The present disclosure is related generally to computer network architectures for machine learning and artificial intelligence, and more specifically, to computer network architectures for the automated completion of estimated final cost data for claims for healthcare clinical episodes using incomplete data for healthcare insurance claims and costs, known to date.
This architecture enables enable a computer to learn by its own experience with users and by the passage of time, thus enabling computer hardware to perform better and produce superior results for the user. Various embodiments comprise practical applications in the healthcare field and constitute a healthcare robot with artificial intelligence.
In the healthcare industry, healthcare insurance claims are submitted, adjudicated and finalized using a process that can take a great deal of time after the related services are rendered. In this process, the insured party submits claims to an insurance company. These claims are adjudicated by the insurance company or other parties, and the claims are often adjusted. When the claims are finalized at the end of the process, the insurance company pays that amount. This final claims amount may be paid directly to the healthcare provider that provided the healthcare to the insured, or the final amount may be paid to the insured as a reimbursement. Some claims are finalized and paid quickly, while other can take more than a year. This duration is dependent upon, among other factors, the type of facility or practitioner performing the services, the billing practices of the associated facility, and the particular insurance payer involved.
This delay in final adjudication means that at any point in time, known historical claims for any given historical period are incomplete (i.e., not finally adjudicated) to some degree. It also means that the more time elapses, the more mature the picture of historical claims becomes (i.e., more claims are finally adjudicated).
For business analysis and risk management purposes, any entity bearing financial risk for the cost of a patient population's care is interested not just in total cost known to date (i.e., claims finally adjudicated), but also in anticipated total cost once all claims are fully mature (i.e., all claims are finally adjudicated).
There is a longstanding frustrated need to have programmable hardware function in a better manner, and particularly to have hardware be smarter and learn from its own experience, with machine learning and artificial intelligence.
The longstanding frustrated need is particularly acute in the healthcare field. For example, in healthcare there is a need for computer architecture to accurately forecast aspects of healthcare to provide practical applications for healthcare providers. Systems that generate such forecasts automatically without human intervention would, of course, operate more efficiently, reliably, and faster than systems requiring constant human input to make such practical applications.
The present disclosure is related generally to computer network architectures for machine learning, and more specifically, to computer network architectures for the automated completion of estimated final cost data for claims for healthcare clinical episodes using incomplete data for healthcare insurance claims and costs, known to date.
Embodiments of the present invention automatically project, estimate or predict future mature claim totals (i.e., eventual finalized claims), and generate and transmit reports of the future total claims to users. These reports of future claims are based on the type of healthcare facility providing the healthcare and based on the type of clinical episode, using the applicable data for claims known to date (which may include some claims made but not adjudicated and finalized, and some claims adjudicated and finalized).
The present disclosure is related generally to computer network architectures for machine learning and artificial intelligence that enable a computer to learn by its own experience with users and by the passage of time, thus enabling superior performance by the computer hardware. Various embodiments apply to the healthcare field and constitute a healthcare robot with artificial intelligence in practical applications to healthcare. Various embodiments perform certain steps automatically without human intervention, and hence are more efficient, reliable and faster than systems without these automatic aspects.
The present disclosure is related generally to computer network architectures for machine learning, and more specifically, to computer network architectures for the automated completion of estimated final cost data for healthcare clinical episodes using incomplete data for healthcare insurance claims and costs, known to date.
The following, with reference to
Embodiments automatically use the pipeline of healthcare outcome models to select and apply the optimal predictive model for claims to be finalized in the future.
In an embodiment, a computer network architecture with automated claims completion, artificial intelligence and machine learning, comprises (1) a prediction module with a prediction generator and an updated database, (2) a learning module with a training submodule, in electronic communication with the prediction module and the updated database, (3) an automatic claims completion web application ACCWA in electronic communication with the prediction module, the learning module, the updated database, and a user device, and wherein (4) the web application is adapted to receive and respond to user requests, and to automatically generate estimated final cost data for claims for healthcare clinical episodes based on incomplete data for healthcare insurance claims and costs.
The ACCWA automatically collects data for the patient episodes from a common prior time period, from various data bases which may include third-party data bases 4109 or existing data bases in the embodiment 114, 132, 152. The ACCWA loads the data into data bases 114, 132, 150 in the embodiment.
The ACCWA accesses a prediction application 106, which generates and transmits a request 192 to a protection module 110, to automatically fit a statistical model to the data, to predict final completed claims. The prediction model 110 automatically saves the resulting prediction model definition, and transmits a response 192 to the prediction application 106 with the prediction model definition. The prediction application 106 applies the model to the data and generates and transmits the complete final claim predictions to the ACCWA.
The ACCWA generates a report of the final claim predictions and transmits the report to the user device in response 182.
In an example of the automatic claims completion process, for which the computer network architecture is structured, the ACCWA may receive 801 a user request from a user device for automated completion of claims for a number of pending incomplete patient clinical episodes, that is, a request for a prediction of the final total of claims for a number of currently incomplete patient care episodes.
The ACCWA automatically collects 802 claims completion data for patient episodes, for complete and incomplete data sets for patient episodes.
The ACCWA loads 803 the collected data into the embodiment into various databases in the embodiment.
The ACCWA matches 804 complete data sets with incomplete data sets for similar patient care episodes.
The ACCWA accesses 805 the prediction application 106, which accesses the prediction module 110, which fits a statistical model to the collected data.
The prediction module then saves 806 the prediction model definition.
The ACCWA collects and loads 807 other incomplete data sets from various data sources.
The prediction application automatically applies 808 the model to the incomplete data sets and predicts final complete claims for the currently incomplete data sets for patient clinical episodes.
The prediction application transmits 809 the complete claims predictions for currently pending patient clinical episodes, to the ACCWA.
The ACCWA generates 810 a report of the predicted complete claims and transmits the report to the user device 108.
Three Layers
Embodiments of the present invention may be viewed as having an underlying data science layer. In the machine learning prediction system 104, this may be viewed as the first layer. This layer automatically adds new data streams (data capture) from all sources to the prediction database 114, learning database 132, and training database 152. This large expanding database contains all the data captured for all patients. The prediction application 104 of the system makes predictions (e.g., the patient risk profile for a patient with a patient risk scoring web application as the prediction application 104) based on the databases. The predictions can be subtle with a large number of parameters and detailed demographic disaggregation of the patient groups.
The prediction web application 106 may be thought of as a second layer that operates on top of the first layer 104, making requests 190 to the first layer, and receiving responses 192 back.
An Automatic Claims Completion Web Application ACCWA 180 may operate together as a third layer on top of the second layer.
The present disclosure describes a novel network architecture configured to produce and automatically improve predictive models, particularly for patient healthcare outcomes regarding for example, cost, quality and patient satisfaction. According to various embodiments, the network architecture allows automatic machine learning as new data sources are added, and new data is collected, and the predictive algorithms are recalibrated and reselected using the expanded, and hence more reliable, data. This architecture addresses, for example, new waste identification problems in healthcare without having to proportionally increase labor cost. The above may enable users of the architecture to quickly realize the value of new data.
Utilization of the novel architecture described herein may enable more efficient operation, hosting, maintenance, integration and use of new data and data sources, and continuous automatic improvement of predictive models by recalibration and reselection of algorithms based on the expanding data.
Machine learning may incorporate predictive model algorithms to execute predictive analytical operations. Learning may be supervised or unsupervised. In general, a predictive model analyzes historical data to identify patterns in the data. The patterns identified may include relationships between various events, characteristics, or other attributes of the data being analyzed. Modeling of such patterns may provide a predictive model whereby predictions may be made. Development of predictive models may employ mathematical or statistical modeling techniques such as curve fitting, smoothing, and regression analysis to fit or train the data. Such techniques may be used to model the distribution and relationships of the variables, e.g., how one or more events, characteristics, or circumstances (which may be referred to as “independent variables” or “predictor variables”) relate to an event or outcome (which may be referred to as a “dependent variable” or “response”).
A machine learning process may include developing a predictive model. For example, a dataset comprising observed data may be input into a modeling process for mapping of the variables within the data. The mapped data may be used to develop a predictive model. The machine learning process may also include utilizing the predictive model to make predictions regarding a specified outcome that is a dependent variable with respect to the predictive model. The machine may then be provided an input of one or more observed predictor variables upon which the output or response is requested. By executing the machine learning algorithm utilizing the input, the requested response may be generated and outputted. Thus, based on the presence or occurrence of a known predictor variable, the machine learning algorithm may be used to predict a related future event or the probability of the future event.
The present invention applies generally to predictive algorithms informed by data. When the data relates to medical patients and treatments, and the algorithms correlate patient data with treatments and healthcare outcomes, then the system produces a pipeline of predictive models for healthcare outcomes that constantly improve and evolve with the machine learning and artificial intelligence of the present invention. For clarity and economy of description, the terms machine learning and artificial intelligence are often used interchangeably herein.
Machine Learning and Artificial Intelligence Architecture
The network architecture 100 may also include one or more prediction applications 106. In some embodiments, the prediction application 106 includes one or more web applications. The prediction application 106 or web applications thereof may include or access episode request templates for defining prediction requests. Prediction applications 106 may be generic for any end user, semi-generic, or customized for a single end user.
In one embodiment, the web applications (i.e., prediction applications), the prediction module, and the learning module may reside in the cloud and be administered by the invention's operator. Web applications may be structured for specific users, classes of users, templates or related data and predictions. In the healthcare industry, web applications may seek prediction reports for outcomes of specific patient episodes, and focus on factors such as patient guidance, automated healthcare performance benchmarks, estimates of missing clinical episode costs related to incomplete health insurance claims, patient risk scoring, forecasting episode and annual costs, facilities rating systems adjusted for risk, and other points.
In a healthcare embodiment, when an end user provides a request for a prediction report, the end user may give specific patient data (e.g., patient metrics and treatment) to populate a selected episode profile or template in a selected web application. The user requests a prediction report, e.g., a prognosis (i.e., predicted outcomes). Other web applications may give other reports. e.g., possible treatments, probability of each possible result of each treatment choice, economics, and schedule of prognosis and treatment. The web application then automatically accesses the prediction module and transmits the episode data with the request, and receives back a prediction report from the prediction module.
Prediction applications 106 may be configured to interface one or more user devices 108 with the machine learning prediction system 104. The user device 108 may access or run the prediction application 106, which may be stored locally, remotely, or reside in the cloud. For example, the user device 108 may remotely access the prediction application 106 with a browser through a network connection, such as an Internet, extranet, or VPN connection. User devices 108 may include an electronic communication device such as a computer, laptop, PDA, tablet, wearable computing device, or smart phone. The user device 108 will typically function as an end point with respect to the network architecture 100 and interface users with the network architecture 100. Users may include customer or client end users, which may be in a commercial environment such as in a SaaS environment. In an example healthcare application, a user or user device 108 may be associated with a customer or customer grouping, which may be a specific hospital, doctor office, insurance company, or grouping of multiple hospitals, doctor offices, or insurance companies, for example.
The machine learning prediction system 104 may include a prediction module 110 and a learning module 130. The prediction module 110 may include a prediction generator 112 configured to process prediction requests and generate prediction responses with respect to specified outcomes. For example, prediction requests, which may also be referred to as report requests, may include episode data comprising observed data related to a response variable. Using the episode data, the prediction generator 112 may execute machine learning algorithms comprising prediction models stored in a prediction database 114.
The learning module 130 may be configured to execute computer learning tasks including generating, calibrating, updating and evaluating system prediction models. The learning module 130 may include a training submodule 150 configured to execute training functions, which may include executing training requests, to fit data comprising observed episode data to one or more model types defined in the machine learning prediction system 104.
The machine learning prediction system 104 may include one or more databases, each of which may comprise one or more member databases, for storing episode data, metrics, model types, current best predictive models, archived predictive models, and other data from which the machine learning prediction system 104 may access to perform system operations. For example, the training submodule 150 may include a training database 152 that stores observed episode data for training system prediction models. The training database 152 may collect and accumulate data over time for use by the training submodule 150. In a healthcare application, for example, the training database 152 may include data relating to individual patients, episodes, and outcomes. In some embodiments, the training database 152 includes a customer's institutional data, which the training submodule 150 may access. In one embodiment, the customer's institutional data is restricted to use only for the customer.
These third-party databases may include medical claims data, prescription refill data, publicly available social media data, credit agency data, marketing data, travel website data, e-commerce website data, search engine data, credit card data, credit score and credit history data, lending data, mortgage data, financial data, travel data, geolocation data, telecommunications usage data, and other third-party data bases. This unique combination of third-party databases gives the system 100 an edge in the quality of the predictive models 142 used and in the forecasts produced. Further embodiments may also use this unique combination of third-party data, further combined with data generated by users 108 of the system 100 prediction applications 106, to further enhance the quality of the predictive models 142 and the forecasts produced.
The training database 152 may also include third party databases that may be used by the training submodule 150 to generate, update, or recalibrate system predictive models. Thus, the training submodule 150 may access or integrate third party databases. In some embodiments, the training database 152 includes one or more database repositories for medical data and socioeconomic data, such as financial credit histories from credit rating companies. The training database 152 may reside in a single or multiple locations, be local or distributed, and may include third party databases accessible by the learning module 130 or training submodule 150. In one embodiment, the training database 152 may be divided or physically or logically partitioned into episode specific data libraries.
A library of system prediction models and associated metrics generated by the training submodule 150 may be stored in a learning database 132. The library may be used for evaluation of the archived models. For example, the learning database 132 may include multiple generated predictive models types of each of a plurality of specific episodes types. The learning module 130 may periodically or upon the occurrence of a predefined event, such as addition of new data or generation of new predictive models, compare metrics across the predictive models in the library to determine if a predictive model for a specific episode performs better than the predictive model for the episode that is stored in the prediction database 114, which the prediction module 110 currently uses for responding to prediction requests for the episode. If, based on the metrics comparison, the learning module 130 determines that another predictive model performs better than the one stored in the prediction database 114, the learning module 130 may replace the predictive model with the model determined to be the current best. The learning database 132 may maintain a record of the metrics and associated predicative models for future evaluations. In some embodiments, the learning module 130 may test or update the regression calibration of one or more algorithm to be used, and add episode data to the database to be kept and expanded and used by this and all subsequent queries.
The Administration Module 170 and Administrative Application 172
In various embodiments, the network architecture 100 or machine learning prediction system 104 may also include an administration module 170. The administration module 170 may be configured to allow systems administrators to develop and control the machine learning prediction system 104, such as defining episode types. For example, system administrators may define patient episode types such as types of patient conditions to be handled, their metrics, and training databases 152 used. The machine learning prediction system 104 may be employed for machine learning and predictive analysis for any episode type. Episode types may include, for example, “hip and knee replacements under Medicare rules” or “pneumonia under Aetna Insurance” rules. The administration module 170 may provide, incorporate, or update episode type definitions, e.g., types of patients, metrics and parameters, and illness types. The administration module 170 may also allow system administrators to define which training database 152 or prediction database 114 each episode type will use.
In some embodiments, system administrators may utilize the administration module 170 to add or integrate additional databases, including third party databases, into the prediction database 114, learning database 132, and training databases 152. In the above or another embodiment, system administrators may utilize the administration module 170 to define which data each patient episode profile type will use. For example, utilizing the administration module 170, system administrators may define which parameters or attributes, such as predictor variables, associated with observed episode data is to be used for generating predictive models or predictions for specific episode types. An administrative application 172 (see, e.g.,
Further Network Architecture
As introduced above with respect to
The learning module 130 or training submodule 150 thereof may include a model type library. In the illustrated embodiment, the training database 152 includes a model type library 153. In various embodiments, model types (sometimes referred to as algorithms) may include one or more of multi-level models, random forest regression models, logistical regression models, gamma-distributed regression models, linear regression models, or combinations thereof. Other model types may be used. In one embodiment, the learning module 130 may be configured to expand utilization of additional model types or redefine or update current model types.
In various embodiments, the learning mode 130 may be configured to automatically, such as periodically or upon the occurrence of a system event, generate training requests. For example, system events may include addition of accessible episode data corresponding to specific episodes, such as addition of or updating of data in the training database 152. The learning module 130 may generate training requests that include all, multiple, or portions of one, more, or all datasets in an episode specific training database 152a, 152b, 152c. For example, the learning module 130 may include a scheduling unit 136 configured to automatically initiate a training process to test or update calibration of the predictive models. For example, training may be initiated with each query, number or volume of queries, periodically, upon introduction of new data, or other time period or event.
In some embodiments, system administrators may submit requests to the learning module 130 instead of or in addition to automatic request generation. In the illustrated embodiment, an administer device 109 interfaces with the scheduling unit 136 via an administrator application 172 to initiate the training process. The administrator device 109 may access or run the administrator application 172, which may be stored locally, remotely, or reside in the cloud. In one embodiment, the administrative application 172 comprises a web application remotely accessible by the administrator device via a browser through a network connection, such as an Internet, extranet, or VPN connection. The administrator device 109 may include an electronic communication device such as a terminal, computer, laptop, PDA, tablet, or smart phone. The training request initiated by the administrator device 109 via the administrator application 172 may include a request template providing a model definition and dataset.
The model definition may identify parameters and episode attributes to use for model fitting in the training process. The dataset may include observed episode data including attribute values of predictor variables and response variables. In some embodiments, the dataset provided in the request may identify all or a portion of episode data stored in the training database 152 for the episode type that corresponds to the model definition, which may be in addition to or instead of transmission of new data in the request.
As introduced above, the dataset may include observed episode attribute values from a plurality of episodes corresponding to the episode type indicated in the model definition. For example, the model definition may specify post-op knee replacement prognosis (Y) and identify attributes, e.g., age, weight, education, income, hospital location, procedure variation, or other attributes (X), to be used in the modeling. The dataset may then include a plurality of data points comprising episode attribute values y0, x01, . . . , x0n corresponding to the model definition. In various embodiments, multiple pairs of model definitions and datasets may be submitted to the training submodule 150, e.g., a list of requests comprising pairs of model definitions and datasets.
In various embodiments, the model definition may be generic to multiple types of models in the model type library 153. For example, the model definition may be applicable to two, more, or all types of models or category of types of models corresponding to an episode type. The model definition may directly or indirectly identify applicable model types. Alternatively or additionally, the learning module 130 may be configured such that model definitions are assumed to be applicable to all system model types unless indicated otherwise.
The training submodule 150 may include a model processing unit 154 configured to process training requests and generate predictive models according to the requests. The model processing unit 154 may employ one or more worker processes 156a, 156b, 156c to generate a fitted model, per the parameters of the request. Worker processes 156a, 156b, 156c may include modeling software programs, software functions, processes, or subroutines, for example. The model processing unit 154, or worker processes 156a, 156b, 156c thereof, may access the model type library 153 to obtain executable instructions for generating one or more predictive models for one or more of the model types using the dataset specified in the request.
In various embodiments, the model processing unit 154 may farm each model definition/dataset pair to a worker process 156a, 156b, 156c to generate fitted models. In one example of the training process, one or more worker process 156a, 156b, 156c may each be tasked with generating a fitted model, per the parameters of the request. As introduced above, the machine learning prediction system 104 may be configured to generically support multiple types of models. Accordingly, multiple fitted models may be generated corresponding to one or more model types. In some embodiments, each worker process 156a, 156b, 156c may run a script, such as a Python script, to generate respective fitted models. In one embodiment, worker processes 156a, 156b, 156c may access or be provided data specified in the request comprising one or more datasets or portions thereof from one or more specified episode specific training databases 152a, 152b, 152c. Alternately, worker processes 156a, 156b, 156c may be provided new or additional data identified by or included in the dataset of the training request. For example, the new or additional data may be accesses or obtained from a third party database. In some instances, all or a portion of the new or additional data may be retained and stored in the training database 152 and thereby expand available episode data for future fine tuning of the system's 104 episode modeling algorithms.
As introduced above, the model processing unit 154 may test generated predictive models, the metrics of which may be stored in the learning database 132. In the illustrated embodiment, the worker processes 156a, 156b, 156c test the fitted models generated from the episode definition/dataset pair in the request. The testing process may include one or more metric evaluations or other testing or evaluation schemes. In one example, worker processes 156a, 156b, 156c may perform k-fold cross-validation to estimate prediction ability of the generated fitted models. In a further example, k-fold cross-validation may include testing of a model via 5-fold cross-validation. K-fold cross-validation may include partitioning a dataset into k different partitions or folds. Fitting the model may therefore include training k−1 subsets as training data consistent with a model type and then testing the trained model with the remaining subset as a validation or test data. This may be repeated until all k subsets have been used as the validation or test data. Calculated cross-validation errors across the folds may be averaged. K-fold cross-validation may also be used to select the fitted model that is likely to be the most predictive with out-of-sample data.
The results of the model fitting and testing may be calculated by the model processing unit 154. For example, worker processes 156a, 156b, 156c may calculate evaluation metrics descriptive of the generated models, which may be used to compare models. Evaluation metrics may include, for example, one or more metrics selected from RMSE, mean bias, mean absolute error, area under the curve on a receiver operating characteristic plot, or combinations thereof. Other evaluation metrics may be used. For example, suitable evaluation metrics may be selected that correspond to the model task or design such as binary classification, multiclass, or multi-label (ROC, Cohen's Kappa, confusion matrix, MCC, accuracy classification score, F1 score, F-beta score, average Hamming loss, Jaccard similarly coefficient score). The results of the model fitting and testing may be captured in a learning database 132. The status of each model fitting worker process 156a, 156b, 156c may also be logged centrally, along with the status of the overall request process.
The learning module 130 may also include a model evaluation unit 134 configured to evaluate generated models in the predictive model library 142. For example, an RMSE, mean bias, mean absolute error, area under the curve on a receiver operating characteristic plot, and a date-time of model fit for each generated predictive model may be sent to the learning database 132 and stored in the predictive model library 142. The model evaluation unit 134 may use this captured test data to screen the predictive quality of the model versus captured test data of previous model versions applicable to a corresponding episode type, which may include comparisons across model versions within a model type or within multiple model types. If the model passes this screening, the model evaluation unit 134 may automatically incorporate the model into the prediction database 114 for use moving forward for future customer environments as the current best, or best-of-class, model for the corresponding episode type or possibly additional episode types.
The learning module 130 may be configured to periodically or upon occurrence of a predefined event test prediction accuracy of models in the predictive model library 142 against new, current, or existing episode data. For example, episode data may be dated in the training database to track currentness of the data to test ability of generated predictive algorithms to predict responses based on current episode data. Periodic prediction analysis may address drift, concept drift, or unintended honing or targeting of the model algorithm over time. Current or new episode data may also be obtained from a third party database or may be submitted in a request or imported into or associated with the training database 152, which may be an event that initiates analysis or evaluation of one or more system models, which may include best models, models that are not classified as current best models, or combinations thereof. A low calculated prediction accuracy may initiate retraining, updating, recalibration, or replacement of a current best model with a better performing model from the predictive model library 142 or a newly generated model fit and tested as above from new or existing episode data.
Accordingly, the machine learning prediction system 104 may be configured for machine learning and artificial intelligence during actual use by constant automatic fine-tuning of the calibration of the algorithm and expansion of the databases. The learning module 130 may test or update the regression calibration of system algorithms in the prediction database 114 and predictive model library 142 and add episode data to the training database 152 to be kept and expanded and used by subsequent queries.
Machine Learning Architecture
Automatically Expanding Data, Machine Learning and Artificial Intelligence
The data used to calibrate, evaluate and select the algorithms (i.e., the models) is constantly updated by addition of applicable databases from users, and from third-party vendors, such as hospitals, medical practices, insurance companies, credit scoring agencies, and credit history agencies. Also, the data available is updated by the data from each user request for an individual prediction for a patient episode. The accumulating data continually expands over time and shall become an increasingly valuable asset. The training module may be automatically rerun periodically, occasionally, or with each individual user request for an individual prediction. That is, the algorithms may be automatically recalibrated, re-evaluated and reselected periodically, occasionally, or otherwise, using the then current expanding database. More data is more statistically reliable, so automatically re-running of the training module with more data makes the model predictions increasingly better, that is this machine learning is improving the prediction results. This provides an automatic artificial intelligence pipeline of ever-improving predictive models.
Network Architecture with Web Applications
The request may be provide to a prediction generator 112 to generate the requested response. Using the desired response specified in the report request and available episode data, the prediction generator 112 selects the appropriate model in the prediction database 114. An identified client model may also be used. The prediction generator 112 may then execute the selected model utilizing the episode data specified in the request to generate a prediction response for the specified outcome in the request. The prediction module 110 may then provide the requested report including the prediction response, indicated by arrow 192, to the prediction application 106, via the application interface 116, for presentation to the user with the user device 108.
Prediction Architecture
Further Machine Learning Architecture
The process may further include receiving service requests for predictions relating to a specified outcome 607. Next, cased on the client request, a selection of the appropriate model from which to generate the requested prediction response may be made 608. With the appropriate model selected, the next step may include refresh/generate requested predictions 609.
Components of an Architecture
Referring to
In some embodiments, the computer apparatus 700 or a machine thereof may operate as a standalone device. In some embodiments, the computer apparatus 700 or a machine 701 thereof may be connected via a communication network 735 to and assist with operations performed by other apparatuses, machines, or systems. For example, the computer apparatus 700 or a machine 701 thereof may be connected with any component in the network or system. In a networked deployment, the computer apparatus 700 or a machine thereof may operate in the capacity of a server, server stack, or a client, such as a client user machine, in a server-client user network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The computer apparatus 701 or a machine thereof may comprise a server computer, a client user computer, a personal computer (PC), a tablet PC, a laptop computer, a desktop computer, a control system, a network router, switch or bridge, or any machine capable of executing a set of instructions 724 (sequential or otherwise) that specify actions to be taken by that computer apparatus 700 or a machine 701 thereof. In one example, the computer apparatus 700 comprising one or more machines 701 is arranged to have a distributed architecture employing a suitable model such as a client-server model. In one embodiment, the computer apparatus 700 may include a web service, e.g., service oriented architecture (SOA) or simple object access protocol (SOAP). The computer apparatus 700 may include a software as a service (SaaS) platform. The computer apparatus 700 may include representational state transfer style resources or resource oriented architecture (ROA). The computer apparatus 700 may include one or more processors, servers, databases, networks or network devices, and peripherals configured to obtain and transmit data and initiate operations configured to perform in whole or in part the operations of the system or platform thereof.
The computer apparatus 700 may include a processor 702, e.g., a central processing unit (CPU), a graphics processing unit (GPU, or both), a main memory 704 and a static memory 706, which communicate with each other via a bus 708. The computer apparatus 700 may further include a video display unit 710, which may be, but is not limited to, a liquid crystal display (LCD), a flat panel, a solid state display, or a cathode ray tube (CRT). The computer apparatus 700 may include an input device 712, such as, but not limited to, a keyboard, a cursor control device 714, such as, but not limited to, a mouse, a disk drive unit 716, a signal generation device 718, such as, but not limited to, a speaker or remote control, and a network interface device 720.
The disk drive unit 716 may include a machine-readable medium 722 on which is stored one or more sets of instructions 724, such as, but not limited to, software embodying any one or more of the methodologies or functions described herein, including those methods illustrated above. For example, the learning or training submodule 130, 150 described herein may include messaging or queuing processes or functions. Thus, the learning or training submodule may include Kue library to handle task distribution and coordination between worker process submodules, which may be backed by Redis in-memory key-value store with pub/sub. The instructions 724 may also reside, completely or at least partially, within the main memory 704, the static memory 706, or within the processor 702, or a combination thereof, during execution thereof by the computer system 700. The main memory 704 and the processor 702 also may constitute machine-readable media.
Various embodiments may be used in systems with various types of prediction web applications 106, including for example: a patient risk scoring web application, a dynamic patient guidance web application, a benchmark automation web application, an automatic claims completion web application, an automated insight generation web application, an active updates of outcomes web application, a performance ranking web application, or other prediction web applications.
For ease of explanation, Layer 2, the prediction application, is sometimes discussed here using a patient risk scoring application 106 embodiment as an example; however it is understood that the principles of the interaction of Layer 2 with Layer 1 are as illustrated by this discussion of this one possible embodiment of Layer 2, but apply equally to all Layer 2 prediction application 106 embodiments.
The following is an example of the work flow of an embodiment of a prediction application for a user experience with the embodiment:
Embodiments automatically use the Layer 1 pipeline of healthcare outcome models to select and apply the optimal predictive model for the prediction application.
In an embodiment, a computer network architecture with a prediction application, artificial intelligence and machine learning, comprises (1) a prediction module with a prediction generator and an updated database, (2) a learning module with a training submodule, in electronic communication with the prediction module and the updated database, (3) a web application in electronic communication with both the prediction module, and a user device, and wherein, (4) the web prediction application is adapted to:
In various embodiments, in
In various embodiments,
Combined Data
The web application 106 may combine data of patient interactions with the healthcare team and with the present invention, and proprietary third-party patient attributes data and model attributes. All data from all sources is kept in the various databases 114, 132, 152 and all such data may be found in each database 114, 132, 152 or partitioned among the databases. Each prediction request 190 collects data for at least one patient, which is added to the databases with each request. As described herein, the databases are also augmented with new data by administrators 108, through an administrator application 172 or administration module 170. Each patient interaction with the system 100 through some other web application 106, or other interaction with the healthcare team such as a doctor visit, may also add data to the databases. Also, proprietary data from third party vendors about patient data for medical, financial, economic, and other data may be added to the databases. Individual data from whatever source is treated confidentially, and is not disclosed; however, the data can be aggregated across demographic groups, and also used to calculate derivative data, all of which can be used as data for the variables in the functional equations of optimal prediction models. The prediction requested may then be disclosed to the individual patient and his clinician, without disclosing any confidential data of any other individual.
A Focus of the Robot's Intelligence and Learning
This ongoing updating and recalibration of the optimal forecasting model using augmented data is an important source of machine learning (i.e., artificial intelligence) in embodiments. One indicia of artificial intelligence and machine learning is that when the system receives the same user request on two different days, the response on the second day will be different and smarter because the database has been expanded (i.e., the machine has learned), and the answer will be more accurate because the model is recalibrated with more data (i.e., the intelligence of the machine is increased). Because data is augmented in part from the machine's own receipt of user requests (i.e., each request with more patient data is added to the system databases), the machine is learning from its own experience and use. The more the computer architecture experiences, the more the computer architecture learns, because it incorporates more data.
Other Matters
Although various embodiments may be referred to as robots because of their machine learning and artificial intelligence, these embodiments might not have an anthropomorphic form or sound.
Dedicated hardware implementations including, but not limited to, application specific integrated circuits, programmable logic arrays and other hardware devices can likewise be constructed to implement the methods described herein. Applications that may include the apparatus and systems of various embodiments broadly include a variety of electronic and computer systems. Some embodiments implement functions in two or more specific interconnected hardware modules or devices with related control and data signals communicated between and through the modules, or as portions of an application-specific integrated circuit. Thus, the example network or system is applicable to software, firmware, and hardware implementations.
In accordance with various embodiments of the present disclosure, the processes described herein are intended for operation as software programs running on a computer processor. Furthermore, software implementations can include, but are not limited to, distributed processing or component/object distributed processing, parallel processing, or virtual machine processing that may be constructed to implement the methods described herein.
The present disclosure describes various modules, which may also be referred to as sub-modules, generators, engines, systems, subsystems, components, units, and the like. Such modules may include functionally related hardware, instructions, firmware, or software. Modules may include physical or logical grouping of functionally related applications, services, resources, assets, systems, programs, databases, or the like. Modules or hardware storing instructions or configured to execute functionalities of the modules may be physically located in one or more physical locations. For example, modules may be distributed across one or more networks, systems, devices, or combination thereof. It will be appreciated that the various functionalities of these features may be modular, distributed, and/or integrated over one or more physical devices. While the learning database as well as the training databases are illustrated individually as discrete elements, it will be appreciated that such logical partitions may not correspond to physical partitions of the data. For example, all or portions of the databases may reside or be distributed among one or more hardware locations. Further, while screening for best-of-class or current best model is generally described as including scanning the predictive model library 142, this library 142 includes the models currently used by the prediction module 110, references to such models, evaluation metrics related to such data, or other data from which the screening may be performed to properly screen the library 142 of predictive models for the best-of-class or current best model.
The present disclosure contemplates a machine-readable medium 722 containing instructions 724 so that a device connected to the communications network 735, another network, or a combination thereof, can send or receive voice, video or data, and to communicate over the communications network 735, another network, or a combination thereof, using the instructions. The instructions 724 may further be transmitted or received over the communications network 735, another network, or a combination thereof, via the network interface device 720.
While the machine-readable medium 722 is shown in an example embodiment to be a single medium, the term “machine-readable medium” should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, and/or associated caches and servers) that store the one or more sets of instructions. The term “machine-readable medium” shall also be taken to include any medium that is capable of storing, encoding or carrying a set of instructions for execution by the machine and that causes the machine to perform any one or more of the methodologies of the present disclosure.
The terms “machine-readable medium,” “machine-readable device,” or “computer-readable device” shall accordingly be taken to include, but not be limited to: memory devices, solid-state memories such as a memory card or other package that houses one or more read-only (non-volatile) memories, random access memories, or other re-writable (volatile) memories; magneto-optical or optical medium such as a disk or tape; or other self-contained information archive or set of archives is considered a distribution medium equivalent to a tangible storage medium. The “machine-readable medium,” “machine-readable device,” or “computer-readable device” may be non-transitory, and, in certain embodiments, may not include a wave or signal per se. Accordingly, the disclosure is considered to include any one or more of a machine-readable medium or a distribution medium, as listed herein and including art-recognized equivalents and successor media, in which the software implementations herein are stored.
The illustrations of arrangements described herein are intended to provide a general understanding of the structure of various embodiments, and they are not intended to serve as a complete description of all the elements and features of the network architecture, systems, and processes that might make use of the structures described herein. While the present disclosure generally describes the inventive network architecture, systems, and process with respect to healthcare applications, the healthcare field is but only one of many potential applications. Indeed, those having skill in the art will appreciate that the network architecture, systems, and processes described herein may find application in many industries. Other arrangements may be utilized and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Figures are also merely representational and may not be drawn to scale. Certain proportions thereof may be exaggerated, while others may be minimized. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense.
Thus, although specific arrangements have been illustrated and described herein, it should be appreciated that any arrangement calculated to achieve the same purpose may be substituted for the specific arrangement shown. This disclosure is intended to cover any and all adaptations or variations of various embodiments and arrangements of the invention. Combinations of the above arrangements, and other arrangements not specifically described herein, will be apparent to those of skill in the art upon reviewing the above description. Therefore, it is intended that the disclosure not be limited to the particular arrangement(s) disclosed as the best mode contemplated for carrying out this invention, but that the invention will include all embodiments and arrangements falling within the scope of the appended claims.
The foregoing is provided for purposes of illustrating, explaining, and describing embodiments of this invention. Modifications and adaptations to these embodiments will be apparent to those skilled in the art and may be made without departing from the scope or spirit of this invention. Upon reviewing the aforementioned embodiments, it would be evident to an artisan with ordinary skill in the art that said embodiments can be modified, reduced, or enhanced without departing from the scope and spirit of the claims described below.
Number | Name | Date | Kind |
---|---|---|---|
6381576 | Gilbert | Apr 2002 | B1 |
6687685 | Sadeghi | Feb 2004 | B1 |
7685000 | Petit | Mar 2010 | B1 |
7809585 | Ghouri | Oct 2010 | B1 |
7917377 | Rao | Mar 2011 | B2 |
8412537 | Fenton | Apr 2013 | B1 |
8489412 | Gliklich | Jul 2013 | B1 |
8508336 | Giobbi | Aug 2013 | B2 |
8732096 | Glukhov | May 2014 | B1 |
8805759 | Cha | Aug 2014 | B1 |
9002719 | Tofte | Apr 2015 | B2 |
9202253 | Macoviak | Dec 2015 | B2 |
9536052 | Amarasingham | Jan 2017 | B2 |
9607058 | Gupta | Mar 2017 | B1 |
9619841 | DiRienzo | Apr 2017 | B2 |
9953372 | Dziabiak | Apr 2018 | B1 |
9996666 | Wilson | Jun 2018 | B1 |
10102926 | Leonardi | Oct 2018 | B1 |
10127234 | Krishnan | Nov 2018 | B1 |
10262374 | Sultan | Apr 2019 | B2 |
10650928 | Larson | May 2020 | B1 |
10734101 | Rajan | Aug 2020 | B2 |
10783998 | Perlin | Sep 2020 | B1 |
10910107 | Larson | Feb 2021 | B1 |
10910113 | Drouin | Feb 2021 | B1 |
10990904 | Drouin | Apr 2021 | B1 |
10998104 | Warner | May 2021 | B1 |
11157808 | Wetta | Oct 2021 | B2 |
11302426 | Winlo | Apr 2022 | B1 |
20020087358 | Gilbert | Jul 2002 | A1 |
20030191669 | Fitzgerald | Oct 2003 | A1 |
20050038669 | Sachdeva | Feb 2005 | A1 |
20050137912 | Rao | Jun 2005 | A1 |
20050222867 | Underwood | Oct 2005 | A1 |
20060052945 | Rabinowitz et al. | Mar 2006 | A1 |
20060080139 | Mainzer | Apr 2006 | A1 |
20060161456 | Baker | Jul 2006 | A1 |
20060178915 | Chao | Aug 2006 | A1 |
20070005154 | Lancaster | Jan 2007 | A1 |
20070260492 | Feied | Nov 2007 | A1 |
20080010097 | Williams | Jan 2008 | A1 |
20080120133 | Krishnaswami | May 2008 | A1 |
20080146893 | Levendowski | Jun 2008 | A1 |
20080183508 | Harker | Jul 2008 | A1 |
20080201172 | McNamar | Aug 2008 | A1 |
20090006419 | Savitsky | Jan 2009 | A1 |
20090048877 | Binns | Feb 2009 | A1 |
20090192828 | Yunker | Jul 2009 | A1 |
20090206992 | Giobbi | Aug 2009 | A1 |
20090222290 | Crowe | Sep 2009 | A1 |
20110112871 | Simonowski | May 2011 | A1 |
20110166883 | Palmer | Jul 2011 | A1 |
20110213625 | Joao | Sep 2011 | A1 |
20110270630 | Green, III | Nov 2011 | A1 |
20110289035 | Stojadinovic et al. | Nov 2011 | A1 |
20120047105 | Saigal | Feb 2012 | A1 |
20120179478 | Ross | Jul 2012 | A1 |
20120271612 | Barsoum | Oct 2012 | A1 |
20120310661 | Greene | Dec 2012 | A1 |
20130054259 | Wojtusiak | Feb 2013 | A1 |
20130096937 | Campbell | Apr 2013 | A1 |
20130159023 | Srinivas | Jun 2013 | A1 |
20130197936 | Willich | Aug 2013 | A1 |
20130226624 | Blessman | Aug 2013 | A1 |
20130290028 | McGuigan | Oct 2013 | A1 |
20130290037 | Hu | Oct 2013 | A1 |
20130304618 | Mastrogiovanni | Nov 2013 | A1 |
20140046682 | Soto | Feb 2014 | A1 |
20140058755 | Macoviak | Feb 2014 | A1 |
20140058763 | Zizzamia | Feb 2014 | A1 |
20140067406 | Hyatt | Mar 2014 | A1 |
20140081652 | Klindworth | Mar 2014 | A1 |
20140095201 | Farooq | Apr 2014 | A1 |
20140122100 | Fillmore | May 2014 | A1 |
20140316810 | Oliver | Oct 2014 | A1 |
20150088541 | Yao | Mar 2015 | A1 |
20150095069 | Waljee | Apr 2015 | A1 |
20150100336 | Ford | Apr 2015 | A1 |
20150127370 | Cornelis | May 2015 | A1 |
20150213223 | Amarasingham | Jul 2015 | A1 |
20150213225 | Amarasingham | Jul 2015 | A1 |
20150242583 | Edson | Aug 2015 | A1 |
20150261922 | Nawana | Sep 2015 | A1 |
20150302178 | Patel | Oct 2015 | A1 |
20150347705 | Simon | Dec 2015 | A1 |
20160048766 | McMahon | Feb 2016 | A1 |
20160048780 | Sethumadhavan | Feb 2016 | A1 |
20160063212 | Monier | Mar 2016 | A1 |
20160071171 | Cancelliere | Mar 2016 | A1 |
20160078578 | Tumma | Mar 2016 | A1 |
20160092641 | Delaney | Mar 2016 | A1 |
20160098533 | Jackson | Apr 2016 | A1 |
20160132646 | Jones | May 2016 | A1 |
20160154937 | Hennenfent | Jun 2016 | A1 |
20160196398 | Vivero | Jul 2016 | A1 |
20160285876 | Perez | Sep 2016 | A1 |
20160321412 | Basri | Nov 2016 | A1 |
20160328526 | Park | Nov 2016 | A1 |
20170006135 | Siebel | Jan 2017 | A1 |
20170017760 | Freese | Jan 2017 | A1 |
20170029894 | Damask | Feb 2017 | A1 |
20170053080 | Geppert | Feb 2017 | A1 |
20170100213 | Kuo | Apr 2017 | A1 |
20170103180 | Jiao | Apr 2017 | A1 |
20170103181 | Czerska | Apr 2017 | A1 |
20170140114 | Are | May 2017 | A1 |
20170161439 | Raduchel | Jun 2017 | A1 |
20170169173 | Snow, Jr. | Jun 2017 | A1 |
20170236063 | Dorris | Aug 2017 | A1 |
20170249434 | Brunner | Aug 2017 | A1 |
20170270257 | St. Clair | Sep 2017 | A1 |
20170293722 | Valverde, Jr. | Oct 2017 | A1 |
20170308650 | Brill | Oct 2017 | A1 |
20170308652 | Ligon | Oct 2017 | A1 |
20170308981 | Razavian | Oct 2017 | A1 |
20170351821 | Tanner, Jr. | Dec 2017 | A1 |
20170372028 | Zhou | Dec 2017 | A1 |
20180011976 | Lewis | Jan 2018 | A1 |
20180025126 | Barnard | Jan 2018 | A1 |
20180060521 | Connolly | Mar 2018 | A1 |
20180060744 | Achin | Mar 2018 | A1 |
20180089376 | Tucker | Mar 2018 | A1 |
20180137247 | Bore | May 2018 | A1 |
20180182475 | Cossler | Jun 2018 | A1 |
20180211727 | Zarkoob | Jul 2018 | A1 |
20180268320 | Shekhar | Sep 2018 | A1 |
20180277246 | Zhong | Sep 2018 | A1 |
20180289387 | Khajavi | Oct 2018 | A1 |
20180336638 | Dziabiak | Nov 2018 | A1 |
20180336639 | Dziabiak | Nov 2018 | A1 |
20180336640 | Dziabiak | Nov 2018 | A1 |
20180342323 | Shankar | Nov 2018 | A1 |
20180344215 | Ohnemus | Dec 2018 | A1 |
20190005115 | Warrier | Jan 2019 | A1 |
20190005195 | Peterson | Jan 2019 | A1 |
20190005410 | Shekhar | Jan 2019 | A1 |
20190034590 | Oren | Jan 2019 | A1 |
20190043605 | Hegarty | Feb 2019 | A1 |
20190080416 | Smith | Mar 2019 | A1 |
20190172564 | Chandra | Jun 2019 | A1 |
20190244683 | Francois | Aug 2019 | A1 |
20190272602 | Ray | Sep 2019 | A1 |
20190311812 | Sweeney | Oct 2019 | A1 |
20190325995 | Malone | Oct 2019 | A1 |
20190333636 | Bennett | Oct 2019 | A1 |
20190355481 | Lamb | Nov 2019 | A1 |
20200005080 | Cogan | Jan 2020 | A1 |
20200058403 | Barrett | Feb 2020 | A1 |
20200075139 | Master | Mar 2020 | A1 |
20200105413 | Vladimirova | Apr 2020 | A1 |
20200111545 | Syeda-Mahmood | Apr 2020 | A1 |
20200152320 | Ghazaleh | May 2020 | A1 |
20200152332 | Yang | May 2020 | A1 |
20200160998 | Ward | May 2020 | A1 |
20200176119 | Spagnolo | Jun 2020 | A1 |
20200211692 | Kalafut | Jul 2020 | A1 |
20200227172 | Perkins | Jul 2020 | A1 |
20200253547 | Harris | Aug 2020 | A1 |
20200258511 | Barkol | Aug 2020 | A1 |
20200279635 | Letterie | Sep 2020 | A1 |
20200286616 | Dunn | Sep 2020 | A1 |
20200293887 | De Brouwer | Sep 2020 | A1 |
20200311300 | Callcut | Oct 2020 | A1 |
20210076960 | Fornwalt | Mar 2021 | A1 |
20210082554 | Kalia | Mar 2021 | A1 |
20210193320 | Shukla | Jun 2021 | A1 |
20210241905 | Hoar | Aug 2021 | A1 |
20210350910 | Dastmalchi | Nov 2021 | A1 |
20210391079 | Clifton | Dec 2021 | A1 |
20220051773 | Appelbaum | Feb 2022 | A1 |
Number | Date | Country |
---|---|---|
101689220 | Mar 2010 | CN |
3471107 | Apr 2019 | EP |
2016094330 | Jun 2016 | WO |
WO-2016201212 | Dec 2016 | WO |
WO-2018149397 | Aug 2018 | WO |
WO-2020089576 | May 2020 | WO |
2021003485 | Jan 2021 | WO |
Entry |
---|
Sheth et al., “Active Semantic Electronic Medical Records” 2006. (Year: 2006). |
Thesmar et al., “Combining the Power of Artificial Intelligence with the Richness of Healthcare Claims Data: Opportunities and Challenges” Mar. 8, 2019, PharmoEconomics, No. 37, pp. 745-752. (Year: 2019). |
Luo et al., “Automating Construction of Machine Learning Models with Clinical Big Data: Proposal Rationale and Methods” Aug. 29, 2017, JMIR Res Protoc, vol. 6, No. 8, pp. 1-18. (Year: 2017). |
Bjarnadottir, Margret Vilborg, “Data-Driven Approach to Health Care: Applications Using Claims Data” Sep. 2008, Doctoral Thesis, MIT, pp. i-130. (Year: 2008). |
Lix et al., “A Prediction model to estimate completeness of electronic physician claims databases” May 11, 2015, BMJ Open, pp. 1-8. (Year: 2015). |
Bertsimas et al., “Algorithmic Prediction of Health-Care Costs” Dec. 2008, Operations Research, vol. 56, No. 6, pp. 1382-1392. (Year: 2008). |
Xie et al., “Predicting Days in Hospital Using Health Insurance Claims” Jul. 2015, IEEE Journal of Biomedical and Health Informatics, vol. 19, No. 4, pp. 1224-1233. (Year: 2015). |
Hamad et al., “Using ‘Big Data’ to Capture Overall Health Status: Properties and Predictive Value of a Claims-Based Health Risk Score” May 7, 2015, PLoS One, vol. 10, No. 5, pp. 1-14. (Year: 2015). |
Yen et al., “Association Between Wellness Score from a Health Risk Appraisal and Prospective Medical Claims Costs” Oct. 2003, JOEM, vol. 45, No. 10, pp. 1049-1057. (Year: 2003). |
Sushmita et al., “Population Cost Prediction on Public Healthcare Datasets” May 18-20, 2015, pp. 87-94. (Year: 2015). |
Sears et al., “Predicting work-related disability and medical cost outcomes: A comparison of severity scoring methods” Dec. 28, 2012, Injury, Int J Care Injured, vol. 45, pp. 16-22. (Year: 2012). |
Ramamurthy et al., “A Configurable, Big Data System for On-Demand Healthcare Cost Prediction” 2017, IEEE International Conference on Big Data, pp. 1524-1533. (Year: 2017). |
Layton, Timothy James, “Risk Selection and Risk Adjustment in Competitive Health Insurance Markets” 2014, Doctoral Dissertation, Boston University, pp. i-179. (Year: 2014). |
Wang, Ye, “Competition in the Market of Health Insurance and Health Care Utilization”, 2017, Doctoral Dissertation, Boston University, pp. i-123. (Year: 2017). |
Silverio et al., “Big Health Data and Cardiovascular Diseases: A Challenge for Research, an Opportunity for Clinical Care” Feb. 25, 2019, Frontiers in Medicine, vol. 6, Art 36, pp. 1-10. (Year: 2019). |
Lentz et al., “Prediction of healthcare utilization following an episode of physical therapy for musculoskeletal pain” 2018, BMC Health Services Research, pp. 1-14. (Year: 2018). |
Johnson et al., “Machine Learning and Decision Support in Critical Care” Jan. 19, 2016, Proceedings of the IEEE, vol. 104, No. 2, pp. 444-466. (Year: 2016). |
Akhmetov et al., “Assessing value of innovative molecular diagnostic tests in the concept of predictive, preventative, and personalized medicine” 2015, EPMA Journal, pp. 1-12. (Year: 2015). |
Reyes-Garcia et al., “Evaluation of the Impact of Data Uncertainty on the Prediction of Physiological Patient Deterioration” 2018, IEEE, pp. 38595-38606. (Year: 2018). |
Razzaghi et al., “Multilevel Weighted Support Vector Machine for Classification on Healthcare Data with Missing Values” Apr. 7, 2016, pp. 1-20. (Year: 2016). |
Xie et al., “Predicting Days in Hospital Using Health Insurance Claims” Jul. 2015, IEEE, pp. 1224-1233. (Year: 2015). |
Rhon et al., “Leveraging healthcare utilization to explore outcomes from musculoskeletal disorders: methodology for defining relevant variables from a health services data repository” 2018, pp. 1-11. (Year: 2018). |
Lipton et al., “Modeling Missing Data in Clinical Time Series with RNNs” Nov. 11, 2016, Proceedings of Machine Learning for Healthcare. (Year: 2016). |
Liu et al., “Learning Hierarchical Representations of Electronic Health Records for Clinical Outcome Prediction” Mar. 20, 2019. (Year: 2019). |
Nezhad et al., “A Predictive Approach Using Deep Feature Learning for Electronic Medical Records: A Comparative Study” Jan. 6, 2018, pp. 1-24. (Year: 2018). |
Jiang et al., “Cost-Sensitive Parallel Learning Framework for Insurance Intelligence Operation” Oct. 10, 2018. (Year: 2018). |
Kanchinadam et al., “Using Discriminative Graphical Models for Insurance Recommender Systems” Dec. 17, 2018, pp. 421-428. (Year: 2018). |
Qazi et al., “An Insurance Recommendation System Using Bayesian Networks” Aug. 2017, pp. 274-278. (Year: 2017). |
De Brouwer et al., “System and Method with Federated Learning Model for Medical Research Applications” Mar. 11, 2019, U.S. Appl. No. 62/816,880. (Year: 2019). |
Beam et al., “Clinical Concept Embeddings Learned from Massive Sources of Multimodal Medical Data” May 18, 2018, pp. 1-11. (Year: 2018). |
Henckaerts et al., “Tree-based machine learning for insurance pricing” Sep. 11, 2018. (Year: 2018). |
Liu et al., “FADL: Federated-Autonomous Deep Learning for Distributed Electronic Health Record” Dec. 3, 2018, pp. 1-5. (Year: 2018). |
Zhang et al., “Patient2Vec: A Personalized Interpretable Deep Representation of the Longitudinal Electronic Health Record” Oct. 25, 2018, arXiv: 1810.04793v3, pp. 1-14. (Year: 2018). |
Bai et al., “Interpretable Representation Learning for Healthcare via Capturing Disease Progression through Time” Aug. 2018, pp. 43-51. (Year: 2018). |
Kanchinadam et al., “Using Discriminative Graphical Models for Insurance Recommender Systems” 2018, pp. 421-428. (Year: 2018). |
Fadhil, Ahmed “A Conversational Interface to Improve Medication Adherence: Towards AI Support in Patient's Treatment” 2018. (Year: 2018). |
Huang et al., “ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmission” Apr. 10, 2019, pp. 1-19. (Year: 2019). |
Wang et Xu, “Inpatient2Vec: Medical Representation Learning for Inpatients” Apr. 18, 2019. (Year: 2019). |
Zhang et al., “MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records” May 8, 2019. (Year: 2019). |
Gottula et al., U.S. Appl. No. 16/387,713, filed Apr. 18, 2019. |
Talvola et al., U.S. Appl. No. 16/404,446, filed May 6, 2019. |
Drouin et al., U.S. Appl. No. 16/584,326, filed Sep. 26, 2019. |
Warner et al., U.S. Appl. No. 16/587,780, filed Sep. 30, 2019. |
Drouin et al., U.S. Appl. No. 16/533,465, filed Aug. 6, 2019. |
Larson et al., U.S. Appl. No. 15/845,633, filed Dec. 18, 2017. |
Larson et al., U.S. Appl. No. 15/998,771, filed Aug. 16, 2018. |
Wang et al., U.S. Appl. No. 16/007,819, filed Jun. 13, 2018. |
Wang et al., U.S. Appl. No. 16/694,292, filed Nov. 25, 2019. |
Larson et al., U.S. Appl. No. 16/694,330, filed Nov. 25, 2019. |
Talvola et al., U.S. Appl. No. 16/697,773, filed Nov. 27, 2019. |
Talvola et al., U.S. Appl. No. 17/677,580, filed Feb. 22, 2022. |
Zou et al., “Regularization and variable selection via the elestic net,” J. of the Royal Statistical Soc., Series B (statistical methodology), 67:2, pp. 301-320 (2005). |
Jung et al., “Predicting need for advanced illness or palliative care in a primary care population using electronic health record data,” J. Biomedical Informatics 92(2019) 103115. |
Khalifa et al., “Utilizing health analytics in improving the performance of healthcare services: A case study on a tertiary care hospital,” J. of Infection and Public Health (2016) 9, 757-765. |