Diabetes onset and progression prediction using a computerized model

1 . A system for improving accuracy in predicting the onset, progression, and severity of diabetes in an insured population using population segment specific modeling based on otherwise disparate, siloed sources, said system comprising:

databases comprising medical data for a plurality of members of an insured population, wherein said medical data is derived from a plurality of otherwise disparate, siloed data sources, including from a plurality of different providers for at least some of the plurality of members, including insurance claims data, lab test result data, clinical data, risk data, data indicating participation in health improvement programs, survey data, and demographic data, wherein at least one of said databases is publicly accessible and at least one of said databases is privately accessible, is associated with a health insurance provider for the insured population, and includes at least some of the insurance claims data;

one or more non-transitory electronic storage devices comprising software instructions, which when executed, configure the one or more processors to:

retrieve said medical data for the members of the population from the databases;

pre-process the retrieved data, including, summarizing, standardizing, and filtering said data received from each of said databases to increase homogeneity of said data;

perform a feature extraction subroutine on said retrieved medical data to extract characteristics for said members of the population;

segment the population into a plurality of segments by at least some of the extracted characteristics, wherein the characteristics comprise (i) date of membership to the population such that segmentation is performed between new members and existing members, (ii) line of business associated with the member such that segmentation is performed between members associated with different lines of business, and (iii) data availability of certain types of said medical data such that segmentation is performed between members associated with different types of data;

for each segment, apply a plurality of different machine learning models to the segment, apply each of the plurality of different machine learning models to a test set of data, determine which of said plurality of different machine learning models provides a highest level of accuracy relative to the test set of data using holdout data, and select the highest accuracy one of the plurality of different machine learning models for the segment;

apply the selected highest accuracy machine learning model for the segment to generate a score each of said members in the segment;

generate a scored member list for electronic display comprising identifying information for each of the members provided in association with the score for each of the members; and

automatically initiate medical interventions for at least some of the members in said scored member list based on the member's score, including automatically scheduling a visit by a healthcare provider for at least each member having a score in a first risk category, and automatically scheduling a call by a healthcare provider for at least each member having a score in a second risk category.

2 . The system of claim 1 wherein:

said medical data comprises health risk alerts, membership information, consumer information, health program information, CMS data, medical claims, and pharmaceutical claims.

3 . The system of claim 1 wherein:

said clinical data comprises claim counts, drug class counts, physician visit counts, and test costs;

said risk data comprises obesity, smoking, prescription risk score, and global risk;

said demographic data comprises race, education level, and active month information; and

said features comprise said claim counts, drug class counts, physician visit counts, test costs, obesity, smoking, prescription risk score, global risk, race, education level, and active month information.

4 . The system of claim 1 wherein:

the extracted features comprise a demographic profile, clinical profile, behavior profile, medication profile, and disease progression profile.

5 . The system of claim 4 wherein:

said demographic profile comprises age, gender, race and socio-economic status;

said clinical profile comprises chronic conditions, mental health conditions, hospitalizations, and medication;

said behavior profile comprises health program participations; and

said medication profile comprises adherence to various medications, including diabetes, heart failure, coronary artery disease.

6 . The system of claim 1 wherein:

each of said different machine learning models comprise at least one of: a neural network, logistic regression, and decision tree.

7 . The system of claim 6 wherein:

at least one of said different machine learning models comprise an ensemble model.

8 . The system of claim 1 wherein:

said one or more non-transitory electronic storage devices comprise additional software instructions, which when executed, configure the one or more processors to:

analyze said medical data for standardized codes indicating various diagnosed conditions;

said score is generated, at least in part, by applying a weight to each of said standardized codes associated with any one of: cardiovascular, cerebrovascular, metabolic, nephropathy, neuropathy, peripheral vascular disease, and retinopathy conditions; and

categorize each of said members of said population into a high, medium, or low complications risk category based on said score.

9 . The system of claim 8 wherein:

said weight comprises a one or a two for each of said conditions, except for neuropathy which is weighted a one;

each of said members having said diabetes complication score ranging from 0-3 are assigned into said low complications risk category;

each of said members having said diabetes complication score ranging from 4-7 are assigned into said medium complications risk category; and

each of said members having said diabetes complication score ranging from 8-13 are assigned into said high complications risk category.

10 . The system of claim 8 wherein:

said one or more non-transitory electronic storage devices comprise additional software instructions, which when executed, configure the one or more processors to automatically assign a first one of said medical interventions to each of said members categorized into said high complications risk category, assign a second one of said medical interventions to each of said members categorized into said medium complications risk category, and assign a third one of said medical interventions to each of said members categorized into said low complications risk category.

11 . The system of claim 10 wherein:

said one or more non-transitory electronic storage devices comprise additional software instructions, which when executed, configure the one or more processors to automatically schedule the visit by the healthcare provider to each of said members categorized into said high complications risk category, and automatically schedule the call by the healthcare provider to each of said members categorized into said medium complications risk category.

12 . A system for improving accuracy in predicting the onset, progression, and severity of diabetes complications in a population using population segment specific modeling based on otherwise disparate, siloed sources and automatically initiate member-specific interventions, said system comprising:

databases comprising medical data for a plurality of members of a population, said medical data comprising standardized codes for various diagnoses as part of medical and pharmaceutical insurance claims data and demographic data, said databases including information from a plurality of different, otherwise disparate, siloed data sources, including from a plurality of different providers for at least some of the plurality of members, with different types of available information, at least one publicly accessible source, and at least one privately accessible source;

one or more non-transitory electronic storage devices comprising software instructions, which when executed, configure the one or more processors to:

retrieve said medical data for the members of the population from each of the databases, including from the plurality of otherwise disparate, siloed data sources, and including data from multiple providers for at least some of the members;

perform a pre-processing subroutine on the data received from each of said databases, said pre-processing subroutine comprising multiple stages including summarizing, standardizing, and filtering said data received from each of said plurality of different otherwise disparate, siloed data sources which increases homogeneity of said data;

perform a domain-specific feature extraction subroutine on said pre-processed data which extracts clinically-significant characteristics for said members of the population, said clinically-significant characteristics comprising date of membership to the population, line of business associated with the members, and availability of certain types of said medical data;

segment the population into a plurality of different segments by the characteristics, wherein the characteristics comprise (i) date of membership to the population such that segmentation is performed between new members and existing members, (ii) line of business associated with the member such that segmentation is performed between members associated with different lines of business, and (iii) data availability of certain types of said medical data such that segmentation is performed between members associated with different types of data;

for each of the segments, engage a model selection engine to apply a plurality of different machine learning models to the data for the segment and to test data, said plurality of different machine learning models comprising at least one neural network based model, at least one logistic regression based model, at least one decision tree based model, and at least one ensemble model including a combination of the at least one neural network based model, the at least one logistic regression based model, and the at least one decision tree based model, and each of the plurality of different machine learning models implementing a multi-stage analysis, and select a most accurate one of the plurality of different machine learning models on a segment specific basis, where the most accurate model being determined by model performance against said test data using holdout data, derive a score for each of the members of the segment based on application of the most accurate one of the plurality of different machine learning models to the data for the member;

generate a scored member list comprising identifying information for each of the members provided in association with the score for each of the members;

generate a diabetes complication score for each of said members by applying a weight to each of said standardized codes in the medical data for the respective member associated with any one of: cardiovascular, cerebrovascular, metabolic, nephropathy, neuropathy, peripheral vascular disease, and retinopathy conditions, where the weight applied varies with the condition, and summing the weights;

categorize each of said members within said scored member list into a high, medium, or low complications risk category based on said diabetes complication score;

assign a first intervention to each of said members categorized into said high complications risk category, said first intervention comprising automatically scheduling a provider visit;

assign a second intervention to each of said members categorized into said medium complications risk category, said second intervention comprising automatically scheduling a provider call;

generate a graphical display at one or more remote electronic devices comprising said scored member list, indication of categorization of each of said members, and indication of assignment of said interventions; and

automatically initiate said interventions, where assigned.

13 . A system for improving accuracy in predicting the onset, progression, and severity of diabetes in a population using population segment specific machine learning modeling based on otherwise disparate, siloed sources, said system comprising:

databases comprising medical data for a plurality of members of a population, wherein said medical data is derived from a plurality of otherwise disparate, siloed data sources, at least one of which is publicly accessible database and at least one of which is privately accessible, said data sources including from a plurality of different providers for at least some of the plurality of members, health risk alerts, membership information, survey information, consumer information, health program information, CMS data, medical claims, pharmaceutical claims, and lab and test result information;

one or more non-transitory electronic storage devices comprising software instructions, which when executed, configure the one or more processors to:

retrieve said medical data for the members of the population from the databases;

pre-process the medical data as part of a multi-stage process including summarizing, standardizing, and filtering to increase homogeneity of the data;

perform a domain-specific feature extraction subroutine on said pre-processed medical data to extract clinically-significant characteristics for said members of the population, the clinically-significant extracted features comprising a demographic profile, clinical profile, behavior profile, medication profile, and disease progression profile, and the characteristics comprising date of membership to the population such that the segmentation is performed between new members and existing members, line of business associated with the member such that the segmentation is performed between members associated with different lines of business, and data availability of certain types of said medical data such that the segmentation is performed between members associated with different types of data;

segment the population into a plurality of segments by at least some of the extracted characteristics;

for each of the segments:

apply each of a plurality of different models to the pre-processed medical data for the respective segment;

apply each of the plurality of different models to a test set of data;

determine which of said plurality of different models provides a highest level of accuracy relative to the test set of data using holdout data; and

select and apply the highest accuracy one of the plurality of different models for the respective segment such that a different one of the plurality of different models is applied to the pre-processed medical data for each of the segments, respectively to generate a predicted onset of diabetes score for each of the members;

analyze the pre-processed to extract international classification of diseases (IDC) codes;

analyze the IDC codes to identify if any one or more of: cardiovascular, cerebrovascular, metabolic, nephropathy, neuropathy, peripheral vascular disease, and retinopathy conditions are indicated;

assign a point value to each of identified one of the conditions in the IDC codes, where said point values comprises a one or a two for each of said conditions, except for neuropathy which is weighted a one;

sum the point values to arrive at a diabetes progression score;

categorize the members based on the diabetes progression score, including categorizing each of said members having the diabetes progression score ranging from 0-3 into a low complications risk category, each of said members having the diabetes progression score ranging from 4-7 into a medium complications risk category, and each of said members having diabetes progression score ranging from 8-13 into said high complications risk category;

assign a first intervention to each of said members categorized into said high complications risk category, assign a second intervention to at least each of said members categorized into said medium complications risk category, and assign a third intervention to at least each of said members categorized into said low complications risk category, said first intervention including automatically scheduling a provider visit, and said second intervention including automatically scheduling a provider call;

generate a scored member list for electronic display for the population comprising identifying information for each of the members provided in association with the predicted onset of diabetes score and the diabetes progression score for each of the members;

cause electronic display of said scored member list; and

cause the assigned interventions to be carried out for the members.

14 . The system of claim 13 wherein:

said one or more non-transitory electronic storage devices comprise additional software instructions, which when executed, configure the one or more processors to:

use a plurality of different ones of the characteristics to segment the population;

use a plurality of different ones of the characteristics to segment a test set of data;

determine which of said plurality of different characteristics provides a highest level of accuracy relative to the test set of data;

select the highest accuracy one of the plurality of different characteristics to segment the population;

said medical data comprises health risk alerts, membership information, survey information, consumer information, health program information, CMS data, medical claims, pharmaceutical claims, and lab and test result information;

said clinical data comprises claim counts, drug class counts, physician visit counts, and test costs;

said risk data comprises obesity, smoking, prescription risk score, and global risk;

said demographic data comprises race, education level, and active month information;

said demographic profile comprises age, gender, race and socio-economic status;

said clinical profile comprises chronic conditions, mental health conditions, hospitalizations, and medication;

said behavior profile comprises health program participations;

said medication profile comprises adherence to various medications, including diabetes, heart failure, coronary artery disease; and

the feature comprise said claim counts, drug class counts, physician visit counts, test costs, obesity, smoking, prescription risk score, global risk, race, education level, and active month information.