Unsupervised and Contextual Anomaly Detection with Application to Cardiology

Reliable clinical decision support requires quantitative tools that can assess whether a measurement is “normal” for a given patient and, crucially, whether such an assessment is trustworthy in the presence of heterogeneous populations, acquisition variability, and limited reference data. In many cardiology applications, obtaining unambiguous labels is inherently difficult. Intracardiac electrograms (EGMs) acquired during ablation procedures exhibit complex and patient-specific morphologies that are not easily mapped to discrete classes. At the same time, echocardiographic reference equations for aortic diameters may become unreliable in sparsely sampled or distribution-shifted patient contexts. These challenges motivate the investigation of data-driven approaches based on unsupervised learning, anomaly detection, and explicit uncertainty quantification. This thesis explores complementary methodological directions for anomaly detection with applications to cardiology. First, we propose an unsupervised deep anomaly detection framework to characterize atrial EGM morphology directly from raw waveforms. The framework produces robust anomaly scores that correlate with established electrophysiological indicators, including voltage, fractionation, and duration, while yielding coherent electroanatomical maps without the need for manually tuned thresholds or ad hoc combinations of handcrafted features. This provides a more synthetic and morphology-oriented description of atrial substrate. Second, we introduce the normalcy score (NS), a probabilistic generalization of Z-score reasoning for contextual anomaly detection, in which the score itself is treated as a random variable rather than as a deterministic quantity. NS leverages heteroscedastic Gaussian processes to model context-dependent mean and variance, thereby distinguishing between aleatoric and epistemic uncertainty. This formulation enables uncertainty-aware and interpretable assessments, particularly in borderline or poorly supported regions of the input space, where overconfident point estimates may be misleading. Third, we instantiate the same uncertainty-aware principle in echocardiography by reformulating the classical aortic Z-score. The resulting score provides clinicians with both an expected score and a highest-density interval, explicitly signaling when limited reference support makes the assessment less reliable. In this way, the proposed approach extends a familiar clinical tool while improving transparency and supporting a more cautious interpretation of abnormality. Finally, we include a complementary methodological study on few-shot source attribution of AI-generated images. By training compact tiny autoencoders, we show that reconstruction residuals can be exploited as lightweight and discriminative signatures, while remaining compatible with class-incremental updates under severe data constraints. Although this chapter is not centered on a cardiology application, it reinforces a broader methodological message of the thesis: reconstruction-based representation pipelines can provide effective and practical solutions in settings where labels are scarce and compact models are desirable. Overall, the thesis shows how unsupervised scoring, contextual probabilistic modeling, and uncertainty-aware inference can support more reliable quantitative assessment across heterogeneous biomedical settings, while also highlighting the importance of interpretable scores and explicit reliability estimates in high-stakes decision-making.