April 16, 2014

June 7, 2013
Guelph Biomathematics and Biostatistics Symposium

Adaptive Strategies in Epidemiology, Ecology, and Engineering



Adaptive Representations for Parameter Optimization
- Daniel Ashlock

Selecting parameters that permit a model to best fit a set of data is a very old problem. For many models there are rapid, optimal regression techniques. Consider a man searching under the streetlight for a key he lost in the park because the light is better - even though the key was not lost near the streetlight. Our choice of models is influenced by the availability of good regression algorithms, even when they lead to less-than-appropriate models. Evolutionary computation is an example of a relatively low-speed but extremely general model fitting technique. In this talk non-adaptive and adaptive techniques for fitting model parameters with evolutionary computation are defined, explained, and contrasted. This talk touches on the issue of representation in evolutionary computation; that is, choosing good ways to represent problems for computer solution.

Evaluating Epoetin dosing strategies using observational longitudinal data
- Cecilia Cotton

Epoetin is commonly used to treat anemia in Chronic Kidney Disease and End Stage Renal Disease subjects undergoing dialysis, however, there is considerable uncertainty as what level of hemoglobin or hematocrit should be targeted in these subjects. In order to address this problem we treat epoetin dosing strategies as a type of dynamic treatment regimen. We present a methodology for comparing the causal effects of multiple treatment regimens on survival based on observational data. This problem is complicated by the fact that depending on the regimen definitions, subjects may have been adherent to multiple regimens at the same time. We present a methodology in which each subject is cloned (or replicated) and contributes follow- up data to each regimen to which they were continuously adherent before being artificially censored. We provide an inverse probability weighted log-rank test with variance estimate that can be used to compare survival under two regimens. For comparing multiple regimens we propose several marginal structural Cox proportional hazards models with robust variance estimation. The methods are illustrated through simulations and applied in an analysis comparing epoetin dosing regimens in a cohort of 33,873 adult hemodialysis patients from the United States Renal Data System.

Swarm Intelligence for Unmanned Aerial Vehicles
- David Howden

Wildfires are destructive conflagrations that occur in areas of wilderness and their remote location serves as a barrier to rapid detection or response. Due to their inaccessibility, these fires can grow to insuppressible proportions and not only cause significant economic damage to an area, but also endanger the lives of communities and their fire fighters. Fast and effective detection is a key factor in bushfire fighting. Accurate knowledge of the status of a bushfire is indispensable for enabling accurate fire prediction modelling, maintaining the safety of fire fighting crews, and allowing efforts to be focused on areas of the highest risk such as urban areas with strong human presence. This problem is ideal for the application of multiple unmanned aerial vehicles. In this presentation, a swarm intelligence approach to exhaustive and continuous surveillance of large areas is introduced. Using a pheromone inspired technique, landmarks and features can be assigned priorities relative to their value or risk, either on deployment or dynamically in reaction to observations. The presented approach is fully distributed, resilient against loss, and not reliant on cooperative decision making or long range communication.

Bio-inspired complex systems engineering
- Taras Kowaliw

The use of complex systems in engineering design promises to open new areas of productivity. We start with a brief description of some of the strengths and weaknesses of using evolutionary computation for design. We next discuss artificial development, the use of a growth stage in the optimization process. Rather than directly specify a design, we aim to grow it. Artificial development is a potential means of incorporating several biologically-motivated design metaphors, ones that might generate designs of greater complexity, robustness, and resilience. Two concrete examples are presented: firstly, the design of forms in the domain of structural engineering, where we recover a form of artificial polymorphism; secondly, the generation of soft-bodied virtual robots. Finally, we discuss potential applications to the automation of synthetic biology.

A Clinical Trial Design for Constructing and Evaluating Individualized Real-Time Treatment Policies
- Susan Murphy

Mobile devices, including mobile phones, are increasingly used to both passively and actively collect patient symptoms, where the patient is, who the patient is with, level of social activity. At the same time, mobile devices are beginning to be used to deliver a variety of real-time behavioral interventions (motivational assistance, cognitive assistance, suggestions concerning social interactions). However only a few researchers have begun to use the real-time patient information to adapt and re-adapt the behavioral interventions to the patient. And for the most part, this adaptation is primarily based on behavioral theory, clinical experience and expert opinion. Data-based evidence is, at best, indirectly used in this process. In this talk we sketch out the outline for, and solicit feedback on, a new clinical trial design for the purpose of providing/using patient data to inform the development of Individualized Real-Time Treatment Policies.

Generating candidate adaptive dosing strategies through simulation and g-estimation
- Ben Rich

Warfarin, a commonly prescribed oral anticoagulant, is a drug which requires adaptive individualized dosing. While much is known about the properties of the drug, an optimal dynamic dosing strategy has remained elusive. We propose the use of a realistic pharmacokinetic-pharmacodynamic model to generate simulated patient data, to which g-estimation is applied to generate a candidate dynamic strategy. Using the same data-generating model, different strategies can be compared on future subjects drawn from the same population or a different one. Our findings suggest that despite partial model misspecification this methodology can lead to a dosing strategy which performs well both within and across populations with varying pharmacokinetic and pharmacodynamic characteristics.


Statistical methods for comparing adaptive treatment strategies in SMART designs with time-to-event endpoints
- Abdus S. Wahed

Adaptive treatment strategies (ATSs; aka dynamic treatment regimes) are individually tailored treatments, with treatment type and/or dosage changing according to patient’s intermediate response at different stages of therapy. Availability of multiple treatment options at each stage and possibility of variable intermediate responses to these treatments may result in many different adaptive treatment strategies. Sequential multiple assignment randomization trials (SMARTs) are often used to study ATSs in the treatment of cancer, leukemia, depression, and AIDS. Frequently the goal is to compare multiple strategies based on time-to-event outcome such as overall survival. In this talk we will discuss some semi-parametric approaches to compare two-stage ATSs based on time-to-event data collected from SMART designs. Specifically, traditional methods of estimation such as Empirical CDF, Kaplan-Meier, and Nelso-Aalen, and test of hypothesis such as Log-Rank test will be adapted using inverse-probability-weighted methods to be applicable to this specific trial design. Application of these methods will be demonstrated by applying to a cancer clinical trial dataset.


Back to top