Hybrid model for principal diagnosis prediction

Abstract

A system for determining a principal diagnosis includes a computer processor configured to 5 receive structured medical claims data including at least a plurality of diagnosis codes. The computer processor is further configured to provide the medical claims data to a machine learning model and receive a plurality of partial codes and a plurality of first confidence scores. The computer processor is further configured to provide the plurality of partial codes and the plurality of first confidence scores to a computational model and receive a plurality of full codes and a plurality of second confidence scores. 0 The computational model is configured to filter and rank each full code from the plurality of diagnosis codes, such that each full code includes a corresponding partial code. The computer processor is further configured to output the plurality of full codes in an ordered recommendation based on the plurality of second confidence scores.

Description

[0001] HYBRID MODEL FOR PRINCIPAL DIAGNOSIS PREDICTION

[0002] Technical Field

[0003] The present disclosure generally relates to a system and a method for determining principal diagnosis.

[0004] Background

[0005] Medical coding of diagnoses and procedures for reimbursement is a central process in health care in several countries and is usually conducted by medical controllers who are specialized medical personnel. In the process of reimbursement, structured patient data is collected by the medical coding specialist from medical documentation and every medical encounter is then assigned a diagnosis related group (DRG), which provides the amount of reimbursement that a health care provider will receive for a given patient. The DRG is computed by complex deterministic algorithms, so called DRG-groupers, which receive patient medical data as input, i.e., age, sex, weight, medical diagnosis codes, e.g., international classification of diseases (ICD) codes, medical procedures codes, e.g., operations and procedure classification (OPS) codes, a principal diagnosis (which is one of the medical diagnosis codes), etc.

[0006] The principal diagnosis is the diagnosis that encodes the main reason of admission of the patient to a hospital and may have a very high impact on the final DRG. Usually, the principal diagnosis is selected by the medical coding specialist of the hospital after completion of the medical procedures and discharge of the patient. High mental resources are required for determining the principal diagnosis, since it is hard to determine, especially for complex cases where there are many diagnoses and procedures. Further, high computing resources may be required by software solutions that replace this task of the medical coding specialist. This may limit a number of coded cases, a quality of coding, and a speed of the reimbursement process which may influence a financial condition of the health care provider.

[0007] Summary

[0008] In a first aspect, the present disclosure provides a system for determining principal diagnosis code. The system includes at least one non-transitory computer-readable storage medium having instructions stored thereon. The system further includes at least one computer processor coupled to the at least one non-transitory computer-readable storage medium and configured to execute the instructions to receive structured medical claims data including at least a plurality of diagnosis codes associated with an inpatient hospital stay. The structured medical claims data may further include other information, e.g., procedure codes, age, sex, gender, demographic data, etc. The computer processor is further configmed to execute the instructions to provide the structured medical claims data to a machine learning model. The computer processor is further configmed to execute the instructions to receive a plurality of partial codes and a corresponding plurality of first confidence scores from the machine learning model. Each partial code from the plurality of partial codes may be a portion of a corresponding potential principal diagnosis code from a plurality of potential principal diagnosis codes. The computer processor is further configured to execute the instructions to provide the plurality of partial codes and the plurality of first confidence scores to a computational model. The computer processor is further configured to execute the instructions to receive a plurality of full codes and a corresponding plurality of second confidence scores from the computational model. The computational model is configured to filter and rank, based at least in part on the plurality of second confidence scores, each of the plurality of full codes from the plurality of diagnosis codes, such that each full code includes a corresponding partial code from the plurality of partial codes. The computer processor is further configmed to execute the instructions to output the plurality of full codes in an ordered recommendation based at least in part on the plurality of second confidence scores to at least one of a user interface, a storage device, and the at least one non-transitory computer-readable storage medium.

[0009] In a second aspect, the present disclosure provides a method for determining principal diagnosis code. The method includes receiving stmctured medical claims data including at least a plurality of diagnosis codes associated with an inpatient hospital stay. The structured medical claims data may further include other information, e.g., procedure codes, age, sex, gender, demographic data, etc. The method further includes providing the structured medical claims data to a machine learning model. The method further includes receiving a plurality of partial codes and a corresponding plurality of first confidence scores from the machine learning model. Each partial code from the plurality of partial codes may be a portion of a corresponding potential principal diagnosis code from a plurality of potential principal diagnosis codes. The method further includes providing the plurality of partial codes and the plurality of first confidence scores to a computational model. The method further includes receiving a plurality of full codes and a corresponding plurality of second confidence scores from the computational model. The computational model is configured to filter and rank, based at least in part on the plurality of second confidence scores, each of the plurality of full codes from the plurality of diagnosis codes, such that each full code includes a corresponding partial code from the plurality of partial codes. The method further includes outputting the plurality of full codes in an ordered recommendation based at least in part on the plurality of second confidence scores to at least one of a user interface, a storage device, and the at least one non-transitory computer-readable storage medium.

[0010] 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.

[0011] Brief Description of the Drawings

[0012] 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.

[0013] FIG. 1 is a schematic block diagram of a system for determining a principal diagnosis code, according to an embodiment of the present disclosure;

[0014] FIG. 2 is a schematic block flow diagram illustrating components of the system, according to another embodiment of the present disclosure;

[0015] FIG. 3 is a schematic block diagram of the system, according to another embodiment of the present disclosure;

[0016] FIG. 4 is a schematic block flow diagram illustrating components of the system for training a machine learning model, according to another embodiment of the present disclosure;

[0017] FIG. 5 is a schematic block diagram of the system, according to another embodiment of the present disclosure;

[0018] FIG. 6 is a schematic block flow diagram illustrating a computational model of the system for determining a plurality of full codes and a corresponding plurality of second confidence scores, according to an embodiment of the present disclosure;

[0019] FIG. 7 is a schematic block diagram of the system, according to another embodiment of the present disclosure; and

[0020] FIG. 8 is a flowchart illustrating a method for determining the principal diagnosis code, according to an embodiment of the present disclosure.

[0021] Detailed Description

[0022] 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.

[0023] In the following disclosure, the following definitions are adopted.

[0024] 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. As used herein, the term “medical record” may refer to medical data generated by a provider for an individual person. The medical record may be in the form of a text document. The medical record may also be referred to as an electronic health record (EHR).

[0025] 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.

[0026] 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.

[0027] As used herein, the terms “medical claims” or “medical claims data” may refer to a combination of both physician claims and hospital claims. The claims may contain information submitted by a provider requesting payment of fees or costs of services provided to a patient. The claims may contain information about the patient and information describing types of services provided, including dates on which such services were provided. For example, the claims may include a date of service, a physician identification, an admission department, patient information (e.g., age, sex, weight, etc.), medical status, morbidities, reason(s) for visit, and a description of service(s) provided. The medical status and morbidities of the patient are represented by diagnosis codes. The reason for visit is typically represent by one of the diagnosis codes, called a principal diagnosis code. The description of the services is provided by operation and procedure codes.

[0028] As used herein, the term “medical coding” may refer to a process of assigning appropriate codes to medical terms in patient records. Additionally, the term may also be directed to assigning operational codes associated with one or more medical treatments relating to a medical diagnosis code. In the medical coding process, documented information regarding a patient encounter, such as the patient’s diagnoses and medical procedures performed, is classified according to one or more standardized sets of codes for reporting to various entities, such as payment providers (e.g., health insurance companies that reimburse health care providers for their services).

[0029] As used herein, the term “diagnosis code” may refer to an entry in a standardized domain specific code table for medical diagnoses. A typical example is ICD-10 code (International Classification of Diseases).

[0030] As used herein, the term “full code” may refer to any representation of a diagnosis code from a classification of standardized codes that may or may not be in a hierarchical structure, for instance, a code from the ICD-10 codes. For example, in ICD-10 coding system, a full code is any code that is allowed to be used without further specification, e.g., “110.00”, “ZU”, “J96.0”. These codes may be referred to as “terminal codes” or “terminated codes”, however, this is specific to ICD-10 classification and there may be differences in other coding systems.

[0031] As used herein, the term “partial code” may refer to a full code itself, or any representation of a parent code of the full code, or a group that contains a parent code of the full code. A partial code may therefore be any representation that may be on a same or a higher level in a hierarchical coding classification of standardized codes as compared to the full code. For example, considering the full code “C92.00”, possible partial codes could be a direct parent in the hierarchy, i.e., “C92”, a class of the full code, i.e., “C”, or a chapter of the full code “II” (roman 2) that groups multiple classes together. The partial code may also be any representation that may be a description of a more general representation of “C92.00“, e.g., “leukemia” for the code itself or “neoplasms” for the chapter.

[0032] As used herein, the term “potential principal diagnosis code” may refer to any full code that may be potentially used as a principal diagnosis code. The full codes that are allowed to be used as the principal diagnosis code are defined by the used coding system. In general, some full codes may not be used as the principal diagnosis code, e.g., in the ICD-10GM from the year 2023, the full code “U99.0” cannot be used as a principal diagnosis code.

[0033] As used herein, the term “procedure code” may refer to one of a catalog of standardized medical treatments. One example of such a catalog used in Germany, is denoted as OPS (operation and procedure), which are codes that are used to classify various tests and treatments. Other examples are CHOP (the Swiss catalog of operations and procedures), current procedural terminology (CPT) code set owned by the American Medical Association, Health Care Procedural Code (“HCPC”) (promulgated by the Health Care Financing Administration), etc. Other countries may have other catalogs in their respective medical systems.

[0034] As used herein, the term “diagnosis related group (DRG)” may refer to groups of different medical diagnosis and procedure code combinations with similar economic and medical resource allocation that form a basis for financing, budgeting, and billing of medical services. DRG systems and DRG grouper are used to determine the DRG for a given medical case. The DRG grouper deterministically maps medical information to a pre-defined DRG from a corresponding DRG system. The DRG systems and DRG grouper may underly regional and longitudinal specific variations and changes.

[0035] 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 or classification 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”. 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 (“GNNs”), generative adversarial neural networks (“GANs”), and single-shot detect networks (“SSDs”). 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 training data to tune parameters of the neural network.

[0036] 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.

[0037] 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.

[0038] As used herein, the term “communicably coupled” generally refers to any type of connection or coupling that allows for the exchange or sharing 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.

[0039] 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.

[0040] 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.

[0041] 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). 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.

[0042] In the process of medical reimbursement, structured patient data is collected by a medical coding specialist from medical documentation and every medical encounter is then assigned a diagnosis related group (DRG), which provides an amount of reimbursement that a health care provider will receive for a given patient. The DRG is computed based on patient medical data as input, i.e., age, sex, weight, medical diagnosis codes, e.g., international classification of diseases (ICD) codes, medical procedures codes, e.g., OPS codes, a principal diagnosis (which is one of the medical diagnosis codes), etc. The principal diagnosis encodes the main reason of admission of the patient to the hospital and may have a very high impact on the final DRG. Usually, the principal diagnosis is selected by the medical coding specialist after completion of the medical procedures and discharge of the patient. High mental resources are required for determining the principal diagnosis, since it is hard to determine, especially for complex cases where there are many diagnoses and procedures. Further, high computing resources may be required by software solutions that replace this task of the medical coding specialist. This may limit a number of coded cases, a quality of coding, and a speed of the medical reimbursement process which may influence a financial condition of the health care provider.

[0043] The present disclosure provides a system for determining principal diagnosis code. The system includes at least one non-transitory computer-readable storage medium having instructions stored thereon. The system further includes at least one computer processor coupled to the at least one non- transitory computer-readable storage medium and configured to execute the instructions to receive structured medical claims data including at least a plurality of diagnosis codes associated with an inpatient hospital stay. The structured medical claims data may further include other information, e.g., procedure codes, age, sex, gender, demographic data, etc. The computer processor is further configured to execute the instructions to provide the structured medical claims data to a machine learning model. The computer processor is further configured to execute the instructions to receive a plurality of partial codes and a corresponding plurality of first confidence scores from the machine learning model. Each partial code from the plurality of partial codes may be a portion of a corresponding potential principal diagnosis code from a plurality of potential principal diagnosis codes. The computer processor is further configured to execute the instructions to provide the plurality of partial codes and the plurality of first confidence scores to a computational model. The computer processor is further configured to execute the instructions to receive a plurality of full codes and a corresponding plurality of second confidence scores from the computational model. The computational model is configured to fdter and rank, based at least in part on the plurality of second confidence scores, each of the plurality of full codes from the plurality of diagnosis codes, such that each full code includes a corresponding partial code from the plurality of partial codes. The computer processor is further configured to execute the instructions to output the plurality of full codes in an ordered recommendation based at least in part on the plurality of second confidence scores to at least one of a user interface, a storage device, and the at least one non- transitory computer-readable storage medium.

[0044] The system of the present disclosure utilizes a hybrid approach in which the machine learning model (i.e., a supervised machine learning classifier) first provides the plurality of partial codes and the corresponding plurality of first confidence scores. Subsequently, the computational model (i.e., a statistical model) provides the plurality of full codes and the corresponding plurality of second confidence scores. Thus, the system may divide the task of determining the principal diagnosis into two sub-tasks, i.e., prediction of the plurality of partial codes via the machine learning model and determination of the plurality of full codes via the computational model. Each partial code may be a portion of the corresponding potential principal diagnosis code. Hence, this proposed approach may allow reduction in a complexity of the system since a target space (i.e., a number of the partial codes) of the machine learning model is reduced. This also improves a performance and an efficiency of the system as compared to other approaches that directly predict the full code of the principal diagnosis.

[0045] Mapping of the plurality of partial codes to the plurality of full codes is done by the computational model using collected and aggregated data in the form of a data source. Examples of the data source may include a reference database, a lookup database, a data catalog (such as the catalog described elsewhere in this specification), data collection, or any other appropriate repository of information. That is, “catalog” is to be understood to be used interchangeably with “data source” as used herein. The catalog may be prepared during training of the machine learning model. Specifically, a frequency of diagnosis codes used as principal diagnosis at admission department level is used to identify a correct principal diagnosis. This is relevant when the plurality of partial codes proposed by the machine learning model matches with two or more full codes determined by the computational model. The catalog may also be used to determine the plurality of diagnosis codes that were never used as a principal diagnosis and filter them accordingly.

[0046] The system may predict the plurality of full codes using the structured medical claims data, without the usage of clinical text documentation. In general, the system may be utilized to predict the principal diagnosis out of the structured medical claims data during a stay of the patient as well, e.g., to forecast required resources of a health care provider. The system may automatically provide the ordered recommendation of the plurality of full codes for determining the principal diagnosis given the structured medical claims data. The ordered recommendation is based on the plurality of second confidence scores. Proposing multiple reasonable candidates not only addresses an unambiguity of the principal diagnosis but may also help to optimize reimbursement (upcoding), or to satisfy the health care providers’ different coding preferences.

[0047] The system may be directly integrated into an existing health care software as a new feature. Such integration may improve an acceptance of the software and increase a user satisfaction. Furthermore, the proposed approach may be used in allied applications as well, for instance, for reimbursement review by health care insurance companies. The proposed solution is data-driven and based on machine learning, which makes it easier to maintain and extend. Moreover, the system provides higher quality recommendations than previous approaches.

[0048] FIG. 1 is a schematic block diagram of a system 100 for determining a principal diagnosis code 106 (or a principal diagnosis). In some embodiments, the principal diagnosis may be one of diagnosis codes associated with a patient. The diagnosis codes may be one or more standardized sets of codes assigned based on documented information regarding a patient encounter, such as the patient’s diagnoses. The principal diagnosis (or main diagnosis) provides the main reason for the patient’s visit to a health care provider, such as a hospital, a clinic, a doctor’s office, and the like. Subsequently, the principal diagnosis may be used for determining a diagnosis related group (DRG) of the patient, which in turn determines a financial reimbursement for the health care provider.

[0049] 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 computer 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 computer processor 104” is interchangeably referred to herein as “the computer processor 104”.

[0050] 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 non-volatile 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 computer processor 104. Alternatively, or in addition, the storage medium 102 may be separate from the computer 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.

[0051] In some embodiments, the computer processor 104 may be embodied in a number of different ways. For example, the computer 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 computer 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 computer processor 104 is embodied as an executor of software instructions, the instructions may specifically configure the computer processor 104 to perform the operations described herein. Alternatively, as another example, when the computer processor 104 is embodied as an ASIC, FPGA, or the like, the computer processor 104 may have specifically configured hardware for conducting the operations described herein.

[0052] 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.

[0053] The at least one computer processor 104 is configured to execute the instructions to receive structured medical claims data 109 including at least a plurality of diagnosis codes 111 associated with a patient 108. In some embodiments, the patient 108 is an inpatient hospital stay. In some embodiments, the structured medical claims data 109 includes preprocessed data in a predefined format. For example, the structured medical claims data 109 includes the plurality of diagnosis codes 111 and not clinical text documentation. Specifically, the system 100 may not require entire clinical text documentation for determining the principal diagnosis. However, in alternative embodiments, the clinical text documentation may also be provided as input data. In some embodiments, the structured medical claims data 109 may also be provided by the storage medium 102 or a user interface. In some embodiments, the plurality of diagnosis codes 111 includes multiple diagnosis codes 110. In some embodiments, the structured medical claims data 109 corresponds to a current medical case and does not include historical hospitalization data associated with the patient 108.

[0054] In some embodiments, the principal diagnosis code 106 is one of the plurality of diagnosis codes 111. In some embodiments, the plurality of diagnosis codes 111 may be obtained directly from a medical record associated with the patient 108. The plurality of diagnosis codes 111 may be from a classification of standardized codes, which may or may not be in a hierarchical structure (e.g., International Classification of Diseases, ICD-10 codes). Examples of such codes may include F41.9, R10, J96.00, R13.1, F32.9, K80.01, etc. Generally, ICD-10 codes consist of up to 5 relevant characters with a format “cnn.nn” (c = an alphabetic character, n = a character including numbers and special characters). Local modifications of the ICD-10 system, e.g., the ICD-10GM for the German healthcare market may allow additional characters. The ICD-10 codes are organized in a hierarchical tree structure.

[0055] The at least one computer processor 104 is configmed to execute the instructions to provide the structured medical claims data 109 to a machine learning model 120. Specifically, the computer processor 104 may provide the structured medical claims data 109 to the machine learning model 120 in the form of a vector representation. The vector representation may be computed based on the structured medical claims data 109.

[0056] The at least one computer processor 104 is configured to execute the instructions to receive a plurality of partial codes 121 and a corresponding plurality of first confidence scores 123 from the machine learning model 120. Each partial code 122 from the plurality of partial codes 121 has a corresponding first confidence score 124 from the plurality of first confidence scores 123. Further, each partial code 122 from the plurality of partial codes 121 may be a portion of a corresponding potential principal diagnosis code 114 from a plurality of potential principal diagnosis codes 113. The plurality of potential principal diagnosis codes 113 includes multiple potential principal diagnosis codes 114. In some embodiments, the plurality of potential principal diagnosis codes 113 may be obtained from a plurality of all potential principal diagnosis codes that can possibly be a principal diagnosis, such as the principal diagnosis code 106.

[0057] Each partial code 122 may represent a portion of a standardized code, i.e., a portion of an ICD- 10 code. Each partial code 122 may include, e.g., the alpha-numeric character of a corresponding ICD- 10 code or may include other characters (e.g., roman characters). For example, the partial code 122 corresponding to code “F41.9” is “F41”, the partial code 122 corresponding to code “120.0” is “120”, the partial code 122 corresponding to code “E10.91” (diabetes mellitus) is “E10”, and so on. Thus, each partial code 122 may be an ICD-10 code in a shortened form. Generally, ICD-10 codes that are higher in the hierarchy represent a disease more generally, while ICD-10 codes that are lower in the hierarchy further refine the medical diagnosis. Hence, the partial codes 122 may carry the most relevant information.

[0058] In some embodiments, the machine learning model 120 predicts the plurality of partial codes 121 based on the structured medical claims data 109 including the plurality of diagnosis codes 111. Each of the plurality of partial codes 121 may be a portion of the corresponding potential principal diagnosis code 114 and each first confidence score 124 represents a probability of the corresponding partial code 122 being a portion of the principal diagnosis code 106. The plurality of first confidence scores 123 reflect a belief of the machine learning model 120 that the corresponding partial code 122 is the principal diagnosis of the encounter. In some embodiments, the machine learning model 120 may be trained to utilize the structured medical claims data 109 where the only required input is the plurality of diagnosis codes 111 containing at least a single ICD-10 code that can be used as the principal diagnosis. Thus, complete medical text documentation may not be required by the machine learning model 120.

[0059] In some embodiments, the computer processor 104 is further configured to receive an admission code 128 of an admission department associated with the patient 108. Specifically, the structured medical claims data 109 further includes the admission code 128. The admission code 128 may represent a department of the healthcare provider where the patient 108 is admitted. When the admission code 128 is not received, the computer processor 104 is further configured to use a predetermined default admission code 129. Another optional and beneficial input for the machine learning model 120 is a plurality of procedure codes 125 for each encounter. Specifically, the structured medical claims data 109 further includes the plurality of procedure codes 125. In some embodiments, the plurality of procedure codes 125 includes multiple procedure codes 126.

[0060] In some embodiments, each of the plurality of procedure codes 125 may be from a classification of standardized codes, e.g., OPS CHOP, CPT codes, HCPC, etc. In the case of OPS codes, each procedure code 126 consists of up to 6 relevant characters with a format “n-nn(n / c).(n / c)(n / c)” (c = an alphabetic character, n = a numerical character), e.g., 8-805, 8-805.40, 1-208.1, 8-010.3, 8-930, 9-262.0, etc. In some embodiments, the structured medical claims data 109 further includes other information such as age, sex, gender, demographic data, and other patient data.

[0061] In some embodiments, the machine learning model 120 is a classification model. However, the machine learning model 120 may also be any other type of machine learning model, including a supervised or unsupervised model, that may efficiently utilize a vector representation, e.g., term frequency - inverse document frequency (TF-IDF), as input and a multi-class target (i.e., the plurality of partial codes 121) as output, and further provides probabilities or confidence scores for ranking of the multi-class target. In some embodiments, the supervised machine learning classifier may be a “Passive-Aggressive Classifier” (See, Crammer, K., et al., “Online Passive-Aggressive Algorithms,” J. Machine Learning Res., 7:551-585 (2006).

[0062] The at least one computer processor 104 is configmed to execute the instructions to provide the plurality of partial codes 121 and the plurality of first confidence scores 123 to a computational model 130. In some embodiments, the computational model 130 is different from the machine learning model 120. The at least one computer processor 104 is configured to execute the instructions to receive a plurality of full codes 131 and a corresponding plurality of second confidence scores 133 from the computational model 130. Each full code 132 from the plurality of full codes 131 has a corresponding second confidence score 134 from the plurality of second confidence scores 133. The computational model 130 is configured to filter and rank, based on the plurality of second confidence scores 133, each of the plurality of full codes 131 from the plurality of diagnosis codes 111, such that each full code 132 includes a corresponding partial code 122 from the plurality of partial codes 121.

[0063] In some embodiments, the computational model 130 may include a deterministic postprocessing algorithm. In some embodiments, the computational model 130 is a statistical model. In some embodiments, the plurality of full codes 131 may be a subset of the plurality of diagnosis codes 111 having their standard form. For example, the plurality of diagnosis codes 111 that are never used as principal diagnosis code 106 may be fdtered out to generate the subset. This information is provided by a catalog 136. The catalog 136 will be discussed in greater detail later in the disclosure.

[0064] Each of the plurality of full codes 131 is assigned the corresponding second confidence scores 134 from the plurality of second confidence scores 133. In some embodiments, the plurality of second confidence scores 133 may be determined based on the plurality of first confidence scores 123 received from the machine learning model 120. For example, each of the plurality of second confidence scores 133 is determined by assigning the first confidence score 124 of the corresponding partial code 122 to the corresponding full code 132. This will be described in greater detail later in the disclosure.

[0065] The at least one computer processor 104 is configured to execute the instructions to output the plurality of full codes 131 in an ordered recommendation based at least in part on the plurality of second confidence scores 133 to at least one of a user interface 140, a storage device 142, and the at least one non-transitory computer-readable storage medium 102. In some embodiments, the computer processor 104 is further configured to output the plurality of second confidence scores 133 corresponding to the plurality of full codes 131. In some embodiments, the computer processor 104 is further configmed to select the principal diagnosis code 106 from the ordered recommendation based on the plurality of second confidence scores 133. In some embodiments, the plurality of full codes 131 is ranked based on the plurality of second confidence scores 133. Further, the full code 132 having the highest second confidence score 134 may be selected as the principal diagnosis code 106.

[0066] In some embodiments, the user interface 140 may include a display, e.g., a touch screen type Liquid Crystal Display (LCD) that may enable the user to interact with the user interface 140. In alternative embodiments, a user may refer to the plurality of full codes 131 and the plurality of second confidence scores 133 through at least one of the user interface 140, the storage device 142, and the at least one non-transitory computer-readable storage medium 102 to determine the principal diagnosis code 106.

[0067] FIG. 2 is a schematic block flow diagram illustrating components of the system 100 for determining the principal diagnosis. Referring to FIGS. 1 and 2, inputs Xi, X2, X3, ... , Xn(collectively, inputs X) to the system 100 consist of numerical and categorical features, where n is a positive integer corresponding to a total number of the inputs X. In some embodiments, the inputs X form a part of the structured medical claims data 109.

[0068] For example, the input Xi represents the plurality of diagnosis codes 111 (e.g., the ICD-10 codes), e.g., E78.5, 110.00, 124.8, U99.0, R06.0, etc. The input X2represents the admission code 128. The input X3represents the plurality of procedure codes 125 (e.g., the OPS codes), e.g., 5-530.32, 8- 800.0, etc. Other inputs X4 - Xnto the system 100 may be age in years, weight, gender code, admission information, demographic information, etc. It should be understood that there may be other inputs X to the system 100 based on application requirements. In some embodiments, the input X2(i.e., the admission code 128) may be an alpha-numeric string while the inputs Xi and X3may be represented by lists of alpha-numeric strings.

[0069] A pre-processing module 144 transforms the inputs X (i.e., the structured medical claims data 109) into an appropriate format to prepare the inputs X for the machine learning model 120. For example, the inputs X are represented as strings and concatenated into a textual representation. In some implementations, the textual representation can be a syntax-less representation. This pre-processing includes basic natural language processing (NLP) techniques, such as tokenization, stop-word removal, feature generation, and discrimination. For example, stop words such as “Test for Corona” (U99.0), which is mandatory and has no additional information may be excluded when processing the inputs X. In a preferred embodiment, textual input data is represented as a sparse TF-IDF matrix representation.

[0070] A pre-processing module 146 processes the input X2, i.e., the admission code 128. During preprocessing, the admission code 128 is mapped to several pre-defined department groups. The mapping itself is based on domain knowledge, e.g., a code “0100” is mapped to “internal medicine department” because many hospitals use the code “0100” to refer to their internal medicine department. If the input X2is unknown or not present, a default group name may be used without significant loss of performance. The predetermined default admission code 129 may be associated with the default group name.

[0071] In some embodiments, the machine learning model 120 predicts a partial code of the principal diagnosis as the plurality of partial codes 121 and provides probabilities or softmax scores for each partial code 122 as the corresponding first confidence score 124. These probability values or other score values may allow the plurality of partial codes 121 to be sorted by their most likely occurrence. In some embodiments, the machine learning model 120 may provide the first confidence score 124 corresponding to each partial code 122 that can possibly be a principal diagnosis. The computational model 130 utilizes the input Xi, i.e., the plurality of diagnosis codes 111, to determine the plurality of full codes 131 based on the plurality of partial codes 121, the plurality of first confidence scores 123, inputs from the catalog 136, and the input X2, i.e., the admission code 128.

[0072] FIG. 3 is a schematic block diagram of the system 100, according to another embodiment of the present disclosure. In some embodiments, the computer processor 104 is further configured to train the machine learning model 120 using a training medical claims data 150 including a plurality of sets of training diagnosis codes 151 and a plurality of training principal diagnosis codes 153 corresponding to the plurality of sets of training diagnosis codes 151. The plurality of sets of training diagnosis codes

[0073] 151 includes multiple sets of training diagnosis codes 152. Further, the set of training diagnosis codes

[0074] 152 includes multiple training diagnosis codes 152a. The plurality of training principal diagnosis codes

[0075] 153 includes multiple training principal diagnosis codes 154. In some embodiments, each potential principal diagnosis code 114 (shown in FIG. 1) is one of the plurality of training principal diagnosis codes 153.

[0076] Each training principal diagnosis code 154 from the plurality of training principal diagnosis codes 153 is associated with a corresponding set of training diagnosis codes 152 from the plurality of sets of training diagnosis codes 151. For example, each training principal diagnosis code 154 may represent a principal diagnosis code 106 assigned by a medical controller of a health care provider for a patient based on the corresponding set of training diagnosis codes 152 associated with the patient. In some embodiments, each training principal diagnosis code 154 is one training diagnosis code 152a included in the corresponding set of training diagnosis codes 152. In some embodiments, the training medical claims data 150 further includes a plurality of admission department codes 155 for a corresponding plurality of admission departments 157. The plurality of admission department codes 155 includes multiple admission department codes 156. Further, the plurality of admission departments 157 includes multiple admission departments 158. In some embodiment, the training medical claims data 150 may further include other inputs, such as age, weight, sex, demographic information, etc. Such inputs may ensure robustness and high performance of the machine learning model 120. Similar to classical machine learning models, the machine learning model 120 is a data driven model that needs to be trained on a sufficient amount of representative data, i.e., the training medical claims data 150. The training medical claims data 150 may include at least the plurality of training principal diagnosis codes 153 and the corresponding plurality of sets of training diagnosis codes 151.

[0077] FIG. 4 is a schematic block flow diagram illustrating components of the system 100 for training the machine learning model 120. Referring to FIGS. 3 and 4, training inputs Ti, T2, T3, ..., Tm(collectively, training inputs T) for training the machine learning model 120 consist of numerical and categorical features, where m is a positive integer corresponding to a total number of the training inputs T. The training input Ti represents the plurality of sets of training diagnosis codes 151 (e.g., the ICD- 10 codes). The training input T2represents the plurality of admission department codes 155. The training input T3represents a plurality of sets of training procedure codes (e.g., the OPS codes) corresponding to the plurality of sets of training diagnosis codes 151. In some embodiments, the training input T2(i.e., the plurality of admission department codes 155) may be an alpha-numeric string while the training inputs T i and T3may be represented by lists of alpha-numeric strings. Other training inputs T4- Tmmay be age, weight, gender code, demographic information, etc. It should be understood that there can be other training inputs T based on application requirements. Input Y represents the plurality of training principal diagnosis codes 153.

[0078] A pre-processing module 161 processes the training input T2, i.e., the plurality of admission department codes 155, in the similar manner as done by the pre-processing module 146 (shown in FIG. 2). A pre-processing module 162 transforms the training inputs T (i.e., the training medical claims data 150) into an appropriate format to prepare the training inputs T for training the machine learning model 120. This processing is similar to the processing performed by the pre-processing module 144 (shown in FIG. 2). For example, the training inputs T are represented as strings and concatenated into a single text document. This pre-processing includes basic natural language processing (NLP) techniques, such as tokenization, stop-word removal, feature generation, and discrimination. In a preferred embodiment, textual input data is represented as a sparse TF-IDF matrix representation.

[0079] In some embodiments, the computer processor 104 is further configured to transform the plurality of training principal diagnosis codes 153 (i.e., the input Y) into a corresponding plurality of training partial codes 159 prior to training the machine learning model 120. The plurality of training partial codes 159 includes multiple training partial codes 160. For example, a pre-processing module 163 transforms the training principal diagnosis code 154 represented as “F41.9” into the training partial code 160 represented as “F41”. In some embodiments, the machine learning model 120 may be trained to predict a partial code (e.g., the plurality of training partial codes 159) of the corresponding training principal diagnosis code 154 given the TF-IDF representation of a corresponding medical case including at least the set of training diagnosis codes 152. Output from the pre-processing modules 162, 163 is fed to a module 164 which trains the machine learning model 120. The machine learning model 120 is then saved and stored.

[0080] In some embodiments, the computer processor 104 is further configured to determine the catalog 136 (schematically shown in FIG. 3) including, for each of the plurality of admission departments 157, an assignment frequency F of each of the plurality of training principal diagnosis codes 153. The assignment frequency F of each of the plurality of training principal diagnosis codes 153 represents a frequency of usage of the corresponding training principal diagnosis code 154 by the corresponding admission department 158.

[0081] For example, the plurality of training principal diagnosis codes 153 includes the training principal diagnosis codes 154-1, 154-2. 154-3, ..., 154-q (collectively, training principal diagnosis codes 154), where q is a positive integer corresponding to a total number of the training principal diagnosis codes 154. Further, the patient may be assigned to the admission departments 158-1, 158-2, 158-3, ..., 158-p (collectively, admission departments 158), where p is a positive integer corresponding to a total number of the admission departments 158. The assignment frequency Fpqrepresents the frequency of usage of the training principal diagnosis code 154-p in the admission department 158-q. A module 165 generates the catalog 136 based on the plurality of training principal diagnosis codes 153 and the plurality of admission departments 157. The catalog 136 is then saved and stored.

[0082] Table 1 shown below provides an example of the catalog 136.

[0083] Table 1

[0084] In some embodiments, the catalog 136 contains statistical data about how the plurality of training principal diagnosis codes 153 are used in the plurality of admission departments 157. Diagnosis codes from the plurality of sets of training diagnosis codes 151 that were never used as the training principal diagnosis code 154 or the diagnosis codes from the plurality of sets of training diagnosis codes 151 that are less frequent (e.g., below a threshold) are ignored for the purpose of the catalog 136. The plurality of training principal diagnosis codes 153 are used both for training the machine learning model 120 (after pre-processing by the pre-processing module 163) and for generating the catalog 136. In some embodiments, the catalog 136 is used by the computational model 130.

[0085] Referring to FIG. 3, in some embodiments, the catalog 136 may be prepared during training of the machine learning model 120. In some embodiments, the computer processor 104 is further configured to receive a mapping data 137 including a correlation between the plurality of training principal diagnosis codes 153 and a plurality of subsequent training principal diagnosis codes 138 of a subsequent time period ST. In some embodiments, the computer processor 104 is further configured to update one or more training principal diagnosis codes 154 from the plurality of training principal diagnosis codes 153 based on the plurality of subsequent training principal diagnosis codes 138.

[0086] For example, the machine learning model 120 may be trained using the training medical claims data 150 from the past. To exemplify, the machine learning model 120 may be trained on the training medical claims data 150 from a year (e.g., year 1) and the catalog 136 may be generated based on this data. In the subsequent time period ST (e.g., year 2), diagnosis codes and procedure codes may get updated. However, training data based on the updated codes may not be readily available for training the machine learning model 120. The plurality of training principal diagnosis codes 154 in the catalog 136 may need to be updated.

[0087] In such cases, the mapping data 137 may be used for updating the one or more training principal diagnosis codes 154 from the plurality of training principal diagnosis codes 153. The mapping data 137 may be used to translate input data, e.g.., diagnosis and procedure codes, from a coding system of the year 2 to a coding system of the year 1. For example, a new code “E66.07” (in the year 2) may be mapped to an old code “E66.02” (from the year 1).

[0088] In some embodiments, the computer processor 104 is further configured to remove one or more training principal diagnosis codes 154 from the plurality of training principal diagnosis codes 153 that are not present in the plurality of subsequent training principal diagnosis codes 138. Subsequently, the catalog 136 may be replaced with an adapted version that supports the codes from the present coding system. For instance, a code that is no longer used may be removed from the catalog 136. This may allow the disclosed system 100, including the machine learning model 120 and the updated catalog 136, to be used for subsequent years. This approach is not limited to bridging just the gap between consecutive years but may also be applied to larger gaps.

[0089] FIG. 5 is a schematic block diagram of the system 100, according to another embodiment of the present disclosure. In some embodiments, the computer processor 104 is further configmed to determine a plurality of vectors 166 corresponding to the training medical claims data 150. For example, a text corresponding to the plurality of sets of training diagnosis codes 151 may be translated into the corresponding plurality of vectors 166.

[0090] In some embodiments, the computer processor 104 is further configured to provide the plurality of vectors 166 to the machine learning model 120. Specifically, the computer processor 104 may utilize the plurality of vectors 166 to influence or otherwise modify the machine learning model 120. For example, the plurality of vectors 166 may be passed into the machine learning model 120 (e.g., a neural network) and one or more weights of the machine learning model 120 may be altered by virtue of processing the plurality of vectors 166.

[0091] In some examples, the plurality of vectors 166 corresponding to the training medical claims data 150 may be obtained via embedding techniques, such as Word2Vec (a natural language processing technique). In some embodiments, the computer processor 104 is further configured to form a plurality of separation criteria 170 for the plurality of training partial codes 159 within the machine learning model 120 based on the plurality of vectors 166. For example, each of the plurality of separation criteria 170 includes a separation hyperplane for the plurality of training partial codes 159. The machine learning model 120 may be trained to classify structured medical claims data (e.g., the structured medical claims data 109 shown in FIG. 1) and predict a partial code of the principal diagnosis.

[0092] FIG. 6 is a schematic block flow diagram illustrating the computational model 130 for determining the plurality of full codes 131 and the corresponding plurality of second confidence scores 133. Referring to FIGS. 1 and 6, in some embodiments, inputs for the computational model 130 include the plurality of diagnosis codes 111, the admission code 128, and the output from the machine learning model 120, i.e., the plurality of partial codes 121 and the plurality of first confidence scores 123. Task of the computational model 130 is to determine the plurality of full codes 131 from the plurality of diagnoses codes 111 based on the output of the machine learning model 120, which is just a partial code.

[0093] In some embodiments, the computational model 130 is further configmed to determine an overlap between the plurality of diagnosis codes 111 and an output of the machine learning model 120. Specifically, the computational model 130 is further configured to determine the overlap between the plurality of diagnosis codes 111 and the plurality of training principal diagnosis codes 153 from the catalog 136. In some embodiments, the output of the machine learning model 120 includes the plurality of training principal diagnosis codes 153. In some embodiments, the computational model 130 is further configured to select each full code 132 from the overlap, such that each full code 132 is one of the plurality of training principal diagnosis codes 153. Specifically, in some embodiments, the computational model 130 is further configured to remove all codes from the plurality of diagnosis codes 111 that are not potential principal diagnosis codes 114.

[0094] For example, a module 172 may determine the overlap between the plurality of diagnosis codes 111 and the plurality of training principal diagnosis codes 153 based on information stored in the catalog 136. The module 172 determines the overlap by performing a lookup from the catalog 136. Specifically, the module 172 may filter out one or more diagnosis codes 110 from the plurality of diagnosis codes 111 that are not present in the catalog 136 to select each full code 132. Therefore, each full code 132 is one of the plurality of diagnosis codes 111 that is also present in the plurality of training principal diagnosis codes 153. In some implementations, the principal diagnosis code 106 may be present in the plurality of diagnosis codes 111 from the encounter.

[0095] In some embodiments, the computational model 130 receives the diagnosis codes 110-1, 110- 2, 110-3, ..., 110-i (collectively, diagnosis codes 110) from the plurality of diagnosis codes 111, wherein i is a positive integer corresponding to a total number of the diagnosis codes 110. The module 172 filters the plurality of diagnosis codes 111 based on the overlap with the plurality of training principal diagnosis codes 153 to select the plurality of full codes 131. Specifically, the plurality of full codes 131 represent the diagnosis codes 110-1, 110-2, ..., 110-(i-k), where k is a positive integer corresponding to a total number of the diagnosis codes 110 that are not present in the plurality of training principal diagnosis codes 153.

[0096] In some embodiments, the computational model 130 is further configmed to determine, using the first confidence score 124 of the respective partial code 122, each second confidence score 134 of the respective full code 132 from the plurality of full codes 131. Specifically, a module 174 determines each second confidence score 134 by assigning the first confidence score 124 of the corresponding partial code 122 present in the corresponding full code 132. In some embodiments, the module 174 may first reduce each of the plurality of full codes 131 to a corresponding partial code before determining the corresponding second confidence score 134 by assigning (e.g., performing a lookup) the corresponding first confidence score 124 since the machine learning model 120 provides the plurality of first confidence scores 123 only for the plurality of partial codes 121. Thus, the plurality of first confidence scores 123 obtained from the machine learning model 120 may be assigned to the plurality of full codes 131.

[0097] In some embodiments, the computational model 130 is further configured to determine if two or more full codes 132 from the plurality of full codes 131 include the same partial code 122. In such a case, the second confidence scores 134 of the two or more full codes 132 may be same since the second confidence scores 134 of the two or more full codes 132 may be mapped to the first confidence score 124 of the same partial code 122 and the machine learning model 120 provides only one first confidence score 124 to each partial code 122. In some embodiments, the computational model 130 is further configured to modify the second confidence score 134 of each of the two or more full codes 132 such that the second confidence scores 134 are not equal to each other.

[0098] Specifically, a module 176 determines if two or more full codes 132 include the same partial code 122. In some embodiments, the computational model 130 (or the module 176) is further configured to determine, using the catalog 136, the assignment frequency F of each of the two or more full codes 132 for the admission department, e.g., using the admission code 128. In some embodiments, the computational model 130 (or the module 176) is further configured to rank each of the plurality of full codes 131 based on the respective modified second confidence scores 134. The above approach may allow the computational model 130 to perform a meaningful ranking of each of the plurality of full codes 131 based on the respective modified second confidence scores 134, even when the two or more full codes 132 include the same partial code 122. If the two or more full codes 132 do not include the same partial code 122, the computational model 130 may directly proceed to output the plurality of full codes 131 and the plurality of second confidence scores 133.

[0099] In some embodiments, the computational model 130 (or the module 176) is further configured to determine the full code 132 from the two or more full codes 132 having the highest assignment frequency F for the admission department. In some embodiments, the computational model 130 (or the module 176) is further configured to modify the second confidence score 134 of each of the two or more full codes 132 other than the full code 132 having the highest assignment frequency F by subtracting the second confidence score 134 by a weighted inverse of the assignment frequency F of the corresponding full code 132. In other words, the computational model 130 (or the module 176) may penalize the second confidence score 134 based on the assignment frequency F of the corresponding full code 132. In some embodiments, the penalty may be larger if the corresponding full code 132 has a lower assignment frequency F. Further, the second confidence score 134 of the full code 132 having the highest assignment frequency F may remain unchanged. It should be noted that this approach is described for exemplary purposes only and alternative approaches may also be utilized for modifying the second confidence score 134 of each of the two or more full codes 132 such that the second confidence scores 134 are not equal to each other.

[0100] Table 2 shown below provides an example of the plurality of full codes 131 and the plurality of second confidence scores 133. The plurality of second confidence scores 133 include second confidence scores 134-1, 134-2, 134-3. The plurality of first confidence scores 123 include first confidence scores 124-1, 124-2, 124-3.

[0101] Table 2

[0102] In the above table, the full codes 132 (after filtering) include [C79.83, C79.30, C79.50, D90.00, 131.30], The corresponding partial codes include [C79, C79, C79, D90, 131], All the partial codes are not unique codes, for instance the partial code “C79” is present three times. Once each full code 132 is transformed into the corresponding partial code, the computational model 130 assigns the first confidence scores 124 obtained from the machine learning model 120 to the corresponding partial codes (of each full code 132) using the partial codes 122, thereby obtaining the second confidence scores 134- 1, 134-2, 134-3. Thus, the same first confidence score 124-1 is assigned to the full codes [C79.83, C79.30, C79.50] since they include the same partial code “C79”.

[0103] Subsequently, the catalog 136 is used to obtain the assignment frequency F of each full code 132 in the respective admission department. Experiments have shown that the catalog 136 improves an overall performance of the system 100. The assignment frequency F may be used to adjust the second confidence scores 134. The second confidence score 134 of the full code 132 having the highest assignment frequency F remains unchanged. In the above table, the second confidence score 134-1 of the full code [C79.83] remains unchanged while the second confidence score 134-1 of the full codes [C79.30, C79.50] are penalized (or reduced) to obtain the respective modified second confidence scores 134-1(F2) and 134-1(F3). A maximum amount of the reduction (width) may be defined as a hyperparameter of the computational model 130.

[0104] FIG. 7 is a schematic block diagram of the system 100, according to another embodiment of the present disclosure. Some components of the system 100 are not shown for the purpose of illustration. In an alternative approach, the machine learning model 120 may itself and directly provide a plurality of full codes (e.g., the plurality of full codes 131 shown in FIG. 1) and a corresponding plurality of second confidence scores (e.g., the plurality of second confidence scores 133 shown in FIG. 1), instead of the plurality of partial codes (e.g., the plurality of partial codes 121 shown in FIG. 1) and the plurality of first confidence scores (e.g., the plurality of first confidence scores 123 shown in FIG. 1) as predictions.

[0105] In some implementations, the machine learning model 120 is further configured to generate a plurality of model full codes 180 and a corresponding plurality of third confidence scores 182. For example, the machine learning model 120 receives the structured medical claims data 109 and directly predicts the plurality of model full codes 180 and the corresponding plurality of third confidence scores 182 in a single step approach. In some embodiments, the plurality of model full codes 180 may be equivalent to the plurality of full codes 131 and the plurality of third confidence scores 182 may be equivalent to the plurality of second confidence scores 133.

[0106] Therefore, the ordered recommendation of the plurality of full codes (e.g., the plurality of full codes 131) may be determined directly by the machine learning model 120 as the plurality of model full codes 180 and there would be no need to apply the computational model 130. In such cases, the computational model 130 may be optional and may be used, e.g., to ensure that only those codes are proposed that are present in the plurality of diagnosis codes 111 (shown in FIG. 1) from the structured medical claims data 109, or to adjust the plurality of third confidence scores 182 according to expert knowledge. The described adaptation may be easily implemented by changing an output of the machine learning model 120 from the plurality of partial codes to the plurality of model full codes 180 without needing changes to the machine learning model 120. In some other embodiments, the machine learning model 120 may be trained specifically to predict the plurality of model full codes 180 and the corresponding plurality of third confidence scores 182.

[0107] In some embodiments, the computer processor 104 is further configured to output the plurality of model full codes 180 and the plurality of third confidence scores 182 to at least one of the user interface 140, the storage device 142, and the at least one non-transitory computer-readable storage medium 102. The plurality of model full codes 180 may be used to determine a principal diagnosis code (e.g., the principal diagnosis code 106 shown in FIG. 1) based on the plurality of third confidence scores 182. The plurality of third confidence scores 182 may be different from each other since the plurality of third confidence scores 182 are determined individually by the machine learning model 120.

[0108] FIG. 8 is a flowchart illustrating a method 200 for determining the principal diagnosis code 106. The method 200 may be implemented using the systems 100 of FIGS. 1-7. Referring now to FIGS. 1-8, at step 202, the method 200 includes receiving structured medical claims data 109 including at least the plurality of diagnosis codes 111 associated with the patient 108. In some embodiments, the structured medical claims data 109 further includes the plurality of procedure codes 125. In some embodiments, the structured medical claims data 109 corresponds to a current medical case and does not include historical hospitalization data associated with the patient 108. At step 204, the method 200 further includes providing the structured medical claims data 109 to the machine learning model 120. In some embodiments, the machine learning model 120 is a classification model.

[0109] At step 206, the method 200 further includes receiving the plurality of partial codes 121 and the corresponding plurality of first confidence scores 123 from the machine learning model 120. Each partial code 122 from the plurality of partial codes 121 may be a portion of the corresponding potential principal diagnosis code 114 from the plurality of potential principal diagnosis codes 113.

[0110] In some embodiments, the method 200 further includes training the machine learning model 120 using the training medical claims data 150 including the plurality of sets of training diagnosis codes 151 and the plurality of training principal diagnosis codes 153 corresponding to the plurality of sets of training diagnosis codes 151. In some embodiments, the training data medical claims 150 further includes the plurality of admission department codes 155 for the corresponding plurality of admission departments 157.

[0111] Each training principal diagnosis code 154 from the plurality of training principal diagnosis codes 153 is associated with the corresponding set of training diagnosis codes 152 from the plurality of sets of training diagnosis codes 151. Each training principal diagnosis code 154 is one training diagnosis code 152a included in the corresponding set of training diagnosis codes 152. In some embodiments, the method 200 further includes transforming the plurality of training principal diagnosis codes 153 into the corresponding plurality of training partial codes 159 prior to training the machine learning model 120. In some embodiments, the method 200 further includes determining the plurality of vectors 166 corresponding to the training medical claims data 150. In some embodiments, the method 200 further includes providing the plurality of vectors 166 to the machine learning model 120. In some embodiments, the method 200 further includes forming the plurality of separation criteria 170 of the plurality of training diagnosis partial codes 153 within the machine learning model 120 based on the plurality of vectors 166.

[0112] In some embodiments, the method 200 further includes determining the catalog 136 including, for each of the plurality of admission departments 157, the assignment frequency F of each of the plurality of training principal diagnosis codes 153. The assignment frequency F of each of the plurality of training principal diagnosis codes 153 represents the frequency of usage of the corresponding training principal diagnosis code 154 by the corresponding admission department 158.

[0113] In some embodiments, the method 200 further includes receiving the mapping data 137 including the correlationbetween the plurality of training principal diagnosis codes 153 and the plurality of subsequent training principal diagnosis codes 138 of the subsequent time period ST. In some embodiments, the method 200 further includes updating the one or more training principal diagnosis codes 154 from the plurality of training principal diagnosis codes 153 based on the plurality of subsequent training principal diagnosis codes 138. In some embodiments, the method 200 further includes removing the one or more training principal diagnosis codes 154 from the plurality of training principal diagnosis codes 153 that are not present in the plurality of subsequent training principal diagnosis codes 138.

[0114] At step 208, the method 200 further includes providing the plurality of partial codes 121 and the plurality of first confidence scores 123 to the computational model 130. In some embodiments, the computational model 130 is different from the machine learning model 120. In some embodiments, the computational model 130 is a statistical model.

[0115] At step 210, the method 200 further includes receiving the plurality of full codes 131 and the corresponding plurality of second confidence scores 133 from the computational model 130. The computational model 130 is configured to filter and rank, based at least in part on the plurality of second confidence scores 133, each of the plurality of full codes 131 from the plurality of diagnosis codes 111, such that each full code 132 includes the corresponding partial code 122 from the plurality of partial codes 121.

[0116] In some embodiments, the method 200 further includes determining the overlap between the plurality of diagnosis codes 111 and an output of the machine learning model 120. In some embodiments, the output of the machine learning model 120 includes the plurality of training principal diagnosis codes 153. In some embodiments, the method 200 further includes selecting each full code 132 from the overlap. In some embodiments, the method 200 fiirther includes receiving the admission code 128 of the admission department 158 associated with the patient 108. In some embodiments, the method 200 further includes using the predetermined default admission code 129 when the admission code 128 is not received.

[0117] In some embodiments, the method 200 further includes determining, using the first confidence score 124 of the respective partial code 122, each second confidence score 134 of the respective full code 132 from the plurality of full codes 131. In some embodiments, the method 200 further includes determining if two or more full codes 132 from the plurality of full codes 121 include the same partial code 122. In some embodiments, the method 200 further includes modifying the second confidence score 134 of each of the two or more full codes 132 such that the second confidence scores 134 are not equal to each other.

[0118] In some embodiments, the method 200 further includes determining, using the catalog 136, the assignment frequency F of each of the two or more full codes 132 for the admission department. In some embodiments, the method 200 further includes determining the full code 132 from the two or more full codes 132 having the highest assignment frequency F for the admission department. In some embodiments, the method 200 further includes modifying the second confidence score 134 of each of the two or more full codes 132 other than the full code 132 having the highest assignment frequency F by subtracting the second confidence score 134 by the weighted inverse of the assignment frequency F of the corresponding full code 132. In some embodiments, the method 200 further includes ranking each of the plurality of full codes 131 based on the respective modified second confidence scores 134.

[0119] At step 212, the method 200 further includes outputting the plurality of full codes 131 in an ordered recommendation based at least in part on the plurality of second confidence scores 133 to at least one of the user interface 140, the storage device 142, and the at least one non-transitory computer- readable storage medium 102. In some embodiments, the method 200 further includes outputting the plurality of second confidence scores 133 corresponding to the plurality of full codes 131. In some embodiments, the method 200 further includes selecting the principal diagnosis code 106 from the ordered recommendation based on the plurality of second confidence scores 133.

[0120] In some embodiments, the method 200 further includes generating, via the machine learning model 120, the plurality of model full codes 108 and the corresponding plurality of third confidence scores 182. In some embodiments, the method 200 further includes outputting, via the computer processor 104, the plurality of model full codes 180 and the plurality of third confidence scores 182 to at least one of the user interface 140, the storage device 142, and the at least one non-transitory computer-readable storage medium 102.

[0121] It should be understood that steps of the method 200 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 200 can be ended at any time. Referring to FIGS. 1-7, the system 100 and the method 200 of the present disclosure utilizes a hybrid approach in which the machine learning model 120 (i.e., a supervised machine learning classifier) first provides the plurality of partial codes 121 and the corresponding plurality of first confidence scores 123. Subsequently, the computational model 130 (i.e., a statistical model) provides the plurality of full codes 131 and the corresponding plurality of second confidence scores 133. Thus, the system 100 and the method 200 may divide the task of determining the principal diagnosis into two sub-tasks, i.e., prediction of the plurality of partial codes 121 via the machine learning model 120 and determination of the plurality of full codes 131 via the computational model 130. Each partial code 122 may be a portion of the corresponding potential principal diagnosis code 114. Hence, this approach may allow reduction in a complexity of the system 100 and the method 200 since a target space (i.e., a number of the partial codes 122) of the machine learning model 120 is reduced. This also improves a performance and an efficiency of the system 100 and the method 200 as compared to other approaches that directly predict the full code 132 of the principal diagnosis.

[0122] Mapping of the plurality of partial codes 122 to the plurality of full codes 131 is done by the computational model 130 using collected and aggregated data in the form of the catalog 136. The catalog 136 may be prepared during training of the machine learning model 120. In some cases, the catalog 136 may be used to filter the plurality of diagnosis codes 111. In some cases, the catalog 136 may be used to modify the second confidence scores 134 of the two or more full codes 132 where the second confidence scores 134 corresponding to the two or more full codes 132 are equal. This is relevant when the plurality of partial codes 122 proposed by the machine learning model 120 matches with the two or more full codes 132.

[0123] The system 100 and the method 200 may predict the plurality of full codes 131 using the structured medical claims data 109, without the usage of clinical text documentation. In general, the system 100 and the method 200 may be utilized to predict the principal diagnosis out of the structured medical claims data 109 during a stay of the patient 108 as well, e.g., to forecast required resources of a health care provider. In some embodiments, input data that is only available after discharge of the patient 108 may be excluded in this approach to account for desired usage during the patient's stay, where this information is not yet available.

[0124] The system 100 and the method 200 may automatically provide the ordered recommendation of the plurality of full codes 131 for determining the principal diagnosis given the structured medical claims data 109. The ordered recommendation is based on the plurality of second confidence scores 133. Proposing multiple reasonable candidates not only addresses an unambiguity of the principal diagnosis but may also help to optimize reimbursement (upcoding), or to satisfy the health care providers' different coding preferences.

[0125] In an alternative approach, the machine learning model 120 may be utilized to predict the plurality of model full codes 180 and the corresponding plurality of third confidence scores 182 directly in a single step, thereby eliminating the computational model 130. The system 100 and the method 200 may be directly integrated into an existing health care software as a new feature. Such integration may improve an acceptance of the software and increase a user satisfaction. Furthermore, the proposed approach may be used in allied applications as well, for instance, for reimbursement review by health care insurance companies. The proposed solution is data- driven and based on machine learning, which makes it easier to maintain and extend. Moreover, the system 100 and the method 200 provide higher quality recommendations than previous approaches.

[0126] In the above disclosure, all terms refer to German health care market, for instance, medical procedures are coded using the OPS classification. In principle, however, the proposed approach may be transferred to any market if diagnoses and procedures are encoded as ICD-10 codes and some procedure code system where identification of a main diagnosis is required. Moreover, the general idea may also be transferred to other hierarchical code representation systems.

[0127] 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.

[0128] 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.

[0129] 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.

[0130] 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. It should be appreciated that the innovations disclosed herein are generally adapted to configure computer-based systems to address technical problems related to determining a principal diagnosis, and in particular, the technical problem of determining a principal diagnosis from structured medical claims data. That is, the innovations described herein are inextricably linked to computing systems that generate structured medical claims data (e.g., for healthcare information systems).

[0131] For instance, according to particular implementations, receiving structured medical claims data for a single patient, each element of the data must be transformed into a high-dimensional vector representation before being processed by the machine learning and computational model. To obtain the predictions from the machine learning model alone, millions or even billions of computations are required, e.g., multiplication of the high-dimensional vector representations with the high dimensional model parameters. For instance, the GPT-3 model utilizes a model with more than 100 billion parameters. In other words, using a GPT-3 model (or something similarly sized), each single element from the structured medical claims data must be processed using more than 100 billion discrete calculations. Even in situations where the model is not as large as the GPT-3 Model, for the model to be sufficiently accurate, the model parameters must be determined using thousands or millions of examples of structured medical claims data, each involving the described computations. Thus, in order for a model (or models) to be sufficiently accurate to address the technical problems disclosed herein, no less than millions of discrete calculations must be performed to accurately predict a primary diagnosis from the structured medical claims. In short, the sheer magnitude of the parameters and calculations involved for each single case that is coded takes the disclosed techniques outside the realm of what is practicable for a human to perform (either in the mind of a human or with the aid of pen and paper) and inextricably links the technology to the realm of computers.

[0132] A number of advantages may be realized by implementing one or more aspects of the disclosed systems, methods, and techniques. For instance, utilizing conventional techniques, the principal diagnosis may be hard to determine, particularly for complex cases, which tends to result in ambiguities in the principal diagnosis. Providing an ordered recommendation of possible principal diagnoses and confidence scores enables the most suitable or most valuable principal diagnosis to be chosen from a set of valid diagnoses. This set is not only reduced to diagnoses that are allowed to be principal diagnosis in the given patient context but also highlights the most probable principal diagnoses such that the complexity of the given case is reduced and the ambiguity can be resolved.

[0133] Furthermore, the quality of coding is improved due at least in part to the predicative accuracy of the model and because the present disclosure provides an automatic, data-based second opinion for each coded case that is difficult, if not impracticable for a human coding specialist to emulate. For example, the present disclosure can be trained on a vast amount of medical claims data from multiple sources, e.g., healthcare providers, the trained models accumulate experience that exceeds the experience of even the most experienced medical coding specialist and it incorporates statistical knowledge on a global level, e.g., different coding preferences and reimbursement optimizations, that are generally not available to a medical coding specialist.

[0134] In short, aspects of the present disclosure improve the coding quality by providing recommendations that are valid, minimize or otherwise address the ambiguity inherent in principal diagnosis determination, and reduce the complexity of a case with a performance better than that achieved by conventional techniques.

[0135] As another example, unlike conventional techniques, the present disclosure is capable of adapting to changes in coding systems, preferences, and coding seasonality by training the models disclosed herein on structured medical data that reflects these changes. This data is naturally generated by healthcare providers during the process of reimbursement, hence generally available. For instance, other solutions may require additional training for medical coding specialists, adjustment / deletion / addition of new rules in a rule-based system or adjustments to the program code of a deterministic algorithm. Moreover, the knowledge required to update the systems / train medical coding specialists must first be manually extracted and processed by experts before it can be used to adapt the existing solutions. The present disclosure suffers from none of these disadvantages. Furthermore, aspects of the present disclosure, like the machine learning model or the computational model may be updated within a very short period of time, e.g., horns, compared to other solutions that require days to weeks. Additionally, the mapping correlations described in the present disclosure allow the update of the present disclosure even if no corresponding new training data is available yet.

[0136] For these and other reasons, experiments using the present disclosure have yielded substantial improvements over other conventional techniques. For instance, the disclosed techniques have been evaluated against a number of different approaches, including heuristic selection of the principal diagnosis, pattern matching approaches with data-driven dictionaries and even deep learning architectures. All these approaches were outperformed by the current hybrid model (using both a machine learning model and a computational model as disclosed herein), often by a huge margin. For instance, the present disclosure achieves a predictive accuracy rate of more than 83%, improving the performance of other evaluated techniques for the task of determining a principal diagnosis by about 5-10%. The disclosed techniques also improve performance compared to the same machine learning model that is trained to directly process full-codes and second confidence scores without application of a computational model. That is, it is the hybrid model (or the combination of using both the machine learning model and the computational model) that achieves the largest improvement in the predictive accuracy.

[0137] Various examples have been described. These and other examples are within the scope of the following claims.

Claims

CLAIMS1. A system for determining a principal diagnosis code, the system comprising: at least one non-transitory computer-readable storage medium having instructions stored thereon; and at least one computer processor coupled to the at least one non-transitory computer- readable storage medium and configured to execute the instructions to: receive structured medical claims data including at least a plurality of diagnosis codes associated with a patient; provide the stmctured medical claims data to a machine learning model; receive a plurality of partial codes and a corresponding plurality of first confidence scores from the machine learning model, wherein each partial code from the plurality of partial codes is a portion of a corresponding potential principal diagnosis code from a plurality of potential principal diagnosis codes; provide the plurality of partial codes and the plurality of first confidence scores to a computational model; receive a plurality of full codes and a corresponding plurality of second confidence scores from the computational model, wherein the computational model is configured to fdter and rank, based at least in part on the plurality of second confidence scores, each of the plurality of full codes from the plurality of diagnosis codes, such that each full code comprises a corresponding partial code from the plurality of partial codes; and output the plurality of full codes in an ordered recommendation based at least in part on the plurality of second confidence scores to at least one of a user interface, a storage device, and the at least one non-transitory computer-readable storage medium.

2. The system of claim 1, wherein the computer processor is further configured to train the machine learning model using training medical claims data comprising a plurality of sets of training diagnosis codes and a plurality of training principal diagnosis codes corresponding to the plurality of sets of training diagnosis codes, wherein each training principal diagnosis code from the plurality of training principal diagnosis codes is associated with a corresponding set of training diagnosis codes from the plurality of sets of training diagnosis codes, and wherein each training principal diagnosis code is one training diagnosis code included in the corresponding set of training diagnosis codes.

3. The system of claim 2, wherein the computer processor is further configmed to transform the plurality of training principal diagnosis codes into a corresponding plurality of training partial codes prior to training the machine learning model.

4. The system of claim 3, wherein the computer processor is further configured to: determine a plurality of vectors corresponding to the training medical claims data; provide the plurality of vectors to the machine learning model; and form a plurality of separation criteria for the plurality of training partial codes within the machine learning model based on the plurality of vectors.

5. The system of claim 2, wherein the training medical claims data further comprises a plurality of admission department codes for a corresponding plurality of admission departments.

6. The system of claim 5, wherein the computer processor is further configured to determine a catalog comprising, for each of the plurality of admission departments, an assignment frequency of each of the plurality of training principal diagnosis codes, and wherein the assignment frequency of each of the plurality of training principal diagnosis codes represents a frequency of usage of the corresponding training principal diagnosis code by the corresponding admission department.

7. The system of claim 6, wherein the computer processor is further configured to: receive a mapping data comprising a correlation between the plurality of training principal diagnosis codes and a plurality of subsequent training principal diagnosis codes of a subsequent time period; and update one or more training principal diagnosis codes from the plurality of training principal diagnosis codes based on the plurality of subsequent training principal diagnosis codes.

8. The system of claim 7, wherein the computer processor is further configured to remove one or more training principal diagnosis codes from the plurality of training principal diagnosis codes that are not present in the plurality of subsequent training principal diagnosis codes.

9. The system of claim 1, wherein the computational model is further configured to: determine an overlap between the plurality of diagnosis codes and an output of the machine learning model; and select each full code from the overlap.

10. The system of claim 9, wherein the output of the machine learning model comprises a plurality of training principal diagnosis codes.

11. The system of claim 6, wherein the computer processor is further configured to receive an admission code of an admission department associated with the patient, and wherein when the admission code is not received, the computer processor is further configured to use a predetermined default admission code.

12. The system of claim 11, wherein the computational model is further configured to: determine, using the first confidence score of the respective partial code, each second confidence score of a respective full code from the plurality of full codes; determine if two or more full codes from the plurality of full codes comprise the same partial code; modify the second confidence score of each of the two or more full codes such that the second confidence scores are not equal to each other; and rank each of the plurality of full codes based on the respective modified second confidence scores.

13. The system of claim 12, wherein the computational model is further configured to: determine, using the catalog, the assignment frequency of each of the two or more full codes for the admission department; determine the full code from the two or more full codes having the highest assignment frequency for the admission department; and modify the second confidence score of each of the two or more full codes other than the full code having the highest assignment frequency by subtracting the second confidence score by a weighted inverse of the assignment frequency of the corresponding full code.

14. The system of claim 1, wherein the structured medical claims data further includes a plurality of procedure codes.

15. The system of claim 1, wherein the structured medical claims data corresponds to a current medical case and does not include historical hospitalization data associated with the patient.

16. The system of claim 1, wherein the computer processor is further configured to select a principal diagnosis code from the ordered recommendation based on the plurality of second confidence scores.

17. The system of claim 1, wherein the computer processor is further configured to output the plurality of second confidence scores corresponding to the plurality of full codes.

18. The system of claim 1 , wherein the machine learning model is a classification model.

19. The system of claim 1, wherein the computational model is a statistical model.

20. The system of claim 1, wherein the computational model is different from the machine learning model.

21. The system of claim 1, wherein the machine learning model is further configured to generate a plurality of model full codes and a corresponding plurality of third confidence scores, and wherein the computer processor is further configured to output the plurality of model full codes and the plurality of third confidence scores to at least one of the user interface, the storage device, and the at least one non-transitory computer-readable storage medium.

22. A method for determining a principal diagnosis code, the method comprising: receiving structured medical claims data including at least a plurality of diagnosis codes associated with a patient; providing the structured medical claims data to a machine learning model; receiving a plurality of partial codes and a corresponding plurality of first confidence scores from the machine learning model, wherein each partial code from the plurality of partial codes is a portion of a corresponding potential principal diagnosis code from a plurality of potential principal diagnosis codes; providing the plurality of partial codes and the plurality of first confidence scores to a computational model; receiving a plurality of full codes and a corresponding plurality of second confidence scores from the computational model, wherein the computational model is configured to fdter and rank, based at least in part on the plurality of second confidence scores, each of the plurality of full codes from the plurality of diagnosis codes, such that each full code comprises a corresponding partial code from the plurality of partial codes; andoutputting the plurality of full codes in an ordered recommendation based at least in part on the plurality of second confidence scores to at least one of a user interface, a storage device, and the at least one non-transitory computer-readable storage medium.

23. The method of claim 22, further comprising training the machine learning model using training medical claims data comprising a plurality of sets of training diagnosis codes and a plurality of training principal diagnosis codes corresponding to the plurality of sets of training diagnosis codes, wherein each training principal diagnosis code from the plurality of training principal diagnosis codes is associated with a corresponding set of training diagnosis codes from the plurality of sets of training diagnosis codes, and wherein each training principal diagnosis code is one training diagnosis code included in the corresponding set of training diagnosis codes.

24. The method of claim 23, further comprising transforming the plurality of training principal diagnosis codes into a corresponding plurality of training partial codes prior to training the machine learning model.

25. The method of claim 24, further comprising: determining a plurality of vectors corresponding to the training medical claims data; providing the plurality of vectors to the machine learning model; and forming a plurality of separation criteria for the plurality of training partial codes within the machine learning model based on the plurality of vectors.

26. The method of claim 23, wherein the training data medical claims further comprises a plurality of admission department codes for a corresponding plurality of admission departments.

27. The method of claim 26, further comprising determining a catalog comprising, for each of the plurality of admission departments, an assignment frequency of each of the plurality of training principal diagnosis codes, wherein the assignment frequency of each of the plurality of training principal diagnosis codes represents a frequency of usage of the corresponding training principal diagnosis code by the corresponding admission department.

28. The method of claim 27, further comprising: receiving a mapping data comprising a correlation between the plurality of training principal diagnosis codes and a plurality of subsequent training principal diagnosis codes of a subsequent time period; andupdating one or more training principal diagnosis codes from the plurality of training principal diagnosis codes based on the plurality of subsequent training principal diagnosis codes.

29. The method of claim 28, further comprising removing one or more training principal diagnosis codes from the plurality of training principal diagnosis codes that are not present in the plurality of subsequent training principal diagnosis codes.

30. The method of claim 22, further comprising: determining an overlap between the plurality of diagnosis codes and an output of the machine learning model; and selecting each full code from the overlap.

31. The method of claim 30, wherein the output of the machine learning model comprises a plurality of training principal diagnosis codes.

32. The method of claim 27, further comprising: receiving an admission code of an admission department associated with the patient; and using a predetermined default admission code when the admission code is not received.

33. The method of claim 32, further comprising: determining, using the first confidence score of the respective partial code, each second confidence score of a respective full code from the plurality of full codes; determining if two or more full codes from the plurality of full codes comprise the same partial code; determining, using the catalog, the assignment frequency of each of the two or more full codes for the admission department; determining the full code from the two or more full codes having the highest assignment frequency for the admission department; modifying the second confidence score of each of the two or more full codes such that the second confidence scores are not equal to each other; and ranking each of the plurality of full codes based on the respective modified second confidence scores.

34. The method of claim 33, further comprising:determining, using the catalog, the assignment frequency of each of the two or more full codes for the admission department; determining the full code from the two or more full codes having the highest assignment frequency for the admission department; and modifying the second confidence score of each of the two or more full codes other than the full code having the highest assignment frequency by subtracting the second confidence score by a weighted inverse of the assignment frequency of the corresponding full code.

35. The method of claim 22, wherein the structured medical claims data further includes a plurality of procedure codes.

36. The method of claim 22, wherein the structured medical claims data corresponds to a current medical case and does not include historical hospitalization data associated with the patient.

37. The method of claim 22, further comprising selecting a principal diagnosis code from the ordered recommendation based on the plurality of second confidence scores.

38. The method of claim 22, further comprising outputting the plurality of second confidence scores corresponding to the plurality of full codes.

39. The method of claim 22, wherein the machine learning model is a classification model.

40. The method of claim 22, wherein the computational model is a statistical model.

41. The method of claim 22, wherein the computational model is different from the machine learning model.

42. The method of claim 22, further comprising: generating, via the machine learning model, a plurality of model full codes and a corresponding plurality of third confidence scores; and outputting, via the computer processor, the plurality of model full codes and the plurality of third confidence scores to at least one of the user interface, the storage device, and the at least one non-transitory computer-readable storage medium.

Images