research statement | Georgios (George) Kontogiannis

My research builds machine-learning systems that turn everyday devices—smartphones and wearables—into accessible tools for health monitoring and early screening. It is organized around a single question: what does it take for a model to stay accurate, calibrated, and trustworthy not on a curated benchmark, but under the messy, heterogeneous conditions of real-world use?

Answering this requires progress on three coupled fronts—learning robust representations from scarce and noisy medical signals, generating data where little exists, and deploying models that behave predictably on-device and across patient populations. I pursue these questions primarily in respiratory and acoustic health, in close collaboration with clinicians and industry, and I validate them in deployed systems rather than in simulation alone. The sections below trace this arc from the core problem, through the methods I develop, to the clinical settings in which they are tested, and finally to the agenda I intend to pursue as a postdoctoral researcher.

Acoustic Biomarkers for Respiratory Health

Auscultation remains subjective and clinic-bound, yet the underlying acoustic signals are rich in diagnostic information. My core line of work asks whether commodity microphones can capture clinically meaningful respiratory biomarkers. I design pipelines that detect and segment coughs from continuous audio and classify respiratory conditions—such as asthma versus COPD—directly from tracheal- and chest-wall–acquired lung sounds, with models lightweight enough to run on-device, in real time, without specialized hardware. This program spans a systematic framing of the field (Sensors, 2024), a deployed on-device detection-and-classification system (Computers in Biology and Medicine, 2026), and edge-AI screening for obstructive disease (JBHI, under review).

Domain Generalization & Robust Deployment

Models that excel in the lab routinely fail in deployment: differences in recording devices, auscultation sites, and patient populations introduce systematic variability unrelated to the underlying pathology. Making models survive this shift is the methodological core of my work. I am developing a dedicated domain-adaptation framework for cross-dataset heterogeneity—so that a single model trained on one cohort, site, or device transfers to others without costly re-collection or retraining—together with deep denoising and signal-restoration architectures tailored to the very low signal-to-noise conditions of real-world medical recordings. This builds on metadata-conditioning mechanisms that let a transformer adjust to recording context (EUSIPCO, 2026) and on signal-level mitigation such as percussive-noise removal and frequency-spectrum correction across heterogeneous microphones and auscultation sites. Throughout, I treat calibration, explainability, and behavior on underrepresented patient subgroups as first-class objectives, because these properties determine whether a model can be trusted at the point of care.

Generative Modeling under Clinical Data Scarcity

Clinical audio datasets are small, imbalanced, expensive to label, and privacy-sensitive—conditions under which conventional augmentation falls short. I develop deep generative models (VAE and GAN families) that synthesize clinically meaningful cough and lung sounds and use them as a principled augmentation strategy to improve classifier robustness. A distinctive element of this work is cross-domain generation: optimizing latent representations in the time–frequency domain while reconstructing acoustically verifiable samples in the time domain, so that synthetic biomarkers remain clinically inspectable. I also study, more broadly, how synthetic data affects downstream recognition performance (DCOSS-IoT, 2024; CBM, 2026).

Multimodal & Clinical AI in Practice

Real diagnosis is multimodal, and my methods are tested in genuine clinical and consortium settings. With the University General Hospital of Patras, I build multimodal models that fuse paired lung-sound recordings, chest imaging, and inflammatory biomarkers—for example, to distinguish bacterial from viral pneumonia—and I develop the clinical data-acquisition tooling that makes such studies possible. Within the EU Horizon SynAir-G project I study causal relationships between indoor environmental signals, wearable physiology, and health outcomes. Through the Pfizer CDI EyeAI project I extend the same edge-deployment philosophy to computer vision, including sclera segmentation and non-invasive jaundice assessment from smartphone images (survey, under review).

Future Directions

As a postdoctoral researcher I want to push edge-deployed health AI from proof-of-concept toward dependable tools. The first two directions are already underway:

A cross-dataset domain-adaptation framework—a single model that transfers across cohorts, recording sites, and devices without re-collection or retraining, with robustness to distribution shift treated as a measurable design goal rather than a hoped-for side effect.
Denoising and signal-restoration architectures for low-SNR medical signals—recovering diagnostic structure from recordings captured on consumer hardware in uncontrolled, real-world environments.
Generation that improves calibration, not just accuracy—coupling generative augmentation with uncertainty-aware, robust training so synthetic data strengthens reliability rather than masking it.
Calibrated multimodal fusion under shift—principled handling of missing or degraded modalities at the point of care, with equitable behavior across underrepresented patient subgroups.
Decisions that are explainable and auditable to clinicians—closing the loop between model outputs and clinical accountability.

My longer-term goal is a general methodology for trustworthy edge health AI that transfers across acoustic, visual, and physiological signals. I am actively seeking postdoctoral positions and research collaborations where I can develop this agenda.