Affective computing is an interdisciplinary field that aims to automatically recognize and interpret emotions. Recent research has focused on using physiological signals (e.g., electrodermal activity) to improve emotion recognition. However, little attention has been paid to the theoretical emotion models underlying these systems. Here, we conducted a systematic review and meta-analysis of the literature on automatic emotion recognition systems using electrodermal activity. We found that models predicting arousal generally outperformed those predicting valence, which is consistent with our pre-registered hypothesis. This finding aligns well with the conceptual framework that views arousal as a psychological and physiological state linked to autonomic nervous system activity, making it more directly related to electrodermal activity. Furthermore, we observed a discrepancy between the types of machine learning models used, mainly classification models, and the emotional models adopted, often of a dimensional nature. Specifically, despite the increased use of dimensional affective models, there has been no corresponding increase in the use of regression models, which would be consistent with the continuous nature of these data. We conclude that a comprehensive understanding of affective states requires consideration of both psychological and computational perspectives in affective computing research.