Similar text retrieval method and system

Abstract

The invention discloses a similar text retrieval method and system, and relates to the technical field of data processing and information retrieval, and the similar text retrieval method comprises the following steps: analyzing artery invasive blood pressure and cuff blood pressure readings in the same time window from a diagnosis and treatment text fragment, calculating fragment-level dynamic cuff pressure and pulse pressure anomalous quantity, and calculating the fragment-level dynamic cuff pressure and pulse pressure anomalous quantity; carrying out statistics and aggregation to generate a patient-level dynamic cuff pressure and pulse pressure abnormal consistency index; basic text similarity of the target test condition and the diagnosis and treatment text is calculated, constraint processing is conducted on the basic text similarity through the consistency index to obtain a retrieval score, and finally the patients are sorted according to the score. According to the method, the virtual height of text evidence caused by dynamic cuff pressure and pulse pressure anomalous due to measurement link artifacts can be recognized and inhibited, candidate error recall is reduced, and the accuracy and reliability of screening and sorting are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and information retrieval, in particular to a similar text retrieval method and system. BACKGROUND

[0002] In the field of intensive care medicine and clinical data mining, drug trials or hemodynamic studies often require screening of patients from massive hospital databases that meet specific inclusion criteria, such as locating potential subjects with specific mean arterial pressure maintenance levels or using pressor drugs; this process mainly relies on searching and analyzing the diagnosis and treatment text in electronic medical records, calculating the text similarity between the target trial conditions and the patient's clinical records, and recalling text segments containing relevant keywords and numerical descriptions to construct a candidate patient list.

[0003] In the intensive care unit, there are two parallel measurement links for arterial invasive blood pressure and non-invasive cuff blood pressure, which are easily affected by physical factors such as pipeline damping or sensor position, resulting in measurement artifacts. A typical manifestation is that the difference between systolic and diastolic pressures displayed by the invasive measurement link is abnormally narrowed, while the non-invasive measurement link reads normally, resulting in a significant difference in mean pressure values between the two. Existing retrieval methods cannot identify this inconsistency in readings caused by the measurement system itself, and are prone to misjudging such artifact records containing false low blood pressure values as strong evidence of meeting the criteria for circulatory collapse, thereby giving such text an excessively high similarity score, resulting in a large amount of data noise in the final output candidate patient list that does not meet the true inclusion criteria, seriously affecting the ranking accuracy of the screening results and the inclusion efficiency of clinical research. SUMMARY

[0004] The purpose of the present application is to solve the problem in the prior art that the arterial invasive blood pressure and cuff blood pressure readings in the same time window are not physically consistent, resulting in the inability to identify abnormal phenomena caused by measurement link artifacts, and thus the text containing false clinical values is misassigned a high similarity score, resulting in a high false positive rate and low ranking accuracy of the candidate patient screening results. The present application provides a similar text retrieval method and system.

[0005] To solve the problems existing in the prior art, the present application adopts the following technical solutions:

[0006] The similar text retrieval method comprises:

[0007] S1, obtaining a set of diagnosis and treatment text segments of a patient to be screened, and parsing the arterial invasive blood pressure readings and cuff blood pressure readings in the same text segment from the set of diagnosis and treatment text segments;

[0008] S2, calculating segment-level dynamic cuff pressure pulse pressure abnormality constant based on the arterial invasive blood pressure reading and the cuff blood pressure reading;

[0009] S3, statistically aggregating the segment-level dynamic cuff pressure pulse pressure abnormality constant to generate patient-level dynamic cuff pressure pulse pressure abnormality consistency index;

[0010] S4, calculating basic text similarity according to target test condition text and diagnosis and treatment text segment set;

[0011] S5, performing constraint processing on the basic text similarity according to the patient-level dynamic cuff pressure pulse pressure abnormality consistency index to obtain a retrieval score;

[0012] S6, sorting patients according to the retrieval score to obtain a candidate list.

[0013] Preferably, the arterial invasive blood pressure reading includes arterial systolic pressure, arterial diastolic pressure and arterial mean pressure; the cuff blood pressure reading includes cuff systolic pressure, cuff diastolic pressure and cuff mean pressure; when the cuff mean pressure is not contained in the text segment, the cuff mean pressure is determined based on the cuff systolic pressure and the cuff diastolic pressure.

[0014] Preferably, determining the cuff mean pressure based on the cuff systolic pressure and the cuff diastolic pressure comprises:

[0015] calculating the difference between the cuff systolic pressure and the cuff diastolic pressure, adding one-third of the difference to the cuff diastolic pressure to obtain the cuff mean pressure.

[0016] Preferably, calculating the segment-level dynamic cuff pressure pulse pressure abnormality constant based on the arterial invasive blood pressure reading and the cuff blood pressure reading comprises:

[0017] calculating the difference between the arterial systolic pressure and the arterial diastolic pressure to obtain arterial invasive pulse pressure;

[0018] calculating the difference between the cuff systolic pressure and the cuff diastolic pressure to obtain cuff pulse pressure;

[0019] taking the natural logarithm of the ratio of the cuff pulse pressure to the arterial invasive pulse pressure to obtain a first component;

[0020] calculating the absolute value of the difference between the arterial mean pressure and the cuff mean pressure, and taking the ratio of the absolute value to the cuff mean pressure as a second component;

[0021] adding the first component and the second component to obtain the segment-level dynamic cuff pressure pulse pressure abnormality constant.

[0022] Preferably, statistically aggregating the segment-level dynamic cuff pressure pulse pressure abnormality constant to generate the patient-level dynamic cuff pressure pulse pressure abnormality consistency index comprises:

[0023] taking the median of all segment-level dynamic cuff pressure pulse pressure abnormality constants of the patient as a central value;

[0024] Calculate the median of the absolute value of the difference between the central value and the scale value of all segment-level dynamic sleeve pressure pulse pressure abnormality constants to obtain a scale value;

[0025] According to the central value and the scale value, each segment-level dynamic sleeve pressure pulse pressure abnormality constant is standardized, and the standardized result is mapped to the interval of zero to one through the Sigmoid function;

[0026] The median of the mapped result is taken as the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index.

[0027] Preferably, the basic text similarity is calculated according to the target test condition text and the set of diagnosis and treatment text segments, including:

[0028] Obtain the target test condition text;

[0029] Respectively construct the query vector of the target test condition text and the document vector of each diagnosis and treatment text segment;

[0030] Calculate the cosine similarity of the query vector and each document vector;

[0031] The maximum value of the cosine similarity corresponding to all diagnosis and treatment text segments of the patient is determined as the basic text similarity of the patient.

[0032] Preferably, the basic text similarity is constrained according to the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index to obtain a retrieval score, including:

[0033] Calculate the difference between 1 and the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index to obtain a reliability coefficient;

[0034] The product of the basic text similarity and the reliability coefficient is taken as the retrieval score.

[0035] In order to solve the above problems, the present application also provides a similar text retrieval system, the system comprises:

[0036] The data analysis module is used for obtaining the set of diagnosis and treatment text segments of the patient to be screened, and analyzing the arterial invasive blood pressure reading and the cuff blood pressure reading in the same text segment from the set of diagnosis and treatment text segments;

[0037] The anomaly calculation module is used for calculating the segment-level dynamic sleeve pressure pulse pressure abnormality constant based on the arterial invasive blood pressure reading and the cuff blood pressure reading;

[0038] The index construction module is used for statistically aggregating the segment-level dynamic sleeve pressure pulse pressure abnormality constant to generate the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index;

[0039] The text matching module is used for calculating the basic text similarity according to the target test condition text and the set of diagnosis and treatment text segments;

[0040] a score correction module configured to constrain the base text similarity according to the patient-level arterial-sleeve pressure pulse pressure abnormality consistency index, to obtain a retrieval score;

[0041] a result ranking module configured to rank the patients according to the retrieval score, to obtain a candidate list.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] 1. The present application introduces a physical consistency verification mechanism for arterial invasive blood pressure and cuff blood pressure readings in the text retrieval process. By analyzing the blood pressure data of the two measurement links within the same time window, a segment-level arterial-sleeve pressure pulse pressure abnormality is constructed. This abnormality can quantify the abnormal collapse degree of invasive pulse pressure relative to cuff pulse pressure and the difference in average pressure values, thereby accurately identifying the arterial-sleeve pressure pulse pressure abnormality phenomenon caused by measurement link artifacts. Further, the patient-level arterial-sleeve pressure pulse pressure abnormality consistency index generated by robust statistical aggregation can evaluate the reliability of blood pressure data in patient diagnosis and treatment text from a global perspective, effectively avoiding false positives caused by single measurement link artifacts or occasional data anomalies, and significantly improving the ability to identify artifact records containing false low blood pressure values.

[0044] 2. The present application converts the patient-level arterial-sleeve pressure pulse pressure abnormality consistency index into a reliability coefficient and uses the coefficient to constrain the content-based base text similarity, thereby achieving dynamic correction of the retrieval score. In the case of high text content matching but poor physical consistency, the retrieval score will be correspondingly reduced, thereby suppressing the false high ranking of artifact samples. In the case of text content matching and good physical consistency, the retrieval score remains at a high level. The final candidate list is sorted according to the corrected retrieval score, ensuring that the top-ranked patients are more consistent with the real clinical enrollment conditions, significantly reducing the false positive rate of candidate patient screening results, and improving the ranking accuracy of the screening results and the enrollment efficiency of the clinical study. BRIEF DESCRIPTION OF DRAWINGS

[0045] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application. In the drawings:

[0046] Figure 1 A flowchart of a similar text retrieval method provided by an embodiment of the present application is shown in the figure;

[0047] Figure 2 A functional module diagram of a similar text retrieval system provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application.

[0049] Embodiment: The present embodiment provides a similar text retrieval method, see Figure 1 , specifically, comprising:

[0050] S1, obtaining a diagnosis and treatment text segment set of a patient to be screened, and parsing an arterial invasive blood pressure reading and a cuff blood pressure reading in a same text segment from the diagnosis and treatment text segment set;

[0051] S2, calculating a segment-level arterial-cuff pressure pulse pressure abnormality constant based on the arterial invasive blood pressure reading and the cuff blood pressure reading;

[0052] S3, statistically aggregating the segment-level arterial-cuff pressure pulse pressure abnormality constant to generate a patient-level arterial-cuff pressure pulse pressure abnormality consistency index;

[0053] S4, calculating a basic text similarity according to a target test condition text and the diagnosis and treatment text segment set;

[0054] S5, performing constraint processing on the basic text similarity according to the patient-level arterial-cuff pressure pulse pressure abnormality consistency index to obtain a retrieval score;

[0055] S6, sorting the patients according to the retrieval score to obtain a candidate list.

[0056] In the embodiments of the present application, the diagnosis and treatment text segment set of the patient to be screened is obtained, and the arterial invasive blood pressure reading and the cuff blood pressure reading in the same text segment are parsed from the diagnosis and treatment text segment set, comprising:

[0057] From the database of the hospital, the diagnosis and treatment text segment set of each patient to be screened that has been authorized by the patient for use is obtained. Each diagnosis and treatment text segment in the set is derived from the clinical diagnosis and treatment related records generated by the patient during the hospitalization in the ward, specifically including the vital sign monitoring records generated by the monitor every hour, the daily physician ward round records, the medical order execution records, and the laboratory examination associated records related to the patient's condition. Each patient to be screened corresponds to a unique diagnosis and treatment text segment set, and each text segment contains time stamp information accurate to minutes to ensure that different text segments of the same patient can be associated in chronological order. At the same time, the time window of the same text segment is set to be consecutive ten minutes, and only the text segments with simultaneous arterial invasive blood pressure and cuff blood pressure related descriptions in the time window are retained for subsequent analysis. The text segments containing only a single blood pressure type reading are marked as invalid and excluded.

[0058] The regular expression-based regularization parsing algorithm is used to traverse and parse each valid text segment in the set of diagnosis and treatment text segments, to identify and extract the arterial invasive blood pressure reading and cuff blood pressure reading existing in the same text segment. In the parsing process, three groups of core regular expression templates are used for keyword matching. The first group is used to match arterial invasive blood pressure keywords, specifically, regular expressions matching IBP or Aline or invasive arterial pressure or arterial pressure, etc. When the blood pressure value near these keywords is parsed, the part of the blood pressure value is the arterial invasive blood pressure reading. The second group is used to match cuff blood pressure keywords, specifically, regular expressions matching NIBP or cuff or non-invasive cuff pressure or cuff blood pressure, etc. When the blood pressure value near these keywords is parsed, the part of the blood pressure value is the cuff blood pressure reading. The third group is used to match the common writing forms of blood pressure values in clinical practice, specifically, regular expressions matching the writing form of BP followed by the systolic and diastolic pressure value combination separated by a slash, and the writing form of separately labeling systolic, diastolic and mean pressure. Through the cooperative matching of the above three groups of templates, the arterial invasive blood pressure reading and cuff blood pressure reading can be accurately extracted.

[0059] For the case of multiple groups of blood pressure values appearing in the same text segment, the blood pressure value group with the timestamp closest to the start time of the text segment is taken as the valid data. If there are multiple groups of blood pressure values with the same timestamp, the average value of each group of values is calculated as the final parsing result. After parsing, the validity is checked by comparing whether the difference between the systolic and diastolic pressures of the arterial invasive blood pressure is within the clinically reasonable range of 20 to 80 mmHg, and whether the difference between the systolic and diastolic pressures of the cuff blood pressure is within the clinically reasonable range of 30 to 90 mmHg. Text segments that fail the check are marked as abnormal and excluded. The arterial invasive blood pressure reading includes arterial systolic pressure, arterial diastolic pressure and arterial mean pressure, which are directly extracted from the corresponding text segment. The cuff blood pressure reading includes cuff systolic pressure, cuff diastolic pressure and cuff mean pressure, of which the cuff systolic pressure and cuff diastolic pressure are directly extracted from the corresponding text segment.

[0060] When it is found in the parsing process that a text segment only contains cuff systolic pressure and cuff diastolic pressure without explicitly containing cuff mean pressure, the cuff mean pressure is determined based on the extracted cuff systolic pressure and cuff diastolic pressure through a pre-set calculation method. The specific calculation method uses the formula:

[0061]

[0062] wherein represents the cuff mean pressure; represents the cuff diastolic pressure parsed from the text segment, with the unit of mmHg; Represent the cuff systolic pressure parsed in the text segment; in the field of clinical medicine, the physiological definition of mean arterial pressure is the average of arterial blood pressure in a cardiac cycle, and its value is directly related to the cardiac pumping function and peripheral vascular resistance, and according to the clinically recognized blood pressure calculation specification, the mean arterial pressure of adult patient population is equal to one third of the diastolic pressure plus the pulse pressure, wherein the pulse pressure is the difference between the systolic pressure and the diastolic pressure, and the calculation method is verified by a large amount of clinical data, and the formula calculation process is simple and does not need complex operation, which is convenient for rapid implementation in clinical information system, and ensures that each calculation based on arterial invasive blood pressure reading and cuff blood pressure reading has complete and accurate parameter basis.

[0063] It should be noted that the patient to be screened refers to a potential patient in the hospital intensive care unit who meets the target drug test inclusion criteria, the set of diagnosis and treatment text segments is a unique set of clinical diagnosis and treatment related text segments corresponding to each patient to be screened, and the text segments are derived from various clinical records generated by the patient during hospitalization in the intensive care unit. The diagnosis and treatment text segment is a single text unit in the diagnosis and treatment text segment set, each text segment includes time stamp information accurate to minutes, the time window is a continuous time period defining the time range of the same text segment, which is set to ten minutes in this embodiment. Only the text segment that simultaneously appears in the time period is retained as an effective text segment. The invalid text segment refers to a text segment that only contains a single blood pressure type reading and does not meet the time window requirement.

[0064] It should be noted that the arterial invasive blood pressure reading refers to the blood pressure value set obtained by transmitting the pressure in the patient's arterial lumen to the pressure sensor through the arterial catheter and the liquid pressure transmission pipeline, which usually includes arterial systolic pressure, arterial diastolic pressure and arterial mean pressure, and is directly represented by the invasive measurement link to represent the peak, valley and low frequency baseline component of the arterial pressure waveform in a cardiac cycle. The cuff blood pressure reading refers to the blood pressure value set obtained by applying external pressure to the limb artery through the inflatable cuff and calculating the blood flow oscillation or related signal change, which usually includes cuff systolic pressure, cuff diastolic pressure and cuff mean pressure, and is an indirect estimation result of the same circulatory system pressure state by non-invasive measurement link. Since the physical measurement links corresponding to the two types of readings are different, one is more significantly affected by the equivalent compliance and damping of the liquid pressure transmission system, and the other is more significantly affected by the sleeve position and tissue mechanical conditions. Therefore, when both types of readings are parsed in the diagnosis and treatment text, they can be used to characterize the relative consistency of the two measurement links for the same blood pressure state.

[0065] It should be noted that the arterial systolic pressure is the value in the arterial invasive blood pressure reading that reflects the highest value of the arterial blood pressure in the systolic phase of the heart, the arterial diastolic pressure is the value in the arterial invasive blood pressure reading that reflects the lowest value of the arterial blood pressure in the diastolic phase of the heart, and the arterial mean pressure is the value in the arterial invasive blood pressure reading that reflects the average value of the arterial blood pressure in one cardiac cycle of the heart; the cuff systolic pressure is the value in the cuff blood pressure reading that reflects the highest value of the cuff measured blood pressure in the systolic phase of the heart, the cuff diastolic pressure is the value in the cuff blood pressure reading that reflects the lowest value of the cuff measured blood pressure in the diastolic phase of the heart, and the cuff mean pressure is the value in the cuff blood pressure reading that reflects the average value of the cuff measured blood pressure in one cardiac cycle of the heart.

[0066] In the embodiments of the present application, the segment-level arterial-cuff pulse pressure abnormality quantity is calculated based on the arterial invasive blood pressure reading and the cuff blood pressure reading, comprising:

[0067] After the arterial invasive blood pressure reading and the cuff blood pressure reading are extracted from the effective diagnosis and treatment text segment, the segment-level arterial-cuff pulse pressure abnormality quantity is calculated, and the specific calculation process is as follows: first, the arterial invasive pulse pressure is calculated using the formula , wherein represents the arterial invasive pulse pressure, represents the arterial systolic pressure parsed from the effective text segment, represents the arterial diastolic pressure parsed from the effective text segment; in the clinical medical field, the physiological definition of pulse pressure is the difference between the systolic pressure and the diastolic pressure in one cardiac cycle of the heart, and the arterial invasive pulse pressure can directly reflect the fluctuation amplitude of the blood pressure in the arterial invasive measurement link, which is a core quantitative index for judging whether the arterial invasive pulse pressure appears abnormal collapse. This calculation method completely follows the clinically recognized definition of pulse pressure, is simple and intuitive in logic, can be directly calculated by parsing the arterial systolic pressure and the arterial diastolic pressure, and can ensure the accuracy of the calculation result.

[0068] The cuff pulse pressure is calculated using the formula , wherein represents the cuff pulse pressure, represents the cuff systolic pressure parsed from the effective text segment, represents the cuff diastolic pressure parsed from the effective text segment. The reason for calculating the cuff pulse pressure using this formula is consistent with that for calculating the arterial invasive pulse pressure. The cuff pulse pressure is used to reflect the fluctuation amplitude of the blood pressure in the cuff non-invasive measurement link, and forms a corresponding comparison with the arterial invasive pulse pressure, thereby providing a basis for subsequently judging whether the pulse pressures of the two exist abnormal relative relationship.

[0069] The first component is calculated using the formula , wherein represents the first component, represents the cuff pulse pressure calculated above, represents the arterial invasive pulse pressure calculated above, represents a very small positive number, and the specific value is 0.000001, which is much smaller than the conventional scale of blood pressure values and can be ignored compared with blood pressure values; the reason for calculating the first component using this formula is that the core role of the first component is to quantify the relative difference between cuff pulse pressure and arterial invasive pulse pressure. When the arterial invasive pulse pressure collapses abnormally, the arterial invasive pulse pressure will be significantly smaller than the cuff pulse pressure. At this time, the ratio of the cuff pulse pressure and the arterial invasive pulse pressure after being added together will be greater than 1, and the natural logarithm result will be a positive number. The more significant the relative difference between the two, the larger the ratio, and the larger the value of the first component, which can accurately capture the characteristics of the abnormal collapse of the arterial invasive pulse pressure. The purpose of introducing is to avoid division by zero when the arterial invasive pulse pressure or the cuff pulse pressure is 0, to prevent calculation deviation caused by extreme values. The reason for using the natural logarithm is that it can convert the nonlinear ratio relationship between the cuff pulse pressure and the arterial invasive pulse pressure into a linear logarithmic relationship, which is convenient for subsequent numerical processing and aggregation. The monotonicity of the natural logarithm can ensure the positive correlation between the relative difference and the value of the first component, ensuring accurate quantification of the abnormal difference in pulse pressure.

[0070] The second component is calculated using the formula wherein represents the second component, represents the arterial mean pressure parsed in the effective text segment, represents the cuff mean pressure parsed in the effective text segment, represents the same very small positive number 0.000001 as described above. The reason for calculating the second component using this formula is that the second component is used to quantify the divergence between the arterial mean pressure and the cuff mean pressure. The absolute value of the difference between the arterial mean pressure and the cuff mean pressure can directly reflect the absolute divergence between the two. The ratio of the absolute value of the difference and the absolute value of the cuff mean pressure after being added can eliminate the influence of the size of the cuff mean pressure itself, achieve normalization, keep the numerical range of the second component stable, adapt to patients with different basic blood pressure levels, and avoid inaccurate quantification of divergence due to differences in patients' basic blood pressure. The purpose of introducing is also to avoid division by zero when the cuff mean pressure is 0, to ensure the stability of the calculation process.

[0071] The first component and the second component are added to obtain a segment-level dynamic sleeve pressure pulse pressure abnormality constant corresponding to the effective diagnosis and treatment text segment, the segment-level dynamic sleeve pressure pulse pressure abnormality constant needs to comprehensively and comprehensively reflect the abnormal characteristics of the dynamic sleeve pressure pulse pressure in the same text segment, and the abnormal characteristics mainly manifest in two dimensions of abnormal collapse of the arterial invasive pulse pressure and significant difference of the dynamic sleeve pressure average pressure. The first component has accurately quantified the relative abnormal difference of the pulse pressure, and the second component has accurately quantified the difference degree of the average pressure. Adding the two can integrate the abnormal characteristics of the two dimensions into a single quantitative index, which not only guarantees the comprehensiveness of capturing the abnormal characteristics, but also makes the quantitative result simple and intuitive, facilitating subsequent statistical aggregation of the abnormality constants corresponding to multiple text segments of the same patient to generate a patient-level dynamic sleeve pressure pulse pressure abnormality consistency index that can reflect the overall abnormality of the patient.

[0072] It should be noted that the arterial invasive pulse pressure is the difference between the systolic pressure and the diastolic pressure of the artery in the same text segment, which is the pressure amplitude caused by the ejection of the arterial pressure wave in a cardiac cycle, reflecting the pressure fluctuation amplitude formed by the instantaneous elastic energy storage and release in the blood vessel. The cuff pulse pressure is the difference between the systolic pressure and the diastolic pressure of the cuff in the same text segment, which is the characterization of the same pulsatile pressure amplitude by the non-invasive cuff measurement link. The first component is the natural logarithm of the ratio of the cuff pulse pressure to the arterial invasive pulse pressure, which is a logarithmic scale that characterizes the relative scaling relationship of the two measurement links to the pulsatile amplitude. When the arterial invasive link is compressed due to changes in the equivalent compliance or damping of the fluid pressure transmission system, the logarithmic value of the ratio increases, indicating that the invasive pulse pressure is relatively collapsed. The second component is the absolute value of the difference between the arterial mean pressure and the cuff mean pressure, which is normalized with respect to the cuff mean pressure, which characterizes the deviation of the two measurement links from the same static pressure baseline or low-frequency pressure component, and is used to represent the divergence strength of the invasive link reading from the non-invasive link in the mean pressure level. The segment-level dynamic sleeve pressure pulse pressure abnormality constant is a comprehensive quantity obtained by adding the first component and the second component, which simultaneously measures the relative collapse of the pulsatile amplitude and the relative divergence of the mean pressure in the same text segment, thereby forming a quantitative representation of the abnormal consistency degree between the arterial invasive blood pressure and the cuff blood pressure, indicating that the blood pressure readings in the segment are more likely to be affected by the physical transmission characteristics of the measurement link and produce apparent changes inconsistent with the true circulation state.

[0073] In an embodiment of the present application, the segment-level dynamic sleeve pressure pulse pressure abnormality constant is statistically aggregated to generate a patient-level dynamic sleeve pressure pulse pressure abnormality consistency index, comprising:

[0074] After calculating the segment-level dynamic-systolic-pulse pressure abnormality of all valid diagnosis and treatment text fragments of a certain patient to be screened, first, the segment-level dynamic-systolic-pulse pressure abnormality corresponding to all valid text fragments of the patient is collected to form a segment-level abnormality set of the patient, and then the patient-level dynamic-systolic-pulse pressure abnormality consistency index is calculated. The specific process is as follows. First, the central value is calculated. All segment-level dynamic-systolic-pulse pressure abnormalities of the patient are sorted in ascending order. If the number of elements after sorting is odd, the number at the middle position is taken as the central value . If the number of elements after sorting is even, the arithmetic mean of the two middle numbers is taken as the central value . The median is a robust statistic. Compared with the arithmetic mean, the median is not affected by the extreme abnormal segment-level abnormality and can accurately reflect the central tendency of the segment-level abnormality of the patient. It is suitable for scenarios where the number of valid text fragments of different patients is significantly different and the distribution of abnormality is different. It avoids the deviation of the central value from the true level caused by the abnormality of individual extreme outliers and ensures the reliability of the subsequent calculation.

[0075] The scale value is calculated according to the formula , where represents the scale value of the patient, represents the absolute value of the difference between each segment-level abnormality and the central value, represents a very small positive number, specifically 0.000001. This value is much smaller than the normal value scale of the abnormality and can be ignored compared with the abnormality value. The result of calculating the median of all absolute values is taken as the median absolute deviation, and then is added to obtain the scale value. The reason is that the median absolute deviation is also a robust statistic. Compared with the standard deviation, it is more resistant to extreme values and can accurately reflect the dispersion of the segment-level abnormality of the patient around the central value. The core purpose of adding is to avoid the division by zero error in the subsequent standardization process when the median absolute deviation is 0. Then each segment-level abnormality is standardized by the formula , where represents the standardized segment-level abnormality, u represents the unique identifier of the patient to be screened, and i represents the index of the valid text fragment. The reason for standardizing by this formula is that the segment-level abnormality of different patients to be screened is significantly different in value range and distribution characteristics due to differences in basic blood pressure level, recording quantity, writing habit, etc. By subtracting the central value and dividing by the scale value, the dimensional and distribution differences of the abnormality of different patients can be eliminated, the normalized processing of the patient is realized, the standardized abnormality has cross-patient comparability, and the deviation of each segment-level abnormality from the overall level of the patient is preserved.

[0076] The normalized result is mapped to the interval of zero to one by a Sigmoid function, using the formula wherein represents the mapped abnormal intensity value, represents a natural constant, and the reason for using the Sigmoid function for mapping is that the normalized can take any real number in theory, while the abnormal intensity needs to be quantified as a value in the interval of zero to one to represent the abnormal degree of the segment-level abnormality, the Sigmoid function has the characteristic of monotonically increasing, which can ensure that the larger the normalized value, the closer the mapped abnormal intensity value to 1, and the smaller the normalized value, the closer the mapped abnormal intensity value to 0, and the change rate of the function near zero is moderate, which can accurately distinguish different degrees of abnormal characteristics, and the function value converges to 0 and 1, avoiding overflow of the abnormal intensity value out of the reasonable interval caused by extreme normalized values.

[0077] The median of the abnormal intensity values of each segment-level abnormality after normalization and mapping is taken as the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index, and the reason for using the median to calculate the mapped result is to avoid the deviation of individual extreme abnormal intensity values on the overall abnormality consistency evaluation of the patient, and the median can accurately reflect the overall abnormality consistency level of all effective text segments of the patient, and the larger the value of the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index, the more significant the dynamic sleeve pressure pulse pressure abnormality consistency of the patient, and vice versa. The index can comprehensively and stably quantify the patient's dynamic sleeve pressure pulse pressure abnormality, providing a reliable quantitative basis for subsequent constraint processing of the basis text similarity.

[0078] It should be noted that the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index is a comprehensive indication quantity in the range of zero to one obtained by stably statistically aggregating the segment-level dynamic sleeve pressure pulse pressure abnormality of the same patient in multiple diagnosis and treatment text segments, which quantitatively represents the degree of systemic inconsistency between the arterial invasive blood pressure measurement link and the cuff blood pressure measurement link of the patient within the observation period, wherein the segment-level dynamic sleeve pressure pulse pressure abnormality reflects the relative collapse of the pulsatile pressure amplitude level and the relative divergence of the average pressure baseline level, so when the patient-level index has a higher value, it means that the combination mode of significant contraction of the invasive pulse pressure relative to the cuff pulse pressure and more obvious deviation of the invasive average pressure from the cuff average pressure repeatedly appears in multiple segments. This combination mode corresponds to the transmission characteristics of the equivalent compliance and damping of the invasive pressure transmission system, which more frequently shows compression of the pressure waveform peak and valley and deviation of the baseline component in the patient's record, so that the invasive reading is more likely to carry apparent deviations caused by the transmission characteristics of the measurement link rather than real circulatory state changes, and therefore the index can be used as a consistency indication quantity for measuring the reliability of the related blood pressure text evidence of the patient.

[0079] In the embodiments of the present application, the similarity of the basic text is calculated according to the target test condition text and the set of diagnosis and treatment text segments, including:

[0080] The target test condition text is obtained, which is a structured clinical screening condition description text based on the enrollment criteria of the ICU inotrope or volume reactivity related test, containing test enrollment core keywords such as inotrope name, mean arterial pressure value range, volume reactivity evaluation related description, etc.; then the query vector of the target test condition text and the document vector of each diagnosis and treatment text segment are constructed, and the construction process is realized based on the TF-IDF algorithm. First, a unified medical vocabulary table is constructed, which is composed of medical terms, numerical descriptions, drug names and other core vocabularies in the diagnosis and treatment text segments of all patients to be screened and the target test condition text. Each term in the vocabulary table corresponds to a unique dimension index. Then, the target test condition text and each diagnosis and treatment text segment are subjected to word segmentation processing. The accurate word segmentation algorithm based on the medical field special word segmentation dictionary can accurately identify medical professional terms such as norepinephrine and mean arterial pressure, avoiding vector construction deviation caused by word segmentation errors.

[0081] The term frequency (TF) and inverse document frequency (IDF) of each term are calculated to determine the TF-IDF weight, wherein the term frequency calculation formula is , represents the term frequency of term t in text d, represents the number of times that term t appears in text d, and the denominator is the sum of the number of times that all terms appear in text d. This formula quantifies the frequency of the term in a single text. The higher the term frequency, the more core the term is in the text; the inverse document frequency calculation formula is , represents the inverse document frequency of term t, N is the total number of all texts to be processed (target test condition text + all diagnosis and treatment text segments), is the number of texts containing term t, is a very small positive number, which is 0.000001, used to avoid calculation errors when the denominator is 0. This formula quantifies the scarcity of the term in all texts. The higher the inverse document frequency, the stronger the discrimination of the term; based on the term frequency and the inverse document frequency, the TF-IDF weight of each term is calculated, and the formula is , represents the TF-IDF weight of term t in text d. This weight takes into account the local frequency and global scarcity of the term, and can accurately represent the core semantic features of the text; based on this weight, the query vector of the target test condition text and the document vector of the i-th diagnosis and treatment text segment of patient u , both of which have the same dimension as the aforementioned medical vocabulary table, and each dimension in the vector corresponds to the TF-IDF weight of the corresponding term in the vocabulary table. The weight of a term not appearing in the text is 0.

[0082] The cosine similarity between the query vector and each document vector is then calculated, and the formula is , where the numerator is the dot product of the query vector and the document vector, that is, the sum of the products of the corresponding dimension values of the two vectors, and the denominator and are the L2 norms of the query vector and the document vector, respectively, and the formula for calculating the L2 norm is , which represents the square root of the sum of the squares of the values of each dimension of the vector. The reason for using cosine similarity calculation is that it can effectively measure the angle between two vectors, with a value range of 0 to 1. The closer the value is to 1, the higher the semantic similarity between the two texts, and it is not affected by the vector dimension and length. It is suitable for medical text vocabulary with high dimension and large text length difference.

[0083] Finally, the maximum value of the cosine similarity corresponding to all the patient's diagnosis and treatment text segments is determined as the patient's basic text similarity. The reason for using the maximum value as the basic similarity is that in the trial enrollment screening scenario, as long as one of the patient's diagnosis and treatment text segments meets the core features of the trial condition text, the patient has the text basis to become a potential candidate. Taking the maximum value can accurately capture the patient's most matching trial condition diagnosis and treatment segment features, avoiding the overall score being dragged down by other low similarity segments. It conforms to the core logic that in clinical screening, a record that meets the conditions is considered a potential candidate. At the same time, it ensures that the basic similarity can intuitively reflect the text matching degree of the patient and the trial condition, providing accurate basic numerical values for subsequent reliability coefficient constraints.

[0084] It should be noted that the basic text similarity refers to the similarity quantification result calculated based on the semantic and term matching degree between the target trial condition text and the diagnosis and treatment text segment set of the patient to be screened, without introducing the dynamic sleeve pressure pulse pressure abnormal consistency constraint. It is used to represent the closeness of the patient's text in the content level and the information point level of the trial condition description. In practice, the target trial condition text can be constructed as a query vector and each diagnosis and treatment text segment can be constructed as a document vector. The cosine similarity between the query vector and the document vector is calculated to obtain the segment-level matching degree, and the maximum value of all segment-level matching degrees of the patient is taken as the basic text similarity of the patient, so that the basic text similarity reflects the text compliance degree of the most consistent segment in the patient's text set to the target trial condition.

[0085] In the embodiment of the present application, the basic text similarity is constrained according to the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index to obtain a retrieval score, which comprises:

[0086] After the calculation of the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index and the basic text similarity is completed, the calculation of the retrieval score is entered, and 1 is subtracted from the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index, and the difference obtained is used as a reliability coefficient, wherein the numerical range of the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index is zero to one, and the numerical size is positively correlated with the significant degree of the patient's dynamic sleeve pressure pulse pressure abnormality consistency, that is, the larger the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index, the more significant the patient's dynamic sleeve pressure pulse pressure abnormality phenomenon, and the higher the risk of false high text evidence caused by blood pressure measurement artifacts in the patient's diagnosis and treatment text segment, and the lower the retrieval reliability; on the contrary, the smaller, the weaker the patient's dynamic sleeve pressure pulse pressure abnormality phenomenon, and the higher the reliability of the diagnosis and treatment text segment, and by calculating the difference between 1 and the patient-level dynamic sleeve pressure pulse pressure abnormality consistency index, the reliability coefficient is positively correlated with the reliability of the patient's diagnosis and treatment text, and the numerical range is also zero to one, when the patient has no dynamic sleeve pressure pulse pressure abnormality phenomenon, the reliability coefficient tends to one, at this time, the basic text similarity can be completely preserved and does not need to be constrained; when the patient's dynamic sleeve pressure pulse pressure abnormality phenomenon is extremely significant, the reliability coefficient tends to zero, which can maximize the inhibition of the false high of the basic text similarity, realize the accurate constraint of the artifact sample, and quickly realize the quantitative evaluation of the reliability, and improve the core purpose of retrieval reliability.

[0087] The basic text similarity is multiplied by the reliability coefficient to obtain the final retrieval score of the patient, wherein the basic text similarity only reflects the semantic matching degree between the patient's diagnosis and treatment text and the target test condition text, and does not consider the false high similarity caused by blood pressure measurement artifacts, while the reliability coefficient has accurately quantified the reliability of the patient's diagnosis and treatment text, and by multiplying the two, the organic combination of the basic text similarity and the reliability can be realized, for the patient with dynamic sleeve pressure pulse pressure abnormality phenomenon, the reliability coefficient is less than one, and after multiplication, the final retrieval score can be reduced to suppress the false high similarity of the artifact sample; for the patient without abnormal phenomenon, the reliability coefficient tends to one, and the basic text similarity can be completely preserved to ensure that the patient who truly meets the inclusion criteria will not be misjudged.

[0088] It should be noted that the retrieval score is the final quantitative evaluation value of whether the patient to be screened is suitable for the target test condition, which is determined by the basic text similarity and the reliability coefficient. The basic text similarity is used to represent the matching degree of the patient's diagnosis and treatment text set with the target test condition text in the content level. The reliability coefficient is obtained by subtracting the patient's level from the arterial invasive blood pressure pulse pressure abnormality consistency index, which is used to represent the reliability of the blood pressure evidence in the patient's diagnosis and treatment text in the physical consistency level. Therefore, the meaning of the retrieval score is to jointly measure the similarity of the text content and the reliability of the blood pressure measurement link consistency. When the patient's text is highly similar to the test condition but also presents stronger arterial invasive blood pressure pulse pressure abnormality consistency, the retrieval score is correspondingly reduced due to the smaller reliability coefficient, thereby inhibiting the misleading of the apparent blood pressure abnormality caused by the measurement link transfer characteristics on the candidate ranking. When the patient's text is similar and the arterial invasive blood pressure pulse pressure abnormality consistency is weak, the retrieval score is higher to improve the priority of the real matching patient in the ranking.

[0089] In the embodiment of the present application, the patients are ranked according to the retrieval score to obtain a candidate list, comprising:

[0090] All the final retrieval scores corresponding to the patients to be screened are summarized, each final retrieval score is bound to a unique patient to be screened, ensuring a one-to-one correspondence between the patient and the retrieval score. Then, the final retrieval scores are arranged in descending order according to their numerical values from large to small. The patient to be screened corresponding to the final retrieval score with a larger numerical value is arranged in a more forward position, and the patient to be screened corresponding to the final retrieval score with a smaller numerical value is arranged in a more backward position. During the arrangement process, invalid patient data with repeated identification or missing retrieval score are removed to ensure the uniqueness and validity of the ranking result. After the ranking is completed, all the patients to be screened are integrated in the arrangement order to form a candidate list. The candidate list can be directly output as a screening candidate set. The patient to be screened in the front position of the list has a higher real matching degree, and the patient's diagnosis and treatment text is less disturbed by the arterial invasive blood pressure pulse pressure abnormality phenomenon, which can effectively reduce the risk of false recall caused by the pseudo-difference sample. The patient to be screened in the rear position of the list is reasonably restrained due to the more significant arterial invasive blood pressure pulse pressure abnormality phenomenon.

[0091] As shown in Figure 2 , it is a functional module diagram of a similar text retrieval system provided by an embodiment of the present application.

[0092] In this embodiment, the functions of each module / unit are as follows:

[0093] The data analysis module is used to obtain the diagnosis and treatment text segment set of the patient to be screened, and analyze the arterial invasive blood pressure reading and cuff blood pressure reading in the same text segment from the diagnosis and treatment text segment set;

[0094] an abnormality calculation module, configured to calculate a segment-level dynamic cuff pressure pulse pressure abnormality based on the arterial invasive blood pressure reading and the cuff blood pressure reading;

[0095] an index construction module, configured to statistically aggregate the segment-level dynamic cuff pressure pulse pressure abnormality to generate a patient-level dynamic cuff pressure pulse pressure abnormality consistency index;

[0096] a text matching module, configured to calculate a basic text similarity based on the target test condition text and the set of diagnosis and treatment text segments;

[0097] a score correction module, configured to perform constraint processing on the basic text similarity based on the patient-level dynamic cuff pressure pulse pressure abnormality consistency index to obtain a retrieval score;

[0098] a result sorting module, configured to sort the patients based on the retrieval score to obtain a candidate list.

[0099] The above merely provides the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and the inventive concept of the present application, can make equivalent replacements or changes within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method of similar text retrieval, characterized by, The method comprises the following steps: S1, obtaining a set of diagnosis and treatment text segments of a patient to be screened, and parsing the arterial invasive blood pressure readings and cuff blood pressure readings in the same text segment from the set of diagnosis and treatment text segments; S2, calculating the segment-level arterial-cuff pressure abnormality constant based on the arterial invasive blood pressure readings and the cuff blood pressure readings; S3, statistically aggregating the segment-level arterial-cuff pressure abnormality constant to generate a patient-level arterial-cuff pressure abnormality consistency index; S4, calculating a basic text similarity based on the target test condition text and the set of diagnosis and treatment text segments; S5, performing constraint processing on the basic text similarity based on the patient-level arterial-cuff pressure abnormality consistency index to obtain a retrieval score; S6, sorting the patients according to the retrieval score to obtain a candidate list.

2. The similar text retrieval method according to claim 1, characterized by, The arterial invasive blood pressure readings include arterial systolic pressure, arterial diastolic pressure and arterial mean pressure; the cuff blood pressure readings include cuff systolic pressure, cuff diastolic pressure and cuff mean pressure; when the text segment does not contain the cuff mean pressure, the cuff mean pressure is determined based on the cuff systolic pressure and the cuff diastolic pressure.

3. The similar text retrieval method according to claim 2, wherein Determining the cuff mean pressure based on the cuff systolic pressure and the cuff diastolic pressure comprises: calculating the difference between the cuff systolic pressure and the cuff diastolic pressure, adding one-third of the difference to the cuff diastolic pressure, and taking the addition result as the cuff mean pressure.

4. The similar text retrieval method according to claim 2, characterized by, Calculating the segment-level arterial-cuff pressure abnormality constant based on the arterial invasive blood pressure readings and the cuff blood pressure readings comprises: calculating the difference between the arterial systolic pressure and the arterial diastolic pressure to obtain an arterial invasive pulse pressure; calculating the difference between the cuff systolic pressure and the cuff diastolic pressure to obtain a cuff pulse pressure; taking the natural logarithm of the ratio of the cuff pulse pressure to the arterial invasive pulse pressure to obtain a first component; calculating the absolute value of the difference between the arterial mean pressure and the cuff mean pressure, and taking the ratio of the absolute value to the cuff mean pressure as a second component; adding the first component and the second component to obtain the segment-level arterial-cuff pressure abnormality constant.

5. The similar text retrieval method according to claim 4, wherein Statistically aggregating the segment-level arterial-cuff pressure abnormality constant to generate a patient-level arterial-cuff pressure abnormality consistency index comprises: taking the median of all segment-level arterial-cuff pressure abnormality constants of the patient as a center value; calculating the median of the absolute values of the differences between all segment-level arterial-cuff pressure abnormality constants and the center value to obtain a scale value; standardizing each segment-level arterial-cuff pressure abnormality constant based on the center value and the scale value, and mapping the standardized result to the interval of zero to one through a Sigmoid function; taking the median of the mapped result as the patient-level arterial-cuff pressure abnormality consistency index.

6. The similar text retrieval method according to claim 5, wherein Calculating the basic text similarity based on the target test condition text and the set of diagnosis and treatment text segments comprises: obtaining the target test condition text; constructing a query vector of the target test condition text and a document vector of each diagnosis and treatment text segment; calculating the cosine similarity of the query vector and each document vector; determining the maximum value of the cosine similarities corresponding to all diagnosis and treatment text segments of the patient as the basic text similarity of the patient.

7. The similar text retrieval method according to claim 6, wherein Performing constraint processing on the basic text similarity based on the patient-level arterial-cuff pressure abnormality consistency index to obtain a retrieval score comprises: calculating the difference between 1 and the patient-level arterial-cuff pressure abnormality consistency index to obtain a reliability coefficient; The product of the base text similarity and the reliability coefficient is used as a retrieval score.

8. A similar text retrieval system for use in the similar text retrieval method according to any one of claims 1 to 7, characterized by The system comprises: a data analysis module configured to obtain a set of diagnosis and treatment text segments of a patient to be screened, and analyze arterial invasive blood pressure readings and cuff blood pressure readings in the same text segment from the set of diagnosis and treatment text segments; an abnormality calculation module configured to calculate segment-level arterial-cuff pressure abnormality based on the arterial invasive blood pressure readings and the cuff blood pressure readings; an index construction module configured to statistically aggregate the segment-level arterial-cuff pressure abnormality to generate a patient-level arterial-cuff pressure abnormality consistency index; a text matching module configured to calculate a base text similarity based on a target test condition text and the set of diagnosis and treatment text segments; a score correction module configured to constrain the base text similarity based on the patient-level arterial-cuff pressure abnormality consistency index to obtain a retrieval score; a result sorting module configured to sort the patients based on the retrieval score to obtain a candidate list.

Images