Unsupervised Deep Representation Learning for Clinical Stratification in Substance Use Disorders
Unsupervised Deep Representation Learning for Clinical Stratification in Substance Use Disorders
Substance use disorders (SUDs) show considerable clinical heterogeneity that limits the usefulness of traditional categorical diagnoses. This observational, cross-sectional study aims to apply an unsupervised deep learning method - an autoencoder - to learn continuous latent representations from standardised psychometric data and to explore whether those representations can help stratify clinical subpopulations. The investigators will recruit 155 adults undergoing residential treatment for SUD. Participants will complete six validated instruments assessing impulsivity (BIS-11), anger regulation (STAXI-2), behavioural activation/avoidance (BADS), borderline symptomatology (BSL-23), generalised anxiety (GAD-7), and environmental reward (EROS). Demographic and clinical variables (age, sex, primary substance, years of use, prior treatments) will also be recorded.
After data cleaning and standardisation (z-scores), a symmetric autoencoder with a 12-dimensional bottleneck (architecture 21-32-24-12-24-32-21) will be trained using mean squared error loss. Regularisation includes L2 weight decay and dropout. The model will be trained 30 times with different random seeds to assess stability; the five best models (by validation pseudo-R²) will be combined into a weighted ensemble. Five-fold cross-validation will evaluate generalisation. For comparison, principal component analysis (PCA) will be applied to the same data. Gaussian mixture models (GMM) will be fitted on the latent space to explore potential clinical subgroups.
The primary outcome is the stability of the latent representation (coefficient of variation of validation MSE across runs). Secondary outcomes include reconstruction performance (pseudo-R²) of the ensemble, comparison with PCA, and the interpretability of latent dimensions via correlations with original variables. GMM results will be described using BIC, silhouette width, bootstrap stability, and clinical characterisation of clusters.
This study does not involve any intervention. Results will be hypothesis-generating and require external validation. No automated clinical decisions will be made.
Substance use disorders (SUDs) are characterised by substantial heterogeneity in clinical presentation, behavioural patterns, emotional regulation difficulties, impulsivity, and treatment response. Individuals with the same categorical diagnosis may differ considerably in symptom severity, comorbid psychopathology, and psychosocial functioning. This variability limits the explanatory value of traditional diagnostic classifications and supports the development of dimensional and data-driven approaches for patient characterisation.
Recent advances in machine learning provide methods capable of identifying latent structures within complex clinical datasets. Autoencoders, a form of unsupervised deep learning, can learn compact nonlinear representations of multidimensional data while preserving relevant information from the original variables. Compared with traditional linear dimensionality reduction methods such as principal component analysis (PCA), autoencoders may better capture complex interactions among psychological and behavioural variables. When combined with probabilistic clustering approaches such as Gaussian mixture models (GMM), these latent representations may facilitate the identification of clinically meaningful patient subgroups.
The purpose of this observational study is to apply an autoencoder model to psychometric and clinical data obtained from adults receiving residential treatment for substance use disorders. The study aims to explore latent dimensions underlying symptom and behavioural variability and to evaluate whether these dimensions support stable subgroup identification.
Primary Objective:
To learn a 12-dimensional latent representation from standardised psychometric and clinical variables using an autoencoder model and evaluate the stability of this representation across repeated training procedures.
Secondary Objectives:
To compare the reconstruction performance of the autoencoder with principal component analysis (PCA).
To characterise the clinical meaning of the latent dimensions through correlations with the original variables.
To explore potential patient subgroups using Gaussian mixture models (GMM) applied to the latent space.
To assess the stability and interpretability of the identified subgroups.
Study Design:
This is a single-centre, observational, cross-sectional, non-interventional study conducted in a residential addiction treatment facility. Recruitment is planned from February 2024 through December 2025. The study is registered prior to dissemination of results.
Study Population:
Approximately 155 adults diagnosed with substance use disorder according to DSM-5 criteria will be included. Eligible participants must be 18 years of age or older, currently receiving residential treatment, capable of completing study questionnaires, and willing to provide written informed consent.
Participants with active psychotic disorders, severe cognitive impairment, significant language or literacy barriers, or imminent discharge from treatment will be excluded.
Measures and Data Collection:
Participants will complete a battery of validated self-report instruments assessing impulsivity, anger regulation, behavioural activation and avoidance, borderline symptomatology, anxiety, and environmental reward. Additional demographic and clinical variables will include age, sex, primary substance of use, years of substance use, and prior treatment history.
Questionnaires include:
Barratt Impulsiveness Scale (BIS-11) State-Trait Anger Expression Inventory-2 (STAXI-2) Behavioral Activation for Depression Scale (BADS) Borderline Symptom List-23 (BSL-23) Generalized Anxiety Disorder-7 (GAD-7) Environmental Reward Observation Scale (EROS)
Data Analysis:
Clinical variables will be standardised prior to analysis. Missing values are expected to be minimal and will be handled using median imputation procedures. Redundant variables with excessive multicollinearity may be removed before modelling.
An autoencoder neural network will be trained to generate a reduced latent representation of the clinical data. Model performance and stability will be evaluated across repeated training runs and cross-validation procedures. Reconstruction accuracy will be compared with PCA using equivalent dimensionality.
The resulting latent space will subsequently be analysed using Gaussian mixture models to explore potential patient subgroups. Model selection will consider statistical fit, cluster stability, and clinical interpretability. Correlations between latent dimensions and original clinical variables will be examined to facilitate interpretation of the learned representations.
Ethical Considerations:
The study protocol has been approved by the corresponding Institutional Ethics Committee. All participants will provide written informed consent prior to participation. Data will be anonymised after collection, and no direct identifiers will be retained.
This study is observational and will not modify routine clinical treatment. No automated clinical decisions will be made based on model outputs. Participants may experience mild emotional discomfort or fatigue while completing questionnaires; psychological support will be available if needed.
The study will be conducted in accordance with the Declaration of Helsinki and applicable local ethical regulations.
Dissemination:
Results will be submitted for publication in peer-reviewed scientific journals and presented at academic conferences. De-identified data and analysis code may be shared publicly after publication to support transparency and reproducibility.
Inclusion Criteria:
Exclusion Criteria: