Retinopathy of Prematurity Screening Method, Screening Device, and Trained Model

Abstract

The purpose of the present invention is to provide: a method for screening for retinopathy of prematurity, which makes it possible to perform the prediction of the progression of retinopathy of prematurity at a proper timing and with high accuracy and which has broad utility; an apparatus for screening for retinopathy of prematurity; and a learned model. Provided is a method for screening for retinopathy of prematurity by performing the prediction of the progression of retinopathy of prematurity, the method including a treatment determination step of determining as to whether or not a treatment for retinopathy of prematurity can be applied to a premature baby on or after a predetermined number of days after birth on the basis of information about the premature baby, in which the premature baby is one whose fetus week number is smaller than a predetermined number of weeks, and the information includes time-series data after birth which are associated with the body weight, body height and vital signs of the premature baby.

Description

Technical Field

[0001] The present disclosure relates to a retinopathy of prematurity screening method, a screening device, and a trained model for predicting the progression of retinopathy of prematurity, which are used to assist physicians in diagnosing retinopathy of prematurity.

Background Art

[0002] Retinopathy of prematurity (ROP) is a major cause of blindness in childhood. It is estimated that 50,000 people worldwide go blind annually, and the number is expected to increase in the future. Although many cases can recover spontaneously even if they develop, especially in extremely premature infants, blindness may occur due to intraocular hemorrhage or retinal detachment. For such severe cases, retinal photocoagulation is performed as a standard treatment. It has been previously shown that retinal photocoagulation is effective in suppressing the progression of retinopathy of prematurity, but it avoids blindness at the cost of tissue destruction and is not a preventive treatment method. Recently, intravitreal administration of a vascular endothelial growth factor (VEGF) inhibitor (anti-VEGF drug), which has an effect equivalent to that of photocoagulation, has been carried out as a treatment, but there are concerns about its impact on overall development, and it is still not suitable for preventive treatment. On the other hand, it has been clarified that a delay in the treatment time results in the inability to obtain the effect of photocoagulation due to the deterioration of the disease condition, and the risk of blindness increases rapidly. Therefore, currently, treatment is carried out for cases that have reached a certain severity according to the disease stage determination based on the international classification and the treatment criteria based on the results of the Early Treatment for Retinopathy of Prematurity Study (ETROP) in the United States. In order to correspond to the therapeutic time window in which treatment is successful, frequent fundus examinations and prompt treatment are required.

[0003] One example of a screening method for retinopathy of prematurity is the technique described in Patent Document 1. This screening method detects tryptase, which can be released by mast cell degranulation, as a marker substance from the blood of a subject, and determines whether or not treatment for retinopathy of prematurity is necessary.

[0004] Furthermore, two models are known for screening for retinopathy of prematurity: the WINROP model developed in Sweden and the CHOP-ROP model reported from the United States. WINROP targets infants born between 23 and 32 weeks of gestation. By inputting gestational age, birth weight, and postnatal weight every week, an alarm is displayed for cases that may be at risk of worsening. The CHOP-ROP model, like WINROP, evaluates infants based on their weight gain every week, and can reduce the number of consultations in the low-risk group. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2014-208601 [Overview of the project] [Problems that the invention aims to solve]

[0006] However, the method described in Patent Document 1 requires an invasive procedure called blood sampling, which is not practical. Furthermore, even if treatment for retinopathy of prematurity is determined to be necessary using the method described in Patent Document 1, spontaneous recovery is possible. Conversely, even if treatment for retinopathy of prematurity is deemed unnecessary, it is possible that retinopathy of prematurity may develop and become severe within a few days. For this reason, frequent fundus examinations are important for predicting the progression of retinopathy of prematurity.

[0007] In Sweden and the United States, predictive models have been developed to reduce the number of screenings for non-progressive cases. However, these models all track postnatal weight gain, and while they have high sensitivity, their specificity is a challenge. Furthermore, accuracy has been shown to be low when the weight of the population differs (e.g., premature infants in developing countries, premature infants with a birth weight of less than 1000g who are at high risk of severe disease).

[0008] Therefore, there is a need for a highly versatile screening method, screening device, and pre-trained model for retinopathy of prematurity that can accurately predict the progression of retinopathy of prematurity at the appropriate time. [Means for solving the problem]

[0009] One aspect of this disclosure is a retinopathy of prematurity screening method for predicting the progression of retinopathy of prematurity, comprising a treatment determination step that determines whether or not treatment for retinopathy of prematurity will be indicated after a predetermined number of days after birth, based on premature infant information including postnatal time-series data on weight, height, and vital signs of premature infants born less than a predetermined gestational age. Another aspect of this disclosure is a screening device for predicting the progression of retinopathy of prematurity, comprising a treatment determination unit that determines whether or not treatment for retinopathy of prematurity will be indicated after a predetermined number of days after birth, based on premature infant information including postnatal time-series data on weight, height, and vital signs of premature infants born less than a predetermined gestational age.

[0010] This method or device can predict the progression of retinopathy of prematurity by determining whether treatment is indicated for retinopathy of prematurity using postnatal time-series data on weight, height, and vital signs. Many factors are intricately related to the progression of retinopathy of prematurity from its onset to the point where treatment is indicated, and these factors change over time. This configuration provides a method or device that predicts the progression of retinopathy of prematurity with high accuracy using multiple time-series data.

[0011] This method or device utilizes information including changes in immaturity and overall condition over time. For example, immaturity is assessed using indicators such as gestational age at birth, while postnatal overall condition is assessed using time-series data including vital signs such as heart rate, respiration, and blood oxygen saturation, as well as weight and height. It also determines whether or not the infant is eligible for treatment of retinopathy of prematurity after a predetermined number of days postpartum. A specific example of an application of this method or device is a diagnostic support device in a neonatal intensive care unit that incorporates a program into equipment that monitors the vital signs of premature infants over time and displays a warning sign for cases that will soon be eligible for treatment of retinopathy of prematurity.

[0012] This method or device allows for the prediction of retinopathy of prematurity progression with higher accuracy compared to existing models (WINROP and CHOP-ROP models). Prompt and appropriate treatment for retinopathy of prematurity reduces the risk of blindness. However, the shortage of ophthalmologists skilled in accurate staging and treatment is a serious global problem. This method or device, however, enables even those without highly specialized knowledge and experience to determine treatment eligibility and initiate treatment at the appropriate time. Furthermore, while retinopathy of prematurity is generally staging based on images obtained from indirect ophthalmoscope examinations or contact-type fundus cameras performed by ophthalmologists, both methods place a significant burden on vulnerable premature infants and carry risks such as heart rate drops during examinations and imaging. This method or device eliminates the need for fundus examinations or imaging, determining treatment eligibility based on premature infant information, including postnatal time-series data on weight, height, and vital signs, thus reducing unnecessary examinations. In other words, it not only contributes to healthcare, but also significantly reduces medical expenses, productivity losses, and social care costs associated with visual impairment, thus benefiting the socio-economic system.

[0013] Thus, this is a highly versatile screening method or device for retinopathy of prematurity that can accurately predict the progression of retinopathy of prematurity at the appropriate time.

[0014] Another aspect of this method further includes a risk assessment step of determining the risk of progression of retinopathy of prematurity based on a Type 1 ROP or APROP score calculated at a predetermined number of days after birth, wherein the treatment assessment step is performed only on premature infants who are determined to have the risk of progression in the risk assessment step. Another aspect of this device further includes a risk assessment unit of determining the risk of progression of retinopathy of prematurity based on a Type 1 ROP or APROP score calculated at a predetermined number of days after birth, wherein the treatment assessment unit determines whether the treatment is appropriate only for premature infants who are determined to have the risk of progression in the risk assessment unit.

[0015] This method or device can estimate the potential progression of retinopathy of prematurity in order to determine the risk of progression of retinopathy of prematurity. Furthermore, it is a highly accurate retinopathy of prematurity screening method or device that uses a two-step determination process to identify whether or not treatment is indicated only for premature infants determined to be at risk of progression.

[0016] Another aspect of this method is that the vital signs are at least one of the premature infant's heart rate, respiratory rate, and arterial blood oxygen saturation.

[0017] Since these vital signs can be obtained using existing monitoring equipment installed in the neonatal intensive care unit, there is no need to develop new equipment, making it efficient.

[0018] Another aspect of this method is that the premature infant information includes at least one of the premature infant's gestational age and Apgar score.

[0019] Thus, by using information on the gestational age and Apgar score of premature infants, in addition to their overall physical condition after birth, it is possible to predict the progression of retinopathy of prematurity with greater accuracy.

[0020] One aspect of the present disclosure is a learned model that functions by a computer, which is composed of a decision tree consisting of a plurality of branch points arranged in a tree structure, and feature quantities calculated based on premature infant information including postnatal time-series data on the weight, height, and vital signs of premature infants whose gestational age at the time of treatment for retinopathy of prematurity is less than a predetermined week are input, and by adding up the evaluation values at each of the branch points, a score for the necessity of treatment for retinopathy of prematurity is output.

[0021] If it is a learned model that has undergone machine learning by a decision tree as in this configuration, it can accurately predict the progression of retinopathy of prematurity at an appropriate timing and is highly versatile. Also, if premature infant information is processed to calculate feature quantities in order to input them into the learned model, the prediction of the progression of retinopathy of prematurity can be performed more accurately.

[0022] One aspect of the present disclosure is a learned model that functions by a computer, which is generated by deep learning including a convolutional neural network, and premature infant information including postnatal time-series data on the weight, height, and vital signs of premature infants whose gestational age at the time of treatment for retinopathy of prematurity is less than a predetermined week is input, and a score for the necessity of treatment for retinopathy of prematurity is output.

[0023] If it is a learned model that has undergone deep learning including a convolutional neural network as in this configuration, it can accurately predict the progression of retinopathy of prematurity at an appropriate timing and is highly versatile.

[0024] Another aspect regarding this model is that the premature infant information is information obtained from premature infants whose total gestational age and gestational age at the time of treatment are 40 weeks or less.

[0025] If the total of the gestational age and the gestational age at the time of treatment is greater than 40 weeks, the risk of progression of retinopathy of prematurity is extremely low. Therefore, the learned model in this configuration can accurately predict the progression of retinopathy of prematurity.

[0026] Another aspect of this model is that the premature infant information excludes the information obtained from the premature infants who were treated early based on the doctor's judgment.

[0027] By excluding such specific cases that were treated early based on the doctor's judgment, a good trained model can be provided.

[0028] Another aspect of this model is that the vital signs are at least one of the heart rate, respiratory rate, and arterial oxygen saturation of the premature infant.

[0029] With such vital signs, since they are obtained by the existing monitoring equipment installed in the neonatal intensive care unit, a large amount of input data for constructing the trained model can be secured.

[0030] Another aspect of this model is that the premature infant information includes at least one of the gestational age and Apgar score of the premature infant.

[0031] In this way, in addition to the overall postnatal condition, by constructing a trained model using the gestational age and Apgar score of the premature infant, the prediction of the progression of retinopathy of prematurity can be performed more accurately.

Brief Description of Drawings

[0032] [Figure 1] is an overall system diagram for realizing the screening method according to this embodiment. [Figure 2] is a block diagram of the screening device according to this embodiment. [Figure 3] is a flowchart for realizing the screening method according to this embodiment. [Figure 4] is an explanatory diagram of the screening method according to this embodiment. [Figure 5] is a diagram showing the relationship between gestational age and treatment results. [Figure 6] is a diagram showing the relationship between the gestational week or gestational week at treatment and the total gestational week at treatment and treatment results. [Figure 7] This is a diagram showing an example of the risk assessment process. [Figure 8] This is a diagram showing the contribution of each feature in machine learning. [Figure 9] This is an example of an ROC curve plot where the treatment determination process was performed using a pre-trained machine learning model. [Figure 10] This is an AUC diagram of an example where a treatment decision process was performed using a pre-trained model developed through deep learning. [Modes for carrying out the invention]

[0033] The embodiments of the retinopathy of prematurity screening method, screening apparatus, and trained model relating to this disclosure will be described below with reference to the drawings. However, the embodiments are not limited to those described below, and various modifications are possible without departing from the gist of the invention.

[0034] Figures 1 and 2 illustrate the system configuration used in the retinopathy screening method for prematurity.

[0035] One or more monitoring devices 1 for acquiring premature infant information, including postnatal time-series data on weight, height, and vital signs, are connected to an internet connection 2. A trained model generator 3, a screening device 4, and AI9 (artificial intelligence) are also connected to the internet connection 2. Here, AI9 may be located on the internet connection 2 or within the trained model generator 3. Furthermore, the trained model generator 3 and the screening device 4 may be the same device, or the screening device 4 may be built into the monitoring device 1, and each function can be used individually or in combination. The screening device 4 may also be a monitoring device installed in an incubator or a dedicated device in a neonatal intensive care unit, and can be used as various diagnostic support devices for predicting the progression of retinopathy of prematurity.

[0036] Monitoring device 1 is a device for monitoring the vital signs of premature infants over time in a neonatal intensive care unit and for periodically measuring the weight and height of premature infants. The premature infant information in this embodiment includes the weight, height, and vital signs of premature infants whose gestational age is less than a predetermined number of weeks (e.g., 28 weeks). This predetermined number of weeks is 36 weeks or less (so-called premature infant), preferably 32 weeks or less, and more preferably 28 weeks or less (so-called extremely premature infant) (the same applies hereinafter). In this embodiment, gestational age at birth is used as an indicator of prematurity, but birth weight may also be used as an indicator. This birth weight is less than 2500g (so-called low birth weight infant), preferably less than 1500g (so-called very low birth weight infant), and more preferably less than 1000g (so-called extremely low birth weight infant). Vital signs include at least one of the premature infant's heart rate, respiratory rate, and arterial blood oxygen saturation, and are taken at least every minute. These vital signs may also include blood pressure, etc., as long as they provide vital information for the premature infant. Premature infant information preferably includes at least one of the premature infant's weight taken three times a week, the premature infant's height taken once a week, the premature infant's gestational age, and the Apgar score taken one minute and five minutes after birth. The frequency of taking premature infant information is not particularly limited; for example, vital signs may be taken every second or every ten minutes, and weight and height may be taken daily. The Apgar score evaluates the condition of a newborn immediately after birth and is an evaluation of five items: skin color, heart rate, responsiveness, muscle tone, and respiration, with a total score out of 10.

[0037] As shown in Figure 2, the trained model generation device 3 comprises a first communication unit 31, a model generation unit 32, a learning feature calculation unit 33, and a first storage unit 34.

[0038] The first communication unit 31 is an interface that transmits and receives data between the monitoring device 1, the screening device 4, or the AI ​​9, etc., via the internet line 2. The first communication unit 31 may receive data directly from the monitoring device 1, or it may store the data acquired by the monitoring device 1 in a server (not shown) and receive the stored data from this server.

[0039] The first storage unit 34 is composed of a non-temporary storage medium such as an HDD or SSD, or a temporary storage medium such as RAM, and stores programs and applications executed by the processor. This first storage unit 34 stores the premature infant learning information 34a of the monitoring device 1 acquired via the first communication unit 31. The premature infant learning information 34a includes postnatal time-series data regarding the weight, height, and vital signs of premature infants whose gestational age is less than a predetermined number of weeks (e.g., 28 weeks). The premature infant learning information 34a also includes the gestational age and Apgar score of the premature infants. Preferably, this premature infant learning information 34a is information obtained from premature infants whose gestational age and gestational age at treatment are 40 weeks or less. Preferably, the premature infant learning information 34a excludes information obtained from premature infants who received early treatment at the discretion of a physician.

[0040] The first memory unit 34 stores treatment information 34c associated with learning premature infant information 34a. This treatment information 34c is categorized into no treatment, treated Type 1 ROP, and treated APROP (Aggressive Posterior Retinopathy of Prematurity). The international classification of retinopathy of prematurity, Classic ROP (typical ROP), has Type 1 ROP and Type 2 ROP, with Type 1 being a treatment indication and Type 2 being less than a treatment indication (no treatment). On the other hand, a type that rapidly worsens separately from Classic ROP is called APROP, and this is also a treatment indication. Type 1 ROP (international classification) corresponds to Type 1 Stage 3 (Ministry of Health classification), and APROP (international classification) corresponds to Type 2 (fulminant type). Hereinafter, treatment indication means that treatment is carried out within 72 hours of diagnosis of Type 1 ROP, or that treatment is carried out promptly if there are early signs of APROP (the same applies below). ROPs that are eligible for treatment are any of the following: Type 1 ROPs, specifically zone 1 ROPs with plus disease, zone 1 stage 3 ROPs without plus disease, and zone 2 stage 3 ROPs with plus disease, or APROPs. Here, plus disease refers to retinal vascular dilation or tortuosity, zone 1 is the region within a circle with a radius of twice the optic disc-macula distance centered on the optic disc, zone 2 is the region within a circle with a radius from the optic disc to the nasal serrate edge, and stage 3 refers to extraretinal fibrovascular proliferation.

[0041] The first memory unit 34 stores the trained model 10. The trained model 10 is a computer-operated model obtained through machine learning or deep learning with supervised data. Furthermore, the first memory unit 34 stores the training features 34b calculated by the training feature calculation unit 33 for machine learning.

[0042] Machine learning is structured around a decision tree, which consists of multiple branching points arranged in a tree structure. This machine learning model evaluates features at each branching point of the decision tree, and assigns an evaluation value to each branching point according to the evaluation result. The evaluation values ​​are then summed up along the branches of the decision tree to obtain information predicting the progression of retinopathy of prematurity. Machine learning can also be performed using ensemble models such as XGBoost, Random Forest, LightGBM, CatBoost, and AdaBoost, in which multiple decision trees are related to each other.

[0043] Deep learning is performed by AI9, which includes well-known convolutional neural networks (CNN, DCGAN, etc.). Convolutional neural networks construct deep, layered models that mimic human neural circuits to infer the progression of retinopathy of prematurity. This deep learning consists of well-known applications provided via Internet connection 2.

[0044] The model generation unit 32 is equipped with a processor and generates a trained model 10. The processor includes an ASIC, FPGA, CPU, or other hardware for executing applications stored in the first memory unit 34 (hereinafter the same applies). When generating a trained model 10 by machine learning, the model generation unit 32 uses the input data as training features 34b and the training data as treatment information 34c (no treatment, treatment Type 1 ROP, treatment APROP) to perform reinforcement learning and generate the trained model 10. The training features 34b are calculated by processing training premature infant information 34a (time-series data after birth regarding the weight, height, and vital signs of premature infants, the gestational age of premature infants, and Apgar scores, etc.), but the details will be described later. On the other hand, when generating a trained model 10 using deep learning, the model generation unit 32 uses training premature infant information 34a (time-series data after birth regarding the weight, height, and vital signs of premature infants, etc.) as input data and treatment information 34c (no treatment, treatment with Type 1 ROP, treatment with APROP) as training data to perform reinforcement learning and generate a trained model 10.

[0045] The learning feature calculation unit 33 is equipped with a processor and calculates multiple learning features 34b from the learning premature infant information 34a. The calculated learning features 34b are input into the trained model 10 for analysis, and the degree of influence of each feature on the treatment information 34c is determined. Figure 8 shows an example of the degree of influence of multiple learning features 34b on the treatment information 34c. The 34b learning features, in order of their impact, are weight, height, height_SD, gestational age, Apgar score at 5 minutes postnatal (Apgar Score; 5min), weight_SD, Apgar score at 1 minute postnatal (Apgar Score; 1min), arterial oxygen saturation (SpO2%), heart rate (HR bpm), sex (M or F), onset, birth type (singleborn, twins, triplet), count, respiratory rate (RESP / min), difference in respiratory rate (RESP / min.delta), difference in heart rate (HR bpm.delta), difference in weight_SD, difference in weight, difference in arterial oxygen saturation (SpO2%.delta), difference in height_SD, and difference in height. The learning feature calculation unit 33 may extract multiple indicators (for example, 10) in order of their influence, and the model generation unit 32 may use these extracted indicators as learning features 34b.

[0046] Weight is calculated by forward interpolation of the daily average (or hourly average) weight of premature infants taken three times a week after birth. Height is calculated by forward interpolation of the daily average (or hourly average) height of premature infants taken once a week after birth. SD represents the magnitude of variability from the mean, i.e., the width of the distribution, expressed as a numerical value called SD (standard deviation). Arterial blood oxygen saturation, heart rate, and respiratory rate are calculated by removing zero values ​​from vital sign data taken every minute after birth and interpolating them as a daily average (or hourly average). Onset refers to the findings of retinopathy of prematurity (whether or not retinopathy of prematurity has occurred) by a physician performed a predetermined number of times after birth. Count is the number of days elapsed since birth converted to units of time. Difference is calculated as the difference for each parameter for each measurement and used as a difference feature. In this embodiment, the input data for weight, height, and vital signs uses 1-day average values, but any 1-minute average to 2-day average values ​​are acceptable, preferably 1-hour average to 1-day average values, and more preferably 1-hour average values ​​or 1-day average values ​​(the same applies hereinafter). If average values ​​exceeding 2 days are used as input data, the prediction accuracy will be poor, and if average values ​​less than 1 minute are used, the amount of data will be large, which may lead to a decrease in calculation speed and the introduction of noise.

[0047] Furthermore, when a trained model 10 is generated by deep learning, the training feature calculation unit 33 is built into the trained model 10. Specifically, the trained model 10 performs weighting using time-series feature information within its internal (convolutional layer) and weighting using past predicted features within its internal (attention mechanism). Here, time-series feature information refers to features obtained by arranging the training premature infant information 34a in a time series, and past predicted features refer to weighted features extracted by the trained model 10 itself as a result of its own predictions, in order to use the weighting of past time-series feature information in the current prediction.

[0048] The trained model 10 generated in this way outputs a score indicating whether treatment for retinopathy of prematurity is necessary. In this embodiment, the necessity of treatment is categorized into no treatment, treatment Type 1 ROP, and treatment APROP, similar to the treatment information 34c, and the score for the necessity of treatment is expressed as the AUC (Area Under the Curve) of the time series for each of the three categories: no treatment, treatment Type 1 ROP, and treatment APROP. When premature infant information, including postnatal time series data on weight, height, and vital signs of premature infants born before a predetermined gestational age (e.g., 28 weeks), is input to the trained model 10, it can output scores for no treatment, treatment Type 1 ROP, and treatment APROP from a predetermined number of days after birth (e.g., 20 days). This predetermined number of days after birth is between 1 week and 5 weeks, preferably between 2 weeks and 4 weeks, and more preferably around 3 weeks (the same applies hereinafter). The score is expressed as a daily or hourly value (e.g., AUC of the time series). In this embodiment, the score is calculated daily, but the calculation interval is every minute to every two days, preferably every hour to every day, and more preferably every hour or every day (the same applies hereinafter). If the calculation interval is longer than two days, the prediction accuracy will be poor, and if it is less than one minute, the amount of data will be large, which may lead to a decrease in calculation speed and the introduction of noise. This score is calculated for no treatment, treated Type 1 ROP, and treated APROP, and if the value for treated is the highest among the scores for each, the treatment time range will be reached in the near future (treatment indication). Here, the treatment time range means that treatment will be carried out within 72 hours in the case of treated Type 1 ROP, or that treatment will be carried out promptly in the case of treated APROP. Treatment is selected from one of the following methods: retinal photocoagulation, retinal cryocoagulation, intravitreal administration of anti-VEGF drugs, and vitrectomy, but retinal photocoagulation or intravitreal administration of anti-VEGF drugs is preferred. Vitrectomy is performed when retinal detachment develops after insufficient response to retinal photocoagulation or anti-VEGF drug therapy. Furthermore, the trained model 10 may be configured to output not only a score indicating whether treatment will be performed in the future (i.e., whether treatment is indicated), but also the timing or time range of treatment.

[0049] As shown in Figure 2, the screening device 4 includes a second communication unit 41, a predictive feature calculation unit 42, a risk determination unit 43, a treatment determination unit 44, a notification unit 45, and a second storage unit 46.

[0050] The second communication unit 41 is an interface that transmits and receives data between the monitoring device 1, the trained model generation device 3, etc., via the internet line 2. The second communication unit 41 may receive data directly from the monitoring device 1, or it may store the data acquired by the monitoring device 1 in a server (not shown) and receive the stored data from this server.

[0051] The second storage unit 46 is composed of a non-temporary storage medium such as an HDD or SSD, or a temporary storage medium such as RAM, and stores programs and applications executed by the processor. The second storage unit 46 stores the predictive premature infant information 46a acquired via the second communication unit 41 from the monitoring device 1 and the trained model 10 generated by the trained model generation device 3. The predictive premature infant information 46a includes postnatal time-series data regarding the weight, height, and vital signs of premature infants whose gestational age is less than a predetermined number of weeks (e.g., 28 weeks). The predictive premature infant information 46a also includes the gestational age and Apgar score of the premature infant. Preferably, this predictive premature infant information 46a is information obtained from premature infants whose total gestational age and gestational age at treatment is 40 weeks or less.

[0052] The second memory unit 46 stores prediction features 46b calculated by the prediction feature calculation unit 42 for input into the trained model 10 that has been machine-learned by the trained model generation device 3.

[0053] The second memory unit 46 stores the judgment result 46c output from the trained model 10. This judgment result 46c is time-series data categorized into no treatment, treatment Type 1 ROP, and treatment APROP. In this embodiment, the judgment result 46c includes the ROC (Receiver Operating Characteristic) curves for no treatment, treatment Type 1 ROP, and treatment APROP. The judgment result 46c also includes the AUC (Area Under the Curve) of the time series calculated from this ROC curve.

[0054] The feature calculation unit 42 for prediction is equipped with a processor and calculates multiple predictive features 46b from the predictive premature infant information 46a. The 46 predictive features used are the same as the 34 trainee features, except for the onset of disease. The predictive features 46b, in order of increasing influence, are weight, height, height_SD, gestational age, Apgar score at 5 minutes postnatal (Apgar Score; 5min), weight_SD, Apgar score at 1 minute postnatal (Apgar Score; 1min), arterial oxygen saturation (SpO2%), heart rate (HR bpm), sex (M or F), birth type (singleborn, twins, triplet), respiratory rate (RESP / min), difference in respiratory rate (RESP / min.delta), difference in heart rate (HR bpm.delta), difference in weight_SD, difference in weight, difference in arterial oxygen saturation (SpO2%.delta), difference in height_SD, and difference in height.

[0055] The risk determination unit 43 is equipped with a processor, and at a predetermined number of days after birth (e.g., 20 days after birth), the trained model 10, into which the predictive premature infant information 46a has been input, outputs the risk of progression of retinopathy of prematurity. The risk of progression is expressed as a score on a daily or hourly basis. This score is calculated for each of the following: no treatment, treated Type 1 ROP, and treated APROP. At a predetermined number of days after birth, if the value of the treated score is higher than a predetermined value, it is determined that there is a risk of progression. As an example, the risk determination unit 43 extracts the judgment index with the highest AUC at a predetermined number of days after birth from the time-series AUC of multiple judgment indices consisting of no treatment, treated Type 1 ROP, and treated APROP. If the AUC of treated Type 1 ROP or treated APROP is higher than a predetermined value (e.g., 0.3), it is determined that there is a risk of progression. This predetermined value is set from between 0.1 and 0.8, preferably 0.2 and 0.6, and more preferably 0.3 and 0.5.

[0056] The treatment determination unit 44 is equipped with a processor, and when predictive premature infant information 46a, which includes postnatal time-series data on weight, height, and vital signs of premature infants born before a predetermined gestational age (e.g., 28 weeks), is input to the trained model 10, it determines whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth (e.g., 20 days) based on the output value of the trained model 10. Preferably, the treatment determination unit 44 determines whether or not treatment is indicated only for premature infants that the risk determination unit 43 has determined to be at risk of progression. Whether or not treatment for retinopathy of prematurity is indicated is expressed as a score on a daily or hourly basis. This score is calculated for no treatment, treatment with Type 1 ROP, and treatment with APROP, and if the value for treatment is the highest among the respective scores, the treatment determination unit 44 determines that the treatment time range using retinal photocoagulation, etc., will be reached in the near future (treatment is indicated). As an example, the treatment determination unit 44 extracts the determination index with the highest AUC from the time-series AUC of multiple determination indexes consisting of no treatment, treatment type 1 ROP, and treatment APROP, and determines that treatment is indicated if the AUC of treatment type 1 ROP or treatment APROP exceeds the AUC of no treatment. As another example, the treatment determination unit 44 determines that treatment is indicated if the AUC of treatment type 1 ROP or treatment APROP exceeds a treatment threshold (e.g., 0.8) from the time-series AUC of multiple determination indexes consisting of no treatment, treatment type 1 ROP, and treatment APROP. This treatment threshold is set between 0.5 and 0.9, preferably 0.6 and 0.9, and more preferably 0.7 and 0.8.

[0057] The notification unit 45 outputs a warning signal when the treatment determination unit 44 determines that treatment is indicated. The notification unit 45 may consist of a warning lamp or warning sound mounted on equipment that monitors the vital signs of premature infants over time in a neonatal intensive care unit, or it may consist of a predetermined notification device installed in the nurses' station.

[0058] Next, using Figures 3 to 10, an example of a retinopathy of prematurity screening method (program) executed by a computer to predict the progression of retinopathy of prematurity using the trained model 10 according to this embodiment will be described.

[0059] The trained model generator 3 acquires premature infant training information 34a and treatment information 34c from each monitoring device 1 over a predetermined period via the internet connection 2 (#31 in Figure 3). As shown in Figure 5, according to data from 719 premature infants at Hospital A, the probability of developing retinopathy of prematurity decreases to about 10% when the gestational age reaches 27 weeks. In addition, among premature infants born before 28 weeks of gestation, approximately 40% of those with a birth weight of less than 1000g are candidates for treatment, and premature infants with a birth weight of less than 1000g are at high risk of developing retinopathy of prematurity (APROP), which rapidly worsens. Therefore, in this embodiment, premature infants born before 28 weeks of gestation are targeted for screening. Accordingly, the trained model generator 3 extracts data on premature infants born before 28 weeks of gestation as training information 34a and treatment information 34c (#32 in Figure 3, filtering).

[0060] As shown in the left panel of Figure 6, data from 206 premature infants treated at Hospital A indicates that the period for treatment of retinopathy of prematurity is from 6 to 16 weeks after birth, when treatment using retinal photocoagulation is performed. Also, as shown in the right panel of Figure 6, data from 206 premature infants treated at Hospital A indicates that the period for treatment using retinal photocoagulation is from 30 to 39 weeks of gestation, when the sum of gestational age and gestational age at treatment is 30 to 39 weeks. Therefore, the trained model generator 3 extracts data for premature infants whose sum of gestational age and gestational age at treatment is 40 weeks or less as training premature infant information 34a and treatment information 34c (Figure 3, #32, filtering). Alternatively, the trained model generator 3 may extract data for premature infants whose sum of gestational age and gestational age at treatment is between 29 weeks and 40 weeks as training premature infant information 34a and treatment information 34c. Furthermore, the trained model generator 3 excludes the training premature infant information 34a and treatment information 34c, which are unique cases where early treatment was administered based on the physician's judgment (Figure 3, #32, filtering). As a result, the training premature infant information 34a in this embodiment consists of time-series data of gestational days, weight, height, respiratory rate, heart rate, and arterial blood oxygen saturation for 206 premature infants treated at Hospital A, as well as the Apgar score at 5 minutes postpartum, the Apgar score at 1 minute postpartum, sex, birth type, count, and onset (definitions of terms are as described above).

[0061] Next, the model generation unit 32 of the trained model generation device 3 performs reinforcement learning using premature infant information 34a (time-series data after birth regarding the weight, height, and vital signs of premature infants, etc.) as input data and treatment information 34c (no treatment, treatment Type 1 ROP, treatment APROP) as training data, and generates a trained model 10 (Figure 3, #33~#36). When the model generation unit 32 performs machine learning (Figure 3, #33 Yes), the training feature calculation unit 33 calculates multiple training features 34b from the premature infant information 34a (Figure 3, #34, feature calculation process). These training features 34b include gestational days, daily average weight, weight difference, weight_SD, weight_SD difference, daily average height, height difference, height_SD, height_SD difference, daily average respiratory rate, respiratory rate difference, daily average heart rate, heart rate difference, daily average arterial oxygen saturation, arterial oxygen saturation difference, Apgar score at 5 minutes postnatal, Apgar score at 1 minute postnatal, sex, birth type, count, and onset (definitions of terms are as described above). As shown in Figure 8, when multiple training features 34b calculated using training premature infant information 34a from 206 premature infants treated at Hospital A are input into the trained model 10 and analyzed, the degree of influence of each feature on treatment information 34c can be determined. Then, the model generation unit 32 performs reinforcement learning using the input data as training features 34b and the training data as treatment information 34c (no treatment, treatment Type 1 ROP, treatment APROP) to generate a trained model 10 (#36 in Figure 3).

[0062] On the other hand, if machine learning is not performed (Figure 3, #33No), the model generation unit 32 performs deep learning including a convolutional neural network (Figure 3, #35). In this deep learning, the model generation unit 32 uses training premature infant information 34a (time-series data after birth regarding the weight, height, and vital signs of premature infants, etc.) as input data and treatment information 34c (no treatment, treatment with Type 1 ROP, treatment with APROP) as training data to perform reinforcement learning and generate a trained model 10 (Figure 3, #36).

[0063] Next, at 20 days postpartum, the screening device 4 outputs a daily score (time-series AUC calculated from the ROC curve) to the trained model 10 into which the predictive premature infant information 46a has been input (see Figure 7), and the risk determination unit 43 determines the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score (Figure 3, #37, risk determination process). This predictive premature infant information 46a includes gestational days, time-series data of weight, height, respiratory rate, heart rate and arterial blood oxygen saturation, Apgar score at 5 minutes postpartum, Apgar score at 1 minute postpartum, sex, birth type and count (definitions of terms are described above). This score is calculated for untreated, treated Type 1 ROP, and treated APROP. At 20 days postpartum, if the treated score is higher than a predetermined value (e.g., 0.3), the patient is judged to be at risk of progression (Figure 3, #38 Yes; Figure 7, left panel). On the other hand, if the treated score is below the predetermined value, the patient is judged to be at no risk of progression (Figure 3, #38 No; Figure 7, right panel).

[0064] Next, as shown in Figure 4, the predictive premature infant information 46a, which has been determined to have a risk of progression, is set as onset data. The treatment determination unit 44 inputs the predictive premature infant information 46a, which is the onset data, into the trained model 10 and determines whether or not treatment for retinopathy of prematurity is indicated after 20 days of age based on the output value of the trained model 10 (#39 in Figure 3, treatment determination step). In this embodiment, the trained model 10 can distinguish between cases of spontaneous regression and cases of disease progression among the predictive premature infant information 46a that has been determined to have a risk of progression. In the case of a machine learning-trained model 10, the treatment determination unit 44 inputs prediction features 46b, which consist of days of gestation, daily average weight, weight difference, weight_SD, weight_SD difference, daily average height, height difference, height_SD, height_SD difference, daily average respiratory rate, respiratory rate difference, daily average heart rate, heart rate difference, daily average arterial oxygen saturation, arterial oxygen saturation difference, Apgar score 5 minutes after birth, Apgar score 1 minute after birth, sex, birth type, and count, into the trained model 10.

[0065] When the premature infant prediction information 46a for premature infants determined to have a progression risk by the risk determination unit 43 is input to the trained model 10, the treatment determination unit 44 determines whether or not treatment for retinopathy of prematurity is indicated after 20 days of age, based on the output value of the trained model 10. More specifically, the treatment determination unit 44 determines that the infant will soon reach the treatment time range using retinal photocoagulation, etc. (treatment is indicated) if the value for "treatment" among the scores for "no treatment," "treatment with Type 1 ROP," and "treatment with APROP" is the highest (Figure 3, #40 Yes), and the notification unit 45 notifies the infant by a predetermined means (Figure 3, #41). In the example on the left of Figure 7, the score (AUC) for "treatment with APROP" was highest at about 3 weeks of age, so it is determined that treatment is indicated.

[0066] Figure 9 shows the retinopathy of prematurity progression prediction performance of a trained model 10 that was trained using machine learning with the aforementioned training feature 34b as input data at Hospital A (206 premature infants). Figure 10 also shows the retinopathy of prematurity progression prediction performance of a trained model 10 that was trained using deep learning with the aforementioned training premature infant information 34a as input data at Hospital A (206 premature infants). The validation data shown in Figure 9 was obtained by inputting the aforementioned prediction feature 46b into the trained model 10, and the retinopathy of prematurity progression prediction performance was expressed as a score (time-series AUC calculated from the ROC curve) for no treatment, treated Type 1 ROP, and treated APROP, respectively. The validation data shown in Figure 10 was obtained by inputting the aforementioned premature infant information 46a for prediction into the trained model 10, and the prediction performance of retinopathy of prematurity progression was expressed as a score (time-series AUC calculated from the ROC curve) for untreated, treated Type 1 ROP, and treated APROP, respectively. The upper part of Figure 9 shows the progression prediction performance (ROC curve) at 20 days postpartum when a trained model 10, which was machine-trained using 34b training features from Hospital A (206 premature infants), is input with 46b prediction features from Hospital A (206 premature infants). The lower part of Figure 9 shows the progression prediction performance (ROC curve) at 20 days postpartum when a trained model 10, which was machine-trained using 34b training features from Hospital A (206 premature infants), is input with 46b prediction features from Hospital B (59 premature infants). As shown in the upper part of Figure 9, the area under the ROC curve (AUC) for untreated patients at Hospital A was 0.69, the AUC for treated APROP was 0.82, and the AUC for treated Type 1 ROP was 0.58. These results show that progression can be predicted with good accuracy for progressive retinopathy of prematurity. Furthermore, as shown in the lower part of Figure 9, the area under the ROC curve (AUC) for no treatment at Hospital B was 0.66, the AUC for APROP with treatment was 0.83, and the AUC for Type 1 ROP with treatment was 0.58, which were almost the same as the progression prediction performance at Hospital A. From this, it can be concluded that the trained model 10, which was machine-learned using the training features 34b from Hospital A, is a highly versatile model capable of predicting the progression of retinopathy of prematurity at Hospital B. Figure 10 shows the progression prediction performance (time series data of AUC) when the trained model 10, which was deep-learned using the training premature infant information 34a from Hospital A (206 premature infants), is input with the prediction premature infant information 46a from Hospital B (59 premature infants). This is the time series prediction performance calculated backward from the day of treatment or discharge. As shown in the figure, the AUC is 0.8 or higher at least 50 days before treatment, indicating that cases with a high potential for progression to treatment-appropriate retinopathy of prematurity can be identified without delay. Therefore, the trained model 10, which was deep-trained using the premature infant information 34a from Hospital A, is a highly versatile model capable of predicting the progression of retinopathy of prematurity at Hospital B. Thus, this embodiment can predict the progression of retinopathy of prematurity with higher accuracy than existing models (WINROP and CHOP-ROP models). There are reports of applying existing models to various countries, but it has been found that there is a significant variation in accuracy from country to country. This is presumed to be due to differences in the level of medical care for neonatal management.The trained model 10 of this embodiment can be used in facilities with different neonatal care systems. As a result, it becomes possible to determine the indication for treatment and initiate treatment at the appropriate time, even for those who do not possess highly specialized knowledge and experience.

[0067] The trained model 10 in this embodiment is highly versatile and capable of accurately predicting the progression of retinopathy of prematurity at the appropriate time. Furthermore, since the risk of developing retinopathy of prematurity is extremely low when the sum of gestational age and gestational age at treatment is greater than 40 weeks, the trained model 10 in this embodiment, which is trained using information on premature infants whose sum of gestational age and gestational age at treatment is 40 weeks or less, can accurately predict the progression of retinopathy of prematurity.

[0068] Furthermore, by using a trained model 10 that outputs scores for no treatment, treated Type 1 ROP, and treated APROP, it is possible to predict the progression of retinopathy of prematurity by determining whether or not treatment is indicated after a predetermined number of days after birth. Many factors are intricately related to the progression process of retinopathy of prematurity from onset to indication of treatment, and these factors change over time. However, this embodiment uses multiple time-series data to predict the progression of retinopathy of prematurity with high accuracy. Moreover, in this embodiment, since the trained model 10 outputs the risk of progression of retinopathy of prematurity, it is possible to estimate the potential degree of progression of retinopathy of prematurity. And, since only premature infants determined to have a risk of progression are identified as being eligible for treatment, this is a highly accurate retinopathy of prematurity screening method with a two-stage determination process.

[0069] [Other embodiments] (1) The risk assessment step in which the trained model 10 outputs the risk of progression of retinopathy of prematurity may be omitted. Even in this case, the progression of retinopathy of prematurity can be accurately predicted by a treatment assessment step that uses the trained model 10, which outputs scores for no treatment, treated Type 1 ROP, and treated APROP, to determine whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth. (2) The trained model 10 may be generated by machine learning methods other than decision trees, or by deep learning methods other than convolutional neural networks. For example, known learning methods such as support vector machines and logistic regression can be used. (3) Premature infant information may include other parameters, provided that it includes postnatal time-series data on weight, height, and vital signs. [Industrial applicability]

[0070] This disclosure can be used in a retinopathy of prematurity screening method, screening device, and trained model for predicting the progression of retinopathy of prematurity. [Explanation of symbols]

[0071] 4: Screening device 10: Pre-trained model 34a: Information on premature infants for learning purposes (information on premature infants) 34b: Training features (features) 46a: Predictive premature birth information (premature birth information)

Claims

1. A method for operating a retinopathy of prematurity screening device for predicting the progression of retinopathy of prematurity, The aforementioned retinopathy screening device for prematurity includes a treatment determination unit, A method for operating a retinopathy of prematurity screening device, wherein the treatment determination unit operates to determine whether or not the infant is eligible for treatment of retinopathy of prematurity after a predetermined number of days after birth, based on premature infant information including postnatal time-series data regarding the weight, height, and vital signs of a premature infant whose gestational age is less than a predetermined number of weeks.

2. The retinopathy of prematurity screening device further comprises a risk determination unit, The risk assessment unit operates to determine the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth. If the risk determination unit determines that there is a risk of progression, the treatment determination unit is activated only for the premature infant, in the manner described in claim 1, the method for operating the retinopathy screening device for prematurity.

3. The method for operating a retinopathy of prematurity screening device according to claim 1 or 2, wherein the vital signs are at least one of the heart rate, respiratory rate, and arterial blood oxygen saturation of the premature infant.

4. A method for operating a retinopathy of prematurity screening device according to any one of claims 1 to 3, wherein the premature infant information includes at least one of the gestational days of the premature infant and the Apgar score.

5. The method for operating a retinopathy of prematurity screening device according to any one of claims 1 to 4, wherein the predetermined number of days after birth is one week or more and five weeks or less.

6. The method of operating the retinopathy of prematurity screening device according to claim 2, wherein the calculation of the score is performed every minute to every two days.

7. A screening device for predicting the progression of retinopathy of prematurity, A screening device equipped with a treatment determination unit that determines whether or not a premature infant is eligible for treatment of retinopathy of prematurity after a predetermined number of days after birth, based on premature infant information including postnatal time-series data on weight, height, and vital signs of premature infants born before a predetermined gestational age.

8. The system further includes a risk assessment unit that determines the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth. The screening device according to claim 7, wherein the treatment determination unit determines whether or not the treatment is appropriate for only the premature infants who have been determined by the risk determination unit to have a risk of progression.

9. A computer-trained model that functions as follows: It consists of a decision tree with multiple branching points arranged in a tree structure, A trained model that takes features calculated based on premature infant information, including postnatal time-series data on weight, height, and vital signs of premature infants born before a specified gestational age who received treatment for retinopathy of prematurity, as input, and outputs a score indicating whether or not treatment for retinopathy of prematurity is necessary by summing the evaluation values ​​at each of the aforementioned branching points.

10. A computer-trained model that functions as follows: Generated by deep learning, including convolutional neural networks, A pre-trained model that takes premature infant information, including postnatal time-series data on weight, height, and vital signs of premature infants born before a specified gestational age who received treatment for retinopathy of prematurity, as input and outputs a score indicating whether treatment for retinopathy of prematurity is necessary.

11. The trained model according to claim 9 or 10, wherein the premature infant information is information obtained from premature infants whose total gestational age and gestational age at treatment is 40 weeks or less.

12. The trained model according to any one of claims 9 to 11, wherein the vital signs are at least one of the heart rate, respiratory rate, and arterial blood oxygen saturation of the premature infant.

13. The pre-trained model according to any one of claims 9 to 12, wherein the premature infant information includes at least one of the gestational days and Apgar score of the premature infant.

Images