About Me

  • Full Name:Israel Campero Jurado
  • Phone:+31 620994782
  • Email:israelcamperojurado@gmail.com

Hello There!

I am a researcher based in the Netherlands. I hold postdoctoral experience focusing on time-series modelling for agentic systems in the context of sleep apnea. I have worked with the University of Groningen (Top 100), ETH Zurich (Top 10), and University Hospital Zurich. I completed my PhD with the AMOR/e Group in the Department of Computer Science at Eindhoven University of Technology (TU/e), and contributed to the European ITEA project INNO4HEALTH. My interests include automated machine learning (AutoML), time series, foundation models, LLMs, and healthcare applications, with a focus on solving real-world problems through challenging benchmarks.

My Resume

  • Work Experience

  • AI Developer for Healthcare Applications / Freelance

    MicroHealth B.V., The Netherlands - June 2026 – Present

    Developing machine learning and LLM-based models for digital healthcare applications, focusing on the prediction and ranking of gut microbiome composition from lifestyle-related variables such as food intake, physical activity, sleep, and other behavioral factors. Working on data-driven modeling approaches to support personalized health insights and decision-making in microbiome-based healthcare.

  • Postdoctoral

    University of Groningen, The Netherlands - November 2025 – Present

    I Developed Data-driven control solutions for medical applications. Investigate the use of machine learning-based algorithms and foundation models for time series analysis in OSA therapy with unobtrusive sensors. I developed strategies to personalise data-driven interventions in close collaboration with clinical partners.

  • PhD candidate

    Eindhoven University of Technology, The Netherlands - May 2021 – November 2025

    Conducting research in Automated Machine Learning (AutoML) with applications in healthcare as part of the INNO4HEALTH project.

  • Visiting Student

    University of Cambridge, United Kingdom - May 2025 – June 2025

    I collaborated on expanding a Machine Learning Benchmark framework. I analyzed diverse datasets, particularly in healthcare and socially impactful domains, to enhance AutoML/ML tool performance. My work addressed challenges posed by complex datasets, ensuring robust and effective real-world applications.

  • Network Scientist

    École Normale Supérieure de Lyon, France - Oct 2020 – March 2021

    Research stay at the ENS in Lyon, where I worked with complex networks and graph signal processing fundamentals.

  • Artificial Intelligence Engineer

    Mirai Innovation Research Institute, Japan - Mar 2020 – May 2020

    Research stay where I developed a project for hand rehabilitation with machine learning and virtual reality: https://ieeexplore.ieee.org/document/9283413 With this project I was winner of the "Premio Hidalgo de Ciencia, Tecnología e Innovación 2020", which is the highest award in science for any state in Mexico.

  • Artificial Intelligence Programmer

    Universidad de Salamanca, Spain - May 2019 – Jul 2019

    Implemented machine learning algorithms for academic research projects.


  • Education

  • PhD Candidate in Computer Science

    Eindhoven University of Technology, The Netherlands - May 2021 – Present

    Thesis: "Automated Machine Learning, Multi-fidelity optimization and benchmarking"

  • Master's degree, Information Technology

    Universidad Politécnica de Pachuca, Mexico - Sep 2018 - Aug 2020 | Grade: 9.9 (10 max)

    Thesis: "Predictive model of hypertension in pregnant women"

  • Bachelor's degree, Telecommunications Engineering

    Universidad Politécnica de Pachuca, Mexico - Sep 2014 - Feb 2018 | Grade: 9.5 (10 max)

    Thesis: "Embedded system for gas consumption calculation"

Projects & Portfolio

Selected projects across AutoML, optimization, time-series/ECG analytics, and applied AI in healthcare & sports. Click any card to expand methods and conclusions.

Repo Paper

Methods & conclusions

I developed an AutoML approach for the CASH problem that runs multiple bio-inspired optimizers in parallel using a generalized island model, combined with multi-fidelity evaluation to reduce wasted compute on expensive candidates. I studied three island topologies (fully-connected, unconnected, ring) to quantify the trade-off between information sharing and maintaining diverse search behavior.

The approach improved convergence time versus asynchronous evolutionary baselines and successive-halving-style methods, while remaining competitive against established AutoML frameworks in an objective OpenML-based comparison—showing that diversity-aware parallel search can deliver strong results under realistic time budgets.

Repo Paper

Methods & conclusions

I extended the island-model AutoML approach with a dynamic migration strategy and multi-objective, multi-fidelity evaluation. The goal was to avoid “discarding too fast” under multi-fidelity methods and to better handle changing data conditions by combining evolutionary search, diversity preservation, and adaptive information exchange.

I benchmarked 10 network topologies and found that hypercube and grid (2D/3D) layouts consistently provided the best balance between connectivity and diversity. These structures remained competitive against state-of-the-art AutoML systems, supporting distributed designs that scale efficiently without collapsing into early dominant solutions.

Repo Paper

Methods & conclusions

I investigated how to use Large Language Models as a practical layer in AutoML: generating candidate pipelines with domain guidance and leveraging empirical metadata from prior experiments. I compared two dataset similarity metrics to retrieve relevant prior runs and improve LLM-guided pipeline proposals, then stress-tested prompt strategies across multiple “challenge levels” to map where LLM guidance helps (and where it breaks).

The study provides a realistic integration blueprint for industry: how to constrain prompts, how to select prior metadata, and how to combine LLM suggestions with established optimizers (including Bayesian optimization) while mitigating failure modes such as invalid pipelines or overconfident but poor defaults.

Project site Partners

Methods & conclusions

I served as the main responsible for AI and data-driven applications within a large European consortium involving 26 partners (including Philips, TNO, imec, PSV and clinical stakeholders). I coordinated problem framing, data workflows, and model development across multiple use-cases—translating real constraints (privacy, noisy signals, deployment realities) into implementable analytics.

The outcome was an applied, multi-stakeholder AI portfolio with measurable value: wearable sensor analytics for athlete monitoring, arrhythmia-related modeling for clinical settings, and reusable data pipelines that supported cross-partner collaboration and reproducibility beyond a single proof-of-concept.

Context

Methods & conclusions

Within INNO4HEALTH, I collaborated with PSV’s women’s youth team and the Canadian company XCO to analyze multi-day wearable data (heart rate, breathing-related signals, and physiological measures). I built a practical analytics workflow: data cleaning, segmentation by training context, feature extraction, and reporting that coaches and stakeholders could interpret.

The analysis highlighted peaks during training/exercise and patterns across days, and linked physiological responses with questionnaires to contextualize external influences (sleep, fatigue, perceived exertion). This demonstrated a scalable pipeline for turning noisy real-world sensor streams into actionable insights.

Paper

Methods & conclusions

I analyzed signal quality for single-lead ECG patches with self-adhesive dry electrodes, collected from six cardiac patients over five days. The work focused on practical reliability: measuring quality degradation, identifying daily patterns, and assessing whether changes were statistically meaningful—key for deploying ML models that depend on stable physiological inputs.

We observed that ECG quality tends to be higher at night (consistent with reduced motion artifacts) and that day-to-day changes were present but not statistically significant. These findings help set realistic expectations for ambulatory monitoring and improve downstream AF/CVD detection performance in real deployment conditions.

Paper

Methods & conclusions

I built an end-to-end methodology to detect ST-segment deviation and estimate its magnitude (up to 1 mV) from single-lead ECG. The pipeline extracted statistical and beat-level characteristics along with signal quality indices, and used AutoML to select strong classification and regression models with tuned hyperparameters.

Validated on the PhysioNet ST-T database, the approach achieved very high performance (98.30% accuracy for multiclass and 99.87% for binary). The key takeaway is that careful feature/SQI design combined with automated model selection can match expert-level screening while keeping the workflow scalable.

Repo

Methods & conclusions

I built a lightweight system to “chat with” OpenML documentation, addressing a common industry pain point: valuable platforms with steep documentation overhead. The workflow crawls and structures docs, indexes them for retrieval, and uses RAG to generate grounded answers while pointing users to the relevant source sections.

The result is a practical productivity tool for faster onboarding and fewer context-switches—useful for teams that rely on OpenML for benchmarking, dataset discovery, or reproducible experimentation.

Website

Methods & conclusions

In collaboration with Newcastle University, I evaluated state-of-the-art computer vision methods for counting marine organisms from images as a scalable alternative to time-consuming lab workflows. We compared multiple approaches (including prompt-based segmentation/detection and few-shot-style counting) across devices, organism aggregation levels, and against human counts.

We found that high-resolution imaging improves reliability, and that some methods excel when organisms are centrally aggregated. However, accuracy plateaus at high densities, highlighting the need for standardized imaging protocols and adaptation for mixed-species scenarios—an actionable roadmap for deploying CV in ecological monitoring.

Repo

Methods & conclusions

I investigated how to make AutoML evaluation more practical by reducing compute in the AutoML Benchmark (AMLB). I ran a large comparison across 11 AutoML frameworks on 104 tasks under multiple time limits and assessed how rankings shift when budgets shrink, as well as how early-stopping mechanisms change outcomes across tools.

The results show framework rankings are often reasonably stable under shorter budgets—supporting “faster, cheaper” benchmark configurations for industry and academia. Early stopping increases variability in outcomes, which is valuable: it reveals differences in robustness and budget-efficiency that matter in real retraining pipelines.