System for predicting the risk of diarrhea in enteral nutrition in critical patients

Abstract

The present application relates to the technical field of critical nutrition medical treatment, and discloses a severe patient enteral nutrition diarrhea risk prediction model system. The nutrition parameter acquisition module of the system is used for acquiring enteral nutrition formula parameters of the severe patient, including nutrition solution osmotic pressure, infusion speed and nutrition solution temperature; the physiological index analysis module is used for collecting intestinal flora abundance data, serum albumin level and gastrointestinal peristalsis frequency of the severe patient; the risk feature extraction module is used for calculating intestinal tolerance characteristic values according to the enteral nutrition formula parameters and the intestinal flora abundance data; and the prediction model construction module is used for constructing a diarrhea risk prediction model based on the intestinal tolerance characteristic values and the serum albumin level. The system can realize comprehensive collection and analysis of severe patient enteral nutrition related data, improve the objectivity and accuracy of diarrhea risk prediction, and provide support for clinical enteral nutrition scheme adjustment and diarrhea prevention.

Description

Technical Field

[0001] This invention relates to the field of critical care nutrition and medical technology, specifically to a risk prediction model system for enteral nutrition diarrhea in critically ill patients. Background Technology

[0002] In intensive care settings, enteral nutrition support is a crucial intervention for maintaining vital signs and ensuring the body's metabolic needs in critically ill patients. It provides essential nutrients through the intestinal route, helping patients overcome the critical stage of their illness. However, due to factors such as high stress levels, impaired intestinal barrier function, and decreased immune levels, diarrhea is a common and challenging complication during enteral nutrition support in critically ill patients. Diarrhea not only significantly reduces the efficiency of intestinal nutrient absorption, hindering the body's recovery, but it can also lead to further gut microbiota dysbiosis, electrolyte imbalance, dehydration, and in severe cases, worsen the patient's condition, prolong hospital stays, and increase the difficulty and resource consumption of medical care.

[0003] The assessment and prediction of diarrhea risk in critically ill patients receiving enteral nutrition often relies on subjective judgment based on the clinical experience of healthcare professionals, lacking a systematic and standardized assessment framework. While some medical institutions attempt to preliminarily assess diarrhea risk by monitoring single indicators (such as the infusion rate of the nutritional solution or the patient's body temperature), this approach fails to comprehensively consider the correlation between the patient's physiological state and the parameters of the enteral nutrition formula, resulting in insufficient accuracy and reliability in risk prediction. For example, focusing solely on the infusion rate while ignoring changes in the patient's gut microbiota abundance makes it difficult to accurately identify the risk of diarrhea caused by microbiota imbalance; similarly, monitoring only serum albumin levels without considering intestinal tolerance characteristics fails to comprehensively assess the patient's gut's adaptability to nutrients, thus hindering early warning of diarrhea.

[0004] Current technologies lack a dedicated system for integrating multi-dimensional data and building predictive models. After obtaining patients' nutritional parameters and physiological indicators, healthcare professionals must manually organize and analyze the data, which is not only time-consuming and labor-intensive but also prone to human error leading to biased data processing results, affecting the timeliness and accuracy of risk assessment. Significant individual differences exist among critically ill patients in terms of disease severity, underlying diseases, and intestinal function. Single assessment criteria or predictive methods are insufficient to meet the diverse needs of patients and cannot provide effective references for healthcare professionals to develop personalized enteral nutrition plans and take preventative measures against diarrhea. Consequently, the incidence of enteral nutrition-related diarrhea in critically ill patients remains high in clinical practice, urgently requiring a systematic and scientific approach to address this issue. Summary of the Invention

[0005] The purpose of this invention is to provide a model system for predicting the risk of diarrhea in critically ill patients receiving enteral nutrition, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a risk prediction model system for diarrhea in critically ill patients receiving enteral nutrition, the system comprising:

[0007] The module includes a nutritional parameter acquisition module, a physiological indicator analysis module, a risk feature extraction module, and a prediction model construction module.

[0008] The nutrition parameter acquisition module is used to acquire the enteral nutrition formula parameters for critically ill patients, including the osmotic pressure of the nutrient solution, the infusion rate, and the temperature of the nutrient solution.

[0009] The physiological index analysis module is used to collect data on the abundance of gut microbiota, serum albumin level, and gastrointestinal motility frequency of the critically ill patients.

[0010] The risk feature extraction module is used to calculate intestinal tolerance feature values ​​based on the enteral nutrition formula parameters and the intestinal flora abundance data;

[0011] The prediction model building module is used to construct a diarrhea risk prediction model based on the intestinal tolerance characteristic value and the serum albumin level.

[0012] Preferably, when the nutrient parameter acquisition module acquires the enteral nutrition formula parameters, the steps include:

[0013] Collect historical data of the osmotic pressure of the nutrient solution and extract the osmotic pressure fluctuation range;

[0014] Monitor the real-time change curve of the infusion rate and calculate the rate stability coefficient;

[0015] Record the temperature deviation of the nutrient solution and generate a temperature deviation index.

[0016] Preferably, when the physiological indicator analysis module collects the gut microbiota abundance data, the following steps are performed:

[0017] Based on the osmotic pressure fluctuation range and the velocity stability coefficient, the time points for collecting gut microbiota samples were selected.

[0018] The gut microbiota abundance data is obtained by adjusting the microbiota abundance detection threshold based on the temperature deviation index.

[0019] Preferably, when the risk feature extraction module calculates the intestinal tolerance feature value, the steps include:

[0020] The gut microbiota abundance data is matched with a preset microbiota baseline distribution to generate a microbiota imbalance score.

[0021] The intestinal mucosal integrity index was calculated by combining the serum albumin level and the gastrointestinal motility frequency.

[0022] The intestinal tolerance characteristic value is generated based on the gut microbiota imbalance score and the intestinal mucosal integrity index.

[0023] Preferably, when the prediction model building module builds the diarrhea risk prediction model, the steps include:

[0024] Using the gut tolerance characteristics as input variables and historical diarrhea occurrence records as supervision labels, an initial risk prediction model is trained.

[0025] The weight parameters of the initial risk prediction model are adjusted based on the dynamic changes in serum albumin levels.

[0026] The diarrhea risk prediction model is generated by integrating multiple adjusted weight parameters.

[0027] Preferably, when the prediction model building module adjusts the weight parameters of the initial risk prediction model, the steps include:

[0028] Extract the rate of decrease in serum albumin levels and calculate the albumin decay coefficient;

[0029] The intestinal tolerance characteristic value is weighted and corrected based on the albumin attenuation coefficient.

[0030] The weight parameters are updated based on the corrected gut tolerance feature values.

[0031] Preferably, when the prediction model construction module integrates multiple adjusted weight parameters, the execution steps include:

[0032] The albumin decay coefficients were collected at different time periods to generate a decay coefficient sequence;

[0033] The number of weight parameters integrated is determined based on the variance of the attenuation coefficient sequence.

[0034] The average value of the weight parameters within the integrated number of weight parameters is taken to generate the final weight parameters.

[0035] Preferably, when the risk feature extraction module generates the intestinal tolerance feature value, it further performs the following steps:

[0036] Monitor the number of abnormal fluctuations in the frequency of gastrointestinal peristalsis and calculate the peristalsis disorder index;

[0037] The peristalsis disorder index and the microbial imbalance score are combined and normalized to generate a comprehensive tolerance score;

[0038] The intestinal tolerance characteristic value is corrected based on the comprehensive tolerance score.

[0039] Preferably, when the physiological indicator analysis module collects the serum albumin level, the following steps are performed:

[0040] Based on the comprehensive tolerance score, a time window for serum albumin detection was selected;

[0041] Within the serum albumin detection time window, serum albumin concentration gradient data are extracted;

[0042] The serum albumin level is generated based on the serum albumin concentration gradient data.

[0043] Preferably, when the prediction model building module trains the initial risk prediction model, the following steps are performed:

[0044] Based on the comprehensive tolerance score and the serum albumin concentration gradient data, the training dataset and the validation dataset are divided.

[0045] The initial risk prediction model is iteratively optimized based on the training dataset.

[0046] The performance of the optimized initial risk prediction model was validated using the validation dataset.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The nutritional parameter acquisition module can specifically acquire enteral nutrition formula parameters for critically ill patients, including osmotic pressure, infusion rate, and temperature of the nutrient solution. These parameters are key factors affecting intestinal tolerance to enteral nutrition. Effective acquisition of these parameters allows for a comprehensive understanding of the core variables in the enteral nutrition process, avoiding incomplete risk assessments due to missing key nutritional parameters. Compared to traditional methods relying on scattered manual recording of nutritional parameters, this module enables centralized and standardized acquisition of these parameters, reducing data omissions and recording errors, and providing accurate basic data support for subsequent risk analysis.

[0049] The physiological indicator analysis module focuses on collecting data on gut microbiota abundance, serum albumin levels, and gastrointestinal motility frequency from critically ill patients. These physiological indicators directly reflect the patient's intestinal function and overall nutritional status. Gut microbiota abundance data reflects the balance of the gut microbiota, and microbiota imbalance is one of the important causes of diarrhea. Serum albumin levels reflect the patient's nutritional reserves and intestinal absorption function. Gastrointestinal motility frequency is closely related to intestinal digestion and excretion. By collecting these physiological indicators, the actual state of the patient's gut and body can be understood from different dimensions, providing multi-dimensional physiological data for subsequent risk feature extraction and predictive model construction, breaking the limitations of traditional risk assessment relying solely on a single physiological indicator.

[0050] The risk feature extraction module calculates intestinal tolerance characteristic values ​​based on enteral nutrition formula parameters acquired by the nutritional parameter acquisition module and gut microbiota abundance data collected by the physiological indicator analysis module. This process organically combines nutritional supply parameters with patient gut physiological status data. By performing correlation analysis on these two types of data and extracting intestinal tolerance characteristic values, the scattered raw data can be transformed into feature information that directly reflects the patient's gut's current tolerance to enteral nutrition. These characteristic values ​​can intuitively reflect the patient's gut's adaptation to enteral nutrition, providing a core analytical object for subsequent construction of a diarrhea risk prediction model. This makes risk prediction more targeted and scientific, avoiding the problem of inaccurate risk feature identification due to a lack of effective correlation between data.

[0051] The predictive model construction module is based on the intestinal tolerance feature values ​​obtained from the risk feature extraction module and the serum albumin levels collected from the physiological indicator analysis module to construct a diarrhea risk prediction model. This module integrates key feature values ​​and physiological indicators, transforming multi-dimensional data into a predictive model that can systematically predict the risk of diarrhea during enteral nutrition in critically ill patients. Through this model, healthcare professionals can anticipate the likelihood of diarrhea in patients receiving enteral nutrition support and adjust enteral nutrition protocols accordingly, such as optimizing nutrient solution formulations, adjusting infusion parameters, or taking targeted preventative measures to reduce the likelihood of diarrhea. This model avoids the drawbacks of traditional risk assessments relying on the subjective experience of healthcare professionals, reduces the impact of human factors on risk prediction results, and improves the objectivity and accuracy of risk prediction. It helps provide tailored enteral nutrition risk assessment services for critically ill patients with different individual differences, adapting to diverse clinical needs. This enables clinical healthcare professionals to conduct enteral nutrition support and diarrhea prevention more efficiently and accurately, improves the therapeutic effect of enteral nutrition in critically ill patients, reduces various complications caused by diarrhea, and alleviates patient suffering and the burden of medical care. Attached Figure Description

[0052] Figure 1This is a timeline diagram of the enteral nutrition diarrhea risk prediction model system for critically ill patients as described in this invention.

[0053] Figure 2 A schematic diagram illustrating the working principle of the nutrient parameter acquisition module for obtaining enteral nutrition formula parameters.

[0054] Figure 3 This diagram illustrates the working principle of the risk feature extraction module in calculating intestinal tolerance feature values. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 This invention provides a diarrhea risk prediction model system for critically ill patients undergoing enteral nutrition. The system integrates multiple modules that work collaboratively to achieve accurate prediction of diarrhea risk during enteral nutrition for critically ill patients. The system mainly includes a nutritional parameter acquisition module, a physiological indicator analysis module, a risk feature extraction module, and a prediction model construction module. The nutritional parameter acquisition module is responsible for acquiring enteral nutrition formula parameters for critically ill patients, including nutrient solution osmotic pressure, infusion rate, and nutrient solution temperature. These parameters are fundamental data for assessing intestinal tolerance. The physiological indicator analysis module collects data on intestinal flora abundance, serum albumin levels, and gastrointestinal motility frequency from critically ill patients. These physiological indicators reflect the patient's intestinal health and metabolic state. The risk feature extraction module calculates intestinal tolerance feature values ​​based on the enteral nutrition formula parameters and intestinal flora abundance data. These feature values ​​comprehensively characterize the intestine's adaptability to the nutrient solution. The prediction model construction module constructs a diarrhea risk prediction model based on the intestinal tolerance feature values ​​and serum albumin levels. It optimizes the model parameters using machine learning methods to achieve dynamic prediction of diarrhea risk. Through modular design, the system automates the data acquisition, feature extraction, and model building processes, thereby improving the accuracy and efficiency of predictions.

[0057] Example 1: See Figure 2After the nutrient parameter acquisition module is activated, it first continuously monitors the osmotic pressure of the nutrient solution flowing through the tubing using high-precision osmotic pressure sensors integrated into the enteral nutrition infusion system. These sensors collect data at fixed intervals and mark the data with timestamps, forming a structured historical data sequence. The system performs statistical analysis on the historical osmotic pressure data over time, not by simply calculating the average value, but by identifying the peak and trough values ​​throughout the entire nutritional support cycle. Algorithms are then used to calculate the range and coefficient of variation, thereby extracting the dynamic parameter of osmotic pressure fluctuation range. This range reflects the changing trend of the potential osmotic load exerted by the nutrient solution on the intestines.

[0058] Infusion rate monitoring is achieved using an embedded flow meter. This device records the volume of nutrient solution flowing per unit time at a high-frequency sampling rate, generating a continuous real-time change curve. The system preprocesses this curve, employing digital filtering technology to eliminate instantaneous noise interference caused by pipeline vibration or pumping pulses. Subsequently, piecewise linear analysis is performed on the smoothed velocity curve, calculating the ratio of the standard deviation to the average value of the velocity within each time period to derive a velocity stability coefficient. This coefficient quantitatively describes the smoothness of the infusion process; a lower coefficient value indicates more uniform infusion. Nutrient solution temperature data is collected by a temperature probe attached to the outer wall of the delivery pipeline. The probe compares the measured temperature with a preset ideal temperature range (typically around 37 degrees Celsius) in real time. The system records the absolute value of the temperature deviation at each sampling moment and integrates all deviation values ​​over a period of time to generate a comprehensive temperature deviation index. This index cumulatively reflects the inaccuracy of temperature control. After completing the nutrient parameter acquisition, the physiological index analysis module begins operation. Its process of collecting intestinal flora abundance data is closely related to the above parameters.

[0059] The system's built-in decision logic receives osmotic pressure fluctuation range and velocity stability coefficient from the nutrient parameter acquisition module. When the osmotic pressure fluctuation range exceeds a preset threshold or the velocity stability coefficient shows significant instability, the system automatically marks this period as a period of intestinal environmental stress and sets these time points as the optimal collection time for intestinal flora samples. This is to obtain the flora status at the moment when the intestine is most likely to be affected by changes in nutrient solution parameters, making the samples more representative and timely. Sample collection is completed through an indwelling intestinal drainage tube or fecal sampling, and the sampling operation is prompted by the system to be performed by medical staff. When performing flora abundance detection in the laboratory, the system introduces the temperature deviation index as an environmental correction factor. The raw signal output of the detection equipment (such as a quantitative PCR instrument or metagenomic sequencing platform) undergoes a threshold adjustment process, which is dynamically adjusted by the temperature deviation index. When the temperature deviation index is high, indicating poor nutrient solution temperature control, the system instructs the detection equipment to lower the signal threshold for species abundance determination, so as to capture the signals of weak flora that may have decreased in number or activity due to temperature stress, avoiding missed detections. When the temperature is well controlled, the standard threshold is used.

[0060] The gut microbiota abundance data acquired in this way is the result after correction for nutrient solution temperature parameters. It includes the relative abundance information of key bacterial groups such as Firmicutes and Bacteroidetes, and is ultimately output in the form of a digital matrix. Nutritional parameters are not only independent monitoring indicators, but also the basis for decision-making in guiding physiological indicator sampling and analysis strategies. The osmotic pressure fluctuation range and velocity stability coefficient jointly define the "opportunity window" for gut microbiota monitoring, while the temperature deviation index finely controls the "sensitivity knob" of microbiota detection. This dynamic adaptation mechanism ensures that the obtained gut microbiota abundance data is no longer an isolated static snapshot, but contains dynamic response information in the context of nutritional intervention, providing high-quality and context-rich input data for subsequent calculation of gut tolerance characteristic values. Data flow between modules is automatic and requires no manual intervention, forming a seamless conversion chain from physical parameters to biological indicators.

[0061] Example 2: See Figure 3After the risk feature extraction module is activated, it first processes the gut microbiota abundance data transmitted from the physiological indicator analysis module. This data is typically presented as a multidimensional vector composed of relative abundance percentages at different bacterial phyla, classes, orders, families, and genera. The module internally stores a preset baseline distribution of gut microbiota, derived from the aggregated analysis of gut microbiota data from a large number of clinically healthy individuals or critically ill patients in a stable nutritional support phase, representing an ideal ecological balance. The system performs similarity matching calculations between the real-time acquired patient gut microbiota abundance data vector and this baseline distribution vector. This calculation is not a simple difference comparison, but rather uses metrics such as the cosine of the angle between spatial vectors or the reciprocal of the Euclidean distance to assess the closeness of the two in multidimensional space. The matching result outputs a similarity score between 0 and 1. This score is then subtracted from 1 and multiplied by a standardized coefficient to generate an easily interpretable gut microbiota imbalance score. A higher score indicates a greater deviation from the healthy baseline and a more significant gut microbiota imbalance.

[0062] While generating a gut microbiota imbalance score, the module simultaneously reads the specific values ​​of serum albumin levels and time-series data on gastrointestinal motility frequency. Serum albumin level is a single-point quantitative indicator, usually measured in grams per liter, while gastrointestinal motility frequency is the average number of peristalsis over a period of time (e.g., 24 hours). The system incorporates these two physiological indicators into a multivariate fusion algorithm, which assigns different contribution weights to albumin level and motility frequency. For example, serum albumin level may be given a higher weight due to its strong correlation with intestinal mucosal repair capacity, while motility frequency, as a reflection of intestinal motility, is given a secondary weight. The algorithm performs a weighted combination process to generate a comprehensive intestinal mucosal integrity index, which is a dimensionless scalar value whose magnitude is positively correlated with the inferred intestinal mucosal health.

[0063] The module integrates the gut microbiota imbalance score and intestinal mucosal integrity index obtained from the preceding steps to generate the final gut tolerance feature value. The integration process is not a simple arithmetic average but takes into account the potential nonlinear interaction between the two. The system uses a pre-defined transformation function, whose parameters are trained using historical clinical data, to map the gut microbiota imbalance score (generally, a higher value indicates a worse condition) and the intestinal mucosal integrity index (a higher value indicates a better condition) onto a unified metric. The output of the transformation function is the gut tolerance feature value, which serves as a condensed and comprehensive indicator summarizing the patient's gut's current intrinsic capacity to withstand nutritional load. Subsequently, the prediction model building module begins operation, using the gut tolerance feature value calculated in real-time by the risk feature extraction module as the primary input variable. The module retrieves the patient's historical records from the hospital information system's electronic medical record database, focusing on extracting information such as whether diarrheal events occurred, the specific time of occurrence, the duration, and the severity level. This information is then converted into binary or multi-class supervised labels (e.g., 0 represents no diarrhea, 1 represents mild diarrhea, and 2 represents moderate to severe diarrhea).

[0064] The system employs a supervised machine learning algorithm, such as logistic regression or support vector machine, using intestinal tolerance features as characteristics and diarrhea occurrence records as labels to train the model and generate an initial risk prediction model. This initial model establishes a preliminary mapping relationship from intestinal tolerance status to diarrhea risk probability. After the initial model is established, the module incorporates dynamic change data of serum albumin levels for model optimization. This dynamic data is represented by a series of serum albumin test values ​​arranged in chronological order. The system analyzes this time series to calculate the trend of albumin level changes, such as whether it is decreasing, stable, or increasing, especially calculating its rate of decrease within a certain time window. Based on the dynamic information of albumin changes, the module initiates a weight adjustment routine. This routine assesses the inadequacy of the current initial model's sensitivity to albumin changes and fine-tunes the weight parameters connecting the intestinal tolerance features and the output results within the model using an iterative algorithm (such as gradient descent), enabling the model to better respond to the increased diarrhea risk caused by deteriorating nutritional status (manifested as a decrease in albumin).

[0065] The module performs a model ensemble step. Instead of relying solely on weight parameters adjusted at a single time point, it repeats the weight adjustment process at different time points during the patient's hospitalization (e.g., every 24 hours or whenever new albumin data becomes available), thus obtaining multiple sets of adjusted weight parameters. The system aggregates these weight parameters, adjusted at different times based on slightly different physiological states, using an ensemble learning strategy, such as averaging all these weight parameters or weighted averaging based on the quality assessment of the data used in each adjustment, ultimately generating a set of integrated, more robust weight parameters, thus forming the final diarrhea risk prediction model. This final model comprehensively considers the immediate intestinal tolerance state and dynamic changes in serum nutritional indicators, achieving more accurate and stable risk prediction. The entire process, from biomarker extraction to statistical model optimization, forms a closed-loop, adaptively updated computational system.

[0066] Example 3: The module first processes time-series data on serum albumin levels. This data comes from periodic blood biochemistry test results, arranged chronologically to form a concentration curve. The system does not simply focus on the instantaneous absolute value of albumin, but rather emphasizes analyzing its changing trend, particularly the decreasing trend. The system uses a linear regression method to fit the albumin concentration values ​​at several recent time points. The slope of the resulting fitted line is defined as the albumin decrease rate, which quantifies the amount of decrease in albumin concentration per unit time. To transform this instantaneous rate into a more universally applicable indicator, the system compares the decrease rate with a preset reference decrease threshold and normalizes it using a proportional function, generating a dimensionless albumin decay coefficient. The magnitude of this coefficient directly reflects the urgency of the patient's nutritional status deterioration.

[0067] The system uses this albumin decay coefficient to correct the intestinal tolerance feature values ​​transmitted from the risk feature extraction module. The correction logic embodies a risk-weighted approach, assuming that under the physiological background of rapid albumin decline, the actual diarrhea risk corresponding to the same intestinal tolerance feature value will increase. Therefore, the correction operation is not a simple arithmetic multiplication, but rather incorporates the decay coefficient as a regulating factor into a weighting function. This function can be expressed as:

[0068] ;

[0069] in: Represents the original intestinal tolerance characteristic value. This represents the calculated albumin attenuation coefficient. It is an empirical proportionality constant obtained through learning from historical data, used to control the influence of the attenuation coefficient, while This is the corrected intestinal tolerance characteristic value. Through this relationship, the albumin decay is effectively integrated into the assessment of intestinal status, so that the characteristic values ​​input to the model include information on the dynamic changes in nutrient metabolism.

[0070] The module is based on the modified gut tolerance metric. The system updates the weight parameters of the initial risk prediction model, which were originally obtained by minimizing the prediction error on the training dataset. Now, the system uses the corrected feature values ​​and corresponding diarrhea event labels at different time points to form new training samples, performing an additional training iteration on the initial model. The learning objective of this iteration is to fine-tune the weight parameters while maintaining the original model structure, making the model more sensitive to the feature values ​​corrected by albumin decay. Common optimization algorithms such as gradient descent are reused to calculate the gradient of the loss function with respect to the weights and make small updates in the opposite direction of the gradient, allowing the model's decision boundary to adapt to this new change in input features. This process is equivalent to allowing the model to relearn how to more accurately interpret the risk information implied by intestinal tolerance features in the specific physiological context of "albumin decay."

[0071] After adjusting the weights at a single time point, the module further performs a weight ensemble operation to enhance the model's stability. The system continuously monitors albumin levels throughout the patient's enteral nutrition support period and triggers the aforementioned weight adjustment process at each pre-defined assessment period (e.g., every 24 hours) or whenever a significant change in albumin is detected. In this way, over time, the system accumulates multiple sets of adjusted weight parameters, each reflecting the model's optimized state based on the latest albumin dynamics at a specific time point. These weight parameters constitute a weight set.

[0072] The system doesn't randomly select weights from this set; instead, it determines the integration strategy based on the sequence characteristics of albumin decay coefficients. It analyzes a historical sequence of albumin decay coefficients and calculates the variance of that sequence. This variance measures the volatility of the albumin decline pattern; a larger variance indicates that the patient's nutritional status is highly unstable and fluctuates drastically, requiring the model to integrate more diverse weight parameters to cover various possible states. Therefore, the system determines the number N of weight parameters to be integrated from a pre-defined reference table based on the variance value. For example, N is 5 when the variance falls within a certain interval, and 7 when it falls into a higher interval. After determining the number N, the system starts with the most recent weight parameters and backtracks to select the most recent N sets of weight parameters. Finally, the module integrates these N sets of weight parameters using a simple arithmetic mean. That is, it adds up the weights at corresponding positions in the N sets of weight parameters and then divides by N to obtain a new set of integrated weight parameters. This final set of weight parameters incorporates optimization information from recent physiological states, making its output more robust than weights adjusted at a single time point. It better addresses fluctuations in patients' physiological states, thus forming the final diarrhea risk prediction model for practical use. The entire implementation process demonstrates the gradual deepening and improvement of the model, from static initialization to dynamic adaptation, and then to multi-time-point knowledge fusion.

[0073] Example 4: The risk feature extraction module, based on the existing process, adds abnormal pattern recognition for time-series data on gastrointestinal motility frequency. The module receives raw motility frequency data collected from a borborygmus sensor or bioelectrical impedance monitoring device. This data is typically a continuous record in units of peristalsis counts per hour. The system first establishes a normal motility frequency range for critically ill patients based on extensive clinical data, for example, 3 to 8 times per hour. Subsequently, the system scans the entire data sequence using sliding time windows (e.g., every 4 hours), identifying and recording the number of times the motility frequency value exceeds the above normal range within each time window. For example, if the recorded frequency values ​​are 2, 9, 4, 10, 3, and 7 times / hour within a 4-hour window, then values ​​2, 9, and 10 are all abnormal, and the number of abnormal fluctuations in that window is 3. The module accumulates the total number of abnormal fluctuations across all time windows within a certain period (e.g., 24 hours) and calculates the ratio of this sum to the theoretical total number of monitoring counts within that period, ultimately obtaining a standardized motility disorder index. This index reflects the frequency of bowel motility rhythm disorders.

[0074] The module needs to integrate the newly calculated peristaltic disorder index with the previously existing microbiota imbalance score. Since these two indicators may have different dimensions and numerical distribution ranges, direct merging would introduce bias. Therefore, the system performs joint normalization on them. This process uses an improved min-maximum method, but instead of normalizing each indicator individually, it treats the data points of both indicators as a two-dimensional dataset, calculates the global minimum and maximum values ​​of this dataset, and then maps each data point (i.e., the data pair consisting of each patient's peristaltic disorder index and microbiota imbalance score) to a preset common range, such as the [0,1] interval. After this joint normalization, the two indicators become comparable, and their relative relationship is preserved. The system assigns weights to the normalized peristaltic disorder index and microbiota imbalance score; the weight coefficients can be derived based on historical data analysis, reflecting their relative importance to intestinal tolerance. Finally, a comprehensive tolerance score is generated through weighted summation. This score integrates information from both intestinal microbiome ecology and mechanomotor function.

[0075] The module uses this newly generated comprehensive tolerance score to correct the previously calculated intestinal tolerance feature values. The correction mechanism can be a linear or non-linear function. For example, a rule can be set: when the comprehensive tolerance score exceeds a certain threshold, it indicates that the intestine is in a state of high stress, and the original intestinal tolerance feature value needs to be appropriately amplified to improve the risk level it indicates; conversely, if the score is very low, the feature value may be moderately downgraded. This correction is dynamic and adaptive, so that the final output intestinal tolerance feature value not only includes the original flora and mucosal information, but also incorporates the real-time state of intestinal dynamics, thus becoming more comprehensive and sensitive. While the risk feature extraction module performs the above calculations, the physiological indicator analysis module also adjusts its serum albumin level collection strategy. The module receives the comprehensive tolerance score calculated in real time from the risk feature extraction module. Different score threshold ranges are preset within the system, and these ranges correspond to different intestinal state levels. When the overall tolerability score remains high or rises rapidly, the system determines that the patient's gut environment may be in a period of significant instability or pre-deterioration. In this case, more frequent monitoring of serum albumin levels is needed to detect potential changes in nutritional status. Therefore, the module dynamically selects and determines the detection window for serum albumin based on the score. When the score exceeds threshold A, the system will advance the planned next detection time and increase the detection frequency from twice a week to three times a week; this adjusted period is considered the critical detection window.

[0076] Within this defined serum albumin testing time window, the module schedules multiple blood sampling tests, not just one. For example, during a critical three-day window, blood samples are collected each morning to test albumin levels. This yields a series of serum albumin concentration values ​​arranged chronologically. The module extracts these concentration values ​​and calculates their gradient data, i.e., the concentration change between adjacent test points. See Table 1 for an example record of serum albumin concentration gradient data:

[0077] Table 1: Serum albumin concentration gradient within the key time window

[0078]

[0079] Based on the serum albumin concentration gradient data acquired within this time window, the module generates a serum albumin level for the predictive model. This "level" is not a simple arithmetic mean, but may contain information from multiple dimensions, such as the average concentration within the window, the lowest concentration, the average of the gradient data (i.e., the average daily rate of change), and the decreasing trend of the concentration (determined by the slope of linear regression). The system integrates this information into a comprehensive serum albumin level assessment result, which can more profoundly reflect the dynamic trajectory of the patient's protein nutritional status during specific stages of poor intestinal tolerance, providing more timely and targeted input variables for the predictive model. It demonstrates a closed-loop feedback system that starts with monitoring gastrointestinal motility abnormalities, generates a comprehensive score to correct core feature values, and then guides the frequency and interpretation of key physiological indicator collection. It emphasizes the intrinsic relationship between different physiological parameters and uses algorithms to transform this relationship into an operable and optimized data acquisition and feature engineering process.

[0080] Example 5: When the module starts the training process, it first needs to construct a dataset for model learning. At this time, the comprehensive tolerance score and serum albumin concentration gradient data become the core basis for dataset partitioning. The system collects complete data records from multiple historical patients. Each record contains the patient's comprehensive tolerance score at a specific time point, serum albumin concentration gradient data (such as average rate of change, downward trend indicators, etc.), and an outcome label indicating whether diarrhea occurred after that time point. The system adopts a partitioning strategy based on time series features, rather than simple random partitioning. It first analyzes the time series of the comprehensive tolerance score to identify the patient phase with consistently high scores and the stable phase with lower scores. At the same time, combined with the serum albumin concentration gradient data, it classifies data points showing a rapid downward trend into relatively stable data points. Then, the system ensures that both the training dataset and the validation dataset contain a sufficient proportion of different types of data points. For example, the training set contains 70% of high scores accompanied by a decrease in albumin and 70% of low scores with stable albumin, while the validation set contains the remaining 30% of data from each category. This stratified sampling method ensures that both the training and validation sets cover a wide range of key clinical scenarios, avoiding model evaluation bias caused by uneven data distribution. After the dataset is partitioned, the module uses the training dataset to iteratively optimize the initial risk prediction model. The initial model may choose a basic logistic regression model or a simple neural network as a starting point.

[0081] The iterative optimization process is a cyclical computational sequence. The system inputs a batch of data from the training dataset (e.g., data from a set of 32 patients at different time points) into the model. The model calculates the predicted diarrhea risk probability for each sample based on its current internal parameters (weights and biases). The system calculates the sum of differences between the model's predicted values ​​and the actual diarrhea outcome labels using a loss function (such as cross-entropy loss); this difference is called the loss value. The optimization goal is to minimize this loss value. The model automatically calculates the gradient of the loss value with respect to each internal parameter using the backpropagation algorithm. The gradient indicates the direction and magnitude in which each parameter should be adjusted to reduce the loss. The model uses an optimization algorithm (such as the Adam optimizer) to make small updates to all weight and bias parameters based on the calculated gradients. This process is called a training step or an iteration. The entire training dataset is typically passed through the model completely multiple times; each complete pass is called a training epoch. The system may perform hundreds or even thousands of training epochs. In each epoch, the training data is shuffled and input into the model to ensure that the model learns general patterns rather than noise in a specific order. As the iterations continue, the loss value will gradually decrease, indicating that the model's predictive ability is improving.

[0082] After the preset number of iterations for optimization is reached, or the loss value begins to decrease gradually, the module pauses the training process and enters the performance verification phase. At this point, the optimized initial risk prediction model is fixed and awaits evaluation. The system inputs a previously reserved validation dataset that was not used during training into this optimized model. The model outputs the predicted probability of diarrhea risk for each sample in the validation set. The system then compares the model's prediction results with the actual outcome labels in the validation set, generating a series of objective performance metrics. These metrics include, but are not limited to: accuracy, which is the proportion of samples correctly predicted by the model (whether predicting diarrhea or not); recall, which focuses on the proportion of samples that actually experience diarrhea, especially important for disease early warning; precision, which focuses on the proportion of samples predicted as diarrhea that actually experience diarrhea; and the area under the receiver operating characteristic (AUC), which comprehensively reflects the model's ability to distinguish between diarrhea and non-diarrhea samples; a higher AUC indicates better model performance.

[0083] The performance validation results guide subsequent operations. If the validation metrics meet preset pass / fail standards (e.g., AUC greater than 0.8 and recall higher than 0.75), the optimized model is considered a usable candidate model and may be archived or entered into broader testing. If the performance fails to meet the standards, the module initiates further tuning processes. Tuning may involve several aspects: First, returning to the iterative optimization steps to adjust training hyperparameters, such as lowering the learning rate to prevent oscillations, increasing the number of training epochs, or trying to use a more complex model structure (e.g., increasing the number of layers in the neural network). Second, re-examining the data partitioning strategy to check for data leakage or unreasonable partitioning. Third, collecting more training data to enrich the sample diversity for model learning. This "training-validation-tuning" cycle may be repeated multiple times until the model performance meets the requirements or resources are exhausted.

[0084] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A predictive model system for the risk of diarrhea in critically ill patients receiving enteral nutrition, characterized in that, It includes a nutritional parameter acquisition module, a physiological indicator analysis module, a risk feature extraction module, and a prediction model construction module; The nutrition parameter acquisition module is used to acquire the enteral nutrition formula parameters for critically ill patients, including the osmotic pressure of the nutrient solution, the infusion rate, and the temperature of the nutrient solution. The physiological index analysis module is used to collect data on the abundance of gut microbiota, serum albumin level, and gastrointestinal motility frequency of the critically ill patients. The risk feature extraction module is used to calculate intestinal tolerance feature values ​​based on the enteral nutrition formula parameters and the intestinal flora abundance data; The prediction model building module is used to build a diarrhea risk prediction model based on the intestinal tolerance characteristic value and the serum albumin level; When the prediction model building module builds the diarrhea risk prediction model, the following steps are performed: Using the gut tolerance characteristics as input variables and historical diarrhea occurrence records as supervision labels, an initial risk prediction model is trained. The weight parameters of the initial risk prediction model are adjusted based on the dynamic changes in serum albumin levels. The diarrhea risk prediction model is generated by integrating multiple adjusted weight parameters. When the prediction model building module adjusts the weight parameters of the initial risk prediction model, the steps include: Extract the rate of decrease in serum albumin levels and calculate the albumin decay coefficient; The intestinal tolerance characteristic value is weighted and corrected based on the albumin attenuation coefficient. The weight parameters are updated based on the corrected gut tolerance feature values.

2. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 1, characterized in that, When the nutrient parameter acquisition module acquires the enteral nutrition formula parameters, the following steps are performed: Collect historical data of the osmotic pressure of the nutrient solution and extract the osmotic pressure fluctuation range; Monitor the real-time change curve of the infusion rate and calculate the rate stability coefficient; Record the temperature deviation of the nutrient solution and generate a temperature deviation index.

3. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 2, characterized in that, When the physiological indicator analysis module collects the gut microbiota abundance data, the following steps are performed: Based on the osmotic pressure fluctuation range and the velocity stability coefficient, the time points for collecting gut microbiota samples were selected. The gut microbiota abundance data is obtained by adjusting the microbiota abundance detection threshold based on the temperature deviation index.

4. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 3, characterized in that, When the risk feature extraction module calculates the intestinal tolerance feature value, the following steps are performed: The gut microbiota abundance data is matched with a preset microbiota baseline distribution to generate a microbiota imbalance score. The intestinal mucosal integrity index was calculated by combining the serum albumin level and the gastrointestinal motility frequency. The intestinal tolerance characteristic value is generated based on the gut microbiota imbalance score and the intestinal mucosal integrity index.

5. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 4, characterized in that, When the prediction model building module integrates multiple adjusted weight parameters, the execution steps include: The albumin decay coefficients were collected at different time periods to generate a decay coefficient sequence; The number of weight parameters integrated is determined based on the variance of the attenuation coefficient sequence. The average value of the weight parameters within the integrated number of weight parameters is taken to generate the final weight parameters.

6. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 5, characterized in that, When the risk feature extraction module generates the intestinal tolerance feature value, it also performs the following steps: Monitor the number of abnormal fluctuations in the frequency of gastrointestinal peristalsis and calculate the peristalsis disorder index; The peristalsis disorder index and the microbial imbalance score are combined and normalized to generate a comprehensive tolerance score; The intestinal tolerance characteristic value is corrected based on the comprehensive tolerance score.

7. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 6, characterized in that, When the physiological indicator analysis module collects the serum albumin level, the following steps are performed: Based on the comprehensive tolerance score, a time window for serum albumin detection was selected; Within the serum albumin detection time window, serum albumin concentration gradient data are extracted; The serum albumin level is generated based on the serum albumin concentration gradient data.

8. The enteral nutrition diarrhea risk prediction model system for critically ill patients as described in claim 7, characterized in that, When the prediction model building module trains the initial risk prediction model, the following steps are performed: Based on the comprehensive tolerance score and the serum albumin concentration gradient data, the training dataset and the validation dataset are divided. The initial risk prediction model is iteratively optimized based on the training dataset. The performance of the optimized initial risk prediction model was validated using the validation dataset.

Images