Suicide is a leading cause of death in the US, yet remains difficult to predict due to its complex, individualized causes—genetic, physiological, lifestyle, and environmental. Recent studies use smartphone-based surveys to capture real-time affects (e.g., happiness, anger) and behaviors (e.g., suicidal ideation), but low compliance makes prediction difficult with limited data. Our goal is to improve “cold start” prediction: forecasting suicide-related events (SREs) for patients with little recorded history. Our first finding is counterintuitive: training standard classifiers (e.g., KNN, Logistic Regression) on more patient data does not improve predictive performance. We explain this by showing that precursors to SREs differ across patients—one patient’s risk may stem from anxiety, another’s from hopelessness—leading to conflicting patterns when data is pooled. This suggests a need to model patient variability. To address this, we propose using Meta-Learning Gaussian Processes (MLGPs), which incorporate a latent variable capturing patient similarity. This allows the model to adapt quickly to new patients by leveraging patterns from similar individuals, improving prediction even with sparse data. On synthetic data mimicking real-world structure, MLGPs outperformed baselines in both positive predictive value and sensitivity. We plan to apply MLGPs to real longitudinal data to aid early identification and intervention for at-risk patients.
