The Translation Gap

The promise of medical technologies and their failure to migrate to the medical field.

Medical technologies promise to alleviate the burden on the clinician, speed up workflows and deliver better clinical outcomes. However, these systems are not widely accepted and seamlessly deployed in clinical practice. Systems unexpectedly fail in real-world settings, complicate processes and leave end users frustrated.

In my research, I investigate the nature of this "translation gap", understand the context of use, and design interventions that would help these technologies deliver what they promise. My goal is to re-align technological advancements with the needs of end users, taking a step towards closing this gap.


Closing the Gap: Designing better Technology

Designing tools and methods to support better interactions

Current technologies are limited in their scope. They do not allow users to interrogate them, probe them or analyze them in different scenarios. Moreover, they do not give a truthful presentation of their failure modes.

This creates a need to rigorously evaluate these systems before they are deployed, in scenarios that are reflective of the real-world. I design methods that add a layer of interrogation, such that users can be informed of the limitations of these technologies.


Closing the Gap: Human-centered Approach

Involving stakeholders and understanding their needs and abilities

Designers of technology are very often different from the users. For technologies to be useful, the first step is understand what the users want. A Human-centered design approach places the user at the forefront, and any interventions are aimed at mitigating the problems the user is facing.

However, technologies are often developed without this perspective, leading to greater harm than good. In my research, I seek to understand what effects do such technologies have on the users, whether it empowers them or leads to more harm.


Research Projects

Holistic Evaluation of Tumor Segmentation Algorithm

Summary statistics are not sufficient to understand the extent of algorithmic failure. We explored how algorithms can be better evaluated, considering robustness, data quality, confidence and diagnostic performance.

Uncertainty Quantification with Bayesian Neural Networks

Communicating uncertainty allows algorithms to be transparent about decisions. However, computing uncertainty is very costly. We designed a method that computes uncertainty at a very low computational cost.

Understanding decision-makers' perception of uncertainty.

Humans make decisions without explicitly thinking of uncertainty in their daily lives. Do things change when such uncertainty is made explicit? How? We're conducting a study to investigate the effects of uncertainty communication.

Research Group Affiliations


Computer Science and Engineering,
University of Michigan

Uncertainty Quantification & Scientific Machine Learning Group

Mechanical Engineering,
University of Michigan

Systems Imaging &
Bioinformatics Lab

Computational Medicine and Bioinformatics,
Michigan Medicine