Towards an Explainable Longitudinal Representation Learning for Organizing Trajectories of Adolescent Mental Health Onset

Mental health conditions emerging in childhood and adolescence can significantly impact well-being across the lifespan. Understanding their developmental trajectories is critical for early diagnosis, intervention, and treatment. However, mental disorders are highly heterogeneous and often co-occur, necessitating longitudinal predictive models that not only identify conditions but also capture shared mental health trajectories. Existing longitudinal representation approaches mostly prioritize predictive accuracy at the expense of interpretability. In this work, we develop and compare two longitudinal and interpretable modeling approaches, a novel Multi-wave Integration for Multi-domain Encoding (MIME) and an adaptation of Agglomerative Clustering for longitudinal data, to derive representations that balance predictive utility and interpretability from longitudinal mental health data. On the predictive side, both methods outperform baselines at forecasting development of depression, anxiety, ADHD, and substance / alcohol use disorders. Furthermore, we found our proposed two approaches to be interpretable based on human judgments of coherence over an intrusion task with random clustering as a baseline. By bridging predictive utility and interpretability, this work provides a foundation for more clinically meaningful longitudinal modeling of mental health conditions.