Embedded Files
Performance Measures for Maintenance of Roadside Features
Performance Measures for Maintenance of Roadside Features
Developing performance measures to evaluate and compare safety risks of maintenance operations on roadside features. The ultimate goal is to develop a risk and difficulty indices to be used in prioritizing and scheduling of maintenance operations.
Adapted multiple statistical methods to identify major factors contributing to work zone injuries of the maintenance crew working alongside highways.
Developed a data pipeline for Caltrans to manage and connect Work Order, Lane Closure, Traffic Volumes, Road Features and Collision Reports.
Developed a classification pipeline for highly imbalanced data accurately predicting the probability a work zone collision.
Prototyped a desktop application to visualize historical trends, estimate levels of difficulty and injury, and visualize work orders range, traffic, and collision data on Google Maps
The most frequent, costly, and time consuming activities
Sample Figure 1. demonstrates an example of the projects' analyses and visualizations in order to develop an index of difficulty.
In this figure, the colored bubbles show maintenance activities that are among the top 25% in duration, cost, and frequency.
Sample Figure 2. demonstrates an example of analysis on injury risk factors.
In this figure, the effect of daily traffic volume on the severity of work zone collisions is demonstrated. Width of the area for each level of severity correspond to the proportion of collisions at different levels of traffic.
Distribution of collision severity by average daily traffic
Feature importance
Sample Figure 3. demonstrates an example of classification analysis
In this figure, an extreme gradient boosting algorithm identifies the most important features to the risk of collision at roadside work zones.
Decision app prototype
Application Prototype (In progress)
The user inputs the activity and its location and will see major factors affecting the risk level visualize as a result such as traffic volumes and historical collision densities.
Real-Time Ambulance Dispatching and Relocation
Real-Time Ambulance Dispatching and Relocation
Proposing a new dispatching and relocation algorithm for ambulance assignments that achieves 18% reduction in average response times of emergency calls.
Formulated a new dynamic programming for real-time design of emergency vehicle management.
Developed a stochastic process that predicts future state of the system and reacts to both current and future demand.
Implemented an adaptive regression method that approximates the solution of a dynamic design in real-time.
Simulated the new design based on four years of historical data from the Mecklenburg County in North Carolina.
An overview of the ambulance dispatching dynamics
Sample Figure 1. shows a simplified overview of the problem.
When an emergency call arrives, the decision maker must decide whether to send an ambulance or delay the call. If an ambulance finishes serving the call, then the decision maker must choose which base the ambulance is redeployed to.
Sample Figure 2. shows the distribution of response times for different dispatching policies.
ADP stands for the proposed policy which adaptively assign ambulances, reallocates them to cover for the already dispatched ones, and redeploys them optimally after their service is finished.
The black line represent the current practice of always sending the closest available ambulance and redeploying them back to their original base.
Distribution of ambulance response time
Benchmark 1 represents a policy that dispatches dynamically but redeploys according to the current practice. This figure shows that the full dynamic policy shift the distribution of response time shifts to the left which means that the average response times of all types of calls decreases under the proposed policy.
Optimal Stopping of Dose-Finding Clinical Trials
Optimal Stopping of Dose-Finding Clinical Trials
Proposing a robust stopping rule for sequential sampling processes in clinical trials that accurately identifies when enough information is gathered in order to stop a clinical trial.
Developed a new adaptive technique for clinical trials with respect to Net Present Value of the drug under study.
Proposed a diffusion approximation solution to the adaptive problem independent of the shape of the underlying process.
Adapted a reinforcement learning algorithm evaluated by diffusion approximation to estimate stopping times.
The plots above show the solution to the diffusion approximation of the online stopping formulation. The colored lines represent different simulations of two clinical trials where each line shows the unit advantage of the treatment over placebo. The left figure shows simulations of a treatment which produces advantageous responses and the right figures shows simulations of a treatment with no significant advantage. The three areas after the dashed vertical line represent the solution of the partial differential equation of the diffusion approximation. If the simulation paths end up in the lower area, it means that the trial should be abandoned because no evidence of advantage has been found. The middle area translates to continuation which means that some evidence is found but its significance is not established. The trial which end up within the upper area show significant positive evidence of advantage over placebo.
In practical implementation, these areas can be evaluated before the start of the trial because they are independent of the shape of the dose-response curve and the dose allocation procedure. When the trial start, the response of the participants who undergo the treatment can be mapped into diffused paths and the decision maker can terminate, abandon, or continue the trial based on the area that the path ends up in.
Cost-Effectiveness Analysis of Pre-diabetes Lifestyle Intervention
Cost-Effectiveness Analysis of Pre-diabetes Lifestyle Intervention
Demonstrated that lifestyle intervention in pre-diabetes cases reduce the probability of progressing to diabetes while being a cost-effective solution in the long run
Coupled population-based simulation and an agent-based simulation to simulate the population of the U.S. and the life progression of an individual.
Validated the simulation mode by the data from CDC.
Adapted different regression and Cox hazard models to simulate the progression of an individual.
Overview of the simulation model
Sample Figure 1. shows the components of the simulation model. The population-based simulation model had a birth module to simulate population birth rates. The microsimulation models simulates the progression of an individual to his/her death. The intervention sub-module simulated the effects of three lifestyle interventions.
Optimal Learning of Dose-Response Curves in Dose-Finding Clinical Trials
Optimal Learning of Dose-Response Curves in Dose-Finding Clinical Trials
Proposed a new and efficient machine learning technique for learning dose-response relationship in clinical trials.
Adapted a time-series general linear model to approximate the unknown function and allowing for unknown correlation structure and unknown measuring variance.
Developed a reinforcement learning algorithm to sequentially update the constructed estimation.
Reduced the complexity of the design and significantly improved the computational effort required to derive new allocation decisions.
Overview of the dose-finding clinical trial
After a dose is assigned to a participant in the trial, an observation of the response becomes available. This observation may be used to modify the characteristics of the trial while the trial is still ongoing.
In this study, only the dose assignment was subject to change with respect to interim observations of the participants responses.
Google Sites
Report abuse