Personalized Therapy Design
Treatment protocols for most complex diseases, such as cancer, have traditionally been devised based on the results of clinical trials averaged over a cohort of patients. In fact, the current standard of care for such diseases consists of applying a common therapy to any given patient; unfortunately patient relapse frequently ensues. By shifting the paradigm of treatment to genetically tailored protocols, the concept of personalized medicine is emerging as the state-of- the-art in patient care.
Personalized treatments involve therapies whose design (choice of drug(s), dosage, duration of treatment, etc.) is individualized based on clinical biomarkers. In this context, the problem of personalizing a treatment scheme can be cast as the search for the optimal therapy that satisfies some performance criterion. By viewing a therapy as a controlled process that is characterized by the parameter vector and whose effect can be quantified in terms of performance metrics of the form , it is possible to efficiently estimate the sensitivity of the system’s performance with respect to using Infinitesimal Perturbation Analysis (IPA). This is accomplished by extracting data from a sample path (simulated or actual) of the observed system based on which an unbiased estimate of the sensitivity can be obtained. Adopting the view that (i) cell-biologic changes necessary for the development of certain diseases may be schematized as a series of discrete states, and (ii) transitions between states may be delayed or prevented by appropriate treatment, the critical observation that motivates this work is that these are precisely the characteristics of a Discrete Event System (DES). In other words, DES models are ideally suited to the view of a “disease of stages” and can subsequently lead to more elaborate Stochastic Hybrid Automata (SHA) models capturing additional details.
Case Study: Prostate Cancer
We have thus far focused on cancer and proposed a methodology applicable to stochastic models of cancer progression, which we illustrate with a case study of optimal therapy design for advanced prostate cancer. We developed a threshold-based policy for optimal therapy design that is associated with a cost metric that combines clinically relevant measures of therapy success. We used a SHA model of prostate cancer evolution and performed IPA to adaptively adjust the threshold values so as to improve therapy outcomes. A Matlab implementation of the IPA-based optimization algorithm for personalized prostate cancer therapy design is available in the following links.
Matlab Implementation