by Alfredo Gioacchino MariaPio Vecchio, Nicole Dalia Cilia, Giuseppe Maniaci, Adriano Schimmenti, Antonino Costanzo

Abstract

Objective: Machine-learning studies on gambling have mostly relied on behavioral or account-based data from online gambling platforms. This study examined whether clinical-demographic and psychometric variables could predict past-year gambling involvement and gambling-related problem severity in psychiatric outpatients, and whether predictability varied across target formulations.
Method: The analytic sample included 340 psychiatric outpatients; 174 reported gambling in the previous 12 months. Predictors included sociodemographic and clinical variables, psychiatric indicators, recent substance and alcohol use, emotion dysregulation, and impulsivity. Five supervised machine-learning formulations were tested: binary past-year gambling involvement, a three-class involvement/problem-risk stratification in the full cohort, four-class severity classification among past-year gamblers, continuous severity regression among gamblers, and full-cohort continuous severity regression with and without the past-year gambling indicator. Up to 17 classification and 15 regression estimators were compared across ten repeated stratified 80/20 train/test splits, with preprocessing fitted on training data only.
Results: Predictive performance varied by target formulation. As expected, binary past-year gambling involvement showed the highest predictive performance, with the best model reaching macro F1 = 0.654 and ROC AUC = 0.707. The three-class formulation showed lower performance, macro F1 = 0.465, and the four-class severity task among gamblers was limited, macro F1 = 0.340. Continuous severity prediction among gamblers was weak across estimators. In the full-cohort regression, retaining the past-year gambling indicator improved rank-based performance, whereas removing it substantially reduced the signal.
Conclusions: In psychiatric outpatients, clinical-demographic and psychometric variables appeared more informative for identifying past-year gambling involvement than for estimating fine-grained gambling-related problem severity. Clinically, this pattern supports a stepped assessment approach: asking about gambling directly in routine psychiatric assessment and, when gambling is present, evaluating gambling-related consequences, impaired control, and functional impact rather than inferring severity from broader clinical features.

Key words: computational psychiatry, emotion dysregulation, gambling, impulsivity, machine learning