Behind the PhD: Israel Campero on AutoML in healthcare
Posted on
Israel Campero’s PhD research looks at how we can make AutoML truly usable in healthcare settings. Although AutoML promises to make machine learning easier and more accessible, hospitals face practical challenges: limited computing resources, messy real world data, and the need for transparent, clinically meaningful results. This thesis develops methods that make AutoML more efficient, more resource aware, and better aligned with clinical needs. The work includes new optimization strategies, a smarter way to start AutoML searches using prior knowledge, and thorough testing in end to end ECG analysis pipelines. The end goal: bring AutoML closer to real, everyday use in healthcare.
Why did you choose this research subject and what makes it so fascinating?
I chose this topic because AutoML promises to make machine learning available to people who are not experts. But in healthcare, that promise is still hard to achieve. Hospitals often have limited computing power, messy data, and strict expectations about transparency. You can’t just give clinicians a black box model.
What fascinates me is that these constraints aren’t small technical details; they completely change what methods are practical in the real world. My work focuses on designing AutoML strategies that are efficient and reliable under these real life conditions, and then testing them in full clinical workflows.
Which challenges did you meet along the way, and how did you overcome them?
One big challenge was making sure my research claims also hold under realistic time and resource limits. Many AutoML methods look great when you have unlimited compute, but their behavior changes drastically when you restrict time. I addressed this by designing experiments that use short, repeatable time budgets and by applying multi fidelity techniques to quickly eliminate bad options.
Another challenge was connecting algorithmic ideas with clinical needs. For ECG analysis, good accuracy is not enough; cardiologists need outputs they can interpret, such as findings specific to certain leads or time periods. This meant working in close alignment with clinical constraints and selecting evaluation methods that translate into results clinicians can actually use.
What is the impact of your work in the real world?
My work offers practical ways to build AutoML systems that are more usable in high stakes settings like healthcare. Algorithmically, I use GEISHA, an asynchronous island model that speeds up the search for good ML pipelines while staying within strict time and compute budgets. I also develop a way to use knowledge from previous datasets, via large language models, to make AutoML start faster and avoid cold starts.
On the evaluation side, I show that using short-budget benchmarks can still give reliable rankings while lowering computational cost, which helps make experimentation more sustainable.
Clinically, I deliver and validate several ECG decision support components, such as long term signal quality analysis and an interpretable detector for ST segment deviations. These tools, developed within the INNO4HEALTH project, aim to make long term ECG monitoring more robust and more actionable in real healthcare environments.
What are your plans after your PhD?
After finishing the PhD, I plan to continue working at the intersection of optimization, machine learning, and healthcare. I enjoy combining method development with real deployment challenges. That’s why I started a Postdoctoral Researcher position focused on individualized sleep apnea diagnosis using AI.
