Introduction The advent of small and portable devices, with high storage and processing capabilities have allowed physiological data to be acquired outside clinical environments in a reliable way, often across several days. These sources of data include ECG, Respiratory Inductance Plethysmography (RIP), oxygen saturation, Actigraphy (ACT), among others. Due to the highly constrained conditions in which the Polysomnography (PSG), the golden standard for sleep disorders diagnosis, is performed, alternative methods are desired. These small and portable devices may provide an alternative for long term monitoring, thus complementing the information provided by the PSG. In this work, we present an algorithm to estimate the Sleep Efficiency (SE), REM and Non-REM sleep percentages from multi-modal data that can easily be acquired from portable devices, combining information from ECG, RIP and ACT. Materials and methods The data was acquired from fifteen healthy volunteers who performed a standard PSG+ Actigraphy. The average SE, REM and NREM percentages, computed from the hypnograms, are 87.2%, 18.87% and 81.13% respectively. An extended feature set was extracted from the RR, RIP and ACT signals. The extracted features were used to train a two Parzen classifiers with a rejection option. The computation of the 3 sleep parameters takes into account the result of each classifier and a final regularization step, which takes includes the a priori training information regarding the performance of the classifier. Results The described method resulted in an estimated average SE of 87.8% corresponding to an average estimation error of 5.62%. The estimated REM and NREM percentages were 18.91% and 81.9%, corresponding to an average estimation error of 8.22% and 5.95% respectively. Conclusion Reliable sleep parameters estimation is typically limited to PSG data, where EEG plays a fundamental role in the determination of the sleep state. The obtained results are encouraging since they were obtained with limited data yielding low estimation errors. The use of classifiers with rejection capability ensures that only non-ambiguous data is used in the estimation process. Acknowledgements This work was supported by FCT (ISR/IST plurianual funding) through the PIDDAC Program funds and FCT project ”Detection of Brain Micro-states in Fibromyalgia” (PTDC/SAU-BEB/104948/2008).