Objectives Passive sensing of cellphone data has been proposed as a promising yet unproven approach for detecting escalations in suicidal thoughts and behaviors (STB). This study investigated the predictive power of cellphone usage data for STB in high-risk adolescents, comparing it to clinical measures and intensive longitudinal self-reports. Methods High-risk adolescents with recent STBs were monitored for six months at an outpatient clinic. Data collection included continuous passive cellphone usage tracking and weekly digital self-reports on STB and prominent risk factors. In-clinic visits included standard clinical assessments of STB risk. We applied machine learning and deep learning models to prospectively predict high-risk STB assessed at clinical visits (i.e., suicidal intent, plan, or behaviors) using three data sources: the previous week's passive cellphone data, the last week's digital self-report, and prior clinical assessment. Results The pilot sample included 33 participants with 176 clinical visits, split into train and test samples. Response and retention rates over six months were high (72% and 81%, respectively). Deep learning models using cellphone data outperformed self-report assessments and standard clinical assessment in predicting high-risk STB class (Overall accuracy=.92,.78,.90, Precision=.64,.68,.67, Recall=1.00,.79,.57, F1=.78,.71.,62 respectively). Conclusions Our results highlight the potential clinical value of passive cellphone monitoring in accurately identifying imminent STB in high-risk adolescents. This method shows promise beyond traditional approaches and intensive self-reports, offering scalability and minimal burden to patients and providers. Digital monitoring could enhance clinical management and follow-up for at-risk youth over extended periods.