Seismocardiography (SCG) is a promising alternative approach for heart rate variability (HRV) monitoring; however, nocturnal recordings are affected by motion, posture, and changes in signal morphology. This study evaluated a previously developed deep-learning method for nocturnal HRV estimation from 3D SCG. A U-Net-based model trained on short resting recordings from over 5,500 individuals was evaluated on 38 overnight recordings (~291 h) from the publicly available NightbeatDB dataset. Heart rate (HR) and RMSSD were calculated from 10-min segments and compared with time-synchronized ECG references. SCG-derived HR showed excellent agreement with ECG (r = 0.982, bias = -0.32 bpm, 95% LoA: [-1.7; 1.0] bpm). RMSSD also showed strong agreement (r = 0.929, bias = 0.57 ms, 95% LoA: [-10.0; 11.1] ms). These results demonstrate that deep learning-based beat detection enables accurate nocturnal HR and RMSSD estimation from 3D SCG and shows potential for long-term home monitoring.