Neonatal jaundice four-level early warning system and method based on multi-modal analysis

Abstract

The invention relates to the technical field of medical data mining, in particular to a neonatal jaundice four-level early warning system and method based on multi-modal analysis. The method comprises the steps of firstly obtaining feeding record data; further obtaining a bilirubin increment-day age-growth increment change curved surface according to the growth increment of the historical neonates of the same day age and the distribution of the corresponding bilirubin increment; historical predicted bilirubin is further obtained based on the change curved surface; further obtaining network prediction bilirubin based on the pre-trained neural network; and finally, according to the difference between the historically predicted bilirubin and the network predicted bilirubin, in combination with the difference between the historically predicted feeding interval and the network predicted feeding interval, four-level early warning of jaundice is carried out. According to the scheme, the influence of breast feeding is considered, the physiological model and the neural network are fused, dynamic personalized prediction and multi-source comparison of bilirubin are realized, the problems of poor model generalization and early warning lag in traditional jaundice monitoring are solved, and the accuracy and timeliness of jaundice risk identification are improved.

Description

Technical Field

[0001] This invention relates to the field of medical data mining technology, specifically to a four-level early warning system and method for neonatal jaundice based on multimodal analysis. Background Technology

[0002] Neonatal jaundice is a common physiological phenomenon in newborns, mainly because their livers are not fully developed and have limited capacity to process bilirubin, making them prone to jaundice. Other major factors affecting bilirubin absorption and breakdown include breastfeeding. While breast milk is rich in antibodies, immune factors, and key nutrients, it also contains female hormones and β-glucuronidase, which can affect the newborn's breakdown and absorption of bilirubin, exacerbating changes in bilirubin levels and causing a physiological increase in bilirubin levels.

[0003] Most existing jaundice risk grading technologies rely solely on jaundice meters to identify bilirubin levels in newborns using images at specific wavelengths. This ignores physiological bilirubin changes caused by breastfeeding, reduces the utilization of multi-source information, and can easily lead to inaccurate jaundice risk predictions, resulting in either overestimation or underestimation of jaundice risk, and causing either over-intervention or under-intervention in neonatal jaundice. Summary of the Invention

[0004] To address the technical problem of inaccurate jaundice early warning caused by neglecting bilirubin changes due to breastfeeding in existing jaundice risk assessment methods, this invention aims to provide a four-level early warning system and method for neonatal jaundice based on multimodal analysis. The specific technical solution adopted is as follows: A four-level early warning method for neonatal jaundice based on multimodal analysis, the method comprising: Obtain the age, weight, bilirubin levels, bilirubin increase and time interval between two consecutive feedings for current and historical newborns; For each newborn, the growth increment of the newborn at each feeding is obtained based on the change in weight during adjacent feedings; based on the distribution of the growth increment and the corresponding bilirubin increment of historical newborns of the same age, the change surface of bilirubin increment-age-growth increment is obtained. Based on the growth increment during feeding within the current newborn's preset time-domain neighborhood, combined with the change surface and the current bilirubin index, historical predicted bilirubin is obtained; based on the time interval during feeding within the current newborn's preset time-domain neighborhood, historical predicted feeding interval is obtained; the sequence of the current newborn's bilirubin index is input into a pre-trained neural network to obtain network predicted bilirubin and network predicted feeding interval. Based on the difference between the historical predicted bilirubin and the network predicted bilirubin, and combined with the difference between the historical predicted feeding interval and the network predicted feeding interval, a four-level jaundice warning is issued.

[0005] Furthermore, the method for obtaining the changing surface includes: Select the age of each historical newborn as the target age, classify the growth increment by data value under the target age, and obtain the bilirubin increment corresponding to each growth increment based on the central tendency of all bilirubin increments corresponding to each growth increment. Curve fitting is performed on the growth increment and the corresponding bilirubin increment for all species at the target age to obtain the change curve of bilirubin increment-growth increment; surface fitting is performed on the change curves for all ages to obtain the change surface of bilirubin increment-age-growth increment.

[0006] Furthermore, the method for obtaining the historical predicted bilirubin includes: Based on the central tendency of the growth increments during feeding within a preset time-domain neighborhood, the predicted growth increment for the next feeding of the current newborn is obtained; the predicted growth increment and age of the current newborn are substituted into the change surface, and combined with the current bilirubin index, the historical predicted bilirubin is obtained.

[0007] Furthermore, the method for conducting a four-level jaundice early warning includes: Based on the ratio of the network-predicted feeding interval to the historical predicted feeding interval, the historical predicted bilirubin is corrected to obtain the corrected historical predicted bilirubin. Based on the degree to which the network-predicted bilirubin level is higher than the corrected historical bilirubin level, a jaundice risk coefficient is obtained; and a four-level jaundice warning is issued based on the jaundice risk coefficient.

[0008] Furthermore, the method for providing a four-level jaundice early warning system based on the jaundice risk coefficient includes: When the jaundice risk coefficient is within the preset low-risk threshold range, it is determined to be low-risk; when the jaundice risk coefficient is within the preset low-to-medium-risk threshold range, it is determined to be low-to-medium-risk; when the jaundice risk coefficient is within the preset high-to-medium-risk threshold range, it is determined to be high-to-medium-risk; when the jaundice risk coefficient is within the preset high-risk threshold range, it is determined to be high-risk.

[0009] Furthermore, the method for obtaining the historical predicted feeding interval includes: The average of the time intervals within the current newborn's preset time domain neighborhood is used as the historical predicted feeding interval.

[0010] Furthermore, the method for obtaining the network prediction feeding interval includes: Based on the central tendency of all time intervals in the current sequence of bilirubin levels in newborns, the network predicts feeding intervals.

[0011] Furthermore, the method for obtaining the growth increment includes: The increase in body weight at each feeding compared to the body weight at the previous feeding is taken as the growth increment at each feeding.

[0012] Furthermore, the pre-trained neural network is an LSTM network.

[0013] The present invention also proposes a four-level early warning system for neonatal jaundice based on multimodal analysis. The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the steps of the four-level early warning method for neonatal jaundice based on multimodal analysis.

[0014] The present invention has the following beneficial effects: This invention first acquires feeding records of current and historical newborns, and then analyzes the data. Next, it obtains the growth increment of each feeding, reflecting individual growth and development from the perspective of weight change. Further, based on the distribution of growth increment and corresponding bilirubin increment of historical newborns of the same age, it obtains a surface curve of bilirubin increment-age-growth increment, reflecting the statistical pattern of bilirubin changes under typical growth and development paths, facilitating subsequent prediction of current newborns' bilirubin levels from a historical data perspective. Based on this surface curve, it predicts the bilirubin increment under the current developmental trend using historical data, and then combines this with the current bilirubin level to obtain historical predicted bilirubin from a historical physiological model perspective, providing a historical reference. To eliminate the influence of different predicted feeding intervals in different prediction methods, it accurately compares the prediction results of the two methods, obtaining historical predicted feeding intervals and network predicted feeding intervals. Furthermore, leveraging the more individualized and real-time characteristics of neural network prediction, it obtains network predicted bilirubin based on a pre-trained neural network, providing individualized prediction results for current newborns. Finally, based on the difference between historical predicted bilirubin and network predicted bilirubin, and combined with the difference between historical predicted feeding intervals and network predicted feeding intervals, it provides a four-level jaundice warning system. This solution takes into account the impact of breastfeeding, integrates physiological models and neural networks, and enables personalized prediction and multi-source comparison of bilirubin dynamics. It solves the problems of poor model generalization and delayed early warning in traditional jaundice monitoring, and improves the accuracy and timeliness of jaundice risk identification. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a four-level early warning method for neonatal jaundice based on multimodal analysis, provided as an embodiment of the present invention; Figure 2 This is an example diagram of a surface curve showing the change in bilirubin increment-age-growth increment according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a four-level early warning system and method for neonatal jaundice based on multimodal analysis proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of a four-level early warning system and method for neonatal jaundice based on multimodal analysis provided by the present invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a four-level early warning method for neonatal jaundice based on multimodal analysis, provided by an embodiment of the present invention, specifically including: Step S1: Obtain the age, weight, bilirubin levels, bilirubin increment, and time interval between two consecutive feedings for the current newborn and historical newborns.

[0021] In this embodiment of the invention, in order to perform multimodal analysis, various feeding data of historical normal newborns are first collected. Since age and weight are important indicators of newborn growth and development, bilirubin and bilirubin increase are important indicators of jaundice risk, and feeding interval is the time change corresponding to changes in bilirubin and weight, it includes at least the age, weight, bilirubin index at each feeding, and the bilirubin increase and time interval between two adjacent feedings. At the same time, various feeding data of the current newborn are also obtained to facilitate subsequent comparative analysis of the current newborn with historical data.

[0022] Among them, bilirubin values ​​can be obtained with the help of a multispectral imaging instrument, mainly by measuring transcutaneous bilirubin (TCB) values ​​using a transcutaneous jaundice tester; the feeding time and bilirubin index data values ​​at each feeding are recorded to obtain the time interval and bilirubin increment between two adjacent feedings, and the feeding time needs to be accurate to at least the minute.

[0023] Step S2: For each newborn, obtain the growth increment of the newborn for each feeding based on the change in weight during adjacent feedings; obtain the change surface of bilirubin increment-age-growth increment based on the distribution of growth increment and corresponding bilirubin increment of historical newborns of the same age.

[0024] Considering that the main cause of jaundice caused by bilirubin is that the newborn's liver is not fully developed, and that the increase in bilirubin changes continuously as the newborn grows and develops, existing data shows that in exclusively breastfed newborns, jaundice appears one week after birth, peaks around two weeks, and can last until four to twelve weeks before subsiding. Therefore, the historical data for newborns is set as the feeding data for the first four weeks, which can be adjusted by the implementer.

[0025] Considering the individual differences in growth among newborns of the same age, we first obtain the growth increment of each feeding based on the weight change during adjacent feedings to reflect individual growth and development. Then, based on the distribution of growth increment and corresponding bilirubin increment of historical newborns of the same age, we obtain the bilirubin increment-age-growth increment change surface and construct the relationship between growth change and bilirubin change. This allows us to obtain a statistical pattern reflecting bilirubin changes under typical growth and development paths, thus providing a reference baseline for prediction and anomaly detection for the current individual and enabling more accurate early warning using multimodal data.

[0026] Preferably, in one embodiment of the present invention, the increase in body weight at each feeding compared to the body weight at the previous feeding is taken as the growth increment at each feeding.

[0027] As an example, the difference between the weight at each feeding and the weight at the previous feeding is used as the growth increment for each feeding.

[0028] The calculation of growth increments includes all historical newborns and current newborns.

[0029] In another embodiment of the present invention, considering that the rate of breast milk consumption can reflect changes in stomach capacity and indirectly reflect the growth of newborns, the amount of breast milk at each feeding can also be collected, and the breast milk can be put into a milk bag or bottle that can assess the feeding amount. Medical staff can then carefully feed the baby and record the amount of breast milk fed.

[0030] The ratio of the amount of breast milk at the i-th feeding to the time interval between the i-th and i+1-th feedings is taken as the breast milk consumption rate at the i-th feeding; the difference between the breast milk consumption rate at the i-th feeding and the breast milk consumption rate at the (i-1)-th feeding is taken as the first growth factor at the i-th feeding. The increase in body weight at each feeding compared to the body weight at the previous feeding is used as the second growth factor for each feeding. The first and second growth factors were weighted and summed with weights of 0.3 and 0.7, respectively, and the sum was used as the growth increment for each feeding.

[0031] Compared to analyzing only weight changes to obtain growth increments, adding the perspective of breast milk volume increases the number of angles for measuring growth increments, making the growth increments more accurate. However, it also increases the amount of data required and the difficulty of obtaining the data. Implementers can choose according to their own preferences.

[0032] Using the same method, the growth increment of each newborn at each feeding was obtained.

[0033] Preferably, in one embodiment of the present invention, the age of each historical newborn is selected as the target age and analyzed one by one. Considering that the growth increment of historical newborns of the same age is different, the growth increment is classified by data value at the target age. Considering that the central tendency of all bilirubin increments corresponding to a certain growth increment reflects the typical performance of bilirubin metabolism under similar growth conditions, the bilirubin increment corresponding to each growth increment is obtained based on the central tendency of all bilirubin increments corresponding to each growth increment.

[0034] As an example, the central tendency of the data is represented by the average value, which is the average of all bilirubin increments corresponding to each growth increment at the target age.

[0035] Further, curve fitting was performed on the growth increment and corresponding bilirubin increment of all species at the target age to obtain the change curve of bilirubin increment-growth increment, and the distribution of daily growth increment and corresponding bilirubin increment was represented by the change curve. Finally, surface fitting was performed on all the age-related curves to obtain the surface curve of bilirubin increment-age-growth increment.

[0036] Please see Figure 2 This diagram illustrates an example of a surface graph showing the variation of bilirubin increment-age-growth increment according to an embodiment of the present invention. Figure 2 As can be seen, with increasing age, the increase in bilirubin generally shows a trend of first increasing and then decreasing.

[0037] It should be noted that, in one embodiment of the present invention, the least squares method is used for curve fitting, and the surface fitting based on the B-spline curve is used to obtain the surface of change of bilirubin increment-age-growth increment. These are all technical means well known to those skilled in the art and will not be described in detail here.

[0038] Step S3: Based on the growth increment during feeding within the current newborn's preset time domain neighborhood, combined with the change surface and the current bilirubin index, obtain the historical predicted bilirubin; based on the feeding time interval within the current newborn's preset time domain neighborhood, obtain the historical predicted feeding interval; input the sequence of the current newborn's bilirubin index into the pre-trained neural network to obtain the network predicted bilirubin and the network predicted feeding interval.

[0039] The growth increments of newborns during their most recent feedings reflect their current growth and development trends. The change surface contains the classic relationship between historical newborn growth and bilirubin increases. By combining the change surface with historical data, the bilirubin increase under the current developmental trend can be predicted. Furthermore, by combining the current bilirubin index, historical predicted bilirubin can be obtained from the perspective of historical physiological models. Therefore, based on the growth increments of newborns during feedings within a preset time domain, combined with the change surface and the current bilirubin index, historical predicted bilirubin can be obtained, providing a historical reference. This facilitates subsequent comparison with the prediction results of neural network models for the current individual newborns and analyzes the risk of jaundice.

[0040] Preferably, in one embodiment of the present invention, the predicted growth increment of the current newborn at the next feeding is obtained based on the central trend of the growth increment of feeding within a preset time domain neighborhood; the predicted growth increment of the current newborn and its age are substituted into the change surface, and combined with the current bilirubin index, the historical predicted bilirubin is obtained.

[0041] As an example, the preset time-domain neighborhood includes the current three most recent feedings, such as the current i-th, i-1, and i-2 feedings. Considering that the feeding time of newborns is relatively short and the growth and development trends in close time domains are similar, the growth increment of the current newborn at the next feeding can be represented by the mean value to represent the central tendency of the growth increment after at least three feedings. The mean value of the growth increment of the current feedings in the preset time-domain neighborhood can be used as the predicted growth increment of the current newborn at the next feeding.

[0042] At this point, the newborn's age can be easily obtained from the feeding record data. Substituting the newborn's predicted growth increment and age into the change surface, the predicted bilirubin increment for the next feeding can be obtained. Then, using the current bilirubin level as a benchmark, the sum of the current bilirubin level and the predicted bilirubin increment for the next feeding is used as the historical predicted bilirubin. The change surface obtained from historical newborn data is used to predict the current newborn's bilirubin level, providing a historical reference.

[0043] Considering that subsequent predictions of bilirubin levels in newborns will be based on neural networks, and changes in bilirubin are closely related to breastfeeding, in order to eliminate the influence of different predicted feeding intervals in different prediction methods and to accurately compare the differences in prediction results between the two methods, it is also necessary to obtain the time intervals predicted by each method separately. Therefore, historical predicted feeding intervals are obtained based on the feeding intervals within the current newborn's preset time domain neighborhood.

[0044] Preferably, in one embodiment of the present invention, considering that the feeding behavior of newborns remains relatively stable in a short period of time, the average of the time intervals within the current newborn's preset time domain neighborhood is used as the historical predicted feeding interval.

[0045] Neural networks model historical bilirubin data for individual newborns, reflecting their development and bilirubin metabolism levels, making predictions more individualized and real-time. By inputting the current newborn's bilirubin index sequence into a pre-trained neural network, the network can predict bilirubin and feeding intervals. By comparing the predictions with historical bilirubin, the current risk of jaundice development can be determined, thus providing an early warning.

[0046] Preferably, in one embodiment of the present invention, the pre-trained neural network is an LSTM network, the loss function is cross-entropy loss, the length of the training sequence is 3, and the sequence of the current newborn's bilirubin index is decomposed, for example [a, b, c, d, e, f]. The training set will be decomposed into [a, b, c], with a target value of d; [b, c, d], with a target value of e; [c, d, e], with a target value of f. If the training length is insufficient, a zero-padding operation is performed, such as [a, b, c], which can be decomposed into [0, a, b], with a target value of c. The number of zeros padded cannot exceed half of the training length.

[0047] Based on the central tendency of all time intervals in the current sequence of bilirubin levels in newborns, the network predicts feeding intervals.

[0048] As an example, to represent the central tendency of the time intervals, the mean of all time intervals in the sequence of current newborn bilirubin levels is used as the basis for obtaining the network's predicted feeding intervals.

[0049] In another embodiment of the present invention, considering that the closer the time interval is to the current time point, the greater the reference value, the corresponding time interval is assigned a weight based on the time distance between the smaller feeding time and the current time among the two feeding times corresponding to each time interval. For example, the time interval between the i-10th and i-9th feedings is weighted according to the time distance between the i-10th and i-th feedings. Specifically, for each time interval in the current bilirubin index sequence, the reciprocal of the time distance between the smaller feeding time in each time interval and the current time is used as the independent variable. After mapping through the softmax function, it is used as the weight of the corresponding time interval. The weighted sum of all time intervals is used to obtain the network's predicted feeding interval.

[0050] The logical relationship is adjusted by taking the reciprocal, normalization is performed by the softmax function, and the sum of all weighted weights is guaranteed to be 1. As the bilirubin index sequence increases, the sample size of the softmax function increases accordingly. The time interval and time distance are the intervals between two time points, and different names are used for easy distinction.

[0051] It should be noted that the specific training process of LSTM networks using the cross-entropy loss function is already existing technology and will not be described in detail here.

[0052] Step S4: Based on the difference between historical predicted bilirubin and network predicted bilirubin, and combined with the difference between historical predicted feeding intervals and network predicted feeding intervals, a four-level jaundice warning is issued.

[0053] Historical prediction of bilirubin is based on the change surface of the classic historical model, which reflects the physiological model trend summarized from a large amount of historical data; network prediction of bilirubin comes from neural networks, and the prediction is more individualized and real-time; based on the difference between historical prediction of bilirubin and network prediction of bilirubin, we can judge the deviation of the current newborn from the typical developmental trend and predict the risk of bilirubin metabolism. Furthermore, considering that feeding interval is an important influencing factor of breast milk jaundice, in order to eliminate the impact of different predicted feeding intervals in different prediction methods and accurately compare the differences in prediction results between the two prediction methods, the differences between historical predicted feeding intervals and online predicted feeding intervals are also combined to carry out a four-level jaundice early warning system and multimodal analysis early warning, thereby improving the reliability of prediction results and the timeliness of medical intervention.

[0054] Preferably, in one embodiment of the present invention, in order to make the historical predicted bilirubin more closely reflect the current individual feeding trends of newborns, the historical predicted bilirubin is corrected according to the ratio of the network predicted feeding interval to the historical predicted feeding interval to obtain the corrected historical predicted bilirubin.

[0055] As an example, the product of the ratio of the network-predicted feeding interval to the historically predicted feeding interval and the historically predicted bilirubin is used as the corrected historically predicted bilirubin.

[0056] The network-predicted feeding interval is in the numerator of the ratio. The formula can also be transformed into the ratio of historical predicted feeding interval to historical predicted bilirubin, and the product of the ratio and the network-predicted feeding interval is the corrected historical predicted bilirubin. The ratio of historical predicted feeding interval to historical predicted bilirubin represents the "historical predicted bilirubin per unit time interval". Multiplying it by the network-predicted feeding interval gives the corrected historical predicted bilirubin for the time interval of the network-predicted feeding interval.

[0057] Considering that the higher the predicted bilirubin level is compared to the revised historical predicted bilirubin level, it indicates that under the same developmental conditions, the current newborn has a higher bilirubin level than historical newborns, and the risk of jaundice is higher. Therefore, a jaundice risk coefficient is obtained based on the degree to which the predicted bilirubin level is higher than the revised historical predicted bilirubin level; and a four-level jaundice warning is based on the jaundice risk coefficient.

[0058] As an example, the difference between the network-predicted bilirubin and the corrected historical bilirubin is used as the independent variable. A positive correlation mapping is performed using the exp(x) function, and the mapped value is then linearly normalized. The normalized result is used as the jaundice risk coefficient.

[0059] Where x is the independent variable; linear normalization is the normalization performed on the corresponding data dimension. For example, the jaundice risk coefficient is the data dimension of the mapping value of the exp(x) function that predicts the jaundice risk of all newborns. The normalization used in the embodiments of the present invention can all adopt this method, which is a technical means well known to those skilled in the art and will not be described in detail here.

[0060] In one embodiment of the present invention, when the jaundice risk coefficient is in a preset low-risk threshold range, it is determined to be low-risk; when the jaundice risk coefficient is in a preset low-to-medium-risk threshold range, it is determined to be low-to-medium-risk; when the jaundice risk coefficient is in a preset high-to-medium-risk threshold range, it is determined to be high-to-medium-risk; and when the jaundice risk coefficient is in a preset high-risk threshold range, it is determined to be high-risk.

[0061] As an example, the preset low-risk threshold range, preset low-to-medium-risk threshold range, preset medium-to-high-risk threshold range, and preset high-risk threshold range are [0,0.3), [0.3,0.6), [0.6,0.8), and [0.8,1.0], respectively. The corresponding measures are: home observation, community follow-up, outpatient re-examination, and emergency hospitalization.

[0062] In other embodiments of the present invention, the implementer may adjust each threshold interval as needed, but must ensure that the preset low-risk threshold interval, preset medium-low-risk threshold interval, preset medium-high-risk threshold interval, and preset high-risk threshold interval are logically related from small to large.

[0063] It should be noted that the embodiments of the present invention only predict abnormal trends of bilirubin through multimodal analysis to provide risk classification and early warning for jaundice, and do not involve the diagnostic process of jaundice.

[0064] An embodiment of the present invention also provides a four-level early warning system for neonatal jaundice based on multimodal analysis. The system includes a memory, a processor, and a computer program. The memory is used to store the corresponding computer program, and the processor is used to run the corresponding computer program. When the computer program runs in the processor, it can implement the four-level early warning method for neonatal jaundice based on multimodal analysis described in steps S1-S4.

[0065] In summary, to address the technical problem of inaccurate jaundice early warning caused by neglecting bilirubin changes induced by breastfeeding in existing jaundice risk assessment methods, this invention proposes a four-level early warning system and method for neonatal jaundice based on multimodal analysis. This invention first acquires feeding data of current and historical newborns; then, based on the distribution of growth increments and corresponding bilirubin increments of historical newborns of the same age, it obtains a surface curve of bilirubin increment-age-growth increment; further, it obtains historical predicted bilirubin based on the surface curve; further, it obtains network predicted bilirubin based on a pre-trained neural network; finally, it performs a four-level jaundice early warning based on the difference between historical predicted bilirubin and network predicted bilirubin, combined with the difference between historical predicted feeding intervals and network predicted feeding intervals. This scheme considers the impact of breastfeeding, integrates physiological models and neural networks, and achieves personalized prediction and multi-source comparison of bilirubin dynamics, solving the problems of poor model generalization and delayed early warning in traditional jaundice monitoring, and improving the accuracy and timeliness of jaundice risk identification.

[0066] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0067] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A four-stage warning method for neonatal jaundice based on multi-modal analysis, characterized in that, The method comprises: acquiring the day age, weight, bilirubin index of the current newborn and historical newborns at feeding time, and bilirubin increment and time interval between adjacent feedings; for each newborn, acquiring the growth increment of the newborn at each feeding according to the change of the weight at adjacent feeding time; acquiring the variation surface of bilirubin increment-day age-growth increment according to the distribution of the growth increment and the corresponding bilirubin increment of the historical newborns of the same day age; acquiring the historical predicted bilirubin according to the growth increment at feeding time within the preset time domain neighborhood of the current newborn, combining the variation surface and the current bilirubin index; acquiring the historical predicted feeding interval according to the time interval of feeding within the preset time domain neighborhood of the current newborn; inputting the sequence of the bilirubin index of the current newborn into a pre-trained neural network to acquire the network predicted bilirubin and the network predicted feeding interval; carrying out jaundice four-level early warning according to the difference between the historical predicted bilirubin and the network predicted bilirubin, combining the difference between the historical predicted feeding interval and the network predicted feeding interval.

2. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The method for acquiring the variation surface comprises: selecting the day age of the historical newborns one by one as the target day age, classifying the growth increment with data values at the target day age, and acquiring the bilirubin increment corresponding to each growth increment according to the centralized trend of all the bilirubin increments corresponding to each growth increment; carrying out curve fitting on all the growth increments and the corresponding bilirubin increments of the target day age to obtain the variation curve of bilirubin increment-growth increment; and carrying out surface fitting on the variation curves of all the day ages to obtain the variation surface of bilirubin increment-day age-growth increment.

3. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The method for acquiring the historical predicted bilirubin comprises: acquiring the predicted growth increment at the next feeding time of the current newborn according to the centralized trend of the growth increment at feeding time within the preset time domain neighborhood; and inputting the predicted growth increment and the day age of the current newborn into the variation surface to acquire the historical predicted bilirubin in combination with the current bilirubin index.

4. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The method for carrying out jaundice four-level early warning comprises: correcting the historical predicted bilirubin according to the ratio of the network predicted feeding interval to the historical predicted feeding interval to obtain the corrected historical predicted bilirubin; acquiring the jaundice risk coefficient according to the extent of the network predicted bilirubin being larger than the corrected historical predicted bilirubin; and carrying out jaundice four-level early warning based on the jaundice risk coefficient.

5. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 4, characterized in that, The method for carrying out jaundice four-level early warning based on the jaundice risk coefficient comprises: when the jaundice risk coefficient is within a preset low-risk threshold interval, determining as low-risk; when the jaundice risk coefficient is within a preset medium-low-risk threshold interval, determining as medium-low-risk; when the jaundice risk coefficient is within a preset medium-high-risk threshold interval, determining as medium-high-risk; and when the jaundice risk coefficient is within a preset high-risk threshold interval, determining as high-risk.

6. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The method for acquiring the historical predicted feeding interval comprises: taking the mean value of the time intervals within the preset time domain neighborhood of the current newborn as the historical predicted feeding interval.

7. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The method for acquiring the network predicted feeding interval comprises: According to the sequence of the bilirubin indicators of the current newborn, the central tendency of all the time intervals is obtained to predict the feeding interval.

8. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The growth increment acquisition method comprises: The increment of the body weight at each feeding compared to the body weight at the last feeding is taken as the growth increment of each feeding.

9. The four-stage warning method for neonatal jaundice based on multi-modal analysis according to claim 1, characterized in that, The pre-trained neural network is an LSTM network.

10. A multi-modal analysis based four-stage warning system for neonatal jaundice, the system comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the steps of the four-level early warning method for neonatal jaundice based on multi-modal analysis according to any one of claims 1-9 when executing the computer program.

Images