The objective of this paper is to investigate if signal analysis and machine learning can be used to develop an accurate sleepiness warning system. The developed system was trained using the supposedly most reliable sleepiness indicators available, extracted from electroencephalography, electrocardiography, electrooculography, and driving performance data (steering behavior and lane positioning). Sequential forward floating selection was used to select the most descriptive features, and five different classifiers were tested. A unique data set with 86 drivers, obtained while driving on real roads in real traffic, both in alert and sleep deprived conditions, was used to train and test the classifiers. Subjective ratings using the Karolinska sleepiness scale (KSS) was used to split the data as sufficiently alert (KSS amp;lt;= 6) or sleepy (KSS amp;gt;= 8). KSS = 7 was excluded to get a clearer distinction between the groups. A random forest classifier was found to be the most robust classifier with an accuracy of 94.1% (sensitivity 86.5%, specificity 95.7%). The results further showed the importance of personalizing a sleepiness detection system. When testing the classifier on data from a person that it had not been trained on, the sensitivity dropped to 41.4%. One way to improve the sensitivity was to add a biomathematical model of sleepiness amongst the features, which increased the sensitivity to 66.2% for participant-independent classification. Future works include taking contextual features into account, using classifiers that takes full advantage of sequential data, and to develop models that adapt to individual drivers.