Reduction of feature space for extracting events from medical data

1. A computer-implemented method for predicting healthcare outcomes, the method comprising:

receiving, by a processor, a plurality of patient temporal traces, each of the temporal traces comprising a unique identifier, an ordered index, and a healthcare event;

conducting, by the processor, sequential pattern mining on the plurality of patient temporal traces to produce a plurality of sequential features, by:

determining, by the processor, a number of healthcare events having a minimum support, the minimum support comprising a number of transactions in the temporal traces that include the healthcare event;

generating, by the processor, a seed set based on the number of healthcare events having the minimum support;

generating, by the processor, candidate sequences based on the seed set;

determining, for each of the candidate sequences, a frequency of the candidate sequence;

selecting one or more of the candidate sequences based on the frequency of each respective candidate sequence; and

generating, by the processor, the plurality of sequential features based on the one or more selected candidate sequences;

clustering, by the processor, the sequential features with a similarity metric to create a plurality of clustered features, wherein clustering the sequential features comprises applying a Smith-Waterman Distance Matrix to the plurality of sequential features and performing nonnegative matrix factorization, wherein the sequential features comprise features of a feature space of a predictive model, wherein clustering the sequential features reduces a total number of sequential pattern features to reduce the feature space of the predictive model and create a reduced feature space of the predictive model, and wherein the plurality of clustered features comprise features of the reduced feature space of the predictive model; and

analyzing, by the processor, the clustered features to predict a healthcare outcome, wherein predicting the healthcare outcome comprises applying the predictive model to the reduced feature space of the predictive model and outputting a predicted outcome to a display, wherein the predictive model identifies a relationship between a respective clustered feature of the reduced feature space and a respective healthcare outcome, and wherein the predicted healthcare outcome represents a likelihood that a specified drug treatment will result in recovery for a condition.

2. The computer-implemented method of claim 1 , wherein the ordered index comprises a time stamp.

3. The computer-implemented method of claim 1 , wherein clustering the sequential features comprises grouping a plurality of similar longitudinal sequential patterns together.

4. The computer-implemented method of claim 3 , wherein grouping the plurality of similar longitudinal sequential patterns together comprises applying a symmetric nonnegative matrix factorization to the plurality of sequential features.

5. A computer program product for predicting healthcare outcomes, the computer program product comprising:

a non-transitory storage medium readable by a processing circuit and storing instructions for execution by the processing circuit for performing a method comprising:

receiving a plurality of patient temporal traces, each of the temporal traces comprising a unique identifier, an ordered index, and a healthcare event;

conducting sequential pattern mining on the plurality of patient temporal traces to produce a plurality of sequential features, by:

determining a number of healthcare events having a minimum support, the minimum support comprising a number of transactions in the temporal traces that include the healthcare event;

generating a seed set based on the number of healthcare events having the minimum support;

generating candidate sequences based on the seed set;

determining, for each of the candidate sequences, a frequency of the candidate sequence;

selecting one or more of the candidate sequences based on the frequency of each respective candidate sequence; and

generating the plurality of sequential features based on the one or more selected candidate sequences;

clustering the sequential features with a similarity metric to create a plurality of clustered features, wherein clustering the sequential features comprises applying a Smith-Waterman Distance Matrix to the plurality of sequential features and performing nonnegative matrix factorization, wherein the sequential features comprise features of a feature space of a predictive model, wherein clustering the sequential features reduces a total number of sequential pattern features to reduce the feature space of the predictive model and create a reduced feature space of the predictive model, and wherein the plurality of clustered features comprise features of the reduced feature space of the predictive model; and

analyzing the clustered features to predict a healthcare outcome, wherein predicting the healthcare outcome comprises applying the predictive model to the reduced feature space of the predictive model and outputting a predicted outcome to a display, wherein the predictive model identifies a relationship between a respective clustered feature of the reduced feature space and a respective healthcare outcome, and wherein the predicted healthcare outcome represents a likelihood that a specified drug treatment will result in recovery for a condition.

6. The computer program product of claim 5 , wherein the ordered index comprises a time stamp.

7. The computer program product of claim 5 , wherein clustering the sequential features comprises grouping a plurality of similar longitudinal sequential patterns together.

8. The computer program product of claim 7 , wherein grouping the plurality of similar longitudinal sequential patterns together comprises applying a symmetric nonnegative matrix factorization to the plurality of sequential features.

9. A processing system for predicting healthcare outcomes, comprising:

a processor in communication with one or more types of memory, the processor configured to:

receive a plurality of patient temporal traces, each of the temporal traces comprising a unique identifier, an ordered index, and a healthcare event;

conduct sequential pattern mining on the plurality of patient temporal traces to produce a plurality of sequential features, by:

determining a number of healthcare events having a minimum support, the minimum support comprising a number of transactions in the temporal traces that include the healthcare event;

generating a seed set based on the number of healthcare events having the minimum support;

generating candidate sequences based on the seed set;

determining, for each of the candidate sequences, a frequency of the candidate sequence;

selecting one or more of the candidate sequences based on the frequency of each respective candidate sequence; and

generating the plurality of sequential features based on the one or more candidate sequences;

cluster the sequential features with a similarity metric to create a plurality of clustered features, wherein clustering the sequential features comprises applying a Smith-Waterman Distance Matrix to the plurality of sequential features and performing nonnegative matrix factorization, wherein the sequential features comprise features of a feature space of a predictive model, wherein clustering the sequential features reduces a total number of sequential pattern features to reduce the feature space of the predictive model and create a reduced feature space of the predictive model, and wherein the plurality of clustered features comprise features of the reduced feature space of the predictive model; and

analyze the clustered features to predict a healthcare outcome, wherein predicting the healthcare outcome comprises applying the predictive model to the reduced feature space of the predictive model and outputting a predicted outcome to a display, wherein the predictive model identifies a relationship between a respective clustered feature of the reduced feature space and a respective healthcare outcome, and wherein the predicted healthcare outcome represents a likelihood that a specified drug treatment will result in recovery for a condition.

10. The processing system of claim 9 , wherein the ordered index comprises a time stamp.

11. The processing system of claim 9 , wherein clustering the sequential features comprises grouping a plurality of similar longitudinal sequential patterns together.

12. The processing system of claim 11 , wherein grouping the plurality of similar longitudinal sequential patterns together comprises applying a symmetric nonnegative matrix factorization to the plurality of sequential features.