This study describes a theory-informed application of data science methods to analyze the quality of reflections made in a health professions education program over time. One thousand five hundred reflections written by a cohort of 369 dental students over 4 years of academic study were evaluated for an overall measure of reflection depth (No, Shallow, Deep) and the presence of six theoretically-indicated elements of reflection quality (Description, Analysis, Feeling, Perspective, Evaluation, Outcome). Machine learning models were then built to automatically detect these qualities based on linguistic features in the reflections. Results showed a dramatic increase from No to Shallow reflections from the start to end of year one (20% [right arrow] 66%), but only a limited gradual rise in Deep reflections across all four years (2% [right arrow] 26%). The presence of all six reflection elements increased over time, but inclusion of Feelings and Analysis remained relatively low even at the end of year four (found in 44% and 60% of reflections respectively). Models were able to reliably detect the presence of Description ([kappa]TEST = 0.70) and Evaluation ([kappa]TEST = 0.65) in reflections; models to detect the presence of Analysis ([kappa]TEST = 0.50), Feelings ([kappa]TEST = 0.54), and Perspectives ([kappa]TEST = 0.53) showed moderate performance; the model to detect Outcomes suffered from overfitting ([kappa]TRAIN = 0.90, [kappa]TEST = 0.53). A classifier for overall depth built on the reflection elements showed moderate performance across all time periods ([kappa]TEST > 0.60) but relied almost exclusively on the presence of Description. Implications for the conceptualization of reflection quality and providing personalized learning support to help students develop reflective skills are discussed.