Granted Patent B2

US 12,725,707 · App. 18/482,741 · Granted Sep 1, 2026

Inventors: Severence M. MacLaughlin (Palm Beach, FL); Ram Prasad Bora (Palm Beach, FL); Dhiraj Sharma (Palm Beach, FL)

Assignee: DeLorean Artificial Intelligence, Inc.

G16H50/20G06N5/022G16H20/40G16H50/30G16H50/50

Abstract

Claims (157)

1 . A method comprising:

receiving, by a computer system from a healthcare database, patient data, the patient data comprising structured healthcare data and unstructured healthcare data;

receiving, by a computer system, historical healthcare data, the historical healthcare data comprising:

historical structured healthcare data;

historical unstructured healthcare data; and

labeled historical outcome data;

tuning one or more hyperparameters of a risk stratification model to increase predictive power and improve processing speed of the risk stratification model;

training, by the computer system, the risk stratification model using the historical healthcare data;

generating, by a first determination engine based on application of the patient data to the risk stratification model, a patient risk stratification, the generating of the patient risk stratification comprising:

generating stratification scoring of the patient based on the patient data, the stratification scoring comprising:

a stratification score for the patient that is indicative of the patient's susceptibility to developing a medical condition; and a binary classification for the patient that is indicative of a positive or negative treatment trend for the patient;

turning one or more hyperparameters of a risk level assignment model to increase predictive power and improve porcessing speed of the risk level assignment model;

training, by the computer system, the risk level assignment model using the historical healthcare data and associated stratification scores and binary classifications;

generating, by a second determination engine based on application of the patient data to the risk level assignment model, a patient risk level assignment, the generating of the patient risk level assignment comprising:

generating risk level scoring of the patient; and

determining, based on the risk level scoring of the patient, a risk category for the patient;

tuning one or more hyperparameters of a next best actions model to increase predictive power and improve processing speed of the next best actions model;

training, by the computer system, the next best action model using labeled historical outcome data of the historical healthcare data and associated stratification scores, binary classifications, and risk categories for the patients;

generating, by a third determination engine based on application of the patient data and the stratification score, binary classification, and risk category for the patient, a set of patient next best actions, the generating of the set of patient next best actions comprising:

determining, based on the patient data, a patient outcome prediction; and

generating, based on the patient outcome prediction, the set of patient next best actions;

generating, by a fourth determination engine, a patient disease state transition prediction, the generating of the patient disease state transition prediction comprising:

determining, based on application of the patient data to a disease state transition prediction model, a set of transition probabilities for the patient;

generating, by a fifth determination engine, a patient unknown identification prediction, the generating of the patient unknown identification prediction comprising:

determining, based on application of the patient data to an unknown patient identification prediction model, a disease propensity score for the patient; and

generating, by a sixth determination engine based on the patient risk stratification, patient risk level assignment, patient next best actions, patient disease state transition predictions, and patient unknown identification predictions, a patient diagnosis report comprising a treatment action,

the treatment action for use in treating a patient.

2 . The method of claim 1 , wherein the first determination engine comprises the risk stratification model, and the training of the risk stratification model, comprising:

generating a binary risk classifier comprising determining, based on the historical patient healthcare data, key risk features and thresholds, wherein the binary risk classifier comprises the key risk features and thresholds, and

wherein the risk stratification model is trained to generate the patient risk stratification based on the binary risk classifier.

3 . The method of claim 2 , further comprising re-training the risk stratification model, comprising:

receiving, by a computer system, updated historical patient data, the updated historical patient data comprising updated structured historical healthcare data and unstructured historical healthcare data;

generating an updated binary risk classifier comprising determining, based on the updated historical patient healthcare data, key risk features and thresholds, wherein the binary risk classifier comprises the key risk features and thresholds;

generating an accuracy score of the updated binary risk classifier and the binary risk classifier;

determining, based on comparing the accuracy score of the updated binary risk classifier with the accuracy score of the binary risk classifier, that the updated risk classifier is more accurate than the binary risk classifier; and

overwriting the binary risk classifier with the updated risk classifier in response to determining that the updated binary risk classifier is more accurate than the binary risk classifier,

wherein the risk stratification model is re-trained to generate the patient risk stratification based on the updated binary risk classifier.

4 . The method of claim 1 , wherein the second determination engine comprises the risk level assignment model, and the training of the risk level assignment model, comprising:

generating a multi-class risk level classifier comprising determining, based on the historical healthcare patient data and the binary classification of the patient, key risk features, wherein the multi-class risk level classifier comprises the key risk features; and

wherein the risk level assignment model is trained to generate the patient risk level assignment based on the multi-class risk level classifier.

5 . The method of claim 4 , further comprising re-training the risk level assignment model, comprising:

receiving, by a computer system, updated historical patient data, the updated historical patient data comprising updated structured historical healthcare data and unstructured historical healthcare data;

generating an updated multi-class risk level classifier comprising determining, based on the updated historical patient healthcare data and the binary classification of the patient, key risk features, wherein the multi-class risk level classifier comprises the key risk features;

generating an accuracy score of the updated multi-class risk level classifier and the multi-class risk level classifier;

determining, based on comparing the accuracy score of the updated multi-class risk level classifier with the accuracy score of the multi-class risk level classifier, that the updated multi-class risk level classifier is more accurate than the multi-class risk level risk classifier; and

overwriting the multi-class risk level classifier with the updated multi-class risk level classifier in response to determining that the updated multi-class risk level classifier is more accurate than the multi-class risk level risk classifier,

wherein the risk level assignment model is re-trained to generate the patient risk level assignment based on the updated multi-class risk level classifier.

6 . The method of claim 1 , wherein the third determination engine comprises the next best actions model, and the training of the next best actions model, comprising:

generating a next best actions classifier comprising determining, based on the historical healthcare patient data, key risk features and thresholds, wherein the next best actions classifier comprises the key risk features and thresholds; and

wherein the next best actions model is trained to generate the set of patient next best actions based on the next best actions classifier.

7 . The method of claim 6 , further comprising re-training the next best actions model, comprising:

receiving, by a computer system, updated historical patient data, the updated historical patient data comprising updated structured historical healthcare data and unstructured historical healthcare data;

generating an updated next best actions classifier comprising determining, based on the updated historical patient healthcare data, key risk features and thresholds, wherein the next best actions classifier comprises the key risk features and thresholds;

generating an accuracy score of the updated next best actions classifier and the next best actions classifier;

determining, based on comparing the accuracy score of the updated next best actions classifier with the accuracy score of the next best actions classifier, that the next best actions classifier is more accurate than the next best actions classifier; and

overwriting the next best actions classifier with the updated next best actions classifier in response to determining that the updated next best actions classifier is more accurate than the next best actions classifier,

wherein the next best actions model is re-trained to generate the set of patient next best actions based on the updated next best actions classifier.

8 . The method of claim 1 , wherein the fourth determination engine comprises the disease state transition prediction model, and the method further comprising training the disease state transition prediction model, comprising:

generating a disease state transition classifier comprising determining, based on the historical healthcare patient data, key disease state features and thresholds, wherein the disease state transition classifier comprises the key disease state features and thresholds; and

wherein the disease state transition prediction model is trained to generate the patient disease state transition prediction based on the disease state transition classifier.

9 . The method of claim 8 , further comprising re-training the disease state transition prediction model, comprising:

receiving, by a computer system, updated historical patient data, the updated historical patient data comprising updated structured historical healthcare data and unstructured historical healthcare data;

generating an updated disease state transition classifier comprising determining, based on the updated historical patient healthcare data, key disease state features and thresholds, wherein the disease state transition classifier comprises the key disease state features and thresholds;

generating an accuracy score of the updated disease state transition classifier and the disease state transition classifier;

determining, based on comparing the accuracy score of the updated disease state transition classifier with the accuracy score of the disease state transition classifier, that the disease state transition classifier is more accurate than the disease state transition classifier; and

overwriting the disease state transition classifier with the updated disease state transition classifier in response to determining that the updated disease state transition classifier is more accurate than the disease state transition classifier,

wherein the disease state transition prediction model is re-trained to the patient disease state transition prediction based on the updated disease state transition classifier.

10 . The method of claim 1 , wherein the fifth determination engine comprises the unknown patient identification prediction model, and the method further comprising training the unknown patient identification prediction model, comprising:

generating a patient identification classifier comprising determining, based on the historical healthcare patient data, key patient features and thresholds, wherein the patient identification classifier comprises the key patient features and thresholds; and

wherein the unknown patient identification prediction model is trained to generate the patient unknown identification prediction based on the patient identification classifier.

11 . The method of claim 10 , further comprising re-training the unknown patient identification prediction model, comprising:

receiving, by a computer system, updated historical patient data, the updated historical patient data comprising updated structured historical healthcare data and unstructured historical healthcare data;

generating an updated patient identification classifier comprising determining, based on the updated historical patient healthcare data, key patient features and thresholds, wherein the patient identification classifier comprises the key patient features and thresholds;

generating an accuracy score of the updated patient identification classifier and the patient identification classifier;

determining, based on comparing the accuracy score of the updated a-patient identification classifier with the accuracy score of the patient identification classifier, that the patient identification classifier is more accurate than the patient identification classifier; and

overwriting the patient identification classifier with the updated patient identification classifier in response to determining that the updated patient identification classifier is more accurate than the patient identification classifier,

wherein the unknown patient identification prediction model is re-trained to generate the set of patient unknown identification prediction based on the updated patient identification classifier.

12 . The method of claim 1 , the method further comprising training the next best actions model to determine next best actions for ESRD patients comprising:

generating a next best actions classifier comprising:

determining, based on historical healthcare patient data, likelihood of patient hospitalization in patients due to fluid overload;

determining, based on the historical healthcare patient data, likelihood of patient missing hospital appointment;

determining, based on the historical healthcare patient data and a dialysis adequacy criterion, likelihood of patient having abnormal dialysis adequacy;

determining, based on the historical healthcare patient data, optimal dry weight in patients due to dialysis; and

determining, based on the historical healthcare patient data, likelihood of patient ESA or IS dosages required to be altered,

wherein the next best actions classifier comprises the likelihood of patient hospitalization in patients due to fluid overload, likelihood of patient missing hospital appointment, likelihood of patient having abnormal dialysis adequacy, optimal dry weight in patients due to dialysis, and likelihood of patient ESA or IS dosages required to be altered; and

wherein the next best actions model is trained to generate the set of patient next best actions based on the next best actions classifier.

13 . The method of claim 12 , further comprising re-training the next best actions model, comprising:

receiving, by a computer system, updated historical patient data, the updated historical patient data comprising updated structured historical healthcare data and unstructured historical healthcare data;

generating an updated next best actions classifier comprising:

determining, based on updated historical healthcare patient data, likelihood of patient hospitalization in patients due to fluid overload;

determining, based on the updated historical healthcare patient data, likelihood of patient missing hospital appointment;

determining, based on the updated historical healthcare patient data and a dialysis adequacy criterion, likelihood of patient having abnormal dialysis adequacy;

determining, based on the updated historical healthcare patient data, optimal dry weight in patients due to dialysis;

determining, based on the updated historical healthcare patient data, likelihood of patient ESA or IS dosages required to be altered;

wherein the next best actions classifier comprises the likelihood of patient hospitalization in patients due to fluid overload, likelihood of patient missing hospital appointment, likelihood of patient having abnormal dialysis adequacy, optimal dry weight in patients due to dialysis, and likelihood of patient ESA or IS dosages required to be altered;

generating an accuracy score of the updated next best actions classifier and the next best actions classifier;

determining, based on comparing the accuracy score of the updated next best actions classifier with the accuracy score of the next best actions classifier, that the next best actions classifier is more accurate than the next best actions classifier; and

overwriting the next best actions classifier with the updated next best actions classifier in response to determining that the updated next best actions classifier is more accurate than the next best actions classifier,

wherein the next best actions model is re-trained to generate the set of patient next best actions based on the updated next best actions classifier.

14 . A non-transitory computer readable storage medium comprising program instructions stored thereon that are executable by a processor to cause the following operations:

receiving, by a computer system from a healthcare database, patient data, the patient data comprising structured healthcare data and unstructured healthcare data;

receiving, by a computer system, historical healthcare data, the historical healthcare data comprising:

historical structured healthcare data;

historical unstructured healthcare data; and

labeled historical outcome data;

tuning one or more hyperparameters of a risk stratication model to increase predictive power and improve processing speed of the risk stratification model;

training, by the computer system, the risk stratification model using the historical healthcare data;

generating, by a first determination engine based on application of the patient data to the risk stratification model, a patient risk stratification, the generating of the patient risk stratification comprising:

generating stratification scoring of the patient based on the patient data, the stratification scoring comprising:

a stratification score for the patient that is indicative of the patient's susceptibility to developing a medical condition; and

a binary classification for the patient that is indicative of a positive or negative treatment trend for the patient;

tuning one or more hyperarameters of a risk level assignment model to increase predictive power and improve processing spped of the risk level assignment model;

training, by the computer system, the risk level assignment model using the historical healthcare data and associated stratification scores and binary classifications;

generating, by a second determination engine based on application of the patient data to the risk level assignment model, a patient risk level assignment, the generating of the patient risk level assignment comprising:

generating risk level scoring of the patient; and

determining, based on the risk level scoring of the patient, a risk category for the patient;

tuning one or more hyperparameters of a next best actions model to incrase predictive power and improve processing speed of the next best actions model;

training, by the computer system, the next best actions model using labeled historical outcome data of the historical healthcare data and associated stratification scores, binary classifications, and risk categories for the patients;

generating, by a third determination engine based on application of the patient data and the stratification score, binary classification, and risk category for the patient, a set of patient next best actions, the generating of the set of patient next best actions comprising:

determining, based on the patient data, a patient outcome prediction; and

generating, based on the patient outcomes prediction, the set of patient next best actions;

generating, by a fourth determination engine, a patient disease state transition prediction, the generating of the patient disease state transition prediction comprising:

determining, based on application of the patient data to a disease state transition prediction model, a set of transition probabilities for the patient;

generating, by a fifth determination engine, a patient unknown identification prediction, the generating of the patient unknown identification prediction comprising:

determining, based on application of the patient data to an unknown patient identification prediction model, a disease propensity score for the patient; and

generating, by a sixth determination engine based on the patient risk stratification, patent risk level assignment, patient next best actions, patient disease state transition predictions, and patient unknown identification predictions, a patient diagnosis report comprising a treatment action,

the treatment action for use in treating a patient.

15 . A system comprising:

a processor; and

non-transitory computer readable storage medium comprising program instructions stored thereon that are executable by the processor to cause the following operations:

receiving, by a computer system from a healthcare database, patient data, the patient data comprising structured healthcare data and unstructured healthcare data;

receiving, by a computer system, historical healthcare data, the historical healthcare data comprising:

historical structured healthcare data;

historical unstructured healthcare data; and

labeled historical outcome data;

tuning one or more hyperparameters of a risk stratification model to increase predictive power and improve processing speed of the risk stratification model;

training, by the computer system, the risk stratification model using the historical healthcare data;

generating, by a first determination engine based on application of the patient data to the risk stratification model, a patient risk stratification, the generating of the patient risk stratification comprising:

generating stratification scoring of the patient based on the patient data, the stratification scoring comprising:

a stratification score for the patient that is indicative of the patient's susceptibility to developing a medical condition; and

a binary classification for the patient that is indicative of a positive or negative treatment trend for the patient;

tuning one or more hyperparameters of a risk level assignment model to increase predictive power and improve processing speed of the risk level assignment model;

training, by the computer system, the risk level assignment model using the historical healthcare data and associated stratification scores and binary classifications;

generating, by a second determination engine based on application of the patient data to the risk level assignment model, a patient risk level assignment, the generating of the patient risk level assignment comprising:

generating risk level scoring of the patient; and

determining, based on the risk level scoring of the patient, a risk category for the patient;

tuning one or more hyperparameters of a next best actions model to increase predictive power and improve processing speed of the next best actions model;

training, by the computer system, the next best actions model using labeled historical outcome data of the historical healthcare data and associated stratification scores, binary classifications, and risk categories for the patients;

generating, by a third determination engine based on application of the patient data and the stratification score, binary classification, and risk category for the patient, a set of patient next best actions, the generating of the set of patient next best actions comprising:

determining, based on the patient data, a patient outcome prediction; and

generating, based on the predictions of patient outcomes prediction, the set of patient next best actions;

generating, by a fourth determination engine, a patient disease state transition prediction, the generating of the patient disease state transition prediction comprising:

determining, based on application of the patient data to a disease state transition prediction model, a set of transition probabilities for the patient;

generating, by a fifth determination engine, a patient unknown identification prediction, the generating of the patient unknown identification prediction comprising:

determining, based on application of the patient data to an unknown patient identification prediction model, a disease propensity score for the patient; and

generating, by a sixth determination engine based on the patient risk stratification, patent risk level assignment, patient next best actions, patient disease state transition predictions, and patient unknown identification predictions, a patient diagnosis report comprising a treatment action,

the treatment action for use in treating a patient.

Assignments (1)

Continuity (1)

References Cited (118)