PhD Thesis Defense by Theofanis Ganitidis titled "Development of drift-adaptive machine learning methods to support health desicion making in dynamic data contexts"

In May 2026 Theofanis Ganitidis completed his PhD Thesis titled "Development of drift-adaptive machine learning methods to support health desicion making in dynamic data contexts" at the National Technical University of Athens (NTUA).

Abstract: The present thesis focuses on the development of drift-adaptive machine learning methodologies to support healthcare decision making in dynamic data environments. Conventional artificial intelligence (AI) systems in healthcare are often affected by data shift, namely changes between the distributions of data used during model development and those encountered during real-world deployment. Such changes may result from temporal dynamics, differences in acquisition conditions, variations in equipment and protocols, or demographic differences, potentially reducing model accuracy, generalization, and reliability. The thesis therefore investigates methods for detecting, quantifying, and adapting to two major forms of data shift in healthcare: temporal shifts in continuously collected data streams and acquisition-related shifts across different clinical cohorts.

First, temporal data shift is investigated in the context of COVID-19 detection using crowdsourced cough audio recordings from the COVID-19 Sounds and COSWARA datasets. A comprehensive drift-monitoring and adaptive framework for data streams is developed, employing the Maximum Mean Discrepancy (MMD) metric for detecting distributional changes and activating model adaptation mechanisms when drift is identified. The datasets are temporally divided into development and deployment periods, and a baseline Convolutional Neural Network (CNN) is initially trained on the development data. The baseline model achieves AUC values of 69.13% and 66.8%, and balanced accuracy values of 63.38% and 61.64% for the COVID-19 Sounds and COSWARA datasets, respectively. Its subsequent evaluation on deployment data demonstrates substantial performance degradation, confirming the presence of temporal drift. To address this degradation, two model adaptation strategies are examined: Unsupervised Domain Adaptation (UDA) and Active Learning (AL). The application of UDA improves balanced accuracy by up to 22% and 24% for the two datasets, respectively, while Active Learning produces improvements of up to 30% and 60%. These findings demonstrate that combining drift detection with adaptive learning mechanisms can substantially enhance the robustness of COVID-19 detection systems under severe temporal distribution shifts.

The thesis further investigates acquisition-related data shifts in multi-cohort clinical imaging, focusing on cardiovascular disease risk stratification using carotid ultrasound images. Initially, an interpretable ensemble comprising three independently trained deep learning CNN models is developed to discriminate between symptomatic and asymptomatic atherosclerotic cases. The dataset consists of 96 ultrasound images obtained from the General University Hospital “Attikon”. To address class imbalance, subsampling of the training sets, a two-phase training strategy, and cost-sensitive weighting are employed. The resulting ensemble achieves an AUC of 73%, sensitivity of 75%, and specificity of 70%. Interpretability methods are additionally applied to identify anatomical regions associated with increased cardiovascular risk and to provide greater transparency into model predictions. Subsequently, the effect of domain shift between different imaging centers is quantified. A CNN trained using data from “Attikon” is evaluated on the Carotid Ultrasound Boundary Study (CUBS) dataset, which contains 689 samples. Model performance decreases from an AUC of 90.1% and accuracy of 89.7% on the source dataset to an AUC of 45.6% and accuracy of 34.3% on CUBS, demonstrating the substantial impact of acquisition-related data shift. Hellinger distance, Kullback–Leibler (KL) divergence, and MMD are employed to quantify differences between feature distributions at multiple representation depths. The analysis reveals strong negative correlations between the measured degree of distribution shift and model accuracy, including a correlation coefficient of ρ = –0.75 for Hellinger distance.

Finally, an interpretable Active Learning framework is proposed to enable efficient model adaptation under domain shift. The framework combines uncertainty estimation through Monte Carlo Dropout, an MMD-based representativeness criterion, and pseudo-labeling for model retraining. The proposed approach achieves 94.2% of the optimal performance while using only 28.2% of the available data. Interpretability techniques, including Grad-CAM and LIME, are incorporated to provide transparency regarding both model predictions and uncertainty estimates.

Overall, the thesis contributes to the understanding and mitigation of data shift in AI-based healthcare applications by developing methodologies for drift detection, domain adaptation, active learning, uncertainty estimation, and model interpretability. The proposed frameworks aim to support the development of robust, efficient, and interpretable AI systems capable of maintaining reliable performance under evolving real-world clinical conditions.