Research

Predicting Outcomes of Internet-Based Interventions for Adult Problem Drinking using Machine Learning

Background: Problem drinking is a significant public health concern, and face-to-face brief interventions have shown effectiveness in curbing it. However, limited implementation in routine care and the community presents a treatment gap. Internet-based interventions (iAIs) could be a potential solution to overcome this gap. Previous studies have investigated the effectiveness of iAIs in treating adult problem drinking, and it is essential to explore the possibility of predicting treatment outcomes using machine learning algorithms.

Objectives: The primary objective of this project is to develop a machine learning model to predict treatment outcomes in internet-based interventions for adult problem drinking. Specifically, the model will aim to predict mean weekly alcohol consumption in standard units (SUs) and treatment response (TR), defined as less than 14/21 SUs for women/men weekly.

Methodology: The proposed project will use the data from the study mentioned above, where a one-stage individual patient data meta-analysis (IPDMA) was conducted with a linear mixed model complete-case approach, using baseline and first follow-up data. The project will follow the following methodology:

Data Preprocessing: The raw data from the study will be preprocessed by cleaning, feature engineering, and transformation to make it suitable for machine learning algorithms. Feature Selection: The relevant features for predicting the treatment outcomes will be selected. Model Selection: Various machine learning models will be evaluated (cross-validation), and the best-performing model will be selected. Hyperparameter Tuning: The hyperparameters of the models will be tuned to achieve optimal performance. Prediction: The final model will be used to predict the treatment outcomes in a test-set.

Expected Results: The expected results of the project are a machine learning model that can predict treatment outcomes in internet-based interventions for adult problem drinking. The model can aid healthcare professionals in identifying patients who are likely to benefit from iAIs and help researchers in designing more effective iAIs.

Conclusion: The proposed project aims to develop a machine learning model that can predict treatment outcomes in internet-based interventions for adult problem drinking. The project builds on the existing study’s findings and will use state-of-the-art machine learning algorithms to achieve the objectives. The results of the project will be valuable in improving the effectiveness of iAIs and addressing the treatment gap for problem drinking.

MLAUD: A Machine Learning-Based Approach to Predict Post-Treatment Drinking Behavior in Patients with Alcohol Use Disorder

Efficient clinical practice requires tools that can accurately predict post-treatment symptom trajectories, thereby helping patients and clinicians make informed decisions regarding the type, intensity, and treatment duration. Despite a large body of research (Adamson et al., 2009) on associations between specific predictors and post-treatment outcomes in alcohol use disorder (AUD), these inference statistic-based models lack the ability to provide adequate predictions for individual patients.

The recent integration of machine learning (ML) algorithms in the clinical research domain has demonstrated the efficiency and reliability of ML-based models for various applications, such as research in AUD populations (Mak et al., 2019). However, it remains to be investigated how ML algorithms can be used to improve individual, post-treatment outcome prediction for patients with AUD. The proposed project aims to address this clinically relevant gap through two phases:

(i) Develop and train ML algorithms to generate valid and reliable post-treatment outcome prediction models for AUD patients’ drinking behavior. Secondary analysis of longitudinal data—agregated from various clinical studies—will be used to develop these computational models.

(ii) Validate the developed ML models’ clinical applicability on a newly collected, independent sample of patients with AUD (N >= 250).

The proposed project will provide much-needed prognostic models to improve AUD patient care in favor of informed, evidence-based, and individualized treatment planning in clinical settings. This research will further generate insights into identifying critical predictors of post-treatment outcomes using data-driven approaches and provide valuable insights into clinical applications of predictive computational tools across mental disorders in general.

A mediation analysis on the relationship between antidepressants, depressive symptoms, and alcohol use disorder

Background: Antidepressants (AD) are commonly prescribed to treat depressive symptoms in patients with alcohol use disorder (AUD), although the evidence of their effect on alcohol use is inconsistent. It is suspected that, besides the direct effects of ADs on alcohol use, the effects are partially related to their modulation of depressive symptoms. In total, these two pathways may mutually mask each other.

Methods: Observations of 153 detoxified, recently abstinent patients (18-60 years of age; 27.5% female) with AUD attending a twelve-week residential treatment program were included in the study. In a mediation model the direct and by altered depressive symptoms mediated effect, of ADs on alcohol use was estimated. Change in alcohol use was assessed by the percentage of days abstinent (PDA) 3-month before and after residential treatment. Depressive symptoms were assessed with the Brief Symptom Checklist (BSCL) at admission and discharge from treatment.