| CPC G16H 50/20 (2018.01) [G16H 10/60 (2018.01); G16H 50/30 (2018.01); G16H 50/50 (2018.01); G06N 3/084 (2013.01)] | 20 Claims |
|
1. A method for providing a patient prediction, the method comprising:
responsive to receiving an indication of initiation of a medical event for a patient, initiating a deep learning machine learning model for the patient, wherein the deep learning machine learning model (a) comprises a modified recurrent neural network (RNN) with gated recurrent units (GRUs) and a self-attention mechanism, (b) is executed by an assessment computing entity, (c) has been trained using machine learning, and (d) is configured to generate a prediction for the patient comprising at least one of an acuity score or a mortality prediction;
responsive to determining that a configurable time period has elapsed since a previous execution of the deep learning machine learning model for the patient, identifying, by the assessment computing entity, a prediction trigger;
responsive to identifying the prediction trigger, requesting, by the assessment computing entity, updated medical data for the patient;
responsive to receiving the updated medical data for the patient, providing, by the assessment computing entity, the updated medical data as input to the deep learning machine learning model;
executing, by the assessment computing entity, the deep learning machine learning model for the patient to cause the deep learning machine learning model to:
determine a respective hidden state for each time step of one or more time steps of medical data, wherein each time step of medical data of the one or more time steps of medical data comprises a respective plurality of measurements corresponding to the time step and corresponding to the patient and the one or more times steps of medical data includes the update medical data,
determine a respective attention parameter for each time step of the one or more time steps,
determine a weighted average hidden state by aggregating a time series of respective weighted hidden states of the deep learning machine learning model for the patient, each respective weighted hidden state of the time series of respective weighted hidden states corresponding to a respective time step of the one or more time steps, wherein the respective weighted hidden state is determined based at least in part on a respective weight determined based at least in part on the respective attention parameter for a respective time step corresponding to the respective hidden state,
wherein the self-attention mechanism determines the attention parameter for each time step of the one or more time steps and the attention parameter for a time step indicates an influence of the time step on the prediction, and
automatically generate the prediction by providing the weighted average hidden state to a classification layer of the deep learning machine learning model;
automatically providing, by the assessment computing entity, at least a portion of the prediction and a self-attention parameter distribution such that at least one of (a) the at least a portion of the prediction and the self-attention parameter distribution is used to update an electronic health record corresponding to the patient or (b) displaying to a clinician for review;
determining an updated configurable time period based at least in part on the prediction; and
responsive to determining that the updated configurable time period has elapsed since the prediction was determined, identifying another prediction trigger, requesting new medical data, receiving the new medical data in response to the request, and updating the deep learning machine learning model for the patient, based at least in part on the new medical data, to automatically determine and provide a new prediction and a new self-attention parameter distribution.
|