System, method and storage medium for extracting target medical information from clinical remarks

Abstract

A system extracts target medical information from clinical remarks stored in memory. The system includes a pre-processing module configured to retrieve a sequence of clinical text of an electronic case from the memory and lexical-element the sequence of clinical text to obtain a sequence of input lexical elements. The system also includes a sequence-to-structure model module configured to transform the sequence of input lemmas into a sequence of structured output lemmas using a trained natural language-based transformer. The system also includes a post-processing unit configured to obtain annotated text-tag pairs of the clinical text in accordance with the structural output lexical elements.

Description

[0001] Cross - reference to related applications

[0002] Claims the priority of U.S. Provisional Application No. 63 / 399,237, filed on August 19, 2022, the entire disclosure of which is incorporated herein by reference. Background Art

[0003] Social determinants of health (SDOH) are the living conditions that affect people's quality of life and health outcomes. SDOH encompasses a wide range of conditions, such as substance use, living situation, employment, education, racial attitudes, geography, pollution, and so on. Understanding SDOH, including the behaviors influenced by these social factors, can inform clinical decision - making. However, most detailed SDOH are characterized in unstructured clinical text in electronic medical records. This text - encoded information must be automatically extracted for secondary use applications such as large - scale retrospective studies and clinical decision - support systems.

[0004] Currently, most event extraction methods adopt a decomposition strategy, that is, the prediction of complex event structures is decomposed into multiple independent sub - tasks (mainly including entity recognition, trigger detection, and argument classification), and then the components of different sub - tasks are combined to predict the entire event structure (e.g., pipeline modeling, joint modeling, or joint inference). A major drawback of these decomposition - based methods is that they require a large number of fine - grained annotations for different sub - tasks, which usually leads to the problem of low data efficiency. For example, they require different fine - grained annotations for trigger detection, type classification, status classification, etc. And usually, separate systems are implemented for each of these sub - tasks to extract the corresponding annotations, which results in a very complex pipeline system.

[0005] Another drawback of decomposition - based methods is that it is very challenging to manually design the optimal combination architecture for different sub - tasks. For example, pipeline models often cause error propagation. Additionally, joint models require heuristically predefined information sharing and decision dependencies between trigger detection, argument classification, and entity recognition, which usually leads to sub - optimal and inflexible architectures. Summary of the Invention

[0006] According to one aspect of the present inventive concept, there is provided a computer-implemented method for extracting target medical information from clinical notes stored in a memory. The method includes: retrieving a sequence of clinical texts of an electronic case from the memory, and tokenizing the sequence of clinical texts to obtain a sequence of input tokens. The method further includes transforming the sequence of input tokens into a sequence of structured output tokens using a trained natural language-based transducer. The method further includes post-processing the structured output tokens to obtain an annotated text-label pair of the clinical text.

[0007] The natural language-based transducer may be a T5 transducer. The T5 transducer may include an encoder and a decoder. The encoder receives a sequence as input and generates a sequence of representations. The decoder receives the sequence of representations and previously generated tokens as input to generate an output token at each time step.

[0008] The post-processing may further include converting the text-label pair into a table format.

[0009] The target medical information may be social determinants of health (SDOH) information.

[0010] According to another aspect of the present inventive concept, there is provided a system for extracting target medical information from clinical notes stored in a memory. The system includes: a preprocessing module, a sequence-to-structure model module, and a post-processing module. The preprocessing module is configured to retrieve a sequence of clinical texts of an electronic case from the memory, and tokenize the sequence of clinical texts to obtain a sequence of input tokens. The sequence-to-structure model module is configured to transform the sequence of input tokens into a sequence of structured output tokens using a trained natural language-based transducer. The post-processing module is configured to obtain an annotated text-label pair of the clinical text based on the structured output tokens.

[0011] The natural language-based transducer of the sequence-to-structure model module may be a T5 transducer. The T5 transducer may include an encoder and a decoder. The encoder receives a sequence as input and generates a sequence of representations. The decoder receives the sequence of representations and previously generated tokens as input to generate an output token at each time step.

[0012] The post-processing module may further be configured to convert the text-label pair into a table format.

[0013] The target medical information may be social determinants of health (SDOH) information.

[0014] According to another aspect of the inventive concept, a non-transitory computer-readable computer medium is encoded with instructions that, when executed, extract target medical information from clinical notes stored in a memory. The medium includes a preprocessing module that, when executed, retrieves a sequence of clinical text of an electronic medical record from the memory and tokenizes the sequence of clinical text to obtain a sequence of input tokens. The medium further includes a sequence-to-structure model module that, when executed, transforms the sequence of input tokens into a sequence of structured output tokens using a trained natural language-based transformer. The medium further includes a post-processing module that, when executed, obtains an annotated text-label pair of the clinical text based on the structured output tokens.

[0015] The natural language-based transformer of the sequence-to-structure model module may be a T5 transformer. The T5 transformer may include an encoder and a decoder. The encoder receives a sequence as input and generates a sequence of representations. The decoder receives the sequence of representations and a previously generated token as input to generate an output token at each time step.

[0016] The post-processing module, when executed, may convert the text-label pair into a table format.

[0017] The target medical information may be social determinants of health (SDOH) information. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] With reference to the accompanying drawings, the above and other aspects and features of the inventive concept will become readily apparent from the following detailed description, in which:

[0019] Figure 1 An example of an SDOH annotation in BRAT format is illustrated;

[0020] Figure 2 is a block diagram illustrating the architecture of a sequence-to-structure model according to one or more embodiments of the inventive concept;

[0021] Figure 3 is a flowchart for reference in describing a sequence-to-structure model according to one or more embodiments of the inventive concept;

[0022] Figure 4 employs Figure 1 as an example to provide an illustrative example of generating an ADE annotation for an input text according to one or more embodiments of the inventive concept;

[0023] Figure 5is a diagram (in a tree structure and a linearized format) illustrating event-based annotations for input text according to one or more embodiments of the inventive concept; and

[0024] Figure 6 is a simplified block diagram of a system for automatically extracting target medical information from clinical notes stored in a memory according to a representative embodiment. Detailed Description

[0025] In the following detailed description, for purposes of explanation and not limitation, representative embodiments that disclose specific details are set forth in order to provide a thorough understanding of embodiments according to the present teachings. Descriptions of known systems, devices, materials, methods of operation, and methods of manufacture may be omitted so as not to obscure the description of the representative embodiments. Nevertheless, systems, devices, materials, and methods within the knowledge of those of ordinary skill in the art are also within the scope of the present teachings and may be used in accordance with the representative embodiments. It should be understood that the terms used herein are merely for the purpose of describing particular embodiments and are not intended to be limiting. These defined terms supplement the scientific and technical meanings of the defined terms that are commonly understood and accepted in the technical field of the present teachings.

[0026] It should be understood that although the terms first, second, third, etc. may be used herein to describe various elements or components, these elements or components should not be limited by these terms. These terms are only used to distinguish one element or component from another. Thus, a first element or component discussed below may be referred to as a second element or component without departing from the teachings of the inventive concept.

[0027] The terms used herein are merely for the purpose of describing particular embodiments and are not intended to be limiting. As used in the specification and the appended claims, the singular forms of the terms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. Additionally, the terms "comprises," "comprising," and / or similar terms specify the presence of the stated features, elements, and / or components, but do not preclude the presence or addition of one or more other features, elements, components, and / or combinations thereof. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0028] Unless otherwise specified, when an element or component is referred to as "connected to", "coupled to", or "adjacent to" another element or component, it should be understood that the element or component can be directly connected or coupled to the other element or component, or there can be intermediate elements or components. That is, these terms and similar terms include situations where one or more intermediate elements or components can be used to connect two elements or components. However, when an element or component is referred to as "directly connected" to another element or component, this only includes the situation where the two elements or components are connected to each other without any intermediate or intervening elements or components.

[0029] Accordingly, the present disclosure is intended to bring one or more of the specifically pointed out advantages by one or more of its aspects, embodiments, and / or specific features or sub-components. For purposes of explanation and not limitation, example embodiments that disclose specific details are set forth in order to provide a thorough understanding of the embodiments according to the present teachings. However, other embodiments that are different from the specific details disclosed herein and that are consistent with the present disclosure are still within the scope of the appended claims. Additionally, descriptions of well-known devices and methods may be omitted so as not to obscure the description of the example embodiments. Such methods and devices are also within the scope of the present disclosure.

[0030] An electronic health record (EHR) generally refers to the digital version of a patient's paper medical record. An EHR is a real-time, patient-centered record that enables authorized users to obtain information instantaneously and securely. While an EHR does contain a patient's medical history and treatment history, an EHR system is built beyond the standard clinical data collected in a provider's office, and an EHR can contain a broader range of patient care observations. An EHR is an important part of health information technology, and among other benefits, it allows access to evidence-based tools that providers can use to make decisions about patient care. A key feature of an EHR is that health information can be created and managed in digital format by authorized providers and can be shared with other providers in more than one healthcare organization. An EHR is built to share information with other healthcare providers and organizations (e.g., laboratories, specialists, medical imaging facilities, pharmacies, emergency care providers, as well as school and workplace clinics).

[0031] At the same time, as previously mentioned, social determinants of health (SDOH) are the living conditions that affect people's quality of life and health outcomes. Such SDOH include a wide range of conditions, such as substance use, living situation, employment, education, racial attitudes, geography, pollution, and so on. SDOH can lead to a shortened life expectancy. For example, substance abuse (including alcohol use, medication use, and smoking) is increasingly recognized as a key factor in morbidity and mortality; more and more people are living alone, which leads to increased social isolation and negative health outcomes; employment and occupation affect income, social status, exposure to risks, and health. Understanding SDOH (including the behaviors influenced by these social factors) can inform clinical decision-making. SDOH are represented in the EHR through structured data and unstructured clinical text; however, clinical text captures detailed descriptions of these determinants, which go beyond the representation of structured data. This text-encoded information must be automatically extracted to be used in secondary use applications (such as large-scale retrospective studies and clinical decision support systems). The automatically extracted data can augment the available structured data to create a more comprehensive patient representation in these downstream applications.

[0032] At least some aspects of the inventive concept relate to extracting SDOH information from, for example, the historical section of clinical notes included in an EHR. Figure 1 An example of an SDOH annotation in BRAT format is illustrated, which was highlighted in the shared task of the 2022 National NLP Clinical Challenge (n2c2). The corpus used in the shared task contains annotated events for social determination events (SDEs), where each social determination event includes a trigger that anchors the event and one or more arguments that characterize the event. The arguments capture status, type, extent, and temporal information. Figure 1 An example of an SDE annotation in BRAT format is shown. BRAT is a tool for text annotation (i.e., for adding annotations to existing text documents). It is designed specifically for "structured" annotation, where the annotations are not free-form text but have a fixed form that can be automatically processed and interpreted by a computer.

[0033] In Figure 1 the example of, three lines or entries of clinical text are shown, namely,

[0034] "Social history: was once a cook; currently unemployed",

[0035] "Smoking: quit smoking 7 years ago; 15 - 20 packs / year", and

[0036] "Drinking: none; Medication use: none"

[0037] In Figure 1Several categories of BART annotations are also shown in the example. "Text span" annotations are those boxes marked, for example, with "tobacco", "alcohol", "drugs", and so on. Figure 1 Another category illustrated in is the "relationship" annotation. For example, the "status" and "quantity" relationships in the example. BRAT also supports annotations of "n-ary associations", which can link any number of other annotations participating in specific roles together. This category of annotation can be used, for example, for event annotation.

[0038] For the purpose of description, Figure 1 the example of will run through the remainder of the following detailed description.

[0039] The present inventive concept provides a mechanism for automatically extracting SDOH from clinical texts. Specifically, a sequence-to-structure generation model is utilized to directly extract all SDOH in an end-to-end manner. The model is based on a Transformer encoder-decoder architecture, where, given a sequence of input tokens, the encoder encodes the input into a sequence of token representations, and the decoder uses these representations and a greedy decoding algorithm to predict the output token by token.

[0040] Even though the sequence-to-structure generation model of the embodiment is designed for extracting SDOH, it can be directly applied to other different information extraction tasks involving identifying triggers and argument spans, normalizing arguments, and predicting the relationships between triggers and argument spans. In fact, such information extraction tasks are prevalent in almost all enterprises that generate or rely on large amounts of text data. For example, removing patient information identifiers from electronic medical records, extracting key issues from complaint data, standardizing radiology procedure descriptions, and so on.

[0041] The embodiments herein improve the SDOH extraction task, which provides a solution for more comprehensive patient representation and can potentially improve patient safety. Such SDOH information is also beneficial to many downstream applications, such as large-scale retrospective studies, cohort selection, clinical decision support systems, and so on. Additionally, the embodiments herein automatically extract structural information from large amounts of text, which provides necessary support for natural language understanding by identifying and parsing concepts, entities, events described in the text and inferring the relationships between them. This automated process can save time and money and improve productivity.

[0042] The inventive concept is directed to a sequence-to-structure generation paradigm for event extraction, which can directly extract events from text in an end-to-end manner. Specifically, instead of decomposing event structure prediction into different subtasks and predicting labels, embodiments herein model the entire event extraction process uniformly in a neural network-based sequence-to-structure architecture, and all triggers, arguments, and their labels are uniformly generated as natural language words. For example, for trigger extraction, a subsequence "(tobacco smoking)" is generated, where the event type "tobacco" and the event trigger "smoking" are both regarded as natural language words. Compared with previous methods, embodiments herein are more data-efficient. That is, embodiments herein can be learned using only rough parallel text record annotations (i.e., sentence pairs, event records) rather than fine-grained token-level annotations. Additionally, the unified architecture facilitates modeling, learning, and exploiting the interactions between different underlying predictions, and knowledge can be seamlessly shared and transferred between different components.

[0043] Figure 2 is a block diagram illustrating the architecture of a sequence-to-structure model according to one or more embodiments of the inventive concept. Figure 3 is a flowchart for reference when describing a sequence-to-structure model according to one or more embodiments of the inventive concept. Figure 4 employs Figure 1 as an example to provide an illustrative example of generating ADE (Adobe Digital Edition) annotations for an input text according to one or more embodiments of the inventive concept. Figure 5 is a diagram (in tree structure and linearized format) illustrating event-based annotations for an input text according to one or more embodiments of the inventive concept.

[0044] With joint reference to Figures 2 - 5 , a pre-trained natural language model, the T5 Transformer 1000, is employed as a Transformer-based encoder-decoder architecture. T5 is an example of a Transformer-based encoder-decoder model developed by Google for text generation. As shown, the T5 Transformer includes an encoder 10 and a decoder 20. The encoder H = Encoder(X) takes a sequence of tokens X = {x0, x1, …, x n} as input and generates a sequence of representations H = {h0, h1, …, h n}, and the decoder y t = Decoder(y t-1 , H) takes the sequence of representations H and the token y t-1 generated at time step t - 1 as input to generate a token at each time step, where y0 = " <bos>".

[0045] At preprocessing step S101, clinical text is retrieved and preprocessed into a sequence of input tokens. As described above, clinical text can be retrieved from the EHRs of one or more patients. Generally, a tokenizer (not shown) converts the incoming text into a digital data structure suitable for machine learning. In the given example, the clinical text is a sequence of words / punctuations: "Smoking: Quit smoking several years ago and 15 - 20 packs / year".

[0046] At step S102, the sequence of input tokens is applied as input to the pre - trained T5 transformer 100. As described above, the encoder X generates a sequence of representations based on the input tokens and generates output tokens based on the representations and the previous output tokens. The result is a sequence of structured output tokens. This constitutes the sequence - to - structure model of the embodiment.

[0047] In post - processing step S103, the generated output (as a sequence of structured output tokens) is converted into text, in which the text within parentheses is a label - text pair.

[0048] In post - processing step S104, the label - text pairs are tabulated and output in tabular format, as Figure 4 shown in the table.

[0049] The main components of the embodiment are the sequence - to - structure model ( Figure 2 and step 102). The trained sequence - to - structure model is capable of generating the SDE of a given text based on the social history section of the clinical note. Such a sequence - to - structure model can be any sequence - to - sequence model, e.g., an RNN - based or a transformer - based model. The inventive concept is not limited to the model itself but includes a generation - based method for extracting SDEs by generating event triggers and their arguments for a given text. To train the model, the inventive concept uses an existing transformer - based sequence - to - sequence (seq2seq) architecture, initializes the seq2seq model using a T5 checkpoint, and further trains it on the collected input - output pairs. Event - based SDOH annotations are converted into a linearized format suitable for the sequence - to - structure model (as Figure 5 shown). Specifically, the inventive concept first parses the SDE annotation ( Figure 1 ) into an event tree and then uses a depth - first traversal of natural language vocabulary to linearize the event tree, where "(” and ")” are structure indicators for representing the semantic structure of the linear expression. Each part of this structure captures the event type, attributes, and the corresponding text span and argument type.

[0050] During training, the open-source tool SpaCy can be used to split clinical notes into sentences. For each sentence, the corresponding SDE annotations are extracted and converted into a linearized format. Then, the model can be trained on the input sentence and its linearized SDE annotations. During inference, for each clinical note, the predictions for each sentence can be concatenated, the offsets of all generated text spans can be identified, and the output can be converted into a table format.

[0051] Figure 6 is a simplified block diagram of a system for automatically extracting target medical information from clinical notes stored in a memory according to a representative embodiment.

[0052] Refer to Figure 6 , system 100 includes a processing unit 110 and a memory 120 for storing instructions executable by the processing unit 110 to implement the processes described herein. Additionally, system 100 includes a user interface 130 for connecting to a user interface, a network interface 140 for connecting to other components and instruments, and a display 150 that may include a graphical user interface (GUI) 155. System 100 also includes or is otherwise connected to a primary data source 160 and an optional secondary data source 170.

[0053] The processing unit 110 represents one or more processing devices and is configured to execute software instructions to perform the functions described in various embodiments herein. The processing unit 110 can be implemented by one or more servers, general-purpose computers, central processing units, processors, microprocessors or microcontrollers, state machines, programmable logic devices, FPGAs, ASICs, or combinations thereof using any combination of hardware, software, firmware, hardwired logic circuits, or combinations thereof. As such, the term "processing unit" encompasses electronic components capable of executing programs or machine-executable instructions and can be construed to include more than one processor or processing core (such as in a multi-core processor and / or parallel processor). The processing unit 110 can also include a collection of processors within a single computer system or distributed across multiple computer systems (e.g., in a cloud-based or other multi-site application). The program has software instructions executed by one or more processors, which can be within the same computing device or distributed across multiple computing devices.

[0054] The processing unit 110 may include an AI engine or module (e.g., the T5 transformer as described above), which may be implemented as software to provide artificial intelligence (e.g., natural language processing (NLP) algorithms), and may apply machine learning (e.g., artificial neural network (ANN), convolutional neural network (CNN), or recurrent neural network (RNN) modeling). The AI engine may reside in any one of various components other than the processing unit 110 (e.g., the memory 120, an external server, and / or the cloud). When the AI engine is implemented in the cloud (e.g., at a data center), the AI engine may be connected to the processing unit 110 via the Internet using one or more wired and / or wireless connections (e.g., via the network interface 140).

[0055] The memory 120 may include a main memory and / or a static memory, where these memories may communicate with each other and with the processing unit 110 via one or more buses. The memory 120 stores instructions for implementing some or all aspects of the methods and processes described herein (including, for example, the functions and methods described above with reference to Figures 2 - 5 the description). The memory 120 may include software modules. In an embodiment of the inventive concept, the memory 120 includes a preprocessing module 120a for performing the above-described preprocessing tasks, a sequence-to-structure model module 120b for performing the above-described sequence-to-structure model, and a postprocessing module 120c for performing the above-described postprocessing tasks.

[0056] For example, the memory 120 may be implemented by any number, type, and combination of random access memory (RAM) and read-only memory (ROM), and may store various types of information (e.g., software algorithms, data-based models (including ANN, CNN, RNN, and other neural network-based models), and computer programs, all of which can be executed by the processing unit 110). Various types of ROM and RAM may include any number, type, and combination of computer-readable storage media (e.g., disk drives, flash memories, electrically programmable read-only memories (EPROMs), electrically erasable programmable read-only memories (EEPROMs), registers, hard disks, removable disks, magnetic tapes, compact disc read-only memories (CD-ROMs), digital versatile discs (DVDs), floppy disks, Blu-ray discs, universal serial bus (USB) drives, or any other form of computer-readable storage media known in the art).

[0057] The memory 120 is a tangible storage medium for storing data and executable software instructions and is non-transitory during the time the software instructions are stored therein. As used herein, the term "non-transitory" should not be construed as an eternal state property but rather a state property that persists for a period of time. The term "non-transitory" specifically disavows transient properties (e.g., carrier waves or signals or other forms of properties that exist only temporarily at any time and any place). A non-transitory storage medium is defined as any medium that constitutes patentable subject matter under 35 U.S.C. § 101 and excludes any medium that does not constitute patentable subject matter under 35 U.S.C. § 101. The memory 120 may store software instructions and / or computer-readable code that enable the performance of various functions. The memory 120 may be secure and / or encrypted, or insecure and / or unencrypted.

[0058] The user interface 130 provides information and data output by the processing unit 110 to the user and / or receives information and data input by the user. That is, the user interface 130 enables the user to enter data and control or manipulate various aspects of the processes described herein and also enables the processing unit 110 to indicate the effects of the user's control or manipulation. All or part of the user interface 130 may be implemented by the GUI 155 visible on the display 150. The user interface 130 may include, for example, a mouse, keyboard, trackball, joystick, haptic device, touchpad, touchscreen, and / or voice or gesture recognition captured by a microphone or video camera, or any other peripheral device or control that allows user feedback to and interaction with the processing unit 110. The display 150 may be a monitor (e.g., a computer monitor, television, liquid crystal display (LCD), organic light-emitting diode (OLED), flat panel display, solid state display, or cathode ray tube (CRT) display, or an electronic whiteboard).

[0059] The network interface 140 provides information and data output by the processing unit 110 to other components and / or instruments (e.g., those that require one or more clock output signals). The network interface 140 may include one or more ports, drivers, or other types of interconnects and / or transceiver circuits. Optionally, clinical text (EHR) may be accessed through the network interface 140.

[0060] For example, the primary data source 160 may include an EHR retrieved through the network interface 140. A secondary data source 170 may be included to make the workflow and performance analysis more complete.

[0061] For purposes of explanation, the memory 120 is described as including modules, each of which includes machine-executable instructions (e.g., in software or a computer program) corresponding to the relevant capabilities of the system 100.

[0062] While the above embodiments can be applied to extracting SDOH information from clinical notes, these embodiments can also be applied to other information extraction tasks involving identifying triggers and argument spans, normalizing arguments, and predicting the relationships between triggers and argument spans.

[0063] In various embodiments of implementing components, systems, and / or methods using a programmable device (e.g., a computer-based system or a programmable logic device), it should be understood that the above systems and methods can be implemented using any of a variety of known or later-developed programming languages (e.g., "C", "C++", "C#", "Java", "Python", etc.). Accordingly, various storage media (e.g., computer disks, optical disks, electronic memories, etc.) can be prepared, which can contain information that can direct a device such as a computer to implement the above systems and / or methods. Once a suitable device can access the information and programs contained on the storage media, the storage media can provide the information and programs to the device, enabling the device to perform the functions of the systems and / or methods described herein. For example, if a computer disk containing appropriate materials such as source files, object files, executable files, etc. is provided to a computer, the computer can receive the information, configure itself appropriately, and perform the various functions of the systems and methods outlined in the above diagrams and flowcharts, thereby implementing various functions. That is, the computer can receive the various parts of the information related to the different elements of the above systems and / or methods from the disk, implement the individual systems and / or methods, and coordinate the functions of the above individual systems and / or methods.

[0064] In view of the present disclosure, it is noted that the various methods and devices described herein can be implemented in hardware, software, and firmware. Additionally, the various methods and parameters are included by way of example only and have no limiting significance. In view of the present disclosure, one of ordinary skill in the art can implement the present teachings when determining their own techniques and the instruments required to affect those techniques, while remaining within the scope of the invention. The functions of one or more of the processors described herein can be incorporated into a smaller number of or a single processing unit (e.g., a CPU) and can be implemented using an application-specific integrated circuit (ASIC) or a general-purpose processing circuit programmed to perform the functions described herein in response to executable instructions.

[0065] Finally, the foregoing discussion is merely intended to illustrate the system and should not be construed as limiting the appended claims to any particular embodiment or group of embodiments. Thus, while the system has been described in particular detail with reference to exemplary embodiments, it should also be understood that numerous modifications and alternative embodiments can be devised by those of ordinary skill in the art without departing from the broader and intended spirit and scope of the system as set forth in the following claims. Accordingly, the specification and drawings are to be regarded as illustrative rather than restrictive of the scope of the appended claims.< / bos>

Claims

1. A computer-implemented method for extracting target medical information from clinical notes stored in a memory, comprising: Retrieving a sequence of clinical text of an electronic case from the memory; Tokenizing the sequence of the clinical text to obtain a sequence of input tokens; Using a trained natural language-based transformer to transform the sequence of the input tokens into a sequence of structured output tokens; and Post-processing the structured output tokens to obtain an annotated text-label pair of the clinical text.

2. The method according to claim 1, wherein, The natural language-based transformer is a T5 transformer.

3. The method according to claim 2, wherein The T5 transformer includes: An encoder that receives a sequence as input and generates a sequence of representations; and A decoder that receives the sequence of representations and previously generated tokens as input to generate an output token at each time step.

4. The method according to claim 1, wherein, The post-processing further includes converting the text-label pair into a table format.

5. The method according to claim 1, wherein The target medical information is social determinants of health (SDOH) information.

6. A system for extracting target medical information from clinical notes stored in a memory, comprising: A preprocessing module configured to retrieve a sequence of clinical text of an electronic case from the memory and tokenize the sequence of the clinical text to obtain a sequence of input tokens; A sequence-to-structure model module configured to use a trained natural language-based transformer to transform the sequence of the input tokens into a sequence of structured output tokens; And A post-processing module configured to obtain an annotated text-label pair of the clinical text based on the structured output tokens.

7. The system according to claim 6, wherein The natural language-based transformer of the sequence-to-structure model module is a T5 transformer.

8. The system according to claim 7, wherein, The T5 transformer includes: An encoder that receives a sequence as input and generates a sequence of representations; and A decoder that receives the sequence of representations and previously generated tokens as input to generate an output token at each time step.

9. The system according to claim 6, wherein The post-processing module is further configured to convert the text-label pair into a table format.

10. The system according to claim 6, wherein, The target medical information is social determinants of health (SDOH) information.

11. A non-transitory computer-readable storage medium encoded with instructions that, when executed, extract target medical information from clinical notes stored in a memory, the non-transitory computer-readable storage medium including: A preprocessing module that, when executed, retrieves a sequence of clinical text of an electronic case from the memory and tokenizes the sequence of the clinical text to obtain a sequence of input tokens; A sequence-to-structure model module that, when executed, uses a trained natural language-based transformer to transform the sequence of the input tokens into a sequence of structured output tokens; And A post-processing module that, when executed, obtains an annotated text-label pair of the clinical text based on the structured output tokens.

12. The non-transitory computer-readable storage medium according to claim 11, wherein, The natural language-based transformer of the sequence-to-structure model module is a T5 transformer.

13. The non-transitory computer-readable storage medium according to claim 11, wherein, The T5 transformer includes: An encoder that receives a sequence as input and generates a sequence of representations; and A decoder that receives the sequence of the representations and previously generated tokens as inputs to generate an output token at each time step.

14. The non-transitory computer-readable storage medium according to claim 11, wherein, The post-processing module, when executed, converts the text-label pairs into a tabular format.

15. The non-transitory computer-readable storage medium according to claim 11, wherein, The target medical information is health social determinants of health (SDOH) information.

Images