Akıllı sistemler ve uygulamaları dergisi

Artificial intelligence–driven data architecture pipelines are increasingly deployed in domains characterized by heterogeneous and evolving datasets, where structural consistency of data plays a critical role in maintaining predictive reliability. However, schema volatility arising from dynamic data sources, evolving annotation standards, and multi-modal integration introduces hidden instabilities that are not addressed by conventional data drift frameworks. Existing literature in medical imaging, microbiology, and public health analytics demonstrates the sensitivity of machine learning models to such structural variations, yet the propagation behavior of schema changes across pipeline stages remains insufficiently quantified. This research addresses this gap by systematically analyzing how schema perturbations propagate through preprocessing, feature extraction, encoding, and model inference layers, and by introducing quantitative metrics to capture distortion and drift accumulation effects. The study presents a formalized pipeline model with schema perturbation operators and evaluates propagation dynamics across representative datasets, revealing nonlinear amplification of errors and critical transformation points that influence system stability. The findings highlight that schema volatility leads to cumulative degradation in predictive performance, particularly in long-lived and compliance-critical AI systems, and that conventional mitigation strategies are inadequate to fully contain these effects. The work concludes by emphasizing the need for schema-aware, adaptive pipeline architectures capable of detecting and mitigating structural inconsistencies in real time. The proposed framework has direct applications in healthcare analytics, microbiological surveillance, and telemedicine systems, where maintaining data integrity and model reliability is essential for accurate and trustworthy decision-making.