Abstract
AIMS: A machine learning approach to identify a multidimensional signature associated with relapse and long-term outcome in alcohol dependence treatment. DESIGN: In this observational naturalistic study, inpatients with alcohol dependence received qualified detoxification plus CBT (Cognitive Behavioral Therapy) and were followed up 6-months after discharge to assess abstinence and drinking behavior. Cross-validated multivariate sparse partial least squares analysis (SPLS) was used to investigate the relationship between clinical features and four long-term outcome variables. SETTING: Germany. PARTICIPANTS: 152 patients (on average 47.8 years old, 72% male) with alcohol dependence, who received inpatient qualified detoxification plus CBT. MEASUREMENTS: 35 clinical features were used to cover all three phases of inpatient treatment (pre-, within-, post-treatment). Among these, sociodemographic characteristics, ICD-10 psychiatric diagnoses, previous detoxification treatments, and somatic measurements as well as inpatient treatment setting such as withdrawal medication, liver ultrasound, further information about the patients´ stay, and post-inpatient care were assessed. The four outcome dimensions included: continuous abstinence, abstinence at follow up, daily alcohol consumption, and days of abstinence after discharge. FINDINGS: Six months after withdrawal treatment 46% of the patients achieved continuous abstinence. Socioeconomic, clinical and somatic features across the treatment timeline were analyzed and summarized into a multivariate signature associated with long-term treatment outcome. Thereby, the SPLS algorithm identified regular completion of withdrawal treatment, higher education, and employment status to be most strongly associated with a positive outcome. Alcohol-related hepatic and hematopoietic damage, number of previous withdrawal treatments and living in a shelter were most profoundly associated with a negative outcome. CONCLUSION: Conceiving treatment outcome as a multidimensional signature and moving beyond simple binary classifications of relapse versus abstinence may improve the understanding of relapse pathways and support more individualized treatment strategies.