A new-born jaundice health management system based on big data

Abstract

The application discloses a new-born jaundice health management system based on big data, and belongs to the technical field of jaundice health management, and specifically comprises the following steps: collecting the transcutaneous jaundice value, hemoglobin concentration, day age and weight data of a new-born through a data collection module; calculating individualized drift indexes based on the hemoglobin concentration and day age data through a physiological parameter calibration module; inputting the transcutaneous jaundice value and the individualized drift indexes into a dynamic correction model through a dynamic prediction analysis module to generate serum bilirubin prediction values; performing smoothing processing on the prediction values and calculating the change amount of adjacent time points by using a time series data processing module; constructing trend indexes and generating serum bilirubin change trend data through a trend analysis detection module; and finally generating a health management scheme based on threshold levels and trend change points through an intelligent decision module; and the application realizes effective evaluation and individualized management of the risk of new-born jaundice.

Description

Technical Field

[0001] This invention relates to the field of jaundice health management technology, specifically to a newborn jaundice health management system based on big data. Background Technology

[0002] Neonatal jaundice is one of the most common clinical problems in newborns, caused by abnormal bilirubin metabolism. Timely and accurate monitoring of bilirubin levels is crucial for preventing serious complications such as bilirubin encephalopathy. Traditional serum bilirubin testing requires blood sample collection, which is an invasive procedure and makes continuous monitoring difficult. With the development of medical technology, percutaneous bilirubin testing, due to its non-invasive and convenient characteristics, has become an important means of clinical screening and dynamic monitoring, and has accumulated a large amount of monitoring data.

[0003] Currently, health management methods based on transcutaneous bilirubin measurement data typically use the transcutaneous bilirubin value directly as an estimate of bilirubin levels, or convert it into a predicted serum bilirubin value using a fixed linear correction formula. However, the relationship between transcutaneous bilirubin values ​​and actual serum bilirubin levels is complexly influenced by various physiological factors, exhibiting significant nonlinear and time-varying characteristics. For example, changes in neonatal hemoglobin concentration directly affect the optical properties of skin tissue, leading to a systematic bias in transcutaneous bilirubin values, and this bias changes dynamically with the newborn's age. Existing methods fail to adequately consider the dynamic differences in the reliability of transcutaneous bilirubin signals under different physiological states, especially lacking mechanisms to quantify and compensate for "signal drift" caused by specific pathophysiological conditions such as anemia. This can lead to biases in trend judgments based on transcutaneous bilirubin values ​​in individuals whose physiological states change. Summary of the Invention

[0004] The purpose of this invention is to provide a big data-based neonatal jaundice health management system to solve the following technical problems:

[0005] Current methods for managing neonatal jaundice generally treat the transcutaneous jaundice measurement value and serum bilirubin level as a simple linear relationship, ignoring the nonlinear characteristics of the relationship's dynamic evolution with age, as well as the systematic drift of the transcutaneous jaundice measurement signal caused by changes in the individual physiological state of the newborn.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A big data-based neonatal jaundice health management system includes:

[0008] The data acquisition module is used to collect transcutaneous jaundice values, hemoglobin concentrations, age, and weight data of individual newborns, and to establish a dataset of individual newborn data.

[0009] The physiological parameter calibration module is used to calculate individualized drift indicators based on the hemoglobin concentration and age data through staged mapping and dynamic interpolation.

[0010] The dynamic prediction and analysis module is used to input transcutaneous bilirubin values ​​and individualized drift indicators into the dynamic correction model. The dynamic correction model performs multi-level nonlinear transformation and generates serum bilirubin prediction values ​​through feature space mapping.

[0011] The time series data processing module is used to obtain the predicted values ​​of serum bilirubin at the current time point and historical time points to form time series data, smooth the time series data and calculate the changes between adjacent time points;

[0012] The trend analysis and detection module is used to construct trend indicators based on changes at adjacent time points, and generate serum bilirubin change trend data by combining trend slope and change point detection.

[0013] The intelligent decision-making module is used to set a threshold level for the predicted serum bilirubin value based on age and weight data, compare the current predicted serum bilirubin value with the threshold level, and generate a health management plan by combining the trend direction change points in the serum bilirubin change trend data.

[0014] As a further aspect of the present invention: the specific process for calculating the individualized drift index in the physiological parameter calibration module is as follows:

[0015] Establish a daily age stage classification standard that includes the neonatal initial stage, physiological adaptation stage and metabolic stability stage, and establish a correspondence table between hemoglobin concentration and baseline drift for each daily age stage. Different functional relationships are used for the correspondence tables of different daily age stages.

[0016] The newborn's age is determined by identifying the age stage. A baseline drift is calculated using linear interpolation based on the current hemoglobin concentration in the corresponding relationship table for that age stage. Within the determined age stage, a stage progress coefficient is calculated based on the progress of the newborn's age relative to the stage's starting age. This coefficient is normalized using an S-shaped function. The baseline drift and stage progress coefficient are then weighted and fused, with the weights dynamically adjusted according to the age stage to generate the final output value of the individualized drift index.

[0017] As a further aspect of the present invention: in the dynamic prediction and analysis module, the dynamic correction model performs multi-layer nonlinear transformations, and the specific process of generating serum bilirubin prediction values ​​through feature space mapping is as follows:

[0018] The dynamic correction model includes a physiological state perception network and a multimodal feature interaction unit. The physiological state perception network receives hemoglobin concentration and age data and outputs a physiological state code. The multimodal feature interaction unit performs cross-attention calculation on the transcutaneous jaundice value and the physiological state code to generate an attention-weighted feature representation.

[0019] Based on attention-weighted feature representation, an individualized drift index is introduced as a bias term to adjust the offset of feature distribution; the temporal feature changes are processed through a gated recurrent unit to capture the dynamic evolution of the relationship between transcutaneous bilirubin values ​​and serum bilirubin; finally, a multilayer perceptron is used to map the temporal features to predicted serum bilirubin values.

[0020] As a further aspect of the present invention: the training process of the dynamic correction model includes:

[0021] A balanced training set containing normal physiological state samples and abnormal physiological state samples was constructed. The abnormal physiological state samples included data from anemic newborns. A multi-task learning framework was adopted, with the main task predicting serum bilirubin values ​​and the auxiliary task identifying physiological state categories.

[0022] During training, a weighted sum of the main task loss and auxiliary task loss is calculated, and the network parameters are updated through gradient backpropagation. During training, the weight coefficient of the auxiliary task loss is linearly decreased as the number of training epochs increases, while the weight coefficient of the main task loss is linearly increased. Adversarial training samples are introduced to simulate the nonlinear relationship between percutaneous jaundice values ​​and serum bilirubin values ​​under different physiological states. The predictive consistency of the model on different physiological state subsets is evaluated on the validation set, and model parameters that are stable under various physiological states are selected.

[0023] As a further aspect of the present invention: the specific process of smoothing the time series data and calculating the changes between adjacent time points in the time series data processing module is as follows:

[0024] A multi-scale sliding window process was applied to the time series data of predicted serum bilirubin values. The multi-scale sliding window includes short-term, medium-term and long-term windows. The moving average of the predicted serum bilirubin values ​​was calculated in different scale windows to generate smoothed time series data.

[0025] Within a short-term window, the first-order difference sequence of predicted serum bilirubin values ​​at adjacent time points is calculated to obtain the instantaneous change. Within a medium-term window, the slope of change between adjacent smoothing points is calculated to obtain the medium-term change. Within a long-term window, the magnitude of change in the trend baseline is calculated to obtain the long-term change. The instantaneous, medium-term, and long-term changes are weighted according to the window scale, with the weighting coefficients being positively correlated with the window length. The changes at each scale are combined to generate a complete set of changes at adjacent time points.

[0026] As a further aspect of the present invention: the specific process for generating serum bilirubin change trend data in the trend analysis and detection module is as follows:

[0027] A multi-dimensional trend feature vector is constructed based on the set of changes at adjacent time points. The multi-dimensional trend feature vector includes the mean, variance, autocorrelation coefficient, and skewness statistical characteristics of the changes. Cluster analysis is used to divide the multi-dimensional trend feature vector into different trend pattern categories, including stable pattern, upward pattern, and downward pattern.

[0028] For each trend pattern category, a representative trend slope is calculated, which is obtained by weighted averaging of all samples within that category. A mutation detection algorithm is applied to the time series data, which identifies points of change in trend direction by analyzing abnormal fluctuations in the distribution of change. The trend pattern categories, representative trend slopes, and points of change in trend direction are integrated to output structured serum bilirubin change trend data.

[0029] As a further aspect of the present invention: the specific process of generating a health management plan in the intelligent decision-making module is as follows:

[0030] A threshold rule base based on age segments and weight classification is constructed, which stores the serum bilirubin threshold levels corresponding to different age segments and weight ranges; a trend change pattern library is established, which includes feature descriptions of upward trend patterns, downward trend patterns and stationary trend patterns.

[0031] The basic risk level is determined by matching the current predicted serum bilirubin value with the threshold levels in the threshold rule base; the similarity between the serum bilirubin change trend data and the pattern features in the trend change pattern base is calculated to identify the trend change type; a preliminary management plan is generated based on the basic risk level and trend change type; the preliminary management plan is then personalized by combining the newborn's feeding method and hemoglobin concentration data to form the final health management plan.

[0032] As a further aspect of the present invention: the health management solution specifically includes:

[0033] Design a management plan framework that includes home monitoring sub-plans, outpatient follow-up sub-plans, and inpatient treatment sub-plans. Select the corresponding management plan framework level according to the basic risk level, and adjust the monitoring frequency and follow-up time configuration in the management plan framework based on the trend change type.

[0034] When an upward trend pattern is identified, the monitoring frequency and re-examination items are increased in the management plan framework; when the hemoglobin concentration data is lower than the lower limit of the normal range, special monitoring requirements related to anemia are added to the management plan; the personalized adjusted management plan framework content is combined with the newborn's basic information to generate a structured health management plan document.

[0035] The beneficial effects of this invention are:

[0036] This invention achieves dynamic quantification and compensation for transcutaneous jaundice signal drift caused by factors such as hemoglobin concentration by constructing an individualized drift index based on age stages. Employing a dynamic correction model that integrates physiological state perception and multimodal feature interaction, it effectively captures the nonlinear mapping relationship between transcutaneous jaundice values ​​and serum bilirubin concentration as they evolve with age, significantly improving the accuracy of predicted serum bilirubin values. Furthermore, through multi-scale time-series analysis, it accurately identifies trend change points and, combined with a threshold rule base and trend pattern base based on clinical guidelines, generates a personalized health management plan that sensitively reflects the dynamic changes in the newborn's physiological state. This method fundamentally overcomes the trend misjudgment problem caused by neglecting the time-varying influence of physiological state in existing technologies, enabling early and accurate identification and intervention decision support for neonatal jaundice risk. Attached Figure Description

[0037] The invention will now be further described with reference to the accompanying drawings.

[0038] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0039] 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.

[0040] Please see Figure 1 As shown, this invention is a big data-based neonatal jaundice health management system, comprising:

[0041] The data acquisition module is used to collect transcutaneous jaundice values, hemoglobin concentrations, age, and weight data of individual newborns, and to establish a dataset of individual newborn data.

[0042] The physiological parameter calibration module is used to calculate individualized drift indicators based on the hemoglobin concentration and age data through staged mapping and dynamic interpolation.

[0043] The dynamic prediction and analysis module is used to input transcutaneous bilirubin values ​​and individualized drift indicators into the dynamic correction model. The dynamic correction model performs multi-level nonlinear transformation and generates serum bilirubin prediction values ​​through feature space mapping.

[0044] The time series data processing module is used to obtain the predicted values ​​of serum bilirubin at the current time point and historical time points to form time series data, smooth the time series data and calculate the changes between adjacent time points;

[0045] The trend analysis and detection module is used to construct trend indicators based on changes at adjacent time points, and generate serum bilirubin change trend data by combining trend slope and change point detection.

[0046] The intelligent decision-making module is used to set a threshold level for the predicted serum bilirubin value based on age and weight data, compare the current predicted serum bilirubin value with the threshold level, and generate a health management plan by combining the trend direction change points in the serum bilirubin change trend data.

[0047] In a preferred embodiment of the present invention, the specific process for calculating the individualized drift index in the physiological parameter calibration module is as follows:

[0048] A age-based stage classification standard based on the developmental characteristics of newborns was established, dividing the newborn's growth process after birth into three key stages: the initial stage (0 to 7 days after birth), during which the newborn's various systems rapidly adjust; the physiological adaptation stage (8 to 28 days after birth), during which the newborn begins to adapt to the external environment; and the metabolic stabilization stage (29 days and above after birth), during which the newborn's physiological functions tend to stabilize.

[0049] Within each age stage, the module is configured with a dedicated table relating hemoglobin concentration to baseline drift. These tables use different mathematical functions to describe the correlation between hemoglobin concentration and baseline drift. For example, in the neonatal stage, where neonatal hemoglobin concentration fluctuates dramatically, a quadratic function table better captures its non-linear characteristics; while in the metabolically stable stage, a linear function table is sufficient to meet the accuracy requirements.

[0050] Upon receiving hemoglobin concentration and age data for a newborn, the module first determines the developmental stage based on the age. For example, a 15-day-old newborn falls within the physiological adaptation stage. The module then retrieves the corresponding hemoglobin concentration-reference drift table for that stage and calculates the reference drift using a linear interpolation algorithm. Specifically, if the newborn's hemoglobin concentration is 120 g / L, the closest reference points in the table are 115 g / L (reference drift of 0.8) and 125 g / L (reference drift of 1.2), respectively. The precise reference drift is then calculated through interpolation.

[0051] Next, the module calculates a stage progress coefficient to reflect the newborn's specific position within the current developmental stage. Taking the physiological adaptation stage as an example, this stage covers 21 days, from 8 to 28 days of age. If the newborn is currently 15 days old, then their stage progress is the 7th day after entering this stage. The module uses an S-shaped function to normalize this progress value, mapping it to the range of 0 to 1, thus obtaining the stage progress coefficient. This approach better reflects the non-linear characteristics of newborn development.

[0052] Finally, the module weights and fuses the baseline drift amount with the stage progress coefficient. The weighting varies across different developmental stages; for example, the baseline drift amount has a higher weight (0.7) in the neonatal stage, while the stage progress coefficient has a correspondingly higher weight in the metabolically stable stage. This dynamic weighting mechanism ensures that optimal individualized drift indices are obtained at different developmental stages. Through this series of precise calculations, the module ultimately outputs drift index values ​​that accurately reflect the individual physiological characteristics of newborns, providing a reliable calibration basis for subsequent serum bilirubin prediction.

[0053] In another preferred embodiment of the present invention, the specific process by which the dynamic prediction analysis module performs multi-layer nonlinear transformation and generates serum bilirubin prediction values ​​through feature space mapping is as follows:

[0054] The dynamic correction model adopts a dual-branch architecture design, which includes two core components: a physiological state perception network and a multimodal feature interaction unit.

[0055] The physiological state awareness network is responsible for processing the physiological characteristic data of newborns. This network receives hemoglobin concentration and age data as input and extracts features through three fully connected layers. The first fully connected layer maps the input 2D physiological data to a 64D feature space, using the ReLU activation function to introduce non-linear transformation capabilities. The second fully connected layer further increases the feature dimension to 128 dimensions, employing batch normalization to stabilize the training process. The third fully connected layer outputs a 32-dimensional physiological state code, which comprehensively represents the newborn's physiological development and blood characteristics.

[0056] The multimodal feature interaction unit is the core innovation of this model, responsible for fusing transcutaneous jaundice values ​​with physiological state encoding. This unit first converts the transcutaneous jaundice values ​​into a 64-dimensional feature vector through an embedding layer, and then performs cross-attention calculation with the physiological state encoding. In the cross-attention mechanism, the transcutaneous jaundice features serve as the query vector, and the physiological state encoding serves as the key-value vector. Attention weights are obtained through scaling dot product attention. For example, when processing a newborn sample with low hemoglobin concentration, the attention mechanism automatically increases the weight of transcutaneous jaundice values ​​associated with anemia features and decreases the weight of values ​​that may be significantly affected by anemia. This design allows the model to dynamically adjust its reliance on transcutaneous jaundice values ​​based on individual physiological state.

[0057] After obtaining the attention-weighted feature representation, the model introduces an individualized drift metric as a bias term to correct the feature distribution. Specifically, the 32-dimensional individualized drift metric is mapped to the same dimension as the attention features through a fully connected layer, and then element-wise added to the attention features. This operation can effectively compensate for feature distribution shifts caused by individual physiological differences; for example, for anemic newborns, this bias term will adjust the feature distribution in the correction direction.

[0058] To capture the dynamic relationship between transcutaneous bilirubin levels and serum bilirubin over time, the model employs a gated recurrent unit (GRU) to process temporal features. The GRU unit contains two core components: an update gate and a reset gate, enabling it to effectively learn long-term dependencies. In practical applications, the model receives transcutaneous bilirubin monitoring data for seven consecutive days, and the hidden state dimension of the GRU unit is set to 64 dimensions. Through GRU processing, the model can identify trends in transcutaneous bilirubin levels, such as a continuous upward trend or a fluctuating downward trend. These temporal patterns are valuable for accurately predicting serum bilirubin levels.

[0059] Finally, the model uses a three-layer multilayer perceptron to map the temporal features output by the GRU units to the final predicted serum bilirubin values. The first perceptron layer contains 128 neurons and uses the ReLU activation function; the second perceptron layer contains 64 neurons, also using the ReLU activation function; the third layer is the output layer, containing one neuron that uses a linear activation function to output the predicted serum bilirubin value. This design ensures that the model can learn complex nonlinear relationships while also outputting continuous predicted values.

[0060] For model training, we constructed a balanced training set containing samples from various physiological states. The training data came from monitoring records of 5,000 newborns from multiple medical centers, with 70% of the samples showing normal physiological states and 30% showing abnormal physiological states. The abnormal samples specifically included data from anemic newborns, whose hemoglobin concentrations were generally below the normal range. To ensure the model's generalization ability, the training set also included representative samples from different age groups and weight ranges.

[0061] The training process employs a multi-task learning framework. The primary task predicts serum bilirubin levels, using mean squared error as the loss function. The auxiliary task identifies physiological state categories, using cross-entropy loss. Initially, the weight coefficient for the auxiliary task loss is set to 0.6, and the weight for the primary task loss is 0.4. This design aims to allow the model to first learn to recognize physiological state features. As the number of training epochs increases, a linear adjustment strategy is adopted: every 10 training epochs, the auxiliary task loss weight decreases by 0.05, and the primary task loss weight increases by 0.05. When 60 training epochs are reached, the auxiliary task loss weight decreases to 0.3, and the primary task loss weight increases to 0.7. This dynamic weight adjustment mechanism ensures that the model gradually strengthens its learning of the serum bilirubin prediction task while mastering the ability to recognize physiological states.

[0062] To improve the model's predictive robustness under abnormal physiological conditions, adversarial training samples were introduced during training. These samples were generated by selectively perturbing the original training data, for example, simulating the nonlinear relationship between transcutaneous jaundice values ​​and serum bilirubin levels under different degrees of anemia. Specifically, for neonatal samples with low hemoglobin concentrations, we added Gaussian noise to simulate abnormal fluctuations in transcutaneous jaundice values ​​while keeping serum bilirubin levels constant, forcing the model to learn to maintain accurate predictions even under such perturbation conditions.

[0063] During the model selection phase, we evaluated the predictive consistency of the models across different physiological state subsets on an independent validation set. The validation set contained data from 1000 newborns, categorized into normal, mild anemia, and severe anemia groups based on their physiological state. We calculated the prediction error distribution of the models across each group and selected model parameters that remained stable across all three groups. Specific evaluation metrics included the mean absolute error between predicted and actual values, the standard deviation of the error, and the distribution of prediction bias. Only models that met the predetermined criteria across all groups were ultimately adopted.

[0064] In a practical application example, consider a 10-day-old newborn with a hemoglobin concentration of 110 g / L. This newborn is in the physiological adaptation stage, and their hemoglobin concentration is slightly below the normal range. The model first generates a physiological state code through a physiological state perception network, highlighting the low hemoglobin level. In the multimodal feature interaction unit, the transcutaneous bilirubin level is 12.5 mg / dL. After cross-attention calculation, the model appropriately reduces the weight of the transcutaneous bilirubin level because the physiological state code indicates anemia. The introduction of an individualized drift index further corrects the feature distribution, compensating for the potential underestimation of transcutaneous bilirubin levels due to anemia. The GRU unit analyzes the transcutaneous bilirubin data sequence over the past 7 days, identifying a slow upward trend in the value. Finally, the multilayer perceptron integrates all this information and outputs a predicted serum bilirubin value of 14.2 mg / dL. Subsequent serum testing confirms the accuracy of this prediction, with a true value of 14.0 mg / dL.

[0065] The design and training method of this dynamic correction model ensures accurate serum bilirubin prediction results under different physiological states. Especially when dealing with newborns in abnormal physiological states, the model effectively overcomes the prediction bias caused by traditional methods neglecting the influence of physiological factors through multimodal feature interaction and individualized drift correction. The model's innovation lies in its organic combination of physiological state perception, multimodal feature fusion, and temporal dynamic analysis, providing reliable technical support for the accurate monitoring of neonatal jaundice.

[0066] In another preferred embodiment of the present invention, the specific process of smoothing the time series data and calculating the changes between adjacent time points in the time series data processing module is as follows:

[0067] A multi-scale sliding window processing technique was employed to capture comprehensive characteristics of serum bilirubin changes through three windows at different time scales. The short-term window was set to 3 days in length, mainly used to capture recent instantaneous fluctuations; the medium-term window was set to 7 days in length, used to observe weekly variation patterns; and the long-term window was set to 14 days in length, used to analyze more macroscopic trend characteristics.

[0068] In actual processing, the module first preprocesses the input serum bilirubin predicted value time series. Taking a set of monitoring data for 15 consecutive days as an example, the module applies three sliding windows of different scales simultaneously. For the short-term window, the module calculates the moving average of every three consecutive data points and smooths it using equal weighting coefficients. For example, when processing the data for day 5, the short-term window uses the values ​​from days 3, 4, and 5 to calculate the moving average. This processing method effectively filters out random fluctuations caused by daily measurement errors while retaining the true short-term variation characteristics.

[0069] The intermediate window employs a similar moving average method, but with the window length extended to 7 days. When processing data on day 10, the intermediate window uses data from days 4 to 10 to calculate a 7-day moving average. Due to the increased window length, the smoothed data curve in the intermediate window is flatter, better reflecting the overall trend of serum bilirubin changes. Notably, the sliding step size of the intermediate window remains consistent with that of the short-term window, ensuring a one-to-one correspondence between data points at different scales.

[0070] The long-term window is processed using a 14-day length, which takes into account the complete cycle of neonatal jaundice development. When processing the data on day 14, the long-term window calculates a moving average of the complete sequence from day 1 to day 14. This long-term smoothing process eliminates the influence of short-term fluctuations and medium-term variability, presenting the most basic trend baseline. Smoothing at three scales is performed simultaneously, generating three complementary smoothed sequences, laying the foundation for subsequent variability analysis.

[0071] After completing multi-scale smoothing, the module begins calculating the change indices at each scale. At the short-term scale, the module calculates the first-order difference sequence between adjacent time points. Specifically, it calculates the difference between each data point and the previous data point. For example, if the predicted serum bilirubin values ​​for three consecutive days are 12.1, 12.3, and 12.6 mg / dL, the corresponding instantaneous changes are 0.2 and 0.3 mg / dL. This instantaneous change can accurately capture the diurnal fluctuations in serum bilirubin.

[0072] The intermediate-term change is calculated based on a 7-day moving average series. The module calculates the slope of change between adjacent smoothed points and uses the least squares method to fit the trend of change of the three most recent smoothed points. For example, if three consecutive 7-day moving averages are 11.8, 11.9, and 12.2 mg / dL, the slope of change calculated by linear regression is 0.2 mg / dL per day. This intermediate-term change reflects the persistence and directionality of serum bilirubin changes.

[0073] Calculating long-term changes is more complex, requiring analysis of the trend baseline changes in the 14-day moving average series. The module employs a piecewise linear fitting method, dividing the long-term series into several equal-length segments and calculating the difference in the means of each segment. For example, the 14-day long-term series can be divided into two 7-day segments, calculating the difference between the mean of the latter 7 days and the mean of the former 7 days, and then dividing by the time span to obtain the long-term change. This method can identify fundamental shifts in the trend baseline.

[0074] After obtaining the changes at three scales, the module performs intelligent weight allocation. The weight coefficients are positively correlated with the window length, with long-term changes having the highest weight (0.5), medium-term changes having the second highest weight (0.35), and instantaneous changes having the lowest weight (0.15). This weight allocation strategy ensures that long-term trends have a decisive influence when assessing overall change, while short-term fluctuations only serve as a supplementary reference. The changes at each scale are then weighted and summed to generate a complete set of changes at adjacent time points.

[0075] In another preferred embodiment of the present invention, the specific process of generating serum bilirubin change trend data in the trend analysis and detection module is as follows:

[0076] First, a multi-dimensional trend feature vector is constructed to extract features from the set of changes from four statistical dimensions. The mean reflects the average level of the changes, the variance characterizes the degree of fluctuation, the autocorrelation coefficient reveals the time dependence of the changes, and the skewness statistic describes the asymmetry of the distribution of the changes. These four statistics together constitute a four-dimensional feature vector, comprehensively characterizing the dynamic characteristics of serum bilirubin changes.

[0077] Next, the module uses cluster analysis to automatically classify multi-dimensional trend feature vectors. Using the K-means clustering algorithm, all change sequences are divided into three typical trend pattern categories. The first category is the stable pattern, characterized by a near-zero mean change, small variance, and low autocorrelation coefficient. The second category is the rising pattern, with a positive mean change, moderate variance, and high autocorrelation coefficient. The third category is the falling pattern, with a negative mean change and other characteristics similar to the rising pattern. Through cluster analysis, the system can automatically identify the trend type of the current serum bilirubin change.

[0078] For each identified trend pattern category, the module calculates its representative trend slope. This calculation uses a weighted average method, taking into account the membership degree of each sample in the cluster. For example, in the rising trend category, samples whose feature vectors are closer to the cluster centers are given higher weights. The representative trend slope calculated using this weighted method can more accurately reflect the typical rate of change for that category.

[0079] Change detection is a crucial step in trend analysis. This module employs a Bayesian inference-based change detection algorithm to identify points of change in trend direction by analyzing anomalous fluctuations in the distribution of changes. The algorithm uses a sliding detection window to continuously monitor the statistical characteristics of the change series. When a significant change in the mean or variance of the change is detected, the algorithm marks that point as a potential change point. For example, when the change series suddenly transitions from a stable low-fluctuation state to a high-fluctuation state, or abruptly changes from a positive mean to a negative mean, the system identifies a trend inflection point.

[0080] To verify the reliability of mutation points, the module employs a multi-level confirmation mechanism. After detecting a suspected mutation point at a short-term scale, it further examines whether coordinated changes occur at medium- and long-term scales. Only when significant statistical changes are observed at multiple scales is the mutation point finally confirmed. This multi-scale verification mechanism effectively reduces the probability of false alarms.

[0081] Finally, the module integrates trend pattern categories, representative trend slopes, and mutation detection results to output structured serum bilirubin change trend data. This data not only includes a description of the current trend state but also provides quantitative indicators of trend strength and early warning information for trend reversals. For example, the system might output a trend analysis result like this: currently in an upward trend, with a representative slope of 0.15 mg / dL / day, no trend reversal detected in the last 24 hours, and a trend stability index of 0.8.

[0082] In practical applications, suppose a newborn's predicted serum bilirubin levels over 15 consecutive days show a stable level for the first 5 days, a slow increase over the middle 5 days, and a rapid increase over the last 5 days. The time-series data processing module captures this gradual change through multi-scale analysis. The short-term window first detects the increase in daily variation, the medium-term window identifies the shift in the weekly trend, and the long-term window confirms the actual shift in the trend baseline. The trend analysis and detection module then categorizes this change as an accelerating upward pattern, calculates the specific trend slope, and identifies two key trend inflection points on days 6 and 11. This structured trend data provides a precise basis for subsequent health management decisions.

[0083] In another preferred embodiment of the present invention, the specific process of generating a health management plan in the intelligent decision-making module is as follows:

[0084] A comprehensive decision support knowledge base was constructed, with a threshold rule base established based on clinical guidelines and expert consensus. Neonatal age is divided into five key stages: days 1-3 (initial stage), days 4-7 (peak stage), days 8-14 (transition stage), days 15-21 (regression stage), and days 22-28 (stabilization stage). Each age stage is further subdivided according to weight range, classifying neonatal weight into low weight group (less than 2500 grams), normal weight group (2500-4000 grams), and high weight group (greater than 4000 grams). For each age-weight combination, the threshold rule base defines serum bilirubin threshold ranges corresponding to four risk levels.

[0085] Taking a 5-day-old newborn weighing 3000 grams as an example, the system will locate the threshold rules corresponding to the peak age and normal weight group. These rules set a low-risk threshold of less than 12 mg / dL, a medium-risk threshold of 12 to 15 mg / dL, a high-risk threshold of 15 to 18 mg / dL, and an extremely high-risk threshold of greater than 18 mg / dL. These threshold data are derived from statistical analysis of a large amount of clinical research data and have been validated by multiple medical centers.

[0086] The trend change pattern library is built based on deep learning of historical cases. This library defines detailed characteristic descriptions for three basic trend patterns. An upward trend pattern is characterized by three consecutive monitoring periods of positive change, a mean change greater than 0.2 mg / dL / day, and an autocorrelation coefficient greater than 0.6. A downward trend pattern is characterized by continuous negative change, with a mean change less than -0.1 mg / dL / day. A stationary trend pattern is characterized by changes fluctuating between ±0.1 mg / dL / day with small variance. Each trend pattern is further subdivided into different intensity levels, such as rapid rise and slow rise.

[0087] In the actual decision-making process, the module first performs a basic risk assessment. The system obtains the current predicted serum bilirubin value. For example, if a 6-day-old newborn has a bilirubin level of 13.5 mg / dL and a weight of 3200g, the system determines that the newborn is in the medium-risk category by querying the threshold rule base. This assessment process considers not only absolute values ​​but also the normal range for newborns of the same age to ensure the clinical rationality of the assessment results.

[0088] Next, a trend risk assessment is performed. The module calculates the similarity between the serum bilirubin change trend data and a trend change pattern library. Assuming that the newborn's monitoring data over the past three days shows that their serum bilirubin level increased from 11.8 to 12.5 and then to 13.5 mg / dL, with a mean change of 0.85 mg / dL per day and an autocorrelation coefficient of 0.72, by comparing it with the pattern library, the system identifies this as a typical upward trend pattern, and a moderate increase.

[0089] Based on the basic risk level and trend change type, the module generates a preliminary management plan. For the medium-risk level combined with an upward trend, the system selects an outpatient follow-up plan as the basic framework. This framework is preset to a monitoring frequency of once every two days, with follow-up items including transcutaneous bilirubin measurement and weight monitoring. However, due to the identification of an upward trend, the system automatically adjusts the monitoring frequency to once daily and adds serum bilirubin testing.

[0090] The personalized adjustment phase takes into account the newborn's feeding method and hemoglobin concentration data. If the newborn is breastfed, the system will add feeding guidance to the plan, including recommended feeding frequency and monitoring of feeding amount. When the hemoglobin concentration is below 110 g / L, the system will determine that there is a risk of anemia and add special monitoring requirements to the management plan, including recommending a complete blood count, increasing nutritional support guidance, and further increasing the monitoring frequency to twice daily.

[0091] The management plan framework adopts a three-tiered design. The home monitoring sub-plan is suitable for low-risk situations and includes basic items such as daily transcutaneous jaundice measurement, weight monitoring, and feeding records. The outpatient follow-up sub-plan is for medium-risk situations, adding professional assessments and serum testing twice a week to the home monitoring plan. The inpatient treatment sub-plan is for high-risk situations and includes comprehensive medical measures such as continuous monitoring, phototherapy intervention, and specialist consultations.

[0092] When the system identifies an upward trend, it automatically adjusts the monitoring plan based on the rate of increase. For a slow increase, the monitoring frequency is adjusted to once every 12 hours; for a rapid increase, monitoring is required once every 6 hours, with the addition of urine output and body temperature monitoring. These adjustments are based on extensive clinical data to ensure timely detection of changes in the patient's condition.

[0093] For newborns with low hemoglobin levels, the system will initiate a special monitoring program. In addition to routine bilirubin monitoring, it will also include observation of skin pallor, monitoring of activity levels, and assessment of feeding tolerance. When the hemoglobin concentration is below 100 g / L, the system will forcibly upgrade the management plan level, and even if the serum bilirubin value does not reach the high-risk threshold, it will recommend outpatient follow-up or inpatient observation.

[0094] During the solution generation process, the system verifies the rationality of each suggestion in real time. For example, when suggesting increasing the monitoring frequency, the feasibility of that frequency will be evaluated simultaneously to avoid providing unrealistic solutions. For suggestions requiring medical intervention, the system checks whether they are compatible with the current availability of medical resources to ensure the operability of the solution.

[0095] The final health management plan document is output in a structured format and includes the following main parts: a basic information summary listing key information such as the newborn's age, weight, and current serum bilirubin level; risk assessment results clearly indicating the baseline risk level and trend risk type; specific management measures detailing monitoring frequency, follow-up examination items, and key points of home care; special precautions providing targeted suggestions based on individual circumstances; and a follow-up plan specifying the next assessment time and early warning indicators. All content is expressed using clear and unambiguous medical terminology to avoid ambiguity.

[0096] To illustrate with a real-world example, consider an 8-day-old newborn weighing 2800 grams with a predicted serum bilirubin level of 14.2 mg / dL, classified as high-risk based on baseline risk assessment. Trend analysis showed a rapid increase in bilirubin from 13.1 mg / dL to 14.2 mg / dL over the past two days. This newborn was mixed-fed, with a hemoglobin concentration of 105 g / L. The system first generated a sub-protocol framework for inpatient treatment based on the high-risk level. Then, based on the rapid upward trend, the monitoring frequency was increased to once every 4 hours, and blood gas analysis was added. Considering the anemia, iron supplementation recommendations and red blood cell parameter monitoring were incorporated into the protocol. The final output health management protocol document included detailed inpatient observation guidelines, indications for phototherapy, and nutritional support plans, providing comprehensive support for clinical decision-making.

[0097] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A big data based neonatal jaundice health management system, characterized in that, The method comprises the following steps: A data acquisition module is used to collect transcutaneous jaundice values, hemoglobin concentration, age and weight data of a newborn individual, and establish a newborn individual data set; A physiological parameter calibration module is used to calculate an individualized drift index based on the hemoglobin concentration and age data through stage-by-stage mapping and dynamic interpolation; A dynamic prediction analysis module is used to input the transcutaneous jaundice values and the individualized drift index into a dynamic correction model, the dynamic correction model performs multi-layer nonlinear transformation, and generates serum bilirubin prediction values through feature space mapping; A time series data processing module is used to obtain serum bilirubin prediction values at the current time point and historical time points to form time series data, and perform smoothing processing on the time series data and calculate the change amount of adjacent time points; A trend analysis detection module is used to construct a trend index based on the change amount of adjacent time points, and generate serum bilirubin change trend data by combining the trend slope and change point detection; An intelligent decision-making module is used to set threshold levels of serum bilirubin prediction values according to the age and weight data, compare the current serum bilirubin prediction values with the threshold levels, and generate a health management scheme by combining the trend direction change point in the serum bilirubin change trend data. In the dynamic prediction analysis module, the specific process of the dynamic correction model performing multi-layer nonlinear transformation and generating serum bilirubin prediction values through feature space mapping is as follows: The dynamic correction model includes a physiological state perception network and a multi-modal feature interaction unit. The physiological state perception network receives the hemoglobin concentration and age data and outputs physiological state encoding. The multi-modal feature interaction unit performs cross-attention calculation on the transcutaneous jaundice values and the physiological state encoding to generate attention-weighted feature representation. On the basis of the attention-weighted feature representation, the individualized drift index is introduced as a bias term to adjust the offset of the feature distribution. The gate recurrent unit is used to process the temporal feature changes and capture the dynamic evolution law of the relationship between the transcutaneous jaundice values and the serum bilirubin. Finally, the multi-layer perception machine is used to map the time series features to the serum bilirubin prediction values.

2. The big data based neonatal jaundice health management system as claimed in claim 1, wherein, In the physiological parameter calibration module, the specific process of calculating the individualized drift index is as follows: Establish a day age stage division standard including the newborn initial stage, physiological adaptation stage and metabolic stability stage, and establish a corresponding relationship table of hemoglobin concentration and reference drift for each day age stage. The corresponding relationship tables of different day age stages adopt different function relationships. According to the day age of the current newborn, the day age stage to which the newborn belongs is determined, and the reference drift is calculated by linear interpolation based on the current hemoglobin concentration value in the corresponding relationship table of the day age stage. In the determined day age stage, the stage progress coefficient is calculated according to the progress of the day age value relative to the starting day age of the stage, the stage progress coefficient is normalized by using the S-shaped function, the reference drift and the stage progress coefficient are weighted and fused, the weighted weight is dynamically adjusted according to the day age stage, and the final output value of the individualized drift index is generated.

3. The big data based neonatal jaundice health management system as claimed in claim 1, wherein, The training process of the dynamic correction model comprises: A balanced training set containing normal physiological state samples and abnormal physiological state samples containing data of anemic newborns is constructed, and a multi-task learning framework is adopted, in which a main task is to predict serum bilirubin values, and an auxiliary task is to identify physiological state categories; In the training process, the weighted sum of the main task loss and the auxiliary task loss is calculated, and the network parameters are updated through gradient back propagation. During the training process, the weight coefficient of the auxiliary task loss is linearly reduced, and the weight coefficient of the main task loss is linearly increased according to the increase of the training cycle number. Anti-training samples are introduced to simulate the nonlinear relationship between the transcutaneous jaundice value and the serum bilirubin value under different physiological states. The prediction consistency of the model on different physiological state subsets is evaluated on the validation set, and the model parameters that are stable under each physiological state are selected.

4. The big data based neonatal jaundice health management system as claimed in claim 1, wherein, In the time series data processing module, the specific process of smoothing the time series data and calculating the change amount between adjacent time points is as follows: Multi-scale sliding window processing is applied to the time series data of the predicted serum bilirubin values, and the multi-scale sliding window includes short-term window, medium-term window and long-term window. The moving average of the predicted serum bilirubin values is calculated in different scale windows to generate smoothed time series data. The first-order difference sequence of the predicted serum bilirubin values between adjacent time points is calculated in the short-term window to obtain the instantaneous change amount. The change slope between adjacent smoothed points is calculated in the medium-term window to obtain the medium-term change amount. The change amplitude of the trend baseline is calculated in the long-term window to obtain the long-term change amount. The instantaneous change amount, medium-term change amount and long-term change amount are weighted and distributed according to the window scale, and the weight coefficient is positively correlated with the window length. The complete adjacent time point change amount set is generated by integrating the change amounts of each scale.

5. The big data based neonatal jaundice health management system as claimed in claim 1, wherein, In the trend analysis detection module, the specific process of generating serum bilirubin trend data is as follows: A multi-dimensional trend feature vector is constructed based on the adjacent time point change amount set, which includes the mean, variance, autocorrelation coefficient and skewness statistical features of the change amount. A clustering analysis method is used to divide the multi-dimensional trend feature vector into different trend pattern categories, including stable pattern, rising pattern and falling pattern. The representative trend slope of each trend pattern category is calculated by weighted average of all samples in the category. A mutation detection algorithm is applied to the time series data, which identifies the trend direction change point by analyzing the abnormal fluctuations of the change amount distribution. The trend pattern category, representative trend slope and trend direction change point are integrated to output structured serum bilirubin trend data.

6. The big data based neonatal jaundice health management system as claimed in claim 1, wherein, In the intelligent decision-making module, the specific process of generating a health management plan is as follows: A threshold rule library based on age segmentation and weight classification is constructed, which stores the serum bilirubin threshold levels corresponding to different age segments and weight ranges. A trend change pattern library is established, which contains the feature descriptions of rising trend pattern, falling trend pattern and stable trend pattern. The current serum bilirubin prediction value is matched with the threshold levels in the threshold rule library to determine the basic risk level. The serum bilirubin change trend data is subjected to similarity calculation with the mode characteristics in the trend change mode library, and the trend change type is identified; A preliminary management scheme is generated based on the basic risk level and the trend change type; The preliminary management scheme is personalized adjusted in combination with the feeding mode and the hemoglobin concentration data of the newborn, and a final health management scheme is formed.

7. The big data based neonatal jaundice health management system as claimed in claim 6, wherein, The health management scheme specifically includes: A management scheme framework including a family monitoring sub-scheme, an outpatient review sub-scheme and an inpatient treatment sub-scheme is designed, a corresponding management scheme framework level is selected according to the basic risk level, and the monitoring frequency and review time configuration in the management scheme framework are adjusted based on the trend change type; When the upward trend mode is identified, the monitoring frequency and review items in the management scheme framework are increased; when the hemoglobin concentration data is lower than the lower limit of the normal range, the special monitoring requirements related to anemia are supplemented in the management scheme; and the content of the personalized adjusted management scheme framework is combined with the basic information of the newborn to generate a structured health management scheme document.

Images