Systems and methods for targeted medical document review
Systems and methods for targeted medical document review are disclosed. A list of medical documents is received. Each medical document is associated with a user enrolled in a plan subject to a medical document review process. A dataset for each medical document, including clinical data, membership data, and provider data associated with the user, is received. A first model is used to determine whether each medical document includes an undocumented condition based on the dataset. The list is reduced to a subset of medical documents determined to include an undocumented condition. A second model is used to determine a risk score associated with each medical document of the subset based on the dataset. The subset of medical documents are ordered in the reduced list based on the risk scores. The ordered, reduced list is provided as input to the medical document review process.
1 . A method for targeted medical document review performed by one or more processors, the method comprising:
receiving a list of a plurality of medical documents, wherein each medical document of the plurality of medical documents is associated with a respective user of a plurality of users enrolled in a plan for which a medical document review process is performed;
receiving a dataset for each medical document of the plurality of medical documents, the dataset including first clinical data, first membership data, and first provider data associated with the respective user, wherein the first clinical data, the first membership data, and the first provider data each include one or more data types from which information associated with an undocumented condition is inferable;
for each medical document of the plurality of medical documents, determining, using a first trained machine learning model to which the one or more data types are input as features, a classification of the medical document as having a presence or an absence of the undocumented condition in the medical document based on the dataset for the medical document, wherein the information inferable from the one or more data types and one or more first feature weights associated with the one or more data types in the first trained machine learning model are leveraged to determine the classification;
reducing the list to a first subset of a plurality of the plurality of medical documents determined as having the presence of the undocumented condition based on an output of the trained first machine learning model;
for each medical document of the first subset, determining, using a second trained machine learning model to which the one or more data types are input as features, a risk score associated with the medical document based on the dataset for the medical document, wherein the information inferable from the one or more data types and one or more second feature weights associated with the one or more data types in the second trained machine learning model are leveraged to determine the risk score;
ordering the first subset based on the determined risk score for each medical document of the first subset;
providing the ordered first subset as input to the medical document review process;
receiving, as feedback from the medical document review process, one or more of (a) an indication of whether the undocumented condition is present in each medical document of the first subset or (b) an actual risk score for each medical document of the first subset; and
generating and providing, to one or more of the first trained machine learning model or the second trained machine learning model for use in retraining, one or more new training samples from the first subset of the plurality of medical documents that, for each new training sample of the one or more new training samples, includes the dataset for a respective medical document of the first subset and the feedback for the respective medical document of the first subset, wherein the retraining includes the one or more of the first trained machine learning model or the second trained machine learning model repeatedly processing at least a portion of the one or more new training samples to cause an adjusting of one or more of the one or more first feature weights or the one or more second feature weights associated with the one or more data types until a determined loss or error associated with the one or more of the first trained machine learning model or the second trained machine learning model is below a predefined threshold.
2 . The method of claim 1 , wherein the undocumented condition is a condition corresponding to a hierarchical condition category (HCC) that has not been previously documented through one or more prospective documentation processes.
3 . The method of claim 1 , wherein the first trained machine learning model is a classification model, and the first trained machine learning model is trained by:
receiving a plurality of training datasets, wherein each of the plurality of training datasets is associated with a previously reviewed medical document of a user and includes second clinical data, second membership data, and second provider data associated with the user and at least a first label indicating whether a particular undocumented condition was included in the previously reviewed medical document; and
providing at least a portion of the plurality of training datasets as input to train the first trained machine learning model to predict a presence or absence of undocumented conditions in medical documents.
4 . The method of claim 3 , wherein the second trained machine learning model is a regression model, and the second trained machine learning model is trained by:
receiving a second subset of one or more of the plurality of training datasets, wherein each training dataset of the second subset is associated with the previously reviewed medical document of a the user including the first label indicating the particular undocumented condition was included in the previously reviewed medical document and further includes a second label indicating a risk score associated with the previously reviewed medical document;
receiving a plurality of historical weights assigned to a plurality of risk factors for a documentation time period associated with the second subset; and
providing at least a portion of the second subset of the plurality of training datasets and the plurality of historical weights as input to train the second trained machine learning model to predict risk scores associated with medical documents.
5 . The method of claim 1 , wherein the second trained machine learning model is a regression model, and the second trained machine learning model is trained by:
receiving a plurality of training datasets, wherein each training dataset of the plurality of training datasets is associated with a previously reviewed medical document of a user including a particular undocumented condition and further includes a second label indicating a risk score associated with the previously reviewed medical document;
receiving a plurality of historical weights assigned to a plurality of risk factors for a documentation time period associated with the plurality of training datasets; and
providing at least a portion of the plurality of training datasets and the plurality of historical weights as input to train the second trained machine learning model to predict risk scores associated with medical documents.
6 . The method of claim 1 , wherein the determined risk score is a weighted hierarchical condition category (HCC) value based, at least in part, on one or more of a plurality of weights assigned to a plurality of risk factors.
7 . The method of claim 1 , wherein ordering the first subset includes ordering the first subset from a highest risk score to a lowest risk score.
8 . The method of claim 1 , wherein the first clinical data includes one or more of suspect data, laboratory data, pharmaceutical data, or metadata of the medical document.
9 . The method of claim 1 , wherein the first membership data includes one or more of claims data, monthly membership record (MMR) data, or model output report (MOR) data.
10 . The method of claim 1 , wherein the first provider data includes one or more of demographic data or behavioral data of a healthcare provider of the respective user that is associated with the medical document.
11 . The method of claim 1 , wherein the dataset further includes one or more of social determinants of health data of the respective user or prospective program data.
12 . The method of claim 1 , wherein the first trained machine learning model is trained using a plurality of first training datasets for a plurality of first medical documents associated with a first documentation time period, and the first trained machine learning model is validated using a plurality of second training datasets for a plurality of medical documents associated with a second documentation time period, different from the first documentation time period, to perform out of time validation.
13 . The method of claim 1 , wherein the list includes, as part of metadata for each medical document of the plurality of medical documents, a storage location of the medical document, and the method further comprising:
initiating the medical document review process by:
retrieving, using the metadata, each first medical document of the ordered first subset from the storage location of the first medical document; and
pre-processing one or more first medical documents of the ordered first subset retrieved.
14 . A system for targeted medical document review, the system comprising:
one or more processors; and
one or more non-transitory computer readable media storing processor-executable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
receiving a list of a plurality of medical documents, wherein each medical document of the plurality of medical documents is associated with a respective user of a plurality of users enrolled in a plan for which a medical document review process is performed;
receiving a dataset for each medical document of the plurality of medical documents, the dataset including first clinical data, first membership data, and first provider data associated with the respective user, wherein the first clinical data, the first membership data, and the first provider data each include one or more data types from which information associated with an undocumented condition is inferable;
for each medical document of the plurality of medical documents, determining, using a first trained machine learning model to which the one or more data types are input as features, a classification of the medical document as having a presence or an absence of the undocumented condition in the medical document based on the dataset for the medical document, wherein the information inferable from the one or more data types and one or more first feature weights associated with the one or more data types in the first trained machine learning model are leveraged to determine the classification;
reducing the list to a first subset of a plurality of the plurality of medical documents determined as having the presence of the undocumented condition based on an output of the first trained machine learning model;
for each medical document of the first subset, determining, using a trained second machine learning model to which the one or more data types are input as features, a risk score associated with the medical document based on the dataset for the medical document, wherein the information inferable from the one or more data types and one or more second feature weights associated with the one or more data types in the second trained machine learning model are leveraged to determine the risk score;
ordering the first subset based on the determined risk score for each medical document of the first subset;
providing the ordered first subset as input to the medical document review process;
receiving, as feedback from the medical document review process, one or more of (a) an indication of whether the undocumented condition is present in each medical document of the first subset or (b) an actual risk score for each medical document of the first subset; and
generating and providing, to one or more of the first trained machine learning model or the second trained machine learning model for use in retraining, one or more new training samples from the first subset of the plurality of medical documents that, for each new training sample of the one or more new training samples, includes the dataset for a respective medical document of the first subset and the feedback for the respective medical document of the subset, wherein the retraining includes the one or more of the first trained machine learning model or the second trained machine learning model repeatedly processing at least a portion of the one or more new training samples to cause an adjusting of one or more of the one or more first feature weights or the one or more second feature weights associated with the one or more data types until a determined loss or error associated with the one or more of the first trained machine learning model or the second trained machine learning model is below a predefined threshold.
15 . The system of claim 14 , wherein the undocumented condition is a condition corresponding to a hierarchical condition category (HCC) that has not been previously documented through one or more prospective documentation processes.
16 . The system of claim 14 , wherein the first trained machine learning model is a classification model, and the first trained machine learning model is trained by:
receiving a plurality of training datasets, wherein each of the plurality of training datasets is associated with a previously reviewed medical document of a user and includes second clinical data, second membership data, and second provider data associated with the user and at least a first label indicating whether a particular undocumented condition was included in the previously reviewed medical document; and
providing at least a portion of the plurality of training datasets as input to train the first trained machine learning model to predict a presence or absence of undocumented conditions in medical documents.
17 . The system of claim 16 , wherein the second trained machine learning model is a regression model, and the second trained machine learning model is trained by:
receiving a second subset of one or more the plurality of training datasets, wherein each training dataset of the second subset is associated with the previously reviewed medical document of the user including the first label indicating the particular undocumented condition was included in the previously reviewed medical document and further includes a second label indicating a risk score associated with the previously reviewed medical document;
receiving a plurality of historical weights assigned to a plurality of risk factors for a documentation time period associated with the second subset; and
providing at least a portion of the second subset of the plurality of training datasets and the plurality of historical weights as input to train the second trained machine learning model to predict risk scores associated with medical documents.
18 . The system of claim 14 , wherein the second trained machine learning model is a regression model, and the second trained machine learning model is trained by:
receiving a plurality of training datasets, wherein each training dataset of the plurality of training datasets is associated with a previously reviewed medical document of a user including a particular undocumented condition and further includes a second label indicating a risk score associated with the previously reviewed medical document;
receiving a plurality of historical weights assigned to a plurality of risk factors for a documentation time period associated with the plurality of training datasets; and
providing at least a portion of the plurality of training datasets and the plurality of historical weights as input to train the second trained machine learning model to predict risk scores associated with medical documents.
19 . The system of claim 14 , wherein the determined risk score is a weighted hierarchical condition category (HCC) value based at least in part on one or more of a plurality of weights assigned to a plurality of risk factors, and the plurality of risk factors are associated with one or more of user demographics, types of conditions, severity of conditions, duration of user enrollment in the plan, or prescription drug use.
20 . One or more non-transitory computer readable media for targeted medical document review, the one or more non-transitory computer readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
receiving a list of a plurality of medical documents, wherein each medical document of the plurality of medical documents is associated with a respective user of a plurality of users enrolled in a plan for which a medical document review process is performed;
receiving a dataset for each medical document of the plurality of medical documents, the dataset including first clinical data, first membership data, and first provider data associated with the respective user, wherein the first clinical data, the first membership data, and the first provider data each include one or more data types from which information associated with an undocumented condition is inferable;
for each medical document of the plurality of medical documents, determining, using a first trained machine learning model to which the one or more data types are input as features, a classification of the medical document as having a presence or an absence of the undocumented condition in the medical document based on the dataset for the medical document, wherein the information inferable from the one or more data types and one or more first feature weights associated with the one or more data types in the first trained machine learning model are leveraged to determine the classification;
reducing the list to a first subset of a plurality of the plurality of medical documents determined as having the presence of the undocumented condition based on an output of the first trained machine learning model;
for each medical document of the first subset, determining, using a second trained machine learning model to which the one or more data types are input as features, a risk score associated with the medical document based on the dataset for the medical document, wherein the information inferable from the one or more data types and one or more second feature weights associated with the one or more data types in the second trained machine learning model are leveraged to determine the risk score;
ordering the first subset based on the determined risk score for each medical document of the first subset;
providing the ordered first subset as input to the medical document review process;
receiving, as feedback from the medical document review process, one or more of (a) an indication of whether the undocumented condition is present in each medical document of the subset or (b) an actual risk score for each medical document of the subset; and
generating and providing, to one or more of the first trained machine learning model or the second trained machine learning model for use in retraining, one or more new training samples from the first subset of the plurality of medical documents that, for each new training sample of the one or more new training samples, includes the dataset for a respective medical document of the first subset and the feedback for the respective medical document of the first subset, wherein the retraining includes the one or more of the first trained machine learning model or the second trained machine learning model repeatedly processing at least a portion of the one or more new training samples to cause an adjusting of one or more of the one or more first feature weights or the one or more second feature weights associated with the one or more data types until a determined loss or error associated with the one or more of the first trained machine learning model or the second trained machine learning model is below a predefined threshold.