Claim analysis with deep learning
Embodiments relate to system for automatically predicting payer response to claims. In an embodiment, the system receives claim data associated with a claim. The system identifies a set of claim features of the claim data, and generates an input vector with at least a portion of the set of claim features. The system applies the input vector to a trained model. A first portion of the neural network is configured to generate an embedding representing the input vector with a lower dimensionality than the input vector. A second portion of the neural network is configured to generate a prediction of whether the claim will be denied based on the embedding. The system provides the prediction for display on a user interface of a user device. The prediction may further include denial reason codes and a response date estimation to indicate if, when, and why a claim will be denied.
1. A method comprising: training a first portion of one or more neural networks implemented on computer, the first portion of the one or more neural networks to generate an embedding based on a first vector of claim features, the first vector comprising at least features of a claim, the embedding comprising a second vector having a lower dimensionality than the first vector, the second vector comprising at least a first sub-vector to express one or more diagnosis tokens and a second sub-vector to express one or more procedure tokens; and
training a second portion of one or more neural networks to generate a prediction of a payer's response to the claim based, at least in part, on the embedding, the prediction of the payer's response to the claim comprising at least a likelihood that the claim would be denied,
wherein training the first portion of the one or more neural networks and training the second portion of the one or more neural networks comprises determining weights for the first and second portions of the one or more neural networks based, at least in part, on a computed prediction of payer's response and a label representing a payer's response.
2. The method of claim 1 , wherein the prediction of the payer's response to the claim further comprises an indication of a response date estimate or a reason for denial, or a combination thereof.
3. The method of claim 1 , wherein the first vector comprises an indication of one or more procedures performed and one or more diagnoses received.
4. The method of claim 3 , wherein the first vector further comprises an indication of patient gender, payer state, total charges, service date, payer identifier, individual relationship code or duration of service, or a combination thereof.
5. The method of claim 1 , wherein a length of the first sub-vector is based, at least in part, on a number of possible diagnoses and a length of the second sub-vector is based, at least in part, on a number of possible procedures.
6. The method of claim 1 , wherein the second vector represents demographic information as binary values, and represents procedures and diagnoses as numeric values.
7. The method of claim 1 , wherein the generated prediction of the payer's response further comprises one or more suspiciousness scores to express a contribution of one or more of the claim features on a denial prediction.
8. The method of claim 7 , wherein the one or more suspiciousness scores are computed based, at least in part, on magnitudes of gradients of prediction scores for values of the features of the claim.
9. The method of claim 7 , wherein at least one of the one or more suspiciousness scores reflect an impact of individual features on a prediction of a claim denial.
10. The method of claim 1 , wherein training the first portion of the one or more neural networks and training the second portion of the one or more neural networks further comprises training a machine-learning model according to a loss function, the loss function being based, at least in part, on a binary cross-entropy loss for a denial probability prediction, a categorical cross-entropy loss for a set of service-level denial reason code classifications or a categorical cross-entropy loss for a set of claim-level denial reason code classifications, or a combination thereof.
11. The method of claim 1 , wherein the second portion of one or more neural networks to generate the prediction of a payer's response based, at least in part, on a first reason code sequence, a second reason code sequence and a response date estimation, wherein the first reason code sequence includes likelihood scores for claim-level reason codes in a set of claim-level reason codes, and wherein the second reason code sequence includes likelihood scores for service-level reason codes in a set of service-level reason codes.
12. The method of claim 11 , wherein the one or more neural networks comprises:
a first set of task-specific output layers to generate the first reason code sequence,
a second set of task-specific output layers to generate the second reason code sequence, and
a third set of task-specific layers to generate the response date estimation.