Antônio H. Ribeiro

Assistant Professor at Uppsala University
Department of Information Technology

My goal is to understand and learn from signals and dynamical systems, building AI methods that meet the standards of evidence required in medicine.

Explore my research

Uppsala University
Sweden

About my work

My research brings together machine learning, automatic control, and signal processing. I work on the development of techniques to extract information and learn the intrinsic behavior of time series, signals, and dynamical systems, with a focus on large-scale models and their robustness and generalization. I am particularly interested in what these advances can bring to medicine, especially cardiology and electrocardiography.

Learning & robustness
Understanding generalization and reliability of large-scale models from first principles.
Signals & systems
Building tools for time series, system identification, and learning dynamics.
AI for medicine
Deploying advanced AI models in medical pipelines.

Research highlights

All publications

Circulation 2026

Opportunistic Screening for Chagas Disease Using an Artificial Intelligence–Enabled ECG: Prospective Evaluation of Feasibility and Diagnostic Accuracy

TL;DR AI-based screening during routine ECG care helped identify previously undiagnosed Chagas disease across 107 Brazilian municipalities. Combining ECGs with three risk-factor questions improved discrimination over either source alone.

Read paper

Screening workflow linking routine primary-care ECGs, AI-generated Chagas risk alerts, confirmatory blood testing, and reports returned to local health teams. — Ribeiro et al. · Figure 2

npj Digital Medicine 2026

CODE-II: a large-scale dataset for artificial intelligence in ECG analysis

TL;DR 2.7 million ECGs from 2.1 million patients, labeled across 66 diagnostic classes. CODE-II provides a foundation for training and evaluating ECG models, with an open subset and a dedicated test set for benchmarking.

Read paper

Map of Brazil showing the geographic distribution of CODE-II ECG recordings, with most collected in Minas Gerais. — Abreu et al. · Figure 1 CC BY 4.0

arXiv 2026

Distributionally robust linear regression through the lens of adversarial training

TL;DR We connect distributionally robust linear regression to adversarial training, bringing adversarial regression and square-root Lasso into one framework. This connection yields error guarantees, tuning that does not require knowing the noise level, and efficient algorithms.

Read paper

Simulation results showing prediction error decreasing with sample size across four Wasserstein orders, matching the predicted fast and slow convergence rates. — Stefansson et al. · Figure 2

NeurIPS 2026

Anytime-Valid Conformal Risk Control

TL;DR We build prediction sets whose risk stays below a chosen level with high probability as calibration data accumulates. The guarantee holds simultaneously at every sample size, so calibration can stop at any time.

Read paper

Standard conformal prediction compared with the anytime-valid method: miscoverage rates fluctuate around the 5 percent target on the left and stay below it for most calibration sequences on the right. — Hultberg et al. · Figure 1

Get in touch

antonio.horta.ribeiro@it.uu.se

Department of Information Technology
Ångströmslaboratoriet, Hus 10
Room 103179

Uppsala University
Box 337
751 05 Uppsala, Sweden