Acute exacerbations are important episodes in the course of chronic obstructive pulmonary disease (COPD) which is associated with a significant increase in mortality, hospitalization and impaired quality of life. An important treatment for COPD is home telehealth-monitoring intervention. Physiological signals monitored continuously with home ventilators would help us address disease condition in time. However, the absence of useful early predictors and poor accuracy and sensitivity of algorithms limit the effectiveness of home telemonitoring interventions. In order to find prediction indicators and improve the accuracy from physiological signals, we developed a prediction method to search for indicators connected with acute exacerbations. In this study, we analyzed one-month physiological data (airflow and oxygen saturation signals) of 22 patients with COPD before acute exacerbations happened. In the analysis we employed non-linear analyses and machine learning. We applied Multiscale entropy analysis (MSE) and Detrend fluctuation analysis (DFA) to extract features from airflow. Random forest (RF), linear discriminant analysis (LDA) and support vector machine (SVM) were used to classify the stable state and acute exacerbations of disease. The results showed that LDA had the best average precision of 62% and SVM had the best average recall of 56%. Additionally, according to the analysis of RF, the most predictive features are mean of airflow, results of DFA and MSE in scale 4. RF shows a highest accuracy of 75% in three methods, when LDA illustrates a highest specificity of 42.9%. This study will provide insights in developing COPD home-monitoring system which can prognose the onset of acute exacerbations, thus reducing the need of hospital admissions and improving the life quality of COPD patients.