Machine Learning Reveals Hidden Craving Profiles in Adolescent-Onset Synthetic Drug Addiction

Craving has long been treated by clinicians and researchers as a single, uniform force—the magnetic pull that draws people back to drugs after periods of abstinence. A new multicenter study from Tianjin, China, challenges that assumption, showing that among people whose substance problems began in adolescence with new psychoactive substances, craving is not one phenomenon but several distinct ones. By combining a statistical technique called latent profile analysis with machine learning, a research team led by Shumei Zhuang and Xiaojuan Che of Tianjin Medical University has mapped three fundamentally different craving profiles among more than a thousand patients, and identified the psychosocial forces that separate them. The work, published in the International Journal of Mental Health and Addiction, offers an early blueprint for a more personalized form of addiction care.

The scale of the problem motivating the study is considerable. New psychoactive substances—a shifting family of synthetic compounds designed to mimic the effects of controlled drugs while evading legal restrictions—have become one of the most difficult fronts in global drug policy. Because their chemical structures change constantly, their health consequences are poorly characterized, and people who begin using them in adolescence face heightened risks of lasting neurobiological and psychological harm. The adolescent brain, with its early-maturing reward circuitry and still-developing self-control systems, is particularly vulnerable to substance initiation and to the entrenched patterns of craving that follow.

To uncover the hidden structure of craving, the researchers assessed 1,008 patients in Tianjin whose substance use disorder began during adolescence. Rather than reducing craving to a single score, the team measured three separate dimensions: the intensity of craving itself, the degree to which it interfered with daily functioning, and the effort patients expended resisting it. Latent profile analysis—a form of statistical modeling that sorts individuals into unobserved subgroups based on patterns across multiple measures—was then applied to see whether distinct combinations of these dimensions existed within the sample.

Three profiles emerged, and they were strikingly uneven in size. The largest group, accounting for 54 percent of participants, was labeled relatively stable, showing moderate craving without overwhelming functional disruption. A second group, 40.2 percent of the sample, fell into a high craving struggle category marked by intense urges and an exhausting internal battle to resist them. The smallest but most concerning group, 6 percent, was described as functionally impaired, in which craving severely disrupted everyday life. That two in five patients occupied the high struggle profile underscores how common severe craving is in this population—and how misleading it would be to treat them all identically.

Identifying the profiles was only the first step. The team then asked whether psychosocial information could predict which profile a patient belonged to, using three machine learning models of increasing sophistication: logistic regression, a classical statistical baseline; random forest, an ensemble method that averages many decision trees; and XGBoost, a gradient-boosted algorithm known for its performance on tabular medical data. Performance was evaluated using macro-AUC, accuracy, and Macro-F1 scores, metrics that account for the imbalanced group sizes. XGBoost achieved the strongest discrimination among the three, although the researchers are candid that overall classification performance remained modest—a limitation they attribute to the exploratory nature of the work and the complexity of the underlying constructs.

Perhaps the most clinically valuable output came from interpreting the models. Using SHAP, an explainability technique that quantifies each feature’s contribution to predictions, the researchers identified the psychosocial variables that most strongly distinguished the craving subgroups. Social support emerged as a leading factor, consistent with a growing body of evidence that supportive networks buffer against relapse. Resistance to peer influence also ranked highly—a finding that resonates with developmental research showing adolescence as the period of peak susceptibility to peer pressure. Coping style, childhood adversity, age, and age at first drug use completed the list of key contributors, weaving together life history, developmental timing, and present-day resources into a single predictive picture.

These findings align closely with the biopsychosocial model of addiction, which holds that substance use disorders arise from interactions among biological vulnerability, psychological processes, and social context rather than from any single cause. The prominence of childhood adversity in the models echoes systematic reviews linking early maltreatment to elevated risk of substance misuse, while the role of social support reflects studies demonstrating that perceived support moderates the relationship between stress and relapse. By embedding these factors in a quantitative, subgroup-aware framework, the study gives that conceptual model a practical, data-driven form.

The authors are careful to frame the work as exploratory rather than definitive. The cross-sectional design captures a single moment in time and cannot establish causal direction—for instance, whether low social support drives intense craving or vice versa. The modest discriminative performance means the models are not yet ready for individual-level clinical decisions, and the data, which contain sensitive personal information, are not publicly available for reasons of privacy protection. Still, the researchers argue that the approach may provide preliminary evidence for risk stratification and for subgroup-sensitive addiction nursing support, in which care is tailored to whether a patient fits the stable, high struggle, or functionally impaired pattern.

If future longitudinal studies confirm and refine these profiles, the implications could be far-reaching. A patient in the functionally impaired group might warrant immediate intensive intervention, while someone in the high craving struggle group might benefit most from coping-skills training and peer-resistance support, and those in the stable group from relapse prevention and social strengthening. In an era when new psychoactive substances continually outpace regulation and traditional treatment models, the ability to sort patients by their actual craving signature—rather than a one-size-fits-all label—could mark a meaningful shift toward precision medicine in addiction care. The study, funded by the Ministry of Education in China, is an early but provocative demonstration that the algorithms of machine learning can find order in one of addiction’s most stubborn and subjective experiences.

Subject of Research: Exploratory risk stratification of adolescent-onset new psychoactive substance use disorder using latent craving profiles and machine learning

Article Title: Exploratory Risk Stratification of Patients With Adolescent-Onset New Psychoactive Substance Use Disorder: Integrating Latent Craving Profiles, Psychosocial Factors, and Machine Learning

Article References: Zhuang, S., Che, X., Song, Y., Li, S., Shang, X., Shi, L., Zhang, X., Song, C., Jing, J., Fan, J., Qin, S., Xu, Z., Zhang, J., Li, J., Li, L., Hou, K., & Jiang, Y. (2026). Exploratory Risk Stratification of Patients With Adolescent-Onset New Psychoactive Substance Use Disorder: Integrating Latent Craving Profiles, Psychosocial Factors, and Machine Learning. International Journal of Mental Health and Addiction. https://doi.org/10.1007/s11469-026-01714-3

Image Credits: AI Generated

DOI: 10.1007/s11469-026-01714-3

Keywords: new psychoactive substances, adolescent-onset substance use disorder, craving, latent profile analysis, machine learning, XGBoost, SHAP, social support, childhood adversity, risk stratification, addiction nursing, peer influence