Georgios (George) Kontogiannis

Background: Cough sounds are accessible, non-invasive biomarkers for respiratory disease assessment and can be captured using consumer-grade smartphones. Existing approaches typically focus solely on cough detection or rely on server-based deep learning for disease classification, which limits deployability and raises privacy concerns. Small, imbalanced cough datasets further hinder model generalization. Objective: To develop a multilayer, smartphone-compatible AI framework for automated cough detection and respiratory disease classification, and to propose a pioneering generative augmentation strategy utilizing a suite of five Variational Autoencoder (VAE) variants and a probabilistic cough-level fusion mechanism to improve disease classification under severe data scarcity and the limitations of conventional audio augmentation techniques. Methods: The proposed framework consists of three AI modules: (1) A Cough Detection Module (CDM) that performs real-time cough event detection and segmentation from continuous audio using lightweight models optimized for on-device execution. (2) A Disease Analysis Module (DAM) that classifies cough events into asthma, COVID-19, or healthy classes using parallel Support Vector Machine classifiers and a probabilistic cough-level fusion strategy. (3) A Generative Augmentation Module (GAM) employing five distinct VAE architectures. This module uniquely operates across the time-frequency domain for latent feature optimization while reconstructing samples in the time-domain to ensure acoustic verifiability. Results: The CDM provides reliable segmentation across heterogeneous recording conditions. The DAM achieves strong discriminability between asthma, COVID-19, and healthy coughs using compact cepstral and spectral features. The multi-variant GAM framework demonstrates superior efficacy in alleviating class imbalance specifically, the cross-domain (time-frequency to time-domain) reconstruction allows for the clinical verification of synthetic biomarkers. All system components operate in real time on commodity Android hardware. Conclusion: This integrated framework addresses key limitations in dataset availability, model generalizability, and deployability, introducing an interpretable generative approach that demonstrates the feasibility of smartphone-based acoustic sensing as a scalable and privacy-preserving tool for respiratory health monitoring.

This paper presents a straightforward yet novel approach to enhance graph-based information retrieval models, by calibrating the relationships between node terms, leading to better evaluation metrics at the retrieval phase, and by reducing the total size of the graph. This is achieved by integrating spectral clustering, embedding-based graph pruning and term re-weighting. Spectral clustering assigns each term to a specific cluster, allowing us to propose two pruning methods: out-cluster and in-cluster pruning based on node similarities. In-cluster pruning refers to pruning edges between terms within the same cluster, while out-cluster pruning refers to edges that connect different clusters. Both methods utilize spectral embeddings to assess node similarities, resulting in more manageable clusters termed concepts. These concepts are likely to contain semantically similar terms, with each term’s concept defined as the centroid of its cluster. We show that this graph pruning strategy significantly enhances the performance and effectiveness of the overall model, reducing, at the same time, its graph sparsity. Moreover, during the retrieval phase, the conceptually calibrated centroids are used to re-weight terms generated by user queries, and the precomputed embeddings enable efficient query expansion through a k-Nearest Neighbors (K-NN) approach, offering substantial enhancement with minimal additional time cost. To the best of our knowledge, this is the first application of spectral clustering and embedding-based conceptualization to prune graph-based IR models. Our approach enhances both retrieval efficiency and performance while enabling effective query expansion with minimal additional computational overhead. Our proposed technique is applied across various graph-based information retrieval models, improving evaluation metrics and producing sparser graphs.

Respiratory sound classification from auscultation recordings is confounded by heterogeneous acquisition conditions, including recording device, auscultation site, and patient characteristics, that introduce systematic variability unrelated to the underlying pathology. We propose a family of metadata-conditioning mechanisms for the Audio Spectrogram Transformer (AST), progressing from gated residual fusion to Feature-wise Linear Modulation (FiLM) adapted to the Transformer architecture, including two novel variants: Token-Aware FiLM (TAFiLM) and Soft-Factorized FiLM (SoftFiLM). Experiments on the ICBHI benchmark demonstrate consistent gains across all conditioning strategies. SoftFiLM achieves 64.11% on the official 4-class task (+3.94% over the baseline) and 72.40% on the 2-class task, establishing a new state-of-the-art on the 2-class task.