CLINICAL PREDICTIVE ANALYTICS SYSTEM

• US 20150112710A1
• Filed: 12/20/2014
• Published: 04/23/2015
• Est. Priority Date: 06/21/2012
• Status: Abandoned Application

• Current Assignee: Battelle Memorial Institute
• Original Assignee: Battelle Memorial Institute
• Inventors: Rust, Steven W., Haber, Daniel
• Application Number:
  US14/578,396
• Publication Number:
  US 20150112710A1
• Time in Patent Office: Days
• US Class Current: 705/2
• CPC Class Codes: G16H 50/30 for calculating health indices; for individual health risk assessment
• CPC Class Codes: G16H 50/50 for simulation or modelling of medical disorders

Abstract

Predictive models are built for the estimation of adverse health likelihood by identifying candidate model risk variables, constructing a model form for an outcome likelihood model that estimates the likelihood of an adverse outcome type using a group of risk variables selected from the set of candidate model risk variables and by classifying each selected risk variable into either a baseline group or a dynamic group. Additionally, predictive models are built by constructing separate baseline and dynamic outcome likelihood model forms and by fitting the constructed model forms to a training data set to produce final models to be used as scoring functions that compute a baseline outcome likelihood and a dynamic outcome likelihood for patient data that is not represented in the training data set. The predictive models can be used with alerting and attribution algorithms to predict the likelihood of an adverse outcome for individuals receiving care.

Claims

1. A method implemented on a computer, for building predictive models for the estimation of adverse health likelihood, comprising:

identifying a set of candidate model risk variables that are associated with an adverse outcome type;

constructing, utilizing the computer, an outcome likelihood model form that estimates the likelihood of the adverse outcome type using risk variables that are selected from the set of candidate model risk variables;

classifying each of the selected risk variables into either a baseline group or a dynamic group, wherein;

the baseline group is composed of those selected risk variables that are non-modifiable based on medical care that is provided to a patient; and

the dynamic group is composed of those selected risk variables that are modifiable based on the medical care that is provided to the patient;

constructing, utilizing the computer, dynamic risk variable model forms that predict values for the selected risk variables in the dynamic group as a function of at least one of the risk variables in the baseline group;

constructing, utilizing the computer, a baseline outcome likelihood model form associated with the adverse outcome type using the outcome likelihood model form and the dynamic risk variable model forms;

constructing, utilizing the computer, a dynamic outcome likelihood model form associated with the adverse outcome type using the outcome likelihood model form and at least one of the selected risk variables; and

fitting the constructed outcome likelihood model form, baseline outcome likelihood model form, and dynamic outcome likelihood model form, to a training data set that includes both outcome data and data values that correspond to the selected risk variables to produce an outcome likelihood model, a baseline outcome likelihood model, and a dynamic outcome likelihood model, which are used as scoring functions to compute a baseline outcome likelihood and a dynamic outcome likelihood for patient data that is not represented in the training data set.

2. The method of claim 1 further comprising:

generating an alerting algorithm that compares computed values for at least one of the overall outcome likelihood, baseline outcome likelihood, and dynamic outcome likelihood against predetermined thresholds, where the alerting algorithm thresholds are based upon a balancing of true positive and false positive behavior; and

generating an attribution assessment for the adverse outcome type by;

evaluating the contribution of the selected risk variables with respect to at least one of the baseline outcome likelihood and the dynamic outcome likelihood associated with an occurrence of an adverse outcome; and

indicating at least one evaluated risk variable that is determined to be a key contributor to the overall outcome likelihood.

3. The method of claim 2, wherein:

evaluating the contributions of the selected risk variables comprises utilizing the computer for producing a vector of indices characterizing the strength of the contributions of individual risk variables to at least one of the baseline outcome likelihood and the dynamic outcome likelihood.

4. The method of claim 3, wherein:

utilizing the computer for producing a vector of indices comprises producing the vector such that the indices sum to 100 and can be interpreted as percentage contributions to the outcome likelihood.

5. The method of claim 2 further comprising:

mapping actual patient data that is not represented in the training data set to the selected risk variables;

applying, by the computer, the mapped patient data in the selected risk variables to the outcome likelihood model, the baseline outcome likelihood model, and the dynamic outcome likelihood model, to compute the baseline outcome likelihood and dynamic outcome likelihood for the actual patient data; and

outputting a presentation of the results as predictions of the likelihoods of the adverse outcome for the patient represented by the actual patient data in response to results from the alerting algorithm.

6. The method of claim 1, wherein identifying a set of candidate risk variables comprises:

identifying a specific physiological condition of interest associated with the adverse outcome type;

obtaining diagnostic information based upon the identified specific physiological condition;

obtaining information about clinical interventions that are likely to be appropriate interventions for the physiological condition of interest;

identifying other physiological conditions that are related to the physiological condition of interest;

populating a knowledgebase that is accessible by the computer, which relates the obtained diagnostic information to the physiological condition of interest and that relates the obtained clinical intervention information to the physiological condition of interest and that relates other identified physiological conditions to the physiological condition of interest; and

selecting the set of candidate risk variables using knowledge extracted from the populated knowledgebase.

7. The method of claim 6, wherein obtaining diagnostic information comprises:

identifying diagnostic criteria employed to diagnose the physiological condition of interest;

identifying corresponding risk metrics of the identified diagnostic criteria; and

identifying electronic sources for data fields necessary to employ the diagnostic criteria including electronic medical record data fields.

8. The method of claim 6 further comprising:

utilizing by the computer, an ontology of triple types allowed by the knowledgebase to identify queries to elicit the diagnostic information and clinical intervention information to build the knowledgebase.

9. The method of claim 1, wherein constructing, utilizing the computer, an outcome likelihood model form comprises at least one of:

selecting at least one risk variable from the set of candidate risk variables that has a reconcilable relationship with an etiology of the adverse outcome type; and

selecting at least one risk variable from the set of candidate risk variables based upon a computed statistical relationship for predicting the adverse outcome type.

10. The method of claim 1, wherein:

constructing, utilizing the computer, an outcome likelihood model form comprises;

receiving by the computer, the training data set;

constructing the outcome likelihood model form as β

1x1+ . . . +β

kxk, where there are k selected risk variables, and where β

1-β

_k represent model coefficients;

further comprising utilizing the computer for;

using the training data set to fit the model form;

determining if β

should be adjusted up or down; and

determining whether factor xi should be dropped, such that the model itself determines which parameters are important to allow prediction of adverse health outcomes.

11. The method of claim 1, wherein:

constructing a dynamic outcome likelihood model form comprises at least one of;

defining an outcome likelihood model component corresponding to each selected risk variable in the dynamic group, and aggregating across comparisons of the component magnitude when calculated with the actual dynamic risk variable value to the component magnitude when calculated with an estimated value of the selected risk variable in the dynamic group that is produced by a corresponding one of the dynamic risk variable model forms; and

basing the dynamic outcome likelihood model form on a set of scores of selected risk variables in the dynamic group, one for each risk variable, where a score is calculated as the magnitude of the corresponding outcome likelihood model component calculated with a known value of a risk variable in the dynamic group minus the magnitude of the same outcome likelihood model component calculated with estimated values of the predictor variable produced by a dynamic risk variable model form.

12. The method of claim 1, wherein:

fitting the constructed outcome likelihood model form, baseline outcome likelihood model form, and dynamic outcome likelihood model form, to a training data set to produce a dynamic outcome likelihood model, comprises at least one of;

computing the dynamic outcome likelihood model form as the outcome likelihood model form minus the baseline outcome likelihood model form;

computing the dynamic outcome likelihood model form as the sum over a set of dynamic risk variable scores; and

computing the dynamic outcome likelihood model form as the sum over a set of dynamic risk variable scores which are greater than zero.

13. The method of claim 1, wherein:

identifying a set of candidate risk variables comprises;

defining the adverse outcome type as a collective adverse outcome that is comprised of at least a first adverse outcome type and a second adverse outcome type; and

identifying candidate risk variables that are determined to be of interest to at least one of the first adverse outcome type and the second adverse outcome type;

constructing a model form for an outcome likelihood model further comprises;

selecting risk variables from the set of candidate risk variables for predicting the first adverse outcome type; and

selecting risk variables from the set of candidate risk variables for predicting the second adverse outcome type;

generating an alerting algorithm that compares a computed value for at least one of the baseline outcome likelihood and dynamic outcome likelihood of the collective adverse outcome type against predetermined thresholds, where the alerting algorithm thresholds are based upon a balancing of true positive and false positive behavior; and

generating an attribution assessment that identifies important risk variables for the collective adverse outcome type by;

evaluating the contribution of the selected risk variables with respect to at least one of the baseline outcome likelihood and the dynamic outcome likelihood, which are associated with an occurrence of the first adverse outcome type, the second adverse outcome type or both by producing a vector of indices characterizing the strength of the contributions of individual selected risk variables to at least one of the baseline outcome likelihood and dynamic outcome likelihood; and

identifying at least one evaluated risk variable that is determined to be a key contributor to at least one of the baseline outcome likelihood and dynamic overall outcome likelihood.

14. The method of claim 13 further comprising:

generating an attribution assessment that identifies important individual adverse outcome types for the collective adverse outcome type by;

evaluating the contribution of the individual adverse outcome types with respect to at least one of the baseline outcome likelihood and the dynamic outcome likelihood associated with an occurrence of the first adverse outcome type, the second adverse outcome type or both by producing a vector of indices characterizing the strength of the contributions of individual adverse outcome types to at least one of the baseline outcome likelihood and dynamic outcome likelihood of the collective adverse outcome type; and

identifying at least one individual adverse outcome type that is a key contributor to at least one of the baseline outcome likelihood and dynamic outcome likelihood for the collective adverse outcome type.

15. A method implemented on a computer for performing clinical likelihood computations to evaluate patient risk, comprising:

collecting electronic patient data about an actual patient to be monitored for an adverse outcome type;

matching, by the computer, the electronic patient data to a set of risk variables for predicting the adverse outcome type where the selected risk variables include at least one variable classified in a baseline group and at least one variable classified in a dynamic group, the baseline group is composed of non-modifiable variables, and the dynamic group is composed of modifiable variables;

utilizing a scoring algorithm associated with an outcome likelihood model to estimate a baseline outcome likelihood and a dynamic outcome likelihood based upon the electronic patient data matched to the set of risk variables;

identifying whether the patient is at-risk based upon the computed baseline outcome likelihood and a dynamic outcome likelihood; and

providing an alert with attribution if at least one of the likelihoods exceeds a predetermined threshold(s).

16. The method of claim 15 further comprising:

utilizing the computer to identify a select variable for which data is missing; and

imputing a value to the missing data based upon other risk variables with known values, where the other risk variables correlate with the missing risk variable.

17. The method of claim 15 further comprising:

performing retrospective risk attribution for a population of patients comprising;

performing computations to produce risk attribution vectors that apply to groups of patients rather than single patients;

utilizing the attribution vectors to draw conclusions across the population as to the likely root cause(s) that lead to eventual adverse outcomes in the patient data; and

utilizing the identified root causes to facilitate clinical policy decisions designed to reduce the incidence rate for adverse patient outcomes.

18. The method of claim 15 further comprising:

outputting, by the computer, risk mitigation information in the form of candidate interventions by linking identified causal factors to underlying conditions and to the output candidate interventions; and

linking identified causal factors to underlying conditions and then to the output candidate interventions based on clinical knowledge extracted from a populated clinical knowledgebase.

19. The method of claim 15 further comprising:

performing what-if risk forecasting by;

simulating, utilizing the computer, at least one intervention;

updating the estimated baseline outcome likelihood and a dynamic outcome likelihood based upon temporary changes to risk variable values required to simulate the intervention; and

identifying whether the patient risk changes based upon the updates to the estimated dynamic outcome likelihood.

20. The method of claim 15 further comprising:

associating uncertainty to forecasted risk based upon the estimated baseline outcome likelihood and a dynamic outcome likelihood;

creating an assessment of missing information;

evaluating the impact of the missing data;

computing, using the computer, an uncertainty reduction that will occur if certain missing pieces of information are obtained; and

recommending, by the computer, prioritized new clinical data types that will likely lead to the greatest uncertainty reduction.

21. The method of claim 15 further comprising:

collecting electronic patient data about an actual patient to be monitored for an adverse outcome;

determining risk variables from the collected electronic patient data, where the risk variables were selected to be predictive of the adverse outcome of interest;

extracting electronic patient data for historical patients having a profile that matches the identified patient profile, and

presenting the identified information from at least one patient history corresponding to a historical patient that matches the identified patient profile.

22. The method of claim 21 further comprising:

extracting electronic patient data from a historical patient archive with full temporal histories on previous patients for which outcomes are known; and

identifying at least one previous type of intervention for a similar patient, and presenting the outcomes of each type of intervention.

23. The method of claim 15 further comprising:

collecting electronic patient data about actual patients monitored for an adverse outcome;

determining risk variables from the collected electronic patient data, where the risk variables are selected to be predictive of the adverse outcome of interest;

identifying patient profiles of similar patients into a historical patient archive; and

evaluating recommended interventions based on the actual histories of the identified patient profiles of similar patients.

24. The method of claim 23 further comprising:

comparing a group of located historical matching patients to evaluate the risk(s) and outcome(s) of those patients that followed a trajectory of a select recommended intervention, to those that received alternative interventions.