviation values; The data processing module is configured to determine the user's nutritional risk level based on the deviation. The data processing module is also configured to input the body data into a nutritional risk model for comparison based on the nutritional risk level. The early warning module is configured to determine the severity based on the comparison results and send corresponding early warning information based on the severity.

[0022] It is understandable that the above-mentioned energy consumption assessment method and system for heat pump systems have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a nutritional risk screening method provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of a nutrition risk screening system provided in an embodiment of the present invention. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] In some embodiments of this application, see Figure 1 As shown, a nutritional risk screening method includes: Collect user body data and historical data and preprocess them.

[0026] Historical data and physical data are compared to obtain several first historical deviation values, and the current user's deviation status is determined based on these first historical deviation values.

[0027] The user's nutritional risk level is determined based on the deviation, and the body data is then input into the nutritional risk model for comparison based on the nutritional risk level.

[0028] The severity level is determined based on the comparison results, and corresponding warning information is sent based on the severity level.

[0029] Specifically, the process begins by collecting the patient's physical data and combining it with their historical data. The patient's current physical data is then compared with their historical data. Since both physical and historical data include various indicators of the patient's health, several first historical deviation values ​​are obtained, representing the degree of deviation between the current and historical data. These first historical deviation values ​​are used to determine the current user's deviation status, and based on this deviation, the user's nutritional risk level is determined. When the user's nutritional risk level is high, the user's current physical data is input into the nutritional risk model for comparison. The comparison results determine the final severity, and based on the severity, corresponding warning information is sent.

[0030] Understandably, by constructing a closed-loop screening process based on individualized historical benchmarks and intelligent stratified judgment, the individualized accuracy and early detection capabilities of risk assessment are improved in the field of clinical nutrition risk management, especially in managing complex therapeutic nutritional interventions such as the ketogenic diet. By comparing current physical data with the patient's own historical baseline data, a "personalized" health and risk coordinate system is effectively established. This self-comparison-based assessment mechanism can keenly capture early abnormal trends that have not yet exceeded the general clinical range but have deviated significantly from the patient's personal baseline. For example, although a ketogenic diet patient's blood ketone level is still within the generally considered "safe" range, if it has declined significantly compared to their personal stable historical level, this method can provide an early warning of the risk of decreased diet adherence or changes in metabolic efficiency, thereby achieving early risk identification and creating a valuable time window for early intervention.

[0031] Furthermore, it optimized the efficiency of clinical resource allocation and reduced unnecessary medical burden. By introducing a hierarchical risk assessment logic—that is, initial screening using historical deviation calculations, followed by in-depth analysis using more complex nutritional risk models only for cases reaching medium- to high-risk levels—it achieved "intelligent triage" of screening work. This avoided the waste of resources caused by indiscriminately inputting all data into complex model calculations, ensuring that high-value computational resources and clinical attention were concentrated on patients truly facing higher risks. At the same time, it reduced the possibility of unnecessary anxiety and over-testing among low-risk patients due to false positives from general tools, making the entire screening and intervention process more focused and efficient.

[0032] Furthermore, it enhances the clinical relevance and operability of risk warnings, improving the timeliness and targetedness of interventions. Traditional risk scoring stops at an abstract number or level, while this invention uses intelligent models to conduct in-depth comparisons and severity assessments of medium- and high-risk cases. Its output can be more closely integrated with specific clinical scenarios and pathophysiological mechanisms. The generated warning information not only indicates the risk level but also links it to specific abnormal indicators, deviation directions, and durations. This makes subsequent warning information and intervention recommendations no longer general health education but rather targeted clinical action guidelines. For example, a warning can directly indicate "insufficient blood ketones accompanied by a hypokalemic trend," thereby guiding physicians to accurately adjust dietary formulas and supplement electrolytes, improving the support for clinical decision-making and the efficiency of intervention implementation.

[0033] Furthermore, its output of early warnings and intervention results can be fed back to update the patient's historical baseline data. This allows the patient's "health baseline" to dynamically evolve with changes in treatment stages and the body's adaptation, and the standards for risk assessment are continuously individualized. This dynamism overcomes the drawbacks of fixed standards that cannot adapt to changes in the patient's disease course, enabling screening to be conducted throughout the entire treatment process, providing risk assessments that are relevant to the current state, whether in the adaptation phase, the stable phase, or the period of complication. In the long term, this accumulated continuous data also provides a valuable resource for understanding the long-term response patterns of individuals to specific nutritional interventions.

[0034] Finally, this method offers profound benefits in improving patient safety and treatment adherence. By finely categorizing risk severity and matching it with differentiated early warning and response mechanisms, it ensures appropriate levels of response to situations of varying urgency. For potential high-risk complications, the system can trigger a rapid clinical response through high-level alerts, maximizing patient safety. For management deviations of medium to low risk, timely reminders can help patients and their families fine-tune their behavior, reinforcing treatment adherence. This tailored communication and intervention approach helps establish a more proactive doctor-patient collaborative management relationship, transforming nutritional risk management from a passive crisis management process into a proactive, collaborative health maintenance process.

[0035] In some embodiments of this application, the collection and preprocessing of user body data and historical data includes: Physical data should include at least blood ketone levels, carbohydrate intake percentage, seizure frequency, blood potassium / magnesium levels, and ketogenic adherence score.

[0036] Historical data is used to filter and remove data from abnormal periods.

[0037] The historical baseline and historical baseline standard deviation are determined based on the preprocessed historical data.

[0038] Specifically, the patient's physical data includes at least the current blood ketone level, fasting blood glucose, daily fat / carbohydrate / protein ratio, and ketosis adherence score—these are the core indicators related to ketosis. It also includes the patient's current weight, body fat percentage, and related physiological and nutritional indicators, as well as current blood lipids (triglycerides, LDL-C), sodium / potassium / magnesium / calcium levels, and liver and kidney function. These are combined with the patient's historical medical data to remove outliers from abnormal periods. The historical baseline is a measure of the central trend of each indicator level when the patient is in a stable state, calculated using a weighted moving average method to reflect recent trends. First, define the time decay weight: ,in, Let i be the weight of the i-th historical data point. This is the attenuation coefficient, which controls the attenuation rate of the weight (set to 1.0). At the current time point, Let i be the timestamp of the i-th historical data point. The time scale parameter is set to 30 days to calculate the historical baseline value: ,in, Here, n represents the weighted historical baseline value of the indicator, and n is the number of data points for this indicator in the valid historical dataset. Let n be the i-th valid historical data value of the indicator. If the number of data points is too small (n < 5), a robust estimation method is adopted, using the median as the historical benchmark value. If the patient is in the early stages of starting a ketogenic diet (<3 months) and historical data is insufficient, a staged baseline is used: statistics from the same patient group in the adaptation phase are used as a reference, combined with the patient's individual data for Bayesian adjustment, while the historical baseline standard deviation is: ,in, The weighted historical benchmark standard deviation of the indicator is given, while the denominator contains... This is the Bessel correction, which ensures unbiased estimation in the case of small samples.

[0039] Understandably, this invention improves the accuracy and clinical relevance of individualized assessments. Traditional general reference ranges often fall short when applied to therapies with highly varying individual responses, such as the ketogenic diet. This invention, by establishing a unique historical baseline for each patient, essentially constructs a "personal health coordinate system." The application of the weighted moving average method, particularly the introduction of time decay weights, ensures that this baseline is not a static snapshot of the past state, but a dynamic trajectory that sensitively reflects the patient's recent stable trend. This allows clinical judgment to transcend the level of "comparing with the average person" and enter a deeper level of "comparing with one's own past stable state." For example, determining whether blood potassium levels are abnormal is no longer based on whether they are below the laboratory-printed general lower limit, but rather on assessing whether they significantly deviate from the patient's typical range during a stable period. This individualized calibration significantly reduces misjudgments caused by differences in individual physiological baselines, making risk warnings truly aligned with the patient's unique physiological background and treatment history. Secondly, it enhances the prospective and early identification capabilities of risk monitoring. By calculating the standard deviation of the historical baseline, this method not only defines the "center point" of an individual's normal state but also quantifies its inherent "fluctuation bandwidth." This provides a scientific benchmark for identifying clinically significant abnormal changes. When a monitored value deviates from its historical mean but remains within the generally normal range, if the deviation significantly exceeds the daily fluctuations represented by the individual's historical standard deviation, the system can issue an early warning, indicating that this change may be unusual. This early warning mechanism based on individual fluctuation patterns can capture subtle but trending signs of deterioration that cannot be identified by general standards, thus significantly advancing the intervention time and shifting the focus from treating existing complications to preventing their occurrence, thereby moving the risk management process forward. Furthermore, the multiple safeguards built into this invention, such as excluding data from clearly abnormal periods, using the median as a robust estimate for small sample data, and setting a minimum standard deviation threshold, collectively construct a defense mechanism. This mechanism can automatically filter noise interference caused by acute clinical events, measurement errors, or data sparsity, ensuring that the calculated historical benchmark and standard deviation can truly and stably reflect the "background signal" of the patient's physiological state. This reduces oversensitivity to occasional outliers or data collection flaws, avoiding unnecessary false alarms triggered by data quality issues, thereby improving the signal-to-noise ratio of the entire screening and early warning process and allowing clinical resources to focus more on real risks. Furthermore, this invention fully considers the complexities of real-world clinical scenarios. For patients in the early stages of treatment, the protocol cleverly utilizes empirical data from similar groups through a "phased baseline" strategy, providing a reasonable initial reference when individual data is insufficient, thus solving the "cold start" problem. Simultaneously, the entire computational framework has good scalability, capable of handling the unified processing of various types of indicators, from blood ketones and electrolytes to blood lipids and liver function.This flexibility and inclusivity ensures that the approach is not only applicable to specific indicators of the ketogenic diet but also allows for the integration of other nutrition and metabolism-related parameters, enabling it to adapt to evolving clinical monitoring needs. Historical benchmarks are not static; they are periodically updated and evolved as new, stable data from patients accumulate. This means that a patient's "normal range" can be adaptively adjusted to their treatment stage, age, or long-term changes in physical condition. This dynamism allows assessment criteria to be synchronized with the patient's long-term health trajectory, avoiding assessment biases caused by using outdated benchmarks. In the long term, this continuous, traceable benchmark change curve itself becomes a valuable clinical resource, helping physicians understand patients' adaptation patterns and response patterns to long-term nutritional interventions.

[0040] In some embodiments of this application, when comparing historical data and physical data to obtain several first historical deviation values, the following are included: The relative and absolute deviations are obtained by comparing the historical baseline values ​​and historical baseline standard deviations with the body data.

[0041] Based on the verification of relative and absolute deviations within the normal range of ketosis adaptation, when the verification is valid, the relevant data is retained; when the verification is invalid, it is marked as normal fluctuation and assigned a value of 0.

[0042] Specifically, the first step is to determine whether the data indicator is a continuous proportion indicator, a count type, or an indicator with more intuitive absolute value changes based on the patient's specific physical data.

[0043] Type A (applicable to relative deviation): Continuous proportional indicators, whose changes relative to the baseline value are more clinically significant. Examples include: blood ketone levels, carbohydrate intake percentage, and blood lipid levels.

[0044] Type B (applicable to absolute deviation): Indicators that are more intuitive than count-based or absolute value changes. For example: frequency of attacks (times / month), certain symptom scores, etc.

[0045] If the indicator is of type A, then calculate the relative deviation: ,like ,but If the indicator is of type B, then calculate the absolute deviation: ,in, This represents the raw deviation of the indicator, indicating the degree to which the current value deviates from the individual's historical baseline. The current measured value of the indicator. For the individual historical baseline value of the indicator, to ensure that the deviation is clinically significant, the current value needs to be validated within the specific target context of ketogenic therapy: if indicator i has a defined normal range for ketogenic therapy. Then check the current value. Does it fall within this range: If If the current value is within the acceptable target range for ketogenic therapy, then any deviation is considered normal fluctuation within the acceptable range for treatment. Therefore, the first historical deviation value of this indicator is set to 0. or If the deviation is found to be within the ideal ketogenic range, then the current value is considered to have exceeded the ideal ketogenic range, and the deviation is clinically significant, thus validating the study.

[0046] Understandably, differentiating indicator types and introducing dual validation improves the accuracy and clinical interpretability of individualized risk assessment. Traditional methods often employ a "one-size-fits-all" approach to deviation calculation, failing to reflect the differentiated clinical significance inherent in changes in different physiological indicators. This invention creatively distinguishes between "continuous proportional indicators" and "count-type absolute indicators," calculating them using relative and absolute deviations respectively. For example, for blood ketone levels, relative change rates more sensitively capture subtle changes in metabolic efficiency. For seizure frequency, absolute differences more directly reflect the actual fluctuations in disease control levels. This differentiated processing based on the inherent attributes of the indicators allows the calculated deviations to more realistically and accurately quantify the clinical meaning of changes in patient condition, enhancing the biological rationality and interpretability of the assessment results.

[0047] Secondly, the verification step of "ketogenic diet within the normal range" enhances the specificity of risk screening and reduces false positive alarms caused by the treatment's inherent characteristics. The treatment logic of the ketogenic diet lies in actively maintaining a specific metabolic state (such as physiological ketosis). This makes many values ​​considered "abnormal" or "deviations" from a conventional medical perspective (such as a certain blood ketone level) precisely the target of effective treatment. If only comparing personal history, a blood ketone value that successfully achieves ideal ketosis may be considered significantly "high" and trigger an alarm. Through the introduced secondary verification mechanism, each deviation is re-examined within the professional context of ketogenic treatment: as long as the current value remains within the ideal target range pursued by treatment, regardless of its variation relative to the personal historical baseline, it is considered "normal fluctuations allowed by treatment" and ignored (assigned a value of 0). This fundamentally eliminates the huge flaw of misinterpreting treatment success signals as risk signals, ensuring that the focus of risk assessment is always firmly locked on truly harmful abnormal changes that deviate from the treatment target, thereby improving the clinical relevance of alarms. By effectively filtering out deviations that fall under the category of "therapeutic normal fluctuations," unnecessary interference and anxiety for patients and healthcare professionals are avoided. This allows clinical focus to be freed from the massive amount of data fluctuations and concentrated on "true positive" risk events that deviate from individual baselines and exceed the treatment safety range. This not only reduces the medical burden and psychological costs caused by over-assessment and false alarms but also makes subsequent warnings and intervention recommendations more targeted. All interventions will directly target the specific goal of bringing indicators "back to the ketogenic treatment safety window," rather than the general "restoration to normal." This ability to accurately identify "signals" from "noise" makes the entire nutritional risk management process more efficient and targeted.

[0048] In some embodiments of this application, when comparing historical data and physical data to obtain several first historical deviation values, the method further includes: The stability coefficient of ketogenic diet implementation and the correlation coefficient of clinical symptoms were determined based on historical data.

[0049] The relative and absolute deviations after validation were re-validated based on the stability coefficient of ketogenic diet implementation and the correlation coefficient of clinical symptoms, resulting in several first historical deviation values.

[0050] Specifically, the stability coefficient of ketogenic diet adherence reflects the consistency and stability of a patient's recent adherence to the ketogenic diet. A higher stability coefficient indicates more reliable daily adherence, and deviations occurring at this level are likely to be more clinically significant. The stability characteristic value is calculated based on historical data from the past 30 days: Dietary stability score: Calculate the coefficient of variation (CV) of daily fat energy ratio (FER). ,in and The mean and standard deviation of daily FER over the past 30 days were used to determine the dietary stability score. ,in, This is a score representing the stability of the dietary structure, ranging from 0 to 1, with values ​​closer to 1 indicating greater stability. To adjust the parameters, a value of 1.0-2.0 was chosen based on the difficulty of controlling the diet.

[0051] Blood ketone stability score: Calculates the coefficient of variation of blood ketone levels (BHB) over the past 30 days. , ,in, This represents the blood ketone stability score. To adjust the parameter, set it to 1.0.

[0052] Finally, there is the adherence stability score: the average AS score of the ketosis adherence over the past 30 days. Assuming a scoring range of 0-10, the scores are then normalized to a stability score: ,in, As the final step, the three scores are weighted and combined to obtain the final ketogenic diet adherence stability score, which represents the adherence stability score. , ,in, The weights for dietary structure, blood ketones, and adherence stability are respectively, satisfying... Furthermore, the weights for dietary structure, blood ketones, and adherence stability were 0.4, 0.3, and 0.3, respectively.

[0053] Calculate the clinical symptom correlation coefficient: This coefficient quantifies the temporal correlation between deviations from a specific nutritional metabolic indicator and clinical symptoms (such as seizures or gastrointestinal reactions). The stronger the correlation, the more attention should be paid to deviations from that indicator. For each indicator to be evaluated, extract the daily or weekly deviation sequence for that indicator over the past 90 days. (The calculation method is the same as before, but historical benchmarks are used for each time point), and at the same time, the clinical symptom severity score sequence for the same period is extracted. This score integrates factors such as the frequency of epileptic seizures and the severity of gastrointestinal symptoms (nausea, vomiting, constipation), normalizing them to a continuous value of 0-1 (1 indicating the most severe symptoms). Then, the cross-correlation function (CCF) between the deviation sequence of the indicators and the clinical symptom score sequence is calculated to find the strongest correlation. ,in, The time lag (in days) is usually considered. (That is, changes in indicators may bring symptoms forward or delay them by up to 7 days). and The correlation coefficient, which is the largest absolute value among the two sequences, is taken as the original association strength of the indicator. .

[0054] Due to the correlation coefficient Therefore, a nonlinear transformation is used to enhance the discriminative power, resulting in a correlation coefficient specific to the index. : ,in To appropriately increase the moderate correlation value, a shape parameter (usually taken as 0.5-0.7) is used. For simplicity, my mother calculates an overall correlation coefficient. This reflects the average correlation between all currently significantly deviated indicators (i.e., indicators where Di>0 after initial validation) and clinical symptoms: ,in, This set of indicators with a deviation value Di > 0 after initial validation uses a weighted average method to ensure that the larger the deviation of an indicator, the greater its correlation coefficient with the overall value. The greater the contribution, the higher the final ranking will be based on... and After secondary verification and correction, the final first historical deviation value is obtained. First, the adjustment factor for each indicator is calculated: ,in, and The historical average value within the patient population is used as a baseline. and To adjust the intensity parameters and control the impact of stability and correlation on the correction amplitude, a value of 0.3 is used. Then let =0.5, if Then let =2.0, to prevent over-adjustment. The final first historical deviation value is determined as follows: ,in, The indicators were adjusted after initial validation. This is the final first historical deviation value that has undergone secondary verification and is used for subsequent risk level determination.

[0055] Understandably, fluctuations in indicators during ketogenic diet management can stem from two fundamentally different causes: one is accidental deviations or laxity in patient adherence (such as slightly exceeding carbohydrate intake on a particular day), and the other is genuine metabolic disturbances or intolerance to treatment. Traditional methods cannot distinguish between these, potentially leading to overreaction to the former or underreaction to the latter. This invention quantifies a patient's recent self-management consistency by calculating a "ketogenic diet adherence stability coefficient (Ks)." Deviations occurring when a patient has consistently maintained stability (high Ks value) are more likely to reflect underlying physiological risks. The system amplifies the weight of these deviations through the coefficient, essentially highlighting them to clinicians: "This patient is typically meticulous; this abnormality requires close attention." Conversely, for patients whose adherence is inherently unstable, some fluctuations can be identified and appropriately downweighted, thereby reducing disruptive alarms caused by behavioral inertia and guiding clinical attention to changes more likely reflecting real health threats.

[0056] Secondly, by quantifying the "clinical symptom correlation coefficient (Kc)," a strong correlation is established between laboratory indicators and patients' actual disease experiences and outcomes, enabling risk assessment to possess both prospective and causal inference. The ultimate clinical significance of changes in nutritional and metabolic indicators lies in whether they trigger or accompany adverse health events. This protocol proactively uncovers the hidden temporal connections between historical deviation patterns of specific indicators and clinical phenotypes such as epileptic seizures and gastrointestinal symptoms through complex time-lag cross-correlation analysis. For example, it might find that a patient's decreasing blood magnesium level consistently precedes an increase in seizure frequency by 2-3 days. When this correlation is quantitatively captured and incorporated into the assessment of current deviation (with weights adjusted via Kc), the system not only reports "low blood magnesium," but also warns of "the reappearance of a warning signal that has historically led to increased seizures." This transforms risk assessment from a static "current situation description" to a dynamic "risk prediction," improving the clinical relevance of the warning and the window of opportunity for intervention. It transforms complex pattern recognition methods (such as "This patient is usually obedient; this change in indicators likely indicates a real problem" or "His blood ketone levels seemed to have spiked after this fluctuation last time") into calculable and reproducible algorithmic logic. This not only reduces reliance on doctors' exceptional personal experience but also provides multidisciplinary teams with consistent and objective supporting judgments. Furthermore, by differentiating between fluctuations caused by inconsistent implementation and dangerous deviations strongly correlated with symptoms, the approach can more precisely guide intervention: alerts to high Ks and high Kc require urgent medical evaluation and protocol adjustments. For fluctuations with low Ks, the primary intervention may be to strengthen patient education and adherence support. This differentiated guidance allows limited medical resources to be allocated more effectively to the most urgent and effective intervention pathways.

[0057] In some embodiments of this application, determining the current user deviation based on several first historical deviation values ​​includes: When the absolute deviation is greater than 30% or the absolute deviation rate exceeds the normal range of ±50%, it is a high-weight indicator and is judged as a serious deviation.

[0058] When the relative deviation rate is greater than 40% or the absolute deviation rate exceeds the normal range of ±60%, it is a medium-weighted indicator and is judged as a moderate deviation.

[0059] When the relative deviation rate is greater than 50% or there is a significant abnormality, it is a low-weight indicator and is judged as a slight deviation.

[0060] In some embodiments of this application, determining a user's nutritional risk level based on deviation includes: When only one low-weight indicator deviates slightly, and the high / medium-weight indicators do not deviate, it is judged as a low-risk level.

[0061] When 2-3 medium / low weighted indicators deviate moderately, or 1 high weighted indicator deviates slightly, it is judged as a medium risk level.

[0062] When ≥2 high-weight indicators deviate significantly, or ≥4 medium / low-weight indicators deviate moderately, the risk level is determined to be high.

[0063] Understandably, firstly, by establishing a two-dimensional matrix of "indicator weight - degree of deviation," the systematization and clinical relevance of risk assessment are enhanced. Traditional methods often treat all indicators equally or simply apply weights, failing to accurately reflect the distinct importance of different physiological parameters in ketogenic therapy. This invention clearly distinguishes between high, medium, and low-weight indicators. For example, parameters directly related to the core treatment mechanism and acute safety, such as blood ketones and blood potassium, are assigned high weights, while certain subjective indicators are assigned low weights. Simultaneously, differentiated deviation thresholds are set for each type of weighted indicator (e.g., a 30% deviation for high-weight indicators is considered severe, while a 50% deviation for low-weight indicators). This ensures that the assessment system is isomorphic with clinical thinking: important indicators, even slight abnormalities, attract high attention. Secondary indicators, however, allow for greater physiological fluctuations. This reduces the probability of unnecessary alarms triggered by slight fluctuations in non-core indicators, while ensuring that abnormalities in core indicators are keenly detected, making the assessment focus highly aligned with clinical priorities.

[0064] Secondly, the adoption of a composite judgment rule based on multiple conditions enhances the robustness of risk level determination and reduces the risk of misjudgment caused by fluctuations in a single indicator or measurement errors. Logical conditions such as "≥2 high-weight indicators showing significant deviation" or "≥4 medium / low-weight indicators showing moderate deviation" are set for a high-risk assessment. This principle of "consensus" or "cumulative" requires that risk signals have a certain breadth and consistency. This brings dual benefits: on the one hand, it reduces the possibility of individual indicators triggering the highest-level alarm due to sporadic occurrences, technical errors (such as inaccurate single tests), or transient physiological fluctuations, improving the system's anti-interference capability. On the other hand, it enhances the specificity of identifying truly high-risk states. Because when multiple indicators, especially high-weight indicators, show significant deviations simultaneously, it often indicates a systemic metabolic imbalance or treatment failure in the patient, rather than a localized, accidental problem. This judgment logic mimics the comprehensive judgment process of senior clinical experts, making the automated risk assessment results more prudent and reliable.

[0065] In some embodiments of this application, when inputting body data into a nutritional risk model for comparison based on nutritional risk levels, the following steps are included: Historical data is divided into training and testing sets, and the encoder is pre-trained using the training set.

[0066] Freeze the encoder parameters and train the main network.

[0067] The encoder parameters are thawed and fine-tuned synchronously with the main network to obtain the nutrition risk model.

[0068] When the nutritional risk level reaches the medium or high risk level, the body data is input into the nutritional risk model for comparison.

[0069] Specifically, longitudinal electronic medical record data of patients with drug-resistant epilepsy receiving ketogenic diet therapy were collected. For each patient's continuous treatment timeline, weekly or monthly segments were used to extract the "current observation point" and the corresponding "historical baseline period." The current observation point included all indicators at that time point (such as average blood ketones this week, total number of seizures this week, etc.). The historical baseline period consisted of 8-12 consecutive weeks of data prior to the observation point during which the patient was in a stable treatment state (no regimen adjustments, no acute complications). The mean and standard deviation of each indicator were calculated as the individual baseline. Based on a retrospective evaluation of each historical data point by experts, and categorized according to the occurrence and severity of nutritional and metabolic adverse events (AEs) within 2-4 weeks after the observation point (for pre-training), the following categorization is used: 0 indicates no adverse events; 1 indicates mild adverse events (e.g., transient constipation, mild somnolence, no medical intervention required); 2 indicates moderate adverse events (e.g., persistent hypokalemia requiring oral supplementation, kidney stones diagnosed but not requiring surgery); 3 indicates severe adverse events (e.g., ketoacidosis requiring emergency treatment, severe hyperlipidemia causing pancreatic...). The severity of the inflammation (for main model training) is labeled as follows: 1 (mild): the deviation indicators are mainly of medium to low weight, the deviation lasts for ≤7 days, and it does not affect the control of the attack; 2 (moderate): there is one high-weight indicator that deviates severely, or the deviation lasts for >7 days, and it has slightly affected the control of the attack; 3 (severe): ≥2 high-weight indicators deviate severely, there are clear complications, and the control of the attack has deteriorated significantly. Finally, the model adopts a two-stage architecture of pre-training-fine-tuning, and incorporates an attention mechanism to enhance interpretability. In the first stage, the feature encoder is pre-trained to learn to extract meaningful low-dimensional feature representations from the original, high-dimensional, multimodal medical data to alleviate the small sample size problem. It is trained using a stacked denoising autoencoder, with a standardized "current-historical" indicator difference vector as input. By minimizing the reconstruction error of the input vector, it forces the encoder to learn the intrinsic distribution and key patterns of the data. In the second stage, the main network is trained using a multi-dimensional fusion neural network. Its core architecture is a multi-input, fully connected network with an attention mechanism. The input layer (three parallel channels): Channel A (deviation features): data output by the pre-trained encoder. Channel B (Context Features): Structured context vector. Channel C (Time-Series Features): Statistical feature vector of recent trends in key indicators (blood ketones, attack frequency). Each channel first undergoes subspace feature extraction through a 2-3 layer fully connected network, with the outputs being... Attention fusion is performed, and the fusion weights of the three sub-features are calculated. , ,in, , For learnable parameters, the output layer is as follows: Task 1 (Main Task - Regression): One neuron outputs a continuous severity score. Linear activation is used. Task 2 (Auxiliary Task - Classification): A three-neuron layer outputs a probability distribution of three severity levels. Activated using Softmax.

[0070] Multi-task learning: Jointly train regression and classification tasks, enabling the model to simultaneously achieve accurate scoring and clear classification.

[0071] Its regression loss The smoothed L1 loss between the predicted score S and the true label is more robust to outliers. Classification loss Cross-entropy loss between predicted probability and true label (one-hot encoding).

[0072] Total loss: ,in, and As a balancing hyperparameter.

[0073] Where N is the number of training samples. For the predicted score of the i-th sample, Let i be the true severity label for the i-th sample. The smoothed L1 loss function is expressed as follows: 1 is an indicator function; its value is 1 when the condition is true, and 0 otherwise. Let be the probability of belonging to category c.

[0074] First, a pre-trained encoder is trained. Then, the pre-trained encoder is loaded, and its parameters are fixed. Samples are input into the encoder to obtain feature vectors, which, along with contextual and temporal features, are input into the main network. The Adam optimizer is then used to optimize the network. Train for the target.

[0075] Finally, the encoder parameters are unfrozen and fine-tuned end-to-end with the main network at a small learning rate, so that feature extraction can better serve the final task.

[0076] The validation set is then used to obtain the final nutritional risk model.

[0077] Understandably, this invention, through its systematic data engineering and expert-driven annotation system, fundamentally improves the quality and representativeness of model training data, thus laying the foundation for high model reliability. Instead of directly using raw, noisy clinical data, this invention creatively constructs a data structure pairing "current observation point" with "individual historical baseline." This "individual self-control" design ensures that the model's learning target is not an abstract, universal standard, but rather each patient's unique, clinically significant "deviation pattern." More importantly, it introduces dual annotation based on prospective clinical outcomes (adverse events in the next 2-4 weeks) by a multidisciplinary expert committee. This not only provides the labels necessary for supervised learning but also ensures that these labels contain profound clinical wisdom and causal logic (i.e., what consequences a certain deviation pattern actually leads to). In this way, the model is guided to learn feature associations that truly have prognostic significance, rather than superficial statistical correlations, enhancing the clinical rationality of the model's internal logic and its long-term predictive value.

[0078] Secondly, the two-stage architecture of "pre-training-fine-tuning" and the multi-dimensional fusion mechanism enhance the model's ability to handle complex, high-dimensional, and small-sample medical data, and improve its generalization and robustness. In the medical field, where data annotation is costly, directly training complex models can easily lead to overfitting. Unsupervised pre-training allows the encoder to learn the general patterns and internal structures of ketogenic patient indicator changes, essentially giving the model the "basic grammar" of the field and utilizing a large amount of information from unlabeled data. Subsequently, the fine-tuning stage combines high-quality labeled samples for targeted optimization. This strategy alleviates the difficulties of small-sample learning, enabling the model to extract key knowledge more efficiently from limited gold-standard cases. The architecture, which integrates multi-channel (deviation features, contextual features, and temporal features) fusion with an attention mechanism, simulates the thought process of clinical experts in comprehensive judgment: focusing on the deviation between the present and the past (channel A), considering the overall context such as treatment stage and medication (channel B), and analyzing the temporal evolution trend of key indicators (channel C). Attention weights automatically identify the most important information sources in each prediction, which not only improves the model's performance but also provides an interpretable window for decision-making, turning the model from a "black box" to a "gray box".

[0079] Furthermore, this invention achieves an optimal balance between accuracy, granularity, and efficiency in risk assessment through multi-task learning and a hierarchical risk triggering mechanism. The model simultaneously performs regression (outputting continuous scores S) and classification (outputting grade probabilities P), enabling it to provide both nuanced risk quantification (e.g., distinguishing between moderate to mild) and clear grade judgments. This addresses the dual clinical needs for accurate assessment and clear decision-making. Crucially, this complex deep learning model is only activated for in-depth comparison when initial screening reaches medium to high risk. This creates an efficient "funnel" mechanism: lightweight rules quickly filter out a large number of low-risk cases, saving valuable computational resources and clinical attention. Resources are then concentrated on in-depth, complex intelligent analysis of medium to high-risk cases. This avoids the waste of resources by using a sledgehammer to crack a nut, ensuring that high-end analytical tools are used where they are most needed, thus optimizing the overall efficiency of technology application.

[0080] In some embodiments of this application, determining the severity based on the comparison results includes: The results are compared to make a preliminary judgment on the severity, and the preliminary severity level is determined based on the probability distribution fusion function.

[0081] Based on historical data, a boundary region is pre-defined, and it is determined whether the comparison result is within the boundary region. When the comparison result is within the boundary region, the comparison result is corrected.

[0082] When the alignment results are corrected, the corrected alignment results are used as the final grade.

[0083] If the comparison results are not corrected, the preliminary severity level will be used as the final level.

[0084] It receives continuous severity scores, classification probability vectors, and fused attention weights from the model. Among them, the continuous scoring threshold is preset in advance: Pre-set probability thresholds: Pre-set the reliability threshold: Define piecewise functions Map continuous ratings to initial levels: Preliminary level: .

[0085] Calculate the probability dominance level: Calculate the maximum probability value: Calculate the entropy (a measure of uncertainty) of the probability distribution: when When the value is greater than 1.0, the model is considered to be uncertain about classification and tends to rely on continuous scoring. .

[0086] like ,and If so, the final level is the probability-dominated level.

[0087] like Then use weighted fusion: like If a conflict occurs, it will be recorded and reviewed.

[0088] Define the boundary region: ,when When in the boundary region, make corrections: in, For continuous score S through function The initial level obtained by direct mapping, For probability vectors The maximum value in, i.e. This represents the model's level of certainty regarding the most likely classification. For probability distribution entropy, To harmonize the decision-making rules through integration and The level obtained after the contradiction In order to be in Based on this, the revised grading system incorporates strict clinical rules (such as the occurrence of complications). The value is a Boolean sign; "true" indicates that the patient currently has a clear complication (such as ketoacidosis or severe electrolyte imbalance). This refers to the number of consecutive days that the current key indicators have deviated from their current state. The weights are pre-set based on clinical importance in the calculation of the historical deviation of the indicators.

[0089] Ultimately, if no correction is triggered, the initial severity level will be used as the final level; if a correction is triggered, the corrected comparison result will be used as the final level.

[0090] In some embodiments of this application, sending corresponding warning information based on severity includes: The severity levels are categorized into Level 1, Level 2, and Level 3 warnings.

[0091] When the final level is mild, a Level 1 warning is triggered; when the final level is moderate, a Level 2 warning is triggered; and when the final level is severe, a Level 3 warning is triggered.

[0092] The confidence level is calculated for the final rating, which includes high confidence, medium confidence, and low confidence.

[0093] When the confidence level is high, the warning level is determined based on the final level.

[0094] When the confidence level is medium, a level 2 warning is triggered.

[0095] When the confidence level is low, a level 3 warning is triggered.

[0096] Specifically, the overall confidence score is: Where, each term represents: probability confidence, consistency of continuous ratings, and probability distribution determinism, respectively. At a high confidence level, when At this point, the confidence level is medium. When the confidence level is low, the formulas mentioned above are as follows: For continuous severity scoring, This is the classification probability vector.

[0097] Understandably, this invention first enhances the robustness and clinical rationality of severity determination by establishing a three-layer fusion and decision-making mechanism of "continuous scoring - discrete probability - clinical rules." It doesn't simply rely on a single output of the model, but creatively allows two different output forms (continuous scoring `S` and categorical probability `P`) to mutually verify and balance each other, introducing explicit decision rules in case of conflict (e.g., following probability at high confidence, conservatively escalating at low confidence). This resolves potential contradictions within the model and reduces the possibility of misjudgment due to accidental fluctuations in a single output. More importantly, it adds a hard correction layer based on clear clinical facts (presence or absence of complications, duration of deviation). This ensures that the final "final grade" is not only derived from the data model but also rooted in inviolable medical principles, guaranteeing that in any situation, the presence of clear complications or long-term uncorrected deviations will receive a sufficient risk level escalation, thus solidifying the clinical safety baseline into the algorithmic logic.

[0098] Secondly, by clearly defining and handling "boundary regions" and "low-confidence" scenarios, the system enhances its ability to process complex or ambiguous cases with greater precision, reducing the decision-making risks caused by the algorithm's "arbitrary" classification. In clinical practice, patients in a "borderline state" are precisely the group most in need of careful evaluation. This invention formally acknowledges and defines the "boundary region" near the scoring threshold and pre-sets additional evaluation and escalation logic for such cases. This is equivalent to the system proactively identifying the "fuzzy zone" of its own judgment and triggering stricter review standards. Simultaneously, by calculating information entropy to quantify the "uncertainty" of the model's classification and integrating the model's own consistency and the degree of agreement between internal and external evidence through a comprehensive confidence score `C`, the solution can clearly distinguish between highly confident judgments and questionable judgments. For low-confidence judgments, instead of hiding or ignoring them, the system proactively "shows weakness" and requests more attention by triggering higher-level warnings (e.g., level two for medium confidence and level three for low confidence). This transparent handling and proactive management of uncertainty improves the system's reliability when facing complex and atypical cases, avoiding the potential risks caused by the algorithm's blind confidence.

[0099] Furthermore, this invention creatively couples "risk level" with "confidence level" to form a dynamic, hierarchical early warning trigger matrix, thereby achieving simultaneous optimization of clinical alert accuracy and resource allocation efficiency. Traditional early warning systems often mechanically trigger fixed responses based solely on risk level. This solution introduces confidence level as a key moderating variable, establishing distinct early warning paths such as "high risk + high confidence" and "high risk + low confidence." A high-confidence high-risk early warning signifies solid evidence, allowing for rapid initiation of standardized emergency intervention procedures. Conversely, a low-confidence high-risk early warning indicates "a suspicious signal but potentially serious consequences," prompting the system to trigger a higher-level (level 3) warning to generate broader and more urgent clinical attention, not only for intervention but also for urgent review. This design offers multiple benefits: it reduces the risk of inappropriate emergency intervention due to low-quality data or model misjudgment; it increases the clinical team's trust in the alerts; and finally, there is a profound and functionally complementary logical correlation between the probability distribution fusion function and the fusion decision rule. Together, they form a critical processing chain from multi-source uncertain information to stable, interpretable, and safe clinical decision-making. Their correlation is not a simple sequential order, but rather a synergistic mechanism of hierarchical verification and cross-verification. The probability distribution fusion function, at its core, is "quantitative verification." It uses data independent of the main model's input sources to perform external cross-verification on the "preliminary comparison results" generated by the main model. It answers the question: "Based on other data sources, is the model's preliminary judgment reliable?" The fusion decision rule, at its core, is "strategic adjudication." It deals with potential contradictions or inconsistencies between the two outputs (continuous score S and classification probability P) within the main model. It answers the question: "When the model's own continuous score and discrete probability point to different conclusions, which should we listen to? What is the final ranking?" Both serve the ultimate goal of "improving the robustness, safety, and interpretability of the final decision." The fusion decision rule is responsible for integrating the model's internal judgments to form a preliminary conclusion. The probability distribution fusion function is responsible for externally verifying the rationality of this preliminary conclusion, forming a closed loop from "internal analysis" to "external verification."

[0100] In summary, the beneficial effects of this invention are as follows: (1) By constructing a closed-loop screening process based on individualized historical benchmarks and intelligent stratification, the individualized accuracy and early detection capability of risk assessment are improved in the field of clinical nutrition risk management, especially in managing complex therapeutic nutritional interventions such as the ketogenic diet. By comparing current physical data with the patient's own historical data baseline, a "personalized" health and risk coordinate system is established. This self-comparison-based assessment mechanism can keenly capture early abnormal trends that have not yet exceeded the general clinical scope but have shown significant deviations from the patient's personal baseline.

[0101] (2) It optimized the efficiency of clinical resource allocation and reduced unnecessary medical burden. By introducing a hierarchical risk assessment logic, namely, preliminary screening is conducted first through historical deviation calculation, and only those cases reaching the medium-to-high risk level are analyzed in depth using more complex nutritional risk models, thus achieving "intelligent triage" of screening work. This avoids the waste of resources caused by indiscriminately inputting all data into complex model calculations, ensuring that high-value computing resources and clinical attention can be concentrated on patients who are truly facing higher risks. At the same time, it reduces the possibility of unnecessary anxiety and over-examination among low-risk patients due to false alarms from general tools, making the entire screening and intervention process more focused and efficient.

[0102] (3) It enhances the clinical relevance and operability of risk warnings, and improves the timeliness and pertinence of intervention measures. This invention uses an intelligent model to conduct in-depth comparisons and severity assessments of medium- and high-risk cases, and its output results can be more closely integrated with specific clinical scenarios and pathophysiological mechanisms. The generated warning information not only indicates the risk level, but also relates to specific abnormal indicators, deviation directions, and durations, so that subsequent warning information and intervention recommendations are no longer general health education, but rather targeted clinical action guidelines.

[0103] (4) Its output of early warning and intervention results can be fed back to update the patient's historical data baseline. This allows the patient's "health baseline" to dynamically evolve with changes in the treatment stage and the body's adaptation, and the risk assessment standards are continuously individualized accordingly. This dynamism overcomes the drawback of fixed standards being unable to adapt to changes in the patient's disease course, enabling screening to be carried out throughout the entire treatment process, whether in the adaptation period, the stable period, or the period of complication, providing risk assessments that fit the current state. In the long term, this accumulated continuous data also provides valuable resources for understanding the long-term response patterns of individuals to specific nutritional interventions.

[0104] (5) By finely classifying the severity of risks and matching them with differentiated early warning and response mechanisms, the system ensures that appropriate levels of response are taken for situations of different levels of urgency. For potential high-risk complications, the system can trigger a rapid clinical response through high-level early warnings, maximizing patient safety. For management deviations of medium and low risk, timely reminders can also help patients and their families fine-tune their behavior and consolidate treatment adherence. This tailored communication and intervention approach helps to establish a more proactive doctor-patient collaborative management relationship, transforming nutritional risk management from a passive crisis management to a proactive, collaborative health maintenance process.

[0105] In another preferred embodiment based on the above embodiments, see [reference] Figure 2As shown, this embodiment provides a nutrition risk screening system for applying the above-described nutrition risk screening method, including: The data acquisition module is configured to collect and preprocess user body data and historical data.

[0106] The feature extraction module is configured to compare historical data and body data to obtain several first historical deviation values, and determine the current user's deviation status based on these first historical deviation values.

[0107] The data processing module is configured to determine the user's nutritional risk level based on the deviation.

[0108] The data processing module is also configured to input body data into a nutritional risk model for comparison based on nutritional risk levels.

[0109] The early warning module is configured to determine the severity based on the comparison results and send corresponding early warning information based on the severity.

[0110] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0112] These computer program instructions may also be stored in a computer-readable storage device that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage device produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A nutritional risk screening method, characterized in that, The method comprises the following steps: Collecting and preprocessing user body data and historical data; Comparing the historical data and the body data to obtain a plurality of first historical deviation values, determining the current user deviation condition based on the plurality of first historical deviation values; Determining the user's nutritional risk level based on the deviation condition, and inputting the body data into a nutritional risk model based on the nutritional risk level for comparison; Determining the severity based on the comparison result, and sending corresponding warning information based on the severity.

2. The nutritional risk screening method according to claim 1, characterized in that, When collecting and preprocessing user body data and historical data, the method comprises the following steps: The body data at least includes blood ketone value, carbohydrate intake proportion, attack frequency, blood potassium / magnesium level and ketogenic compliance score; Filtering and removing abnormal period data based on the historical data; Determining the historical reference value and historical reference standard deviation based on the preprocessed historical data.

3. The nutritional risk screening method according to claim 2, characterized in that, When comparing the historical data and the body data to obtain a plurality of first historical deviation values, the method comprises the following steps: Comparing the historical reference value and the historical reference standard deviation with the body data to obtain relative deviation and absolute deviation; Verifying the relative deviation and the absolute deviation based on the normal range of ketogenic adaptation, when the verification is valid, retaining the related data, when the verification is invalid, marking as normal fluctuation and assigning a value of 0.

4. The nutritional risk screening method according to claim 3, characterized in that, When comparing the historical data and the body data to obtain a plurality of first historical deviation values, the method further comprises the following steps: Determining the ketogenic diet execution stability coefficient and the clinical symptom correlation coefficient based on the historical data; Based on the ketogenic diet execution stability coefficient and the clinical symptom correlation coefficient, the relative deviation and the absolute deviation after verification are verified again to obtain a plurality of first historical deviation values.

5. The nutritional risk screening method according to claim 4, characterized in that, When determining the current user deviation condition based on a plurality of first historical deviation values, the method comprises the following steps: When the absolute deviation is greater than 30% or the absolute deviation rate exceeds the normal range ± 50%, it is a high-weight indicator and is determined as a serious deviation; When the relative deviation rate is greater than 40% or the absolute deviation rate exceeds the normal range ± 60%, it is a medium-weight indicator and is determined as a moderate deviation; When the relative deviation rate is greater than 50% or obvious abnormalities occur, it is a low-weight indicator and is determined as a slight deviation.

6. The nutritional risk screening method according to claim 5, characterized in that, When determining the user's nutritional risk level based on the deviation condition, the method comprises the following steps: When only one low-weight indicator is slightly deviated, and high / medium-weight indicators are not deviated, it is determined as a low risk level; When 2-3 medium / low-weight indicators are moderately deviated, or one high-weight indicator is slightly deviated, it is determined as a medium risk level; When ≥2 high-weight indicators are seriously deviated, or ≥4 medium / low-weight indicators are moderately deviated, it is determined as a high risk level.

7. The nutritional risk screening method according to claim 6, characterized in that, When inputting the body data into a nutritional risk model based on the nutritional risk level for comparison, the method comprises the following steps: Divide the historical data into a training set and a test set, and use the training set to pre-train the encoder; Freeze the encoder parameters and train the main network; Unfreeze the encoder parameters, synchronize with the main network for fine-tuning, and obtain the nutritional risk model; When the nutritional risk level reaches the medium risk level or the high risk level, the body data is input into a nutritional risk model for comparison.

8. The nutritional risk screening method according to claim 7, characterized in that, When determining the severity based on the comparison result, the steps include: performing a preliminary severity judgment on the comparison result, and determining the preliminary severity level based on a probability distribution fusion function; previously setting a boundary region based on historical data and determining whether the comparison result is in the boundary region, and when the comparison result is in the boundary region, correcting the comparison result; when the comparison result is corrected, the corrected comparison result is taken as the final level; when the comparison result is not corrected, the preliminary severity level is taken as the final level.

9. The nutritional risk screening method according to claim 8, characterized in that, When sending the corresponding warning information based on the severity, the steps include: the severity includes a first-level warning, a second-level warning, and a third-level warning; when the final level is mild, a first-level warning is triggered, when the final level is moderate, a second-level warning is triggered, and when the final level is severe, a third-level warning is triggered; performing a confidence calculation on the final level, the confidence including high confidence, medium confidence, and low confidence; when the confidence is high, determining the warning level based on the final level; when the confidence is medium, triggering a second-level warning; when the confidence is low, triggering a third-level warning.

10. A nutritional risk screening system for use in the method of nutritional risk screening according to any one of claims 1 to 9, characterized in that The steps include: a collection module configured to collect user body data and historical data and perform preprocessing; a feature extraction module configured to compare the historical data and the body data to obtain a plurality of first historical deviation values, and determine a current user deviation condition based on the plurality of first historical deviation values; a data processing module configured to determine a user nutritional risk level based on the deviation condition; the data processing module is further configured to input the body data into a nutritional risk model for comparison based on the nutritional risk level; a warning module configured to determine the severity based on the comparison result, and send the corresponding warning information based on the severity.

Images