Using natural language processing to detect adverse events

Abstract

To provide a method for automatically classifying a clinical description of a patient by a computer processor using natural language processing.SOLUTION: A clinical description relates to use of a ventricular assist device for treating a patient. A method includes a step of receiving at least one clinical description including a text. The method also next includes a step of determining a location of a target word in the text. The method further includes a step of determining presence of at least one negative word in an active region, where the active region includes a predetermined number of words in the text that are immediately before and after the subject word, including the subject word. The method then includes a step of determining presence of at least one body part word in the active region. Finally, the method includes a step of determining that the clinical description should be ignored if the active region includes either the negative word or the body part word.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of priority under 35 U.S.C. §119(e) from U.S. Provisional Application No. 62 / 784,192, filed December 21, 2018, the entire contents of which are incorporated herein by reference. [Background technology]

[0002] background Cardiovascular disease can reduce a patient's quality of life. Various treatment options, ranging from pharmaceuticals to mechanical devices and transplants, have been developed to treat such diseased hearts. Ventricular assist devices (VADs), such as cardiac pump systems and catheter systems, are often used in cardiac treatment to provide hemodynamic support and promote recovery. Some cardiac pump systems are inserted percutaneously into the heart and can operate in parallel with the native heart to supplement cardiac output. Such cardiac pump systems include the Impella® family of devices by Abiomed, Inc., Danvers, Massachusetts. Some of these medical procedures can lead to adverse events that occur in patients during treatment. This can be due to incorrect device use or suboptimal device configuration.

[0003] After a patient is treated using a medical device such as a VAD, a detailed description of the treatment provided to the patient, as well as any clinical indications during such treatment, is recorded. Such records are traditionally written by hand in medical abbreviations or typed into a computer. Alternatively, the treatment details are provided to a speech-to-text device (such as a Dictaphone or a laptop microphone running voice recognition software) and stored as a text file. Such files are often stored in patient data repositories and provided to other clinicians who need access to the patient's medical file.

[0004] Typically, clinicians must determine whether a patient experienced an adverse event during treatment before deciding on further treatment. For example, adverse events include bleeding, hemolysis, and ischemia, which may have occurred during the use of a VAD to treat a patient (e.g., due to the use of a new guidewire with an Impella® pump). Such determination involves manually reading and reviewing clinical records to determine whether any such adverse events occurred during treatment. Manual review involves some degree of freedom. For example, certain portions of the text in the clinical description may be overlooked or difficult to read, or interpretation of the clinical description may vary from person to person. Furthermore, to obtain clinical indications of a treatment plan (e.g., success rate), a repository of clinical descriptions must be analyzed to obtain representative indicators. Manual review of a large number of clinical descriptions is time-consuming and may involve some errors due to the aforementioned degrees of freedom.

[0005] Attempts to automate the analysis of clinical descriptions involve the use of natural language preprocessing and machine learning techniques, such as bagging and random forests, logistic regression, and regression trees. Such algorithms are complex and recursive, occupying processor resources on computing systems, especially if the algorithms do not converge easily. Machine learning using such algorithms also often requires a sizable pool of training data before the machine model can be trusted. Thus, for small data repositories, insufficient training data can result in unstable machine learning models, making their output in analyzing clinical descriptions unreliable. Summary of the Invention

[0006] overview The methods and systems described herein use natural language processing and keyword searching by a processor of a computing device to determine whether a clinical statement is relevant to a treatment, including an adverse event. The method begins by receiving at least one clinical statement including text. The processor then determines the location of a target word within the text. The processor then proceeds to determine the presence of at least one negation word within an active area (including a predetermined number of words within the text that include the target word and are located immediately before and after the target word). Next, the processor determines the presence of at least one body part word within the active area. The method then determines that the clinical statement should be ignored if the active area includes either a negation word or a body part word.

[0007] By searching for keywords within the text of the clinical descriptions, training of machine learning algorithms (which require training data) is not required, thereby freeing up system resources of the computing device. The keyword searching nature of the disclosed methods and systems does not monopolize the processor of the computing device performing the analysis of the clinical descriptions.

[0008] In some embodiments, the method further includes processing the text to generate word tokens, determining and grouping word tokens that include inflected forms of a word, and performing a keyword search on the text using the grouped word tokens. In other embodiments, the method includes flagging the clinical description if the active region does not include a negation word and a body part word. In certain embodiments, the method includes writing the flag to a header of the clinical description. In some embodiments, the predetermined number of words in the active region is at least three words. In other embodiments, the predetermined number of words in the active region is three. In certain embodiments, the at least one negation word includes any one of "no," "not," "nor," "non," "without," "never," and "false." In some embodiments, the clinical description is obtained from the Acute Myocardial Infarction Cardiogenic Shock (AMICS) repository.

[0009] In another aspect, a system for automatically classifying clinical descriptions of a patient is provided. The system includes at least one ventricular assist device (VAD) for treating the patient. The system also includes a controller in communication with the VAD and configured to generate at least one clinical description of the patient's treatment with the VAD. Further, the system includes a data repository for storing the clinical description of the treatment. The system also includes a processor in communication with the data repository, the processor configured to perform a method according to any of the aforementioned aspects. In some embodiments, the system disables use of the VAD if the number of clinical descriptions including adverse events exceeds a predetermined threshold.

[0010] In yet another aspect, a system for automatically classifying patient clinical descriptions, each clinical description associated with the use of a ventricular assist device for the patient, is provided, the system including a processor configured to perform a method according to any of the preceding aspects.

[0011] In a further aspect, there is provided a computer program comprising computer-executable instructions that, when executed by a computing device including a processor, causes the computing device to perform a method according to any of the preceding aspects. [Brief explanation of the drawings]

[0012] These and other objects and advantages will become apparent from the following detailed description considered in conjunction with the accompanying drawings, in which like reference characters refer to like parts throughout.

[0013] [Figure 1] FIG. 1 illustrates an exemplary system for discovering adverse events in clinical narratives, according to an aspect of the present disclosure. [Figure 2] 1 is an exemplary flowchart illustrating a method for natural language processing using keyword search. [Figure 3] 1 is an exemplary flowchart illustrating a method for discovering adverse events in a clinical narrative according to an aspect of the present disclosure. [Figure 4] FIG. 2 illustrates the optimization of the length of the action region surrounding the target word in relation to the number of false positives generated using the method of FIG. 1. [Figure 5] 5A and 5B illustrate the use of the method of FIG. 1 on clinical statements containing negative terms. [Figure 6] 6A and 6B illustrate the use of the method of FIG. 1 on clinical statements containing body part terms. DETAILED DESCRIPTION OF THE INVENTION

[0014] Detailed Description To provide an overall understanding of the methods and systems described herein, certain exemplary embodiments will be described. While the embodiments and features described herein are particularly described for use in connection with the use of natural language processing to automatically detect adverse events in clinical descriptions involving the use of ventricular assist devices (VADs), it will be understood that all components and other features outlined below may be combined with one another in any suitable manner and may be adapted and applied to other types of medical therapies having associated clinical descriptions.

[0015] The systems and methods described herein use natural language processing (NLP) to automatically detect the occurrence of adverse events in clinical descriptions. NLP is used to perform a search for keywords within the active region of a target word contained in the clinical description. If a keyword is found, the processor considers the clinical description to be related (or, in some cases, not related) to the treatment in which the adverse event occurred. In some aspects of the present disclosure, a flag is written to the header of the clinical description text file. By searching for keywords within the text of the clinical description, training of machine learning algorithms (which require training data) is not required, thereby freeing up system resources of the computing device. The keyword search nature of the methods and systems of the present disclosure does not monopolize the processor of the computing device performing the analysis of the clinical description.

[0016] FIG. 1 illustrates a block diagram of a system 100 for automatically detecting the occurrence of adverse events in a clinical narrative 110. The system 100 includes a computing device 120, such as a laptop, in communication with a patient data repository 130. For simplicity, only the processor 125 of the computing device 120 is shown in FIG. 1. However, it will be understood that the computing device 120 also includes other components typically associated with a computing device, such as volatile memory (such as random access memory RAM), non-volatile memory (such as read-only memory ROM), a display, and connecting buses that enable communication between these components, all of which are included in the present disclosure.

[0017] The computing device 120 includes a processor 125 capable of executing machine-readable instructions and performing operations on text data using natural language processing. The computing device 120 communicates with a patient data repository 130 containing patient data obtained from various medical institutions. According to certain aspects of the present disclosure, the patient data repository 130 may include an Acute Myocardial Infarction Cardiogenic Shock (AMICS) database compiled and maintained by a CRM such as Salesforce.com, Inc. The AMICS database 130 stores data from the treatment of high-risk percutaneous coronary intervention (PCI) patients and patients with cardiogenic shock. The AMICS database 130 may also store data specific to the VADs 140 available for treatment in a VAD database 135. The VAD database 135 may include operating parameters for each device.

[0018] The patient data includes clinical descriptions 110 stored in the AMICS database 130 after treatment of a patient in cardiogenic shock. Such treatment includes the use of medical devices to alleviate the patient's condition, such as a VAD 140. VADs provide ventricular support to patients in cardiogenic shock and may include, but are not limited to, Impella® pumps, extracorporeal membrane oxygenation (ECMO) pumps, balloon pumps, and Swan-Ganz catheters. Impella® pumps may include the Impella 2.5® pump, Impella 5.0® pump, Impella CP® pump, and Impella LD® pump, all from Abiomed, Inc. of Danvers, Massachusetts.

[0019] The VAD 140 is connected to a controller 150 that allows a physician 160 to operate the VAD 140 when treating a patient 170. Such operation may include guiding the VAD within the patient 170 and adjusting operating parameters of the VAD 140 to suit the condition of the patient 170. Operating parameters include, for example, but are not limited to, purge volume, flow rate, and pump speed. According to certain aspects of the present disclosure, the controller 150 may include an Automated Impella® Controller (AIC) by Abiomed, Inc. of Danvers, Massachusetts.

[0020] Each VAD 140 may include at least one sensor that collects data from the patient 170 while the VAD is being used to treat the patient. The patient data is transmitted as signals to the controller 150. Such data may include, but is not limited to, mean arterial pressure (MAP), left ventricular pressure (LVP), left ventricular end-diastolic pressure (LVEDP), pulmonary artery wedge pressure (PAWP), pulmonary capillary wedge pressure (PCWP), and pulmonary artery occlusion pressure (PAOP). The controller 150 communicates the patient data to the AMICS database 130, which stores the data for post-treatment analysis. The AMICS database 130 may also be provided with additional data from a physician 160 (such as notes from the patient's treatment) that may be stored along with the patient data.

[0021] Data from patients and physicians may be stored in the repository 130 as clinical descriptions 110. In certain aspects of the present disclosure, the clinical descriptions 110 may be stored in the AMICS database 130 as at least one text file with a *.txt extension. The clinical descriptions 110 may include text in any language (e.g., English) and / or abbreviations (e.g., medical abbreviations). An exemplary clinical description is shown in Table 1. The text file may also include header information including identifying data, such as the name of the patient and medical institution, patient demographics, date, time (not shown in Table 1), etc. The above are exemplary aspects of clinical descriptions, and it will be understood that the term "clinical description" encompasses any group of machine-readable characters that include information related to a medical procedure performed on a patient, such as a cardiovascular procedure using a VAD.

[0022] (Table 1) Exemplary clinical descriptions in *.txt format TIFF0007770461000001.tif66128

[0023] The clinical descriptions 110 are evaluated by the computing device 120 to classify various events occurring during each treatment. Events may include, for example, the occurrence of adverse events, malfunctions of a treatment device, and treatment success. The clinical descriptions 110 may be selected based on a set of specified criteria, such as geographic region, time period, diagnosis type, patient age, and the type of treatment device used (e.g., a VAD consisting of an Impella® CP pump). Such criteria may be predetermined or may be entered by a clinician operating the laptop 120 via, for example, a graphical user interface (GUI). Each selected clinical description 110 is then analyzed by the processor 125 of the computing device 120 using software-implemented natural language processing (NLP) algorithms. Examples of NLP software include, but are not limited to, Apache OpenNPL, Mallet, ELIZA, and cTAKES.

[0024] The natural language processing algorithm determines the presence of keywords in each selected clinical description 110. Keywords may include at least one word that acts on or in relation to a subject word in the clinical description 110. Keywords may be used to categorize a selection of clinical descriptions 110. According to one aspect of the present disclosure, subject words may be used to describe adverse events (such as bleeding, hemolysis, or ischemia) that occurred during a patient's cardiac treatment. Examples of subject words may include "bleed," "clot," and "heart," and examples of keywords may include "not," "non," and "no." Such keywords and subject words may be predetermined and stored in the memory of the computing device 120 for a particular type of analysis. Alternatively, keywords and subject words may be entered by a clinician operating the computing device 120 via the GUI of the NLP software. The presence of keywords in an active region surrounding the subject word is then analyzed by the NLP to determine whether a particular event occurred during the respective treatment. For example, the NPL may identify the occurrence of an adverse event and flag the clinical description 110 as including the adverse event 122 or not including the adverse event 124 .

[0025] FIG. 2 shows a flowchart illustrating a method 200 of natural language processing according to one embodiment of the present disclosure. The method 200 of FIG. 2 is executed by the processor 125 of FIG. 1. The method begins at step 210, where the processor 125 of the computing device 120 retrieves a selection of clinical descriptions 110 from the AMICS database 130. As previously mentioned, the selection of clinical descriptions 110 may be based on a set of specified criteria, such as geographic region, time period, diagnosis type, patient age, and type of treatment device used (e.g., VAD). At step 220, an NPL algorithm divides the text of each clinical description 110 into fragments or tokens (this is called tokenizing the text). Depending on the NPL algorithm used, certain characters in the text, such as punctuation characters, may be ignored. Each token is used as a semantic unit for processing the text associated with the selected clinical description.

[0026] After tokenizing the text, method 200 then proceeds to step 230, also known as lemmatization, in which similar tokens are grouped together based on the inflected forms of the tokens so that they can be parsed as a single item. Essentially, lemmatization (or stemming) links tokens that have the same base form (root) and groups them so that they can be processed in a similar manner. For example, in English, the verb "walk" can appear as "walk," "walked," "walks," or "walking." The base form here is "walk," which can be looked up in a dictionary. The output of lemmatization step 230 is a bag-of-words (BOW) containing groups of tokens, each with an associated base form.

[0027] Once the BOW for the selected clinical description 110 is formed, a keyword search can be performed (step 240). The NPL lemmatizes the target word and identifies the root of the target word. The NPL then scans the BOW to determine whether the root of the target word is present in the BOW. If such determination is positive, i.e., if the BOW contains a root that matches the root of the target word, the selected clinical description 110 is deemed to contain the target word. Conversely, if the determination is negative, i.e., if the BOW does not contain the root of the target word, the selected clinical description 110 is deemed to not contain the target word.

[0028] FIG. 3 illustrates a flow diagram of a method 300 for automatically classifying clinical descriptions 110 according to one aspect of the present disclosure. The method 300 of FIG. 3 is executed by the processor 125 of FIG. 1. Similar to method 200, method 300 begins at step 310, where the processor 125 of the computing device 120 retrieves a selection of clinical descriptions 110 from the AMICS database 130. As mentioned above, the selection of the clinical descriptions 110 may be based on a set of specified criteria, such as, for example, geographic region, time period, diagnosis type, patient age, and type of treatment device used (e.g., VAD). The criteria may be specified by a clinician via a GUI of the computing device 120.

[0029] In step 320, the processor 125 determines the location of the target word within each selected clinical description 110 using an NPL algorithm running on the processor 125. Once the location of the target word is identified, the method 300 further uses the NPL process 200 to identify an active area associated with the target word. The active area includes a predetermined number of words within the text of the selected clinical description 110 that immediately precede and follow the target word. The active area also includes the target word. The predetermined number of words may be stored within the computing device 120 or may be provided as input from the clinician via the GUI. The predetermined number of words defines the size of the active area (i.e., the granularity of the method 300), hereinafter referred to as the granularity size.

[0030] The method 300 then continues analyzing the active regions within each selected clinical description 110. Here, the processor 125 uses the NPL method 200 to search each active region of the selected clinical description 110 for keywords. As described above, keywords operate on or in conjunction with subject words within each clinical description 110. According to aspects of the present disclosure, keywords may include negation words or body part words. Negation words may include, but are not limited to, "no," "not," "nor," "non," "without," "never," and "false." The presence of a negation word within an active region of a subject word reverses or negates the normal meaning of the subject word. For example, if the clinical description reads, "... groin site is dry with no evidence of bleeding at all...," the presence of the negation word "no" negates the meaning of the subject word "bleeding" occurring in the groin. Thus, upon detecting the negative word "no" within the active region, processor 125 flags this clinical statement as not being related to an adverse event (which is bleeding in the groin).

[0031] Similarly, body part words may include any body part, such as, for example, "leg," "arm," "abdomen," and "groin." The presence of a body part word in the active region of a target word negates the normal meaning of the target word. Unlike negation words, according to one aspect of the present disclosure, the presence of a body part word indicates that the adverse event (e.g., bleeding) does not occur within the heart. For example, if a clinical statement reads, "...patient is very sick and they feel she is bleeding into her abdomen...," the presence of the body part word "abdomen" negates the meaning of the target word "bleeding" because it is not related to the heart. According to one aspect of the present disclosure, any adverse event in a clinical statement that does not have a body part word acting on the target word is assumed to occur in the patient's heart. Thus, upon detecting the body part word "abdomen" within the active region, processor 125 flags the clinical statement as not related to a bleeding adverse event in the heart. It will be appreciated that the NPL keyword search can be further customized to match words in the active area with specific body part words, for example "heart," and flag clinical descriptions based on the match.

[0032] 3, once the location of the target word within the text of the clinical description 110 has been determined, the method 300 proceeds to step 330, where it is further determined whether a negation word is present in the active area. If a negation word is present in the active area (“Yes” at step 330), the clinical description is flagged as not containing an adverse event. In certain aspects of the present disclosure, a flag may be written (e.g., in ASCII characters) to the header of a text file for each clinical description, and the flagged clinical description 110 may be written back to the AMICS database by the computing device 120.

[0033] If a negation term is not present in the active region (“No” at step 330), method 300 proceeds to step 340, where it is further determined whether a body part term is present in the active region. If a body part term is present (“Yes” at step 340), the clinical statement is flagged as not including an adverse event, as at step 350. If a body part term is not present (“No” at step 340), the clinical statement is flagged as including an adverse event, as at step 360. In certain aspects of the present disclosure, flags for both the “Yes” and “No” answers at step 340 may be stored in the header of a text file for each clinical statement, and the flagged clinical statement 110 may be written back to the AMICS database by computing device 120.

[0034] The strategic keyword search of the present disclosure scans the text within each clinical description 110 for target words and any specified keywords within the target word's active area. Once a keyword is identified, the clinical description is flagged and the analysis continues with the next text file. This is less processor intensive than NPL, which uses machine learning, thus freeing up processing power on the computing device.

[0035] FIG. 4 illustrates an optimization chart 400 of the NPL algorithm based on the size of the active region for detecting negation terms related to a target term, according to one embodiment of the present disclosure. In effect, this optimizes the granularity of the method 300. As previously mentioned, the number of false positives (FP) refers to the number of times the processor 125 fails to detect the presence of a negation term or body part word in the active region when the negation term or body part word is actually present, and a true positive (TP) means that the negation term or body part word is present and detected. In effect, an FP is a missed alarm, for example, the processor 125 misses detecting a bleeding event in the patient's heart according to the clinical description 110 by incorrectly identifying a negation term or body part word in the active region of the clinical description. Similarly, the number of false negatives (FN) refers to the number of times the processor 125 incorrectly detects the presence of a negation term or body part word in the active region when the negation term or body part word is not actually present, and a true negative (TN) means that the negation term or body part word is neither present nor detected.

[0036] In FIG. 4, the size of the active region is referenced in terms of the number of words immediately preceding and following the target word. Optimization is performed in terms of the number of false positives. Line graph 410 of optimization chart 400 shows that the occurrence of false positives shows the least variation for active regions with a granularity of three words or greater. Furthermore, Table 2 shows exemplary values ​​of TP, FP, FN, and TN as the granularity of the active region changes. The values ​​in Table 2 reinforce the trend shown in FIG. 4, in which the number of missed alarms, i.e., FPs, decreases as the size of the active region increases. The rate of decrease in FPs levels off after a critical granularity of the active region. According to an embodiment of the present specification, the critical size of the active region is considered to be three words.

[0037] (Table 2) FP optimization by active area size TIFF0007770461000002.tif55128

[0038] 5A and 5B show examples of automatically detecting the presence of negated words in clinical descriptions 500, 550 using an NPL algorithm according to the systems and methods described above. In FIG. 5A, the target word 510 is "bleeding" and the active region 520 has a granularity of 3. The keyword 530 is the negated word "no." When the negated word 530 is present in the active region 520 of the target word 510, the clinical description 500 is flagged as not containing an adverse event. Similarly, in FIG. 5B, the target word 560 is "bleeding" and the active region 570 has a granularity of 3. The keyword 580 is the negated word "no." When the negated word 580 is present in the active region 570 of the target word 560, the clinical description 550 is flagged as not containing an adverse event. In both of the above examples, NPL processing according to method 200 shown in FIG. 2 is used to identify the target word, the negated word, and the words in the active region.

[0039] 6A and 6B illustrate an example of automatically detecting the presence of body part words in clinical descriptions 600, 650 using the NPL algorithm according to the systems and methods described above. In FIG. 6A, the target word 610 is "bleeding" and the active region 620 has a granularity of 3. The keyword 630 is the body part word "abdomen." When the body part word 630 is present within the active region 620 of the target word 610, the clinical description 600 is flagged as not containing an adverse event. As previously mentioned, adverse events in clinical descriptions that do not have a body part word acting on the target word are assumed to occur in the patient's heart. Thus, upon detecting the body part word "abdomen" within the active region 620, the processor 125 flags the clinical description 600 as not relating to a cardiac bleeding adverse event. Similarly, in FIG. 6B, the target word 660 is "bleeding" and the active region 670 has a granularity of 3. The keyword 680 is the body part word "groin." When the body part word 680 is present within the active region 670 of the target word 660, the clinical description 650 is flagged as not containing an adverse event. In both of the above examples, NPL processing according to method 200 shown in Figure 2 is used to identify the target word, the body part word, and the words within the active region.

[0040] Table 3 shows a confusion matrix that provides an example diagram to illustrate the effectiveness of the methods and systems of the present disclosure. Table 3 compares the results of a full keyword search using NPL on all words in a clinical description to a strategic negation and / or body part keyword search using NPL that searches only negation and / or body part words in the active region of the target word, according to an embodiment of the present disclosure. Metrics associated with the confusion matrix include precision, recall, and accuracy. Precision is determined using the formula: TP / (TP+FP), recall is determined using the formula: TP / (TP+FN), and accuracy is determined using the formula: (TP+TN) / (TP+FP+FN+TN). In the exemplary confusion matrix of Table 3, the precision, recall, and accuracy metrics for full keyword search are 53.4%, 97.5%, and 91.7%, respectively, while the same metrics for strategic negation and / or body part keyword search are 74.0%, 92.5%, and 96.2%, respectively. These metric comparisons indicate that strategic negation and / or body part keyword search identifies keywords with higher precision and accuracy. Such figures demonstrate that the strategic negation and / or body part keyword search of the present disclosure outperforms conventional NPL techniques while also minimizing the use of system resources for performing such natural language processing.

[0041] As described above, after each selected clinical description 110 is flagged, the flag can be stored in the header of each clinical description text file, and the text file can be written back to the AMICS database 130 by the computing device 120. To this end, a clinician can filter flagged clinical descriptions from the AMICS database based on specific criteria (e.g., VAD type, patient age, name of medical institution, etc.) to obtain the percentage of clinical descriptions containing the target term. For example, a clinician operating the computing device 120 can obtain data from the AMICS database related to an Impella® 2.5 heart pump used on a male patient aged 50-55 in Boston, Massachusetts, who suffered intracardiac bleeding during a cardiac procedure. If such data is lower than the statistically represented data for patients matching the criteria, it may indicate various problems. Such problems may include, for example, that the cardiac procedure was performed incorrectly on the patient or that the VAD used has a defect that requires adjustment. If the cardiac procedure was performed incorrectly, such data may trigger further training at the medical institution. If a VAD defect is suspected, such data can be used as quality control during the manufacture of such devices. Additionally, data obtained from flagged clinical statements containing adverse events may be used to trigger a lockout mechanism to warn physicians against VAD use if the number of flagged clinical statements exceeds a predetermined threshold. Such automatic flagging of clinical statements, as described above, may provide clinicians with feedback that can improve the care provided to patients.

[0042] (Table 3) Comparison of confusion matrices TIFF0007770461000003.tif72128

[0043] The foregoing is merely illustrative of the principles of the present disclosure, and the apparatus may be practiced in other than the described embodiments, which are presented for purposes of illustration and not limitation. It should be understood that the methods disclosed herein, while shown for use in an automated ventricular assist system, may also be applied to systems to be used in other automated medical systems.

[0044] Changes and modifications will occur to those skilled in the art after considering this disclosure. The features of the disclosure may be implemented in any combination and subcombination (including multiple subcombinations and subcombinations) with one or more other features described herein. The various features described or illustrated above, including any components thereof, may also be combined or integrated in other systems. Furthermore, certain features may be omitted or not implemented.

[0045] Examples of changes, substitutions, and alterations would be ascertainable by one skilled in the art and could be made without departing from the scope of the information disclosed herein. All references cited herein are incorporated by reference in their entirety and made a part of this application.

Claims

1. 1. A method for automatically classifying a patient's clinical descriptions, comprising: receiving, using one or more processors, at least one clinical descriptor file from a data repository including a plurality of clinical descriptor files, each clinical descriptor file including text related to use of a medical device for a corresponding patient; using the one or more processors to determine a location of a subject term within the text of the at least one clinical description file; using the one or more processors to determine whether at least one of a plurality of predetermined body part words is located within an active area, the active area including the target word and a predetermined number of words in the text immediately preceding and following the target word; using the one or more processors to write a flag in a header of the clinical descriptor file in response to determining that the active region does not include at least one of the plurality of predetermined body part words, the flag indicating that the at least one clinical descriptor file includes an adverse event; A method comprising:

2. determining the location of the subject word; processing the text of the at least one clinical description file to generate word tokens; Determining and grouping word tokens containing inflected forms of a word; performing a keyword search on the text using the grouped word tokens; The method of claim 1 , comprising:

3. The method of claim 1 , wherein the predetermined number of words in the active area is at least three words.

4. using the one or more processors to determine whether at least one of a plurality of predetermined negation words is located within the active region; 2. The method of claim 1, wherein the flag is written to the header of the clinical description file in response to (a) a determination that the active region does not include at least one of the plurality of predetermined negative words, and (b) a determination that the active region does not include at least one of the plurality of predetermined body part words.

5. 5. The method of claim 4, wherein the plurality of predetermined negation words include "no," "not," "nor," "non," "without," "never," and "false."

6. 10. The method of claim 1, wherein the data repository is the Acute Myocardial Infarction Cardiogenic Shock (AMICS) repository.

7. Disabling use of the medical device using the one or more processors in response to determining that a number of clinical description files containing adverse events exceeds a predetermined threshold. The method of claim 1 further comprising:

8. Initiating further training at the medical institution using the one or more processors in response to determining that the number of clinical description files containing adverse events exceeds a predetermined threshold. The method of claim 1 further comprising:

9. Initiating a manufacturing quality control process using the one or more processors in response to determining that the number of clinical description files containing adverse events exceeds a predetermined threshold. The method of claim 1 further comprising:

10. 10. The method of claim 1, wherein each medical device is a ventricular assist device (VAD).

11. 1. A system for automatically classifying clinical descriptions of a patient, comprising: receiving at least one clinical description file from a data repository, the clinical description file including text related to use of a medical device on a patient; determining a location of a subject term within the text of the at least one clinical description file; determining whether at least one of a plurality of predetermined body part words is located within an active area, the active area including the target word and a predetermined number of words in the text immediately before and after the target word; writing a flag in a header of the clinical descriptor file in response to determining that the active region does not include at least one of the plurality of predetermined body part words, the flag indicating that the at least one clinical descriptor file includes an adverse event; one or more processors configured to perform Including, the system.

12. further comprising said data repository; The system of claim 11 , wherein the data repository includes a plurality of clinical description files, each clinical description file including text related to the use of a medical device for a corresponding patient.

13. 12. The system of claim 11, wherein the one or more processors are further configured to disable the medical device in response to determining that a number of clinical description files containing adverse events exceeds a predetermined threshold.

14. The system of claim 13 further comprising the medical device.

15. 12. The system of claim 11, wherein the one or more processors are further configured to initiate further training at a medical institution in response to determining that a number of clinical description files containing adverse events exceeds a predetermined threshold.

16. 12. The system of claim 11, wherein the one or more processors are further configured to initiate a manufacturing quality control process in response to determining that a number of clinical description files containing an adverse event exceeds a predetermined threshold.

17. The system of claim 11 , wherein the medical device is a ventricular assist device (VAD).

18. A non-transitory computer-readable storage medium having stored thereon instructions that, when executed by one or more processors, receiving at least one clinical description file from a data repository, the clinical description file including text related to use of a medical device on a patient; determining a location of a subject term within the text of the at least one clinical description file; determining whether at least one of a plurality of predetermined body part words is located within an active area, the active area including the target word and a predetermined number of words in the text immediately before and after the target word; writing a flag in a header of the clinical descriptor file in response to determining that the active region does not include at least one of the plurality of predetermined body part words, the flag indicating that the at least one clinical descriptor file includes an adverse event; a non-transitory computer-readable storage medium that causes the one or more processors to perform

Images