NIH-funded research project

ORBIL-AI

Orthogonality-Regularized Biomedical Information-theoretic Learning

Funding
National Institutes of Health, National Library of Medicine
Mechanism
R16 Support for Research Excellence (SuRE)
Award number
1R16LM015384-01
Principal Investigator
Dr. Timothy Oladunni
Institution
Department of Computer Science, School of Computer, Mathematical and Natural Sciences, Morgan State University
Project period
September 10, 2026 – July 31, 2030

Project Overview

ORBIL-AI investigates a fundamental problem in multimodal machine learning: redundant information across learned representations. When a model combines several views of the same signal, such as time-domain, frequency-domain, and time–frequency representations of an electrocardiogram (ECG), standard fusion methods optimize only the prediction task and assume that each stream will contribute something new. In practice, separate encoders can learn overlapping features of the same events, so that adding streams need not add information.

The project builds on the published Complementary Feature Domains for Optimal ECG Multimodal Deep Learning Performance (CFD-MED) work, published as “Rethinking Multimodality” (IEEE Access, 2025), and asks whether explicitly encouraging complementary, near-orthogonal representations improves predictive performance, probability calibration, interpretability, and generalization. It does so by adding two kinds of constraint to the learning objective: information-theoretic penalties on redundancy between streams (normalized mutual information and maximum canonical correlation) and a geometric penalty that pushes the latent representations toward orthogonality.

The questions are about representation learning and information theory; ECG classification is the setting in which they are tested. The Research Strategy notes that the principles are not specific to ECG; within this project they are studied with ECG data.

Shared and distinct information Two overlapping circles labeled Observation A and Observation B. The overlap is labeled shared information; each non-overlapping part is labeled distinct information. Distinct to A Shared by A and B Distinct to B Observation A Observation B
Schematic: redundant streams overlap heavily; complementary streams contribute information the others do not.

ORBIL-AI is a research project investigating fundamental scientific questions. It is not a clinical product, a validated technology, or a research center, and its methods have not been clinically validated.

Project Aims

The three aims below are summarized from the project’s Research Strategy. All describe planned work.

Aim 1

Develop and validate ORBIL-AI’s core methodology

Develop a multimodal ECG representation-learning framework with three specialized encoders: a 1D-CNN for the time-domain waveform, a Transformer for the frequency spectrum, and a 2D-CNN for time–frequency scalograms. Investigate a composite learning objective that adds three regularizers to the task loss: normalized mutual information, maximum canonical correlation, and latent-space orthogonality. Evaluate each component through systematic, hypothesis-driven ablation studies.

  • Ablations compare a baseline, each regularizer alone (MI, CCA, orthogonality), and the full model.
  • Regularizer weights tuned by a coarse grid followed by Bayesian optimization with early stopping.
  • Models, training scripts, and evaluation pipelines to be versioned and released publicly.
Aim 2

Evaluate ORBIL-AI’s robustness and trustworthiness

Evaluate generalization across publicly available ECG datasets, including device- and site-held-out testing. Assess predictive performance, probability calibration, explanation fidelity, and robustness to signal disturbances.

  • Calibration: Expected Calibration Error (ECE), Negative Log-Likelihood (NLL), and reliability diagrams.
  • Explanation fidelity: Saliency Alignment Index (SAI) and Attribution Matching Index (AMI), measured against expert-annotated ECG landmarks.
  • Stress tests from baseline wander to lead dropout, with documented failure modes compiled in a public “failure atlas.”
Aim 3

Capacity building and student pipeline at Morgan State University

Develop a structured undergraduate and master’s student research program with paid research positions, collaborative research teams (“pods”), reproducible machine-learning workflows, mentorship, and curriculum development, and support student participation in publications and conference presentations.

  • Paid fellowships for at least two students per year, with emphasis on multi-year retention.
  • Pods for data curation, model development and ablations, and evaluation.
  • Training in reproducible workflows (DVC, MLflow); weekly mentoring meetings and quarterly skills workshops.
  • A new course module, “Multimodal AI for Biomedical Signals,” developed with the School of Computer, Mathematical, and Natural Sciences.

Preliminary Evidence

The Research Strategy reports the following preliminary ECG classification results (macro-F1 on a held-out set). They come from the earlier CFD-MED work that motivated the project; they are not ORBIL-AI project outcomes.

ModelMacro-F1 (scale 0 to 1)
Single-stream 1D-CNN (time) 0.94
Two-stream fusion (time + time–frequency) 0.96
Three-stream fusion (time + frequency + time–frequency) 0.94

Adding a third stream did not improve on the two-stream model. The Research Strategy reports that this coincided with higher redundancy between streams and lower latent orthogonality, and that saliency maps from the two-stream model aligned more consistently with ECG landmarks. These descriptive observations motivate studying representation complementarity directly, rather than assuming that additional feature streams improve performance. They are not tests of statistical significance.

Research & Training Plan

All activities below are proposed; none is reported as complete.

  • Information-theoretic and geometric regularization

    Penalties on mutual information and canonical correlation between streams, combined with a latent orthogonality constraint.

  • Hypothesis-driven ablations

    Isolating the contribution of each regularizer against a baseline and the full model.

  • Out-of-distribution testing

    Device- and site-held-out partitions of public ECG datasets.

  • Calibration and explanation fidelity

    ECE, NLL, and reliability diagrams; SAI and AMI against annotated ECG landmarks.

  • Robustness and failure analysis

    Stress tests for noise, baseline wander, and lead dropout, with a public failure atlas.

  • Reproducible workflows and open source

    Versioned data and code (DVC, MLflow) and a public software release.

  • Student training and curriculum

    Paid fellowships, research pods, mentorship, workshops, and a new course module.

Scientific Significance

Redundancy and complementarity between representations are central questions in pattern recognition and representation learning, yet multimodal models are usually designed by heuristic combination of inputs. ORBIL-AI treats complementarity as an explicit training objective and evaluates its effects on calibration and explanation fidelity as well as accuracy, aiming at principles of how information is represented and integrated in learning systems rather than improvements to a single model or dataset.

Prior findings

What earlier work established

Peer-reviewed work preceding the award (IEEE Access, 2025) found that fusing time and time–frequency representations improved on a single-domain baseline, while adding a frequency-domain stream produced no further gain, consistent with redundancy.

Working hypotheses

What the project will test

That explicitly penalizing redundancy and encouraging near-orthogonal latent representations improves predictive performance, calibration, interpretability, and out-of-distribution generalization. These hypotheses have not yet been tested within the project.

Open questions

What remains unknown

How reliably redundancy can be estimated in high-dimensional learned representations, how the three regularizers interact, and when enforcing separation improves rather than limits what a model learns.

Research Team

  • Dr. Timothy Oladunni

    Principal Investigator

    Assistant Professor, Department of Computer Science, School of Computer, Mathematical and Natural Sciences, Morgan State University · Profile

  • Farouk Ganiyu Adewumi

    Graduate Researcher (PhD)

    Morgan State University · Lab profile

  • Raymond Mbam

    Graduate Researcher (PhD)

    Advanced Computing, Morgan State University · Lab profile

  • Oluwafemi Ifesanmi

    Graduate Researcher

    Advanced Computing, Morgan State University · Lab profile

  • Daniel Owolabi

    Graduate Researcher (M.S.)

    Advanced Computing, Morgan State University · Lab profile

  • Obaloluwa Arulogun

    Undergraduate Researcher

    Computer Science, Morgan State University · Lab profile

  • Aditya Raj Poudel

    Undergraduate Researcher

    Computer Science, Morgan State University · Lab profile

Students participate through the Pattern Recognition & Machine Learning Laboratory. No external collaborating institutions are currently listed for this project.

Timeline & Progress

Progress to date

  1. April 2025Completed

    Preliminary findings on redundancy and complementarity in multimodal ECG, and the Complementary Feature Domains (CFD) perspective, presented at the Yale University Computer Science Colloquium (before the award).

  2. 2025Completed

    Visiting faculty research fellowship at Yale University in cardiovascular science and computational AI in medicine (before the award).

  3. September 10, 2026Completed

    Project period began.

  4. October 29, 2026Completed

    Award announced by Morgan State University.

  5. Fall 2026In progress

    Undergraduate research on whether models that combine time, frequency, and time–frequency ECG representations learn complementary or redundant information, following a research plan specified in advance.

Planned four-year timeline

Adapted from Figure 3 of the Research Strategy. Months are counted from the project start (September 2026) and bar positions are approximate. Milestones at months 24, 36, and 48.

Months 0–24: setup, Aim 1 development, Aim 2 evaluation

Recruit student fellows Approximately months 0 to 2
Environment and repositories (DVC, MLflow) Approximately months 0 to 2.5
Onboarding and responsible conduct of research Approximately months 1.3 to 4.2
Data curation Approximately months 2 to 7.8
Encoders (time, frequency, time–frequency) Approximately months 3.8 to 10
ORBIL composite loss Approximately months 5 to 9.5
Ablations (baseline to full model) Approximately months 7.5 to 16.5
Hyperparameter tuning Approximately months 10 to 14
Out-of-distribution splits (device and site) Approximately months 14 to 17.7
Calibration (ECE, NLL) Approximately months 15 to 21
Explanation fidelity (SAI, AMI) Approximately months 17.7 to 22.5
Manuscript drafting and submission Approximately months 19 to 24

Months 24–48: Aim 3 capacity building and dissemination

Student pipeline and research pods (ongoing) Approximately months 0 to 24
MLOps training (DVC, MLflow) Approximately months 6.5 to 16
Curricular integration and mentorship Approximately months 7.3 to 16
Open-source releases Approximately months 13 to 24
Follow-on proposal Approximately months 17.5 to 24

Publications & Research Outputs

Outputs supported by this award will be listed here as they become available. The project began on September 10, 2026.

Status of project outputs
Output typeStatus
Peer-reviewed publicationsNone yet
PreprintsNone yet
SoftwarePlanned: public release of models, training scripts, and evaluation pipelines
Datasets and resourcesPlanned: public failure atlas from robustness testing
Presentations & other outputsNone yet

Background: prior work that preceded the award

Not supported by this award; listed because the project builds on it.

  1. Rethinking Multimodality: Optimizing Multimodal Deep Learning for Biomedical Signal Classification T. Oladunni, A. Wong IEEE Access, 2025. Peer-reviewed. The CFD-MED work on which ORBIL-AI builds (Complementary Feature Domains for Optimal ECG Multimodal Deep Learning Performance).
  2. Explainable Deep Neural Network for Multimodal ECG Signals: Intermediate Versus Late Fusion T. Oladunni, E. Aneni IEEE Access, 2025. Peer-reviewed.
  3. More Isn’t Always Better: Investigating Redundancy and Complementarity in Multimodal ECG Deep Learning Computer Science Colloquium, Yale University, April 2025. Talk.

Funding Acknowledgment

Research reported on this page is supported by the National Library of Medicine of the National Institutes of Health under Award Number R16LM015384 (1R16LM015384-01), as part of an award totaling approximately $520,000 over four years, with an initial award of $130,134.

The content is solely the responsibility of the author and does not necessarily represent the official views of the National Institutes of Health.