Leveraging Machine Learning to Advance Alcohol Research: Current Applications, Challenges, and Opportunities

Key Takeaways

• Alcohol-related publications using machine learning almost exclusively relied on conventional techniques, whereas current public discourse emphasizes state-of-the-art models.
• A majority of models predicted alcohol consumption or alcohol use disorder (AUD) diagnosis, which is generally easier to forecast than, for example, disorder or treatment outcome.
• Only 40% of models utilized multimodal data, which is needed for encoding the complexity of AUD and related clinical outcomes.
• Addressing the complexity of AUD requires creating machine learning models and quantitative benchmarks that accurately capture the multifaceted nature of alcohol use and its comorbidities.

Introduction

Alcohol use disorder (AUD) currently affects 27.9 million Americans¹ and costs the United States more than $250 billion annually.² It is the fourth leading preventable cause of death,³ underscoring the urgent need for effective strategies to understand, predict, and prevent alcohol misuse. At the mechanistic level, AUD arises from complex interactions between neural, genetic, and environmental factors that collectively influence vulnerability and resilience to alcohol’s effects. Genetic and epigenetic factors account for 40% to 60% of a person’s addiction risk,^4,5 although psychological and environmental factors (such as family history,^6,7 traumatic events,⁸ peer pressure,^9,10 other drug use,^11,12 and externalizing behaviors^13,14) also play crucial roles in shaping drinking behaviors. However, findings about these predictive factors are often inconsistent or weak: many individuals without known risk factors develop AUD,¹⁵ whereas at-risk individuals remain resilient¹⁶ (e.g., approximately 50% of maltreated youth¹⁷). Equally pressing is the challenge of predicting treatment response and long-term abstinence. Although pharmacological and behavioral interventions are available, only 7.9% of people with AUD receive alcohol use treatment in the United States,¹⁸ and of those, only 16% achieve abstinence.¹⁹

One major hurdle in alcohol research has been the predominant reliance on fragmented, small-scale datasets unable to reveal the complex and heterogeneous nature of AUD.^20,21 The need for larger heterogeneous data sets for AUD is recognized by funding agencies.²² Since 2012, the National Institute of Alcohol Abuse and Alcoholism has funded the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) study²³ to annually collect brain magnetic resonance imaging (MRI) data, neuropsychology testing, alcohol use, and related data of 831 individuals who were age 12 to 21 at baseline. NCANDA was the first to report on in vivo disruption due to alcohol of white matter microstructural development during adolescence.²⁴ An even larger sample (more than 11,000 children age 9 to 10 years at baseline) has been recruited by the Adolescent Brain Cognitive Development (ABCD) Study,²⁵ which is tracking brain development, behavior, and environmental changes to gain insights into how early life factors shape substance use and mental health outcomes. Over three decades, the Collaborative Studies on the Genetics of Alcoholism (COGA) has collected phenotypic data²⁶ from nearly 18,000 individuals ages 7 to 97 across more than 2,200 families, as well as DNA and electrophysiological measures on a large subset. The longitudinal data set provides a unique opportunity for researchers to identify genes influencing the risk of AUD and related outcomes. Another key effort is the Monitoring the Future (MTF) study—an ongoing, nationally representative study that has been surveying more than 25,000 8th, 10th, and 12th grade students²⁷ and approximately 20,000 adults ages 19 to 65 each year since 1973.²⁸ By capturing longitudinal behavioral and attitudinal changes, MTF offers a unique resource for understanding population-level patterns and risk factors for substance use. Finally, aiming to advance personalized medicine, the All of Us Research Program²⁹ is planning to collect multimodal data (including genomic profiles, electronic health records, and environmental exposure) from at least 1 million people in the United States to discern how individual differences in lifestyle, environment, and biology affect health outcomes, such as the relationship between risky drinking behavior and cancer diagnosis.³⁰

In addition to increasing sample sizes, alcohol studies have expanded the scope and diversity of data collected from each participant to enable more precise characterization of them. Current studies increasingly integrate more comprehensive self-reports,³¹ medical records,³² genetic profiling,³³ neuropsychological testing,³⁴ and imaging.³⁵ The advent of wearable technologies and ubiquitous mobile devices (such as smartphones³⁶ and biosensors measuring physical activity³⁷ and intoxication³⁸) has made it possible to conduct ecological momentary assessments.^39,40 This surge in data volume introduces substantial computational and methodological challenges, regardless of cohort size.

The emergence of larger and more diverse data sets calls for a transformation in fundamental analytical approaches. Alcohol research has traditionally relied on hypothesis-driven analyses, which involve preselecting a narrow set of variables^41,42 and applying univariate regression models at the group level to test pairwise associations. Although conceptually straightforward and statistically interpretable, this approach fragments understanding into siloed single factors that each explain only a small fraction of variance in alcohol outcomes, failing to capture the intricate multivariate interactions among behavioral traits, cognitive functions, environmental influences, and neural mechanisms that underlie AUD.^43,44 Furthermore, the findings only reflect group-level trends that generally do not translate to individual-level insights.⁴⁵

A promising alternative is machine learning (Figure 1), which is designed to translate complex, multimodal data into predictions on an individual basis.⁴⁶ Machine learning is a branch of artificial intelligence (AI) that enables computers to identify interactions among measurements (i.e., patterns) predictive of the outcome directly from data rather than relying on predefined statistical formulas or human-crafted rules. Typically, a machine learning model needs to be trained on a large number of samples, where each sample is described by a set of features (predictor variables) and a target outcome variable of interest (e.g., diagnosis, symptom severity, drinking level, or treatment response). In alcohol research, predictors may span multiple domains, including neuroimaging (e.g., functional connectivity, cortical thickness, regional volume, or microstructural integrity), behavioral and cognitive traits (e.g., measures of impulsivity, decision-making, or sensation seeking), environmental exposures (e.g., peer drinking, family history, or neighborhood stressors), and genetic or physiological markers (e.g., polygenic risk scores, heart rate, sleep disturbance). The goal of the model is to integrate information across these domains to generate a single prediction of the chosen target outcome specific to each individual.

Image

Figure 1. Machine learning methods can transform complex, multimodal data into individualized predictions of alcohol use disorder (AUD) and related outcomes. Data modalities used by studies reviewed here included, among others, behavior, biological specimens, demographics, sensor data, neuroimaging, genes, and mental health assessments. Raw data were either directly analyzed by deep learning models or first extracted into aggregate measurements for use in conventional machine learning approaches, which were used in most studies reviewed. AUD-related outcomes predicted by these methods encompassed AUD diagnosis, consumption patterns, prognosis, relapse, and treatment response. To obtain the predictions, the studies first trained the machine learning models by determining the parameter setting that minimized the difference between predicted and observed outcomes on the training data set. The studies then measured the accuracy of the predictions and the constellation of measurements (i.e., pattern) that drove the predictions, on a separate test set.

During training, the model learns (or is trained) by iteratively adjusting its internal parameters to minimize the prediction error, which is the difference between actual and predicted outcome of the training samples. Once this error is minimized, the parameterized model is evaluated on a set of test samples (i.e., participants and their data not used for training). This evaluation assesses the degree to which the model generalizes to new individuals. Common evaluation metrics⁴⁷ include accuracy (i.e., quantifying the proportion of correct predictions) and the area under the receiver operating characteristic curve (AUC), which measures how well the model discriminates between outcome classes over the full range of decision thresholds. Lastly, one can identify the pattern driving inference, offering insights into potential mechanisms or risk factors underlying alcohol outcomes.

Conventional machine learning models (e.g., random forest, support vector machine [SVM]) typically confine analysis to tabular data, where summary scores are extracted from raw measurements before training the model. For example, brain MRI studies relied on region-level summaries (such as brain volume or cortical thickness) as input to machine learning models, which confined prediction accuracy to the anatomical granularity of the brain regions defined by human experts beforehand.⁴⁸ Removing this constraint necessitates performing predictions directly from raw higher-dimensional data (e.g., all voxels of a 3D brain MRI). This is the domain of deep learning models (Figure 1), a subfield of machine learning that gained traction starting in 2012.⁴⁹ Deep learning models jointly learn to derive measurements (i.e., features) from the raw data and to predict the outcome using large-scale artificial neural networks. Treating feature extraction as part of the optimization process has led to significant improvements in predictive accuracy across numerous domains, fueling growing interest in applying deep learning approaches to data-driven medical research.⁵⁰

This review provides a systematic overview of how machine learning methods have been applied across the landscape of alcohol research over the past decade. Rather than emphasizing the specific constellations of neurobiological or psychological factors identified by machine learning analyses, the review focuses on the methodological trends, data modalities, and modeling choices that characterize current machine learning practice across alcohol-related research. Specifically, the review examines which populations and clinical outcomes have been most frequently targeted, what level of prediction accuracy has been achieved, what types of data have been used to predict outcomes, and how different machine learning models (including deep learning) have been adopted. The review also highlights the challenges associated with applying these methods to AUD studies and discusses how overcoming those challenges could transform diagnosis and treatment from subjective observations to objective, individualized assessments that lead to accurate precision medicine.

Search Method

In April and July 2025, authors conducted a literature search in PubMed, ScienceDirect, and EBSCO Academic Search Premier for publications that used machine learning to analyze alcohol-related outcomes. Across the three databases, a key word search using (alcohol[Title] OR drinking[Title]) AND (alcohol[Title/Abstract]) AND (“machine learning”[Title/Abstract] OR “deep learning”[Title/Abstract] OR predict[Title/Abstract] OR classify[Title/Abstract]) returned 2,618 manuscripts published from 2015 to April 15, 2025.

A rule-based text-mining Python script (available upon request from the authors) searched for the co-occurrence of prediction-related terms (e.g., predict, model, classify, identify) and alcohol-related terms (e.g., alcohol, drinking, dependence, relapse), while removing articles in which alcohol was a predictor rather than an outcome (e.g., “Alcohol use predicts depression”). This automatic screening reduced the selection to 567 manuscripts. A final manual selection removed eight preprints that had not been peer-reviewed, 21 articles that used animal models, 27 articles that studied fetal alcohol spectrum disorders, 139 articles that did not use machine learning approaches, and 262 articles that did not primarily analyze alcohol consumption behaviors (e.g., withdrawal symptoms, alcohol-related liver diseases, drunk driving). This selection resulted in 110 original peer-reviewed human research studies^7,51-161 that tested the model on data that it was not trained on, which is a key difference between machine learning and population-level statistical analysis.

Results of the Literature Search

The search identified 2,618 articles for initial examination. Of those, 2,508 were excluded as described above, and 110 were included in the review. For each selected article, Appendix 1 records the number of subjects, mean age, sex ratio, prediction task (e.g., predicting relapse, AUD diagnosis classification), number of predictors, machine learning model used (e.g., random forest, SVM), and reported model accuracy (e.g., AUC and classification accuracy). If multiple alcohol-related prediction tasks were explored, the average prediction performance was recorded. If multiple machine learning models were explored, the model with the highest performance was recorded.

With respect to demographic factors, 85 of the 110 reviewed articles (or 77%) reported sex ratios: eight studies (9% of 85 articles) focused exclusively on men, 40 manuscripts (47%) were somewhat balanced between the sexes (i.e., the percentage of men was 40% to 60%), and three only studied women. Of the 110 reviewed articles, 89 manuscripts (81%) reported the mean age of cohorts, with most studies studying samples with a mean age from late childhood to middle-aged adulthood: specifically, 82 articles studied cohorts with mean ages of 10 to 50 years. Only seven studies focused on cohorts with a mean age greater than 50 years.

Among all prediction outcomes (Figure 2A), AUD diagnosis and alcohol consumption patterns were the two categories most often evaluated (both 31% of articles). Articles focusing on classifying AUD diagnosis reported a median AUC of 89% (Figure 2B), which was the highest reported among categories with at least two articles. In comparison, articles on predicting alcohol consumption had a median AUC of 78%. Treatment response was the third most frequently predicted outcome but had the lowest median AUC (71%), followed by prognosis (14%, median AUC: 80%). Relapse and alcohol concentration both were assessed in only 4% of studies but had high median AUCs (85% and 96%, respectively). However, the very high AUC (96%) for sensor-based prediction of alcohol concentration requires confirmation as it was only published by one article.

Image

Figure 2. Characteristics of the studies identified by the review. (A) Articles are categorized based on the clinical outcome targeted by the machine learning prediction: more than half the articles focus on predicting alcohol use disorder (AUD) diagnosis or consumption patterns. (B) Area under the curve (AUC) of the models of various outcomes. The most accurate clinical outcomes to predict were alcohol concentration and AUD diagnosis. (C) Number of studies that used a certain type of input predictors. The most common predictors were neuroimaging data, alcohol and other substance use patterns, demographics, mental health symptoms, and behavioral data from neuropsychological testing and self-reports. (D) Number of studies that used a certain type of machine learning model. The most frequently adopted models were random forest, linear regression, and support vector machine. Note. EHR, electronic health record; kNN, k-nearest neighbor; MLP, multilayer perceptron; SVM, support vector machine.

Of the 110 articles, 46 (42%) used more than one type of predictor to predict outcomes. The most frequently used predictors were imaging data (32% of articles), followed by self-reported substance use (27%), demographics (24%), mental health (24%), and behavioral assessment (23%) (Figure 2C). Other data types were rarely used for prediction (10% or less). Conventional machine learning methods¹⁶² were the most popular approaches for prediction, including the random forest approach (used in 24% of 110 articles), linear regression (17%), and SVM (16%) (Figure 2D). Other approaches, such as logistic regression, gradient boosting, decision tree, multilayer perceptron, ensemble methods, k nearest neighbor, and clustering were used in 10% of studies or less. Deep learning approaches were used in 3% of investigations. Studies with larger sample sizes tended to use higher-dimensional input features for prediction (i.e., used larger numbers of predictors) (Figure 3A; r = .20, p = .05); they also resulted in lower prediction accuracy (Figure 3B; r = −.26, p = .03).

Image

Figure 3. Correlation of sample size with number of predictors and accuracy. The sample size showed (A) a positive correlation with the number of input features to the machine learning models; and (B) a negative correlation with model prediction accuracy.

Results of the Studies Reviewed

Demographics

Historically, AUD has been mostly diagnosed in men.¹⁶³ However, in recent years, the prevalence of AUD among women has been rising. This is a concerning trend from a public health standpoint, as the adverse effects of alcohol misuse tend to be more severe in women than in men.^164,165 This shift is evident in the articles reviewed herein, with only 9% of studies focusing exclusively on men, and the proportion of male participants in studies published since 2023 (n = 49) significantly lower (p = .032) compared to those published before 2023 (n = 36). Accordingly, the reviewed articles identified sex differences in the correlations of neural substrates and behavioral correlates with alcohol misuse and treatment response.^77,97 For example, childhood trauma and microstructural integrity in temporal and motor structures were female-exclusive predictors of alcohol misuse in young adulthood, whereas social recognition ability, personality traits, and microstructural integrity in cerebellar, motor, and occipital structures were more important predictors for young males.^77,97 Liver function test results predicted treatment response exclusively in males, whereas mental health symptoms were more informative predictors for females.¹⁵⁴ Discovery of these distinctions was performed either by training a single machine learning algorithm on both sexes and comparing important predictors between males and females,⁷⁷ or by training separate models for males and females and comparing the outcomes of the two models.⁹⁷

In addition to sex, age is a critical demographic factor in alcohol use.¹⁶⁶ Most recent machine learning studies have focused on adolescents through middle-aged adults, a critical window when drinking behaviors typically initiate^167,168 and consolidate.^169-171 As acknowledged in the reviewed articles, adolescence and early adulthood are critical for brain development^60,146 and encompass major life transitions (e.g., transition from high school to college, entry into the workforce, and early parenthood). These transitions often involve shifts in social norms and heightened stress, both of which can drive alcohol use.^97,140 Notably, binge drinking (defined as five or more drinks per occasion) peaks during the late teens to early 30s¹⁷² and is one of the strongest predictors of developing AUD later in life.¹⁷³ By contrast, alcohol misuse in adulthood poses the greatest public health burden,^77,96 including alcohol-related injuries (e.g., accidents, violence, fatalities)^174-176 and economic costs (e.g., lost productivity and increased health care expenditure).^165,177 Thus, machine learning studies are particularly valuable for simultaneously analyzing tens or hundreds of factors drawn from many conceptually different domains with the goal of identifying early markers of vulnerability and understanding later-life health and social outcomes to inform prevention and intervention strategies.

Prediction Outcome

Distinguishing individuals with AUD from controls was a prevalent prediction task probed by machine learning analysis; it was addressed in 35 of the 110 articles (31%; Figure 2A). The high proportion may be due to the relatively high accuracy achieved by machine learning approaches (median AUC: 89.0%; Figure 2B). Predicting AUD is relatively simple compared to other clinical outcomes because it is a binary decision process based on clear diagnostic framing (i.e., meets the criteria of the Diagnostic and Statistical Manual of Mental Disorders: yes/no). Furthermore, AUD is typically associated with severe and enduring alterations in brain regions and behavior, as also documented by the reviewed articles.^84,129 Accordingly, the prediction is often based on neuroimaging data (20 out of 35 articles). For example, prediction models based on structural and functional MRI have yielded converging evidence of AUD-related abnormalities in prefrontal circuits (executive dysfunction),^55,69,96,138 ventral striatum (reward sensitivity),^84,96 the cingulate cortex,^24,55 default mode network,^138,139 and the sensory cortex^96,107 (Appendix 1). Although these regions also have been previously reported by literature based on traditional group-level analyses, the effect sizes revealed by machine learning approaches far exceed those identified by univariate statistical tests. For example, a neural network classifier achieved an AUC of 79% in classifying 51 people with AUD and 51 control subjects using whole-brain resting-state functional connectivity.⁹⁶ In contrast, individual functional connectivity features typically exhibited group differences of t < 4.0, corresponding to an accuracy of only about 0.65.⁹⁶

Predicting alcohol consumption patterns that did not meet the criteria for AUD was another task frequently investigated by machine learning models (35 studies, or 31% of the articles). Because alcohol misuse is among the strongest predictors of future AUD onset, this task holds significant value for prevention and policy design. Prediction models for this task often involved mixed data sources beyond imaging-based predictors. Common predictors of alcohol misuse identified by machine learning included sociodemographic factors (e.g., male sex or lower socioeconomic status),^{51,119,130,142} poor executive control,^51,119 social behavior change,^83,119 psychological dysfunction,^83,119 and other substance use.^51,83 The accuracy of these predictors in differentiating people who drink heavily from control subjects (median AUC 78%) was substantially higher than the accuracy of single predictors, which typically yielded only marginal improvements above chance.^51,61,119 Collectively, these findings position machine learning approaches as a promising and potentially clinically actionable framework for early risk quantification and targeted prevention of AUD.

The third most predicted outcome of machine learning was treatment response. Common predictors identified by the reviewed articles included severity of dependence and craving at baseline,^7,98,127 self-efficacy,^127,154 and psychiatric comorbidity.^7,98,154 Initial evidence also suggests that the quantitative, individualized prediction by machine learning models aggregating multiple predictors was more accurate than predictions made by human experts.¹²⁷ Despite the potential, this prediction task was associated with the lowest median AUC (71%) among all prediction outcomes. One reason is that treatment response often is assessed using nonstandardized subjective measures (e.g., craving scales and self-reported relapse) that are known to exhibit low test-retest reliability and high individual variability.⁷¹ Further increasing variability was the wide difference in treatment duration, from a few days¹³² to a few years¹³⁶ resulting in inconsistent timing of follow-up assessments. Thus, clear neurobiological and behavioral correlates with treatment outcomes remain elusive. In addition, studies on treatment outcomes also suffer from only a limited proportion of enrolled individuals completing the full course of treatment (e.g., 78% drop out rate⁶⁶). This low completion rate contributes to significant class imbalance between individuals who do or do not respond to treatment. Collectively, these factors make it difficult to robustly train machine learning models and hinder their generalizability,¹⁷⁸ which is reflected in treatment response emerging as the clinical outcome with the lowest average predictive accuracy (Figure 2B).

Types of Data Used for Training Machine Learning Models

This review revealed that predictors of alcohol-related clinical outcomes encompass a diverse array of data types (Figure 2C), with the most popular one being medical imaging data. As mentioned, chronic heavy alcohol use results in significant volume loss in brain regions⁵⁵ and disrupts brain function.^69,84,96 Neuroimaging can capture these alterations noninvasively, making it appealing for investigating addiction mechanisms.¹⁷⁹ Beyond improving the understanding of addiction mechanisms, the objectivity of nonfunctional neuroimaging transcends traditional self-report measures and thus has the potential for providing quantitative biomarkers for identifying individuals at elevated risk for drinking onset, predicting relapse,¹⁸⁰ and assessing treatment outcomes.¹⁸¹

However, neuroimaging data also have some disadvantages. They are relatively expensive to acquire, can be susceptible to measurement variability and artifacts (e.g., head motion, scanner differences, and physiological fluctuations), and are associated with small effect sizes (typically explaining < 5% to 10% of the variance),^182,183 because they are indirect proxies reflecting intermediate phenotypes (such as functional connectivity or activation patterns) rather than direct measures of behaviors, such as craving or relapse.^184,185 In contrast to neuroimaging studies, neuropsychological assessments, self-reports, and demographic variables (e.g., alcohol use history, age of onset, or family history) are more proximally aligned with AUD outcomes and easier to collect, often translating to stronger predictive accuracy. This explains why the 18 studies using only neuroimaging as predictors had a lower median AUC of 79% than the 22 studies using a single type of nonneural predictors (median AUC: 91%).

No single data type (whether neuroimaging, genetic, behavioral, or self-report) fully captures the complexity of AUD and related clinical outcomes. In this review, 42% of studies investigated predictors from multiple modalities, which machine learning models are primed to do because their data-driven search does not require prior knowledge about the modalities they analyze. Each modality—whether it be genetic data, brain imaging, behavioral traits, or environmental exposures—generally captures a distinct yet complementary layer of risk or resilience.^26,128,186 By seamlessly integrating these diverse information sources, the reviewed studies consistently indicated that multimodal machine learning models not only provide a more comprehensive representation of individual differences in alcohol-related outcomes but also enhance prediction accuracy and generalizability compared to single-modal models.^7,55,119,129

Discussion

Model Design Choices

In general, the choice of machine learning models is heavily influenced by data set characteristics, particularly sample size and number of predictors, which are often strongly correlated (Figure 3A).¹⁸⁷ Training highly complex models on a limited number of samples with many predictors can lead to “overfitting,” where the model captures noise patterns in the training data rather than generalizable relationships based on reliable measurements. Therefore, given that most alcohol-related datasets typically only include fewer than 1,000 subjects, conventional machine learning methods (i.e., random forest, SVM, linear regression)¹⁶² are more common because they depend on relatively fewer model parameters to learn. In contrast, state-of-the-art deep learning approaches require far larger data sets to generalize.^188,189 Empirical studies have shown that these “sophisticated” models often have overestimated accuracy and are easily overfitted on small data sets.¹⁸⁹

Moreover, the data in alcohol studies were often structured tabular data (e.g., demographic variables, self-report scores, or brain regional measurements) rather than raw, high-dimensional inputs (e.g., minimally processed structural MRI data). In these tabular-feature scenarios, traditional models often performed on par with or even better than deep learning because the key signals of the precomputed features (e.g., regional brain volumes or cognitive scores) and relatively simple prediction targets (often binary outcomes, such as diagnosis vs. control) did not require complex modeling. Recent evaluations on structured data sets found that tree-based ensembles (such as random forest and XGBoost¹⁹⁰) consistently outperformed deep neural networks (which are considered the state-of-the-art in AI) for both purely numerical features and mixed data types.^191,192 Thus, although deep learning has only begun to tread in alcohol research, its broader potential will depend on the availability of large, well-curated data sets.

Beyond prediction accuracy, the decision process of conventional models is generally easier to interpret compared with deep learning. This is essential for results from alcohol studies to be useful to clinicians and researchers who need to understand why a model makes a prediction to trust and act on it. For example, random forest can provide ranked feature importance¹⁹³ to indicate which features are more influential to the model prediction. By contrast, deep neural networks are often regarded as “black boxes” with complex internal computations that defy straightforward explanations.^188,194 This opacity can lead to reluctance in adopting deep learning models for medical decisions, because clinicians cannot easily verify what the model has learned or identify potential biases.¹⁹⁵

Another important advantage of conventional machine learning methods over deep learning for alcohol researchers is the relatively low barrier to deployment and maintenance.¹⁹⁶ Conventional methods are typically easier to set up, requiring fewer design choices and less parameter tuning than deep learning approaches. They handle mixed data types and occasional missing values with minimal preprocessing, whereas deep networks usually demand careful data cleaning and customization for each new task. Conventional machine learning algorithms also run efficiently on standard computational resources, with training of the models often completed in seconds or minutes on a normal personal computer. In contrast, deep learning approaches necessitate high-performance GPUs or cloud infrastructure, which is expensive and harder to operate.¹⁹⁴

In summary, the dominance of conventional machine learning approaches in alcohol-related neuroimaging and behavioral prediction can be attributed to a combination of data limitations, interpretability needs, and ease of use. Although this observation might be biased by the search criteria of this review, which were confined to nonspecific key words, such as “machine learning” and “deep learning,” none of the deep learning approaches used in the studies were based on state-of-the-art architectures (e.g., transformers, generative models, and large language models) that recently have been the focus of public discourse.

Key Methodological Challenges and Future Directions for AI in Alcohol Research

Despite the increased use of machine learning analyses in alcohol research, several methodological limitations continue to constrain the full potential of this technology. The limiting factor most commonly mentioned among the reviewed articles is the number of available samples (25 articles). As mentioned, generalizability of machine learning methods relies on training them on heterogeneous data sets, which requires collecting data from a large number of samples. Funding institutions have identified this need and have recently funded several large, multicenter studies, such as NCANDA, ABCD, COGA, and MTF. However, simply increasing sample size does not overcome existing limitations; machine learning methodology also needs to be tailored to study such large samples in alcohol research. This review revealed a negative correlation between sample size and model prediction accuracy (Figure 3B), contradicting the general understanding that the accuracy of machine learning should increase with the size of training data. This paradox reflects a broader challenge in psychiatry increasingly recognized in the literature,^197,198 namely the trade-off between predictive accuracy and population heterogeneity.¹⁹⁷ Smaller data sets tend to be more homogeneous, allowing models to learn a single predictive pattern with high accuracy. However, this pattern is likely specific to a restricted subpopulation, thus resulting in a lower accuracy in larger heterogenous data sets; this was pointed out by 35 articles of this review. A promising way forward lies in the development of foundation models.¹⁹⁹ These are large-scale, general-purpose models that are first trained on broader data sets, including those not necessarily related to alcohol research. Once pretrained, these models can be fine-tuned to specific subpopulations or clinical questions (e.g., predicting relapse in young females or treatment response in people with chronic AUD with liver disease). Creating such models could be essential for moving from subjective, population-level generalizations toward truly individualized, quantitative, and context-sensitive predictions that are the foundational promise of precision psychiatry.

Another challenge in using machine learning to advance alcohol research is the modeling of confounders, which were stringently accounted for by only 27 of the 110 articles (25%) reviewed. When confounding variables (e.g., age, sex, or comorbid mental health conditions) influence both the input features (e.g., neural or behavioral measures) and the outcome (e.g., alcohol consumption or treatment response), models may learn spurious associations that lead to misinterpretation of machine learning findings²⁰⁰ as also acknowledged by 24 articles of this review. For instance, age affects both brain connectivity and drinking patterns.²⁰¹ If age is not adjusted for, the model may “predict” alcohol misuse by detecting brain maturation, not alcohol effects. Similarly, alcohol consumption patterns significantly differ between males and females; consequently, a machine learning model not controlling for sex might simply detect sex differences in the input predictors. Another type of confound is caused by comorbidity (mentioned by nine articles), because alcohol use often co-occurs with other physical and psychiatric symptoms (e.g., liver disease and depression) and other substance use. The identified predictors hence might not be linked to alcohol outcomes but to other confounding phenotypes. This type of signal leakage caused by confounds^202,203 often results in misleading model interpretations²⁰⁴ and inflates accuracy scores (e.g., AUC, accuracy). This is especially problematic in nonrandomized, observational data sets typically encountered in alcohol research.

Mitigation of confounding effects is well established in traditional statistics but often underdeveloped or overlooked in machine learning research, which has historically only focused on maximizing predictive accuracy. In particular, modern deep learning models have the capacity to encode complex nonlinear confounding effects that are hard to detect or adjust for.²⁰³ This gap could be addressed by treating modeling confounders as a core component of the machine learning pipeline. Researchers could apply appropriate preprocessing techniques (e.g., residualization, stratification, or harmonization²⁰⁵) before training, and may want to consider models specifically designed to reduce confounding, such as domain-invariant²⁰⁶ or confounder-aware approaches.^61,207,208 Also, evaluation of the models could be enhanced by going beyond reporting a single AUC and incorporating fairness metrics²⁰⁹ (e.g., subgroup AUCs) and sensitivity analyses²¹⁰ (e.g., performance changes with and without confounders). Validation of models on external data sets where the distribution of individual confounders differs would be crucial for ensuring generalizability and robustness.

Beyond modeling confounding effects, six of the 110 articles reviewed here explicitly mentioned the limitation that the adopted machine learning approach could only reveal statistical association between predictors and outcome but could not reveal causality between them. This issue highlights an urgent need for a paradigm shift in machine learning applications within alcohol research. Rather than focusing solely on optimizing prediction accuracy, models could also be designed to reveal dependencies and potential causal pathways among variables. New models designed to discern how external environmental and sociodemographic factors (i.e., those that give rise to population heterogeneity) moderate genetic, neural, and behavioral underpinnings of AUD in a data-driven fashion could move the field forward.^211,212 A promising future direction is the application of canonical correlation analysis²¹³ and its variants, which can uncover shared patterns across domains that are not only interpretable but can also serve as input features for downstream phenotype prediction.²¹⁴ To further capture directional influences, structural equation modeling²¹⁵ can be integrated to model hypothesized causal pathways and test how upstream factors (e.g., genes, environment) propagate their effects through intermediate phenotypes (e.g., brain function) to shape behavioral outcomes. This causal approach might offer a powerful new avenue in generative modeling²¹² that could ultimately enable virtual interventions, allowing researchers to test “what-if” scenarios without carrying out real-world in vivo experiments. For example, a generative model could simulate how removing environmental risk, improving modifiable behaviors, such as sleep hygiene, or neural simulation could predict downstream effects on alcohol-related behaviors. Achieving this vision will require alcohol researchers to closely work together with AI methodologists and implementation scientists, such as facilitated by Stanford’s AI for Mental Health Initiative.²¹⁶

As mentioned, compared to statistical group analysis that only quantifies a population average trend, a strength of machine learning is to condense multivariate patterns across many features to an individualized score that can be mapped directly onto clinical decisions (e.g., risk of AUD onset or likelihood of relapse). However, none of the reviewed studies discussed the deployment of machine learning approaches in clinical settings. This issue is symptomatic for psychiatry in general, whereas other areas of medicine are beginning to integrate AI models into clinical workflows and even clinical trials.^217,218 Unlike other fields of medicine that rely on measurable physiological indicators, psychiatry still relies heavily on subjective reports and assessment for diagnosis and treatment. This lack of objective, quantifiable targets poses a major barrier to training and validating clinically actionable machine learning models. Consequently, as of March 2026, none of the FDA-approved AI-enabled medical devices were tailored toward psychiatric conditions.^219,220 Unique to the field of alcohol research is that studies soon will be able to replace self-reported alcohol use with real-time quantitative measurements provided by noninvasive alcohol biosensors.^221,222 Such real‐time, quantitative monitoring of alcohol use will provide objective, continuous data that can transform model development and clinical validation.

To fully realize this potential, rigorous evaluation standards must be established. The studies reviewed exhibit substantial variability in validation practices: thus, 18% of studies relied on a single train/test split without cross-validating results on all samples, only 12% of studies validated findings on external data sets, and 16% of articles adopted potentially flawed methodological designs, such as double dipping (e.g., feature selection on the full data set before training) and data leakage (e.g., information of test samples were used for training). The establishment of standardized benchmark data sets and data splits, as well as the preregistration of machine learning analysis and evaluation plans,²²³ including training and testing data construction, can help remedy these concerns. Once quantitative assessments are coupled with stringent model evaluation, researchers will be able to easily test the validity of their models, allowing the field to gain a deeper insight into addiction. Ensuring the responsible deployment of this technology in clinical settings will enable practitioners to diagnose patients based on quantitative markers and objectively assess the progress of alcohol treatments. Finally, prevention programs might be able to accurately determine risk of alcohol misuse in individuals. By doing so, the field of alcohol research could be a trailblazer in psychiatry as it would use AI technology to radically improve the diagnosis and treatment of a psychiatric disease, namely AUD.

Conclusion

This review focused on studies employing machine learning methods that, in contrast to hypothesis-driven analyses, can uncover complex multivariate patterns predictive of alcohol-related outcomes on an individual basis. Most of the 110 reviewed articles trained conventional machine learning approaches on a single data modality to predict alcohol consumption or AUD diagnosis in cohorts with mean ages between 12 and 50 years. Studies were limited by the small number of available samples, not being able to gain insights about causality, and failing to account for confounders. Significantly improving the diagnosis and treatment of AUD will require harnessing the full potential of recent advances in deep learning (such as foundation models).²²⁴ Specifically, fostering multidisciplinary research teams to create rigorous and trustworthy models can help analyze large, multimodal data sets that can capture the multifaceted nature of alcohol use and its comorbidities based on quantitative measures. Additionally, setting guidelines for the responsible development and deployment of these machine learning models would help ensure that these approaches will improve precision alcohol treatment.