Empirical Derivation and Prediction of Treatment Trajectories in Harmonized AUD Clinical Trial Datasets

Abstract

In clinical settings targeting alcohol use disorder (AUD), it is often unclear whether a treatment option may best suit a patient's clinical needs. Clinicians providing AUD treatment are often required to predict patients' responses to guide treatment decisions. Recently, machine learning approaches have been used as tools in precision medicine to help guide these clinical decisions. However, the extent of their clinical utility in populations undergoing treatment is largely unknown. Using data from four Phase 2 randomized clinical trials affiliated with the NIAAA Clinical Investigations Group and a Phase 3 trial sponsored by the NIAAA, we developed a machine learning model to predict treatment response phenotypes derived from clustering drinking rates at the end of treatment. Harmonized data included demographics and baseline data from biological and clinical assessments. Follow-up analyses were performed to characterize treatment response phenotypes. Three clusters corresponding to mild (M_SDU = 1.3), moderate (M_SDU = 6.70) and severe (M_SDU = 15.3) alcohol consumption were identified from end-of-treatment drinking data. Performance of the tree-based classifier using out-of-sample test data was 71% with baseline drinking included and 61% without. Exploratory analyses revealed participants clustered as mild drinkers showed reductions in drinking across treatment (M_Difference = −0.731, SE = 0.114, p < 0.001) whereas participants clustered as severe had escalation in use (M_Difference = 6.82, SE = 0.52, p < 0.001). Although males drank more than females at baseline (M_Difference = 1.46, SE = 0.287, p < 0.001), no significant differences in consumption emerged at the end of treatment. Findings from this work indicate that alcohol use derived from patterns of consumption at the end of treatment maps onto unique treatment response trajectories for mild and severe forms of AUD. Furthermore, the identified clusters revealed sex-specific differences in alcohol consumption patterns across different phases of treatment. Overall, this highlights the utility of computational methods for deriving clinically meaningful AUD-related phenotypes across multiple studies, each with different treatments and participant characteristics.