Method, apparatus, and computer program product for using machine learning to encode a healthcare claim as a predefined sized vector
A method, apparatus and computer program product are provided for generating a predefined sized vector representative of a healthcare claim including a variable number of service lines. A neural network is trained to identify hierarchical relationships between claim lines and services lines and to represent the relationships in the predefined sized vector. The predefined sized vector may be used to make predictions regarding adjudication, such as probability of denial (claim-level or service-level), days to pay, probability of being paid as a diagnosis related group (DRG), allowed amount, and/or denial reason. Predictions may also be made from the predefined sized vector relating to the likelihood and/or identification of missing procedure codes.
1. An apparatus for utilizing a trained encoder to encode, as predefined sized vectors, healthcare claims comprising a variable number of service lines, the apparatus comprising at least one processor and at least one memory including computer program code, the at least one memory and the computer program code configured to, with the processor, cause the apparatus to at least:
receive a claim-level record associated with a healthcare claim, wherein the claim-level record comprises attributes describing a patient visit for healthcare services;
receive one or more service lines related to the claim-level record, wherein each service line comprises attributes describing services rendered during the patient visit;
embed claim-level features from the claim-level record to generate a claim-level vector;
embed service line features from each of the service lines to generate a service line vector for each service line;
combine the claim-level vector with the one or more service line vectors to generate a predefined sized vector representative of the healthcare claim, the predefined sized vector reflecting hierarchical relationships between at least some of the service lines and the claim-level record, enabling a prediction engine to generate predictions; and
train the encoder by inputting a plurality of training healthcare claims, each training healthcare claim comprising a claim-level record and one or more services lines related to the claim-level record, and each training healthcare claim further comprising a confirmed characteristic;
wherein each of the training healthcare claims has an associated predefined sized vector that, when plotted on an embedding space, is positioned closer to predefined sized vectors representative of other healthcare claims comprising more similar characteristics, relative to other predefined sized vectors representative of other healthcare claims comprising less similar characteristics.
2. The apparatus of claim 1 , wherein the generated predictions relate to adjudication of at least one of the claim-level record or the one or more service lines.
3. The apparatus of claim 1 , wherein the generated predictions relate to information expected in the claim-level record or one or more service lines.
4. The apparatus of claim 1 , wherein the generated prediction comprises at least one of:
probability of denial;
days to pay;
probability of being paid as a diagnosis related group (DRG);
allowed amount;
denial reason; or
missing procedure codes.
5. The apparatus of claim 1 , wherein the claim-level record comprises at least one of patient information, payer information, or a principal diagnosis code.
6. The apparatus of claim 1 , wherein at least one of the service lines comprises at least one of a procedure code, a procedure modifier, a revenue code, or a type of facility.
7. The apparatus of claim 1 ,
wherein embedding the claim-level features from the claim-level record to generate the claim-level vector comprises:
for each claim-level feature, generating an N-dimensional claim-level vector; and
concatenating the N-dimensional claim-level vectors to generate an M*N-dimensional claim-level vector representing the claim header, wherein M represents a number of claim-level features,
wherein embedding the service line service line features from each of the service lines to generate a service line vector for each service line comprises:
for each service line feature, generating an N-dimensional service line vector; and
for each service line, generate an S*N-dimensional service line vector, wherein S is the number of service line features in each service line;
generate an L-length sequence of S*N-dimensional service line vectors, wherein L is the number of service lines;
wherein generating the predefined sized vector comprises:
processing the M*N-dimensional claim-level vector and the L-length sequence of S*N-dimensional service line vectors through neural network layers to (a) identify the hierarchical relationships, and (b) reduce dimensionality to D by generating a D-dimensional claim header vector and L D-dimensional line vectors;
pooling the L D-dimensional line vectors to generate a D-dimensional service-line vector; and
concatenating the D-dimensional claim header vector with the D-dimensional service-line vector to generate the predefined sized vector as a D*2-dimensional vector.
8. A method for utilizing a trained encoder to encode, as predefined sized vectors, healthcare claims comprising a variable number of service lines, the method comprising:
receiving a claim-level record associated with a healthcare claim, wherein the claim-level record comprises attributes describing a patient visit for healthcare services;
receiving one or more service lines related to the claim-level record, wherein each service line comprises attributes describing services rendered during the patient visit;
embedding, with at least a processor, claim-level features from the claim-level record to generate a claim-level vector;
embedding service line features from each of the service lines to generate a service line vector for each service line;
combining the claim-level vector with the one or more service line vectors to generate a predefined sized vector representative of the healthcare claim, the predefined sized vector reflecting hierarchical relationships between at least some of the service lines and the claim-level record, enabling a prediction engine to generate predictions; and
training the encoder by inputting a plurality of training healthcare claims, each training healthcare claim comprising a claim-level record and one or more services lines related to the claim-level record, and each training healthcare claim further comprising a confirmed characteristic;
wherein each of the training healthcare claims has an associated predefined sized vector that, when plotted on an embedding space, is positioned closer to predefined sized vectors representative of other healthcare claims comprising more similar characteristics, relative to other predefined sized vectors representative of other healthcare claims comprising less similar characteristics.
9. The method of claim 8 , wherein the generated predictions relate to adjudication of at least one of the claim-level record or the one or more service lines.
10. The method of claim 8 , wherein the generated predictions relate to information expected in the claim-level record or one or more service lines.
11. The method of claim 8 , wherein the generated prediction comprises at least one of:
probability of denial;
days to pay;
probability of being paid as a diagnosis related group (DRG);
allowed amount;
denial reason; or
missing procedure codes.
12. The method of claim 8 , wherein the claim-level record comprises at least one of patient information, payer information, or a principal diagnosis code.
13. The method of claim 8 , wherein at least one of the service lines comprises at least one a procedure code, a procedure modifier, a revenue code, or a type of facility.
14. The method of claim 8 , wherein embedding the claim-level features from the claim-level record to generate the claim-level vector comprises:
for each claim-level feature, generating an N-dimensional claim-level vector; and
concatenating the N-dimensional claim-level vectors to generate an M*N-dimensional claim-level vector representing the claim header, wherein M represents a number of claim-level features,
wherein embedding the service line service line features from each of the service lines to generate a service line vector for each service line comprises:
for each service line feature, generating an N-dimensional service line vector; and
for each service line, generate an S*N-dimensional service line vector, wherein S is the number of service line features in each service line;
generate an L-length sequence of S*N-dimensional service line vectors, wherein L is the number of service lines;
wherein generating the predefined sized vector comprises:
processing the M*N-dimensional claim-level vector and the L-length sequence of S*N-dimensional service line vectors through neural network layers to (a) identify the hierarchical relationships, and (b) reduce dimensionality to D by generating a D-dimensional claim header vector and L D-dimensional line vectors;
pooling the L D-dimensional line vectors to generate a D-dimensional service-line vector; and
concatenating the D-dimensional claim header vector with the D-dimensional service-line vector to generate the predefined sized vector as a D*2-dimensional vector.
15. A computer program product for utilizing a trained encoder to encode, as predefined sized vectors, healthcare claims comprising a variable number of service lines, the computer program product comprising at least one non-transitory computer-readable storage medium having computer-executable program code instructions stored therein, the computer-executable program code instructions comprising program code instructions to:
receive a claim-level record associated with a healthcare claim, wherein the claim-level record comprises attributes describing a patient visit for healthcare services;
receive one or more service lines related to the claim-level record, wherein each service line comprises attributes describing services rendered during the patient visit;
embed claim-level features from the claim-level record to generate a claim-level vector;
embed service line features from each of the service lines to generate a service line vector for each service line;
combine the claim-level vector with the one or more service line vectors to generate a predefined sized vector representative of the healthcare claim, the predefined sized vector reflecting hierarchical relationships between at least some of the service lines and the claim-level record, enabling a prediction engine to generate predictions; and
train the encoder by inputting a plurality of training healthcare claims, each training healthcare claim comprising a claim-level record and one or more services lines related to the claim-level record, and each training healthcare claim further comprising a confirmed characteristic;
wherein each of the training healthcare claims has an associated predefined sized vector that, when plotted on an embedding space, is positioned closer to predefined sized vectors representative of other healthcare claims comprising more similar characteristics, relative to other predefined sized vectors representative of other healthcare claims comprising less similar characteristics.
16. The computer program product of claim 15 , wherein embedding the claim-level features from the claim-level record to generate the claim-level vector comprises:
for each claim-level feature, generating an N-dimensional claim-level vector; and
concatenating the N-dimensional claim-level vectors to generate an M*N-dimensional claim-level vector representing the claim header, wherein M represents a number of claim-level features,
wherein embedding the service line service line features from each of the service lines to generate a service line vector for each service line comprises:
for each service line feature, generating an N-dimensional service line vector; and
for each service line, generate an S*N-dimensional service line vector, wherein S is the number of service line features in each service line;
generate an L-length sequence of S*N-dimensional service line vectors, wherein L is the number of service lines;
wherein generating the predefined sized vector comprises:
processing the M*N-dimensional claim-level vector and the L-length sequence of S*N-dimensional service line vectors through neural network layers to (a) identify the hierarchical relationships, and (b) reduce dimensionality to D by generating a D-dimensional claim header vector and L D-dimensional line vectors;
pooling the L D-dimensional line vectors to generate a D-dimensional service-line vector; and
concatenating the D-dimensional claim header vector with the D-dimensional service-line vector to generate the predefined sized vector as a D*2-dimensional vector.