Waist circumference
Men 90 cm or higher,
10

(10)
Women 85 cm or higher

Pulse rate for
60 or less
20

1 min (20)
60 to 80
10

80 or more
0

Weight
BMI 35 or more
40

(BMI) (40)
BMI 34.9 to 25
30

BMI 25 or more
0

Waist circumference
Men 90 cm or higher,
10

(10)
Women 85 cm or higher

Pulse rate for
60 or less
20

1 min (20)
60 to 80
10

80 or more
0

TABLE 3

Item (Total score)
Response item
Score by item

Number of wake-up
3 times or more
20

times during
1 to 2 times
10

sleep (20)
0 time
0

Other sleep
Bed time is after midnight
5

states (20)
Bed time is irregular
5

Wake-up time in the morning
5

is after 8 a.m.

Total sleep time is irregular
5

Pulse rate for
100 or more
20

1 min (20)
80 to 100
10

80 or less
0

Upper and
Chest circumference/hip
5

lower body
circumference >= average

ratio in body
Neck circumference/calf
5

type (10)
circumference >= average

Heat
Usually feel hot in face
5

evaluation (15)
Feet below waist are cold
5

Usually like cold food
5

Usual activity
Usually move around a lot
5

habits (5)

Usual
Usual personality is extrovertive
5

personality (5)
or impatient

Abdominal
Feel tense when pressing
5

condition (5)
abdomen with hand

Meanwhile, in an embodiment, it may be determined whether there is a missing value in the response information about the diagnosee before performing the diagnosis. In addition, in response to the determination that there is the missing value, the missing value may be corrected using an autoencoder-based deep learning model. Accordingly, the convenience of the diagnosee may be improved, and the embodiment will be described in detail later with reference to FIGS. 11 and 12.

In step S22, the diagnosis may be performed based on the response information about the diagnosee. For example, the diagnosis system 10 may diagnose the energy and blood generation function based on the response information to the first survey question (e.g., if the total score of the response items is 50 points or more, it is determined that there is a problem in the corresponding function), diagnose the energy and blood circulation function based on the response information to the second survey question, and diagnose the energy and blood balance adjustment function based on the response information to the third survey question. In addition, the diagnosis system 10 may diagnose the health condition (e.g., presence or absence of disease, type of disease) of the diagnosee by considering the problematic function. Examples of diseases associated with the problematic function will refer to Table 4 below. As another example, the diagnosis system 10 may also diagnose the health condition of the diagnosee by comprehensively considering response information about the energy and blood generation function, circulation function, and adjustment function.

TABLE 4

Classification
Associated diseases

Energy and
Dementia, Parkinson's disease, amnesia, acute and

blood generation
chronic digestive diseases, chronic fatigue, low blood

function
pressure, tinnitus (Meniere), decreased vision,

vertigo, headache, Raynauds syndrome, shoulder pain

Energy and
Obesity, metabolic diseases (hyperlipidemia, blood

blood circulation
pressure, diabetes), heart disease, stroke, respiratory

function
disease, arthritis, back pain (disk, stenosis),

autoimmune diseases (rheumatism, fibromyalgia,

ankylosing spondylitis), various cancers, etc.

Energy and
Reproductive diseases, infertility (sterility), endocrine

blood balance
diseases (thyroid hormones, etc.), urinary diseases

adjustment
(urinary disorders), hair loss, dry eyes, skin

function
diseases (psoriasis), headache, tinnitus, mental

illness (insomnia, panic disorder, depression),

back pain (disk, stenosis), arthritis, plantar

fasciitis, etc.

In an embodiment, the diagnosis system 10 may provide the diagnosee with additional survey information for diagnosing diseases associated with the problematic function, and also predict diseases associated with the diagnosee based on the response information to the additional survey.

Meanwhile, the specific method of performing the diagnosis may vary depending on an embodiment.

In an embodiment, the health condition of the diagnosee may be diagnosed based on the similarity between a state vector generated based on response information about the diagnosee (i.e., a vector representing the health condition of the diagnosee) and a disease vector. The embodiment will be described below with reference to FIGS. 3 to 5.

In another embodiment, the health condition of the diagnosee may be diagnosed using an autoencoder-based deep learning model trained using response information about a normal class. The embodiment will be described below with reference to FIGS. 6 and 7.

In yet another embodiment, the health condition of the diagnosee may be diagnosed using an autoencoder-based deep learning model trained using response information about a patient class. The embodiment will be described below with reference to FIG. 8.

In yet another embodiment, the health condition of the diagnosee may be diagnosed using a classifier that classifies the normal class and the patient class. The embodiment will be described below with reference to FIG. 9.

In yet another embodiment, the health condition (e.g., disease type) of the diagnosee may be diagnosed using a classifier that classifies a disease class. The embodiment will be described below with reference to FIG. 10.

In yet another embodiment, the diagnosis may be performed based on a combination of the embodiments described above. For example, the diagnosis system 10 may also diagnose the health condition of the diagnosee by using a first deep learning model trained with response information about the normal class and a second deep learning model trained with response information of the patient class together. As another example, the diagnosis system 10 may also diagnose the health condition of the diagnosee using a classifier that classifies a normal class and a disease class. As yet another example, the diagnosis system 10 may diagnose the presence or absence of the disease of the diagnosee using the first deep learning model trained with the response information from the normal class, and also diagnose a disease type of the diagnosee using a classier that classifies the disease class in response to the diagnosis result that the disease is present.

In step S24, the diagnosis result may be provided to the diagnosee. For example, the diagnosis system 10 may provide the diagnosis result to the diagnosee terminal 1.

The diagnosis result may include information such as the health condition (e.g., presence or absence of disease, type of disease) of the diagnosee, a treatment method, etc. However, the present disclosure is not limited thereto. Examples of the treatment method will refer to Tables 5 to 7 below. Tables 5 to 7 illustrate life regimens to treat problems with the energy and blood generation, circulation, and balance adjustment functions, respectively.

TABLE 5

Classification
Life regimens

Food
Eat three meals slowly at a set time.

Avoid oily food, flour-based food, and cold food.

Eat soft and warm foods.

Drink ginseng tea, ginger tea, cinnamon tea, mugwort

tea, etc., which help with a stomach function.

Exercise
Avoid strenuous exercise and do exercises such as

stretching, yoga, and walking.

Because a lack of energy and blood can lead to obesity

in the lower body, do upper body exercises (e.g.,

horizontal bar, upper body dumbbell exercises) that

can draw energy upward.

Avoid exercising in cold water, such as swimming.

Hobby
Try to live a bright life by intentionally

listening to upbeat music.

TABLE 6

Classification
Life regimens

Food
When there is a problem with the circulation

function, there is a tendency for metabolic

products to accumulate in the body, so that the

amount of food is controlled to manage obesity.

Eat mainly a low-carbohydrate, high-protein diet.

Drink teas that help with the circulation function,

such as Ophiopogon japonicus tea, mulberry leaf

tea, yulmu tea, and green tea.

Exercise
Avoid continuously lying down or napping for 30

minutes or more because of being tired.

Do exercise that sweats a lot, such as brisk

walking, jogging, hiking, or swimming.

In obese states, the amount of exercise increases

along with fat loss.

Hobby
Do active and dynamic activities, such as

aerobics, dance sports, and nanta.

Enjoy listening to exciting dance music, etc.

TABLE 7

Classification
Life regimens

Food
Drink plenty of water of about 2 liters a day to

avoid lack of yin energy.

Eat fish, seaweed, and cold foods, and avoid spicy

and salty foods or dry side dishes.

Drink barley tea, which has a cool nature, and

Chinese matrimony vina tea, eucommia tea, cornelian

cherry tea, etc. which strengthen the yin energy,

and avoid teas such as coffee or green tea that

stimulate and enhance the body due to caffeine.

Exercise
Avoid exercise that causes a lot of sweating.

Since a lower body will be weak due to a lack of

yin energy, do lower body exercises that can draw

the energy downward (e.g., riding a bicycle, walking

barefoot, walking backwards, climbing stairs,

squats, etc.).

Avoid swimming or other upper body exercises that

do not apply gravity to the lower body.

Hobby
Do static activities, such as meditation, danjeon

breathing, art, and baduk.

Listen to quiet, relaxing music.

So far, a survey-based diagnosis method according to several embodiments of the present disclosure has been described with reference to FIG. 2. According to the method described above, it is possible to diagnose the health condition of a diagnosee using a plurality of survey questions for diagnosing energy and blood generation function, circulation function, and balance adjustment function. From an oriental medicine viewpoint, since most diseases are caused by problems with the functions, it is possible to accurately diagnose the health condition of a diagnosee by using survey questions for diagnosing these functions. Furthermore, it is possible to provide a quick and convenient diagnosis service to a diagnosee by using a survey technique.

Hereinafter, several embodiments of a diagnosis method will be described with reference to the drawings in FIG. 3 and below.

First, a diagnosis method according to an embodiment of the present disclosure will be described with reference to FIGS. 3 and 4.

As illustrated in FIG. 3, the embodiment relates to a method for diagnosing a health condition of a diagnosee based on vector similarity, and may start at step S222 of generating a state vector representing the health condition of the diagnosee based on response information about the diagnosee. A specific method of generating the state vector at step S222 may vary.

For example, the diagnosis system 10 may generate a state vector for the energy and blood generation function based on the scores of the first survey question. As a more specific example, if the first survey question consists of a plurality of questions, a state vector (e.g., a vector having the scores of the response items of the diagnosee as a vector element) for the energy and blood generation function may be generated based on the scores of each question. Further, the diagnosis system 10 may generate a state vector for the energy and blood circulation function based on the scores of the second survey question, and generate a state vector for the energy and blood balance adjustment function based on the scores of the third survey question.

As another example, as illustrated in FIG. 4, the diagnosis system 10 may generate vectors V1 to V3 for the first survey question to third survey question 41 to 43 and generate a state vector V_stateby synthesizing the generated state vector based on predetermined weights W1 to W3. In such a case, a state vector V_statemay be generated by comprehensively considering the energy and blood generation function, circulation function, and balance adjustment function. In this example, the weights W1 to W3 may also vary depending on the demographic characteristics (e.g., sex, age range, age, race, body type, BMI, etc.) of the diagnosee and disease history (e.g., type of disease, number of times suffering from a disease, etc.). For example, since the energy and blood generation function is closely related to the stomach, if there is a history of diseases related to the stomach, the weight W1 for the first survey question may be determined as a higher value. Similarly, when there is a history of cardiopulmonary diseases, the weight W2 for the second survey question may be determined as a higher value. Alternatively, in the case of a lower body obesity body type, since there is a high possibility of having a problem with the energy and blood generation function, the weight for the first survey question may be determined as a higher value. Similarly, in the case of an upper body obesity body type, since there is a high possibility of having a problem with the energy and blood balance adjustment function, the weight for the third survey question may be determined as a higher value. Alternatively, since there is a high possibility of having a problem with the energy and blood circulation function due to a decreased metabolism as the age is higher, the weight for the second survey question may be determined as a higher value as the age range of the diagnosee increases.

As yet another example, as illustrated in FIG. 5, the diagnosis system 10 may generate a state vector using an autoencoder-based deep learning model. Specifically, the state vector may be generated using a latent vector 53 extracted from an encoder 52 of a deep learning model trained with response information about a normal class or a patient class. The autoencoder-based deep learning model may include a basic auto-encoder and its variations (e.g., variational autoencoders). However, for the convenience of understanding, the following description will continue assuming that the deep learning model consists of a basic autoencoder structure. This example will be further described with reference to FIG. 5. FIG. 5 illustrates an autoencoder structure in which a dimensionality d of the latent vector 53 (or bottleneck layer) is 3.

As illustrated in FIG. 5, an autoencoder-based deep learning model may consist of an encoder 52 and a decoder 54, and may be trained based on a reconstruction loss value between input data 51 and output data 55. Those skilled in the art have already known the configuration and operating principle of the autoencoder, and thus, more description of the autoencoder will be omitted.

In this example, the diagnosis system 10 may extract the latent vector 53 by inputting response information about the diagnosee into the encoder 52 of the deep learning model trained with the response information about the normal class or the patient class. Of course, the diagnosis system 10 may convert the response information about the diagnosee into a numerical form to be input to the encoder 52 through an appropriate vectorization technique (e.g., the method described with reference to FIG. 4). In addition, the diagnosis system 10 may generate a state vector by using the extracted latent vector 53 as a state vector of the diagnosee, or performing additional processing (e.g., PCA, etc.) on the extracted latent vector 53.

It will be described with reference to FIG. 3 again.

In step S224, the health condition of the diagnosee may be diagnosed based on the similarity between the generated state vector and a disease vector. For example, as illustrated in FIG. 4, the diagnosis system 10 may diagnose whether there is an abnormality in the health condition or a type of disease based on the similarity (e.g., cosine similarity, etc.) between the state vector V_stateof the diagnosee and disease vectors V_diseaseAand V_diseaseB(e.g., diagnose a disease with the highest similarity as the disease of the diagnosee).

For reference, the disease vectors V_diseaseAand V_diseaseBmay be generated based on response information about patients belonging to the corresponding disease class, and may be generated in the same or similar manner as the generated vector.

So far, the diagnosis method according to an embodiment of the present disclosure has been described with reference to FIGS. 3 to 5. Hereinafter, a diagnosis method according to another embodiment of the present disclosure will be described with reference to FIG. 6.

As illustrated in FIG. 6, the embodiment relates to a method for diagnosing a health condition of a diagnosee using an autoencoder-based deep learning model 60 trained with response information 63 about a normal class (see normal).

Specifically, the deep learning model 60 may be configured to include an encoder 61 and a decoder 62, and may be trained using response information 63 of a normal class. For example, the diagnosis system 10 may input individual response information 64 into the deep learning model 60 and train the deep learning model 60 using the reconstruction loss value between input information 64 and reconstructed information 65 (e.g., the individual response information 64 may be converted into a numerical form and input into the deep learning model 60; see FIG. 12). At this time, in the deep learning model 60, a process of encoding the input information 64 into a latent vector by the encoder 61 and decoding the latent vector by the decoder 62 may be performed, and as the reconstruction loss value is back-propagated, weight parameters of the encoder 61 and the decoder 62 may be updated. As such a training process is repeated, the deep learning model 60 has the ability to compress (encode) and then reconstruct (decode) response information (e.g., 64) of the normal class.

Next, as illustrated in FIG. 7, a reconstruction loss value for response information 71 about the diagnosee may be calculated using the trained deep learning model 60. As described above, the reconstruction loss value may be calculated based on a difference between the response information 71 about the diagnosee and information 72 reconstructed by the deep learning model 60.

Next, the health condition of the diagnosee may be diagnosed based on the calculated reconstruction loss value. For example, if the reconstruction loss value is equal to or greater than a reference value, the diagnosis system 10 may determine that there is an abnormality in the health condition of the diagnosee. A large reconstruction loss value means that the response information 71 about the diagnosee is information that is not well reconstructed by the deep learning model 60, which is because the response information 71 about the diagnosee is significantly different from the response information 63 about the normal class. If the reconstruction loss value is equal to or less than the reference value, the diagnosis system 10 may determine that there is no abnormality in the health condition of the diagnosee.

So far, the diagnosis method according to another embodiment of the present disclosure has been described with reference to FIGS. 6 and 7. Hereinafter, a diagnosis method according to yet another embodiment of the present disclosure will be described with reference to FIG. 8.

As illustrated in FIG. 8, the embodiment relates to a method for diagnosing a health condition of a diagnosee using an autoencoder-based deep learning model 80 trained with response information 83 about a patient class (see disease).

Specifically, the deep learning model 80 may be configured to include an encoder 81 and a decoder 82, and may be trained using response information 83 about a patient class. For example, the diagnosis system 10 may input individual response information 84 into the deep learning model 80 and train the deep learning model 80 using a reconstruction loss value between input information 84 and reconstructed information 85.

Next, a reconstruction loss value for the response information (e.g., 71) about the diagnosee may be calculated using the trained deep learning model 80, and the health condition of the diagnosee may be diagnosed based on the calculated reconstruction loss value. For example, if the reconstruction loss value is equal to or greater than a reference value, the diagnosis system 10 may determine that there is no abnormality in the health condition of the diagnosee. A large reconstruction loss value means that the response information (e.g., 71) about the diagnosee is information that is not well reconstructed by the deep learning model 80, which is because the response information (e.g., 71) about the diagnosee is significantly different from the response information 83 about the patient class. If the reconstruction loss value is equal to or less than the reference value, the diagnosis system 10 may determine that there is an abnormality in the health condition of the diagnosee.

So far, the diagnosis method according to yet another embodiment of the present disclosure has been described with reference to FIG. 8. Hereinafter, a diagnosis method according to yet another embodiment of the present disclosure will be described with reference to FIG. 9.

As illustrated in FIG. 9, the embodiment relates to a method for diagnosing a health condition of a diagnosee using an autoencoder-based deep learning model 90 trained with response information 95 about a normal class (see normal) and a patient class (see disease).

Specifically, the deep learning model 90 may be configured to include an encoder 91, a decoder 92, and a classifier 93. At this time, the classifier 93 may serve to classify an input latent vector 94 into a normal class and a patient class.

As described above, the deep learning model 90 may be trained using response information 95 about the normal class and the patient class. For example, the diagnosis system 10 may train the encoder 91 and decoder 92 using a reconstruction loss value for individual response information 96 about the normal class or the patient class. The reconstruction loss value may be calculated based on a difference between input response information 96 and reconstructed information 97. In addition, the diagnosis system 10 may input a latent vector 94 extracted from the input response information 96 into the classifier 93 to calculate a classification loss value for the latent vector 94. The classification loss value may be calculated based on a difference between a predicted class and a correct class (i.e., a difference in confidence scores for each class) by the classifier 93. In addition, the diagnosis system 10 may train the classifier 93 and the encoder 91 using the classification loss value while freezing the decoder 92. Accordingly, the encoder 91 may be trained to extract a latent vector 94 suitable for class classification.

Next, a latent vector (e.g., 94) may be extracted from the response information (e.g., 71) of the diagnosee using the trained deep learning model 90, and a confidence score for each class (e.g., a confidence score for a normal class or a patient class) for the latent vector (e.g., 94) may be calculated by the classifier 93.

Next, the health condition of the diagnosee may be diagnosed based on the confidence score for each class. For example, if the confidence score of the patient class is equal to or greater than a reference value (or the confidence score of the normal class is equal to or less than a reference value), the diagnosis system 10 may determine that there is an abnormality in the health condition of the diagnosee. If it is visa verse, the diagnosis system 10 may determine that there is no abnormality in the health condition of the diagnosee.

So far, the diagnosis method according to yet another embodiment of the present disclosure has been described with reference to FIG. 9. Hereinafter, a diagnosis method according to yet another embodiment of the present disclosure will be described with reference to FIG. 10.

As illustrated in FIG. 10, the embodiment relates to a method for diagnosing a disease type of a diagnosee using an autoencoder-based deep learning model 100 trained with response information 105 about patients in different disease classes (see disease A and disease B).

Specifically, the deep learning model 100 may be configured to include an encoder 101, a decoder 102, and a classifier 103. At this time, the classifier 103 may serve to classify an input latent vector 104 into a first disease class (see disease A) and a second disease class (see disease B). Of course, the classifier 103 may further perform classification for a third disease class.

As described above, the deep learning model 100 may be trained using response information 105 about patients belonging to different disease classes. For example, the diagnosis system 10 may train the encoder 101 and decoder 102 using a reconstruction loss value for individual response information 106 about a patient in a specific disease class. The reconstruction loss value may be calculated based on a difference between input response information 106 and reconstructed information 107. In addition, the diagnosis system 10 may input a latent vector 104 extracted from the input response information 106 into the classifier 103 to calculate a classification loss value for the latent vector 104. The classification loss value may be calculated based on a difference between a predicted disease class and a correct class (i.e., a difference in confidence scores for each class) by the classifier 103. In addition, the diagnosis system 10 may train the classifier 103 and the encoder 101 using the classification loss value while freezing the decoder 102. Accordingly, the encoder 101 may be trained to extract a latent vector 104 suitable for disease class classification.

Next, a latent vector (e.g., 104) may be extracted from the response information (e.g., 71) of the diagnosee using the trained deep learning model 100, and a confidence score for each disease class for the latent vector (e.g., 104) may be calculated by the classifier 103.

Next, the disease type of the diagnosee may be diagnosed based on the confidence score for each class. For example, when the confidence score of the first disease class is equal to or greater than a reference value, the diagnosis system 10 may determine that the diagnosee suffers from a first disease.

So far, several embodiments of the diagnosis method have been described with reference to FIGS. 3 to 10. According to the aforementioned method, the health condition of the diagnosee may be accurately diagnosed using the autoencoder-based deep learning model trained using similarity in a vector space or using response information about the normal class or the patient class. Furthermore, the type of disease may be accurately diagnosed by using a classifier that classifies disease classes.

Meanwhile, there may be cases where there are missing values (i.e., non-response to specific survey items) in the response information provided by the diagnosee for various reasons. In such a case, the diagnosis system 10 may request the diagnosee to correct the missing values, or may perform automatic missing value correction for the convenience of the diagnosee. Hereinafter, a method for performing automatic missing value correction by the diagnosis system 10 will be described in detail with reference to FIGS. 11 and 12.

FIG. 11 is an exemplary flowchart illustrating a missing value correction method according to an embodiment of the present disclosure. However, it is only a preferred embodiment for achieving the object of the present disclosure, and it is obvious that some steps may be added or deleted as necessary.

As illustrated in FIG. 11, the embodiment may start at step S110 of acquiring an autoencoder-based deep learning model trained using response information about a survey respondent. For example, as illustrated in FIG. 12, the diagnosis system 10 consists of an encoder 121 and a decoder 122 and may acquire a deep learning model 120 trained with response information.

In an example, classes of survey respondents may be classified based on demographic characteristics and/or disease history, and a deep learning model for each class may be constructed. For example, a first deep learning model may be constructed using response information about a survey respondent belonging to a first class, and a second deep learning model may be constructed using response information about a survey respondent belonging to a second class. In this case, the diagnosis system