UC Irvine

Cardiovascular disease remains the leading cause of mortality worldwide, accounting for nearly one in three global deaths. The diagnostic instruments that physicians and biomedical researchers rely on, such as electrocardiography, contrast-enhanced imaging, and high- frequency echocardiography, require trained operators, controlled clinical settings, or direct physical contact with the subject. This dissertation develops, validates, and analyzes computational methods that lower this access barrier across two scales: small-animal cardiac phenotyping and contactless human cardiac monitoring. Section I develops AI-guided cardiac imaging for teleost preclinical models. The Zebrafish Automated Cardiac Analysis Framework (ZACAF) replaces manual ventricular tracing with a U-Net segmentation backbone refined by data augmentation, transfer learning, and test- time augmentation. On the n=41 nrap mutant cohort spanning three genotypes, it recovers manual ejection fraction (EF) and fractional shortening within 2.0% and 1.7% mean absolute error at validation IoU 0.876. A refined U-Net + EfficientNet encoder generalizes the approach to multi-cohort zebrafish echocardiography at validation Dice 0.967 and IoU 0.937, with EF error rate 9.83% on a held-out set of 51 test videos. A spatiotemporal Video Vision Transformer (ViViT) + SegFormer hybrid with a dedicated regression head extends the framework to the African turquoise killifish lifespan, with n=101 ground-truth EF an-notations across 164 longitudinal echocardiographic videos at ages 8 to 20 weeks, attaining Dice 0.871, IoU 0.773, EF Pearson r=0.89 (95% CI [0.74, 0.96]), MAE 5.4%, and intraclass correlation 0.88 (95% CI [0.72, 0.95]). Section II develops contactless cardiac monitoring from ordinary facial video. A deployable end-to-end framework couples a VideoMamba state-space video encoder with a physiology- informed five-wave (P, Q, R, S, T) McSharry Gaussian-mixture decoder that maps upstream cardiac-related representations into a shared ECG-oriented reconstruction space. Evaluated on 5 held-out subjects of a 22-subject UC Irvine cohort (IRB #STUDY00000473) with synchronized facial video and BIOPAC five-lead ECG, the direct video-to-ECG model attains BPM MAE 7.6 bpm, RR MAE 0.14 s, QRS rejection 0.35, P/T rejection 0.63, and reconstruction SNR 18.09 dB, outperforming cascaded video-to-rPPG-to-ECG baselines on every morphology-domain metric. The two sections trace a continuum from preclinical bench to consumer-camera deployment, along which the same methodological problem—inverting an indirect physical measurement to recover a hidden physiological state. When the inferential gap between an indirect, low-burden measurement and a clinically actionable readout closes, the access barrier of cardiovascular assessment will be lower.