Accurate state-of-health estimation is crucial for reliable battery management in electric vehicles and stationary energy systems. Although many studies employ the NASA open battery dataset, most focus on a single pack offering the cleanest data at room temperature with similar C-rates. This limits the applicability of the model to more diverse conditions. In this work, we leverage incremental capacity (IC) analysis as a core feature extraction method while extending to broader usage profiles using transfer learning. Our comprehensive clustering study shows that IC curves alone suffice to capture the aging phenomena, an observation that aligns with earlier single-pack results. To address scenarios where newly encountered operational conditions deviate significantly from the training set, we introduce two enhancements: (1) a distance-based uncertainty estimation mechanism that flags uncertain predictions under conventional transfer learning available in literature without explicit adaptation, and (2) a domain adaptation framework that realigns either individual source instances via Kernel Mean Matching (KMM) or feature representations via Domain-Adversarial Neural Network (DANN) to handle wide temperature and usage variations. Our results reveal that DANN offers higher predictive accuracy across diverse temperatures, while KMM remains a practical intermediary technique that shows fundamental analogies with clustering. Empirical evaluation demonstrates that the proposed approach can be applied across multiple regimes, unlike single-pack baselines, confirming that a model trained with IC-based features and domain adaptation can generalize well beyond the narrow conditions typically targeted in previous work. This method paves the way for more robust, data-driven battery health assessments in real-world applications.