Identifying and ranking risk factors using trained predictive models

Patent metadata

Field	Value
Publication number	US-11355246-B2
Application number	US-201615172512-A
Country	US
Kind code	B2
Filing date	Jun 3, 2016
Priority date	May 3, 2016
Publication date	Jun 7, 2022
Grant date	Jun 7, 2022

Abstract

Embodiments are directed to methodologies, systems and computer program products for generating, for each of a plurality of risk factors in a patient database containing information of a plurality of patients, an index of input values for the risk factor. For each patient P of the plurality of patients, a series of local impact scores is computed for the patient. Computing the series of local impact scores for the patient includes calculating a risk score for the patient with respect to each of the indexed input values for each of the plurality of risk factors. For at least one of the plurality of patients, at least some of the plurality of risk factors are ranked based at least partly on the computed local impact scores for each of the at least some risk factors, and an indication of the ranked risk factors for the at least one patient is provided.

First claim

What is claimed is:

1. A system comprising one or more processors, a system memory, and a bus that couples various system components including the system memory to the one or more processors, the system configured to identify impactful patient-specific risk factors by performing a method that includes:
for each of a plurality of risk factors in a patient database containing information of a plurality of patients, generating, via neural network, an index of input values for the risk factor, wherein at least one risk factor of the plurality of risk factors includes LDL-cholesterol, and wherein a range of the index of input values spans from a lowest observed value to a highest observed value in the patient database;
for each patient P of the plurality of patients, computing, by a trained predictive model of the neural network, a series of local impact scores for the patient, wherein computing the series of local impact scores for the patient includes calculating a risk score for the patient with respect to each of the indexed input values for each of the plurality of risk factors, wherein computing the series of local impact scores for each patient P includes, for each risk factor R in the plurality of risk factors:
creating, for each input value V in the generated index for the risk factor R, a new patient model NP; and
determining, using the trained predictive model and based on each of the created new patient models, a local impact score S for the risk factor R; and
for at least one of the plurality of patients, ranking, by the trained predictive model of the neural network, at least some of the plurality of risk factors based at least in part on the computed local impact scores for each of the at least some risk factors, and providing an indication of the ranked at least some risk factors for the at least one patient, wherein computing the local impact scores comprises:
examining a range of possible predicted scores for each risk factor of the plurality of risk factors for each patient P, wherein each range begins with each patient P's actual risk score, and
weighting each patient P's predictive risk score based on a proximity to each patient P's respective actual risk score
outputting the series of local impact scores.

2. The system of claim 1 , wherein generating the index of input values for a risk factor includes determining all values for the risk factor that are associated with the plurality of patients.

3. The system of claim 1 , wherein computing the series of local impact scores for each patient P includes, for each risk factor R in the plurality of risk factors:
creating, for each input value V in the generated index for the risk factor R, a new patient model NP; and
determining, using the trained predictive model and based on each of the created new patient models, a local impact score S for the risk factor R.

4. The system of claim 1 , wherein computing a series of local impact scores for the patient includes computing the series of local impact scores using the trained predictive model that is based at least in part on one or more of a plurality of statistical classification algorithms that include at least one of logistic regression, decision trees, random forests, support vector machines, neural networks, and Bayesian networks.

5. The system of claim 1 , wherein ranking the at least some risk factors based at least in part on the computed local impact score S for each of the at least some risk factors includes determining, for each of the at least some risk factors, a local importance score by calculating a smallest change in value for the risk factor needed to result in a largest change to a predicted risk for the at least one patient.

Assignees

• IBM

Inventors

• Krause Josua
• Ng Kenney
• Perer Adam

Classifications

• G06N20/00

  Machine learning · CPC title

• G16H50/30 Primary

  for calculating health indices; for individual health risk assessment · CPC title

• G16H10/60

  for patient-specific data, e.g. for electronic patient records · CPC title

• G16H50/20

  for computer-aided diagnosis, e.g. based on medical expert systems · CPC title