Research Advancement: Learning stochastic finite-state transducer to predict
Project Title: Learning stochastic finite-state transducer to predict individual patient outcomes
Institution and State: University of Puerto Rico Medical Sciences Campus, Puerto Rico
Investigators: Patricia Odoñez, Ph.D., UPR Rio Piedras Campus and Abiel Roche-Lima, Ph.D., UPR Medical Sciences Campus
Background: Current methods for measuring the well-being of a patient in the intensive care unit (ICU) acquire patient’s vital signs data at rates that are difficult for a human to analyze (60–500 Hz). These measurements are displayed on a monitor for a few seconds as a collection of univariate time series and then lost to further analysis. Instead, a lower-frequency version of this data is stored in an electronic health record after validation by a medical provider at the rate of once every 15 min to once every several hours. Physicians make life-saving decision based on this lower-frequency data. Recently, there has been interest in storing and analyzing the high-frequency data using automated and semi-automated methods.
Advance: This research introduces a new method to predict patient outcomes in Intensive Care Units which used probabilities of edit distance costs learned by stochastic finite-state transducer models. Time series data were converted to sequence representation to be used as a model input. Several experiments were developed by changing the parameters during the conversion process. Good results were obtained based on the computed prediction metrics. When compared with previous works, the proposed method improved Accuracy, Precision and F-measure metric values. In future work, other implementations of the algorithm in parallel will be used to increase the sequence length and improve efficiency. The final goal is to create a multivariate representation of the algorithm. Additionally, other approaches using finite-state transducers can be used to improve prediction. For example, rational kernels (kernel based on finite-state transducers) can be combined with kernel methods, such as Support Vector Machine.
How NIMHD Grant Enabled Advance: NIMHD RCMI Grant MD007600 provided technical and administrative support, as well as computational facilities for conducting bioinformatics studies. Important to mention that Dr. Abiel Roche-Lima was recruited through the RCMI Program to develop bioinformatics at the UPR Medical Sciences Campus.
Public Health Impact Statement: The proposed model can be included in an intelligent mobile decision-aid tool for practitioners in Intensive Care Units, as affordable and reliable decision support system for predicting patients’ outcomes. For example, based on these models a mobile tool can be able to set off alarms in physician’s cell phones when a patient is positively classified to have an acute episode of hypotension within an hour.
Support: Research supported in part by Natural Sciences and Engineering Research Council of Canada (NSERC) at University of Manitoba, Canada, NIMHD- G12 MD007600 from the National Institutes of Health at University of Puerto Rico Medical Sciences Campus, and R25-MD010399 from the National Institutes of Health at the University of Puerto Rico Río Piedras Campus.
Citation: Ordoñez P, Schwarz N, Figueroa-Jiménez A, Garcia-Lebron LA and Roche-Lima A. Learning stochastic finite-state transducer to predict individual patient outcomes. Health Technol (2016) doi:10.1007/s12553-016-0146-2, http://link.springer.com/article/10.1007/s12553-016-0146-2