Depressive episode is key symptom collection of mood disorders. Early intervention can prevent it from happening or reduce its impact, and close monitoring can greatly improve medical management. However, most current monitoring methods are ex post facto, coarse in time granularity and resource consuming. In this study, we aimed to develop a cost-friendly and high usability depressive episode detection framework. In Phase I, we fitted instantaneous affective state models by using R-R intervals collected with photoplethysmogram sensors in smartwatches from laboratory experiments of 1107 participants. In Phase II we utilized the models from Phase I to record long-term affective experience of 2192 participants. Depressive episode models were fitted with affective experience time series. The best instantaneous affective states models achieved overall accuracies of 91% with 2 classes (neutral/ aroused) and 82% with 3 classes (joy/ neutral/ sadness), and the depressive episode models (less severe/ more severe) achieved an overall accuracy of 76% and a best accuracy of 88%. We investigated and discussed the performance differences of the models with multiple settings. We found person-based feature normalization is effective in improving model performance for subjective affect experience. We also found identification of diurnal mood variation may be critical in depressive episode detection.