CN108648783B

CN108648783B - Method for rapidly judging similarity of medical and technical examination items

Info

Publication number: CN108648783B
Application number: CN201810199061.3A
Authority: CN
Inventors: 何必仕; 江翔; 陈晖�; 沈伟富; 何炜; 高彩英; 陈伟
Original assignee: Zhejiang Radiology Information Technology Co ltd; Hangzhou Dianzi University
Current assignee: Zhejiang Radiology Information Technology Co ltd
Priority date: 2018-03-12
Filing date: 2018-03-12
Publication date: 2021-08-10
Anticipated expiration: 2038-03-12
Also published as: CN108648783A

Abstract

The invention discloses a method for quickly judging similarity of medical and technical examination items. According to the invention, according to the naming characteristics of medical examination projects, parts and an examination method are directly used as semantic spaces, low-dimensional feature vectors are dynamically constructed for similarity judgment, the defects of complex calculation such as massive statistics, dimension reduction decomposition, semantic extraction and the like in the conventional text similarity calculation method are overcome, and the influence of the examination order on the similarity is distinguished. The invention is simple, convenient, quick, accurate and effective, and meets the requirement of practical application.

Description

Method for rapidly judging similarity of medical and technical examination items

Technical Field

The invention relates to the field of medical information retrieval, in particular to a method for quickly judging similarity of medical and technical examination items.

Background

Although the health supervision department makes five applications of three orders require that each medical institution develops the mutual recognition work of the inspection data and the inspection results on the premise of ensuring the medical quality and the medical safety, the health resources are reasonably and effectively utilized, and the burden of patients is reduced. However, local interest drives, insufficient information content of conventional films, evidence inversion, etc., result in a large number of repeat examinations. Many patients often have multiple medical visits to ask for diagnosis when they have trouble with miscellaneous diseases. Just after a series of comprehensive examinations in one hospital, the patient is forced to repeat the examinations while traveling to another hospital. Not only delays the diagnosis and treatment time and wastes sanitary resources, but also consumes the body, and particularly, the repeated radiation examination causes the patient to receive excessive radiation dose in a short period.

With the establishment of health information platforms and regional PACS in various regions, the sharing of electronic medical records and image examination among hospitals becomes a reality, and the mutual recognition work of examination results is technically strongly supported. However, the number of repeated examinations remains high. For this reason, the health supervision department hopes to realize the prior reminding and the after audit of the repeated check through the health information platform.

In actual work, the numbers and names of examination items are compiled differently in hospitals, and no industry standard is converted and unified. This makes the definition of "double check" difficult. Starting from similarity analysis of 'inspection project names', firstly, the 'inspection project names' only consist of a plurality of non-repeated words, such as 'CT head flat scan', and no word frequency TF difference exists; secondly, the examination item name and the examination report are not in the relation of title and content, and the inverse text frequency IDF cannot be counted. In the traditional text similarity analysis technology based on semantics, feature words of key points (1) are obtained by counting word frequency TF and reverse text frequency IDF; (2) the semantic extraction of the original text is realized through LSA or PLSA dimension reduction processing; (3) similarity is calculated using cosine measures of the feature vectors. Obviously, the similarity analysis for the "inspection item name" cannot be applied at all for two points (1) and (2).

Furthermore, the existing text similarity calculation method cannot distinguish the influence of the examination order on the similarity of examination items.

Reference documents: 201310105450.2A method for calculating text similarity based on target text.

Disclosure of Invention

The invention provides a method for rapidly judging similarity of medical examination items in view of the technical problems, which comprises the following steps:

1. and acquiring the current inspection project name A and the historical inspection project name B to be compared, and completing automatic word segmentation of 2 inspection project names.

2. And comparing with a professional name word library of the inspection project, classifying according to the equipment type, the part and the inspection method, and finishing synonym substitution.

3. And checking the device type, and determining a device type matching coefficient alpha. When the types of the two inspection devices are the same, taking alpha as 1; otherwise, α is taken to be 0.5.

4. 2 feature vectors of the parts and the inspection methods are constructed respectively, and the similarity is calculated.

According to the parts obtained by word segmentation, a part vector space is constructed, the parts are columns, and A, B are rows, as shown in FIG. 1. The two examined bit sequences are assigned an initial value of 0 or 1, respectively. According to the examination item professional name library check, if there is an affiliation between columns (parts), most of the attributes extend downward to sub-parts (sub-part attributes cannot extend upward). To distinguish the effect of the checking order on the similarity of the checking items, the current item a is taken to be 1.2 when it extends downward, and the history item B is taken to be 0.8 when it extends downward. This yields the location feature vectors for a and B.

Also, an inspection method vector space is constructed according to the inspection method obtained by word segmentation, the inspection method is column, and A, B is row, as shown in table 1. The inspection method sequences of the two inspections are given initial values of 0 or 1, respectively, thereby obtaining inspection method feature vectors of a and B.

Respectively calculating the part similarity beta and the inspection method similarity gamma according to a cosine measurement formula of the vector:

wherein bd is_a,bd_bFeature vectors, m, representing A and B, respectively_a,m_bThe inspection method feature vectors of a and B are respectively represented.

5. And (5) comprehensively calculating the similarity, and carrying out repeated inspection and judgment.

The total similarity Φ calculation is as follows:

Φ＝α*(w₁β+w₂γ)

wherein, α, β, γ are respectively the device type matching coefficient, the part similarity, and the inspection method similarity. w is a₁And w₂For the weighting factor, generally take w₁＝2/3，w₂＝1/3。

Generally, two thresholds Th1 (0.8-0.9) and Th2 (0.6-0.7) are set, and when the total similarity phi is greater than Th1, the repeated check can be judged; and when the Th1 is larger than or equal to the total similarity phi > Th2, the suspected duplicate check is judged.

The invention provides a method for quickly judging similarity of medical examination items, which is characterized in that according to the naming characteristics of the medical examination items, parts and an examination method are directly used as semantic spaces, low-dimensional feature vectors are dynamically constructed for similarity judgment, the defects of complex calculation such as massive statistics, dimension decomposition and reduction, semantic extraction and the like of the conventional text similarity calculation method are overcome, and the influence of an examination order on the similarity is distinguished. The method is simple, convenient, quick, accurate and effective, and meets the requirements of practical application.

Drawings

FIG. 1 is a schematic view of the working process of the present invention.

Detailed Description

The invention is further described below with reference to the accompanying drawings.

The typical application scenario of the invention is that when a doctor opens an order for examination, the doctor compares the currently opened examination item with the recent examination item of the patient on the health information platform, and if the examination belongs to repeated examination, the doctor gives a prompt in advance. And the second typical application scenario is that similar examinations of a patient in a short period are arranged on a health information platform, and after-the-fact audit of repeated examinations is carried out by combining with an electronic medical record of the patient.

The names of the medical examination items are usually named by a combination of "device type", "part", and "examination method", but the word order is not necessarily required. Now, taking a radiology examination order-opening scene as an example, the current order-opening examination item is "skull pituitary CT flat scan + enhanced scan", and the recent examination item is "CT head flat scan", which will be described.

In the specific implementation of the present invention, as shown in fig. 1, the following steps are included:

1. and acquiring a current inspection item name A and a historical inspection item name B to be compared, and completing automatic word segmentation of 2 inspection item names by adopting an ICTCCLAS word segmentation tool. To increase word segmentation accuracy, the user lexicon may be used to augment the specialized terms.

Here, a is "skull pituitary CT sweep + enhancement scan", and B is "CT head sweep". By means of ICTCCLAS system, A is 'CT flat scan enhanced scan' of head pituitary and B is 'CT head flat scan'.

2. And comparing with a professional name library of the inspection project, classifying according to the equipment type, the part and the inspection method, completing synonym substitution, and discarding redundant words.

Table 1 is an example of a professional name word library of inspection items, now categorizing and substituting synonyms for inspection items a and B, respectively:

a, equipment type is { CT }, part is { skull, pituitary }, and inspection method is { flat scanning, enhancement }

B, equipment type (CT), part (skull), and inspection method (flat scanning)

Table 1 examination of project professional thesaurus examples (synonyms in parentheses)

3. And checking the device type, and determining a device type matching coefficient alpha. When the types of the two inspection devices are the same, taking alpha as 1.0; otherwise, α is taken to be 0.5.

In general, the type of device in the inspection item can be known, and does not need to be participled from the name of the inspection item. When the two checking devices are the same in type, the comparison significance is achieved. Obviously, here, both examination apparatus types are CT, and the apparatus type matching coefficient α is determined to be 1.0.

According to the parts obtained by word segmentation, a part vector space is constructed, the parts are columns, and A, B are rows. The two examined bit sequences are assigned an initial value of 0 or 1, respectively. According to the examination item professional name library check, if there is an affiliation between columns (parts), most of the attributes extend downward to sub-parts (sub-part attributes cannot extend upward). To distinguish the effect of the checking order on the similarity of the checking items, the current item a is taken to be 1.2 when it extends downward, and the history item B is taken to be 0.8 when it extends downward. From this, the location feature vectors of A and B are obtained, as shown in Table 2, BD_a＝[1,1],BD_b＝[1,0.8]。

TABLE 2 site vector space Structure

Examination of item/site	Skull	Pituitary gland
			A	1	1
B	1	0.8

Similarly, an inspection method vector space is constructed according to the inspection method obtained by word segmentation, the inspection method is column, and A, B is row. The inspection method columns of the two inspections are given initial values of 0 or 1, respectively, to thereby obtain inspection method feature vectors of A and B, as shown in Table 3, M_a＝[1,1],M_b＝[1,0]。

TABLE 3 inspection method vector space construction

Inspection item/inspection method	Flat broom	Enhancement
			A	1	1
B	1	0

β＝COS(BD_a,BD_b)＝1.8/(2*1.64)^1/2＝1.8/1.812＝0.993

γ＝COS(M_a,M_b)＝1/1.414＝0.707

The total similarity Φ calculation is as follows:

Φ＝α*(w₁β+w₂γ)

Here, the total similarity Φ is 1.0 (2/3 0.993+1/3 0.707) 0.898. If Th1 takes (0.8-0.9) median value 0.85, Φ is 0.898> Th1, and current inspection a is determined as the repeat inspection of recent inspection B.

In the foregoing exemplary application scenario one, a prompt for repeat examination needs to be given to the order of the doctor, and the doctor is expected to recall the recent examination and decide whether to order the examination as appropriate, so as to reduce unnecessary repeat examinations. In the above-mentioned typical application scenario two, the suspected repeated examination list is automatically listed for medical and political examiner to perform repeated examination post-affair audit by combining with the electronic medical record of the patient, so as to improve the examination efficiency.

The foregoing descriptions of the embodiments of the present invention are provided for illustration purposes and not for the purpose of limiting the invention as defined by the appended claims.

Claims

1. A method for rapidly judging similarity of medical examination items is characterized by comprising the following steps:

step 1, acquiring a current inspection project name A and a historical inspection project name B to be compared, and completing automatic word segmentation of the two inspection project names;

step 2, comparing with a professional name word library of the inspection project, classifying according to the equipment type, the part and the inspection method, and finishing synonym substitution;

step 3, checking the device type, and determining a device type matching coefficient alpha; when the types of the two inspection devices are the same, taking alpha as 1; otherwise, taking alpha as 0.5;

step 4, respectively constructing two characteristic vectors of the part and the inspection method, and calculating the similarity;

constructing a position vector space according to the positions obtained by word segmentation, wherein the positions are columns, and A, B are rows; respectively assigning an initial value of 0 or 1 to the two checked position sequences; checking according to a professional name word library of the inspection item, if the columns have an affiliation, extending most of the attributes downwards to the sub-parts; in order to distinguish the influence of the checking sequence on the similarity of the checking items, the name A of the current checking item is 1.2 when extending downwards, and the name B of the historical checking item is 0.8 when extending downwards; thus obtaining the position feature vectors of A and B;

similarly, according to the examination method obtained by word segmentation, an examination method vector space is constructed, the examination method is column, and A, B is row; assigning initial values of 0 or 1 to the inspection method lists of the two inspections respectively, thereby obtaining inspection method feature vectors of A and B;

respectively calculating the part similarity beta and the inspection method similarity gamma according to a cosine measurement formula of the vector;

step 5, comprehensively calculating the similarity, and carrying out repeated inspection and judgment;

the total similarity Φ calculation is as follows:

Φ＝α*(w₁β+w₂γ)

wherein, w₁And w₂Is a weight coefficient;

two thresholds Th1 and Th2 are established, Th1> Th 2; when the total similarity Φ > Th1, a duplicate check may be determined; and when the Th1 is larger than or equal to the total similarity phi > Th2, the suspected duplicate check is judged.