System and method for generating clinical score based on symbolic regression

Abstract

The present invention relates to a system for generating a clinical score based on symbolic regression. The system for generating a clinical score based on symbolic regression groups a plurality of responses into a plurality of response groups, assigns the same partial weight to responses included in the same response group, generates a plurality of partial weight groups, determines a final partial weight group by iteratively evaluating the partial weight groups, determining top-ranking partial weight groups, and generating new partial weight groups over a plurality of generations, and generates a score table on the basis of the final weight group.

Description

Symbolic Regression-Based Clinical Score Generation System and Method

[0001] The present invention relates to a clinical survey-based score generation system for clinical evaluation and disease risk measurement, and more specifically, to a clinical score generation system and method based on symbolic regression.

[0002] Various clinical questionnaires are utilized to assess and diagnose patients' disease risks. Self-report questionnaires serve as important tools for the rapid evaluation of a patient's condition. However, due to the large number of items in these questionnaires, respondent fatigue can accumulate during the response process. This fatigue can degrade the quality of the survey data. To address this issue of degraded data quality, abbreviated questionnaire techniques have been introduced, reducing the number of items to utilize only core questions. Recently, machine learning-based automated item selection and score calculation methods are being researched.

[0003] While machine learning-based scoring methods for clinical surveys can provide high predictive accuracy, they are considered "black boxes" with opaque internals, resulting in low reliability and difficult results interpretation. Although Explainable AI (XAI) offers tools to understand and interpret the prediction processes of AI models, it does not provide the same level of transparency as traditional self-report surveys (i.e., the uncompromised original questionnaires) that medical professionals can interpret intuitively.

[0004] Therefore, regarding clinical survey evaluation, a method is required that can provide high reliability and interpretability while utilizing abbreviated questionnaires.

[0005] One objective of the present invention is to provide a clinical score generation system and method that can facilitate result interpretation while having high prediction accuracy.

[0006] A symbolic regression-based clinical score generation system according to an embodiment of the present disclosure may include: a response grouping unit that groups multiple responses to each of a plurality of items included in a clinical survey into multiple response groups; a weighting unit that assigns equal partial weights to responses included in the same response group among the multiple responses and generates multiple partial weight groups; a weighting adjustment unit that evaluates the performance of multiple partial weight groups based on symbolic regression, determines upper partial weight groups among the multiple partial weight groups based on the evaluation results, generates new partial weight groups by performing genetic operations on the upper partial weight groups, and determines a final partial weight group by repeating the evaluation of partial weight groups, determination of upper partial weight groups, and generation of new partial weight groups over multiple generations; and a prediction unit that generates a score table including weights for each of the multiple responses to each of the multiple items based on the final partial weight group, receives response data regarding the clinical survey, and calculates a total score regarding the received response data based on the received response data and the score table.

[0007] A method for generating clinical scores based on symbolic regression according to an embodiment of the present disclosure comprises the steps of receiving input features and target features, classifying input features according to type, setting a threshold value for each of the input features, generating binary threshold variables based on the threshold value, generating a plurality of partial weight groups, grouping binary threshold variables, calculating a performance indicator for each of the plurality of partial weight groups, selecting upper partial weight groups based on the performance indicator among the plurality of partial weight groups, generating new partial weight groups from the upper partial weight groups based on genetic operations, determining whether to repeat, and selecting an optimal partial weight group in response to the determination result that the repetition will not be performed, and may perform the steps of grouping binary threshold variables or generating new partial weight groups in response to the determination result that the repetition will be performed.

[0008] The present invention can provide a symbolic regression-based clinical score generation system and method that can derive results with high prediction accuracy and facilitate result interpretation by automatically grouping response items based on symbolic regression and assigning weights to each group to calculate a total score.

[0009] FIG. 1 is a block diagram illustrating a symbolic regression-based clinical score generation system according to an embodiment of the present disclosure.

[0010] Figure 2 is a diagram showing the use of a shortened questionnaire.

[0011] FIG. 3 is a drawing showing an example of a shortened questionnaire according to an embodiment of the present disclosure.

[0012] FIG. 4 is a drawing for explaining grouping items according to an embodiment of the present disclosure.

[0013] FIG. 5 shows the weights for each of the plurality of items according to an embodiment of the present disclosure.

[0014] FIG. 6 shows the weights for each of the plurality of items according to an embodiment of the present disclosure.

[0015] FIG. 7 is a flowchart illustrating the process of grouping responses according to an embodiment of the present disclosure.

[0016] FIG. 8 is a diagram illustrating symbolic regression according to an embodiment of the present disclosure.

[0017] FIG. 9 is a drawing for explaining the operation of a prediction unit according to an embodiment of the present disclosure.

[0018] FIGS. 10a and FIGS. 10b illustrate an example of a symbolic regression-based clinical score generation algorithm according to an embodiment of the present disclosure.

[0019] FIGS. 11a to 11c are drawings showing score tables according to embodiments of the present disclosure.

[0020] FIG. 12 is a diagram comparing a symbolic regression-based clinical score generation system and a machine learning-based prediction model according to an embodiment of the present disclosure.

[0021] FIG. 13 is a block diagram showing a computing system implementing a symbolic regression-based clinical score generation system according to an embodiment of the present disclosure.

[0022] FIG. 1 is a drawing showing the best mode for carrying out the present invention.

[0023] In the following, embodiments of the present invention will be described clearly and in detail so that a person skilled in the art can easily practice the present invention.

[0024] FIG. 1 is a block diagram illustrating a symbolic regression-based clinical score generation system according to an embodiment of the present disclosure.

[0025] Referring to FIG. 1, a symbolic regression-based prize score generation system (100) may include a response grouping unit (110), a weighting unit (120), a weighting adjustment unit (130), and a prediction unit (140).

[0026] The response grouping unit (110) can group multiple responses to each of the multiple items included in the clinical questionnaire into response groups. The clinical questionnaire may be a self-report questionnaire. The clinical questionnaire may be a shortened questionnaire containing selected subsets of items (e.g., key items) from the total number of items included in the original questionnaire. The original questionnaire has m items has, and responses to each item It can be used to determine the risk or severity of a disease based on. The abbreviated questionnaire is a subset of the original questionnaire, consisting of s items smaller than m. Can have. Items of the shortened questionnaire Each response may be categorical or float. The response grouping unit (110) may classify multiple items into categorical or float types. For computational efficiency, the response grouping unit (110) shortens questionnaire items Each response can be grouped. Each response group is (i is 1, ..., s and k is 1, 2, ..., It can be expressed as (Im). is a question This is the number of response groups. If the number of response groups is large, the prediction accuracy may be improved, but the risk of overfitting may relatively increase and computational efficiency may decrease. The response grouping unit (110) can determine the number of response groups based on symbolic regression. Through symbolic regression, the response grouping unit (110) can determine the minimum number of response groups while preserving the prediction accuracy.

[0027] The weighting unit (120) can assign equal partial weights to responses included in the same response group among multiple responses. For example, the i-th response of a shortened questionnaire Weights for It can assign weights. The weight assignment unit (120) assigns weights to the i-th response of the shortened questionnaire. When assigning, the weights as in Equation 1 The sum of the responses in the original questionnaire can be assigned to be equal to or as close as possible to the sum of the responses in the original questionnaire items.

[0028]

[0029] is the weight of the response j for the i-th item of the shortened questionnaire, and is the lth response of the original questionnaire.

[0030] Weights for To assign weights, the weight assigning unit (120) assigns weights response group Corresponding partial weights It can be divided as. That is, weights as in mathematical formula 2 are partial weights It can be calculated as the sum of some or all of it.

[0031]

[0032] is the weight of the response j for the i-th item of the shortened questionnaire, and is the partial weight of the first response group for the i-th item, and is the partial weight of the second response group for the i-th item, and is the partial weight of the k-th response group for the i-th item. Here, response j is the response group It is a response grouped into.

[0033] Specifically, the response to the i-th item is j=3, and the response j=3 is the response group When grouped by, weight is the sum of partial weights It can be calculated as =. The response j=2 and the response j=3 are also response groups When grouped as, weight is the sum of partial weights = It can be calculated as follows. That is, responses included in the same response group have the same weight. When responses with the same weight are grouped, the efficiency of model fitting can be improved.

[0034] If the response value increases along with the disease severity sequence, the weight increases as the response j increases can increase. For monotonically increasing responses, partial weights are restricted to non-negative values, and thus as j increases It also guarantees that it increases. On the other hand, for responses that do not increase monotonically, partial weights may have negative values.

[0035] The weighting unit (120) is an initial partial weight group It can generate initial partial weight sets generated by the weighting unit (120). It can be optimized into final partial weight sets through symbolic regression techniques.

[0036] The weight adjustment unit (130) uses partial weights based on symbolic regression. It can determine. The weight adjustment unit (130) has a plurality of partial weight groups. Each performance can be evaluated. The weight adjustment unit (130) comprises multiple partial weight groups. Mean absolute error (MAE) can be used as a measure of fitness to evaluate each performance. The weight adjustment unit (130) comprises multiple partial weight groups Among them, a top weight group can be determined based on the performance evaluation results. The weight adjustment unit (130) uses the determined top weight group as the parent weight group and the new generation's partial weight groups. It can generate. The weight adjustment unit (130) can repeat the process of generating, evaluating, and determining top weight groups of partial weight groups over multiple generations. The weight adjustment unit (130) can perform the iterative process of generating, evaluating, and determining top weight groups of partial weight groups until the prediction performance is no longer improved. The weight adjustment unit (130) can determine the final partial weight group through iteration. The final partial weight group can be determined as the partial weight group evaluated as having the highest performance during the process of adjusting partial weights through symbolic regression.

[0037] As explained earlier, the weights per response are partial weights adjusted through the process of deriving the final set of partial weights. It can be calculated using... The prediction unit (140) is a plurality of items included in the shortened questionnaire. For each, multiple response-specific weights A score table containing can be generated. The maximum value of the total score on the score table (i.e., when responses with scores are selected) must be equal to the maximum total score of the original questionnaire. This condition can be expressed by Equation 3.

[0038]

[0039] The left side represents the sum of the maximum response values ​​for each item in the original questionnaire, and the right side represents the sum of the maximum weights in the shortened questionnaire. The sum of the maximum response values ​​is equal to the maximum total score of the original questionnaire, and the sum of the maximum weights is equal to the maximum total score of the shortened questionnaire.

[0040] To satisfy the conditions of Equation 3, the maximum total score of the original questionnaire in the symbolic regression process (i.e., Virtual artificial data points set as ) may be used. The artificial data points may be included in the training data set of the symbolic regression-based training process performed by the weight adjustment unit (130). Through this, the model trained through symbolic regression can learn how to effectively predict the maximum total score. However, even if artificial data points are added, the condition of Equation 3 is not always satisfied. The prediction unit (140) may adjust the weights in the regression operation to satisfy the condition of Equation 3.

[0041] The prediction unit (140) [determines] partial weights when there is a difference (greater than 0) between the maximum total score of the original questionnaire and the maximum total score of the shortened questionnaire calculated in the current model during the training process Any one of them can be corrected to increase by the corresponding difference. The prediction unit (140) is all partial weights For each, the performance can be evaluated by increasing it by the corresponding difference, and the combination with the highest performance can be selected. A score table can be generated based on the selected combination.

[0042] The prediction unit (140) receives response data regarding a clinical survey and can calculate a total score regarding the response data based on the received response data and a score table. The response data is data in which responses are checked for multiple items included in the clinical survey. That is, the response data corresponds to the results of checking responses in a shortened questionnaire. The prediction unit (140) can calculate the weights on the score table based on the response data A total score regarding the response data can be calculated using [this method]. The total score can be used to check the severity or presence of the disease based on the survey results.

[0043] The prediction unit (140) is the weights on the score table A total score threshold can be determined based on this. The prediction unit (140) can calculate the probability p based on Equation 4 using a sigmoid function.

[0044]

[0045] s is the number of items in the shortened questionnaire, p is the probability, is the item-specific weight, and b is the constant bias value. The probability can be set to a specific value. For example, the reference probability can be set to 0.5. In the classification task, the prediction unit (140) can calculate the total score threshold based on the inverse sigmoid function. The prediction unit (140) can calculate the threshold based on Equation 5.

[0046]

[0047] Thr is the total score threshold, 0.5 is the standard probability, and b is the constant bias value. The constant bias value b can be determined through the symbolic regression process.

[0048] The prediction unit (140) can compare the total score regarding the response data with the total score threshold. The prediction unit (140) can predict the original questionnaire results based on the total score regarding the response data. The prediction unit (140) can determine whether the total score regarding the response data is greater than or equal to the total score threshold and output the determination result. If it is determined that the total score regarding the response data is greater than or equal to the total score threshold, the response data can be predicted to have developed a disease.

[0049] A symbolic regression-based award score generation system (100) can automatically group survey responses using symbolic regression and assign partial weights optimized for each group. The score table produced by the symbolic regression-based award score generation system (100) is clear and intuitive. Therefore, it can be easily understood whether the weights for each survey response are used to predict severity.

[0050] Figure 2 is a diagram showing the use of a shortened questionnaire.

[0051] Referring to Figure 2, the shortened questionnaire (SQ) can be used to predict the diagnostic results of the original questionnaire (OQ).

[0052] The original questionnaire (OQ) has m items. , , … , It may include m items , , … , Each includes multiple responses. For example, items It can have n1 responses. Item It can have n2 responses. It can have nm of responses. Item inner questions Each response has a unique response value. The test subject has m items , , … , Each response can be checked. The sum of the response values ​​corresponding to each answer is the total score, which can be used to diagnose diseases. However, the original questionnaire contains a large number of items, which may lead to accumulated fatigue among respondents during the response process. This fatigue resulting from the response process can degrade the quality of the survey data.

[0053] The Short Questionnaire (SQ) consists of s items , , … , It may include s items of the Shortened Questionnaire (SQ). , , … , Each original questionnaire (OQ) has m items. , , … , It consists of selected important items. s items of the Short Questionnaire (SQ) , , … , m items of the original questionnaire (OQ) , … , It is a subset of ). s items , , … , Each has a unique weight. The sum of the weights corresponding to each response can be used to predict the diagnosis of a disease. Figure 2 illustrates the case where s is 3. That is, the test for the Shortened Questionnaire (SQ) is completed when the subject responds to only 3 items. The result of the Original Questionnaire (OQ) can be predicted solely by the sum of the weights of the responses to the 3 items. The higher the performance of the Shortened Questionnaire (SQ), the closer the total score calculated through the Shortened Questionnaire (SQ) is to the total score of the Original Questionnaire (OQ).

[0054] FIG. 3 is a drawing showing an example of a shortened questionnaire according to an embodiment of the present disclosure.

[0055] Referring to Fig. 3, multiple items included in the short questionnaire (SQ) , , … , It can be either a categorical response or a float response. For example, sleep multiple items , , … , Among these, difficulty falling asleep, difficulty staying asleep, degree of dissatisfaction with current sleep patterns, interference with daily life due to sleep disorders, and gender may be categorical responses. Multiple items , , … , Among them, age, weight, BMI, etc., may be real-valued responses.

[0056] FIG. 4 is a drawing for explaining grouping items according to an embodiment of the present disclosure.

[0057] Referring to Fig. 4, the item (ISI2) may have categorical responses from 0 to 4.

[0058] Item (ISI2) corresponds to an item included in the abbreviated questionnaire. For example, Item (ISI2) may be an item regarding the degree of dissatisfaction with the current sleep pattern. The test subject may check one of the responses 0 to 4 for Item (ISI2) depending on the intensity they feel.

[0059] Responses 0 through 4 of Item (ISI2) can be grouped into response groups (RG1, RG2, RG3). For example, responses 0 and 1 of Item (ISI2) can be grouped into response group (RG1), responses 2 and 3 into response group (RG2), and response 4 into response group (RG3). Each response group (RG1-RG3) has partial weights , , This can be granted.

[0060] Responses 0 through 4 of Item (ISI2) may have a monotonically increasing property. That is, the response values ​​of the original questionnaire from response 0 to response 4 may have an increasing pattern. The weights of each of responses 0 through 4 of Item (ISI2) , , , , Each can be calculated as the accumulation of partial weights. For example, the weights of Response 0 and Response 1 belonging to Response Group (RG1) , All are identical It may be. The weights of Response 2 and Response 3 belonging to Response Group (RG2) , All are identical It may be. The weight of response 4 belonging to response group (RG3) Is It may be. If the test taker checks response 3 for item (ISI2), the sum of the partial weights Weights calculated through It can be applied.

[0061] FIG. 5 shows the weights for each of the plurality of items according to an embodiment of the present disclosure.

[0062] Referring to Figure 5, the item Response 1 to Response 5, Item is response 1 to 4, item It can have a response of 0 to 4.

[0063] Question Responses 1 and 2 are included in response group (RG11), responses 3 and 4 are included in response group (RG12), and response 5 may be included in response group (RG13). Response group (RG11) includes partial weights A partial weight is assigned, and the response group (RG12) has partial weights. A partial weight is assigned, and the response group (RG13) has partial weights. ...can be assigned. Item The weights of Response 1 and Response 2 , is partial weight is, and the item The weights of responses 3 and 4 , is the sum of partial weights is, and the item The weight of response 5 is the sum of partial weights am.

[0064] Question Responses 1 through 3 are included in response group (RG21), and response 4 may be included in response group (RG22). Response group (RG21) includes partial weights A partial weight is assigned, and the response group (RG22) has partial weights. ...can be assigned. Item The weights of responses 1 through 3 , , is partial weight is, and the item The weight of response 4 is the sum of partial weights am.

[0065] Question Other items including [items] may be grouped, assigned partial weights, and weights determined by the sum of the partial weights in a manner similar to the method described above.

[0066] Question Responses 1 and 2 are included in response group (RG11), responses 3 and 4 are included in response group (RG12), and response 5 may be included in response group (RG13). Response group (RG11) includes partial weights A partial weight is assigned, and the response group (RG12) has partial weights. A partial weight is assigned, and the response group (RG13) has partial weights. ...can be assigned. Item The weights of Response 1 and Response 2 , is partial weight is, and the item The weights of responses 3 and 4 , is the sum of partial weights is, and the item The weight of response 5 is the sum of partial weights am.

[0067] FIG. 6 shows the weights for each of the plurality of items according to an embodiment of the present disclosure.

[0068] Referring to Fig. 6, the item (QF) can have real-valued items.

[0069] Item (QF) corresponds to an item included in the shortened questionnaire. For example, Item (QF) may be an item regarding the subject's weight. The subject may check their weight as real number data. For example, if the subject's weight is 80 kg, the response to Item (QF) may be 80 (kg).

[0070] Responses to the Question Function (QF) can be grouped according to a range. For example, the weight range of 0 kg to 120 kg can be grouped into response groups (RG1, RG2, RG3). For example, the range of the Question Function (QF) from 0 kg to less than 40 kg can be grouped into response group (RG1), the range from 40 kg to less than 70 kg can be grouped into response group (RG2), and the range from 70 kg to less than 120 kg can be grouped into response group (RG3). Each response group (RG1-RG3) has partial weights , , This can be granted.

[0071] The item (QF) can have a non-monotonic increasing property depending on each response group (RG1-RG3). That is, the weight of the item (QF) for each response jkg ( ) can increase or decrease depending on the range, or have a mixed pattern of increase and decrease. In other words, each response group (RG1-RG3) has partial weights , , It can have a range of positive or negative numbers, including 0. For example, , has a positive value, It can have a negative value.

[0072] Weights of 0kg to 120kg for the QF responses Each can be calculated as the accumulation of partial weights. For example, corresponding to the weights of responses of 0kg or more and less than 40kg belonging to response group (RG1). more All less than the same It may be. Corresponding to the weights of responses weighing 40kg or more and less than 70kg belonging to response group (RG2). more Less than is the sum of all equal partial weights It may be. Corresponding to the weights of responses weighing 70kg or more and less than 120kg belonging to response group (RG3). more Less than is the sum of all equal partial weights It could be. In this case, more Weights less than the specified value can have positive values, and more Weights below this level can have negative values. If the test subject checks 80kg as the response for Item (FQ), the weight for Item (FQ) is the sum of partial weights It could be.

[0073] FIG. 7 is a flowchart illustrating the process of grouping responses according to an embodiment of the present disclosure.

[0074] Referring to FIG. 7, a symbolic regression-based clinical score generation method may include a response grouping process. The response grouping process may include a step (S101) of receiving input features and target features. Input features and target feature y can correspond to the items of the shortened questionnaire and the survey results, respectively.

[0075] The response grouping process may include a step (S102) of classifying input features according to type. In step S102, input features may be classified as categorical variables or real-valued variables. For example, input features Among the input features is classified as a categorical variable, and other input features It can be classified as a real number variable.

[0076] The response grouping process may include a threshold setting step (S103). Step S103 is for integer responses for categorical variables. Set it, and for floating-point variables, continuous threshold value You can set it.

[0077] The grouping process of the response may include a step (S104) of generating a binary threshold variable. The binary threshold variable can be determined based on the following mathematical formula 6.

[0078]

[0079] is the i-th input feature It is a binary threshold variable indicating whether this response (or threshold) is greater than or equal to j, and is the i-th input feature This is an inverse binary threshold variable indicating whether the response (or threshold) is less than j.

[0080] The response grouping process may include a step (S105) of generating initial partial weight groups. In step S105, different partial weight groups may be randomly generated to assign weights to each binary threshold variable. For example, 20,000 partial weight groups may be generated in step S105.

[0081] The response grouping process may include a step (S106) of grouping binary threshold variables. In step S106, the same partial weight Binary threshold variables assigned can be grouped.

[0082] The response grouping process may include a step (S107) of calculating performance indicators. In step S107, performance indicators for each partial weight group may be calculated to evaluate the prediction performance of the partial weight groups.

[0083] The grouping process of the response may include a step (S108) of selecting a top partial weight group. In step S108, a pre-set proportion of top partial weight groups may be selected from among the partial weight groups according to performance indicators. According to one embodiment, the pre-set proportion may be the top 30%. The proportion may be less or greater than this.

[0084] The response grouping process may include a step (S109) of generating new generation sub-weight groups from upper weight groups based on genetic operations. In step S109, multiple sub-weight groups may be generated based on genetic operations. According to one embodiment, the genetic operations may include crossover, point mutation, host mutation, etc.

[0085] The grouping process of the response may include a step (S110) for determining whether to repeat. In step S110, the determination of whether to repeat may be based on whether the set maximum number of generations has been reached. In step S110, the determination of whether to repeat may also be based on whether the performance of the partial weight groups has improved compared to the performance of the partial weight groups of the previous generation.

[0086] In step S110, if it is determined to be a repeatable execution (S110-Yes), step S106 may be performed. Subsequently, steps S107, S108, and S109 may be performed in sequence.

[0087] In step S110, if it is determined that iteration is not necessary (S110-No), a step of selecting an optimal partial weight set (S111) may be performed. The optimal partial weight set may be determined as the partial weight set having the highest prediction performance among the partial weight sets of the final generation generated in step S109.

[0088] Steps S101 and S102 can be performed by the response grouping unit (110) of the symbolic regression-based clinical score generation system (100). Steps S103 through S105 can be performed by the weighting unit (120) of the symbolic regression-based clinical score generation system (100). Steps S106 and S111 can be performed by the weighting adjustment unit (130) of the symbolic regression-based clinical score generation system (100).

[0089] The final set of partial weights can be determined through the response grouping process. Through this, the set of partial weights is optimized, and a score table containing weights for each item can be generated based on the optimized partial weights.

[0090] FIG. 8 is a diagram illustrating symbolic regression according to an embodiment of the present disclosure.

[0091] Referring to Fig. 8, symbolic regression may include initialization of a set of candidate solutions, evolution of a set of upper solutions, and selection of a final solution.

[0092] In the initialization (A) of the candidate solution set, a set of input variables may be provided. For example, Input variables such as [list of variables] may be provided. Multiple initial candidate solutions may be generated by combining input variables using randomly selected mathematical operators (e.g., addition, multiplication, etc.).

[0093] For example, candidate solution 1 is With, candidate 2 is With , candidate solution n is It can have. The fitness of each candidate solution can be evaluated using performance metrics such as the Root Mean Squared Error (RMSE). Candidate solutions with relatively high performance are selected for evolution, while those with relatively low performance can be discarded.

[0094] Evolutionary operations (B) can be performed on the selected candidate solutions for evolution. Evolutionary operations may include crossover, mutation, etc. Crossover operations can generate a child solution with the characteristics of the parent solution by selecting and exchanging the subtrees of two candidate solutions. Mutation can introduce a new structure by randomly changing some nodes of the candidate solutions. Evaluation metrics can be calculated for the new solutions generated through evolutionary operations, and the performance of each candidate solution can be evaluated based on the evaluation metrics.

[0095] As an example of cross-operation, of the first generation and Through the intersection of 2 generations can be generated. As an example of a mutation operation, the first generation's Through mutations of the second generation can be generated.

[0096] The evolutionary operation can be performed over multiple generations. In each generation, performance evaluation is performed, specific candidate solutions are selected, and new candidate solutions can be generated. The evolutionary operation can be stopped when specific conditions are met. For example, it may terminate when the performance improvement of candidate solutions converges, a preset performance is achieved, or after a preset number of generations. The termination conditions for the evolutionary operation can be set in various ways.

[0097] After the evolutionary operation is stopped, a final solution (C) can be selected. The final solution can be selected as the combination with the best performance among the candidate solutions of the final generation.

[0098] Symbolic regression according to an embodiment of the present disclosure can be used by a symbolic regression-based granular score generation system (100) to determine the number of optimal response groups. Additionally, symbolic regression according to an embodiment of the present disclosure can be used by a symbolic regression-based granular score generation system (100) to determine a final partial weight group.

[0099] FIG. 9 is a drawing for explaining the operation of a prediction unit according to an embodiment of the present disclosure.

[0100] Referring to FIG. 9, the prediction unit (140) can perform a regression operation (RT) or a classification operation (CT).

[0101] The prediction unit (140) can correct some of the partial weights for fitting with the original questionnaire to the final partial weight group determined by the weight adjustment unit (130). The prediction unit (140) can perform a regression operation (RT) to correct the partial weights.

[0102] The prediction unit (140) can correct one of the final partial weights by increasing it by that difference if there is a difference between the maximum total score of the original questionnaire and the maximum total score of the model based on the final partial weight group when performing the regression operation (RT). The prediction unit (140) can perform correction on all partial weights.

[0103] The prediction unit (140) is the sum of the weights of the received responses ( A total score can be calculated based on ). The total score can be used for performance evaluation when performing calibration. Through calibration, the combination with the highest performance can be selected. A score table for the shortened questionnaire can be generated based on the calibrated partial weights.

[0104] The prediction unit (140) can perform a classification task. A threshold value (Thr) can be determined based on weights on a score table. The threshold value (Thr) can be determined by the method of FIG. 1. The prediction unit (140) sums the weights of the received responses ( ) and the threshold value (Thr) can be compared. Based on the comparison result, the presence or absence of disease can be predicted. For example, the sum of the weights of the received responses ( If ) is equal to or greater than the threshold value (Thr), it can be predicted that the disease has occurred based on the test results.

[0105] FIGS. 10a and FIGS. 10b illustrate an example of a symbolic regression-based clinical score generation algorithm according to an embodiment of the present disclosure.

[0106] Referring to FIGS. 10a and FIGS. 10b, the symbolic regression-based clinical score generation method may include a symbolic regression-based clinical score generation (SymScore) algorithm.

[0107] The symbolic regression-based clinical score generation (SymScore) algorithm is survey items Input can be received including a file containing target features and a task type. The task type can be selected as either a regression task or a classification task.

[0108] The SymScore algorithm, based on symbolic regression, can output a score table that assigns values ​​to each response. The score table may include weights for each response.

[0109] The SymScore algorithm based on symbolic regression may include an initialization step, a data preprocessing step, a conversion of predictor variables to binary threshold variables, a model training step, a check for overfitting, a cross-validation performance check using k-fold cross validation, and a table generation step.

[0110] In the initialization step, an input file containing responses to items is loaded, and it can be identified whether the task to be performed is a regression task or a classification task.

[0111] In the data preprocessing stage, predictor variables (e.g., key items) and response options for each item can be selected. In the data preprocessing stage, target columns (e.g., total scores or diagnoses) can be defined.

[0112] In the step of converting predictor variables into binary critical variables, a binary critical variable can be generated for each predictor variable. The binary critical variable indicates whether a threshold is exceeded.

[0113] In the model training phase, 70% of the total data can be split into a training set and 30% into a test set. When performing regression, additional artificial data can be generated to represent the maximum response value and included in the training.

[0114] In the model training phase, a symbolic regression model can be trained to satisfy Equation 7.

[0115]

[0116] The symbolic regression model is the sum of the scores from the original questionnaire It is trained to approximate. At this time, survey questions Response to It is the weight assigned to.

[0117] In the model training phase, weights Determines the weights is response group Each part weight related to It is divided into.

[0118] During the overfitting check phase, it is determined whether overfitting has occurred in regression or classification tasks. In regression, an overfitting warning may be triggered if the test set MEA exceeds a preset ratio (e.g., 1.5) of the training set MAE. In classification, an overfitting warning may be triggered if the AUROC difference between the test set and the training set exceeds a preset value (e.g., 0.1). Regarding regression, MAE stands for Mean Absolute Error, which is the average of the absolute differences between actual values ​​and values ​​predicted by the model. If the error on the training data is low but the error on the test data is large, it can be determined that the model has been trained too heavily on the training data and is unable to generalize to new data, indicating an overfitting situation. In this case, an overfitting warning may be triggered.

[0119] In the context of classification tasks, the AUROC (area under the receiver operating characteristic curve) is a graph representing the ratio of true positives to false positives, representing the area under the ROC curve. The closer the value is to 1, the higher the classification performance. However, if the difference in classification performance between the training and test data is greater than 0.1, it may indicate overfitting, as the model fits the training data but performs poorly on new data. In this case, an overfitting warning may be triggered.

[0120] The instability of the model can be determined during the cross-validation performance check step using k-fold validation. During regression, if the standard deviation of the cross-validation MAE exceeds a preset percentage (e.g., 15%) of the mean MAE, an instability warning may be generated. During classification, if the standard deviation of the cross-validation AUROC exceeds a preset value (e.g., 0.1), an instability warning may be generated.

[0121] An instability warning indicates that model performance varies significantly with each data split, leading to reduced reliability; there is a possibility that the model will perform well only on specific datasets while underperforming on others.

[0122] In the score table generation step, a score table containing weights for each response can be generated. For the classification task, a threshold value can be generated.

[0123] FIGS. 11a to 11c are drawings showing score tables according to embodiments of the present disclosure.

[0124] Referring to FIGS. 11a to 11c, a score table for obstructive sleep apnea (OSA), insomnia, and co-morbid insomnia and sleep apnea can be generated.

[0125] Obstructive sleep apnea, insomnia, and insomnia-induced sleep apnea can each be grouped into different response groups for the same questions. Additionally, weights (i.e., scores) for each response can also be assigned differently depending on the disease.

[0126] For example, for obstructive sleep apnea, for item ISI1b, a weight of 12 may be assigned to responses 0 to 1, a weight of 9 to response 2, and a weight of 0 to responses 3 to 4. For insomnia, for item ISI1b, a weight of 0 may be assigned to response 0, a weight of 4 to responses 1 to 2, and a weight of 12 to responses 3 to 4. For sleep apnea accompanied by insomnia, a weight of 0 may be assigned to responses 0 to 1 to item ISI1b, and a weight of 15 to responses 2 to 4.

[0127] When classifying, the threshold value may also vary by disease. For example, if the total score of the survey responses for obstructive sleep apnea is greater than or equal to 54, the subject may be predicted to have obstructive sleep apnea. If the total score of the survey responses for insomnia is greater than or equal to 57, the subject may be predicted to have insomnia. If the total score of the survey responses for sleep apnea accompanied by insomnia is greater than or equal to 62, the subject may be predicted to have sleep apnea accompanied by insomnia.

[0128] A symbolic regression-based clinical score generation system according to an embodiment of the present disclosure can generate a score table for each survey. The score table indicates the correlation of survey responses regarding diseases and can derive intuitive and transparent results when interpreting survey results.

[0129] FIG. 12 is a diagram comparing a symbolic regression-based clinical score generation system and a machine learning-based prediction model according to an embodiment of the present disclosure.

[0130] Referring to Fig. 12, a symbolic regression-based clinical score generation system can calculate weights representing scores for each item.

[0131] When machine learning (ML)-based models derive conclusions from input data, the process is considered a black box, making it impossible to verify the internal workings of the computation. Consequently, there are limitations in interpreting the basis for the judgments made when reaching conclusions.

[0132] A symbolic regression-based clinical score generation system according to an embodiment of the present disclosure has weights for each response ( - A score table including ) can be provided. The score table indicates how each response contributed to drawing the conclusion. Therefore, it can provide high reliability and interpretability compared to machine learning (ML).

[0133] FIG. 13 is a block diagram showing a computing system implementing a symbolic regression-based clinical score generation system according to an embodiment of the present disclosure.

[0134] Referring to FIG. 13, the computing system (1000) may include a processor (1010), memory (1020), an interface device (1030), and storage (1040).

[0135] The processor (1010) can control all operations, including data processing of the symbolic regression-based clinical score generation system (100). The processor (1010) can execute firmware or software loaded into memory (1020). The processor (1010) may include at least one general-purpose processor, such as a central processing unit (CPU) or an application processor (AP). The processor (1010) may also include at least one special-purpose processor, such as a neural processing unit (NPU), a neuromorphic processor, or a graphics processing unit (GPU). The processor (1010) may include two or more processors of the same type.

[0136] The memory (1020) can store codes and instructions executed by the processor (1010). The memory (1020) can store data processed by the processor (1010). For example, the memory (1020) can store various data such as input data, output data, a symbolic regression-based clinical score generation model, and a score table. The memory (1020) may include non-volatile memory including at least one of volatile memory such as RAM (Random Access Memory) or SRAM (Static Random Access Memory), flash memory type, hard disk type, multimedia card micro type, card type memory (e.g., SD or XD memory), ROM (Read-Only Memory), and EEPROM (Electrically Erasable Programmable Read-Only Memory) and PROM (Programmable Read-Only Memory).

[0137] The interface circuit (1030) can provide signal or data communication between the computing system (1000) and the outside. For example, the processor (1010) can receive training data or test data, etc. from the outside (e.g., a host) through the interface device (1030). The interface device (1030) can output a score sheet, disease diagnosis results, etc.

[0138] The storage (1040) may further include storage for storing various data. The storage (1040) may load programs, models, data, etc. stored in the storage into memory (1020). The storage may include the non-volatile memory described above.

[0139] The present invention can provide a symbolic regression-based clinical score generation system and a method of operation thereof that can derive results with high prediction accuracy and facilitate result interpretation by automatically grouping response items based on symbolic regression and assigning weights to each group to calculate a total score.

[0140] The above description describes specific embodiments for implementing the present invention. The present invention will include not only the embodiments described above, but also embodiments that can be simply modified or easily modified. Furthermore, the present invention will include technologies that can be easily modified and implemented using the embodiments. Accordingly, the scope of the present invention should not be limited to the embodiments described above, but should be defined by the claims set forth below as well as equivalents to the claims of this invention.

[0141] The present disclosure relates to a clinical questionnaire-based scoring system for clinical evaluation and disease risk measurement. More specifically, a clinical scoring system based on symbolic regression is available for use in the method.

Claims

1. A response grouping unit that groups multiple responses to each of the multiple items included in a clinical questionnaire into multiple response groups; A weighting unit that assigns equal partial weights to responses included in the same response group among the plurality of responses above, and generates a plurality of partial weight groups; A weight adjustment unit that evaluates the performance of the plurality of partial weight groups based on symbolic regression, determines upper partial weight groups among the plurality of partial weight groups based on the evaluation results, generates new partial weight groups by performing genetic operations on the upper partial weight groups, and determines a final partial weight group by repeating the evaluation of the partial weight groups, the determination of the upper partial weight groups, and the generation of the new partial weight groups over a plurality of generations; and A symbolic regression-based clinical score generation system comprising a prediction unit that generates a score table including weights for each of a plurality of responses to each of a plurality of items based on the above final partial weight group, receives response data regarding the clinical survey, and calculates a total score regarding the received response data based on the received response data and the score table.

2. In Paragraph 1, The above response grouping unit is a symbolic regression-based clinical score generation system that determines the number of response groups from the plurality of responses based on symbolic regression.

3. In Paragraph 1, A symbolic regression-based clinical score generation system in which each of the above multiple items is classified as categorical or real.

4. In Paragraph 1, A symbolic regression-based clinical score generation system in which the weight adjustment unit determines the partial weight groups corresponding to the top 30% of the evaluation results among the plurality of partial weight groups as the upper partial weight groups.

5. In Paragraph 1, The above clinical questionnaire is a symbolic regression-based clinical score generation system that is a shortened questionnaire of the original questionnaire.

6. In Paragraph 5, The above prediction unit is a symbolic regression-based clinical score generation system that performs a regression operation to correct the partial weights included in the final partial weight group so that the maximum total score of the original questionnaire and the maximum total score of the shortened questionnaire are the same.

7. In Paragraph 6, A symbolic regression-based clinical score generation system that divides 70% of the total data into a training set and the remaining 30% into a test set, and generates an overfitting warning in response to the average absolute error of the training set exceeding a preset ratio of the average absolute error of the test set during the regression operation.

8. In Paragraph 7, The above prediction unit is a symbolic regression-based clinical score generation system that generates an instability warning in response to the standard deviation of the cross-validation mean absolute error exceeding a preset multiplier of the mean absolute error during the above regression operation.

9. In Paragraph 1, The above prediction unit is a symbolic regression-based clinical score generation system that performs a classification task to calculate a total score threshold based on the inverse sigmoid function.

10. In Paragraph 9, A symbolic regression-based clinical score generation system that compares the total score calculated above with the total score threshold value and predicts the original survey result based on the comparison result.

11. In Paragraph 9, A symbolic regression-based clinical score generation system that divides 70% of the total data into a training set and the remaining 30% into a test set, and generates an overfitting warning in response to the difference between the AUROC (area under the receiver operating characteristic curve) of the training set and the AUROC of the test set exceeding a preset value during a classification task.

12. In Paragraph 11, The above prediction unit is a symbolic regression-based clinical score generation system that generates an instability warning in response to the standard deviation of the cross-validation AUROC being greater than 0.

1.

13. A step of receiving input features and target features; A step of classifying the above input features according to type; A step of setting a threshold value for each of the above input features; A step of generating binary threshold variables based on the above threshold value; A step of generating multiple partial weight groups; A step of grouping the above binary threshold variables; A step of calculating a performance indicator for each of the plurality of partial weight groups; A step of selecting upper partial weight groups based on performance indicators among the plurality of partial weight groups mentioned above; A step of generating new partial weight groups from the upper partial weight groups based on genetic operations; Step for determining whether to repeat; The above judgment result includes a step of selecting an optimal partial weight group in response to not repeating the process, and A symbolic regression-based clinical score generation method that performs the steps of grouping the binary threshold variables or generating the new partial weight groups in response to the above judgment result being repeated.

14. In Paragraph 13, The above input features are a symbolic regression-based clinical score generation method classified into either a categorical type or a real type.

15. In Paragraph 13, The above genetic operation is a symbolic regression-based clinical score generation method including crossover and mutation.

16. In Paragraph 13, A symbolic regression-based clinical score generation method further comprising the step of generating a score table based on the above-mentioned optimal partial weight group.

17. In Paragraph 13, The above input features are a symbolic regression-based clinical score generation method in which the responses correspond to each item of a shortened questionnaire from which only the core items are extracted from the original questionnaire.

18. In Paragraph 17, A symbolic regression-based clinical score generation method further comprising the step of correcting partial weights included in the optimal partial weight group so that the total sum of maximum scores of the original questionnaire and the total sum of maximum scores of the shortened questionnaire are equal.

19. In Paragraph 13, A symbolic regression-based clinical score generation method that further includes a step of calculating a total score threshold based on an inverse sigmoid function.

20. In Paragraph 19, Step of receiving response data; A symbolic regression-based clinical score generation method further comprising the step of predicting whether onset occurs in response to the total score regarding the above response data being greater than or equal to the total score threshold.

Images