EP4774720A1

EP4774720A1 - Method and system for generating clinical notes

Info

Publication number: EP4774720A1
Application number: EP24786155.2A
Authority: EP
Inventors: Nathan J. BRAKE; Thomas Schaaf; Federico Fancellu
Original assignee: Solventum Intellectual Properties Co
Current assignee: Solventum Intellectual Properties Co
Priority date: 2023-09-06
Filing date: 2024-09-05
Publication date: 2026-07-15
Also published as: WO2025052305A1

Abstract

A method for generating clinical notes includes receiving a plurality of existing clinical notes of a plurality of existing physicians and a plurality of audio records of conversations between the plurality of existing physicians and respective patients. The method further includes training a machine learning model using each of the plurality of existing clinical notes and the corresponding audio record. The method further includes receiving a prior clinical note of a new physician and selecting a representative physician from the plurality of existing physicians based on the prior clinical note and the plurality of existing clinical notes. The method further includes receiving an audio information of a new conversation between the new physician and a patient and generating, by the machine learning model, a personalized new clinical note of the new physician based on the new conversation and the one or more existing clinical notes of the representative physician.

Description

METHOD AND SYSTEM FOR GENERATING CLINICAL NOTES

Technical Field

[0001] The present disclosure generally relates to a method and a system for generating clinical notes.

Background

[0002] When documenting doctor-patient encounters, each doctor has their own style of recording a clinical note (or an electronic health record) that captures important information from the encounter. With the maturity of artificial intelligence, natural language understanding (NLU) models are designed to help automatically generate these clinical notes. Such NLU models may be trained based on training data that includes doctor-patient conversations paired with a corresponding clinical note. However, if the resulting clinical note is not in the style of the doctor or an associated scribe, the clinical note may need to be rewritten to match style expectations which requires time and effort.

[0003] NLU models are expected to generate the clinical notes in a way that is customized appropriately to each doctor and that requires minimal corrections or rework. This could be particularly challenging when a new doctor is enrolled and little information may be initially available to train the NLU model on how to generate a clinical note customized appropriately for the new doctor. In some cases, clinical notes may be initially available for the new doctors that they may have submitted in the past, but a corresponding transcript or audio record of the conversation that produced the clinical note may not be available. Thus, it may be difficult to generate a model that may be able to produce personalized clinical notes for the new doctors with only prior clinical notes available.

Summary

[0004] In a first aspect, the present disclosure provides a method for generating clinical notes. The method includes receiving a plurality of existing clinical notes of a plurality of existing physicians and a plurality of audio records of conversations between the plurality of existing physicians and respective patients. Each existing physician from the plurality of existing physicians has one or more existing clinical notes from the plurality of clinical notes. Each existing clinical note is obtained based on a corresponding audio record from the plurality of audio records. The method further includes training a machine learning model using each of the plurality of existing clinical notes and the corresponding audio record, such that the machine learning model is configured to generate a personalized clinical note for an existing physician from the plurality of existing physicians upon receiving an audio information of a conversation between the existing physician and a patient. The method further includes receiving a prior clinical note of a new physician absent in the plurality of existing physicians. The method further includes selecting a representative physician from the plurality of existing physicians based on the prior clinical note of the new physician and the plurality of existing clinical notes. The method further includes receiving audio information of a new conversation between the new physician and a patient. The method further includes generating, by the machine learning model, a personalized new clinical note for the new physician based on the new conversation and the one or more existing clinical notes of the representative physician.

[0005] In a second aspect, the present disclosure provides a system for generating clinical notes. The system includes at least one non-transitory computer-readable storage medium having instructions stored thereon. The system further includes at least one processor coupled to the at least one non-transitory computer-readable storage medium and configured to execute the instructions to receive a plurality of existing clinical notes of a plurality of existing physicians and a plurality of audio records of conversations between the plurality of existing physicians and respective patients. Each existing physician from the plurality of existing physicians has one or more existing clinical notes from the plurality of clinical notes. Each existing clinical note is obtained based on a corresponding audio record from the plurality of audio records. The processor is further configured to execute the instructions to train a machine learning model using each of the plurality of existing clinical notes and the corresponding audio record, such that the machine learning model is configured to generate a personalized clinical note for an existing physician from the plurality of existing physicians upon receiving an audio information of a conversation between the existing physician and a patient. The processor is further configured to execute the instructions to receive a prior clinical note of a new physician absent in the plurality of existing physicians. The processor is further configured to execute the instructions to select a representative physician from the plurality of existing physicians based on the prior clinical note of the new physician and the plurality of existing clinical notes. The processor is further configured to execute the instructions to receive an audio information of a new conversation between the new physician and a patient. The processor is further configured to execute the instructions to generate, by the machine learning model, a personalized new clinical note of the new physician based on the new conversation and the one or more existing clinical notes of the representative physician. [0006] The details of one or more examples of the disclosure are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the disclosure will be apparent from the description and drawings, and from the claims.

Brief Description of the Drawings

[0007] Exemplary embodiments disclosed herein may be more completely understood in consideration of the following detailed description in connection with the following figures. The figures are not necessarily drawn to scale. Like numbers used in the figures refer to like components. However, it will be understood that the use of a number to refer to a component in a given figure is not intended to limit the component in another figure labeled with the same number.

[0008] FIG. 1 is a schematic view of a system for generating clinical notes, according to an embodiment of the present disclosure;

[0009] FIG. 2 is a schematic view of a system for generating clinical notes, according to another embodiment of the present disclosure;

[0010] FIG. 3 is a schematic view of a system for generating clinical notes, according to another embodiment of the present disclosure;

[0011] FIG. 4 is a schematic view of a system for generating clinical notes, according to another embodiment of the present disclosure;

[0012] FIG. 5 is a schematic view of the system of FIG. 4, according to another embodiment of the present disclosure;

[0013] FIG. 6 is a schematic view of a system for generating clinical notes, according to another embodiment of the present disclosure; and

[0014] FIG. 7 is a flowchart illustrating a method for generating clinical notes, according to an embodiment of the present disclosure.

[0015] FIG. 8 is a flowchart illustrating a method for generating clinical notes, according to an embodiment of the present disclosure.

Detailed Description

[0016] In the following description, reference is made to the accompanying figures that form a part thereof and in which various embodiments are shown by way of illustration. It is to be understood that other embodiments are contemplated and may be made without departing from the scope or spirit of the present disclosure. The following detailed description, therefore, is not to be taken in a limiting sense.

[0017] In the following disclosure, the following definitions are adopted. [0018] As used herein, the term “patient”, and its equivalents, may refer to an individual being monitored and/or cared for within a clinical environment or who has been previously monitored and/or cared for within the clinical environment. In various examples, a patient is a human, but implementations of this disclosure are not so limited. Clinical environment may include, but are not limited to, a doctor's office, a medical facility, a medical practice, a medical lab, an urgent care facility, a medical clinic, an emergency room, an operating room, a hospital, a long term care facility, a rehabilitation facility, a nursing home, and a hospice facility.

[0019] As used herein, the term “clinical note” may refer to medical data or a medical record generated by a provider for an individual person. The clinical note may be in the form of a text document. The clinical note may also be referred to as an electronic health record (EHR).

[0020] As used herein, the term “medical data” may refer to data describing an individual person's medical history or medical condition, including lab test results, medication history, immunization history, and so forth. The term may also include a personal identifier of a patient. [0021] As used herein, the term “provider” may refer to, for example, a physician (including, but not limited to, a radiologist, a surgeon, a primary care physician, and a medical specialist), a physician assistant, a nursing professional, a medical laboratory technician, medical clinics, hospitals, health insurance providers, diagnostic sites, imaging sites, pharmacies, and the like. The term “provider” may also refer to an academic institution, a government research laboratory, a non-profit entity, or a for-profit entity, such as a pharmaceutical, health insurance, biotechnology, wearable device, physiological monitoring, or medical device company.

[0022] As used herein, the term “physician” may refer to a health care provider or a medical professional, such as a doctor, a nurse, or other appropriate clinician.

[0023] As used herein, the term “audio record” may refer to audio data stored in a digital format.

[0024] As used herein, the term “machine-learning model” may refer to a computer model or a computer representation that may be tuned (e.g., trained) based on inputs to approximate unknown functions. For example, the machine-learning model may include one or more of vectorization machine-learning models, sequence-to-sequence models, transformer models, a decision tree (e.g., a gradient boosted decision tree), a linear regression model, a logistic regression model, association rule learning, inductive logic programming, support vector learning, a Bayesian network, a regression-based model, a neural network, or a combination thereof. The process of building or optimizing a machine learning model is referred to herein as “training”.

[0025] As used herein, the term “neural network” may refer to one example of a machine learning model that can be tuned (e.g., trained) based on inputs to approximate unknown functions. In particular, the neural network may include a model of interconnected neurons (arranged in layers) that communicate and learn to approximate complex functions and generate outputs based on a plurality of inputs provided to the model. For example, the neural network may include deep convolutional neural networks (CNN), Region-CNN (R-CNN), Faster R-CNN, Mask R-CNN, fully convolutional neural networks, recurrent neural networks (“RNNs”), such as long short-term memory neural networks (“LSTMs”), graph neural networks, generative adversarial neural networks (GAN), and single-shot detect (SSD) networks. In other words, a neural network is an algorithm that implements deep learning techniques, which utilize a set of learned parameters arranged in layers according to a particular architecture to attempt to model high-level abstractions in data using supervisory data to tune parameters of the neural network. [0026] As used herein, the term “Wi-Fi” refers generally to a bi-directional radio communication technology that operates based on one or more of the ‘Institute of Electrical and Electronics Engineers’ (“IEEE”) 802. 11 family of standards, which are incorporated herein by reference. The IEEE 802.11 standards specify the radio frequency (RF) and protocol characteristics of a bi-directional radio communication system.

[0027] As used herein, the term “coupled” generally means either a direct connection between two or more elements that are connected or an indirect connection through one or more passive or active intermediary devices.

[0028] As used herein, the term “communicably coupled” generally refers to any type of connection or coupling that allows for communication of information. The term communicably coupled may include, but is not limited to, electrically coupled (e.g., through a wire), optically coupled (e.g., through an optical cable), wirelessly coupled (e.g., through a radio frequency or other similar technologies), and/or the like. The technology by which the information is transmitted is not material to the meaning of communicably coupled.

[0029] As used herein, all numbers should be considered modified by the term “about”. As used herein, “a,” “an,” “the,” “at least one,” and “one or more” are used interchangeably.

[0030] The term “about”, unless otherwise specifically defined, means to a high degree of approximation (e.g., within +/- 5% for quantifiable properties) but again without requiring absolute precision or a perfect match.

[0031] As used herein as a modifier to a property or attribute, the term “generally”, unless otherwise specifically defined, means that the property or attribute would be readily recognizable by a person of ordinary skill but without requiring absolute precision or a perfect match (e.g., within +/- 20 % for quantifiable properties). [0032] As used herein, the term “configured to” and like is at least as restrictive as the term “adapted to” and requires actual design intention to perform the specified function rather than mere physical capability of performing such a function.

[0033] Conventionally, natural language understanding (NLU) models are designed to help automatically generate clinical notes (or electronic health record) from doctor-patient conversations. However, if the resulting clinical note is not in a style of the doctor or an associated scribe, the clinical note may need to be rewritten to match the expectations which requires time and effort. NLU models may be trained based on training data that includes doctor-patient conversations paired with a corresponding clinical note. It could be particularly challenging when a new doctor is enrolled and little information may be initially available to train the NLU model on how to generate a clinical note customized appropriately for the new doctor. In some cases, clinical notes may be initially available for the new doctors that they may have submitted in the past, but a corresponding transcript or audio record of the conversation that produced the clinical note may not be available. Thus, it may be difficult to generate a model that may be able to produce personalized clinical notes for new doctors with only prior clinical notes available.

[0034] The present disclosure provides a method for generating clinical notes. The method includes receiving a plurality of existing clinical notes of a plurality of existing physicians and a plurality of audio records of conversations between the plurality of existing physicians and respective patients. Each existing physician from the plurality of existing physicians has one or more existing clinical notes from the plurality of clinical notes. Each existing clinical note is obtained based on a corresponding audio record from the plurality of audio records. The method further includes training a machine learning model using each of the plurality of existing clinical notes and the corresponding audio record, such that the machine learning model is configured to generate a personalized clinical note for an existing physician from the plurality of existing physicians upon receiving an audio information of a conversation between the existing physician and a patient. The method further includes receiving a prior clinical note of a new physician absent in the plurality of existing physicians. The method further includes selecting a representative physician from the plurality of existing physicians based on the prior clinical note of the new physician and the plurality of existing clinical notes. The method further includes receiving an audio information of a new conversation between the new physician and a patient. The method further includes generating, by the machine learning model, a personalized new clinical note of the new physician based on the new conversation and the one or more existing clinical notes of the representative physician. [0035] The method of the present disclosure includes receiving the prior clinical note of the new physician and selecting the representative physician from the plurality of existing physicians based on the prior clinical note of the new physician and the plurality of existing clinical notes. This may allow the machine learning model to generate the personalized new clinical note of the new physician based on the one or more existing clinical notes of the representative physician upon receiving the audio information of the new conversation between the new physician and a patient. Thus, the method may allow generation of the personalized new clinical note of the new physician immediately upon enrollment. Further, the personalized new clinical note of the new physician may require minimal manual post editing after generation to match the style of the new physician.

[0036] The machine learning model may be updated as audio information for more new conversations becomes available for the new physician paired with the corresponding personalized new clinical notes, thereby improving accuracy. Since the machine leaning model is able to generate the personalized new clinical note of the new physician immediately upon enrolment, the training required by the machine learning model to accurately generate the personalized new clinical note may be less as compared to conventional methods where a machine learning model is trained only when training data, i.e., doctor-patient conversations paired with a corresponding clinical note, for the new physician becomes available.

[0037] For the above reasons, the methods and techniques of the present disclosure are specifically adapted to improve computerized systems that are designed to address the technical problem of automatically adjusting a machine learning model for the purpose of clinical note personalization. That is, the present disclosure is inextricably linked to the underlying computer technology for training machine learning models. As such, it should be appreciated that the present disclosure cannot be performed by a human, with or without the aid of pen and paper. For instance, training a machine learning model as described in the present disclosure cannot be performed using pen and paper and must be performed by the computing system on which the model to be trained resides. Specifically, it would be impracticable for a human to manually perform gradient descent on hundreds of thousands, let alone, tens of millions of matrix weights in a timely manner to train the machine learning model as disclosed herein. Likewise, generating clinical notes using the trade machine learning model cannot be performed by a human with or without the aid of pen and paper for similar reasons.

[0038] FIG. 1 is a schematic view of a system 100 for generating clinical notes. In some embodiments, the system 100 is configured to automate collection and processing of clinical encounter information to generate/store/distribute the clinical notes. The system 100 includes at least one non-transitory computer-readable storage medium 102 having instructions stored thereon. The system 100 further includes at least one processor 104 coupled to the at least one non-transitory computer-readable storage medium 102. The term “at least one non-transitory computer-readable storage medium 102” is interchangeably referred to herein as “the storage medium 102”. The term “at least one processor 104” is interchangeably referred to herein as “the processor 104”.

[0039] In some embodiments, the storage medium 102 may include any type of computer readable storage media, including, but not be limited to, various types of volatile and nonvolatile storage media, including, but not limited to, random access memory, read-only memory, programmable read-only memory, electrically programmable read-only memory, electrically erasable read-only memory, flash memory, magnetic tape or disk, optical media, and the like. In some cases, the storage medium 102 may include a cache or random access memory for the processor 104. Alternatively, or in addition, the storage medium 102 may be separate from the processor 104, such as a cache memory, a system memory, or other memory. In some embodiments, the storage medium 102 may be an external storage device or a database for storing data. Examples may include a hard drive, compact disc (“CD”), digital video disc (“DVD”), memory card, memory stick, floppy disc, universal serial bus (“USB”) memory device, or any other device operative to store data.

[0040] In some embodiments, the processor 104 may be embodied in a number of different ways. For example, the processor 104 may be embodied as various processing means, such as one or more of a microprocessor or other processing elements, a coprocessor, or various other computing or processing devices including integrated circuits, such as, e.g., an ASIC (application specific integrated circuit), an FPGA (field programmable gate array), or the like. As such, whether configured by hardware or by a combination of hardware and software, the processor 104 may represent an entity (e.g., physically embodied in circuitry - in the form of processing circuitry) capable of performing operations according to some embodiments while configured accordingly. Thus, for example, when the processor 104 is embodied as an executor of software instructions, the instructions may specifically configure the processor 104 to perform the operations described herein. Alternatively, as another example, when the processor 104 is embodied as an ASIC, FPGA, or the like, the processor 104 may have specifically configured hardware for conducting the operations described herein.

[0041] In some embodiments, the system 100 may be implemented as a server-side application, a client-side application, or a hybrid server-side/client-side application, and may be connected to a network (e.g., the Internet or a local area network). In some embodiments, the system 100 may include various components, examples of which may include, but are not limited to, a personal computer, a server computer, a series of server computers, a mini computer, a mainframe computer, one or more Network Attached Storage (NAS) systems, one or more Storage Area Network (SAN) systems, one or more Platform as a Service (PaaS) systems, one or more Infrastructure as a Service (laaS) systems, one or more Software as a Service (SaaS) systems, a cloud-based computational system, and a cloud-based storage platform.

[0042] The at least one processor 104 is configured to execute the instructions to receive a plurality of existing clinical notes 112 of a plurality of existing physicians 110-1, 110-2, . . . , 110- N (collectively, physicians 110) and a plurality of audio records 114 of conversations between the plurality of existing physicians 110 and respective patients 108, where N is a positive integer corresponding to a total number of the existing physicians 110 (e.g., N = 2, 5, etc.). Each existing physician 110 from the plurality of existing physicians 110 has one or more existing clinical notes 112 from the plurality of clinical notes 112. The term "plurality of clinical notes 112" is interchangeably referred to herein as “the clinical notes 112”.

[0043] Each existing clinical note 112 is obtained based on a corresponding audio record 114 from the plurality of audio records 114. In some embodiments, each existing clinical note 112 may be associated with the corresponding audio record 114 representative of the conversation between the corresponding existing physician 110 and the respective patient 108. For example, each existing clinical note 112 may be obtained after processing information obtained from the conversation between the corresponding existing physician 110 and the respective patient 108. In some embodiments, the conversation between the existing physician 110 and the respective patient 108 may be stored in the form of the audio records 114. The term "plurality of audio records 114" is interchangeably referred to herein as “the audio records 114”.

[0044] In some embodiments, each existing clinical note 112 may represent a clinical note written in a style of the corresponding existing physician 110. For example, each existing clinical note 112 may be obtained after processing the conversation between the corresponding existing physician 110 and the respective patient 108 in the style or personal preference of the corresponding existing physician 110. In some embodiments, the style or personal preferences of the plurality of existing physicians 110 may include personal habits of recording medical information, e.g., presence of bullet points, vocabulary, number of sentences, etc.

[0045] In some examples, each existing clinical note 112 may be obtained by correcting or updating an intermediate clinical note or a transcript of the conversation between the corresponding existing physician 110 and the respective patient 108 automatically generated via natural language understanding models. Alternatively, the existing clinical note 112 may be manually recorded by the existing physician 110. In some embodiments, the plurality of existing physicians 110 may have a database of the one or more existing clinical notes 112 that are recorded based on the style and personal preferences of the corresponding existing physician 110. Such a database may be generated over time based on conversations between the existing physicians 110 and their respective prior patients 108.

[0046] The at least one processor 104 is further configured to execute the instructions to train a machine learning model 106 using each of the plurality of existing clinical notes 112 and the corresponding audio record 114, such that the machine learning model 106 is configured to generate a personalized clinical note PC for an existing physician 110 from the plurality of existing physicians 110 upon receiving an audio information Al of a conversation between the existing physician 110 and a patient 109. A clinical note generated in the style or personal preference of a physician is referred to herein as a “personalized clinical note”. In some embodiments, the machine learning model 106 may be trained to generate any section of a clinical note, e.g., history of present illness, medication and dosage history, social history, family history, etc.

[0047] In some embodiments, the machine learning model 106 may be implemented in a local facility within the clinical environment, such as in a general purpose computer, or a set of computers on a local area network. In some embodiments, the machine learning model 106 may be communicably coupled to the processor 104 by way of one or more wired and/or wireless communication interfaces. In some examples, the wireless communication interface may communicate data via one or more wireless communication protocols, such as Bluetooth©, infrared, Wi-Fi, Zigbee, wireless universal serial bus (USB), radio frequency, near-field communication (NFC), RFID protocols, IEEE 802.11a, 802.11b, 802.11g, 802.1 In, or generally any wireless communication protocol.

[0048] In some embodiments, the machine learning model 106 is trained based on each of the preexisting clinical notes (i.e., the existing clinical notes 112) written in the style and personal preferences of the corresponding existing physician 110 and the corresponding audio record 114. Thus, each of the plurality of existing clinical notes 112 paired with the corresponding audio record 114 may be utilized as training data for generating the machine learning model 106. Availability of a large amount of the training data means that the machine learning model 106 may reliably represent the style and personal preferences of the existing physician 110. For example, the machine leaning model 106 may be trained based on multiple instances of such existing clinical notes 112 and the corresponding audio records 114, such that the machine learning model 106 may represent the style and personal preferences of each existing physician 110.

[0049] After being trained, the machine learning model 106 is able to generate the personalized clinical note PC for the existing physician 110 upon receiving the audio information Al of the conversation between the existing physician 110 and the patient 109. In some embodiments, the audio information Al may be captured by an audio input device (not shown), e.g., a handheld microphone, a lapel microphone, an embedded microphone, such as those embedded within eyeglasses, smart phones, tablet computers and/or watches, or an audio recording device.

[0050] In some embodiments, the audio input device may include a plurality of discrete audio devices disposed within a clinical environment for capturing physician-patient encounter. Such audio input devices may be communicably coupled (wired or wirelessly) to the processor 104. The processor 104 may transmit the audio information Al to the machine learning model 106 for generating the personalized clinical note PC. In some embodiments, the machine learning model 106 may first generate a transcript of the audio information Al and then extract words or phrases in the transcript for generating the personalized clinical note PC.

[0051] In some embodiments, the machine learning model 106 may be a combination of discrete models. For example, the machine learning model 106 may include a model which separates the audio information Al into speech by different speakers, in this case the existing physician 110 and the patient 109. In some embodiments, the machine learning model 106 may further include a model that implements traditional speech to text functionality. Such a model may be trained using supervised learning techniques and labeled training speech data to recognize medical-related terminology in speech, including medical terms such as symptoms, medications, human anatomical terms, etc.

[0052] In some embodiments, the machine learning model 106 may further include a model (e.g., a named-entity recognition model) which processes the text generated using the speech to text functionality to recognize medically relevant words or phrases. Such models are well known in the field of machine learning and are described extensively in scientific literature. Application of this model to the text generated using the speech to text functionality may result in a highlighted transcript of the audio information Al with relevant words or phrases highlighted as recognized by the named entity recognition model as well as extraction of such highlighted words or text as data for generation of the personalized clinical note PC. The highlighted words or phrases may be classified for populating different regions or fields of the personalized clinical note PC, e.g., history of patient illness, medication and dosage history, social history, family history, etc.

[0053] In some embodiments, the machine learning model 106 may include another model (e.g., a deep learning model such as a sequence-to-sequence model) that converts the text to a draft clinical note. In some embodiments, various attribution techniques may be employed by the machine learning model 106 that may effectively softly assign responsibility for a given output (e.g., the draft clinical note) word to input (e.g., conversation transcript text) words according to the machine learning model 106, e.g., attention weights, vocabulary embedding weights, integrated gradient, etc. As a result, this may provide a soft mapping from transcript word positions to clinical note word positions. In some embodiments, the transcript word position assigned maximal attribution for a given output word may be interpreted as being aligned (linked) to that output. A word in the draft medical report, aligned to a word in the conversation transcript text, may now be associated with an audio time interval of the audio information Al of the encounter. It should be understood that alternative approaches may also be utilized for generating the draft medical report.

[0054] It may also be possible that the machine learning model 106 may process the audio information Al to directly generate the draft medical report and also that the machine learning model 106 may have access to other relevant info sources. In some embodiments, the machine learning model 106 may directly generate the draft medical report (i.e., the personalized clinical note PC) in the style and personal preferences of the corresponding existing physician 110 or the draft medical report may be updated based on the style and personal preferences of the corresponding existing physician 110 to obtain the personalized clinical note PC.

[0055] In some embodiments, the machine learning model 106 may include a plurality of sub-models corresponding to the plurality of existing physicians 110. Each sub-model may be trained for a corresponding existing physician 110 from the plurality of existing physician 110 to generate the personalized clinical note PC for the corresponding existing physician 110. In other words, a dedicated sub-model may be trained for each existing physician 110. This approach may leverage existing state-of-the-art generation models, e.g., based on neural transformers, or neural sequence-to-sequence architectures. Alternatively, in some embodiments, a pre-trained model may be selected and only some parts of the pre-trained model may be adopted for each existing physician 110 (e.g., a shared encoder, or shared bottom-layers of a decoder in a transformer architecture), allowing to reduce an amount of the training data required.

[0056] In some embodiments, the machine learning model 106 may be trained for each existing physician 110 where the training data is available. Training data here refers to the audio record 114 (representing the conversation between the existing physician and the respective patients 108) paired with a corresponding existing clinical note 112, written in the style of the corresponding existing physician 110. However, training of the machine learning model 106 may be limited in case a new physician 120 is enrolled into the system 100 whose training data may not be available during enrollment. In other words, the new physician 120 may be absent in the plurality of existing physicians 110. Therefore, the machine learning model 106 is not trained for generating personalized clinical notes for the new physician 120. Initially, the machine leaning model 106 may have very limited or no training data available for the new physician 120 since audio records for the new physician 120 may not be available. However, one or more prior clinical notes 122 written in the style and personal preferences of the new physician 120 may be available that the new physician 120 may have submitted in the past. The term “one or more prior clinical notes 122” is interchangeably referred to herein as “the prior clinical note 122”.

[0057] The at least one processor 104 is further configured to execute the instructions to receive the prior clinical note 122 of the new physician 120. The at least one processor 104 is further configured to execute the instructions to select a representative physician RP from the plurality of existing physicians 110 based on the prior clinical note 122 of the new physician 120 and the plurality of existing clinical notes 112. In some embodiments, the representative physician RP may be selected by analyzing similarity between the prior clinical note 122 of the new physician 120 and each existing clinical note 112 of the corresponding existing physicians 110. In other words, the representative physician RP may be selected from the plurality of existing physicians 110 having the one or more existing clinical notes 112 most similar (in style) to that of the prior clinical note 122 of the new physician 120.

[0058] In some embodiments, the similarity may be analyzed by using a statistical language model, e.g., N-gram, generative pre-trained transformers (GPT), or similar. For example, the statistical language model may compare the prior clinical note 122 of the new physician 120 with each existing clinical note 112 of the plurality of existing physicians 110 by computing a likelihood function. Alternatively, other suitable models, such as rule-based models, machine- learned regressor models, machine-learned classifier models, or the like may also be used. Further, the similarity may be analyzed based on stylistic features of the prior clinical note 122 and the one or more existing clinical notes 112.

[0059] The at least one processor 104 is further configured to execute the instructions to receive an audio information A2 of a new conversation NC between the new physician 120 and a patient 124. In some embodiments, the audio information A2 may be captured by the audio input device described above. The new conversation NC may be representative of an encounter between the new physician 120 and the patient 124 after the new physician 120 is enrolled into the system 100 and the machine learning model 106 is untrained on the style and personal preferences of the new physician 120.

[0060] The at least one processor 104 is further configured to execute the instructions to generate, by the machine learning model 106, a personalized new clinical note PN of the new physician 120 based on the new conversation NC and the one or more existing clinical notes 112 of the representative physician RP. In some embodiments, the machine learning model 106 may use the style and personal preferences of the representative physician RP (through the corresponding the one or more existing clinical notes 112) while processing the new conversation NC to generate the personalized new clinical note PN. Thus, the machine learning model 106 may be able to generate the personalized new clinical note PN with the style and personal preferences of the new physician 120 without needing prior training data for the new physician 120. This may save time and effort of the new physician 120 or associated medical staff to generate the personalized new clinical note PN of the new physician 120.

[0061] In some embodiments, the at least one processor 104 is further configured to execute the instructions to receive changes 128 on the personalized new clinical note PN from the new physician 120. In some embodiments, the personalized new clinical note PN may be reviewed by the new physician 120 (or another medical professional involved with the patient encounter) to determine an accuracy and correctness of the personalized new clinical note PN and/or make corrections to the personalized new clinical note PN. Such corrections may be provided as the changes 128 to the at least one processor 104.

[0062] In some embodiments, the at least one processor 104 may display the personalized new clinical note PN generated by the machine learning model 106 via a display means (not shown) accessible to the new physician 120. Further, the display means may allow the new physician 120 to correct or make changes to the personalized new clinical note PN generated by the machine learning model 106 based on the style and personal preferences of the new physician 120. In some embodiments, the at least one processor 104 is further configured to execute the instructions to update the machine learning model 106 based on the changes 128. For example, the machine learning model 106 may consider the changes 128 made by the new physician 120, e.g., through error backpropagation.

[0063] In some embodiments, the at least one processor 104 is further configured to execute the instructions to generate a similarity score 129 based on a comparison between the personalized new clinical note PN and the prior clinical note 122 of the new physician 120. In some embodiments, the similarity score 129 may be generated based on a comparison of the stylistic features of the personalized new clinical note PN and the prior clinical note 122. In some embodiments, the similarity score 129 may rely on encoding the personalized new clinical note PN and the prior clinical note 122 in a continuous space, using document and/or word embedding methods. The encoding may be based on stylistic features.

[0064] In some embodiments, the stylistic features may include parameters such as a number of sentences (to capture verbosity), reading scores (to capture complexity of language use), and term frequency scores (e.g., term frequency - inverse document frequency or TF-IDF). Other stylistic features that are more fine-grained may also be used, e.g., a presence of bullet points. In some embodiments, the at least one processor 104 is further configured to execute the instructions to update the machine learning model 106 based on the similarity score 129. For example, the representative physician RP for the new physician 120 may be updated based on the similarity score 129.

[0065] FIG. 2 is a schematic view of a system 200 for generating clinical notes, according to another embodiment of the present disclosure. The system 200 is substantially similar and functionally equivalent to the system 100 shown in FIG. 1, and same components in this embodiment are referred to by same reference numerals and differences between the embodiments are discussed.

[0066] In some embodiments, training the machine learning model 106 further includes determining one or more attributes 216 of each existing physician 110 based on the one or more existing clinical notes 112 of the corresponding existing physician 110. In some embodiments, the one or more attributes 216 may be determined based on stylistic features of the one or more existing clinical notes 112 of the corresponding existing physician 110. For example, the one or more existing clinical notes 112 may be analyzed to obtain the stylistic features of the corresponding existing physician 110.

[0067] In some embodiments, training the machine learning model 106 further includes generating a plurality of physician tokens 218. Each physician token 218 from the plurality of physician tokens 218 is generated based on the one or more attributes 216 of the corresponding existing physician 110. In some embodiments, each physician token 218 may provide the stylistic features of the corresponding existing physician 110 in the form of a vector representation. In some embodiments, training the machine learning model 106 further includes embedding the plurality of physician tokens 218 in the machine learning model 106. For example, the plurality of physician tokens 218 may be mapped to an array of vectors and passed into the machine learning model 106 (e.g., a neural network) as an embedding 230.

[0068] In some embodiments, the embedding 230 may be a dense representation of the one or more attributes 216 of each existing physician 110. This may allow the machine learning model 106 to leverage potential similarities between the one or more attributes 216 (or clinical styles/preferences) of each existing physician 110 and gain advantage from the common styles and overlapping preferences of the existing physicians 110. In other words, this may allow effective parameter sharing and generation of a single model, i.e., the machine learning model 106, that may learn the style and personal preferences of different existing physicians 110.

[0069] In case of a transformer (or encoder-decoder) architecture, the embedding 230 may have a same dimensionality as that of the vocabulary used, or the dimension could be different if a projection or a feed-forward-network is used. In alternative embodiments, the embedding 230 (i.e., the physician tokens 218) for the plurality of existing physicians 110 may be randomly initialized and updated during training of the machine learning model 106. In some embodiments, additional embeddings representing hospitals, hospital systems, clinical note systems, scribing organizations, and scribes may also be utilized to model additional factors that may be considered stylistic or personalized.

[0070] In some embodiments, training the machine learning model 106 further includes forming a plurality of physician clusters 232-1, 232-2, . . . , 232-M (collectively, physician clusters 232) of the plurality of existing physicians 110 based on the plurality of physician tokens 218, where M is a positive integer corresponding to a total number of the physician clusters 232 (e.g., M = 2, 5, etc.). Each physician cluster 232 from the plurality of physician clusters 232 includes one or more physician tokens 218 from the plurality of physician tokens 218 and a centroid 234 of the one or more physician tokens 218.

[0071] In some embodiments, during training, the existing physicians 110 with similar styles and personal preferences, i.e., similar physician tokens 218, will move closer to each other in the embedding space (also called a dimensional space), thereby forming the plurality of physician clusters 232. Each physician cluster 232 from the plurality of physician clusters 232 may represent similar physician tokens 218, i.e., existing physicians 110 with similar one or more attributes 216.

[0072] In some embodiments, each physician clusters 232 may be represented by the centroid 234. In some embodiments, the centroid 234 may be a central point in the embedding space for the corresponding physician cluster 232. In some embodiments, the plurality of physician clusters 232 may be generated using, e.g., k-means, or another appropriate algorithm. Clustering methods use similarity formulas to assess similarity between points in an embedding space. It should be understood that other techniques for clustering may also be employed, such as density -based clustering, hierarchical clustering, connectivity-based clustering, etc.

[0073] In some embodiments, selecting the representative physician RP further includes determining one or more attributes 226 of the new physician 120 based on the prior clinical note 122. In some embodiments, the one or more attributes 226 may be determined in a similar manner as the one or more attributes 216. In some embodiments, selecting the representative physician RP further includes selecting the representative physician RP based on a comparison between the one or more attributes 226 of the new physician 120 and the centroid 234 of each physician cluster 232.

[0074] In some embodiments, the at least one processor 104 may determine the selected centroid 234 by evaluating language models trained on the one or more existing clinical notes 112 of the corresponding existing physicians 110 of each physician cluster 232. Such language models may compare the one or more attributes 226 of the new physician 120 with the centroid 234 of each physician cluster 234 and the centroid 234 with the highest language model likelihood is chosen. Subsequently, the selected centroid 234 may be used to initialize embedding for the new physician 120 within the machine learning model 106. This allows the new physician 120 not seen during the training of the machine leaning model 106 to be mapped to one of the physician clusters 232. In some embodiments, in cases where the at least one processor 104 may find multiple results (i.e., centroids 234) with similar probabilities, a (weighted) interpolation of the centroids 234 (e.g., from top 2 or 3 physician clusters 232) may be obtained.

[0075] As soon as the new physician 120 is in production, the machine learning model 106 may generate the personalized new clinical note PN for the new physician 120 that may be paired with new audio records generated from the audio information A2. This may be used to update the embedding for the new physician 120 that was initially assigned based on the representative physician RP, until there is enough training data available to apply a model technique that may utilize larger amount of training data for the new physician 120. An advantage of clustering is that the number of physician clusters 232 will be much smaller than the number of existing physicians 110 in the training data of the machine learning model 106. This may speed up the process of finding an embedding to initialize for the new physician 120. [0076] FIG. 3 is a schematic view of a system 300 for generating clinical notes, according to another embodiment of the present disclosure. The system 300 is substantially similar and functionally equivalent to the system 100 shown in FIG. 1, and same components in this embodiment are referred to by same reference numerals and differences between the embodiments are discussed.

[0077] In some embodiments, training the machine learning model 106 further includes determining a plurality of unique identifiers 316 of the plurality of existing physicians 110. Each unique identifier 316 from the plurality of unique identifiers 316 is representative of an existing physician 110 from the plurality of existing physicians 110. For example, each unique identifier 316 may include a textual representation of a physician identification (physician ID) of the corresponding existing physician 110.

[0078] In some embodiments, training the machine learning model 106 further includes encoding the plurality unique identifiers 316 using a one-hot encoding scheme to generate an encoded matrix 318. One-hot encoding scheme is a process by which categorical variables (the existing physicians 110 in this case) are converted into a format in which a machine learning model is better able to use the categorical variables for training and/or prediction. For example, one-hot encoding scheme may be used to represent categorical variables as binary vectors so that categorized inputs may be included as a feature for a machine learning model. As one example, each received input may be categorized by indicating a “1” if the category is met or a “0” if a category is not met.

The encoded matrix 318 may be represented as an array of vectors in which all elements of the vectors are “0” except one, which has “ 1” as its value. For example, [0 00 1 00] may be a one-hot encoded vector in which the fourth category matches, e.g., with the unique identifier 316 of one of the existing physicians 110 in a group of six unique identifiers 316 (or six existing physicians 110). In some embodiments, the encoded matrix 318 may represent an array of all the one-hot encoded vectors with each column representing the corresponding unique identifier 316 of the corresponding existing physician 110. Generally, one-hot encoding scheme may result in a high-dimensional sparse matrix.

[0079] In some embodiments, training the machine learning model 106 further includes feeding the encoded matrix 318 to the machine learning model 106. In case of a transformer (or encoder-decoder) architecture, the encoded matrix 318 may be accessed by the encoder (e.g., through a projection or a feed-forward-network). In some embodiments, the encoded matrix 318 provides the unique identifiers 316 of each existing physician 110 as a side -information. In some other embodiments, the encoded matrix 318 may be fed to the decoder instead of the encoder.

[0080] In some embodiments, selecting the representative physician RP further includes selecting the representative physician RP based on a comparison between the prior clinical note 122 of the new physician 120 and the one or more existing clinical notes 112 of each existing physician 110 whose unique identifier 316 is encoded in the encoded matrix 318. In some embodiments, the at least one processor 104 may analyze similarity between the prior clinical note 122 of the new physician 120 and the one or more existing clinical notes 112 of each existing physician 110 by using a statistical language model, e.g., N-gram, generative pre-trained transformers (GPT), or similar. For example, the statistical language model may compare the prior clinical note 122 of the new physician 120 with the one or more existing clinical notes 112 of each existing physician 110 by computing a likelihood function.

[0081] After more personalized new clinical notes PN and the corresponding audio records for the new physician 120 become available, the assignment of the representative physician RP from the plurality of existing physicians 110 may be evaluated by computing a likelihood of the personalized new clinical notes PN, and the unique identifier 316 of the existing physician 110 corresponding to the best likelihood may be used. This decision may be made by a weighting of the likelihoods derived from the statistical language models and clinical note generation models within the machine learning model 106.

[0082] FIG. 4 is a schematic view of a system 400 for generating clinical notes, according to another embodiment of the present disclosure. The system 400 is substantially similar and functionally equivalent to the system 100 shown in FIG. 1, and same components in this embodiment are referred to by same reference numerals and differences between the embodiments are discussed.

[0083] In some embodiments, the machine learning model 106 includes an agnostic model 440 untrained by the plurality of existing clinical notes 112 and the plurality of audio records 114. In some embodiments, the machine learning model 106 further includes a personalized model 442 trained by each of the plurality of existing clinical notes 112 and the corresponding audio record 114. In some embodiments, the at least one processor is further configured to execute the instructions to generate, by the agnostic model 440, an agnostic clinical note AC based on the audio information Al of the conversation between the existing physician 110 and the patient 109. In some embodiments, the at least one processor 104 is further configured to execute the instructions to modify, by the personalized model 442, the agnostic clinical note AC to generate the personalized clinical note PC of the existing physician 110.

[0084] In some embodiments, the agnostic model 440 may be any preexisting model suitable for generating the agnostic clinical notes AC not necessarily specific to the style and personal preferences of the corresponding existing physician 110. In some embodiments, the agnostic model 440 may leverage existing natural language understanding techniques for generating the agnostic clinical note AC. In some embodiments, the agnostic model 440 may either be trained or provided by an outside / third party provider that does not allow direct access to the personalized model 442.

[0085] In some embodiments, the personalized model 442 may be trained (e.g., using each of the plurality of existing clinical notes 112 and corresponding audio record 114) on the style and personal preferences of each existing physician 110. The personalized model 442 may re-write the agnostic clinical note AC for generating the personalized clinical note PC of corresponding existing physician 110. This technique may require less training data to train the personalized model 442 as would be required to generate a model that receives doctor-patient conversations and processes it into personalized clinical notes. Further, this technique has the advantage of utilizing a large physician-agnostic model (i.e., agnostic model 440) that pools together all training data (e.g., medical vocabulary, etc.) that is generally available, while requiring only physician-wise training of the personalized model 442.

[0086] For the personalized model 442, learning to re-write the agnostic clinical note AC into the personalized clinical note PC may require a mapping between the agnostic clinical note AC and the personalized clinical note PC. This may be generated by using the existing clinical notes 112 of the existing physicians 110, or from having a scribe or the existing physician 110 post-edit the agnostic clinical note AC into a personalized clinical note of the corresponding existing physician 110.

[0087] FIG. 5 is a schematic view of the system 400, according to another embodiment of the present disclosure. In some embodiments, training the machine learning model 106 further includes determining one or more attributes 416 of each existing physician 110 based on the one or more existing clinical notes 112 of the corresponding existing physician 110. In some embodiments, the one or more attributes 416 may be determined based on stylistic features of the one or more existing clinical notes 112 of the corresponding existing physician 110.

[0088] In some embodiments, training the machine learning model 106 further includes generating a plurality of physician tokens 418. Each physician token 418 from the plurality of physician tokens 418 is generated based on the one or more attributes 416 of the corresponding existing physician 110. In some embodiments, each physician token 418 may provide the stylistic features of the corresponding existing physician 110 in the form of a vector representation. In some embodiments, training the machine learning model 106 further includes embedding the plurality of physician tokens 418 in the personalized model 442 of the machine learning model 106. For example, the plurality of physician tokens 418 may be mapped to an array of vectors and passed into the personalized model 442 of the machine learning model 106 (e.g., a neural network) as an embedding 430.

[0089] In some embodiments, training the machine learning model 106 further includes forming a plurality of physician clusters 432-1, 432-2, . . . , 432-P (collectively, physician clusters 432) of the plurality of existing physicians 110 based on the plurality of physician tokens 418, where P is a positive integer corresponding to a total number of the physician clusters 432 (e.g., P = 2, 5, etc.). Each physician cluster 432 from the plurality of physician clusters 432 includes one or more physician tokens 418 from the plurality of physician tokens 418 and a centroid 434 of the one or more physician tokens 418. In some embodiments, the plurality of physician clusters 432 may be generated using, e.g., k-means, or another appropriate algorithm.

[0090] In some embodiments, selecting the representative physician RP further includes determining one or more attributes 426 of the new physician 120 based on the prior clinical note 122. In some embodiments, the one or more attributes 426 may be determined in a similar manner as the one or more attributes 416. In some embodiments, selecting the representative physician RP further includes selecting the representative physician RP based on a comparison between the one or more attributes 426 of the new physician 120 and the centroid 434 of each physician cluster 432. Subsequently, the selected centroid 434 may be used to initialize embedding for the new physician 120 within the personalized model 442 of the machine learning model 106. [0091] As soon as the new physician 120 is in production, the agnostic model 440 of the machine learning model 106 may generate the agnostic clinical note AC based on the audio information A2 of the new conversation between the new physician 120 and the patient 124. Subsequently, the personalized model 442 of the machine learning model 106 may modify or rewrite the agnostic clinical note AC to generate the personalized new clinical note PN of the new physician 120 based on the initialized embedding for the new physician 120. If an initial small amount of training data for the new physician 120 becomes available, i.e., pairs of the agnostic clinical note AC and the personalized new clinical note PN, the embedding for the new physician 120 may be updated, e.g., through error backpropagation while freezing other parameters of the personalized model 442.

[0092] FIG. 6 is a schematic view of a system 500 for generating clinical notes, according to another embodiment of the present disclosure. The system 500 is substantially similar and functionally equivalent to the system 100 shown in FIG. 1, and same components in this embodiment are referred to by same reference numerals and differences between the embodiments are discussed.

[0093] In some embodiments, the at least one processor 104 is further configured to execute the instructions to generate a synthetic audio information 550 based on the prior clinical note 122 of the new physician 120. In some embodiments, the synthetic audio information 550 is representative of a synthetic conversation between the new physician 120 and a patient (e.g., a hypothetical patient). In some embodiments, a model may be trained that is capable of generating a conversation from a clinical note. The model may then be used to generate the synthetic audio information 550 from the prior clinical note 122.

[0094] In some embodiments, the synthetic audio information 550 may be in the form of an audio record or a transcript of the synthetic conversation. In some embodiments, one or more templates of the synthetic conversation may be prepared earlier. The prior clinical note 122 may then be used to fill in the template, thereby generating the synthetic conversation between the new physician 120 and a patient.

[0095] In some embodiments, the at least one processor 104 is further configured to execute the instructions to train the machine learning model 106 further based on the synthetic audio information 550 and the prior clinical note 122. For example, once enough number of pairs of the synthetic audio information 550 and the corresponding prior clinical notes 122 become available, the machine learning model 106 may be trained with this data.

[0096] FIG. 7 is a flowchart illustrating a method 600 for generating clinical notes. The method 600 may be implemented using any one of the systems 100, 200, 300, 400, 500 of FIGS. 1-6. Referring now to FIGS. 1-7, at step 602, the method 600 includes receiving the plurality of existing clinical notes 112 of the plurality of existing physicians 110 and the plurality of audio records 114 of the conversations between the plurality of existing physicians 110 and the respective patients 108. Each existing physician 110 from the plurality of existing physicians 110 has the one or more existing clinical notes 112 from the plurality of clinical notes 112. Each existing clinical note 112 is obtained based on the corresponding audio record 114 from the plurality of audio records 114.

[0097] At step 604, the method 600 further includes training the machine learning model 106 using each of the plurality of existing clinical notes 112 and the corresponding audio record 114, such that the machine learning model 106 is configured to generate the personalized clinical note PC for the existing physician 110 from the plurality of existing physicians 110 upon receiving the audio information A 1 of the conversation between the existing physician 110 and the patient 109.

[0098] At step 606, the method 600 further includes receiving the prior clinical note 122 of the new physician 120 absent in the plurality of existing physicians 110. At step 608, the method 600 further includes selecting the representative physician RP from the plurality of existing physicians 110 based on the prior clinical note 122 of the new physician 120 and the plurality of existing clinical notes 112. At step 610, the method 600 further includes receiving the audio information A2 of the new conversation between the new physician 120 and the patient 124. At step 612, the method 600 further includes generating, by the machine learning model 106, the personalized new clinical note PN of the new physician 120 based on the new conversation NC and the one or more existing clinical notes 112 of the representative physician RP.

[0099] In some embodiments, the method 600 further includes receiving the changes 128 on the personalized new clinical note PN from the new physician 120. In some embodiments, the method 600 further includes updating the machine learning model 106 based on the changes 128. [00100] In some embodiments, the method 600 further includes generating the similarity score 129 based on the comparison between the personalized new clinical note PN and the prior clinical note 122 of the new physician 120. In some embodiments, the method 600 further includes updating the machine learning model 106 based on the similarity score 129.

[00101] In some embodiments, training the machine learning model 106 further includes determining the one or more attributes 216 of each existing physician 110 based on the one or more existing clinical notes 112 of the corresponding existing physician 110. In some embodiments, training the machine learning model 106 further includes generating the plurality of physician tokens 218. Each physician token 218 from the plurality of physician tokens 218 is generated based on the one or more attributes 216 of the corresponding existing physician 110. In some embodiments, training the machine learning model 106 further includes embedding the plurality of physician tokens 218 in the machine learning model 106.

[00102] In some embodiments, training the machine learning model 106 further includes forming the plurality of physician clusters 232 of the plurality of existing physicians 110 based on the plurality of physician tokens 218. Each physician cluster 232 from the plurality of physician clusters 232 includes the one or more physician tokens 218 from the plurality of physician tokens 218 and the centroid 234 of the one or more physician tokens 218.

[00103] In some embodiments, selecting the representative physician RP further includes determining the one or more attributes 226 of the new physician 120 based on the prior clinical note 122. In some embodiments, selecting the representative physician RP further includes selecting the representative physician RP based on the comparison between the one or more attributes 226 of the new physician 120 and the centroid 234 of each physician cluster 232.

[00104] In some embodiments, training the machine learning model 106 further includes determining the plurality of unique identifiers 316 of the plurality of existing physicians 110. Each unique identifier 316 from the plurality of unique identifiers 316 is representative of the existing physician 110 from the plurality of existing physicians 110. In some embodiments, training the machine learning model 106 further includes encoding the plurality unique identifiers 316 using the one-hot encoding scheme to generate the encoded matrix 318. In some embodiments, training the machine learning model 106 further includes feeding the encoded matrix 318 to the machine learning model 106.

[00105] In some embodiments, selecting the representative physician RP further includes selecting the representative physician RP based on the comparison between the prior clinical note 122 of the new physician 120 and the one or more existing clinical notes 112 of each existing physician 110 whose unique identifier 316 is encoded in the encoded matrix 318.

[00106] In some embodiments, the machine learning model 106 includes the agnostic model 440 untrained by the plurality of existing clinical notes 112 and the plurality of audio records 114. In some embodiments, the machine learning model 106 further includes the personalized model 442 trained by each of the plurality of existing clinical notes 112 and the corresponding audio record 114. In some embodiments, the method 600 further includes generating, by the agnostic model 440, the agnostic clinical note AC based on the audio information A 1 of the conversation between the existing physician 110 and the patient 109. In some embodiments, the method 600 further includes modifying, by the personalized model 442, the agnostic clinical note AC to generate the personalized clinical note PC of the existing physician 110.

[00107] In some embodiments, training the machine learning model 106 further includes determining the one or more attributes 416 of each existing physician 110 based on the one or more existing clinical notes 112 of the corresponding existing physician 110. In some embodiments, training the machine learning model 106 further includes generating the plurality of physician tokens 418. Each physician token 418 from the plurality of physician tokens 418 is generated based on the one or more attributes 416 of the corresponding existing physician 110. In some embodiments, training the machine learning model 106 further includes embedding the plurality of physician tokens 418 in the personalized model 442 of the machine learning model 106.

[00108] In some embodiments, training the machine learning model 106 further includes forming the plurality of physician clusters 432 of the plurality of existing physicians 110 based on the plurality of physician tokens 418. Each physician cluster 432 from the plurality of physician clusters 432 includes the one or more physician tokens 418 from the plurality of physician tokens 418 and the centroid 434 of the one or more physician tokens 418.

[00109] In some embodiments, selecting the representative physician RP further includes determining the one or more attributes 426 of the new physician 120 based on the prior clinical note 122. In some embodiments, selecting the representative physician RP further includes selecting the representative physician RP based on the comparison between the one or more attributes 426 of the new physician 120 and the centroid 434 of each physician cluster 432. [00110] In some embodiments, the method 600 further includes generating the synthetic audio information 550 based on the prior clinical note 122 of the new physician 120. In some embodiments, the synthetic audio information 550 is representative of the synthetic conversation between the new physician 120 and a patient. In some embodiments, the method 600 further includes training the machine learning model 106 further based on the synthetic audio information 550 and the prior clinical note 122.

[00111] It should be understood that steps of the method 600 are not necessarily presented in any particular order and that performance of some or all the steps in an alternative order(s) is possible and is contemplated. The steps have been presented in the demonstrated order for ease of description and illustration. Further, it should be understood that steps can be added, omitted, and/or performed simultaneously without departing from the scope of the appended claims. Moreover, it should also be understood that the illustrated method 600 can be ended at any time. [00112] FIG. 8 is a flowchart illustrating a method 700 for generating clinical notes. As depicted, method 700 is similar to method 600. The primary differences between method 700 and method 600 is that method 700 is performed iteratively and at step 706 may determine whether to use an existing physician representation or select a representative physician when generating the personalized new clinical note for the physician, although other differences should be apparent from the depiction and the following description. That said, it is also possible for the method 600 to implement one or both of these features. Similarly, aspects of the method 600 described above that are not described in connection with the method 700 can be used according to particular embodiments.

[00113] Like method 600, the method 700 may be implemented using any one of the systems 100, 200, 300, 400, 500 of FIGS. 1-6. Referring now to FIGS. 1-8, at step 702, the method 700 includes receiving or otherwise using the plurality of existing clinical notes 112 of the plurality of existing physicians 110 and the plurality of audio records 114 of the conversations between the plurality of existing physicians 110 and the respective patients 108. Each existing physician 110 from the plurality of existing physicians 110 has the one or more existing clinical notes 112 from the plurality of clinical notes 112. Each existing clinical note 112 is obtained based on the corresponding audio record 114 from the plurality of audio records 114.

[00114] At step 704, the method 700 further includes training the machine learning model 106 using each of the plurality of existing clinical notes 112 and the corresponding audio record 114, such that the machine learning model 106 is configured to generate the personalized clinical note PC for the existing physician 110 from the plurality of existing physicians 110 upon receiving the audio information A 1 of the conversation between the existing physician 110 and the patient 109.

[00115] At step 706, the method 700 further includes determining, based at least in part on the output of the machine learning model, whether existing clinical notes and corresponding audio records are associated with a new physician or an existing physician. At step 708, if it was determined in step 706 that the clinical notes and corresponding audio records are not associated a new physician, then an existing physician representation is retrieved. On the other hand, at step 710, it was determined in step 706 that the clinical notes and corresponding audio records are associated a new physician, then a representative physician from the plurality of exiting physicians is selected. According to particular embodiments, the representative physician can be selected using one or more style features. These style features can range from the number of sentences (to capture verbosity), reading scores (to capture complexity of language use) and term frequency scores (e.g. TF-IDF). Other features can also be more fine-grained; e.g., the presence of bullet points, how the physician indicates emphasis (e.g., using bolded versus italicized characters), to name a few examples. These features can be used as a distance metric, where physicians that use similar features are closer (i.e., have a shorter computed distance) and physicians that use dissimilar features are farther (i.e., have a larger computed distance). In general, the physician with the shorter distance is selected as the representative physician, although other techniques for weighting, or biasing, which representative physician to select can also be used. At step 712, the method 700 further includes receiving the audio information A2 of the new conversation between the new physician 120 and the patient 124. At step 714, the method 700 further includes generating, by the machine learning model 106, the personalized new clinical note PN of the new physician 120 based on the new conversation NC and the one or more existing clinical notes 112 of the representative physician RP. At step 716, the method 700 further includes generating a final clinical note. For instance, the clinical note generated in step 714 can be presented to the physician and input can be received from the physician corresponding to one or more edits to the new clinical note. According to particular embodiments, the edits made to the clinical note by the physician can be used to further modify the machine learning model. For instance, stylistic changes or changes in word choice can be used to modify the machine learning model so that subsequent generations of the personalized clinical note more accurately reflect the preferences of the particular physician.

[00116] It should be understood that steps of the method 700 are not necessarily presented in any particular order and that performance of some or all the steps in an alternative order(s) is possible and is contemplated. The steps have been presented in the demonstrated order for ease of description and illustration. Further, it should be understood that steps can be added, omitted, and/or performed simultaneously without departing from the scope of the appended claims. Moreover, it should also be understood that the illustrated method 700 can be ended at any time. [00117] Referring to FIG. 1-8, the systems 100, 200, 300, 400, 500 the method 600, and the method 700 of the present disclosure may allow selection of the representative physician RP from the plurality of existing physicians 110 based on the prior clinical note 122 of the new physician 120 and the plurality of existing clinical notes 112, thereby enabling generation of the personalized new clinical note PN of the new physician 120 immediately upon enrolment into the systems 100, 200, 300, 400, 500.

[00118] The machine learning model 106 may be able to generate the personalized new clinical note PN of the new physician 120 with the style and personal preferences of the new physician 120 without needing prior training data for the new physician 120. Further, the personalized new clinical note PN may require minimal manual editing after generation to match the style of the new physician 120. This may save time and effort of the new physician 120 or associated medical staff to generate the personalized new clinical note PN.

[00119] The machine learning model 106 may be subsequently updated as audio information for more new conversations become available for the new physician 120 paired with the corresponding personalized new clinical note PN to improve accuracy. Since the machine leaning model 106 is able to generate the personalized new clinical note PN for the new physician 120 immediately upon enrolment, the training required by the machine learning model 106 to accurately generate the personalized new clinical note PN may be less as compared to conventional methods.

[00120] It should be noted that the system 100, 200, 300, 400, 500, the method 600, and the method 700 of the present disclosure may also be applied in context of hospital systems where new hospitals may be onboarded using a machine learning model based on prior hospital representations. Each hospital may include a plurality of physicians.

[00121] Unless otherwise indicated, all numbers expressing feature sizes, amounts, and physical properties used in the specification and claims are to be understood as being modified by the term “about”. Accordingly, unless indicated to the contrary, the numerical parameters set forth in the foregoing specification and attached claims are approximations that can vary depending upon the desired properties sought to be obtained by those skilled in the art utilizing the teachings disclosed herein.

[00122] As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” encompass embodiments having plural referents, unless the content clearly dictates otherwise. As used in this specification and the appended claims, the term “or” is generally employed in its sense including “and/or” unless the content clearly dictates otherwise.

[00123] Spatially related terms, including but not limited to, “proximate,” “distal,” “lower,” “upper,” “beneath,” “below,” “above,” and “on top,” if used herein, are utilized for ease of description to describe spatial relationships of an element(s) to another. Such spatially related terms encompass different orientations of the device in use or operation in addition to the particular orientations depicted in the figures and described herein. For example, if an object depicted in the figures is turned over or flipped over, portions previously described as below, or beneath other elements would then be above or on top of those other elements.

[00124] As used herein, when an element, component, or layer for example is described as forming a “coincident interface” with, or being “on,” “connected to,” “coupled with,” “stacked on” or “in contact with” another element, component, or layer, it can be directly on, directly connected to, directly coupled with, directly stacked on, in direct contact with, or intervening elements, components or layers may be on, connected, coupled or in contact with the particular element, component, or layer, for example. When an element, component, or layer for example is referred to as being “directly on,” “directly connected to,” “directly coupled with,” or “directly in contact with” another element, there are no intervening elements, components or layers for example.

[00125] Various examples have been described. These and other examples are within the scope of the following claims.

Claims

1. A method for generating clinical notes, the method comprising: receiving a plurality of existing clinical notes of a plurality of existing physicians and a plurality of audio records of conversations between the plurality of existing physicians and respective patients, wherein each existing physician from the plurality of existing physicians has one or more existing clinical notes from the plurality of clinical notes, and wherein each existing clinical note is obtained based on a corresponding audio record from the plurality of audio records; training a machine learning model using each of the plurality of existing clinical notes and the corresponding audio record, such that the machine learning model is configured to generate a personalized clinical note for an existing physician from the plurality of existing physicians upon receiving an audio information of a conversation between the existing physician and a patient; receiving a prior clinical note of a new physician absent in the plurality of existing physicians; selecting a representative physician from the plurality of existing physicians based on the prior clinical note of the new physician and the plurality of existing clinical notes; receiving an audio information of a new conversation between the new physician and a patient; and generating, by the machine learning model, a personalized new clinical note of the new physician based on the new conversation and the one or more existing clinical notes of the representative physician.

2. The method of claim 1 , wherein training the machine learning model further comprises: determining one or more attributes of each existing physician based on the one or more existing clinical notes of the corresponding existing physician; generating a plurality of physician tokens, wherein each physician token from the plurality of physician tokens is generated based on the one or more attributes of the corresponding existing physician; embedding the plurality of physician tokens in the machine learning model; and forming a plurality of physician clusters of the plurality of existing physicians based on the plurality of physician tokens, wherein each physician cluster from the plurality of physician clusters comprises one or more physician tokens from the plurality of physician tokens and a centroid of the one or more physician tokens.

3. The method of claim 2, wherein selecting the representative physician further comprises: determining one or more attributes of the new physician based on the prior clinical note; and selecting the representative physician based on a comparison between the one or more attributes of the new physician and the centroid of each physician cluster.

4. The method of claim 1 , wherein training the machine learning model further comprises: determining a plurality of unique identifiers of the plurality of existing physicians, wherein each unique identifier from the plurality of unique identifiers is representative of an existing physician from the plurality of existing physicians; encoding the plurality unique identifiers using a one-hot encoding scheme to generate an encoded matrix; and feeding the encoded matrix to the machine learning model.

5. The method of claim 4, wherein selecting the representative physician further comprises selecting the representative physician based on a comparison between the prior clinical note of the new physician and the one or more existing clinical notes of each existing physician whose unique identifier is encoded in the encoded matrix.

6. The method of claim 1, wherein the machine learning model comprises: an agnostic model untrained by the plurality of existing clinical notes and the plurality of audio records; and a personalized model trained by each of the plurality of existing clinical notes and the corresponding audio record; and wherein the method further comprises: generating, by the agnostic model, an agnostic clinical note based on the audio information of the conversation between the existing physician and the patient; and modifying, by the personalized model, the agnostic clinical note to generate the personalized clinical note of the existing physician.

7. The method of claim 6, wherein training the machine learning model further comprises: determining one or more attributes of each existing physician based on the one or more existing clinical notes of the corresponding existing physician; generating a plurality of physician tokens, wherein each physician token from the plurality of physician tokens is generated based on the one or more attributes of the corresponding existing physician; embedding the plurality of physician tokens in the personalized model of the machine learning model; and forming a plurality of physician clusters of the plurality of existing physicians based on the plurality of physician tokens, wherein each physician cluster from the plurality of physician clusters comprises one or more physician tokens from the plurality of physician tokens and a centroid of the one or more physician tokens.

8. The method of claim 7, wherein selecting the representative physician further comprises: determining one or more attributes of the new physician based on the prior clinical note; and selecting the representative physician based on a comparison between the one or more attributes of the new physician and the centroid of each physician cluster.

9. The method of claim 1, further comprising: receiving changes on the personalized new clinical note from the new physician; and updating the machine learning model based on the changes.

10. The method of claim 1, further comprising: generating a similarity score based on a comparison between the personalized new clinical note and the prior clinical note of the new physician; and updating the machine learning model based on the similarity score.

11. The method of claim 1 , further comprising: generating a synthetic audio information based on the prior clinical note of the new physician, wherein the synthetic audio information is representative of a synthetic conversation between the new physician and a patient; and training the machine learning model further based on the synthetic audio information and the prior clinical note.

12. A system for generating clinical notes, the system comprising: at least one non-transitory computer-readable storage medium having instructions stored thereon; and at least one processor coupled to the at least one non-transitory computer-readable storage medium and configmed to execute the instructions to: receive a plurality of existing clinical notes of a plurality of existing physicians and a plurality of audio records of conversations between the plurality of existing physicians and respective patients, wherein each existing physician from the plurality of existing physicians has one or more existing clinical notes from the plurality of clinical notes, and wherein each existing clinical note is obtained based on a corresponding audio record from the plurality of audio records; train a machine learning model using each of the plurality of existing clinical notes and the corresponding audio record, such that the machine learning model is configured to generate a personalized clinical note for an existing physician from the plurality of existing physicians upon receiving an audio information of a conversation between the existing physician and a patient; receive a prior clinical note of a new physician absent in the plurality of existing physicians; select a representative physician from the plurality of existing physicians based on the prior clinical note of the new physician and the plurality of existing clinical notes; receive an audio information of a new conversation between the new physician and a patient; and generate, by the machine learning model, a personalized new clinical note of the new physician based on the new conversation and the one or more existing clinical notes of the representative physician.

13. The system of claim 12, wherein training the machine learning model further comprises: determining one or more attributes of each existing physician based on the one or more existing clinical notes of the corresponding existing physician; generating a plurality of physician tokens, wherein each physician token from the plurality of physician tokens is generated based on the one or more attributes of the corresponding existing physician; embedding the plurality of physician tokens in the machine learning model; and forming a plurality of physician clusters of the plurality of existing physicians based on the plurality of physician tokens, wherein each physician cluster from the plurality of physician clusters comprises one or more physician tokens from the plurality of physician tokens and a centroid of the one or more physician tokens.

14. The system of claim 13, wherein selecting the representative physician further comprises: determining one or more attributes of the new physician based on the prior clinical note; and selecting the representative physician based on a comparison between the one or more attributes of the new physician and the centroid of each physician cluster.

15. The system of claim 12, wherein training the machine learning model further comprises: determining a plurality of unique identifiers of the plurality of existing physicians, wherein each unique identifier from the plurality of unique identifiers is representative of an existing physician from the plurality of existing physicians; encoding the plurality unique identifiers using a one-hot encoding scheme to generate an encoded matrix; and feeding the encoded matrix to the machine learning model.

16. The system of claim 15, wherein selecting the representative physician further comprises selecting the representative physician based on a comparison between the prior clinical note of the new physician and the one or more existing clinical notes of each existing physician whose unique identifier is encoded in the encoded matrix.

17. The system of claim 12, wherein the machine learning model comprises: an agnostic model untrained by the plurality of existing clinical notes and the plurality of audio records; and a personalized model trained by each of the plurality of existing clinical notes and the corresponding audio record; and wherein the at least one processor is fiirther configmed to execute the instructions to: generate, by the agnostic model, an agnostic clinical note based on the audio information of the conversation between the existing physician and the patient; and modify, by the personalized model, the agnostic clinical note to generate the personalized clinical note of the existing physician.

18. The system of claim 17, wherein training the machine learning model further comprises: determining one or more attributes of each existing physician based on the one or more existing clinical notes of the corresponding existing physician; generating a plurality of physician tokens, wherein each physician token from the plurality of physician tokens is generated based on the one or more attributes of the corresponding existing physician; embedding the plurality of physician tokens in the personalized model of the machine learning model; and forming a plurality of physician clusters of the plurality of existing physicians based on the plurality of physician tokens, wherein each physician cluster from the plurality of physician clusters comprises one or more physician tokens from the plurality of physician tokens and a centroid of the one or more physician tokens.

19. The system of claim 18, wherein selecting the representative physician further comprises: determining one or more attributes of the new physician based on the prior clinical note; and selecting the representative physician based on a comparison between the one or more attributes of the new physician and the centroid of each physician cluster.

20. The system of claim 12, wherein the at least one processor is further configured to execute the instructions to: receive changes on the personalized new clinical note from the new physician; and update the machine learning model based on the changes.

21. The system of claim 12, wherein the at least one processor is further configured to execute the instructions to: generate a similarity score based on a comparison between the personalized new clinical note and the prior clinical note of the new physician; and update the machine learning model based on the similarity score.

22. The system of claim 12, wherein the at least one processor is further configured to execute the instructions to: generate a synthetic audio information based on the prior clinical note of the new physician, wherein the synthetic audio information is representative of a synthetic conversation between the new physician and a patient; and train the machine learning model further based on the synthetic audio information and the prior clinical note.