WO2020113544A1

WO2020113544A1 - Artificial intelligence medical symptom recognition system based on end-to-end learning

Info

Publication number: WO2020113544A1
Application number: PCT/CN2018/119720
Authority: WO
Inventors: Xiaoqing Yang; Zang Li
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: 2020-06-11
Anticipated expiration: 2021-06-07
Also published as: CN112655054B; CN112655054A

Abstract

A method and system for recognizing a medical symptom from a patient description. An exemplary artificial intelligence system includes a patient interaction interface configured to receive the patient description (103) including at least one span. The system also includes a processor (104). The processor (104) is configured to determine word vectors for words in a span and weights associated with the respective word vectors. The processor (104) is further configured to determine a weighted word vector based on the word vectors and the associated weights. The processor (104) is also configured to construct a span representation using the weighted word vector, and determine the medical symptom based on the span representation.

Description

ARTIFICIAL INTELLIGENCE MEDICAL SYMPTOM RECOGNITION SYSTEM BASED ON END-TO-END LEARNING

TECHNICAL FIELD

The present disclosure relates to artificial intelligence (AI) systems and methods for recognizing a patient’s medical symptoms, and more particularly to, AI systems and methods for medical symptom recognition from the patient’s descriptions using end-to-end learning.

BACKGROUND

Pre-diagnosis is usually performed in hospitals to preliminarily determine the illnesses of patients before sending them to the right doctors. Pre-diagnosis is typically based on symptoms described by the patient. For example, if the patient says she has a fever and a running nose, she will be pre-diagnosed as having a cold or a flu and be sent to an internal medicine doctor. If the patient says that she has itchy rashes on her skin, she will be pre-diagnosed as having skin allergies and be sent to a dermatologist.

Pre-diagnosis is typically performed by medical practitioners, such as physicians or nurses. For example, hospitals usually have pre-diagnosis personnel available at the check-in desk to determine where the patient should be sent to. However, having practitioners perform the pre-diagnosis wastes valuable resources. Automated pre-diagnosis methods are used to improve the efficiency. For example, diagnosis robots are being developed to perform the pre-diagnosis. These automated methods provide a preliminary diagnosis based on patient’s described symptoms, e.g., based on preprogramed mappings between diseases and known symptoms.

Patient descriptions are, however, not accurate or clear. For example, the patient may be under the influence of the illness or medicine and could not express herself accurately. In addition, patients are not practitioners and are therefore not familiar with medical terminologies for describing symptoms. Indeed, patients, especially when describing symptoms orally, may use informal language while medical terminologies are usually formal. As a result, existing automated methods could not readily recognize medical symptoms from patient descriptions.

Embodiments of the disclosure address the above problems by providing improved artificial intelligence systems and methods for automatically recognizing medical symptoms from patient’s descriptions using end-to-end learning.

SUMMARY

Embodiments of the disclosure provide an artificial intelligence system for recognizing a medical symptom from a patient description. The artificial intelligence system includes a patient interaction interface configured to receive the patient description including at least one span. The system also includes a processor. The processor is configured to determine word vectors for words in a span and weights associated with the respective word vectors. The processor is further configured to determine a weighted word vector based on the word vectors and the associated weights. The processor is also configured to construct a span representation using the weighted word vector, and determine the medical symptom based on the span representation.

Embodiments of the disclosure also provide an artificial intelligence method for recognizing a medical symptom from a patient description. The artificial intelligence method includes receiving, by a patient interaction interface, the patient description including at least one span. The method further includes determining, by the processor, word vectors for words in a span and weights associated with the respective word vectors. The method also includes determining, by the processor, a weighted word vector based on the word vectors and the associated weights. The method additionally includes constructing, by the processor, a span representation using the weighted word vector, and determining, by the processor, the medical symptom based on the span representation.

Embodiments of the disclosure further provide a non-transitory computer-readable medium having instructions stored thereon that, when executed by a processor, causes the processor to perform an artificial intelligence method for recognizing a medical symptom from a patient description. The artificial intelligence method includes receiving the patient description including at least one span. The method further includes determining word vectors for words in a span and weights associated with the respective word vectors. The method also includes determining a weighted word vector based on the word vectors and the associated weights. The method additionally includes constructing a span representation using the weighted word vector, and determining the medical symptom based on the span 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.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a schematic diagram of an exemplary AI system for recognizing a medical symptom from a patient description, according to embodiments of the disclosure.

FIG. 2 illustrates a schematic diagram of an exemplary end-to-end learning model for learning an entity indicating a medical symptom based on a patient description, according to embodiments of the disclosure.

FIG. 3 illustrates a flowchart of an exemplary method for recognizing a medical symptom from a patient description, according to embodiments of the disclosure.

DETAILED DESCRIPTION

Reference will now be made in detail to the 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 illustrates a block diagram of an exemplary AI system 100 for recognizing a medical symptom from a patient description, according to embodiments of the disclosure. Consistent with the present disclosure, AI system 100 may receive patient description 103 from a patient terminal 120. For example, patient terminal 120 may be a mobile phone, a desktop computer, a laptop, a PDA, a robot, a kiosk, etc. Patient terminal 120 may include a patient interaction interface configured to receive patient description 103 provided by patient 130. In some embodiments, patient terminal 120 may include a keyboard, hard or soft, for patient 130 to type in patient description 103. Patient terminal 120 may additionally or alternatively include a touch screen for patient 130 to handwrite patient description 103. Accordingly, patient terminal 120 may record patient description 103 as texts. If the input is handwriting, patient terminal 120 may automatically recognize the handwriting and convert it to text information. In some other embodiments, patient terminal 120 may include a microphone, for recording patient description 103 provided by patient 130 orally. Patient terminal 120 may automatically transcribe the recorded audio data into texts. In some alternative embodiments, AI system 100 may receive patient description 103 in its original format as captured by patient terminal 120, and the 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, AI system 100 may 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 a field-programmable gate array (FPGA) ) , or separate devices with dedicated functions. In some embodiments, one or more components of AI system 100 may be located in a cloud, or may be alternatively in a single location (such as inside a mobile device) or distributed locations. Components of AI system 100 may be in an integrated device, or distributed at different locations but communicate with each other through a network (not shown) . Consistent with the president disclosure, AI system 100 may be configured to automatically recognize medical symptoms from patient description 103 using end-to-end learning.

Communication interface 102 may send data to and receive data from components such as patient terminal 120 via communication cables, a Wireless Local Area Network (WLAN) , a Wide Area Network (WAN) , wireless networks such as radio waves, a cellular network, and/or a local or short-range wireless network (e.g., Bluetooth ^TM) , or other communication methods. In some embodiments, communication interface 102 may include an integrated services digital network (ISDN) card, cable modem, 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 can also be implemented by communication interface 102. In such an implementation, communication interface 102 can send and receive electrical, electromagnetic or optical signals that carry digital data streams representing various types of information.

Consistent with some embodiments, communication interface 102 may receive data such as patient description 103 from patient terminal 120. Patient description 103 may be received as texts or in its original format as acquired by patient terminal 120, such as an audio or in handwriting. Patient description 103 may include one sentence or multiple sentences that describe the symptoms and feelings of patient 130. For example, patient 130 may describe her symptom as “I am having a recurring pain in the head, also feeling a bit dizzy, and my nose seems running too. ” When patient description 103 is originally provided by patient 130 orally, the description may additionally contain various spoken language words, such as, hmm, well, all right, you know, okay, so, etc. Communication interface 102 may further provide the received data to memory 106 and/or storage 108 for storage or to processor 104 for processing.

Processor 104 may include any appropriate type of general-purpose or special-purpose microprocessor, digital signal processor, or microcontroller. Processor 104 may be configured as a separate processor module dedicated to recognizing medical symptom (s) from patient description 103 by using an end-to-end learning model. Alternatively, processor 104 may be configured as a shared processor module for performing other functions unrelated to medical symptom recognition.

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

In some embodiments, units 140-146 execute a computer program to apply an end-to-end learning model to automatically recognize medical symptoms from patient description 103. For example, FIG. 2 illustrates a schematic diagram of an exemplary end-to-end learning model 200 for learning an entity indicating a medical symptom based on patient description 103, according to embodiments of the disclosure. End-to-end learning model 200 may include several sub-models, such as word embedding models 210, a bi-directional Long Short-Term Memory (LSTM) model 220, span representation models 230, and softmax models 240. FIG. 2 will be described together with units 140-146.

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

I am having a recurring pain in the head.

Also feeling a bit dizzy.

And my nose seems running too.

Word embedding unit 140 is configured to determine a word vector for each word in the sentence. Using the last sentence in the exemplary description above as an example, there are six words in the sentence, “And, ” “my, ” “nose, ” “seems, ” “running, ” and “too. ” A word vector is 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 respective word embedding models 210. Word embedding models 210 map the words to vectors of real numbers (referred to as “word vectors” ) . For example, word embedding models 210 generate word vectors v1, v2, v3, v4, v5, and v6 for the six words w1, w2, w3, w4, w5, and w6, respectively. By mapping a word in a space of one-dimension to a vector of a high-dimension, each word embedding model 210 encodes the meanings and features of the word into the numbers in the vector. In some embodiments, the word vectors may be of several hundred dimensions.

In some embodiments, word embedding unit 140 may perform the mapping using methods such as neural networks, dimensionality reduction on the word co-occurrence matrix, probability models, explainable knowledge base method, and explicit representation in terms of the context in which words appear. For example, word embedding learning model 210 may be implemented as a Continuous Bag of Words (CBOW) learning model or a Glove learning model, etc. In some embodiments, word embedding learning models 210 may be trained using sample words and word vectors. Word embedding learning models 210 may be trained using different language database to accommodate different languages, such as English, Chinese, Spanish, etc.

Attention calculation unit 142 may be configured to determine word representations based on the word vectors and then calculate attentions for the respective words based on the word representations. In some embodiments, a bi-directional learning model, such as bi-directional LSTM model 220, may be used to generate the word representations. Bi-directional LSTM model 220 is a type of recurrent neural network (RNN) and may process data sequentially and keep its hidden state through time. Unlike word vectors that contain meanings and features of the individual words, word representations additionally provide context information of the words, i.e., information of the entire sentence the words are in.

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

In some embodiments, attention calculation unit 142 is further configured to identify a span. Consistent with the disclosure, a “span” is a partition of a sentence that contains a plurality of words of that sentence in their original order. In some embodiments, a span can be identified by selecting a starting word and an ending word in the sentence and the words between the two becomes the identified span. For example, if “my” is selected as the starting word and “running” is selected as the ending word, the span can be identified as “my nose seems running. ” Other spans with “my” as the starting word include “my nose, ” “my nose seems, ” and “my nose seems running too. ”

In some embodiments, attention calculation unit 142 may identify spans that are between two “substantive” words. Consistent with present disclosure, a “substantive word” is a word that has substantive meaning indicating or otherwise related to medical symptoms. A “non-substantive word” is any word that is not a substantive word. For example, AI system 100 may identify span “recurrent pain in my head” between substantive words “recurrent” and “head. ” In some embodiments, attention calculation unit 140 may determine whether a word is a notional word that has substantive meanings or a relational word that merely expresses a grammatical relationship between notional words to express the meanings. A relational word may be determined as “non-substantive. ” For the remaining notional words, attention calculation unit 142 may then determine whether they are related to medical symptoms. Accordingly, certain notional words may be further filtered out, such as nouns used as the subject, e.g., “I, ” “we, ” “you, ” “it” as non-substantive, and verbs and adjectives that do not meaningfully describe a symptom, e.g., “have, ” “seem, ” “look, ” “feel, ” and “a little bit. ”

In the embodiment shown in FIG. 2, there are two spans, first including words (W1, W2, W3) and second including words (W3, W4, W5, W6) . For example, using the third sentence in the exemplary description above, the first span includes “And my nose” and the second span includes “nose seems running too. ”

In some embodiments, attention calculation unit 142 is further configured to calculate an attention for each word in the identified span based on the word representations. An “attention” is also known as an attention weight, which indicates the relative importance of each word in the span. Using the span (W3, W4, W5, and W6) in FIG. 2 (e.g., “nose seems running too” ) as an example, attention calculation unit 142 may calculate attentions a3, a4, a5, and a6 for the words therein, based on word representations R3, R4, R5, and R6. Because W3 ( “nose” ) and W5 ( “running” ) in this span are substantive words that carry meanings more important than the others, a3 and a5 may be larger in value than a4 and a6.

The word vectors and respective attentions are provided to span representation construction unit 144 to construct span representations. In some embodiments, span representation models 230, as shown in FIG. 2, may be applied to assemble the span representations. As a first step, a weighted word vector of the span may be determined as a weighted sum of the word vectors weighted by the respective attentions. Using span (W3, W4, W5, W6) as an example, the weighted word vector is determined as Vs=a3*W3+a4*W4+a5*W5+a6*W6.

In some embodiments, the span representation may be an assembly of the word representation of the starting word in the span, the weighted word vector, and the word representation of the ending word in the span. Assembling the word vector and word representations means lining up the vectors one by one. Again using span (W3, W4, W5, W6) as an example, the span representation will be (R3, Vs, R6) . Other constructions are also contemplated. For example, the span representation may contain the weighted word vector itself, weighted word vector assembled with the word representation of the most important word (e.g., word with highest attention) , weighted word vector assembled with the word representations of two most important words on the two ends, etc.

Diagnosis unit 146 may detect one or more symptoms based on the span representations of patient description 103. In some embodiments, a classification learning model may be used to classify the span representation in a class associated with an entity indicative of a medical symptom. For example, the entities may include “fever, ” “headache, ” “nausea, ” “migraine, ” “joint pain, ” “running nose, ” “bleeding, ” “swelling, ” “upset stomach, ” “vomit, ” etc. For example, a span representation corresponding to the span “recurring pain in the head” may be classified to be associated with entity “migraine. ” As another example, a span representation corresponding to the span “nose seems running too” may be classified to be associated with entity “running nose. ”

In some embodiments, the classification learning model may be a feedforward neural network, such as softmax models 240 shown in FIG. 2. In some embodiments, the feedforward neural network, e.g., softmax models 240, may be trained using sample span representations and entities of known medical symptoms. Sample span representations may be obtained by applying word embedding models 210, bi-directional LSTM model 220 and span representation models 230 on patient descriptions provided by sample patients. The training entities associated may be provided by medical professionals such as physicians or nurses by diagnosing the sample patients.

In some embodiments, based on the recognized symptoms, diagnosis unit 146 may make a pre-diagnosis and provide diagnosis result 105. For example, units 140-144 may recognize symptoms described by several entities detected from patient description 103 “I am having a recurring pain in the head, also feeling a bit dizzy, and my nose seems running too, ” such as “headache, ” “migraine, ” “faint, ” and “running nose. ” Based on the symptoms, diagnosis unit 146 may pre-diagnose the illness sustained by the patient. For example, diagnosis unit 146 may predict that the patient likely has a flu. In some embodiments, diagnosis unit 146 may use a learning model to predict the illness based on the symptoms. The learning model may be trained with sample symptoms of patients and the final diagnosis of the patients made by physicians.

Although the embodiments described above train the various sub-models of end-to-end learning model 200 individually, in some embodiments, end-to-end learning model 200 may also be trained as a whole. That is, the sub-models of end-to-end learning model 200 may be trained jointly, rather than individually. For example, end-to-end learning model 200 may be trained using sample patient descriptions and their corresponding symptoms, e.g., as determined by physicians. End-to-end learning model 200 may be trained using different language database to accommodate different languages, such as English, Chinese, Spanish, etc.

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

Memory 106 and/or storage 108 may be further configured to store information and data used by processor 104. For instance, memory 106 and/or storage 108 may be configured to store the various types of data (e.g., entities associated with known symptoms) . For example, entities may include “fever, ” “headache, ” “nausea, ” “migraine, ” “joint pain, ” “running nose, ” “bleeding, ” “swelling, ” “upset stomach, ” “vomit, ” etc.

In some embodiments, memory 106 and/or storage 108 may also store intermediate data such as the word vectors, word representations, spans, attentions, weighted word vectors, and span representations, etc. Memory 106 and/or storage 108 may additionally store various learning models including their model parameters, such as word embedding models 210, a bi-directional LSTM model 220, span representation models 230, and softmax models 240 that are described above. The various types of data may be stored permanently, removed periodically, or disregarded immediately after the data is processed.

Diagnosis result 105 may be provided to patient 130 through a display 150. 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 provide a Graphical User Interface (GUI) presented on the display for user input and data depiction. The display may include a number of different types of materials, such as plastic or glass, and may be touch-sensitive to receive inputs from the user. For example, the display may include a touch-sensitive material that is substantially rigid, such as Gorilla Glass ^TM, or substantially pliable, such as Willow Glass ^TM. In some embodiments, display 150 may be part of patient terminal 120.

For example, FIG. 3 illustrates a flowchart of an exemplary method 300 for recognizing a medical symptom from a patient description, according to embodiments of the disclosure. Method 300 may be implemented by AI system 100 and particularly processor 104 or a separate processor not shown in FIG. 1. Method 300 may include steps S302-S320 as described below. It is to be appreciated that some of the steps may be optional to perform the disclosure provided herein. Further, some of the steps may be performed simultaneously, or in a different order than shown in FIG. 3.

In step S302, AI system 100 may receive a patient description, e.g., patient description 103. Patient description 103 may be received as texts or in its original format as acquired by patient terminal 120, such as an audio or in handwriting. If received as an audio, patient description 103 may be transcribed into texts. If received in handwriting, patient description 103 may be automatically recognized and convert into texts. Patient description 103 may include one sentence or multiple sentences that describe the symptoms of patient 130. For example, patient 130 may describe her symptom as “I am having a recurring pain in the head, also feeling a bit dizzy, and my nose seems running too. ”

In some embodiments, when patient description 103 contains multiple sentences, AI system 100 may first divide patient description 103 into different sentences. For example, the above exemplary description may be divided into three sentences: “I am having a recurring pain in the head. ” “Also feeling a bit dizzy. ” and “And my nose seems running too. ”

In step S304, AI system 100 may determine a word vector for each word in a sentence of patient description 103. In some embodiments, the word vectors are determined using word embedding, which maps the words to vectors of real numbers. In some embodiments, the word vectors may be of several hundred dimensions. As shown in FIG. 2, six words w1, w2, w3, w4, w5, and w6 are input into respective word embedding models 210. Word embedding models 210 generate word vectors v1, v2, v3, v4, v5, and v6 for the six words w1, w2, w3, w4, w5, and w6, respectively. In some embodiments, word embedding learning model 210 may be implemented as a Continuous Bag of Words (CBOW) learning model or Glove learning model, etc. In some embodiments, word embedding learning models 210 may be trained using sample words and word vectors.

In step S306, AI system 100 may determine a word representation for each word, based on the word vectors. Word representations provide context information of the words, i.e., information of the entire sentence the words are in, in addition to the meanings of the individual words. In some embodiments, a bi-directional learning model, such as bi-directional LSTM model 220 shown in FIG. 2, may be used to generate the word representations R1, R2, R3, R4, R5, and R6.

The bi-directional learning model may include two layers, each designed to let data flow in a different direction. As shown in FIG. 2, bi-directional LSTM model 220 includes two sets of LSTM cells, one in the “forward” direction and the other in the “backward” direction. For example, one set of LSTM cells process word vectors in the order of v1, v2, v3, v4, v5, and v6 so that data flows in the “forward” direction. Another set of LSTM cells process these word vectors in the order of v6, v5, v4, v3, v2, and v1 so that data flows in the “backward” direction. Within each set, the multiple LSTM cells are connected sequentially with each other. In some embodiments, the two sets of LSTM cells are internally connected to provide additional data flow.

In step S308, AI system 100 may identify a span from patient description 103. For example, FIG. 2 shows two identified spans, first including words (W1, W2, W3) and second including words (W3, W4, W5, W6) . In some embodiments, a span can be identified by words between a starting word and an ending word. For example, AI system 100 may select “my” as the starting word and “running” as the ending word to identify a span “my nose seems running. ” In some embodiments, attention calculation unit 142 may identify spans that are between two “substantive” words. Consistent with present disclosure, a “substantive word” is a word that has substantive meaning indicating or otherwise related to medical symptoms. For example, AI system 100 may identify span “recurrent pain in my head” between substantive words “recurrent” and “head. ”

In step S310, AI system 100 may calculate attentions for the words in the identified span based on the word representations of these words. For example, AI system 100 may calculate attentions a3, a4, a5, and a6 for the span (W3, W4, W5, and W6) in FIG. 2, based on word representations R3, R4, R5, and R6 for the respective words W3, W4, W5, and W6. If (W3, W4, W5, and W6) is “nose seems running too, ” a3 and a5 may be larger in value than a4 and a6, as W3 ( “nose” ) and W5 ( “running” ) in this span are substantive words that carry meanings more important than the others.

In step S312, AI system 100 may calculate a weighted word vector for the identified span. In some embodiments, the weighted word vector of the span may be determined as a weighted sum of the word vectors weighted by the respective attentions. Again using span (W3, W4, W5, W6) as an example, the weighted word vector is determined as Vs=a3*W3+a4*W4+a5*W5+a6*W6.

In step S314, AI system 100 may construct a span representation. In some embodiments, the span representation may be an assembly of the weighted word vector and at least one word representation. For example, the starting word representation of the span, the weighted word vector, and the ending word representation of the span may be assembled to form the span representation. Using span (W3, W4, W5, W6) of FIG. 2 as an example, the span representation will be constructed as (R3, Vs, R6) , an assembly of the word representation R3 of the first word W3 in the span, the weighted word vector Vs, and the word representation R6 of the last word W6. Other constructions may include the weighted word vector itself, weighted word vector assembled with the word representation of the most important word (e.g., word with highest attention) , weighted word vector assembled with the word representations of two most important words on the two ends, etc.

In step S316, AI system 100 may apply a classifier on the span representation to determine a matched entity. In some embodiments, a classification learning model may be used to classify the span representation in a class associated with an entity indicative of a medical symptom. In some embodiments, the classification learning model may be a feedforward neural network, such as softmax models 240 shown in FIG. 2. For example, a span representation of the span “recurring pain in the head” may be matched with entity “migraine. ” As another example, a span representation corresponding to the span “nose seems running too” may be matched with entity “running nose. ”

In step S318, AI system 100 may determine if all span have been identified and matched with the entities. If not all spans are accounted for (S318: no) , method 300 returns to step S308 to identify another span. Otherwise, if all spans are accounted for (S318: yes) , method 300 proceeds to step S320, where AI system 100 makes a pre-diagnosis based on symptoms described by the matched entities. For example, medical symptoms detected from patient description 103 “I am having a recurring pain in the head, also feeling a bit dizzy, and my nose seems running too” may include “headache, ” “migraine, ” “faint, ” and “running nose. ” Based on the symptoms, AI system 100 may predict that the patient likely has a flu. In some embodiments, AI system 100 may use a learning model to predict the illness based on the symptoms. The learning model may be trained with sample symptoms of patients and the final diagnosis of the patients made by physicians.

Another aspect of the disclosure is directed to a non-transitory computer-readable medium storing instructions which, when executed, cause one or more processors to perform the methods, as discussed above. The computer-readable medium may include volatile or non-volatile, magnetic, semiconductor, tape, optical, removable, non-removable, or other types of computer-readable medium or computer-readable storage devices. For example, the computer-readable medium may be the storage device or the memory module having the computer instructions stored thereon, as disclosed. In some embodiments, the computer-readable medium may be a disc or a flash drive having the computer instructions stored thereon.

It will be apparent to those skilled in the art that various modifications and variations can be made to the disclosed system and related methods. Other embodiments will be apparent to those skilled in the art from consideration of the specification and practice of the disclosed system and related methods.

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

An artificial intelligence system for recognizing a medical symptom from a patient description, comprising:

a patient interaction interface configured to receive the patient description including at least one span; and

a processor configured to:

determine word vectors for words in a span and weights associated with the respective word vectors;

determine a weighted word vector based on the word vectors and the associated weights;

construct a span representation using the weighted word vector; and

determine the medical symptom based on the span representation.
The artificial intelligence system of claim 1, wherein the word vectors are determined using word embedding.
The artificial intelligence system of claim 1, wherein to determine the weights, the processor is further configured to:

determine word representations for the respective words in the span; and

determine attentions of the respective word representations as the weights associated with the respective word vectors.
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.
The artificial intelligence system of claim 4, wherein the bi-direction learning model is a Bi-LSTM network.
The artificial intelligence system of claim 1, wherein the weighted word vector is a sum of the word vectors each weighted by its associated weight.
The artificial intelligence system of claim 1, wherein the processor is further configured to apply a classification learning model to classify the span representation in a class associated with the medical symptom.
The artificial intelligence system of claim 7, wherein the classification learning model is a feedforward neural network.
The artificial intelligence system of claim 8, wherein the classification learning model is a softmax network.
The artificial intelligence system of claim 1, wherein to construct the span representation, the processor is further configured to assemble the word representation of at least one word in the span and the weighted word vector.
An artificial intelligence method for recognizing a medical symptom from a patient description, comprising:

receiving, by a patient interaction interface, the patient description including at least one span;

determining, by the processor, word vectors for words in a span and weights associated with the respective word vectors;

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

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

determining, by the processor, the medical symptom based on the span representation.
The artificial intelligence method of claim 11, wherein the word vectors are determined using word embedding.
The artificial intelligence method of claim 11, wherein determining the weights further comprises:

determining word representations for the respective words in the span; and

determining attentions of the respective word representations as the weights associated with the respective word vectors.
The artificial intelligence method of claim 13, wherein determining the word representations further comprises applying a bi-directional learning model to the respective word vectors.
The artificial intelligence method of claim 14, wherein the bi-direction learning model is a Bi-LSTM network.
The artificial intelligence method of claim 11, wherein the weighted word vector is a sum of the word vectors each weighted by its associated weight.
The artificial intelligence method of claim 11, wherein the processor is further configured to apply a feedforward neural network to classify the span representation in a class associated with the medical symptom.
The artificial intelligence method of claim 11, wherein constructing the span representation further includes assembling the word representation of at least one word in the span and the weighted word vector.
Anon-transitory computer-readable medium having instructions stored thereon that, when executed by a processor, causes the processor to perform an artificial intelligence method for recognizing a medical symptom from a patient description, the artificial intelligence methods comprising:

receiving the patient description including at least one span;

determining word vectors for words in a span and weights associated with the respective word vectors;

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

constructing a span representation using the weighted word vector; and

determining the medical symptom based on the span representation.
The non-transitory computer-readable medium of claim 19, wherein determining the weights further comprises:

applying a bi-directional learning model to the respective word vectors to determine word representations for the respective words in the span; and

determining attentions of the respective word representations as the weights associated with the respective word vectors.