As ambient systems proliferate, there is an increased need for data-level fusion methods that reflect the challenges of unknown environments, non-deterministic signal behaviours, and movement artifact. Cushioning between the sensor and bed occupant decreases signal to noise power and signal availability, while non-linear signals can masquerade as delayed and reversed signals. The main contributions of this thesis are to study how these challenges affect extraction of a breathing signal from bed-based sensors and to propose more robust fusion techniques.New trend analysis methods effectively corrected polarity reversals, increasing the number of good quality signals by 9% and reducing mean respiratory rate error by 24%. To fuse these signals, selection combining, weighted summation, and blind source separation methods were innovated and compared. None performed best all of the time; some were generally good with some weaknesses, while others had specialized strengths. Contextual ensemble fusion selected the best fusion method in degraded conditions in 55% of records, compared to 36% for the top individual fusion method, providing clinical applications with more reliable data. While sleep medicine is an important application, ambient monitoring is also suited to cognitive medicine and palliative care. Developed methods were applied to monitor patients in palliative care, marking the first long-term, continuous monitoring of this population. Breathing patterns observed in the last weeks of life included Cheyne-Stokes respiration and tachypnea, while breathing variability was associated with survival time.ii