Reliable heart rate (HR) detection is essential for long-term cardiac monitoring, particularly in hospitalized patients with complex conditions. Due to its optical and non-invasive nature, photoplethysmography (PPG) is inherently susceptible to motion artifacts and noise. These challenges intensify under arrhythmic conditions such as atrial fibrillation (AF), where signal distortions may blur the boundary between poor-quality segments and pathological rhythms, potentially impairing downstream tasks like HR estimation. This study developed a signal quality assessment (SQA) algorithm designed for this high-risk clinical population and evaluated its robustness through HR estimation. We collected 24-hour synchronous PPG and electrocardiogram (ECG) recordings from 49 hospitalized cardiac patients, with all PPG segments manually annotated for quality. External validation was conducted using the MIMIC-IV dataset. To avoid dependence on specific segment lengths or classifier types, we assessed SQA performance using seven machine learning models and four segmentation lengths. The SQA framework was then applied to HR estimation to evaluate clinical utility. We implemented a Standard Deviation of Successive Differences (SDSD)-based peak filtering method and compared it with an autocorrelation-based approach under different cardiac rhythm conditions. Threshold tuning in both SQA classification and SDSD filtering was conducted to explore the balance between data usability and reliable HR estimation. The proposed model achieved an AUROC of 96.1% (Sinus Rhythm (SR) ＋ AF), with 90.6% on MIMIC-IV. Predicted SQA labels closely matched manual annotations, with mean absolute error (MAE) differences of 0.08 bpm (SR＋AF), 0.25 bpm (SR), 0.62 bpm (AF), and 0.53 bpm (MIMIC-IV). SDSD reduced MAE by 46.57% for SR＋AF, 41.67% for SR, and 49.69% for AF, further demonstrating the effectiveness of integrating SQA into HR estimation workflows.