TWI836884B - Artificial Intelligence Prediction Method for Uric Acid Stones - Google Patents

Abstract

An artificial intelligence uric acid stone prediction method comprises: establishing an artificial intelligence uric acid stone prediction system, the artificial intelligence uric acid stone prediction system comprises a machine learning prediction model including a plurality of model input variables and a model output variable, the model input variables include a glomerular filtration rate estimation value and a urine acid-base value, the model output variable indicates the probability that the kidney stone component is uric acid crystals; inputting a plurality of model input values of a patient suffering from kidney stones corresponding to the model input variables respectively into the machine learning prediction model; and the machine learning prediction model calculates a model output value corresponding to the model output variable according to the model input values, the model output value indicates the probability that the patient's kidney stone component is uric acid crystals.

Description

Artificial intelligence uric acid stone prediction method

The present invention relates to a stone prediction method, in particular to an artificial intelligence (AI) uric acid stone prediction method.

Uric acid kidney stones account for approximately 10% to 15% of all kidney stones. Unlike other types of kidney stones, uric acid stones are associated with several features of metabolic syndrome and nutrient distribution disorders, including diabetes and obesity. In addition, unlike other types of kidney stones that require surgery, most uric acid stones can be treated conservatively. Therefore, when it is known that a patient has kidney stones, before treatment, it is important to further distinguish whether the kidney stones are pure uric acid stones or non-uric acid stones.

At present, relevant research reports have pointed out that dual-energy computed tomography (DECT) scans can be used to accurately distinguish whether kidney stones are pure uric acid stones or non-uric acid stones before treating kidney stones. However, DECT scans produce a very high radiation dose and the cost of the equipment is relatively high. Therefore, it is necessary to find a solution.

Therefore, the purpose of the present invention is to provide an artificial intelligence method for predicting uric acid stones.

Therefore, the artificial intelligence uric acid stone prediction method of the present invention includes: (A) establishing an artificial intelligence uric acid stone prediction system, wherein the artificial intelligence uric acid stone prediction system includes a machine learning including several model input variables and a model output variable Prediction models. The input variables of these models include an estimated value of glomerular filtration rate and a urine pH value. The output variables of the model represent the probability that the component of kidney stones is uric acid crystals; (B) a patient suffering from kidney stones. A number of patient model input values corresponding to the model input variables are input to the machine learning prediction model; and (C) the machine learning prediction model calculates a corresponding model output based on the model input values. The model output value of the variable represents the probability that the patient's kidney stones are composed of uric acid crystals.

Another purpose of the present invention is to provide another artificial intelligence uric acid stone prediction method.

Therefore, the artificial intelligence uric acid stone prediction method of the present invention comprises: (a) establishing an artificial intelligence uric acid stone prediction system, wherein the artificial intelligence uric acid stone prediction system comprises a plurality of system input variables, and a machine learning prediction model comprising a plurality of model input variables and a model output variable, wherein the system input variables comprise a serum creatinine concentration, a gender, an age, and a urine acid-base value, wherein the model input variables at least comprise the urine acid-base value , the model output variable represents the probability that the kidney stone is composed of uric acid crystals; (b) a plurality of system input values of a patient suffering from kidney stones corresponding to the system input variables are input into the artificial intelligence uric acid stone prediction system; and (c) the machine learning prediction model calculates a model output value corresponding to the model output variable according to the system input values, wherein the model output value represents the probability that the patient's kidney stones are composed of uric acid crystals.

The functions of the present invention are: (1) As long as two easily obtained clinical variables (estimated glomerular filtration rate and urine pH) of patients suffering from kidney stones are used as the machine learning prediction model By inputting variables into the model, the probability that the patient's kidney stones are composed of uric acid crystals can be calculated, allowing medical personnel to determine whether the patient's kidney stones are uric acid stones and to prescribe appropriate treatment as soon as possible; (2) as long as The serum creatinine concentration, gender, age and urine pH of patients with kidney stones are input into the artificial intelligence uric acid stone prediction system as system input values, and the machine learning prediction model can output the model output value. It indicates the probability that the kidney stone component is uric acid crystals, which can be used by medical personnel to determine whether the patient's kidney stone type is uric acid stone, so as to prescribe symptomatic treatment as soon as possible.

2: Artificial intelligence uric acid stone prediction system

21: Estimated glomerular filtration rate (eGFR) calculation unit

22: Machine learning prediction model

S91~S93: steps

Other features and effects of the present invention will be clearly presented in the embodiments with reference to the drawings, in which: Figure 1 is a block diagram illustrating an artificial intelligence (AI) uric acid stone prediction system for implementing some embodiments of the present invention's artificial intelligence (AI) uric acid stone prediction method, wherein the artificial intelligence uric acid stone prediction system includes a device for calculating a kidney Estimated glomerular filtration rate (eGFR) estimated glomerular filtration rate calculation unit, and a machine learning prediction model; Figure 2 is a block diagram illustrating other implementations of the artificial intelligence uric acid stone prediction method of the present invention For example, the artificial intelligence uric acid stone prediction system includes the renal filament estimated value calculation unit; Figure 3 is a block diagram illustrating that the renal filament in Figures 1 and 2 is not set Variations of the ball filtration rate estimation value calculation unit; Figure 4 is a flow chart illustrating the main embodiment of the artificial intelligence uric acid stone prediction method of the present invention; Figure 5 is a graph illustrating the artificial intelligence uric acid stone prediction method of the present invention. In the first embodiment, when several model input variables include an estimated value of glomerular filtration rate and a urine pH value, the machine learning prediction model applies the verification receiver operating characteristic obtained from the verification data of Group A. (Receiver Operating Characteristic, ROC) curve; Figure 6 is a graph illustrating the verification ROC curve obtained when the machine learning prediction model is applied to Group B verification data in the first embodiment; Figure 7 is a graph, It is explained that in the first step of the artificial intelligence uric acid stone prediction method of the present invention In the second embodiment, when the model input variables include the estimated glomerular filtration rate, the urine pH, and a body mass index (BMI), the machine learning prediction model is applied to the validation data of Group A. The obtained verification ROC curve; Figure 8 is a graph illustrating the verification ROC curve obtained when the machine learning prediction model is applied to Group B verification data in the second embodiment; Figure 9 is a graph illustrating the verification ROC curve obtained in this second embodiment. In the third embodiment of the invention of the artificial intelligence uric acid stone prediction method, when the model input variables include an age, the estimated glomerular filtration rate, the urine pH, and the body mass index, the machine learning The verification ROC curve obtained when the prediction model is applied to the verification data of Group A; Figure 10 is a graph illustrating the verification ROC curve obtained when the machine learning prediction model is applied to the verification data of Group B in the third embodiment; Figure 11 is a graph illustrating that in the fourth embodiment of the artificial intelligence uric acid stone prediction method of the present invention, when the model input variables include a gender, the age, the estimated glomerular filtration rate, and the urine acid-base value, and the body mass index, the verification ROC curve obtained when the machine learning prediction model is applied to group A verification data; Figure 12 is a graph illustrating that in the fourth embodiment, the machine learning prediction model is applied to The verification ROC curve obtained when verifying data in Group B; Figure 13 is a graph illustrating that in the fifth embodiment of the artificial intelligence uric acid stone prediction method of the present invention, when the model input variables include the gender, the age, the kidney Estimated silk ball filtration rate, urine pH, body mass index, and whether there is diabetes The validation ROC curve obtained when the machine learning prediction model is applied to the verification data of Group A when the disease history is determined; Figure 14 is a graph illustrating that in the fifth embodiment, the machine learning prediction model is applied to the verification data of Group B The obtained verification ROC curve; Figure 15 is a graph illustrating that in the sixth embodiment of the artificial intelligence uric acid stone prediction method of the present invention, when the model input variables include the gender, the age, and the glomerular filtration rate estimate value, the urine pH, the body mass index, whether there is a history of diabetes, whether there is a history of gout, and whether there is a history of bacteriuria, the validation ROC curve obtained when the machine learning prediction model is applied to the validation data of Group A ; Figure 16 is a graph illustrating the verification ROC curve obtained when the machine learning prediction model is applied to Group B verification data in the sixth embodiment; Figure 17 is a graph illustrating the artificial intelligence uric acid stones of the present invention In the seventh embodiment of the prediction method, when the model input variables include the gender, the age, the estimated glomerular filtration rate, the urine pH, the body mass index, whether there is a history of diabetes, whether there is The validation ROC curve obtained when the machine learning prediction model is applied to the validation data of Group A when there is a history of gout, a history of bacteriuria, and a history of hypertension; and Figure 18 is a graph illustrating the seventh implementation In this example, the validation ROC curve obtained when the machine learning prediction model is applied to Group B validation data.

Referring to Figures 1 and 4, the first embodiment of the artificial intelligence uric acid stone prediction method of the present invention includes the steps shown in the flowchart of Figure 4. First, as shown in step S91 of Figure 4, an artificial intelligence uric acid stone prediction system 2 as shown in Figure 1 is established. Among them, the artificial intelligence uric acid stone prediction system 2 includes a plurality of system input variables, an estimated glomerular filtration rate (eGFR) calculation unit 21 for calculating an estimated glomerular filtration rate (eGFR), and a machine learning prediction model 22 including a plurality of model input variables and a model output variable. Among them, the model output variable represents the probability that the kidney stone component is uric acid crystals.

It should be noted that although one of the input variables x ₅ is shown in Figure 1 as body mass index (BMI), this input variable x ₅ is only used in the second to seventh embodiments (which will be described in detail later). The input variable will be used. Therefore, the first embodiment will be described in detail below while temporarily ignoring the input variable x ₅ in FIG. 1 . That is, in this first embodiment, the system input variables only include x ₀ = "serum creatinine concentration (Scr)", x ₁ = "gender", x ₂ = "age", and x ₄ = "Urine pH".

In the first embodiment, the eGFR calculation unit 21 calculates one of the model input variables x 3 = "eGFR" based on the system input variables x ₀ = "serum creatinine concentration", x ₁ = "sex", x ₂ = "age" using the following formula for isotope dilution mass spectrometry traceable modification of Diet in Renal Disease, so that in the first embodiment, the model input variables include x ₃ = "eGFR", and x ₄ = "urine acid-base value": eGFR [mL/min/1.73m ² ] = 175 × (serum creatinine concentration) ^-1.154 _× (age) ^-0.203 × [0.742 if female].

其中，x代表該等模型輸入變數所形成之向量，而y則是該機器學習預測模型22的輸出變數，其值介於0與1之間，代表「腎結石成分為尿酸結晶的機率」。在本第一實施例中，函數f(x)具有下列形式： f(x)=w₀+w₁(x)f₁(x)+w₂(x)f₂(x)+...+w_n(x)f_n(x)，其中，w₀為一常數參數，而w₁(x)、w₂(x)、...、w_n(x)與f₁(x)、f₂(x)、...、f_n(x)則皆為以全連通之人工神經網路(ANN)所建構之函數。 In this first embodiment, the structure of the machine learning prediction model 22 is as follows:

Among them, x represents the vector formed by the input variables of the model, and y is the output variable of the machine learning prediction model 22, with a value between 0 and 1, representing "the probability that the kidney stone component is uric acid crystals." In this first embodiment, the function f(x) has the following form: f(x)=w ₀ +w ₁ (x)f ₁ (x)+w ₂ (x)f ₂ (x)+... +w _n (x)f _n (x), where w ₀ is a constant parameter, and w ₁ (x), w ₂ (x),..., w _n (x) and f ₁ (x), f ₂ (x),..., f _n (x) are all functions constructed by fully connected artificial neural networks (ANN).

In this first embodiment, the construction of the machine learning prediction model 22 is based on the patient data of Group A (Cohort A) suffering from kidney stones, the patient data of Group B (Cohort B) suffering from kidney stones, etc. Two groups of patient data. Among them, the patient data in Group A are divided into a training set (60%) for model training and a validation set (40%) for model validation.

Among them, during the construction process of the machine learning prediction model 22, patients in group A The patient data includes the data of 1098 patients, and 146 of the 1098 patients (13.3%) suffer from pure uric acid stones, while the patient data of Group B includes the data of 71 patients, and the 71 patients Three of the patients (4.23%) suffered from pure uric acid stones. In addition, by the way, infrared spectroscopy can be used to analyze the surgically removed stones of patients in Group A and Group B suffering from kidney stones to determine whether the kidney stones are pure uric acid stones or non-uric acid stones. Among them, in this first embodiment, pure uric acid stones are defined as 1, and non-uric acid stones are defined as 0.

In this first embodiment, the calculation results obtained by calculating the prevalence of pure uric acid stone patients and non-uric acid stone patients in group A patient data show that the statistical difference between the training set and the validation set of group A patient data is very small. Group B patient data is only used for validation.

Therefore, after the machine learning prediction model 22 is generated based on the training set of patient data of group A, the area under the ROC curve (AUC) analysis method of the receiver operating characteristic (ROC) curve can be further used to obtain the optimal threshold values for predicting negative and positive based on the Youden’s index. Then, based on the obtained optimal threshold values, the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the machine learning prediction model 22 are calculated to evaluate the model performance of the machine learning prediction model 22. Then, the machine learning prediction model 22 is applied to the validation set of patient data of group A and patient data of group B to verify whether the machine learning prediction model 22 can correctly identify patients with uric acid stones.

In the first embodiment, since the model input variables are only x ₃ = "eGFR" and x ₄ = "urine acid-base value", the function f(x) can be simplified to the following form: f(x)=w ₀ +w ₃ f ₃ (x ₃ )+w ₄ f ₄ (x ₄ ), where w ₀ , w ₃ , and w ₄ are constant parameters, and the functions f ₃ (x ₃ ) and f ₄ (x ₄ ) are both nonlinear functions constructed by a fully connected two-layer neural network. The output dimension of the first layer of the two-layer neural network is _mi (i=3,4), where _mi is an integer between 1 and 100, and the output dimension of the second layer is 1. The functions f ₃ (x ₃ ) and f ₄ (x ₄ ) are of the following forms respectively: _fi ( _xi )=σi _,2 (Ai _{,2σi ,1} (Ai _,1xi + _{Bi ,1} )+Bi _,2 ),i=3,4.

個可訓練模型參數。在本第一實施例中，所使用之m_i(i=3,4)皆為20，且B_i,1(i=3,4)皆設為零，共85個可訓練模型參數。 Among them, A _i,1 and B _i,1 (i=3,4) are vector variables of m _i -by-1, and each variable includes m _i adjustable parameters. And A _i,2 (i=3,4) are all 1-by- _mi vector variables, and each variable also includes m _i adjustable parameters. B _3,2 and B _4,2 are both scalar variables. σ _i,1 and σ _i,2 (i=3,4) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

trainable model parameters. In this first embodiment, the m _i (i=3,4) used are all 20, and the Bi _,1 (i=3,4) are all set to zero, resulting in a total of 85 trainable model parameters.

Referring to Figures 4, 5, and 6, when the machine learning prediction model 22 established in step S91 of the first embodiment is applied to the validation set of patient data of group A, its area under the ROC Curve (AUC) can be calculated to be 0.8213, as shown in Figure 5. Similarly, when the machine learning prediction model 22 is applied to patient data of group B, its area under the ROC Curve (AUC) can be calculated to be 0.8382, as shown in Figure 6.

Next, as shown in step S92 of FIG. 1 and FIG. 4 , in the first embodiment, if one wishes to know the probability that the kidney stones of a patient suffering from kidney stones are composed of uric acid crystals, one must first input the patient's four system input values corresponding to the system input variables x ₀ = "serum creatinine concentration", x ₁ = "sex", x ₂ = "age" and x ₄ = "urine acid-base value" into the artificial intelligence uric acid stone prediction system 2, and then the eGFR calculation unit 21 calculates the patient's eGFR based on the system input values x ₀ , x ₁ , x ₂ using the aforementioned eGFR formula, and uses it as the model input value x ₃ .

Then, as shown in step S93, the machine learning prediction model 22 can calculate a model output value corresponding to the model output variable y based on the patient's model input values x ₃ = "eGFR" and x ₄ = "urine acid-base value", which can represent the probability that the patient's kidney stones are composed of uric acid crystals.

Referring to Figures 1, 7, and 8, in a second embodiment of the present invention, the model input variables include x ₃ = "eGFR", x ₄ = "urine pH value", and x ₅ = "body mass" Index (BMI)", so the function f(x) can be simplified to the following form: f(x)=w ₀ +w ₃ f ₃ (x ₃ )+w ₄ f ₄ (x ₄ )+w ₅ f ₅ (x ₅ ), where w ₀ , w ₃ , w ₄ , and w ₅ are constant parameters, and the functions f ₃ (x ₃ ), f ₄ (x ₄ ), and f ₅ (x ₅ ) are all fully connected two-layer Nonlinear functions constructed by neural networks. The output dimension of the first layer of the two-layer neural network is m _i (i=3 to 5), where m _i is an integer between 1 and 100, and the output dimension of the second layer is 1. The functions f ₃ (x ₃ ) to f ₅ (x ₅ ) respectively have the following forms: f _i (x _i )=σ _i,2 (A _i,2 σ _i,1 (A _i,1 x _i +B _{i, 1} )+B _i,2 ),i=3,4,5.

個可訓練模型參數。在本第二實施例中，所使用之m_i(i=3至5)皆為20，且B_i,1(i=3至5)皆設為零，共127個可訓練模型參數。 Among them, A _i,1 and B _i,1 (i=3 to 5) are vector variables of m _i -by-1, and each variable includes m _i adjustable parameters. And A _i,2 (i=3 to 5) are all 1-by-m _i vector variables, each of which also includes m _i adjustable parameters. B _3,2 to B _5,2 are all scalar variables. σ _i,1 and σ _i,2 (i=3 to 5) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

trainable model parameters. In this second embodiment, the m _i (i=3 to 5) used are all 20, and the Bi _,1 (i=3 to 5) are all set to zero, resulting in a total of 127 trainable model parameters.

Referring to Figures 4, 7, and 8, when the machine learning prediction model 22 established in step S91 of the second embodiment is applied to the verification set of group A patient data, its area under the curve ( AUC) is 0.8238, as shown in Figure 7. Similarly, when the machine learning prediction model 22 is applied to the patient data of Group B, the area under the curve (AUC) can be calculated to be 0.848, as shown in Figure 8 .

Next, as shown in step S92 of Figures 1 and 4, in this second embodiment, if one wants to know the probability that the kidney stones of a patient suffering from kidney stones are composed of uric acid crystals, the five system input values of the patient corresponding to the system input variables _x0 = "serum creatinine concentration", _x1 = "gender", _x2 = "age", _x4 = "urine acid-base value" and _x5 = "BMI" are first input into the artificial intelligence uric acid stone prediction system 2, and then the eGFR calculation unit 21 calculates the eGFR of the patient according to the system input values _x0 , _x1 , _x2 using the aforementioned eGFR formula.

Then, as shown in step S93, in the second embodiment, the machine learning prediction model 22 can input values x ₃ = "eGFR" and x ₄ = "urine pH" according to the model input values of the patient. , and x ₅ = "BMI", calculate the model output value y, which represents the probability that the patient's kidney stones are composed of uric acid crystals.

Referring to Figure 2, in a third embodiment of the present invention, the model input variables also include x ₂ = "age". It should be noted that although the system input variable x ₁ "gender" is shown in Figure 2, it can also be used as a model input variable, and the input variables x ₆ "whether there is a history of diabetes" and x ₇ "whether "Do you have a history of gout", x ₈ "Do you have a history of bacteriuria", x ₉ "Do you have a history of hypertension", but the model input variable x ₁ "gender" and the input variables x ₆ ~ x ₉ are in other embodiments (later (will be described in detail later), therefore, the third embodiment will be described in detail below while temporarily ignoring the system input variables x ₁ and x ₆ ~x ₉ . That is, in this third embodiment, the system input variables include x ₀ = "serum creatinine concentration", x ₁ = "gender", x ₂ = "age", x ₄ = "urine pH value" ”, and x ₅ = “BMI”, and the model input variables include x ₂ = “age”, x ₃ = “eGFR”, x ₄ = “urine pH”, and x ₅ = “BMI” .

In the third embodiment, the function f(x) can be simplified to the following form: f(x)=w ₀ +w ₂ f ₂ (x ₂ )+w ₃ f ₃ (x ₃ )+w ₄ f ₄ (x ₄ )+w ₅ f ₅ (x ₅ ), where w ₀ , w ₂ ,...,w ₅ are constant parameters, and the functions f ₂ (x ₂ ),...,f ₅ (x ₅ ) are all nonlinear functions constructed by a fully connected two-layer neural network. The output dimension of the first layer of the two-layer neural network is _mi (i=2 to 5), where _mi is an integer between 1 and 100, and the output dimension of the second layer is 1. Functions f ₂ (x ₂ ) to f ₅ (x ₅ ) are of the following forms respectively: _fi ( _xi )=σi _,2 (Ai _{, 2σi,1} (Ai _,1xi + _{Bi ,1} )+Bi _,2 ),i=2,3,4,5.

個可訓練模型參數。在本第三實施例中，所使用之m_i(i=2至5)皆為20，且B_i,1(i=2至5)皆設為零，共169個可訓練模型參數。 Among them, _Ai,1 and _Bi,1 (i=2 to 5) are _mi -by-1 vector variables, each of which includes _mi adjustable parameters. _Ai,2 (i=2 to 5) are 1-by- _mi vector variables, each of which also includes _mi adjustable parameters. B _2,2 to B _5,2 are all scalar variables. σ _i,1 and σ _i,2 (i=2 to 5) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

In the third embodiment, _mi (i=2 to 5) used is 20, and Bi _,1 (i=2 to 5) is set to zero, for a total of 169 trainable model parameters.

Referring to Figures 2, 4, 9, and 10, when the machine learning prediction model 22 established in step S91 of the third embodiment is applied to the validation set of patient data of group A, its area under the curve (AUC) can be calculated to be 0.8253, as shown in Figure 9. Similarly, when the machine learning prediction model 22 is applied to patient data of group B, its area under the curve (AUC) can be calculated to be 0.8529, as shown in Figure 10.

Next, as shown in step S92 of FIG. 2 and FIG. 4 , in the third embodiment, if you want to know the probability that the kidney stone component of a patient suffering from kidney stones is uric acid crystals, first classify the patients. Corresponding to the system input variables x ₀ = "serum creatinine concentration", x ₁ = "sex", x ₂ = "age", x ₄ = "urine pH" and x ₅ = "BMI" five Each system input value is input to the artificial intelligence uric acid stone prediction system 2, and the eGFR calculation unit 21 calculates the patient's eGFR based on the system input values x ₀ , x ₁ , and x ₂ using the aforementioned eGFR formula.

Then, as shown in step S93, in this third embodiment, the machine learning prediction model 22 calculates the model output value y based on the patient's model input values _x2 = "age", _x3 = "eGFR", _x4 = "urine acid-base value", and _x5 = "BMI", which represents the probability that the patient's kidney stones are composed of uric acid crystals.

Referring to Figure 2, in a fourth embodiment of the present invention, the model input variables also include x ₁ = "gender", but the system input variables and the model input variables do not need to include x ₆ to x ₉ .

In this fourth embodiment, the function f(x) can be simplified to the following form: f(x)=w ₀ +w ₁ f ₁ (x ₂ )+w ₂ f ₂ (x ₂ )+w ₃ f ₃ ( x ₃ )+w ₄ f ₄ (x ₄ )+w ₅ f ₅ (x ₅ ), where w ₀ , w ₁ ,...,w ₅ are constant parameters, f ₁ (x ₁ ),... , f ₅ (x ₅ ) functions are all nonlinear functions constructed with fully connected two-layer neural networks. The output dimension of the first layer of the two-layer neural network is m _i (i=1 to 5), where m _i is an integer between 1 and 100, and the output dimension of the second layer is 1. The functions f ₁ (x ₁ ) to f ₅ (x ₅ ) respectively have the following forms: f _i (x _i )=σ _i,2 (A _i,2 σ _i,1 (A _i,1 x _i +B _{i, 1} )+B _i,2 ),i=1,...,5.

個可訓練模型參數。在本第四實施例中，所使用之m_i(i=1至5)皆為20，且B_i,1(i=1至5)皆設為零，共211個可訓練模型參數。 Among them, _Ai,1 and _Bi,1 (i=1 to 5) are _mi -by-1 vector variables, each of which includes _mi adjustable parameters. _Ai,2 (i=1 to 5) are 1-by- _mi vector variables, each of which also includes _mi adjustable parameters. B _1,2 to B _5,2 are all scalar variables. σ _i,1 and σ _i,2 (i=1 to 5) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

In the fourth embodiment, _mi (i=1 to 5) used is 20, and Bi _,1 (i=1 to 5) is set to zero, for a total of 211 trainable model parameters.

Referring to Figures 2, 4, 11, and 12, when the machine learning prediction model 22 established in step S91 of the fourth embodiment is applied to the verification set of group A patient data, the value under the curve can be calculated The area (AUC) is 0.8259, as shown in Figure 11. Similarly, when the machine learning prediction model 22 is applied to the patient data of Group B, the area under the curve (AUC) can be calculated to be 0.8627, as shown in Figure 12.

Next, as shown in step S92 of FIG. 2 and FIG. 4 , in the fourth embodiment, if you want to know the probability that the kidney stone component of a patient suffering from kidney stones is uric acid crystals, first classify the patients. Corresponding to the system input variables x ₀ = "serum creatinine concentration", x ₁ = "sex", x ₂ = "age", x ₄ = "urine pH" and x ₅ = "BMI" five Each system input value is input to the artificial intelligence uric acid stone prediction system 2, and then the eGFR calculation unit 21 calculates the patient's eGFR based on the system input values x ₀ , x ₁ , and x ₂ using the aforementioned eGFR formula.

Then, as shown in step S93, in the fourth embodiment, the machine learning prediction model 22 inputs values x ₁ = "gender", x ₂ = "age", and x ₃ = " based on the model input values of the patient. eGFR", x ₄ = "urine pH", and x ₅ = "BMI", the model output value y is calculated, which represents the probability that the patient's kidney stones are composed of uric acid crystals.

Referring to Figure 2, in a fifth embodiment of the present invention, the system input variables and the model input variables also include x ₆ = "whether there is a history of diabetes", but the system input variables and the model input variables The variables do not yet need to include x ₇ to x ₉ .

In this fifth embodiment, the function f(x) can be simplified to the following form: f(x)=w ₀ +w ₁ f ₁ (x ₂ )+w ₂ f ₂ (x ₂ )+w ₃ f ₃ ( x ₃ )+w ₄ f ₄ (x ₄ )+w ₅ f ₅ (x ₅ )+w ₆ f ₆ (x ₆ ), where w ₀ , w ₁ ,...,w ₆ are constant parameters, f ₁ (x ₁ ),...,f ₆ (x ₆ ) functions are all nonlinear functions constructed with fully connected two-layer neural networks. The output dimension of the first layer of the two-layer neural network is m _i (i=1 to 6), where m _i is an integer between 1 and 100, and the output dimension of the second layer is 1. The functions f ₁ (x ₁ ) to f ₆ (x ₆ ) respectively have the following forms: f _i (x _i )=σ _i,2 (A _i,2 σ _i,1 (A _i,1 x _i +B _{i, 1} )+B _i,2 ),i=1,...,6.

個可訓練模型參數。在本第五實施例中，所使用之m_i(i=1至6)皆為20，且B_i,1(i=1至6)皆設為零，共253個可訓練模型參數。 Among them, A _i,1 and B _i,1 (i=1 to 6) are vector variables of m _i -by-1, and each variable includes m _i adjustable parameters. And A _i,2 (i=1 to 6) are all 1-by-m _i vector variables, each of which also includes m _i adjustable parameters. B _1,2 to B _6,2 are all scalar variables. σ _i,1 and σ _i,2 (i=1 to 6) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

trainable model parameters. In the fifth embodiment, the m _i (i=1 to 6) used are all 20, and the Bi _,1 (i=1 to 6) are all set to zero, resulting in a total of 253 trainable model parameters.

Referring to Figures 2, 4, 13, and 14, when the machine learning prediction model 22 established in step S91 of the fifth embodiment is applied to the validation set of patient data of group A, its area under the curve (AUC) can be calculated to be 0.829, as shown in Figure 13. Similarly, when the machine learning prediction model 22 is applied to patient data of group B, its area under the curve (AUC) can be calculated to be 0.9363, as shown in Figure 14.

Next, as shown in step S92 of FIG. 2 and FIG. 4 , in the fifth embodiment, if you want to know the probability that the kidney stone component of a patient suffering from kidney stones is uric acid crystals, first classify the patients. Corresponding to the system input variables x ₀ = "serum creatinine concentration", x ₁ = "sex", x ₂ = "age", x ₄ = "urine pH", x ₅ = "BMI", x The six system input values ₆ = "whether there is a history of diabetes" are input into the artificial intelligence uric acid stone prediction system 2, and then the eGFR calculation unit 21 calculates using the aforementioned eGFR formula based on the system input values x ₀ , x ₁ , and x ₂ Find the patient's eGFR.

Then, as shown in step S93, in the fifth embodiment, the machine learning prediction model 22 calculates the model output value y based on the patient's model input values _x1 = "gender", _x2 = "age", _x3 = "eGFR", _x4 = "urine acid-base value", _x5 = "BMI", _x6 = "whether there is a history of diabetes", which represents the probability that the patient's kidney stones are composed of uric acid crystals.

Referring to Figure 2, in a sixth embodiment of the present invention, the system input variables and the model input variables also include x ₇ = "whether there is a history of gout" and x ₈ = "whether there is a history of bacteriuria" , but the system input variables and the model input variables do not need to include x ₉ yet.

In the sixth embodiment, the function f(x) can be simplified to the following form: f(x)= _w0 + _w1f1 ( _{x2 )} +w2f2( _{x2)+w3f3(x3 ) + w4f4 ( x4 )} + _{w5f5 ( x5 )} + _w6f6 ( _x6 , _x7 , _x8 ), wherein _w0 , _w1 ,..., _w6 are constant parameters, and the functions _f1 ( _x1 ),..., _f6 ( _x6 , _x7 , _x8 ) are all nonlinear functions constructed by _a fully _connected two-layer neural network. The output dimension of the first layer of the two-layer neural network is _mi (i=1 to 6), where _mi is an integer between 1 and 100, and the output dimension of the second layer is 1. The functions _f1 to _f6 are of the following forms: _fi ( _xi ) = σi _,2 (Ai _,2 σi _,1 (Ai _{,1 xi} +Bi _,1 ) + Bi _,2 ), i = 1, ..., 5; _f6 ( _x6 , _x7 , _x8 ) = _σ6,2 ( _{A6,2 σ6,1} ( _{A6,1 x6} + _{A7,1 x7} + _{A8,1 x8} + _B6,1 ) + _B6,2 ),

個可訓練模型參數。在本第六實施例中，所使用之m_i(i=1至6)皆為20，且B_i,1(i=1至6)皆設為零，共293個可訓練模型參數。 Among them, A _i,1 and B _i,1 (i=1 to 6) are vector variables of m _i -by-1, and each variable includes m _i adjustable parameters. The vector variables A _7,1 and A _8,1 have the same dimensions as A _6,1 . And A _i,2 (i=1 to 6) are all 1-by-m _i vector variables, each of which also includes m _i adjustable parameters. B _1,2 to B _6,2 are all scalar variables. σ _i,1 and σ _i,2 (i=1 to 6) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

trainable model parameters. In this sixth embodiment, the m _i (i=1 to 6) used are all 20, and the Bi _,1 (i=1 to 6) are all set to zero, resulting in a total of 293 trainable model parameters.

Referring to Figures 2, 4, 15, and 16, when the machine learning prediction model 22 established in step S91 of the sixth embodiment is applied to the validation set of patient data of group A, its area under the curve (AUC) can be calculated to be 0.8417, as shown in Figure 15. Similarly, when the machine learning prediction model 22 is applied to patient data of group B, its area under the curve (AUC) can be calculated to be 0.9461, as shown in Figure 16.

Next, as shown in step S92 of Figures 2 and 4, in the sixth embodiment, if one wants to know the probability that the kidney stones of a patient suffering from kidney stones are composed of uric acid crystals, the eight system input values of the patient corresponding to the system input variables _x0 = "serum creatinine concentration", _x1 = "gender", _x2 = "age", _x4 = "urine acid-base value", _x5 = "BMI", _x6 = "whether there is a history of diabetes", _x7 = "whether there is a history of gout", _x8 = "whether there is a history of bacteriuria" are first input into the artificial intelligence uric acid stone prediction system 2, and then the eGFR calculation unit 21 calculates the eGFR of the patient according to the system input values _x0 , _x1 , _x2 using the aforementioned eGFR formula.

Then, as shown in step S93, in this sixth embodiment, the machine learning prediction model 22 calculates the model output value y based on the patient's model input values _x1 = "gender", _x2 = "age", _x3 = "eGFR", _x4 = "urine acid-base value", _x5 = "BMI", _x6 = "whether there is a history of diabetes", _x7 = "whether there is a history of gout", _x8 = "whether there is a history of bacteriuria", which represents the probability that the patient's kidney stones are composed of uric acid crystals.

Referring to FIG. 2 , in a seventh embodiment of the present invention, the system input variables and the model input variables also include x ₉ = "whether there is a history of hypertension".

In this seventh embodiment, the function f(x) can be simplified to the following form: f(x)=w ₀ +w ₁ f ₁ (x ₂ )+w ₂ f ₂ (x ₂ )+w ₃ f ₃ ( x ₃ )+w ₄ f ₄ (x ₄ )+w ₅ f ₅ (x ₅ )+w ₆ (x ₉ )f ₆ (x ₆ ,x ₇ ,x ₈ ), where w ₀ , w ₁ ,. ..,w ₅ is a constant parameter, f ₁ (x ₁ ),...,f ₅ (x ₅ ),f ₆ (x ₆ ,x ₇ ,x ₈ ) functions are all based on fully connected two-layer neural Nonlinear functions constructed by the network. The output dimension of the first layer of the two-layer neural network is m _i (i=1 to 6), where m _i is an integer between 1 and 100, and the output dimension of the second layer is 1. Function w ₆ (x ₉ ) is also a nonlinear function constructed with a fully connected two-layer neural network. The output dimension of the first layer is m ₇ , m ₇ is an integer between 1 and 100, and the output dimension of the second layer The output dimension of the layer is 1. The functions f ₁ to f ₆ respectively have the following forms: f _i (x _i )=σ _i,2 (A _i,2 σ _i,1 (A _i,1 x _i +B _i,1 )+B _i,2 ) ,i=1,...,5，f ₆ (x ₆ ,x ₇ ,x ₈ )=σ _6,2 (A _6,2 σ _6,1 (A _6,1 x ₆ +A _7,1 x ₇ +A _8,1 x ₈ +B _6,1 )+B _6,2 ), w ₆ (x ₉ )=w _6,0 ． σ _7,2 (A _7,2 σ _7,1 (A _9,1 x ₉ +B _7,1 )+B _7,2 ).

個可訓練模型參數。在本第七實施例中，所使用之m_i(i=1至6)皆為20，m₇為5，且B_i,1(i=1至7)皆設為零，共304個可訓練模型參數。 Among them, _Ai,1 and _Bi,1 (i=1 to 6) are both _mi -by-1 vector variables, each of which includes _mi adjustable parameters. The dimensions of vector variables A _7,1 and A _8,1 are the same as A _6,1 . The dimensions of vector variables A _9,1 and B _7,1 are m ₇ -by-1, each of which includes m ₇ adjustable parameters. _Ai,2 (i=1 to 7) are all 1-by- _mi vector variables, each of which also includes _mi adjustable parameters. B _1,2 to B _7,2 and w _6,0 are all scalar variables. σ _i,1 and σ _i,2 (i=1 to 7) are standard nonlinear activation functions commonly used in constructing neural networks. The nonlinear function f(x) includes

In the seventh embodiment, _mi (i=1 to 6) is 20, _m7 is 5, and Bi _,1 (i=1 to 7) is set to zero, for a total of 304 trainable model parameters.

Referring to Figures 2, 4, 17, and 18, when the machine learning prediction model 22 established in step S91 of the seventh embodiment is applied to the validation set of patient data of group A, its area under the curve (AUC) can be calculated to be 0.8446, as shown in Figure 17. Similarly, when the machine learning prediction model 22 is applied to patient data of group B, its area under the curve (AUC) can be calculated to be 0.951, as shown in Figure 18.

Next, as shown in step S92 of Figures 2 and 4, in the seventh embodiment, if one wants to know the probability that the kidney stones of a patient suffering from kidney stones are composed of uric acid crystals, the patient's corresponding system input variables _x0 ~ _x9 are first input into the artificial intelligence uric acid stone prediction system 2, and then the eGFR calculation unit 21 calculates the patient's eGFR according to the system input values _x0 , _x1 , _x2 using the aforementioned eGFR formula.

Then, as shown in step S93, in the seventh embodiment, the machine learning prediction model 22 calculates the model output value y based on the model input values _x1 to _x9 of the patient, which represents the probability that the composition of the patient's kidney stones is uric acid crystals.

Referring to the variation example of Figure 3, it should be noted that although in the aforementioned embodiments of Figures 1 and 2, the eGFR calculation unit 21 must first be used to input the variables "serum creatinine concentration", "gender" and "Age" is used to calculate the model input variable "eGFR". However, in variations of the embodiments of the present invention, it is not necessary to set the eGFR calculation unit 21 in the artificial intelligence uric acid stone prediction system 2, but instead The system input variables and the model input variables can be exactly the same, and both include at least the eGFR and the urine pH value. For example, as shown in Figure 3, the system input variables and the model input variables include x ₁ = "Gender", x ₂ = "Age", x ₃ = "eGFR", x ₄ = "Urine pH", x ₅ = "BMI", x ₆ = "Do you have a history of diabetes", x ₇ = "Do you have a history of gout?", x ₈ = "Do you have a history of bacteriuria", and x ₉ = "Do you have a history of hypertension".

In summary, the advantages and effectiveness of the artificial intelligence uric acid stone prediction method of the present invention are that, based on the high AUC shown in Figures 5 to 18, it can be seen that the machine learning prediction model 22 has a highly accurate prediction ability. Therefore, as long as 2 to 9 easily obtained clinical variables of patients suffering from kidney stones are used as model input variables of the machine learning prediction model 22, the machine learning prediction model 22 can calculate the probability y that the patient's kidney stones are composed of uric acid crystals. The probability y can be used by medical personnel to determine whether the patient's kidney stones are uric acid stones, so as to prescribe symptomatic treatment as soon as possible; therefore, the purpose of the present invention can be achieved.

However, the above are only examples of the present invention and should not be used to limit the scope of the present invention. All simple equivalent changes and modifications made based on the patent scope of the present invention and the content of the patent specification are still within the scope of the present invention. within the scope covered by the patent of this invention.

S91~S93: Steps

Claims (10)

An artificial intelligence uric acid stone prediction method comprises the following steps: (A) establishing an artificial intelligence uric acid stone prediction system, wherein the artificial intelligence uric acid stone prediction system comprises a machine learning prediction model including a plurality of model input variables and a model output variable, wherein the model input variables include a glomerular filtration rate estimation value and a urine acid-base value, and the model output variable represents the composition of the kidney stone The probability of uric acid crystals; (B) inputting several model input values of a patient suffering from kidney stones, which respectively correspond to the model input variables, into the machine learning prediction model; and (C) the machine learning prediction model calculates a model output value corresponding to the model output variable based on the model input values, and the model output value indicates the probability that the composition of the patient's kidney stones is uric acid crystals. The artificial intelligence uric acid stone prediction method as described in claim 1, wherein the model input variables also include a body mass index. The artificial intelligence method for predicting uric acid stones as described in claim 2, wherein the model input variables also include an age. The artificial intelligence method for predicting uric acid stones as described in claim 3, wherein the model input variables also include a gender. The artificial intelligence method for predicting uric acid stones as described in claim 4, wherein the model input variables also include whether there is a history of diabetes. As described in claim 5, the artificial intelligence uric acid stone prediction method, wherein the model input variables also include whether there is a history of gout and whether there is a history of bacteriuria. The artificial intelligence uric acid stone prediction method as described in claim 6, wherein the model input variables also include whether there is a history of hypertension. The artificial intelligence uric acid stone prediction method as described in any one of claims 1 to 7, wherein the structure of the machine learning prediction model is

, where x represents the vector formed by the input variables of the model, and y is the output variable of the model, representing the probability that the kidney stones are composed of uric acid crystals, f(x)= _w0 + _w1 (x) _f1 (x)+ _w2 (x) _f2 (x)+...+ _wn (x) _fn (x), where _w0 is a constant parameter, and _w1 (x), _w2 (x),..., _wn (x) and _f1 (x), _f2 (x),..., _fn (x) are all functions constructed by a fully connected artificial neural network. An artificial intelligence uric acid stone prediction method comprises the following steps: (a) establishing an artificial intelligence uric acid stone prediction system, wherein the artificial intelligence uric acid stone prediction system comprises a plurality of system input variables and a machine learning prediction model comprising a plurality of model input variables and a model output variable, wherein the system input variables comprise a serum creatinine concentration, a gender, an age, and a urine pH value, and the model input variables at least comprise the urine pH value, The model output variable indicates the probability that the kidney stone is composed of uric acid crystals; (b) a plurality of system input values of a patient suffering from kidney stones, which respectively correspond to the system input variables, are input into the artificial intelligence uric acid stone prediction system; and (c) the machine learning prediction model calculates a model output value corresponding to the model output variable according to the system input values, wherein the model output value indicates the probability that the patient's kidney stones are composed of uric acid crystals. The artificial intelligence uric acid stone prediction method as described in claim 9, wherein: in step (a), the artificial intelligence uric acid stone prediction system also includes a glomerular filter for calculating a glomerular filtration rate estimate. Rate estimate calculation unit, the system input variables also include a body mass index, whether there is a history of diabetes, whether there is a history of gout, whether there is a history of bacteriuria, and whether there is a history of hypertension, the model input variables also include the Gender, age, estimated glomerular filtration rate, body mass index, history of diabetes, history of gout, history of bacteriuria, and history of hypertension; estimated glomerular filtration rate The value calculation unit calculates the glomerular filtration rate estimate based on the patient's serum creatinine concentration, gender, and age, and then the machine learning prediction model estimates the glomerular filtration rate based on the gender, age, and age. value, the urine pH value, the body mass index, whether there is a history of diabetes, whether there is a history of gout, whether there is a history of bacteriuria, and whether there is a history of hypertension, calculate the model output value; and the machine learning prediction model The structure is

, where x represents the vector formed by the input variables of the model, and y is the output variable of the model, which represents the probability that the kidney stone component is uric acid crystals, f(x)=w ₀ +w ₁ (x)f ₁ ( x)+w ₂ (x)f ₂ (x)+...+w _n (x)f _n (x), where w ₀ is a constant parameter, and w ₁ (x), w ₂ (x) ,..., w _n (x) and f ₁ (x), f ₂ (x),..., f _n (x) are all functions constructed by fully connected artificial neural networks.

Images