Presentation
DH4 - Assessing Alcohol Use Disorder: Insights From Lifestyle, Background, and Family History With Machine Learning Techniques
SessionPoster Session 1
DescriptionIntroduction
The number of alcohol consumers remains substantial, as do the associated consequences. In 2023, around 85% of U.S. adults aged 18 and above reported alcohol consumption. That same year, 28 million people developed Alcohol Use Disorder (AUD), 4 million emergency department visits were alcohol-related, and around 178k deaths were attributed to alcohol use. The situation among underage populations regarding alcohol use is particularly alarming. Around 10.6 million teenagers aged 12 to 20 reported drinking in the past year, and 5.6 million in the past month. Furthermore, 663k engaged in heavy alcohol use during the past month. Alcohol use in teenagers is reported to affect normal brain development and increase the risk of AUD. Alcohol use during this formative time can disrupt these developmental processes, potentially causing long-term changes in brain function and structure. Given the challenges teenagers face with self-regulation and decision-making, factors such as family background and social interactions play significant roles in shaping their behaviors regarding alcohol use, which is crucial in preventing AUD among this demographic.
AUD is a chronic, relapsing brain disorder marked by an inability to regulate alcohol consumption, even when it leads to negative social, work-related, or health outcomes. It is associated with various symptoms, including intense cravings for alcohol, difficulty controlling intake, increased tolerance (requiring more alcohol to achieve the same effect), and withdrawal symptoms when alcohol use is reduced or stopped. AUD is affected by a variety of genetic, environmental, psychological, and social factors. Thus, it is crucial to comprehend how each of these factors contributes to AUD and to recommend various strategies or approaches for prevention associated with each factor.
In this work, we aimed to explore various factors, including individuals’ background information, lifestyle choices, and family history of AUD. Our analysis utilizes real-world survey data collected by The All of Us Research Program from the National Institutes of Health (NIH). We analyzed the data to examine how each factor influences alcohol consumption through classical machine learning techniques. The specific objectives of this work are outlined below.
• Identify the most influential determinants of AUD to understand how lifestyle, background, and family history contribute to an individual's risk of developing AUD. This can enable parents and educators to guide students with strategies to mitigate their AUD development. Similarly, adults may modify their behaviors or environments to minimize their own risk from AUD.
• Predict an individual's likelihood of developing AUD based on these factors. This approach enables a proactive prediction of the individual’s risk of developing AUD, allowing for early intervention and targeted prevention.
Approaches
Data: The data used in this work is from the All of Us Research Program’s Registered Tier Dataset V7. This dataset includes surveys, electronic health records, physical measurements, and genetic samples exploring the impact of biology, environment, and lifestyle on health. Three surveys were selected in this study with 33 factors (i.e. features) extracted: 1) The Basics for demographic information; 2) Lifestyle for participants’ use of tobacco, alcohol, and recreational drugs and 3) Personal and Family Health History with information about participants past medical history and those of their family members. Sex at birth and gender identity were combined into one feature due to their identical responses. Additionally, the question, "In your entire life, have you had at least one drink?" received 99.6% "yes" responses, leading to its exclusion. In total, 31 features were included in our study. We focused on participants who completed all three surveys in full (6,016 participants), among which 1,442 (24%) have experienced AUD, while 4,574 (76%) have never had AUD in their lifetime.
Models: To understand the impact of lifestyle, background, and family history on the development of AUD (Objective 1), the importance of features was analyzed using decision trees due to their ability to naturally rank features based on their contribution to the model. Top factors were examined individually using data visualization and the Chi-Square Test of Independence. Afterward, three classification approaches were employed to predict an individual’s likelihood of developing AUD: 1) Decision Trees; 2) Random Forests; and 3) Gaussian Naive Bayes. To ensure sufficient information for feature analysis, we used 80% of the data for training and the remaining 20% for testing. In addition to the standard train-test split (i.e., hold-out validation, which splits the data once), we also perform classification using 5-fold cross-validation, reducing the risk of overfitting to a specific data split.
Findings
Our analysis reveals that the most significant factors associated with AUD are individual annual income (0.0882), recreational drug use (0.0773), and the AUD status of the participant’s grandparents (0.0698) (belonging to health history).
We find that the factors from health history (i.e., AUD-Grandparent, AUD-Father, AUD-Mother, AUD-Sibling, AUD-Son, AUD-Daughter) do not exhibit a linear relationship with AUD. However, the importance of most of these features is high, such as AUD Grandparent, AUD-Mother, AUD-Sibling, and AUD-Son. This may point to a non-linear relationship, where the effect of family history on AUD may not follow a simple, proportional pattern. Instead, it likely interacts with other factors, such as the individual’s environment, upbringing, or coping mechanisms, which may amplify or mitigate the development of AUD.
We also find the other top factors from The Basics Survey exhibit a comparably strong linear relationship with AUD.
• Annual Income: individuals with a lower annual income are more likely to develop AUD compared with others with a higher annual income.
• Recreation drug use: individuals with no history of recreational drug use have the lowest likelihood of developing AUD compared to those who have used recreational drugs.
• Living years: there are no significant differences in AUD ratios across residence lengths, but participants living at their current address for less than one year have the highest AUD ratio.
• Sex: females are less likely to develop AUD compared with males.
• Marriage status: individuals who are never married or separated have a higher likelihood of developing AUD compared to those who are married or widowed.
• Education level: individuals with lower educational attainment have a higher likelihood of developing AUD compared to those with a college degree or higher.
The prediction results indicate that Random Forests outperformed the other two machine learning techniques. The model achieves an accuracy of 81% which performs well in identifying cases without AUD but still struggles to accurately predict the cases with AUD.
The imbalance is likely due to the data disparity (284 with AUD, 920 without AUD). Addressing this, along with model optimization, could improve minority class performance. One effective approach is to use cross-validation, which led to improved performance for all three methods, achieving 82% accuracy for random forests. However, further improvement is needed, and future work could explore resampling and stratified cross-validation.
Takeaways
By using Decision Trees, we first identified the most important features influencing alcohol use disorder (AUD): annual income, recreational drug use, length of residence, sex/gender, marital status, education level, and family history of AUD. This information can be valuable for parents, healthcare professionals, and educators in guiding individuals to reduce the risk of developing AUD before it becomes a serious issue. We then utilized data visualization and Chi-Square Tests of Independence to analyze the associations between individual factors and AUD, aiming to enhance our understanding of these relationships. Last but not least, we applied three different machine learning techniques to predict whether an individual has developed AUD based on 31 factors from lifestyle, personal information, and family history of AUD. Among these methods, random forests achieved the highest accuracy (82%), compared to Decision Trees and Naive Bayes.
The number of alcohol consumers remains substantial, as do the associated consequences. In 2023, around 85% of U.S. adults aged 18 and above reported alcohol consumption. That same year, 28 million people developed Alcohol Use Disorder (AUD), 4 million emergency department visits were alcohol-related, and around 178k deaths were attributed to alcohol use. The situation among underage populations regarding alcohol use is particularly alarming. Around 10.6 million teenagers aged 12 to 20 reported drinking in the past year, and 5.6 million in the past month. Furthermore, 663k engaged in heavy alcohol use during the past month. Alcohol use in teenagers is reported to affect normal brain development and increase the risk of AUD. Alcohol use during this formative time can disrupt these developmental processes, potentially causing long-term changes in brain function and structure. Given the challenges teenagers face with self-regulation and decision-making, factors such as family background and social interactions play significant roles in shaping their behaviors regarding alcohol use, which is crucial in preventing AUD among this demographic.
AUD is a chronic, relapsing brain disorder marked by an inability to regulate alcohol consumption, even when it leads to negative social, work-related, or health outcomes. It is associated with various symptoms, including intense cravings for alcohol, difficulty controlling intake, increased tolerance (requiring more alcohol to achieve the same effect), and withdrawal symptoms when alcohol use is reduced or stopped. AUD is affected by a variety of genetic, environmental, psychological, and social factors. Thus, it is crucial to comprehend how each of these factors contributes to AUD and to recommend various strategies or approaches for prevention associated with each factor.
In this work, we aimed to explore various factors, including individuals’ background information, lifestyle choices, and family history of AUD. Our analysis utilizes real-world survey data collected by The All of Us Research Program from the National Institutes of Health (NIH). We analyzed the data to examine how each factor influences alcohol consumption through classical machine learning techniques. The specific objectives of this work are outlined below.
• Identify the most influential determinants of AUD to understand how lifestyle, background, and family history contribute to an individual's risk of developing AUD. This can enable parents and educators to guide students with strategies to mitigate their AUD development. Similarly, adults may modify their behaviors or environments to minimize their own risk from AUD.
• Predict an individual's likelihood of developing AUD based on these factors. This approach enables a proactive prediction of the individual’s risk of developing AUD, allowing for early intervention and targeted prevention.
Approaches
Data: The data used in this work is from the All of Us Research Program’s Registered Tier Dataset V7. This dataset includes surveys, electronic health records, physical measurements, and genetic samples exploring the impact of biology, environment, and lifestyle on health. Three surveys were selected in this study with 33 factors (i.e. features) extracted: 1) The Basics for demographic information; 2) Lifestyle for participants’ use of tobacco, alcohol, and recreational drugs and 3) Personal and Family Health History with information about participants past medical history and those of their family members. Sex at birth and gender identity were combined into one feature due to their identical responses. Additionally, the question, "In your entire life, have you had at least one drink?" received 99.6% "yes" responses, leading to its exclusion. In total, 31 features were included in our study. We focused on participants who completed all three surveys in full (6,016 participants), among which 1,442 (24%) have experienced AUD, while 4,574 (76%) have never had AUD in their lifetime.
Models: To understand the impact of lifestyle, background, and family history on the development of AUD (Objective 1), the importance of features was analyzed using decision trees due to their ability to naturally rank features based on their contribution to the model. Top factors were examined individually using data visualization and the Chi-Square Test of Independence. Afterward, three classification approaches were employed to predict an individual’s likelihood of developing AUD: 1) Decision Trees; 2) Random Forests; and 3) Gaussian Naive Bayes. To ensure sufficient information for feature analysis, we used 80% of the data for training and the remaining 20% for testing. In addition to the standard train-test split (i.e., hold-out validation, which splits the data once), we also perform classification using 5-fold cross-validation, reducing the risk of overfitting to a specific data split.
Findings
Our analysis reveals that the most significant factors associated with AUD are individual annual income (0.0882), recreational drug use (0.0773), and the AUD status of the participant’s grandparents (0.0698) (belonging to health history).
We find that the factors from health history (i.e., AUD-Grandparent, AUD-Father, AUD-Mother, AUD-Sibling, AUD-Son, AUD-Daughter) do not exhibit a linear relationship with AUD. However, the importance of most of these features is high, such as AUD Grandparent, AUD-Mother, AUD-Sibling, and AUD-Son. This may point to a non-linear relationship, where the effect of family history on AUD may not follow a simple, proportional pattern. Instead, it likely interacts with other factors, such as the individual’s environment, upbringing, or coping mechanisms, which may amplify or mitigate the development of AUD.
We also find the other top factors from The Basics Survey exhibit a comparably strong linear relationship with AUD.
• Annual Income: individuals with a lower annual income are more likely to develop AUD compared with others with a higher annual income.
• Recreation drug use: individuals with no history of recreational drug use have the lowest likelihood of developing AUD compared to those who have used recreational drugs.
• Living years: there are no significant differences in AUD ratios across residence lengths, but participants living at their current address for less than one year have the highest AUD ratio.
• Sex: females are less likely to develop AUD compared with males.
• Marriage status: individuals who are never married or separated have a higher likelihood of developing AUD compared to those who are married or widowed.
• Education level: individuals with lower educational attainment have a higher likelihood of developing AUD compared to those with a college degree or higher.
The prediction results indicate that Random Forests outperformed the other two machine learning techniques. The model achieves an accuracy of 81% which performs well in identifying cases without AUD but still struggles to accurately predict the cases with AUD.
The imbalance is likely due to the data disparity (284 with AUD, 920 without AUD). Addressing this, along with model optimization, could improve minority class performance. One effective approach is to use cross-validation, which led to improved performance for all three methods, achieving 82% accuracy for random forests. However, further improvement is needed, and future work could explore resampling and stratified cross-validation.
Takeaways
By using Decision Trees, we first identified the most important features influencing alcohol use disorder (AUD): annual income, recreational drug use, length of residence, sex/gender, marital status, education level, and family history of AUD. This information can be valuable for parents, healthcare professionals, and educators in guiding individuals to reduce the risk of developing AUD before it becomes a serious issue. We then utilized data visualization and Chi-Square Tests of Independence to analyze the associations between individual factors and AUD, aiming to enhance our understanding of these relationships. Last but not least, we applied three different machine learning techniques to predict whether an individual has developed AUD based on 31 factors from lifestyle, personal information, and family history of AUD. Among these methods, random forests achieved the highest accuracy (82%), compared to Decision Trees and Naive Bayes.
Event Type
Poster Presentation
TimeMonday, March 314:45pm - 6:15pm EDT
LocationFrontenac Foyer
Digital Health (DH)
Simulation and Education (SE)
Hospital Environments (HE)
Medical and Drug Delivery Devices (MDD)
Patient Safety and Research Initiatives (PS)