DARPA project N66001-11-1-4183

Predictive Modeling of Patient State and Therapy Optimization

Principal Investigator
Obradovic Zoran

Abstract

This project will develop and validate effective predictive modeling technology to achieve the following sepsis treatment related aims on high dimensional and noisy data at a clinically relevant scale:

• AIM 1: Personalized sepsis therapy optimization for an individual patient’s state improvement
• AIM 2: Early diagnosis of sepsis and accurate detection of change in the state of sepsis, and
• AIM 3: Gene expression analysis for sepsis biomarkers identification.

These aims will be achieved by exploiting advanced nonlinear methods for analysis of temporal dependencies in high dimensional multi-source data. Our approach will include the development and application of effective methods for:

• structured regression-based predictive control for personalized optimization of therapeutic strategies (AIM 1)
• early classification on interdependent multivariate time series (AIM 2),
• mining of integrated genomic and clinical temporal data (AIM 3).

For optimization of therapy, a highly expressive nonlinear predictive control algorithm will be developed that models a joint distribution over model outputs and handles high dimensional data of different quality with correlated variables and with missing observations. To obtain a diagnosis of sepsis as early as possible, early classification methods will be developed based on efficient selection of the most informative attributes for learning in high dimensional temporal data when information content of most individual attributes is low. Early classification methods will also be adapted to monitor a sepsis patient and recognize a transition to a different stage of sepsis that would require a change in therapy (e.g. from sepsis to serious sepsis or to septic shock). Methods for uncertainty analysis will help to restrict the application of classifiers to situations consistent with their training experience. Finally, high dimensional genomic data will be analyzed by a modification of an L1 regularized logistic regression that learns from a small number of cases when most of variables are irrelevant.