Name
Title | ||
|---|---|---|
Personalized mortality prediction for the critically ill using a patient similarity metric and bagging |
Abstract | ||
|---|---|---|
Conventional mortality prediction in intensive care is based on severity of illness scores created from heterogeneous patient data that usually perform well at the population level but not necessarily at the patient level. With the emergence of large-scale electronic medical records, it is now feasible to utilize only those past patients that are clinically similar to a given index patient for whom mortality prediction is needed. Identification of similar patients can be achieved via a patient similarity metric (PSM) that quantifies the extent of similarity between two patients. Extending from a previous study, the present study aimed to investigate PSM-based bagging instead of hard-thresholding to mitigate the shortcomings identified in the previous study. Based on intensive care data from 17,152 patients, a cosine-similarity PSM and three predictive models were deployed. Besides bagging, the same methods as the previous study were used to enable a valid comparison. With a bootstrap size of 1000, the results showed that bagging led to a similar predictive performance for logistic regression (mean area under the receiver operating characteristic curve [95% confidence interval]: 0.815 [0.809, 0.821]) but worse performances for death counting (0.663 [0.655, 0.672]) and decision tree (0.510 [0.501, 0.520]), in comparison with hard-thresholding. Decreasing the bootstrap size to 500 had minimal effect on predictive performance. Future research should investigate other PSM types. |
Year | DOI | Venue |
|---|---|---|
2016 | 10.1109/BHI.2016.7455902 | 2016 IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI) |
Keywords | Field | DocType |
personalized mortality prediction,patient similarity metrics,patient similarity bagging,illness score,patient data,electronic medical record,PSM-based bagging,hard-thresholding,bootstrap,logistic regression,decision tree | Severity of illness,Population,Decision tree,Receiver operating characteristic,Computer science,Artificial intelligence,Statistics,Confidence interval,Logistic regression,Intensive care,Machine learning,Bootstrapping (electronics) | Conference |
Citations | PageRank | References |
0 | 0.34 | 4 |
Authors | ||
1 | ||