Methods for evaluating clinical comparative efficacy using real-world health data and artificial intelligence

A system for evaluating clinical comparative efficacy using real-world health data, comprising:

• one or more processors;
• memory;
• and
• one or more programs stored in the memory, wherein the one or more programs are configured for execution by the one or more processors and include instructions for:
• obtaining health trajectories for members of a healthcare system, wherein each health trajectory corresponds to a respective member and represents a time-ordered series of events for the respective member, and each event represents an interaction between the respective member and the healthcare system;
• identifying index events in the health trajectories for the members, wherein an index event is any clinical or health data point;
• segmenting the health trajectories with index events into a plurality of sub-trajectories such that each sub-trajectory ends at a different index event;
• generating a digital fingerprint for each sub-trajectory by applying representation learning to generate an embedding vector, including training a code embedding machine learning model using at least a portion of the plurality of sub-trajectories to generate embedding vectors;
• identifying one or more sub-trajectories that are similar to a patient sub-trajectory by performing a nearest-neighbor search on the embedding vectors of the members, wherein the patient sub-trajectory corresponds to a patient member of the healthcare system, wherein identifying the one or more sub-trajectories comprises:
• selecting, from the plurality of sub-trajectories, one or more treatment sub-trajectories of the members that have taken a same prior treatment option as the patient member;
• grouping the one or more treatment sub-trajectories into a control arm set of trajectories that includes sub-trajectories that continued taking the same treatment option as the same prior treatment option after an indexed event, and one or more treatment arm sets of trajectories that each include sub-trajectories that were prescribed a respective different treatment option after the indexed event;
• computing a control arm average treatment effect for the control arm set of trajectories based on analyzing the sub-trajectories, in the control arm set of trajectories, after the indexed event;
• computing a respective treatment arm average treatment effect, for each of the one or more treatment arm sets of trajectories, based on analyzing the respective sub-trajectories, in the respective treatment arm sets of trajectories, after the indexed event;
• and
• predicting effects of the treatment strategies by averaging the control arm average treatment effect and the treatment arm average treatment effects;
• and
• identifying treatment strategies that are most likely to be efficacious for the patient member based on outcomes associated with each of the treatment strategies, according to the one or more sub-trajectories,
• wherein applying representation learning comprises:
• obtaining code embeddings corresponding to claim codes representing events in a sub-trajectory;
• inputting the code embeddings to a recurrent neural network (RNN) to obtain hidden states for different layers for a terminal index event, wherein the RNN back-propagates error while learning embeddings during training;
• and
• combining the hidden states for different layers of the RNN to obtain the embedding vectors for the sub-trajectory.

A non-transitory computer readable storage medium storing one or more programs configured for execution by an electronic device with a display, the one or more programs comprising instructions for:

• obtaining health trajectories for members of a healthcare system, wherein each health trajectory corresponds to a respective member and represents a time-ordered series of events for the respective member, and each event represents an interaction between the respective member and the healthcare system;
• identifying index events in the health trajectories for the members, wherein an index event is any clinical or health data point;
• segmenting the health trajectories with index events into a plurality of sub-trajectories such that each sub-trajectory ends at a different index event;
• generating a digital fingerprint for each sub-trajectory by applying representation learning to generate an embedding vector, including training a code embedding machine learning model using at least a portion of the plurality of sub-trajectories to generate embedding vectors;
• identifying one or more sub-trajectories that are similar to a patient sub-trajectory by performing a nearest-neighbor search on the embedding vectors of the members, wherein the patient sub-trajectory corresponds to a patient member of the healthcare system, wherein identifying the one or more sub-trajectories comprises:
• selecting, from the plurality of sub-trajectories, one or more treatment sub-trajectories of the members that have taken a same prior treatment option as the patient member;
• grouping the one or more treatment sub-trajectories into a control arm set of trajectories that includes sub-trajectories that continued taking the same treatment option as the same prior treatment option after an indexed event, and one or more treatment arm sets of trajectories that each include sub-trajectories that were prescribed a respective different treatment option after the indexed event;
• computing a control arm average treatment effect for the control arm set of trajectories based on analyzing the sub-trajectories, in the control arm set of trajectories, after the indexed event;
• computing a respective treatment arm average treatment effect, for each of the one or more treatment arm sets of trajectories, based on analyzing the respective sub-trajectories, in the respective treatment arm sets of trajectories, after the indexed event;
• and
• predicting effects of the treatment strategies by averaging the control arm average treatment effect and the treatment arm average treatment effects;
• and
• identifying treatment strategies that are most likely to be efficacious for the patient member based on outcomes associated with each of the treatment strategies, according to the one or more sub-trajectories,
• wherein applying representation learning comprises:
• obtaining code embeddings corresponding to claim codes representing events in a sub-trajectory;
• inputting the code embeddings to a recurrent neural network (RNN) to obtain hidden states for different layers for a terminal index event, wherein the RNN back-propagates error while learning embeddings during training;
• and
• combining the hidden states for different layers of the RNN to obtain the embedding vectors for the sub-trajectory.