Abstract / Summary
Approximately 30% of U.S. adolescents have consumed alcohol, and earlier onset is associated with faster escalation and greater risk of adult alcohol use disorder. Predicting not only who will initiate drinking but also when could enable more timely and targeted prevention. Existing machine-learning approaches have largely classified drinking status or predefined trajectory classes, requiring fully observed outcomes and often excluding adolescents who remain abstinent at last follow-up. Many also rely on neuroimaging or small, hand-selected feature sets, limiting scalability. We developed a stacked-encoder survival framework for predicting drinking-onset timing from baseline clinical data alone. The framework compresses 1,582 clinical features using an unsupervised denoising autoencoder and re-embeds them with FocalTab, a TabPFN encoder fine-tuned with focal loss; the encoders are then frozen for survival modeling using a neural Cox or random survival forest head. Among 661 NCANDA adolescents who were non-drinkers at baseline, 495 initiated drinking within six years. The stacked representation outperformed single-encoder and raw-feature models, achieving an Uno C-index of 0.7654 and time-dependent AUC of 0.8274. Focal-loss encoding outperformed cross-entropy encoding, while SHAP identified alcohol expectancies, substance access, and community prevention engagement among the most influential predictors. Unsupervised clustering further identified five biotypes with a monotonic gradient in mean time to onset (4.30 to 2.86 years) and differential limbic and thalamic neuroimaging phenotypes. These findings support the feasibility of scalable, clinically based prediction of drinking-onset timing and data-driven adolescent risk stratification.