Modeling time-series data and temporal features

The key challenge for analysis of clinical data is that EHRs consist of complex multivariate time series of clinical variables collected for a specific patient, such as laboratory test results, medication orders, physiological parameters, past patient's diagnoses, surgical interventions and their outcomes. A fundamental challenge the researchers and data analysts face is how to summarize and represent this complex time-series data in order to make them amenable to statistical analysis and modeling. Our work has focused primarily on predictive modeling problems and solutions where our goal is to identify or construct temporal features important for prediction of adverse events, patient outcomes, and future patient management decisions. Our work on temporal features covers the following directions:

Fixed feature template maps

The approach relies on a fixed set of feature template mappings defined by domain experts that allows the conversion of complex time series of labs, medications or physiological values to different temporal features. For example, the time series of values for a numerical lab (such as platelet count) were converted to 28 different features reflecting the last lab value observed, the time elapsed since the last value, last trend, apex and horizon values, and their differences from last value, etc. We have developed similar feature sets for medications orders (e.g. aspirin or heparin), physiological parameters, input/output volumes, etc. We have successfully used these feature sets in modeling the patient state for adverse event detection, and in the outlier based alerting projects.

Selected publications:

Features based on temporal predictive patterns

The approach constructs temporal features by extracting predictive patterns characterizing patient subgroups in EHRs that are important for prediction. Such patterns, when explicitly represented, can be used as features one needs to include when defining classification models. Our work focuses on temporal patterns based on temporal abstraction and temporal logic to get a high-level qualitative description of clinical time series. This representation is more flexible and allows us to express various temporal patterns observed in clinical time-series data.

Selected publications:

State-space models of dynamics

The approach we are currently exploring attempts to build models of multivariate time series data and their behaviors by exploring lower-dimensional representations of the patient state with the help of Markov process models. Briefly, our goal is to find a lower dimensional patient state representation that summarizes and compactly encodes all information about past patient's observations that is needed to predict well the behavior of the time series in the future. Such a state representation can yield a compact feature space for various prediction problems in clinical time series and would be an alternative to features based on fixed feature mappings or predictive patterns discussed earlier. We have studied and developed models based on Linear dynamical system and Gaussian processes frameworks, and their hierarchical combination to model time-series of different laboratory tests. Our most recent work focuses on the development of a sparse LDS framework that aims to recover lower dimensional representation of the hidden LDS state.

Selected publications:


The web page is updated by milos.