In the last decades numerous researches have revealed a strong link between emotions and several physiological responses. However, the automatic recognition of emotions still remains a challenge. In this work we describe a novel approach to estimate valence, arousal and dominance values from various biological parameters (derived from electrodermal activity, heart rate variability signal and electroencephalography), by means of multiple linear regression models. The models training was performed by using a set of pictures pre-evaluated in terms of valence, arousal and dominance, selected from the International Affective Picture System (IAPS) database. By using the step-wise regression method, all the possible combinations of considered biological parameters were tested as input variables for the models. The three multiple linear regression models that could provide the best fit for IAPS pictures valence, arousal and dominance values were selected. The features included in the optimal models were the average of the inter-beat duration (mean RR), the EEG spectral power computed in alpha, beta and theta frequency bands (Alpha, Beta and Theta power) and the average value of EDA signal (mean EDA). The obtained models show an overall good performance in predicting valence, arousal and dominance values.