Research

Theoretical frameworks for representation, robustness, and interpretability in machine learning, studied with cardiovascular signals

Research Program

The group’s research is organized around nine theoretical frameworks. They are related, sharing questions about how information is represented, preserved, and integrated by learning systems, but each is a distinct contribution with its own assumptions, claims, and evidence. They are not presented as a single hierarchy, and progress on one does not establish another.

Each entry below separates what has been reported, what is hypothesized, and what is still under investigation.

Funded Projects

Frameworks at a Glance

Peer-reviewed Published in a peer-reviewed venue
Preprint Publicly available; not yet peer reviewed
Ongoing investigation Formulated; results not yet reported

Representation, Complementarity & Trustworthy Learning

Dynamical-Systems Analysis of Cardiovascular Signals

Theoretical Frameworks

Representation, Complementarity & Trustworthy Learning

Complementary Feature Domain (CFD) Theory

When do several representations of the same signal add information, rather than repeat it?

Peer-reviewed

Core idea

Time-domain, frequency-domain, and time–frequency representations of an ECG can carry complementary or redundant information. CFD proposes that the benefit of multimodal fusion depends on the complementarity of the fused domains, not on the number of domains combined.

Reported evidence

In a peer-reviewed study (IEEE Access, 2025), a model fusing time and time–frequency representations outperformed the time–frequency baseline across evaluation metrics under bootstrap and Bayesian analysis, while adding the frequency domain produced no further improvement. A targeted ablation supported the interpretation of representational redundancy. A 2026 preprint formalizes CFD’s notion of predictive contribution and argues that preserving Shannon information in a representation does not by itself preserve how that representation contributes to prediction; this result has not yet completed peer review.

Hypotheses & open questions

How far the complementarity pattern generalizes beyond the ECG datasets and architectures studied, and how complementarity can be measured information-theoretically and enforced during training. The latter is a focus of the NIH-funded ORBIL-AI project.

Explainable-AI Trustworthiness (EAT)

When is an explanation of a model’s prediction trustworthy?

Peer-reviewed basis

Core idea

Explanations are treated as objects to be evaluated, not as illustrations. EAT sets out criteria for trustworthy ECG model explanations, including whether attributions are grounded in clinically meaningful signal segments and whether they remain stable under perturbation.

Reported evidence

A peer-reviewed comparison of intermediate and late fusion for multimodal ECG (IEEE Access, 2025) analyzes model explanations with saliency methods. The accompanying preprint describes testable criteria, including informational grounding to the ST–T segment quantified with adjusted and normalized mutual information, and perturbation tests.

Hypotheses & open questions

How the criteria transfer across architectures, datasets, and signal types, and how explanation quality relates to clinical judgment.

Physiologic Invariance Theory (PIT)

Which changes to the input should leave a model’s prediction unchanged?

Preprint

Core idea

Robustness is defined relative to physiologic drift: natural variation in rhythm, amplitude, and morphology that does not reflect a change in condition and should not change a prediction. PIT proposes metrics and training strategies that keep latent representations stable under such benign variation while remaining sensitive to pathological change.

Reported evidence

Formulated in a 2025 preprint (TechRxiv). The work has not yet completed peer review.

Hypotheses & open questions

Whether invariance learned on benchmark ECG data holds in clinical and wearable recording conditions, and how to separate benign variation from early pathological change.

Physiologic Energy Conservation Theory (PECT)

How far may a model’s internal representation move when the signal’s energy changes?

Preprint

Core idea

An energy-based criterion for concept drift. Under label-preserving variation, the magnitude of latent-representation drift should scale with the normalized change in signal energy; persistent violations of this proportionality indicate real concept drift. The criterion is operationalized as an energy-constrained regularizer (ECRL).

Reported evidence

A 2026 preprint evaluates the criterion on multimodal ECG across seven unimodal and hybrid models and reports reduced representation drift under physiologic perturbations. The work has not yet completed peer review.

Hypotheses & open questions

A related 2026 preprint presents a CFD-guided framework (PECS) for deciding when a model should hold, update, or flag its prediction under physiologic drift in ECG, PPG, and respiration. It reports improved drift classification over the baselines evaluated and the largest benefit from combined ECG and PPG when the two signals disagree, and it notes limits: the best-performing domain pair changed across datasets, and respiration helped only when treated as a selective trust signal. Larger validation with clinically adjudicated drift labels remains future work.

Complementarity-Preserving Generative Theory (CPGT)

Can synthetic signals preserve the cross-domain structure of real ones?

Preprint

Core idea

Generative models can collapse domain-specific information. CPGT imposes explicit complementarity constraints during generation, implemented in a quantum-inspired generative adversarial model (Q-CFD-GAN) for multimodal ECG synthesis.

Reported evidence

A 2026 preprint reports synthetic ECG that preserves time, frequency, and time–frequency structure and examines its use for data augmentation. The work has not yet completed peer review.

Hypotheses & open questions

Whether complementarity-preserving augmentation improves downstream robustness across datasets and tasks, and how to assess the fidelity of synthetic physiological signals.

Dynamical-Systems Analysis of Cardiovascular Signals

Cardiac Stability Theory (CST)

Can cardiovascular stability be defined from the geometry of a reconstructed cardiac attractor?

Preprint

Core idea

A theoretical model, under investigation, that proposes characterizing cardiovascular health as the maintenance of moderate dynamical complexity on a bounded, quasi-periodic cardiac attractor. The framework is stated axiomatically and is a hypothesis to be tested, not an established definition of health. CST proposes a Cardiac Stability Index (CSI) computed from three geometric properties: the largest Lyapunov exponent, recurrence determinism, and signal entropy.

Reported evidence

Formulated in a 2026 preprint, now under review. A related 2026 preprint evaluates a stabilized index (SCSI) on 176,742 segments from four PPG datasets, identifies three evaluation artifacts that inflated a heuristic AUC from 0.573 to 0.752, and reports a pooled AUC of 0.757 (95% CI 0.686–0.828) and NPV of 0.966 for tachypnea screening on 18 held-out records; per-record AUC was 0.497 ± 0.207. Neither preprint has yet completed peer review.

Hypotheses & open questions

Whether the index tracks clinically meaningful change within individuals, given the variability in per-record performance. A 2026 preprint on 655 recordings found that maximal Lyapunov exponents showed essentially no recording-specific correspondence between contact PPG and camera-based rPPG, so attractor properties measured with contact sensors may not carry over to camera-acquired signals. The index is used in a lab prototype application (HeartSpan Score) for research purposes.

Attractor Domain Theory (ADT)

Which properties of a reconstructed attractor carry which cardiovascular information, and why?

Preprint

Core idea

ADT proposes that the information in a reconstructed cardiac attractor partitions into three mutually non-redundant domains: a Geometry Domain (delay embedding; artifact rejection), an Ergodic Domain (asymptotic statistical invariants; stability estimation), and a Variational Domain (finite-time Lyapunov exponent field; hemodynamic inference). The preprint states a Domain Sufficiency Theorem asserting that the three domains are necessary and sufficient; the theorem has not yet undergone independent peer review.

Reported evidence

A 2026 preprint reports Geometry Domain validation through the SCSI index on 176,742 PPG segments from four datasets (AUC 0.757, 95% CI 0.686–0.828; NPV 0.966 for tachypnea screening on held-out records, after correcting three evaluation artifacts). The work has not yet completed peer review.

Hypotheses & open questions

Empirical validation of the Ergodic and Variational domains, and independent review of the Domain Sufficiency Theorem.

Attractor-Vascular Coupling Theory (AVCT)

Can the attractor geometry of a PPG signal support calibrated, cuffless blood pressure estimation?

Preprint

Core idea

AVCT proposes a formal coupling between vascular mechanics and the geometry of the cardiac attractor, as a basis for calibrated cuffless blood pressure estimation, operationalized with Takens delay embedding and attractor morphology features. The preprint develops this coupling formally to generate testable hypotheses.

Reported evidence

The current version of the preprint (v3, October 2026) reports two retrospective evaluations of a calibrated estimator using leave-one-subject-out cross-validation on existing datasets. (1) 46 intensive-care and surgical subjects (BIDMC and VitalDB) with repeated reference measurements: systolic/diastolic MAE 2.05/1.67 mmHg for the ECG + PPG model and 2.02/1.63 mmHg for the PPG-attractor-only model. (2) 217 ambulatory subjects from the public PPG-BP dataset, with one cuff reference per subject, which tests cross-subject calibrated estimation rather than tracking within a person: 1.98/0.96 mmHg. These errors fall within commonly cited AAMI/ISO and BHS thresholds, but, as the preprint states, this does not establish compliance with device-validation standards. No smartphone-acquired PPG was evaluated; the title names the intended application. The work has not yet completed peer review.

Hypotheses & open questions

Prospective validation with smartphone-acquired PPG and repeated reference blood pressure measurements is planned. Performance on smartphone-camera PPG, for tracking within a person over time, and across skin tones has not yet been established; related work on contact-to-camera PPG recoverability (CST) suggests that some attractor properties may not transfer to camera signals. The method is a research result, not a validated medical device.

Physiologic Trajectory Theory (PTT)

What constitutes a valid physiologic trajectory, as opposed to a meaningful departure from one?

Ongoing investigation

Core idea

PTT proposes per-individual admissible corridors for a stability index, estimated with Kalman state-space models, in place of comparison against population norms. Departures are classified with CUSUM sequential detection as Stable, Benign Excursion, or True Drift, guided by a proposed Cardiac Asymmetry Principle.

Reported evidence

No results have been reported yet. PTT is a working framework.

Hypotheses & open questions

Evaluation is planned on longitudinal smartphone PPG data under an IRB protocol that has been submitted (target enrollment 500 participants across Fitzpatrick skin types I–VI). The dataset is intended for public release, subject to IRB approval and participant consent.

Application Areas

Cardiovascular Signal Analysis

ECG and PPG analysis for cardiac abnormality detection, including earlier work on suspected hypertrophic cardiomyopathy from single-lead wearable ECG (IEEE Sensors Letters, 2021) and multimodal ECG classification (IEEE Access, 2025). Longitudinal monitoring (PTT) is under investigation.

Cuffless Blood Pressure Estimation

A proposed method for calibrated blood pressure estimation from attractor features of PPG (AVCT); it is not an experimentally validated capability for everyday use. Results to date come from retrospective evaluation on existing intensive-care, surgical, and ambulatory datasets and are reported in a preprint; no smartphone-acquired PPG has yet been evaluated, and prospective smartphone validation is future work. A related 2026 preprint studies context-augmented, session-calibrated blood pressure tracking from PPG (arXiv:2610.29091).

Equity Across Skin Tones

Optical PPG signals can vary with skin pigmentation. A longitudinal smartphone PPG study spanning Fitzpatrick skin types I–VI has been submitted for IRB review, to support evaluation of performance across demographic groups.