Использование машинного обучения с помощью краткого перечня симптомов для просмотра Comorbid Disabilities: A account of Principle Study.

Introduction

Mental health disorders, including anxiety and mood disorders, rank among the top ten global causes of disease burden [ 1 , 2 ] and are linked to reduced quality of life and broad health, social, and economic harm [ 3 , 4 ]. They often co-occur with substance and behavioral addictions such as gambling and gaming [ 5 – 7 ], which themselves contribute substantially to preventable morbidity, mortality, and socioeconomic costs [ 8 – 10 ]. Individuals with comorbid mental health and addictive disorders have lower quality of life, more complex clinical profiles, and poorer treatment outcomes [ 7 , 11 ]. Comprehensive screening for substance use (SU) and behavioral addictions is therefore crucial to identify those at risk and enable integrated interventions [ 7 , 12 – 14 ]. Yet, due to limited time, awareness, or training, clinicians often fail to screen for addictions [ 13 ], creating treatment gaps and hindering recovery. Efficiently identifying high-risk patients could help address this gap.

The 53-item Brief Symptom Inventory (BSI-53) [ 15 , 16 ] is among the most widely used global measures of psychopathology [ 17 ]. Although it assesses nine domains (e.g., anxiety, depression, hostility), research indicates it is essentially unidimensional and serves as a general indicator of psychological distress [ 17 , 18 ]. The BSI is commonly administered during intake, particularly when a brief assessment is required [ 19 ]. Thus, if the BSI – already routinely used in mental health settings – could also inform about a patient’s risk for substance or behavioral addictions, it would offer a highly efficient screening strategy. Importantly, the BSI does not directly assess SU or other addictive behaviors. Therefore, any predictive value it offers for these outcomes would derive from items indirectly related to such behaviors – highlighting its potential as a tool for identifying individuals at risk.

Beyond mere comorbidity, contemporary frameworks of psychopathology provide a mechanistic basis for expecting general-distress measures to carry information about addiction. The Hierarchical Taxonomy of Psychopathology (HiTOP) and related dimensional approaches show that internalizing distress, externalizing impulsivity, and a broad general-psychopathology dimension share variance with substance and behavioral addictions [ 20 , 21 ]. The BSI explicitly probes negative affectivity, somatic distress, hostility, and impulse-related symptoms, which map onto these transdiagnostic dimensions. Predictive value is therefore expected because the BSI indexes a part of the latent distress and dysregulation that contribute to risk for substance and behavioral addictions.

Some studies have shown associations between the BSI and specific substance [ 22 , 23 ] or behavioral addictions [ 24 ], but its broader predictive capacity remains unclear. Given that substance and behavioral addictions share mechanisms and comorbidity with mental disorders [ 25 , 26 ], the BSI may capture general or addiction-specific risk patterns shaped by legality, norms, or biological activity [ 27 ]. Understanding this could improve identification of high-risk patients and inform tailored treatments.

Machine learning is well suited to this question for three reasons. First, item-level prediction across 53 inter-correlated indicators may benefit from nonlinear and interaction-aware algorithms (decision trees, random forests, boosting) that linear regression cannot capture. Second, regularized and subset-selection methods (LASSO, elastic net, best subset) can identify sparse item combinations predictive of each outcome without prior commitment to the BSI’s nominal subscale structure, particularly important given the essential unidimensionality of the BSI. Third, even simple linear and logistic regression on the GSI serve naturally as machine-learning baselines within this framework, providing direct comparators against more flexible algorithms.

To address this empirical gap, we conducted a cross-sectional study in a general population sample of Israeli adults – a Western population with substance-use prevalence comparable to the USA and Europe. Participants completed the BSI and self-report measures of substance and behavioral addictions. We examined whether the BSI – which does not directly assess SU or addictive behaviors – could predict risk for (1) problematic alcohol and drug use (non-medical use of cannabis, cocaine, amphetamines, inhalants, sedatives/sleeping pills, hallucinogens, street opioids, prescription stimulants, and prescription opioid painkillers) and (2) potential problematic addictive behaviors (compulsive sexual behavior, pornography use, gambling, and online gaming). As a proof of principle, we applied machine learning algorithms to assess the extent to which BSI responses can predict risk for problematic SU and addictive behaviors, optimizing their predictive performance.

Methods

Sample

Data were collected in April 2022 from a cross-sectional general population sample of Jewish, Hebrew speaking (since cultural differences require substantial methodological adaptations [ 28 ]) adults (18–70) in Israel [ 29 ]. Participants were recruited through iPanel [ 30 ], a national online data-collection agency with a diverse panel, enabling quota-based sampling for demographic representativeness. Quotas were set on the joint distribution of age, gender, religiosity, education, and geographic region (Jerusalem area, Tel Aviv and Central area, North, South), with targets derived from the Israeli Central Bureau of Statistics population estimates, and participants were selected either from those surveyed in 2018 who remained registered or randomly from new panel members [ 31 ]. Sampling continued until quotas were met, yielding a quasi-representative sample of Israel’s Jewish adult population. Strict confidentiality was maintained because iPanel lacked access to survey responses, and researchers had no identifying information. Survey methodology follows the ICC/ESOMAR International Code [ 30 ] and was approved by the (Reichman University [Approval No. 2022050_1re_P]) IRB. Written informed consent was obtained electronically from all participants prior to the study via the Qualtrics platform [ 32 ] in accordance with local guidelines, and they subsequently completed an online survey assessing sociodemographics, SU, addictive behaviors, and mental and physical health. Anonymous online surveys often improve reporting of stigmatized behaviors [ 33 ]. Respondents received a 22 ILS gift card. Quality assurance involved participation by invitation only to registered people; three attention checks; and flagging unexpected response patterns. Of 11,750 invited to participate, 4,948 agreed, 1,944 were excluded due to quotas, 505 did not complete the survey (135 failed attention checks, 370 dropped out), and 25 were excluded based on response patterns, for an available sample of 2,474. The forced-response format prevented missing data. Since the BSI measure is calculated separately for men and women, 23 respondents who reported “other” for gender were excluded, leaving 2,451 for analysis.

Measures

Psychopathology

Overall psychopathology was assessed using the 53-item BSI, which measures nine domains: anxiety; depression; hostility; interpersonal sensitivity; obsession compulsion; paranoid ideation; phobic anxiety; psychoticism; and somatization [ 15 , 16 ]. Each item rates distress by a specific problem over the past month on a five-point scale (0 = not at all to 4 = extremely). The BSI shows good internal consistency and validity as a general indicator of psychopathology [ 17 , 18 ]. The Global Severity Index (GSI) is calculated as the mean of all 53 items and standardized to a T score (mean = 50, SD = 10) [ 16 ], using sample norms of each gender (men, mean = 0.52; SD = 0.63; women, mean = 0.62; SD = 0.63). Norms were used from the current sample since previously reported norms in Israel were from over 20 years ago and based on a smaller, less-representative sample [ 34 ]. For the predictive models, the per-gender mean and standard deviation used to construct the GSI T score were fitted on the training set only and applied unchanged to the held-out test set, so that no information from the test data informed the standardization.

Problematic SU

The Alcohol, Smoking and Substance Involvement Screening Test (ASSIST 3.1) was administered and scored per WHO guidelines [ 29 , 35 ]. Respondents reported lifetime nonmedical use of alcohol, cannabis, cocaine, amphetamines, inhalants, sedatives, hallucinogens, street opioids, prescription opioid painkillers, prescription stimulants, and other substances. For each substance, six items assessed frequency of current use, craving, socioeconomic or health problems, or failure to meet expectations due to use and lifetime, if anyone expressed concern about their use and trouble controlling use. Weighted responses produced a substance involvement score [ 35 ]; nonusers scored 0. The any-drug score was defined as the maximum across all drug classes. Binary variables indicated problematic use at moderate/high risk (score ≥4 for all substances; ≥11 for alcohol), cutoffs validated as approximating abuse and dependence [ 35 ]. An additional binary variable represented problematic use of any drug.

Gambling

Problematic gambling in the past 12 months was assessed using the Problem Gambling Severity Index (PGSI) [ 36 ], which includes nine items rated from 0 = never to 3 = almost always. The summed score (0–27) indicated moderate risk/problem gambling at ≥5, consistent with established validity criteria [ 37 , 38 ].

Gaming

Problematic gaming in the past 6 months was assessed using the Game Addiction Scale (GAS) [ 39 ], an established measure with good psychometric properties, which contains seven items rated 1 = never to 5 = very often. Scores were summed (7–35). Binary classification as problematic gaming was assigned when “often” or higher was endorsed on ≥4 items, a threshold supported by previous research [ 39 – 41 ]. Non-players received the minimum score (7) and were coded “no” for the binary variable.

Sex

Problematic compulsive sexual behavior was assessed using the Hypersexual Behavior Inventory (HBI), which showed good psychometric properties [ 42 , 43 ]. The HBI includes 19 items rated 1 = never to 5 = very often (range = 19–95). Scores ≥53 indicated clinically significant hypersexual behavior [ 43 ].

Pornography

Problematic pornography use within the past year was assessed using the Problematic Pornography Use Scale (PPUS), which showed good reliability and validity [ 44 ], containing 12 items rated 0 = never true to 5 = almost always true (range = 0–60). Participants endorsing “often true” or higher on ≥7 items were classified as problematic users, similar to the definition of binary GAS [ 40 , 41 ].

Data Analysis

Descriptive statistics were computed for sociodemographic variables and binary indicators of problematic SU and behavioral addictions. Before main analyses, we verified the unidimensionality of the 53 BSI items (online suppl. Table S1; for all online suppl. material, see https://doi.org/10.1159/000553511 ). A scree plot of eigenvalues was inspected, and confirmatory factor analyses were performed for two models previously identified as optimal [ 17 ]: (1) one-factor model; (2) bifactor model, with an underlying general factor and 9 uncorrelated domains.

Confirmatory factor analyses used variance-adjusted weighted least squares estimators, appropriate for ordinal variables. Model fit was evaluated via CFI and TLI ≥0.95 and root-mean-square error approximation ≤0.06 [ 45 ]. Items loadings <0.40 were considered weak indicators. For the bifactor model, we computed omega hierarchical (proportion of total variance explained by the general factor) and explained common variance (variance explained by the general factor within each subdomain) [ 46 ]. Analyses were conducted in R using lavaan , semTools , and BifactorIndicesCalculator .

To examine whether GSI scores differed between individuals with and without each addiction, t tests assuming unequal variances were conducted. For each outcome, we estimated two parallel sets of predictive models. The continuous-regression track predicted the raw severity score (alcohol total, drug maximum, GAS, HBI, PPUS, PGSI). The binary-classification track predicted the clinical-threshold indicator. Both tracks were fitted with the same six algorithms ( Table 1 ), implemented via the tidymodels [ 47 ] framework in R. Three nonlinear methods: decision trees, random forests, and boosting captured potential nonlinear relationships and interactions among items [ 48 ]. Decision trees provide interpretable rule-based predictions, random forests improve accuracy by averaging across multiple resampled trees, and boosting enhances precision by sequentially correcting prior misclassifications [ 48 ] Three linear regularization methods, LASSO regression, elastic net regression, and subset selection, were implemented and performed item-level selection [ 49 , 50 ]. LASSO uses L1 regularization to perform feature selection by shrinking some coefficients to zero; the elastic net combines L1 and L2 penalties to address correlated predictors; and subset selection identifies the smallest set of BSI items that best predicts each outcome according to the Bayesian information criterion (online suppl. Table S2). Baseline comparators used only the GSI T score as predictor: a linear regression on the continuous track and a logistic regression on the binary track. The full pipeline is summarized in online supplementary Figure S1.

Table 1. Machine learning models utilized to predict problematic SU and addictive behaviors from the BSI Model Description Interpretability R package All optimized hyperparameters Administration options
Nonlinear models
1. Decision trees A decision tree is a flowchart-like model used for both classification and regression tasks. It splits the data into subsets based on feature values, making decisions at each level to maximize information gain or minimize impurity, leading to a tree-like model of decisions. In this study, a decision tree algorithm was employed to predict each addiction score using a series of yes-no decision rules High Rpart The complexity parameter (cp) controls when the tree stops growing, optimized over a log-scale range from 10 −4 to 10 −1 (8 values). Tree depth specifies the maximum number of internal splits, tuned from 3 to 15 levels (5 values). The minimum number of observations required in a terminal node was optimized from 2 to 40 (4 values) Digitally, pen and paper 2. RF RF is an ensemble learning method that constructs multiple decision trees during training. The model enhances accuracy and controls overfitting by introducing randomness in both data sampling (bootstrap samples) and feature selection for each tree split Low Ranger The number of predictors considered at each split (mtry) was optimized from 1 to all 53 BSI items (8 values). The minimum number of observations in a terminal node was tuned from 2 to 40 (5 values). The number of decision trees in the forest was fixed at 200 Digitally 3. Boosting Boosting is a machine learning ensemble technique that aims to convert weak learners into strong learners. It does this by iteratively training decision trees, on the residuals or misclassified data points from the previous model, thereby correcting the errors of its predecessors Low xgboost The learning rate (shrinkage parameter) controls the contribution of each sequential tree and was optimized over a log-scale range from 0.001 to 0.1 (5 values). Tree depth specifies how deep each individual boosted tree can grow, tuned from 1 to 6 levels (4 values). The minimum number of observations in a terminal node was optimized from 2 to 20 (3 values). The number of sequential boosting rounds was fixed at 500 Digitally
Linear models
4. LASSO This model uses L1 (LASSO) regularization, which tends to force some regression coefficient estimates to be exactly zero, effectively performing variable selection and producing a simpler model with fewer predictors Medium Glmnet The regularization penalty (λ, lambda) controls the strength of L1 shrinkage and was optimized over a log-scale range from 10 −4 to 10 (40 values). The mixing parameter (mixture) was fixed at 1.0 to apply pure L1 regularization, which forces some coefficient estimates to exactly zero, performing automatic variable selection Digitally 5. Elastic net This model combines features of both L1 (LASSO) and L2 (ridge) regularization regression. This shrinks the coefficients of less important features, without necessarily forcing them to be exactly zero. It is especially useful when there are correlated predictors as it balances between coefficient shrinkage and feature selection Medium Glmnet The regularization penalty (λ) was optimized over a log-scale range from 10 −4 to 10 (30 values). The mixing parameter (α, mixture) balances between L1 (LASSO) and L2 (ridge) regularization and was tuned from 0 (pure ridge) to 1 (pure LASSO) in 11 steps Digitally 6. Best subset selection This technique systematically explored all possible combinations of predictor variables derived from the BSI items to identify the subset that yielded the best predictive performance for each addiction type. The BIC was chosen as the evaluation criterion to assess the model’s goodness of fit and complexity Medium Leaps This technique systematically evaluates subsets of predictor variables using the BIC to balance model fit and complexity. For continuous outcomes, an exhaustive search examines all possible combinations of the 53 BSI items to identify the subset with the best BIC score Digitally BIC, Bayesian information criterion.

Machine learning models utilized to predict problematic SU and addictive behaviors from the BSI

BIC, Bayesian information criterion.

Predictor sets were specified as follows. The baseline models used the GSI T score as a single predictor. All six machine-learning algorithms used the 53 individual raw BSI item scores as predictors, without demographic or socioeconomic variables, to isolate the instrument’s inherent predictive utility. The baseline and machine-learning models therefore use different predictor sets by design; the comparison is between the unidimensional summary captured by the GSI and the item-level information available in the full BSI.

Model development used a single 80%–20% train/test split ( n = 1,960 training, n = 491 test), stratified on a composite “any addiction” indicator (defined as meeting the clinical threshold for at least one of the six assessed outcomes: problematic alcohol use, problematic drug use, problematic gambling, problematic gaming, compulsive sexual behavior, or problematic pornography use). Hyperparameter tuning used repeated 10-fold cross-validation with 5 repeats within the training set, with the same fold structure shared across algorithms and outcomes to support like-for-like comparison. The continuous track tuned on root-mean-square error; the binary track tuned on recall (sensitivity), chosen because in addiction screening missing a true positive is the operationally more costly error and recall is the loss most directly tied to that cost in the imbalanced setting. The best configuration per algorithm was then refit on the full training set and evaluated once on the held-out test set on the full battery of discrimination metrics (AUPRC, AUC-ROC, precision, recall). Selected hyperparameters are reported in online supplementary Table S3. To verify that test-set performance was not driven by a single random split, we also report the cross-validation out-of-sample estimates alongside the test-set values in online supplementary Table S4; the two were closely consistent across cells.

Class imbalance on the binary track was addressed with synthetic minority over-sampling ([ 51 , 52 ]), applied within each training fold, targeting a moderate minority share (∼30%) rather than full 1:1 balancing, to avoid the over-synthesis that can degrade test-set performance at very low prevalence. For transparency and robustness, we also fitted versions without resampling and with inverse-prevalence class weights; the three are compared in online supplementary Table S5. Binary-track performance was evaluated using AUPRC, area under the receiver-operating-characteristic curve (AUC-ROC), precision, and recall (sensitivity). Operating-point metrics (precision, recall) are reported at the default 0.5 classification threshold, noting that this threshold is not necessarily clinically optimal; threshold tuning was deliberately left to future deployment work. Confidence intervals for all test-set metrics were derived via 1,000-replicate nonparametric bootstrap. Future research may examine whether adding demographic or contextual variables enhances predictive accuracy or generalizability across settings.

Since the study was conceived after data collection, formal pre-registration was not possible. We addressed this by reporting all algorithm and outcome combinations transparently, using a fixed held-out test set evaluated only once, and reporting bootstrap-based confidence intervals on all metrics. Findings are framed as exploratory and as a proof of principle.

Continuous-track performance was summarized as R 2 (the squared correlation between predicted and observed values), representing the proportion of variance explained by the BSI items. Confidence intervals for all performance metrics were derived via 1,000-replicate nonparametric bootstrap on the test set. The full pipeline used the tidymodels package suite (rsample, recipes, parsnip, workflows, tune, yardstick, themis) in R.

Results

Sample Descriptive and Problematic SU and Addictive Behaviors Prevalences

Of the sample, about half were women and had some college or more; 42% were ages 18–34; 57% were married; 72% had any work; 23% had problematic drug or alcohol use while 12% had at least one addictive behavior. Among behavioral addictions, gambling was reported by 5% of participants, compulsive sexual behavior by 6%, problematic pornography use by 3%, and problematic internet gaming by 2% ( Table 2 ).

Table 2. Sample descriptives for dataset used to predict problematic SU and addictive behaviors from the BSI ( N = 2,451) ​
N
Prevalence, % Sociodemographics  Gender   Men 1,233 50.3   Women 1,218 49.7  Age   18–25 years 448 18.3   26–34 years 586 23.9   35–49 years 702 28.6   50–70 years 715 29.2  Educational attainment   High school, non-matriculation 192 7.8   High school, matriculation 496 20.2   Post-high school, technological 527 21.5   Post-high school, academic 1,236 50.4  Marital status   Married 1,408 57.4   Not married 1,043 42.6  Employment status   Any job 1,766 72.1   No job 685 27.9  Religiosit y   Secular 1,276 52.1   Traditional 759 31.0   Religious 344 14.0   Ultra-Orthodox 72 2.9  Residency   Jerusalem area 274 11.2   Tel Aviv and Central area 1,001 40.8   North 639 26.1   South 537 21.9  Ethnicity   Middle Eastern/North African 961 39.2   European, not Former Soviet Union 793 32.4   Former Soviet Union 235 9.6   Mixed 401 16.4   Other 61 2.5 ASSIST 3.1 problematic use a  Alcohol 323 13.2  Cannabis 266 10.9  Sedatives 96 3.9  Prescription stimulants 94 3.8  Prescription opioid painkillers 65 2.7  Other drugs b 50 2.0  Any drug 351 14.3  Any drug or alcohol 567 23.1 Addictive behaviors  Gambling c 130 5.3  Internet gaming d 59 2.4  Sex e 148 6.0  Pornography d 75 3.1  Any addictive behavior 289 11.8
a ASSIST = Alcohol, Smoking and Substance Involvement Screening Test; score 4+ (11 + for alcohol), indicating moderate or high risk levels.
b Cocaine, amphetamines, hallucinogens, inhalants, or street opioids.
c Moderate or severe on the PGSI scale.
d Often or very often on more than half of items for relevant measures: GAS for gaming; PPUS for pornography.
e Score of 53 or more on the HB.

Sample descriptives for dataset used to predict problematic SU and addictive behaviors from the BSI ( N = 2,451)

a ASSIST = Alcohol, Smoking and Substance Involvement Screening Test; score 4+ (11 + for alcohol), indicating moderate or high risk levels.

b Cocaine, amphetamines, hallucinogens, inhalants, or street opioids.

c Moderate or severe on the PGSI scale.

d Often or very often on more than half of items for relevant measures: GAS for gaming; PPUS for pornography.

e Score of 53 or more on the HB.

Unidimensionality of BSI Items

Factor analysis of the 53 BSI items supported essential unidimensionality. The first factor (eigenvalue = 30.98) explained far more variance than the second (2.38), and the scree plot “elbow” occurred at the second factor, suggesting that a majority of the variance was explained by one factor (online suppl. Fig. S2). The data fit well to the single factor model (CFI = 0.991; TLI = 0.990; RMSEA = 0.064) and the bi-factor model (CFI and TLI = 0.990; RMSEA = 0.051). In the bi-factor model, all 53 items showed strong loadings (≥0.40) on the underlying general factor, while each subdomain (except somatization) had few (0–2) items with strong loadings. Indices indicated that almost all of the total variance was explained by the general factor (omega hierarchical = 0.966), while for each subdomain, between 70% and 93% of the common variance was explained by the general factor (online suppl. Table S1).

GSI Scores by Addiction

Across all outcomes, individuals meeting the criteria for problematic SU or behavioral addiction had significantly higher Global Severity Index (GSI) T scores than those who did not ( Table 3 , all p < 0.0001). The pattern was most pronounced for behavioral addictions. For problematic gaming, the mean GSI was 66.4 (SD = 16.4) versus 49.6 (SD = 9.4) in the comparison group, a difference of 16.7 T points (95% CI: 12.4, 21.0; t (58.95) = 7.81). For gambling and compulsive sexual/pornography use, mean differences were 12.5 (95% CI: 9.8, 15.1) and 12.0 (95% CI: 9.8, 14.2) T points, respectively. Differences for substance outcomes were smaller but still substantial: 6.7 (95% CI: 5.3, 8.1) T points for alcohol and 6.6 (95% CI: 5.3, 8.0) T points for drugs. The composite “any behavioral addiction” indicator showed an 11.4-point GSI difference (95% CI: 9.7, 13.0) and “any drug or alcohol” a 6.8-point difference (95% CI: 5.7, 7.9) (see Table 3 for full statistics).

Table 3. Associations between the GSI (BSI) score and problematic SU and addictive behaviors Measurement Mean (SD) transformed GSI score Mean difference (95% CI)
t test
p value with addiction without addiction
t
df Problematic SU  Alcohol 55.8 (12.5) 49.2 (9.3) 6.7 (5.3, 8.1) 9.20 377.22 <0.0001  Drug 55.7 (12.3) 49.1 (9.2) 6.6 (5.3, 8.0) 9.70 418.89 <0.0001  Alcohol or drugs 55.2 (12.3) 48.5 (8.6) 6.8 (5.7, 7.9) 12.25 739.46 <0.0001 Addictive behaviors  Gaming 66.4 (16.4) 49.6 (9.4) 16.7 (12.4, 21.0) 7.81 58.95 <0.0001  Gambling 61.8 (15.1) 49.4 (9.2) 12.5 (9.8, 15.1) 9.33 134.43 <0.0001 Sex or pornography 61.2 (14.4) 49.2 (9.1) 12.0 (9.8, 14.2) 10.73 178.05 <0.0001  Any behavioral 60.0 (14.0) 48.7 (8.5) 11.4 (9.7, 13.0) 13.49 317.63 <0.0001 BSI, Brief Symptom Inventory; GSI, Global Severity Index; CI, confidence interval; df, degrees of freedom.

Associations between the GSI (BSI) score and problematic SU and addictive behaviors

BSI, Brief Symptom Inventory; GSI, Global Severity Index; CI, confidence interval; df, degrees of freedom.

Machine Learning Analysis

The predictive ability of the BSI was evaluated on two parallel tracks. Continuous-track R 2 (proportion of variance explained on the held-out test set) is reported in Table 4 ; binary-track AUPRC, AUC-ROC, precision, and recall at the default 0.5 threshold are reported in Table 5 . All values include 1,000-replicate bootstrap 95% confidence intervals. For at least one machine-learning algorithm per outcome, predictive ability exceeded the GSI-only baseline; the gain over baseline was largest for the rarer behavioral outcomes.

Table 4. Continuous-track predictive performance: R 2 (95% CI) per algorithm and outcome Model Alcohol ASSIST Drugs ASSIST Gaming GAS Sex HBI Pornography PPUS Gambling PGSI 1. Base model (linear, GSI only) 3.42% 9.49% 9.17% 23.16% 13.91% 3.39% 2. Decision tree 6.44% [1.53%, 13.71%] 4.42% [0.92%, 10.66%] 3.83% [0.64%, 10.30%] 22.14% [12.71%, 34.91%] 10.90% [4.12%, 21.49%] 1.41% [0.01%, 9.91%] 3. RF 4.96% [1.12%, 10.47%] 10.05% [3.62%, 18.85%] 8.73% [3.44%, 16.52%] 24.80% [16.92%, 33.93%] 17.07% [9.64%, 25.97%] 4.71% [0.58%, 14.13%] 4. Boosting 6.23% [1.60%, 12.89%] 11.51% [4.38%, 20.83%] 5.08% [1.41%, 11.44%] 23.17% [14.66%, 33.96%] 11.71% [5.48%, 20.76%] 3.38% [0.01%, 15.74%] 5. LASSO 4.74% [0.98%, 11.08%] 11.23% [4.65%, 20.17%] 6.77% [2.04%, 14.36%] 24.08% [14.55%, 35.47%] 17.16% [8.85%, 28.15%] 5.09% [0.65%, 16.34%] 6. Elastic net 4.89% [1.15%, 10.57%] 10.98% [4.31%, 19.71%] 8.36% [3.12%, 16.67%] 23.85% [14.39%, 35.23%] 16.82% [8.75%, 27.38%] 4.45% [0.39%, 15.76%] 7. Subset selection (BIC) 3.75% [0.48%, 9.17%] 10.90% [4.63%, 19.82%] 4.16% [0.91%, 9.54%] 21.89% [12.11%, 33.37%] 16.21% [7.09%, 27.72%] 4.92% [0.56%, 15.57%]
R
2 is the proportion of variance in the continuous outcome score explained by the model on the held-out test set. 95% confidence intervals were computed using 1,000-replicate nonparametric bootstrap. Models were fitted in tidymodels (R) with repeated 10-fold cross-validation × 5 repeats on a training set (80%) and evaluated once on a held-out test set (20%). ASSIST, Alcohol, Smoking and Substance Involvement Screening Test; GAS, Game Addiction Scale; HBI, Hypersexual Behavior Inventory; PPUS, Problematic Pornography Use Scale; PGSI, Problem Gambling Severity Index; LASSO, Least Absolute Shrinkage and Selection Operator; BIC, Bayesian information criterion.

Continuous-track predictive performance: R 2 (95% CI) per algorithm and outcome

R
2 is the proportion of variance in the continuous outcome score explained by the model on the held-out test set. 95% confidence intervals were computed using 1,000-replicate nonparametric bootstrap. Models were fitted in tidymodels (R) with repeated 10-fold cross-validation × 5 repeats on a training set (80%) and evaluated once on a held-out test set (20%). ASSIST, Alcohol, Smoking and Substance Involvement Screening Test; GAS, Game Addiction Scale; HBI, Hypersexual Behavior Inventory; PPUS, Problematic Pornography Use Scale; PGSI, Problem Gambling Severity Index; LASSO, Least Absolute Shrinkage and Selection Operator; BIC, Bayesian information criterion.

Table 5. Binary-track predictive performance: AUPRC, AUC-ROC, precision, and recall (95% CI) at the clinical threshold for each algorithm and outcome Model Model metric Alcohol Drugs Gaming Sex Pornography Gambling Base model AUPRC 0.23 [0.16, 0.32] 0.33 [0.25, 0.44] 0.19 [0.10, 0.34] 0.33 [0.18, 0.49] 0.23 [0.03, 0.50] 0.10 [0.03, 0.31] AUC-ROC 0.66 [0.59, 0.71] 0.68 [0.62, 0.74] 0.73 [0.65, 0.81] 0.80 [0.71, 0.87] 0.83 [0.69, 0.94] 0.72 [0.60, 0.83] Precision 0.50 [0.00, 1.00] 0.71 [0.36, 1.00] 0.75 [0.00, 1.00] 1.00 [1.00, 1.00] NA [NA, NA] 0.50 [0.00, 1.00] Recall 0.03 [0.00, 0.07] 0.06 [0.02, 0.11] 0.09 [0.00, 0.21] 0.06 [0.00, 0.16] 0.00 [0.00, 0.00] 0.06 [0.00, 0.20] Decision tree AUPRC 0.32 [0.24, 0.41] 0.29 [0.22, 0.37] 0.36 [0.24, 0.46] 0.26 [0.14, 0.38] 0.25 [0.08, 0.43] 0.33 [0.16, 0.48] AUC-ROC 0.60 [0.54, 0.67] 0.47 [0.39, 0.54] 0.67 [0.58, 0.76] 0.62 [0.54, 0.71] 0.71 [0.52, 0.88] 0.70 [0.58, 0.83] Precision 0.33 [0.21, 0.46] 0.26 [0.18, 0.34] 0.20 [0.12, 0.29] 0.24 [0.10, 0.38] 0.10 [0.02, 0.19] 0.15 [0.06, 0.24] Recall 0.25 [0.15, 0.35] 0.30 [0.21, 0.40] 0.48 [0.31, 0.66] 0.29 [0.14, 0.47] 0.40 [0.10, 0.75] 0.50 [0.25, 0.75] RF AUPRC 0.23 [0.16, 0.32] 0.35 [0.25, 0.45] 0.16 [0.09, 0.28] 0.18 [0.10, 0.29] 0.20 [0.03, 0.49] 0.09 [0.04, 0.19] AUC-ROC 0.62 [0.56, 0.69] 0.69 [0.63, 0.75] 0.73 [0.65, 0.81] 0.77 [0.69, 0.85] 0.81 [0.68, 0.92] 0.72 [0.58, 0.85] Precision 0.33 [0.13, 0.57] 0.56 [0.31, 0.81] 0.38 [0.00, 0.75] 0.27 [0.00, 0.55] 1.00 [1.00, 1.00] 0.08 [0.00, 0.27] Recall 0.08 [0.03, 0.15] 0.10 [0.04, 0.17] 0.09 [0.00, 0.20] 0.10 [0.00, 0.21] 0.10 [0.00, 0.33] 0.06 [0.00, 0.21] Boosting AUPRC 0.25 [0.18, 0.35] 0.35 [0.26, 0.45] 0.19 [0.09, 0.32] 0.21 [0.12, 0.36] 0.05 [0.02, 0.14] 0.10 [0.04, 0.21] AUC-ROC 0.65 [0.58, 0.72] 0.72 [0.66, 0.77] 0.74 [0.66, 0.81] 0.79 [0.71, 0.86] 0.69 [0.47, 0.88] 0.75 [0.61, 0.87] Precision 0.32 [0.17, 0.48] 0.48 [0.29, 0.68] 0.20 [0.08, 0.32] 0.21 [0.09, 0.34] 0.04 [0.00, 0.12] 0.13 [0.03, 0.24] Recall 0.15 [0.07, 0.24] 0.17 [0.10, 0.26] 0.27 [0.13, 0.41] 0.26 [0.11, 0.43] 0.10 [0.00, 0.36] 0.31 [0.09, 0.56] Elastic net AUPRC 0.28 [0.20, 0.39] 0.34 [0.25, 0.43] 0.18 [0.09, 0.28] 0.28 [0.16, 0.47] 0.04 [0.01, 0.10] 0.15 [0.04, 0.32] AUC-ROC 0.66 [0.59, 0.73] 0.70 [0.64, 0.77] 0.74 [0.65, 0.83] 0.75 [0.64, 0.86] 0.69 [0.53, 0.86] 0.73 [0.60, 0.86] Precision 0.34 [0.20, 0.49] 0.42 [0.25, 0.59] 0.14 [0.05, 0.26] 0.30 [0.17, 0.45] 0.07 [0.00, 0.16] 0.07 [0.00, 0.17] Recall 0.18 [0.09, 0.28] 0.17 [0.10, 0.25] 0.18 [0.06, 0.33] 0.42 [0.24, 0.58] 0.30 [0.00, 0.63] 0.12 [0.00, 0.33] LASSO AUPRC 0.28 [0.20, 0.39] 0.34 [0.25, 0.43] 0.18 [0.09, 0.28] 0.28 [0.16, 0.47] 0.07 [0.02, 0.15] 0.11 [0.04, 0.30] AUC-ROC 0.66 [0.58, 0.73] 0.70 [0.64, 0.77] 0.74 [0.65, 0.83] 0.75 [0.64, 0.86] 0.79 [0.65, 0.92] 0.74 [0.63, 0.85] Precision 0.34 [0.20, 0.49] 0.42 [0.25, 0.59] 0.14 [0.05, 0.26] 0.30 [0.17, 0.45] 0.10 [0.00, 0.21] 0.12 [0.00, 0.25] Recall 0.18 [0.09, 0.28] 0.17 [0.10, 0.25] 0.18 [0.06, 0.33] 0.42 [0.24, 0.58] 0.30 [0.00, 0.60] 0.19 [0.00, 0.40] Models predict the binary indicator of problematic SU or behavioral addiction at the clinical threshold defined in Table 2. SMOTE within training folds at moderate ratio (˜30%) was used to address class imbalance; results without resampling and with class-weighted loss are reported in online supplementary Table S5. All metrics are evaluated on the held-out test set; 95% confidence intervals were computed using 1,000-replicate nonparametric bootstrap. AUPRC, area under the precision-recall curve; AUC-ROC, area under the receiver-operating-characteristic curve; LASSO, Least Absolute Shrinkage and Selection Operator.

Binary-track predictive performance: AUPRC, AUC-ROC, precision, and recall (95% CI) at the clinical threshold for each algorithm and outcome

Models predict the binary indicator of problematic SU or behavioral addiction at the clinical threshold defined in Table 2. SMOTE within training folds at moderate ratio (˜30%) was used to address class imbalance; results without resampling and with class-weighted loss are reported in online supplementary Table S5. All metrics are evaluated on the held-out test set; 95% confidence intervals were computed using 1,000-replicate nonparametric bootstrap. AUPRC, area under the precision-recall curve; AUC-ROC, area under the receiver-operating-characteristic curve; LASSO, Least Absolute Shrinkage and Selection Operator.

On the continuous track ( Table 4 ), the best-performing algorithm and R 2 per outcome were compulsive sexual behavior, random forest, R 2 = 24.80% (95% CI: 16.92, 33.93); pornography use, LASSO, R 2 = 17.16% (8.85, 28.15); problematic drug use, boosting, R 2 = 11.51% (4.38, 20.83); gaming, random forest, R 2 = 8.73% (3.44, 16.52); problematic alcohol use, decision tree, R 2 = 6.44% (1.53, 13.71); and gambling, LASSO, R 2 = 5.09% (0.65, 16.34). Behavioral addictions, particularly compulsive sexual behavior and pornography use, remained the most predictable continuous outcomes, consistent with their stronger GSI mean differences ( Table 3 ).

On the binary classification track ( Table 5 , synthetic minority over-sampling within training folds at moderate ratio; full no-resampling and class-weighted comparisons in online suppl. Table S5), AUPRC ranged from 0.04 (pornography, LASSO) to 0.36 (gaming, decision tree); the highest AUPRC per outcome consistently exceeded its prevalence-based chance baseline by a substantial multiple – for example, the gaming decision-tree AUPRC of 0.36 against a 0.02 prevalence baseline and the gambling decision-tree AUPRC of 0.33 against the corresponding 0.05 prevalence baseline. AUC-ROC values for the best-performing algorithms were in the moderate-to-good range across outcomes ( Table 5 ). Best models’ AUC-ROC ranged from 0.66 (alcohol) to 0.83 (pornography), while models training and optimization optimized recall for rare outcomes, hindering AUC-ROC performance in such settings. The best-performing algorithm differed across outcomes and metrics ( Table 5 ). For rarer outcomes (gaming, pornography), AUPRC gain above prevalence serves as a good discrimination metric.

Operating-point metrics at the default 0.5 classification threshold were modest, especially for rare outcomes: precision and recall were typically below 0.30 even when AUC-ROC and AUPRC indicated good discrimination. From a screening perspective, however, the gain in case detection over the GSI-only baseline was not negligible. For example, the decision-tree models identified 25–50% of true cases (e.g., 48% for gaming, 50% for gambling, 40% for pornography use), whereas the GSI-only logistic baseline identified between 0% (pornography) and 9% (gaming) of true cases across the six outcomes ( Table 5 ). This dissociation reflects the well-known difficulty of operating at low prevalence with an untuned threshold and is reported in Table 5 . In a clinical application, threshold selection would itself require tuning against the local target prevalence and the relative cost of false-positive versus false-negative classifications; the present results document discrimination performance rather than recommend a deployable threshold.

Test-set estimates were consistent with cross-validation out-of-sample estimates across all algorithms and outcomes (online suppl. Table S4). Detailed selected hyperparameters are reported in online supplementary Table S3.

Discussion

The present study provides valuable insight into the untapped potential of the BSI in providing information about diverse addictions. In an adult population sample from Israel, the overall BSI score (GSI) was significantly higher among those with problematic alcohol use, drug use, or other addictive behaviors (Internet gaming, compulsive sexual behavior, gambling, and pornography use) than among those without. The differences in GSI scores reinforce the well-established comorbidity between psychopathology and addiction. These findings highlight the utility of general distress measures, such as the BSI, in identifying individuals at risk for addiction-related comorbidities. Moreover, machine learning algorithms were used to maximize the ability of the BSI items to predict addiction outcomes, supporting the principle that general psychopathology measures contain indirect information about addiction risk.

The positive association between the GSI and each addiction type aligns with previous research linking general psychopathology, as measured by the BSI, and specific addictions. For instance, individuals reporting problematic pornography use in a German population study exhibited higher BSI scores [ 24 ]; Dutch patients treated for gamma-hydroxybutyrate dependence showed greater distress than outpatient controls, especially with comorbid psychiatric disorders [ 53 ]. Stimulant users in Europe [ 22 ] and alcohol users in the USA [ 23 ] were associated with increased psychological distress. Thus, evidence shows that substance and behavioral addictions are related to the overall BSI measure of psychopathology.

To our knowledge, prior research had not investigated these associations by quantifying the extent to which the BSI can predict severity of diverse addictions. As a proof of principle, this study applied a deliberately generic machine-learning workflow to demonstrate that a general psychopathology questionnaire can serve as an informative predictor of addiction outcomes. Crucially, the same pipeline (identical algorithm grids, the same resampling strategy, the same tuning criterion) was applied uniformly across all six addiction outcomes and all algorithms without iterating or calibrating to optimize any single target. The models were not tailored to a specific clinical use case, nor were their decision thresholds adapted post hoc. This generic design was chosen to establish a benchmark demonstrating that the BSI carries item-level signal about addiction risk even under an unoptimized workflow. In any future deployment, the pipeline would need to be calibrated to the clinical context. For example, in a setting where missing a true positive is particularly costly (e.g., addiction screening in psychiatric intake), recall should be prioritized even at the expense of precision; in a setting where false-positive referrals carry high burden, precision should be favored instead. Different clinical objectives would warrant different tuning criteria, thresholds, and resampling strategies, decisions that are inherently context dependent and were therefore deliberately left open in this proof-of-principle study.

Because several of the addiction outcomes are rare in the population (gaming ∼2%, pornography ∼3%), conventional AUC-ROC can overstate a model’s practical discrimination. The AUPRC provides a more informative summary in imbalanced settings because it directly indexes how well a model distinguishes rare positive cases from the majority class. In this study, AUPRC values consistently exceeded the prevalence baseline, indicating that the models extract genuine signal, not merely majority-class prediction.

The most clinically consequential metric in a screening context may be recall (sensitivity): the proportion of true cases the model identifies. When prioritizing such a metric, the advantage of the machine-learning models over the GSI-only baseline can become even more significant. At the default 0.5 classification threshold, the logistic baseline on the GSI identified between 0% and 9% of true cases across the six outcomes ( Table 5 ). By contrast, the decision-tree models, tuned on recall within the same generic framework, identified 25–50% of true cases (e.g., 48% for gaming, 50% for gambling, 40% for pornography use). This gain is particularly meaningful in addiction screening, where missing a case (a false negative) often carries greater cost than generating a referral that turns out to be unnecessary (a false positive). In such contexts, even imperfect sensitivity adds substantial value because each additional detected case represents a patient who might otherwise go unscreened and untreated.

Importantly, these recall values were obtained at the default 0.5 probability threshold, which is not necessarily optimal for any clinical application. In practice, threshold selection is itself a decision that should be tuned against the local prevalence and the relative costs of false-positive versus false-negative errors. Lowering the threshold below 0.5 would further increase recall at the expense of precision, and in high-sensitivity screening contexts this may be the appropriate trade-off. We did not perform post hoc threshold optimization in this study precisely because it would conflate proof-of-principle evidence with deployment-specific tuning. Nevertheless, the full ROC and precision-recall curves (online suppl. Fig. S3) show that substantially higher recall values are achievable by selecting alternative operating points, an option available to any future implementation that defines its target sensitivity and acceptable false-positive rate.

The algorithms performed “best” for different addictions per metric. These results highlight the differential informativeness of the BSI for behavioral addictions compared to SU disorders. The variation in predictive power across different addictions suggests that while there are common underlying mechanisms across addiction types, there may also be differences [ 40 ]. Behavioral addictions, strongly associated with internalizing disorders such as anxiety, depression, and stress disorders [ 21 ], may be more directly captured by BSI dimensions, whereas SU disorders may be more related to externalizing factors such as anti-social behavior and impulsivity [ 20 ]. Future research should explore whether integrating additional psychological, social, or behavioral measures into predictive models could enhance predictive accuracy for SU disorders.

This differential predictive ability may be attributable to varying degrees of association between specific psychopathologies and addiction types. For instance, the strong association between general psychopathology and compulsive sexual behavior or pornography use aligns with previous research highlighting the detrimental consequences of these behaviors [ 24 ]. Future studies should further unpack the nuanced relationships between general and specific psychopathologies and different addictions – for example, the contribution of moral incongruence and related distress to compulsive sexual or pornography use [ 54 ]. While this study examined the potential of a unidimensional psychopathology measure (BSI) to predict addiction risk, demonstrating its utility for behavioral addictions and to a more restricted extent for SU disorders, future research should investigate whether incorporating multidimensional screening tools could enhance predictive accuracy for both.

Across models, predictive performance was largely comparable. Notably, reduced-item approaches (subset selection, LASSO, elastic net) performed as well as or better than models using all 53 BSI items. It is important to note that these models were not designed to create a shortened BSI but to optimize predictive accuracy. The strong performance of reduced-item models suggests that some items may carry greater predictive weight, yet this should not be interpreted clinically. Because BSI items are highly intercorrelated, reflecting the scale’s unidimensional structure, the selected items represent statistical optimization rather than distinct diagnostic markers. Their utility lies in revealing how general psychopathology captures variance relevant to addiction risk. Future research should clarify whether these patterns reflect true addiction-specific differences or shared underlying distress and refine item-selection strategies to maximize predictive value across mental-health domains rather than adhere to subscale boundaries, as shown in recent personality research [ 55 ].

Although decision-tree models explained the least variance overall (except for alcohol), their simplicity and transparency might make them a valuable starting point for the future development of interpretable, rule-based screening tools that could ultimately be deployed in clinical contexts with limited digital infrastructure. The present decision-tree models provide initial, proof-of-principle evidence that such rule-based predictions can be derived from the BSI, and future studies can refine, calibrate, and prospectively validate these models against specific clinical contexts to establish an implementation-ready screening heuristic. Many healthcare systems, especially in low-resource or underfunded settings, lack access to advanced data-driven tools due to financial or logistical barriers [ 56 , 57 ]. In such environments, decision trees might provide a cost-effective and interpretable screening method that, once refined and properly adapted, could enable immediate, rule-based application without specialized software, training, or infrastructure. The present results therefore suggest initial, proof-of-principle evidence supporting the development of such tools in future work.

The integration of machine learning in healthcare presents both opportunities and ethical challenges [ 58 ], particularly in addiction prediction, where bias and stigma are major concerns [ 59 ]. Addiction has historically been viewed through a moral lens, and predictive models risk reinforcing biases if not carefully implemented [ 60 ]. To prevent discrimination and ensure fair application, these tools should be used as clinical decision-support aids rather than definitive diagnostic instruments. Additionally, regulatory oversight, ongoing model auditing, and fairness-focused algorithm development are essential to mitigate bias and safeguard patient welfare, ensuring that AI-driven screening enhances, rather than harms, clinical care.

Study results emphasize the importance of responsively leveraging existing data from widely administered questionnaires for predictive purposes. Clinics routinely administer psychopathology questionnaires, like the BSI, during intake procedures but may not directly assess potential comorbidities. Given the high prevalence and clinical importance of such comorbidities, along with clinicians’ time constraints and the burden of additional screening, it is crucial to maximize the information contained in existing assessments. In light of this need, the potential to predict different addictions using the BSI serves as a proof-of-principle for a broader vision: that any screening tool or clinical measure routinely collected in a healthcare setting – not only the BSI but any standardized questionnaire, structured interview, or routinely recorded clinical information – could be used to inform about comorbid conditions across the full spectrum of psychopathology. In this vision, every piece of information already stored in a patient’s record becomes a potential input to an integrated predictive layer that flags patients warranting targeted follow-up screening. The BSI demonstrated meaningful predictive ability for different types of addictions, making it a promising candidate, after prospective validation, for early identification of at-risk individuals. Future refinement and validation of these predictive models could enable integration into electronic health records and clinical decision-support systems, promoting targeted screening and more personalized interventions. However, the BSI is merely one illustration of this principle; future work should extend it to other instruments and other target conditions. The ultimate aim is not a single BSI-based screening tool but a paradigm in which clinical data are routinely and automatically cross-leveraged to improve detection of comorbid psychopathology, after appropriate prospective validation and calibration to each deployment context.

Limitations

Study limitations are noted. First, the sample is best characterized as an online-panel population, thus only quasi-representative of the general population. Although quota sampling on age, gender, religiosity, education, and geographic region yielded a demographically quasi-representative sample of Hebrew-speaking Jewish adults in Israel, individuals with severe, currently active addictions, who may exhibit chaotic lifestyles, social withdrawal, or low engagement with research panels, are systematically less likely to be represented. The predictive models therefore learn the BSI-addiction relationship primarily in mild-to-moderate community presentations, and their performance in clinical populations enriched for severe phenotypes is unknown and should not be assumed. Certain subpopulations (e.g., ultra-Orthodox individuals, older adults less likely to engage in online research) may also be under-represented.

Second, the study focused on the Jewish population in Israel. Psychopathological expression and its relationship to addictive behaviors are culturally and linguistically mediated, and the models learned patterns specific to this population’s cultural context. Because SU and behavioral addictions are influenced by social and legal norms, results may differ across cultural or religious contexts. For example, alcohol consumption, which is prohibited in Islam, occurs under different social circumstances among Jewish and Muslim populations. Generalizability to other cultural, linguistic, or religious populations therefore remains an empirical question requiring direct replication and cross-cultural validation. Ongoing pilot studies among additional groups aim to expand future data collection for a more representative national sample. Nonetheless, the analyzed sample remains most culturally similar to Western populations, supporting potential generalizability to comparable settings. That said, although Israel is broadly comparable to other Western samples on several aggregate sociodemographic and psychometric dimensions, such aggregate comparability does not by itself guarantee equivalence in the expression of, or the cultural norms surrounding, the specific addictive behaviors examined here (alcohol use, gambling, pornography, internet use), which can vary substantially even across contexts broadly considered Western. Transferability of the present models to other cultural, linguistic, or religious contexts should therefore not be assumed without prior local validation; any application outside the Hebrew-speaking Israeli Jewish adult population studied here would benefit from re-estimation and recalibration against local data and from appropriate cross-cultural psychometric checks. Third, the survey was in Hebrew, but >90% of Jews in Israel are fluent in Hebrew [ 61 ].

Fourth, other biases may exist. Respondents might hesitate to report illegal or stigmatized behaviors; however, the confidential online format likely mitigated this concern [ 33 ]. Additionally, self-reported responses rely on respondent’s understanding of the questions, though standard, validated screening tools were used to enhance accuracy. There may be also recall bias, which was limited by using short timeframes for current outcomes (1 month – 1 year). Fifth, differences in the time frames across measures could introduce variability in responses.

Sixth, pre-registration was not possible. We addressed this by reporting all results transparently, using a single held-out test set evaluated only once, repeated 10-fold cross-validation with 5 repeats during model development, bootstrap-based confidence intervals on all metrics, and explicit comparison of cross-validation and test-set estimates (online suppl. Table S4); the findings are framed as exploratory and proof of principle.

Seventh, the design is cross-sectional and self-report; the BSI-addiction association is plausibly bidirectional (addiction-related distress can inflate BSI scores) and we make no causal-risk-prediction claim. Prospective and external validation in clinical populations is required before any clinical deployment.

Eighth, the binary-classification models provide initial, proof-of-principle evidence that the BSI carries screening-relevant signal substantially exceeding the GSI-only baseline. Operational performance at the default 0.5 classification threshold was modest at this stage, and future studies will be needed to refine threshold calibration against the local prevalence and cost ratio, and to prospectively validate these models, before they can be used as stand-alone screening rules in practice.

Conclusions

In a quasi-representative general population sample from Israel, culturally similar to other Western countries, the BSI provided information about the risk for substance and behavioral addictions. Ideally, everyone with mental health issues should be screened for addictions, but many are not. This proof-of-principle study suggests that existing psychopathology measures can potentially be used to more efficiently identify patients who most likely need to be screened for addictions, to get the appropriate diagnoses and integrated treatment, which may improve prognosis and limit the negative consequences of such comorbidities.

Statement of Ethics

This study protocol was reviewed and approved by the Reichman University Institutional Review Board (IRB) in Israel (Approval No. 2022050_1re_P). Written informed consent was obtained electronically from participants prior to the study in accordance with local guidelines.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

This study was not supported by any sponsor or funder.

Author Contributions

M.D.L. and D.S. made substantial contributions to the conception and design of the work and the acquisition, analysis, and interpretation of data and drafted or substantively revised the manuscript. R.E. contributed to the conception, design, and acquisition of data. S.L.-R. contributed to the design of the work and the acquisition of data and substantively revised the manuscript. M.M. contributed to the conception and design of the work and the acquisition of data and substantively revised the manuscript. All authors read and approved the final manuscript.