International Business Machines Corporation

Massachusetts Institute of Technology

Wangzhi Dai

Uri Kartoun

Kenney Ng

Collin M Stultz

Kristen Ann Severson

Paul D Myers

A computer system selects features of a dataset for predictive modeling. A first set of features that are relevant to outcome are selected from a dataset comprising a plurality of cases and controls. A subset of cases and controls having similar values for the first set of features is identified. The subset is analyzed to select a set of additional features relevant to outcome. A first and second predictive model are evaluated to determine that the second predictive model more accurately predicts outcome, wherein the first predictive model is based on the first set of features and the second predictive model is based on the first set of features and the additional features. The second predictive model is utilized to predict outcomes. Embodiments of the present invention further include a method and program product for selecting features of a dataset for predictive modeling in substantially the same manner described above.

Data model processing in machine learning employing feature selection using sub-population analysis

Aspects of the invention include includes identifying a respective estimated clinical risk score for each of a first group of patients and a second group of patients. An alternative probability estimate is generated using a same set of inputs used to determine each respective estimated clinical risk score. An unreliability of a patient's clinical risk score is determined based at least in part on a feature of the patient and on a difference between the alternative probability estimate and the determined respective estimated clinical risk score.

Assessing unreliability of clinical risk prediction

A computer system trains a predictive model. A plurality of subsets of features are selected from a dataset comprising a plurality of cases and controls and a plurality of features. Cases and controls are matched to select a plurality of case-control subsets for each subset of features, each case-control subset having similar values for the corresponding subset of features. For each case-control subset, a statistical significance of each feature of the plurality of features absent from the subset of features used to match the case-control subset is identified. A final subset of features is selected based on satisfying a statistical significance of each feature for the plurality of case-control subsets. A predictive model is trained using the final subset of features. Embodiments of the present invention further include a method and program product for training a predictive model in substantially the same manner described above.

Growing community of inventors

Cambridge, MA, United States of America

Wangzhi Dai