On-body Sensing of Cocaine Craving, Euphoria and Drug-Seeking Behavior Using Cardiac and Respiratory Signals

Drug addiction is a chronic brain-based disorder that affects a person's behavior and leads to an inability to control drug usage. Ubiquitous physiological sensing technologies to detect illicit drug use have been well studied and understood for different types of drugs. However, we currently lack the ability to continuously and passively measure the user state in ways that might shed light on the complex relationships between cocaine-induced subjective states (e.g., craving and euphoria) and compulsive drug-seeking behavior. More specifically, the applicability of wearable sensors to detect drug-related states is underexplored. In the current work, we take an initial step in the modeling of cocaine craving, euphoria and drug-seeking behavior using electrocardiographic (ECG) and respiratory signals unobtrusively collected from a wearable chest band. Ten experienced cocaine users were studied using a human laboratory paradigm of self-regulated (i.e., "binge") cocaine administration, during which self-reported visual analog scale (VAS) ratings of cocaine-induced subjective effects (i.e., craving and euphoria) and behavioral measures of drug-seeking behavior (i.e., button clicks for drug infusions) are collected. Our results are encouraging and show that self-reported VAS Craving scores are predicted with a normalized root-mean-squared error (NRMSE) of 17.6% and a Pearson correlation coefficient of 0.49. Similarly, for VAS Euphoria prediction, an NRMSE of 16.7% and a Pearson correlation coefficient of 0.73 were achieved. We further analyze the relative importance of different morphology-related ECG and respiratory features for craving and euphoria prediction. Demographic factor analysis reveals how one single factor (i.e., average dollar ($) per cocaine use) can help to further boost the performance of our craving and euphoria models. Lastly, we model drug-seeking behavior using cardiac and respiratory signals. Specifically, we demonstrate that the latter signals can predict participant button clicks with an F1 score of 0.80 and estimate different levels of click density with a correlation coefficient of 0.85 and an NRMSE of 17.9%.