Ali Ajdari

ABOUT ME

I am a postdoctoral research fellow at the department of Radiation Oncology at Harvard Medical School and Massachusetts General Hospital. My areas of interest is applications of optimization techniques, including convex, robust, dynamic, and stochastic optimization, in healthcare systems, especially in the world of oncology. I heavily use state-of-the-art data analytics, predictive modeling and machine learning techniques in advanced data-driven optimization models. In particular, I am very interested in Personalized Medicine -- the personalization of standard treatment protocols based on patients' genetic and epigenetic characteristics.

EDUCATION

Ph.D., Industrial & Systems Engineering, University of Washington, Seattle, WA

Robust, dynamic, and convex optimization in radiation therapy (read more)

M.Sc., Industrial & Systems Engineering, Sharif University of Technology, Iran

Dynamic adaptive experimental design in simulation optimization (read more)

B.Sc., Industrial & Systems Engineering, Sharif University of Technology, Iran

Neural network model for predicting US-EU Forex exchange rate (read more)

partially-observable Markov models

in an era where data rules the game, surprisingly vast amount of this data is subject to uncertainty and ambiguity. Our inferences about the states of things are often “partial”, based on limited observations we have made about the state of the system. Therefore, efficient methods for synthesizing these fragmented partial data should be designed for improving the inference. In my research, I employ partially observable Markov decision processes (POMDP) to model the trajectory of chronic disease evolution and progression based on partial information regarding the disease states (e.g. from radiological images, genetic information, patient’s symptoms, etc.). The key is to design efficient “estimators” of the actual states given the limited (and uncertain) observations about the disease trajectory over time.

personalized medicine

personalized medicine, called by many the “holy grail of medicine”, involves using patient-specific biological characteristics (also known as biomarkers) in designing patient-specific treatment plans (e.g. prescription drugs, radiation dose, surgery margin, etc.). It analyzes high-dimensional medical data such as genome sequencing, radiobiological images, biopsy reports, and blood-based biomarkers (e.g. proteomics, miRNA, etc.) to assess and predict patients’ response to a particular treatment plan, and adapt the plan to maximize therapeutic gains. In my research, I frequently use advanced Machine Learning and Statistics techniques, combined with mathematical programming, optimization, and predictive modeling to adapt and personalize chemo-radiation treatment plans.

simulation modeling & optimization

optimization of any complex dynamic system must at some form or another involve computer simulation. The downside however is that such simulation models are expensive to make and computationally hard to optimize. Simulation Optimization is a bridge between optimization and simulation, which aims at optimizing complex models using computer simulation to evaluate various design policies/strategies. In my research, I have devised novel theoretical models for efficient simulation optimization of complex systems. In addition, I have used computer simulation to model the complex dynamics within emergency departments. With the emergence of (deep) Reinforcement Learning techniques, there is an ever-growing need for more efficient simulation models. This is one area of research I am keen on pursuing in the future.

robust optimization

robust optimization (RO) is a subfield of optimization which deals with uncertainties within the optimization model. These uncertainties can reside in model parameters, data, and/or even the modeling choice itself. The goal of RO is to find an optimal solution that is “robust” with respect to these uncertainties. RO has applications in virtually all the areas wherein optimization is used, including healthcare, manufacturing, finance, security, and more. In my research, I focus on the application of RO in radiation therapy (RT). I use traditional RO (i.e. worst-case and probabilistic RO) as well as more recent techniques such as Adaptive/Adjustable RO and Multi-source RO to arrive at optimal or near-optimal solutions that are robust with respect to various radiobiological or systematic uncertainties.

interpretable machine learning

machine learning (ML) methods are now indispensable tools in many areas of research and practice. They shape the way we talk, sleep, and interact with each other. However, they are traditionally designed as “black box”, which means both designer and user have little to no idea how the algorithm has reached its results. Recently, there has been an increasing interest in interpretable machine learning algorithms. Due to their explanatory capabilities and “transparency”, these algorithms can be adapted far more easily in sensitive decision-making, such as in medicine, fraud, and public health. In my research, I aim to develop efficient interpretable ML-based models in the context of radiation oncology. I mainly focus on the link between Decision Trees and Random Forest (RF) to make the design and output of RF models more interpretable.

dynamic programming

dynamic programming (DP) deals with the temporal aspect of decision making; that is, decision making in non-stationary systems that evolve over time. Dynamic programmer takes into account the decision horizon and aims at optimizing the system such that the expected reward (cost) of the decision making is maximized (minimized) over that horizon. I use DP in chemo-radiation treatment planning. I model the problem as a multi-stage decision-making problem in which treatment planner must choose an action (i.e. radiation or drug dose) at each stage to maximize the probability of disease control. Due to the large size of the problem, this usually involves borrowing techniques from Approximate Dynamic Programming, e.g. certainty equivalent control (CEC), open-loop feedback control (OLFC), and reinforcement learning (RL).

Bayesian networks

bayesian statistics is a subfield of statistics and probability which uses Bayes’ theorem to guide its inferences. It is based on the premise that prior knowledge about an event contains useful information about the underlying rules governing various instantiations of that events (i.e., posterior manifestations). It is particularly useful in settings wherein revealed data exhibits certain level of noise, and a true inference cannot be readily made. Bayesian Networks (BN) are a class of network-based graphical models that uses Bayes’ rule to model the inter-relationships between multiple random variables within the network. I implement various bayesian learning and predictive BN modeling techniques in RT planning to learn and model the behavior of patients’ response over the treatment course.