CN112655054A

CN112655054A - Artificial intelligence medical symptom identification system based on end-to-end learning

Info

Publication number: CN112655054A
Application number: CN201880097271.XA
Authority: CN
Inventors: 杨晓庆; 李奘
Original assignee: Beijing Didi Infinity Technology and Development Co Ltd
Current assignee: Beijing Didi Infinity Technology and Development Co Ltd
Priority date: 2018-12-07
Filing date: 2018-12-07
Publication date: 2021-04-13
Also published as: WO2020113544A1

Abstract

Embodiments of the present application provide methods and systems for identifying medical conditions from a patient description. An exemplary artificial intelligence system includes a patient interactive interface configured to receive a patient description (103) including at least one word segment. The system also includes a processor (104). The processor (104) is configured to determine a word vector for each word within a word segment and a weight associated with the corresponding word vector. The processor (104) is further configured to determine a weighted word vector based on the word vector and the associated weights. The processor (104) is also configured to construct a word segment representation using the weighted word vector and determine a medical symptom based on the word segment representation.

Description

Artificial intelligence medical symptom identification system based on end-to-end learning

Technical Field

The present application relates to Artificial Intelligence (AI) systems and methods for identifying medical symptoms in patients, and more particularly, to AI systems and methods for identifying medical symptoms from patient descriptions using end-to-end learning.

Background

Pre-diagnosis is usually performed at a hospital in order to preliminarily determine the disease of a patient before sending the patient to a suitable physician. Pre-diagnosis is typically based on the symptoms described by the patient. For example, if a patient says she has a fever and a runny nose, she will be pre-diagnosed with a cold or flu and sent to a physician. If the patient says that she has a rash on her skin, she will be pre-diagnosed as having skin allergy and sent to the dermatologist.

Pre-diagnosis is typically performed by a healthcare practitioner (e.g., a doctor or nurse). For example, hospitals typically provide pre-diagnostic personnel at a check-in table to determine where patients should be sent. However, having a practitioner make a pre-diagnosis wastes valuable resources. An automatic pre-diagnosis method may be used to improve efficiency. For example, diagnostic robots are being developed to perform pre-diagnosis. These automated methods provide a preliminary diagnosis based on the symptoms described by the patient, e.g., based on a preprogrammed mapping between the disease and known symptoms.

However, patient description is inaccurate or unclear. For example, a patient may be affected by a disease or drug and may not be able to accurately express himself. Furthermore, patients are not practitioners and are therefore unfamiliar with medical terminology describing symptoms. In fact, patients, particularly when verbally describing symptoms, may use informal language, whereas medical terminology is generally formal. Thus, existing automated methods do not readily identify medical conditions from the patient description.

Embodiments of the present application address the above-mentioned problems by providing improved artificial intelligence systems and methods that automatically identify medical conditions from a patient's description using end-to-end learning.

Disclosure of Invention

Embodiments of the present application provide an artificial intelligence system for identifying a medical condition from a patient description. The artificial intelligence system includes a patient interactive interface configured to receive a patient description including at least one word segment. The system also includes a processor. The processor is configured to determine a word vector for each word within a word segment and a weight associated with the corresponding word vector. The processor is further configured to determine a weighted word vector based on the word vector and the associated weights. The processor is also configured to construct a word segment representation using the weighted word vector and determine a medical symptom based on the word segment representation.

Embodiments of the present application also provide an artificial intelligence method for identifying a medical condition from a patient description. The artificial intelligence method includes receiving, via a patient interactive interface, a patient description including at least one word segment. The method also includes determining, by the processor, a word vector for each word within a word segment and a weight associated with the corresponding word vector. The method also includes determining, by a processor, a weighted word vector based on the word vector and the associated weights. The method also includes constructing, by a processor, a word segment representation using the weighted word vector, and determining, by the processor, a medical symptom based on the word segment representation.

Embodiments of the present application also provide a non-transitory computer readable medium having instructions stored thereon that, when executed by a processor, cause the processor to perform an artificial intelligence method for identifying a medical condition from a patient description. The artificial intelligence method includes receiving a patient description including at least one word segment. The method also includes determining a word vector for each word within a word segment and a weight associated with the corresponding word vector. The method also includes determining a weighted word vector based on the word vector and the associated weights. The method further includes constructing a word segment representation using the weighted word vector and determining a medical symptom based on the word segment representation.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

Drawings

Fig. 1 is a schematic diagram of an exemplary AI system for identifying a medical condition from a patient description, according to some embodiments of the present invention.

Fig. 2 is a schematic diagram of an exemplary end-to-end learning model for learning an entity indicative of a medical condition based on a patient description, shown in accordance with some embodiments of the present invention.

Fig. 3 is a flow diagram illustrating an exemplary method for identifying a medical condition from a patient description according to some embodiments of the invention.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like parts.

Fig. 1 is a block diagram illustrating an exemplary AI system 100 for identifying a medical condition from a patient description according to some embodiments of the present invention. Consistent with the present application, the AI system 100 may receive a patient description 103 from a patient terminal 120. For example, the patient terminal 120 may be a mobile phone, desktop computer, laptop, PDA, robot, kiosk, or the like. The patient terminal 120 may include a patient interaction interface configured to receive a patient description 103 provided by a patient 130. In some embodiments, the patient terminal 120 may include a hard or soft keyboard for the patient 130 to enter the patient description 103. The patient terminal 120 may additionally or alternatively include a touch screen for the patient 130 to hand-write the patient description 103. Thus, the patient terminal 120 may record the patient description 103 as text. If the input is handwriting, the patient terminal 120 may automatically recognize the handwriting and convert it to textual information. In some other embodiments, the patient terminal 120 may include a microphone for recording the patient description 103 orally provided by the patient 130. The patient terminal 120 may automatically transcribe the recorded audio data into text. In some alternative embodiments, AI system 100 may receive patient description 103 captured by patient terminal 120 in its raw format and handwriting recognition and audio transcription may be performed automatically by AI system 100.

In some embodiments, as shown in fig. 1, AI system 100 may include a communication interface 102, a processor 104, a memory 106, and a storage 108. In some embodiments, the AI system 100 can have different modules in a single device, such as an Integrated Circuit (IC) chip (e.g., implemented as an Application Specific Integrated Circuit (ASIC) or Field Programmable Gate Array (FPGA)), or a stand-alone device with dedicated functionality. In some embodiments, one or more components of the AI system 100 may be located in the cloud, or may alternatively be in a single location (such as within a mobile device) or in a distributed location. The components of the AI system 100 may be in an integrated device, or distributed in different locations, but in communication with each other via a network (not shown). Consistent with the present application, the AI system 100 may be configured to automatically identify medical symptoms from the patient description 103 using end-to-end learning.

The communication interface 102 may be through a communication cable, a Wireless Local Area Network (WLAN), a Wide Area Network (WAN), a wireless network such as radio waves, a cellular network, and/or a local or short-range wireless network (e.g., Bluetooth)^TM) Or other communication method, to and from components such as the patient terminal 120. In some embodiments, communication interface 102 may include an Integrated Services Digital Network (ISDN) card, a cable modem, a satellite modem, or a modem to provide a data communication connection. As another example, communication interface 102 may include a Local Area Network (LAN) card to provide a data communication connection to a compatible LAN. Wireless links may also be implemented by the communication interface 102. In such implementations, the communication interface 102 may send and receive electrical, electromagnetic or optical signals that carry digital data streams representing various types of information.

Consistent with some embodiments, the communication interface 102 may receive data, such as a patient description 103, from the patient terminal 120. The patient description 103 may be received as text or in a raw format, such as audio or handwriting, acquired by the patient terminal 120. The patient description 103 may include a sentence or sentences that describe the symptoms and sensations of the patient 130. For example, patient 130 may describe her symptoms as "my head is repeatedly painful, feels somewhat dizziness, and my nose also appears to be runny. When patient description 103 is initially dictated by patient 130, the description may additionally contain various spoken words, such as kay, haoba, rowed, you know, like, and so forth. The communication interface 102 may also provide received data to the memory 106 and/or storage 108 for storage or to the processor 104 for processing.

The processor 104 may comprise any suitable type of general or special purpose microprocessor, digital signal processor, or microcontroller. The processor 104 may be configured as a separate processor module dedicated to identifying medical symptoms from the patient description 103 by using an end-to-end learning model. Alternatively, the processors 104 may be configured to share a processor module for performing other functions unrelated to medical condition identification.

As shown in FIG. 1, processor 104 may include a plurality of modules, such as a word embedding unit 140, an attention calculation unit 142, a word segment representation construction unit 144, a diagnostic unit 146, and the like. These modules (and any corresponding sub-modules or sub-units) may be hardware units (e.g., portions of an integrated circuit) of the processor 104 that are designed for use with other components or software units implemented by the processor 104 by executing at least a portion of a program. The program may be stored on a computer readable medium and when executed by the processor 104, it may perform one or more functions. Although FIG. 1 shows all of the units 140 and 146 within one processor 104, it is contemplated that these units may be distributed among multiple processors located close together or remotely from each other.

In some embodiments, the unit 140 executes 146 a computer program to automatically identify the medical condition from the patient description 103 applying an end-to-end learning model. For example, fig. 2 is a schematic diagram of an exemplary end-to-end learning model 200 for learning an entity indicative of a medical condition based on a patient description 103, shown in accordance with some embodiments of the present invention. The end-to-end learning model 200 may include several sub-models, such as a word embedding model 210, a bidirectional long-short term network (LSTM) model 220, a word segment representation model 230, and a softmax model 240. FIG. 2 will be described in conjunction with cell 140 and 146.

In some embodiments, when the patient description 103 contains multiple sentences, the segmentation unit 140 may divide the patient description 103 into different sentences. For example, the above description may be divided into three sentences, and then the end-to-end learning model 200 is applied to each sentence:

my head is repeatedly painful.

Some dizziness was felt.

And my nose also appears to run.

The word embedding unit 140 is configured to determine a word vector for each word in the sentence. Using the last sentence in the above exemplary description as an example, there are six words in the sentence, "and", "my", "nose", "seem", "runny nose" and "also on". Word vectors are determined for each of the six words using word embedding. As shown in fig. 2, six words w1, w2, w3, w4, w5, and w6 are input into the corresponding word embedding model 210. Word embedding model 210 maps words to real number vectors (referred to as "word vectors"). For example, the word embedding model 210 generates word vectors v1, v2, v3, v4, v5, and v6 for six words w1, w2, w3, w4, w5, and w6, respectively. Each word embedding model 210 encodes the meaning and features of a word as a number in a vector by mapping the word in one-dimensional space to a high-dimensional vector. In some embodiments, the word vector may be several hundred dimensions.

In some embodiments, the word embedding unit 140 may perform the mapping using methods such as neural networks, dimensionality reduction to word co-occurrence matrices, probabilistic models, interpretable knowledge bases, and explicit representations based on the context in which the words occur. For example, the word embedding learning model 210 may be implemented as a continuous vocabulary learning model (CBOW) or a Glove learning model, among others. In some embodiments, the word embedding learning model 210 may be trained using sample words and word vectors. The learning model 210 may be trained using different language databases to accommodate different languages, such as english, chinese, spanish, and the like.

The attention calculation unit 142 may be configured to determine word representations based on the word vectors and then calculate the attention of the respective words based on the word representations. In some embodiments, a two-way learning model, such as two-way LSTM model 220, may be used to generate the word representation. The bi-directional LSTM model 220 is a Recurrent Neural Network (RNN) that can process data in sequence and maintain its hidden state over time. Unlike a word vector, which contains the meaning and characteristics of a single word, a word representation additionally provides context information for the word, i.e., information for the entire sentence in which the word is located.

As shown in FIG. 2, the bi-directional LSTM model 220 may include two sets of LSTM cells designed to flow data in two different directions. For example, a group of LSTM cells processes the word vector in the order v1, v2, v3, v4, v5, and v6, such that data flows in the "forward" direction. Another set of LSTM cells processes these word vectors in the order v6, v5, v4, v3, v2, and v1, so that the data flows in the "backward" direction. In each group, a plurality of LSTM units are sequentially connected to each other. In some embodiments, two sets of LSTM units are connected internally to provide additional data streams. By using a bi-directional model, attention calculation unit 142 may obtain word representations that contain rich "bi-directional" (forward and backward) context information for the words. As shown in fig. 2, the word representations R1, R2, R3, R4, R5, and R6 can be determined.

In some embodiments, the attention calculation unit 142 is further configured to identify word segments. Consistent with this application, a "word segment" is a partition of a sentence that contains at least two words in the original order of the sentence. In some embodiments, the word segments may be identified by selecting a starting word and an ending word in the sentence, and the word between the two is the identified word segment. For example, if "my" is selected as the starting word and "snivel" is selected as the ending word, the segment may be identified as "my nose seems to be snivel too. Other word segments starting with "my" include "my nose", "my nose seems", "my nose also seems to run a nose".

In some embodiments, attention calculation unit 142 may identify word segments between two "substantive" words. Consistent with this application, a "parenchymal word" is a word having a substantial meaning that is indicative of, or otherwise associated with, a medical condition. An "insubstantial word" is any word that is not a substantive word. For example, the AI system 100 may identify a segment between the substantive words "repeat" and "head" that repeatedly pain my head. In some embodiments, the attention calculation unit 140 may determine whether a word is a concept word having a substantial meaning or a relation word that expresses only a grammatical relation between concept words to express a meaning. The relational terms may be determined to be "immaterial". Then, for the remaining concept words, the attention calculation unit 142 may determine whether they are related to a medical symptom. Thus, certain conceptual words, such as nouns used as subjects, e.g., "i", "we", "you", "it" as immaterial, and verbs and adjectives that describe symptoms meaningless, e.g., "have", "seem", "look", "feel", and "little bit", may be further filtered out.

In the embodiment shown in fig. 2, there are two word segments, the first comprising words (W1, W2, W3) and the second comprising words (W3, W4, W5, W6). For example, using the third sentence in the exemplary description above, the first term includes "and my nose" and the second term includes "the nose seems to be also runny".

In some embodiments, the attention calculation unit 142 is further configured to calculate the attention of each word in the identified word segments based on the word representation. "attention" is also referred to as an attention weight, which indicates the relative importance of each word in a word segment. Using the word segments (W3, W4, W5, and W6) in fig. 2 (e.g., "the nose seems to be running a nose") as an example, the attention calculation unit 142 may calculate the attention a3, a4, a5, and a6 of the words therein based on the word representations R3, R4, R5, and R6. Because W3 ("nose") and W5 ("runny nose") in this range are substantive words having more important meanings than others, the values of a3 and a5 may be greater than a4 and a 6.

The word vectors and the respective attentions are provided to a word segment representation construction unit 144 for constructing the word segment representations. In some embodiments, as shown in FIG. 2, a segment representation model 230 may be applied to assemble segment representations. As a first step, the weighted word vector of a word segment may be determined as a weighted sum of the word vectors weighted by the respective attention. Using the word segments (W3, W4, W5, W6) as an example, the weighted word vector is determined as Vs a 3W 3+ a 4W 4+ a 5W 5+ a 6W 6.

In some embodiments, the word segment representation may be a combination of a word representation of a starting word in the word segment, a weighted word vector, and a word representation of an ending word in the word segment. Combining word vectors and word list representations means arranging the vectors one by one. Again using the word segments (W3, W4, W5, W6) as an example, the word segment representation would be (R3, Vs, R6). Other configurations are also contemplated. For example, the segment representations may contain the weighted word vector itself, a weighted word vector combined with the word representations of the most important words (e.g., the words with the highest attention), a weighted word vector combined with the word representations of the two most important words at both ends, and so on.

The diagnostic unit 146 may detect one or more symptoms based on the field representation of the patient description 103. In some embodiments, a classification learning model may be used to classify word segment representations in classes associated with entities indicative of medical symptoms. For example, an entity may include "fever," headache, "" nausea, "" migraine, "" joint pain, "" runny nose, "" bleeding, "" swelling, "" stomach discomfort, "" vomiting, "and so forth, e.g., a word segment representation corresponding to the word segment" pain recurrently in the head "may be classified as associated with the entity" migraine. As another example, a term representation corresponding to the term "the nose appears to be also runny" may be classified as related to the entity "runny".

In some embodiments, the classification learning model may be a feed-forward neural network, such as softmax model 240 shown in fig. 2. In some embodiments, a feed-forward neural network, such as softmax model 240, may be trained using sample segment representations and entities of known medical symptoms. The sample word segment representation may be obtained by applying the word embedding model 210, the bi-directional LSTM model 220, and the word segment representation model 230 to a patient description provided by a sample patient. The relevant training entities may be provided by a medical professional (e.g., a doctor or nurse) by diagnosing the sample patient.

In some embodiments, based on the identified symptoms, the diagnostic unit 146 may perform a pre-diagnosis and provide the diagnostic result 105. For example, element 140-144 may identify the symptoms "my head is repeatedly painful and feels a bit dizzy and my nose seems to be runny," such as "headache," "migraine," "weakness," and "runny nose," detected by the patient description 103. Based on the symptoms, the diagnosis unit 146 may diagnose the disease from which the patient suffers in advance. For example, the diagnostic unit 146 may predict that the patient is likely to have influenza. In some embodiments, the diagnostic unit 146 may predict the disease using a learning model based on the symptoms. The learning model may be trained on the patient's sample symptoms and the physician's final diagnosis.

Although the above-described embodiments train the various sub-models of the end-to-end learning model 200 separately, in some embodiments, the end-to-end learning model 200 may also be trained as a whole. That is, the sub-models of the end-to-end learning model 200 may be trained jointly, rather than individually. For example, the end-to-end learning model 200 may be trained using the sample patient descriptions and their corresponding symptoms (e.g., as determined by a physician). The end-to-end learning model 200 may be trained using different language databases to accommodate different languages, such as english, chinese, spanish, and so on.

Memory 106 and storage 108 may comprise any suitable type of mass storage provided to store any type of information that processor 104 may need to operate. The memory 106 and storage 108 may be volatile or non-volatile, magnetic, semiconductor, tape, optical, removable, non-removable, or other types of storage devices or tangible (i.e., non-transitory) computer-readable media, including but not limited to ROM, flash memory, dynamic RAM, and static RAM. The memory 106 and/or storage 108 may be configured to store one or more computer programs that may be executed by the processor 104 to perform the functions disclosed herein. For example, the memory 106 and/or storage 108 may be configured to store a program that may be executed by the processor 104 to generate the diagnostic results 105 for the patient 130 using the end-to-end learning model 200.

The memory 106 and/or storage 108 may be further configured to store information and data used by the processor 104. For example, memory 106 and/or storage 108 may be configured to store various types of data (e.g., entities related to known symptoms). For example, entities may include "fever", "headache", "nausea", "migraine", "arthralgia", "runny nose", "hemorrhage", "swelling", "stomach discomfort", "vomiting", and the like.

In some embodiments, memory 106 and/or storage 108 may also store intermediate data such as word vectors, word representations, word segments, attention, weighted word vectors and word segment representations, and the like. The memory 106 and/or storage 108 may additionally store various learning models, including model parameters thereof, such as the word embedding model 210, the bi-directional LSTM model 220, the word segment representation model 230, and the softmax model 240 described above. Various types of data may be permanently stored, periodically deleted, or immediately ignored after processing the data. The diagnostic results 105 may be provided to the patient 130 via the display 150.

The display 150 may include a display such as a Liquid Crystal Display (LCD), a light emitting diode display (LED), a plasma display, or any other type of display, and provides a Graphical User Interface (GUI) presented on the display for user input and data description. The display may comprise many different types of materials, such as plastic or glass, and may be touch sensitive to receive input from a user. For example, the display may comprise a substantially rigid touch sensitive material, such as Gorilla Glass^TMOr substantially flexible, e.g. Willow Glass^TM. In some embodiments, the display 150 may be part of the patient terminal 120.

For example, fig. 3 is a flow diagram illustrating an exemplary method 300 for identifying a medical condition from a patient description, according to some embodiments of the invention. The method 300 may be implemented by the AI system 100, in particular the processor 104 or a separate processor not shown in fig. 1. The method 300 may include steps S302-S320 as described below. It should be understood that some steps may be optional to perform the present application as provided herein. Further, some steps may be performed simultaneously, or in a different order than shown in fig. 3.

In step S302, the AI system 100 can receive a patient description, such as the patient description 103. The patient description 103 may be received as text or in a raw format, such as audio or handwriting, acquired by the patient terminal 120. If received as audio, the patient description 103 may be transcribed into text. If received in handwriting, the patient description 103 may be automatically recognized and converted to text. The patient description 103 may include a sentence or sentences that describe the symptoms of the patient 130. For example, patient 130 may describe her symptoms as "my head is repeatedly painful, also feeling somewhat dizziness, and my nose also seems to be runny. "

In some embodiments, when the patient description 103 contains multiple sentences, the AI system 100 can divide the patient description 103 into different sentences. For example, the above exemplary description may be divided into three sentences: "my head is repeatedly experiencing pain. "also felt somewhat dizzy. "and" my nose also appears to be runny. "

In step S304, the AI system 100 can determine a word vector for each word in the sentence of the patient description 103. In some embodiments, word embedding is used to determine a word vector, which maps words to a real number vector. In some embodiments, the word vector may be several hundred dimensions. As shown in fig. 2, six words w1, w2, w3, w4, w5, and w6 are input into the corresponding word embedding model 210. The word embedding model 210 generates word vectors v1, v2, v3, v4, v5, and v6 for six words w1, w2, w3, w4, w5, and w6, respectively. In some embodiments, the word embedding learning model 210 may be implemented as a continuous vocabulary learning model (CBOW) or a Glove learning model, among others. In some embodiments, the word embedding learning model 210 may be trained using sample words and word vectors.

In step S306, the AI system 100 may determine a word representation for each word based on the word vectors. In addition to the meaning of the individual words, the word representations also provide contextual information of the words, i.e. information of the entire sentence in which the word is located. In some embodiments, a two-way learning model, such as the two-way LSTM model 220 shown in fig. 2, may be used to generate the word representations R1, R2, R3, R4, R5, and R6.

The two-way learning model may include two layers, each designed to flow data in a different direction. As shown in FIG. 2, the bi-directional LSTM model 220 includes two sets of LSTM cells, one in the "forward" direction and one in the "backward" direction. For example, a group of LSTM cells processes the word vector in the order v1, v2, v3, v4, v5, and v6, such that data flows in the "forward" direction. Another set of LSTM cells processes these word vectors in the order v6, v5, v4, v3, v2, and v1, so that the data flows in the "backward" direction. In each group, a plurality of LSTM units are sequentially connected to each other. In some embodiments, two sets of LSTM units are connected internally to provide additional data streams.

In step S308, the AI system 100 may identify a word segment from the patient description 103. For example, fig. 2 shows two recognized word segments, the first word segment comprising words (W1, W2, W3), the second word segment comprising words (W3, W4, W5, W6). In some embodiments, word segments may be identified by words between the starting word and the ending word. For example, the AI system 100 may select "my" as the starting word and "snivel" as the ending word to identify the word segment "my nose appears to be snivel". In some embodiments, attention calculation unit 142 may identify word segments between two "substantive" words. Consistent with this application, a "parenchymal word" is a word having a substantial meaning that is indicative of, or otherwise associated with, a medical condition. For example, the AI system 100 may identify a segment between the substantive words "repeat" and "head" that repeatedly pain my head.

In step S310, the AI system 100 can calculate the attention of the words within the identified word segments based on the word representations of the words. For example, the AI system 100 may calculate the attentions a3, a4, a5, and a6 of the word segments (W3, W4, W5, and W6) in fig. 2 based on the word representations R3, R4, R5, and R6 of the respective words W3, W4, W5, and W6. If (W3, W4, W5 and W6) is "the nose also seems to run a lot", the values of a3 and a5 may be greater than a4 and a6, because W3 ("nose") and W5 ("running a lot") are substantive words having more important meaning than others in this paragraph.

In step S312, the AI system 100 can calculate a weighted word vector for the identified word segments. In some embodiments, the weighted word vector for a word segment may be determined as a weighted sum of the word vectors weighted by the respective attentions. Again using the word segments (W3, W4, W5, W6) as an example, the weighted word vector is determined as Vs 3W 3+ a 4W 4+ a 5W 5+ a 6W 6.

In step S314, the AI system 100 can construct a word segment representation. In some embodiments, the word segment representation may be a combination of a weighted word vector and at least one word representation. For example, a starting word representation of a word segment, a weighted word vector, and an ending word representation of the word segment may be combined to form a word segment representation. Using the word segments of fig. 2 (W3, W4, W5, W6) as an example, the word segment representation will be constructed as (R3, Vs, R6), which is a combination of the word representation R3 of the first word W3, the weighted word vector Vs, and the word representation R6 of the last word W6 in the word segment. Other structures may include the weighted word vector itself, a weighted word vector combined with the word representation of the most important word (e.g., the word with the highest attention), a weighted word vector combined with the word representations of the two most important words at both ends, and so on.

In step S316, the AI system 100 can apply a classifier on the word segment representation to determine a matching entity. In some embodiments, a classification learning model may be used to classify word segment representations in classes associated with entities indicative of medical symptoms. In some embodiments, the classification learning model may be a feed-forward neural network, such as softmax model 240 shown in fig. 2. For example, a word segment representation of the word segment "recurrent pain in the head" may be matched to the entity "migraine". As another example, a word segment representation corresponding to the word segment "the nose appears to be runny" may be matched to the entity "runny".

In step S318, the AI system 100 may determine whether all of the word segments have been identified and matched with the entity. If not all of the word segments are considered (S318: NO), the method 300 returns to step S308 to identify another word segment. Otherwise, if all of the word segments are considered (S318: YES), the method 300 proceeds to step S320, where the AI system 100 makes a pre-diagnosis based on the symptoms described by the matching entities. For example, the medical condition detected from the patient description 103 "my head is repeatedly painful, also feels a bit dizzy, my nose also seems to run a nose" may include "headache", "migraine", "dizziness" and "runny nose". Based on these symptoms, the AI system 100 can predict that the patient is likely to have influenza. In some embodiments, the AI system 100 can predict disease using a learning model based on symptoms. The learning model may be trained on the patient's sample symptoms and the physician's final diagnosis.

Another aspect of the application relates to a non-transitory computer-readable medium storing instructions that, when executed, cause one or more processors to perform a method as described above. The computer-readable medium includes volatile or nonvolatile, magnetic, semiconductor, tape, optical, removable, non-removable, or other types of computer-readable medium or computer-readable storage device. For example, a computer-readable medium as in the present application may be a storage device or a memory module having stored thereon computer instructions. In some embodiments, the computer readable medium may be a disk or flash drive having computer instructions stored thereon.

It will be apparent that various modifications and variations can be made in the system and related methods of the present application by those of ordinary skill in the art. Other embodiments will be apparent to those skilled in the art from consideration of the specification and practice of the system and associated method of the present application.

It is intended that the specification and examples be considered as exemplary only, with a true scope being indicated by the following claims and their equivalents.

Claims

1. An artificial intelligence system for identifying a medical condition from a patient description, comprising:

a patient interaction interface configured to receive a patient description including at least one word segment; and

a processor configured to:

determining a word vector for each word in a word segment and a weight associated with the corresponding word vector;

determining a weighted word vector based on the word vector and the associated weights;

constructing a word segment representation using the weighted word vectors; and

based on the word segment representation, a medical symptom is determined.

2. The artificial intelligence system of claim 1 wherein the word vector is determined using word embedding.

3. The artificial intelligence system of claim 1, wherein to determine the weights, the processor is further configured to:

determining a word representation of each of the words within the word fragments; and

determining the attention of each said word representation as said weight associated with the corresponding word vector.

4. The artificial intelligence system of claim 3 wherein to determine the word representations, the processor is further configured to apply a bi-directional learning model to the respective word vectors.

5. The artificial intelligence system of claim 4 wherein the Bi-directional learning model is a Bi-LSTM network.

6. The artificial intelligence system of claim 1 wherein the weighted word vectors are a sum of each of the word vectors weighted by an associated weight.

7. The artificial intelligence system of claim 1 wherein the processor is further configured to apply a classification learning model to classify the word segment representations in categories relating to medical symptoms.

8. The artificial intelligence system of claim 7 wherein the classification learning model is a feed-forward neural network.

9. The artificial intelligence system of claim 8, wherein the classification learning model is a softmax network.

10. The artificial intelligence system of claim 1 wherein to construct the word segment representation, the processor is further configured to combine a word representation of at least one word within the word segment with the weighted word vector.

11. An artificial intelligence method for identifying a medical condition from a patient description, comprising:

receiving, via a patient interactive interface, a patient description including at least one word segment;

determining, by a processor, a word vector for each word in a word segment and a weight associated with the corresponding word vector;

determining, by the processor, a weighted word vector based on the word vector and the associated weights;

constructing, by the processor, a word segment representation using the weighted word vector; and

determining, by the processor, a medical symptom based on the word segment representation.

12. The artificial intelligence method of claim 11 wherein the word vector is determined using word embedding.

13. The artificial intelligence method of claim 11, wherein determining the weight further comprises:

determining the attention of the respective word representation as the weight associated with the corresponding word vector.

14. The artificial intelligence method of claim 13 wherein determining the word representation further comprises applying a bi-directional learning model to the respective word vector.

15. The artificial intelligence method of claim 14 wherein the Bi-directional learning model is a Bi-LSTM network.

16. The artificial intelligence method of claim 11 wherein the weighted word vectors are a sum of each of the word vectors weighted by an associated weight.

17. The artificial intelligence method of claim 11, wherein the processor is further configured to apply a feed-forward neural network to classify the word segment representations in categories related to medical symptoms.

18. The artificial intelligence method of claim 11, wherein constructing the word segment representation further comprises combining a word representation of at least one word within the word segment and the weighted word vector.

19. A non-transitory computer readable medium having instructions stored thereon that, when executed by a processor, cause the processor to perform an artificial intelligence method for identifying a medical condition from a patient description, the artificial intelligence method comprising:

receiving a patient description including at least one word segment;

constructing a word segment representation using the weighted word vectors; and

based on the word segment representation, a medical symptom is determined.

20. The non-transitory computer-readable medium of claim 19, wherein determining the weight further comprises:

applying a bi-directional learning model to the respective word vector to determine a word representation of the respective word within the word segment; and