Image-based deep learning for early detection of adolescent emotional and self-harm risks
Abstract
The early detection of emotional disorders and self-harm risks among adolescents is of critical importance, aligning closely with the emerging needs highlighted by contemporary research on adolescent mental health interventions. Traditional methods, relying heavily on self-reported questionnaires and manual clinical interviews, often suffer from limitations including subjective biases, delayed detection, and inadequate scalability, which are insufficient to address the urgent demand for proactive and large-scale screening solutions in crisis intervention frameworks. In response to these challenges, we propose a novel image-based deep learning methodology rooted in a systematic Medical AI strategy, integrating a foundation model termed MedForm and a strategic diagnostic framework called DiagSolve. MedForm captures complex multimodal clinical patterns through hierarchical attention mechanisms and robust representation fusion, while DiagSolve enables uncertainty-aware, context-sensitive decision-making, ensuring both diagnostic precision and operational resilience under real-world constraints. Experimental evaluations demonstrate that our framework significantly enhances early identification performance for adolescent emotional and self-harm risks, surpassing conventional methods in terms of sensitivity, interpretability, and robustness against missing or noisy inputs. By rigorously addressing the stochastic and partially observable nature of clinical mental health data, our approach offers a scalable and reliable pathway toward effective adolescent emotional crisis intervention, resonating strongly with the priorities of advancing real-world adolescent mental health care.