Dynamic identification of the biomarkers leveraging the dynamics of the biomarker

A system for providing a temporally dynamic model parameter with improved learning with structured sparsity, comprising:

• a memory configured to generate a temporally dynamic model parameter, the memory including:
• a formulation module configured to induce random sparsity on model parameter built from a vector by minimizing a loss function based on patient measurements taken at a plurality of time points in a longitudinal study and stored on a computer readable storage medium;
• a relation module configured to induce temporal sparsity on the model parameter by applying a first norm including a fused least absolute shrinkage and selection operator (LASSO) with a fused term encoded in a matrix identifying temporally related values of longitudinal study data having a same type of patient measurement taken at different time points, the temporally related values being identified according to similarity at neighboring points in time for each type of patient measurement, the first norm being weighted by a first coefficient to provide the temporal sparsity;
• a selection module configured to induce grouped sparsity on the model parameter by applying a second norm including an l1,2-norm of the vector, and selecting at least one value relevant to a disease of the longitudinal study data and the temporally related values of the at least one value to provide a temporally dynamic model parameter, the second norm being weighted with a second coefficient to provide feature selection;
• and
• a solving module including a processor configured to predict, with the temporally dynamic model parameter, a future condition of a patient according to data corresponding to a current condition of the patient, wherein
• the formulation module is further configured to formulate a l1-norm regularization problem by combining the minimized loss function with applying a third norm to the vector, the third norm being weighted by a third coefficient to provide random sparsity, wherein the loss function, together with the applying the third norm, formulates a l1-norm regularization problem.

A method for predicting disease progression with a model parameter using feature selection with improved learning with structured sparsity, the method comprising:

• measuring values for a selected set of biomarkers at multiple time points for each of a plurality of study participants;
• inducing random sparsity on the model parameter by minimizing a loss function based on the selected set of biomarkers;
• identifying at each point in time, from the selected set of biomarkers, temporally smooth values similar to values of corresponding biomarkers at neighboring points in time using a first norm including a fused least absolute shrinkage and selection operator (LASSO) with a fused term encoded in a matrix, the first norm being weighted by a first coefficient to provide the temporal sparsity;
• identifying predictive features from the temporally smooth values that correspond to a target response of a disease by applying a second norm including a l1,2-norm of selected set of biomarkers, the second norm being weighted with a second coefficient to provide feature selection;
• correlating the predictive features to disease progression for the disease;
• predicting a future condition of a patient relative to the disease according to patient biomarkers;
• and
• displaying the future condition via a display to a user, wherein
• inducing random sparsity includes applying a third norm to the selected set of biomarkers, the third norm being weighted by a third coefficient to provide random sparsity, wherein the loss function, together with the applying the third norm, formulates a l1-norm regularization problem.