NLP Analysis of Weekly Narratives for Dynamic Clinical Assessment in SUD

Overview

This prospective observational study follows adults undergoing residential rehabilitation for severe substance use disorders at a specialized treatment center in Mexico. Participants provide weekly written narratives describing their emotions, challenges, coping strategies, and treatment experiences, and complete validated psychological questionnaires every two weeks, including the Generalized Anxiety Disorder-7 (GAD-7), Environmental Reward Observation Scale (EROS), Automatic Thoughts Questionnaire-8 (ATQ-8), and Behavioral Activation for Depression Scale (BADS). The study applies natural language processing (NLP) and machine learning methods to analyze participants' narratives and identify emotional, cognitive, and behavioral patterns associated with clinical change over time. Narrative-derived features are combined with questionnaire scores to generate a dynamic clinical risk representation that may help detect early signs of psychological worsening or improvement during residential treatment. Participants continue receiving standard residential care, and the study does not modify treatment decisions or clinical interventions. Up to 35 participants with sufficient longitudinal follow-up data will be included in the primary analysis. Data collection is expected to continue through September 2026.

Patients undergoing treatment for substance use disorders frequently describe changes in mood, motivation, hopelessness, cognitive rigidity, emotional distress, and coping strategies through spontaneous written language. These narratives may contain clinically meaningful indicators associated with psychological deterioration, treatment progress, or relapse vulnerability. However, systematic manual analysis of longitudinal written narratives is difficult to implement in routine clinical practice because of the volume and complexity of the data. Recent developments in natural language processing (NLP), representation learning, and deep learning provide methods for extracting quantitative linguistic and semantic information from written text. Integrating these features with repeated psychometric assessments may support the development of longitudinal models capable of characterizing changes in mental health status over time in patients receiving residential addiction treatment. Objectives The objectives of this study are: To extract semantic, emotional, and linguistic features from weekly patient narratives using NLP methods, including sentence embeddings, sentiment and emotion classification, and semantic similarity analyses based on Acceptance and Commitment Therapy (ACT) constructs. To integrate narrative-derived variables with repeated psychometric measures (Generalized Anxiety Disorder-7 (GAD-7), Environmental Reward Observation Scale (EROS), Automatic Thoughts Questionnaire-8 (ATQ-8), and Behavioral Activation for Depression Scale (BADS)) using a multimodal deep learning autoencoder capable of generating a low-dimensional representation of longitudinal clinical status. To evaluate whether latent representations generated by the autoencoder are associated with periods of clinical worsening or clinical improvement across time using leave-one-patient-out cross-validation procedures. To develop a dynamic longitudinal risk representation capable of estimating future changes in automatic negative thoughts, measured through subsequent ATQ-8 scores. Study Design This study is a prospective observational cohort conducted at a single residential rehabilitation center in Mexico. Eligible participants are adults aged 18 years or older with a diagnosis of severe substance use disorder who are admitted for residential treatment. Individuals with active psychotic symptoms or cognitive impairment that substantially interferes with the ability to complete written narratives are excluded. Participants complete: Weekly digital written narratives using open-ended prompts focused on emotions, challenges, coping responses, interpersonal experiences, and perceived treatment progress. Biweekly administration of the following validated self-report instruments: GAD-7 EROS ATQ-8 BADS All information is collected through digital forms integrated into routine clinical monitoring procedures at the treatment center. Participants may contribute data for up to 20 weeks, depending on duration of residential stay. The study began on 25 May 2025, and primary completion is anticipated in September 2026. Up to 35 participants with sufficient longitudinal observations will be included in the primary analytic cohort. NLP and Machine Learning Pipeline Text preprocessing Narratives are written in Spanish and undergo preprocessing procedures that include: minimum length filtering, normalization, consolidation of narrative fields into a single weekly text sample, tokenization and linguistic annotation. Feature extraction Narrative features include: Sentiment classification (positive, neutral, negative) Emotion probabilities for joy, sadness, anger, fear, surprise, and disgust using the pysentimiento library Linguistic variables including: type-token ratio, mean sentence length, proportion of first-person pronouns, proportion of past-tense verbs, obtained through udpipe Semantic similarity measures between patient narratives and ACT-related prototype domains including: experiential avoidance, cognitive fusion, rule-governed behavior, helplessness, achievement orientation, hopefulness Semantic similarity is computed through cosine similarity between sentence-transformer embeddings and predefined prototype centroids. Dimensionality reduction procedures using principal component analysis (PCA) are applied to selected linguistic and semantic variables to derive a principal component representing orientation toward internal emotional experience versus external contextual events. Autoencoder Architecture The multimodal model receives concatenated longitudinal feature vectors after robust scaling. The architecture includes: a bidirectional long short-term memory (BiLSTM) encoder, multi-head attention mechanisms, variational regularization using a β-variational autoencoder (β-VAE) framework, a decoder trained to reconstruct the original temporal feature sequence, and an auxiliary classification component predicting next-period clinical worsening. The training objective combines: mean squared reconstruction error, focal loss for classification, Kullback-Leibler divergence, and optional temporal smoothness regularization. Model performance is evaluated using leave-one-patient-out cross-validation procedures. A final model may subsequently be trained using the complete dataset. Primary Analyses Primary analyses include: evaluation of discriminative performance for prediction of subsequent clinical worsening using: area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), F1 score, Matthews correlation coefficient (MCC) examination of associations between latent trajectory representations and longitudinal psychometric changes, development of a Composite Clinical Risk Index (CCRI) derived from latent representations, and comparison of CCRI trajectories with weeks classified as clinical worsening according to predefined multimodal criteria. Ethics and Dissemination The study has received approval from the institutional review board of Under The Tree Miller A.C. All participants provide written informed consent prior to participation. Participation does not alter or replace standard residential treatment. All clinical decisions remain under the responsibility of treating professionals independent of study procedures. Study findings will be submitted for publication in peer-reviewed scientific journals regardless of outcome. De-identified datasets and analysis code may be made available upon reasonable request and in accordance with institutional and ethical requirements. Recruitment Status Recruitment is ongoing. Final data collection for the primary outcome is anticipated in September 2026.