Substance use, e.g. of opioid, cocaine, methamphetamine, and alcohol, has skyrocketed in the US, and
has been declared a public health crisis. Substance use disorder (SUD) leads to a significant impairment
in physical, mental, social, and occupational functioning, coupled with an inability to reduce or control use
and the development of tolerance and withdrawal symptoms. Like any other addiction, the precursor to
substance use episodes are cravings moments, and hence, they will be opportune moments to intervene
if it is possible to detect them. We have the potential to characterize patients’ conditions by combining multimodal sensor measurement in the laboratory like fMRI and clinical assessment, and field measurement via wearable and ubiquitous sensing. However, there are still many challenges to be solved to leverage these multimodal data to extract clinically useful and actionable insights. The proposed research will be executed via two aims:
Aim 1) Develop a multidimensional database from patients with SUD, via laboratory brain and physiological measurements and field physiological and behavioral measurements. We will use fMRI to measure ground truth of craving responses, while recording physiological signature during craving inducing or emotional cues. We will also collect physiological and behavioral responses using mobile phones and wearable sensors in the field. We will analyze the relationships among brain-biobehavioral responses and craving and affective profiles in laboratory and field settings.
Aim 2) Develop robust multimodal data science methods for craving and affective profile detection
and characterization: We will design 1) new approaches that positively transfer knowledge from multiple modalities to fewer modalities: we will combine data from different sensor sources while guaranteeing the model can be run with fewer sensors during testing, 2) craving and affect profile detection using regular/irregular characteristics and sparse weak labels, 3) craving and affect detection models with unlabeled data and a limited number of noisy labels, and 4) personalized craving and affect detection models with unlabeled data. We will evaluate our proposed data science approach using the data collected previously and in non-patient and patient pilot studies.
